diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_tensor([7447, 7447])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_tensor([7447, 7447])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d91c45641a2feeedfd6937030d815fb0e719f3b2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_tensor([7447, 7447])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Through a rapid-cut montage of macro close-ups, the film captures, with an almost scientific precision, the entire process of a droplet of ink, thick as night, plump and plump, plump black ink plunging into still, pure water. The sequence begins with an extreme close-up shot from directly above, showing the plump, jet-black droplet suspended above the water's surface. It then silently touches the surface, stirring a tiny, perfectly formed ripple. The camera then shifts to a horizontal profile, showing the moment the droplet pierces the surface, sinking like a miniature depth charge, leaving a slender, straight black trail. Then, an even closer close-up captures the body of the droplet's sudden yet graceful unfurling within the water, countless silky black tendrils spreading out in all directions to form a complex, evolving spherical cloud. Illuminated by a bright, even backlight, these inky filaments reveal breathtaking detail, some as dense as smoke, others as thin as gauze, interweaving into magnificent patterns reminiscent of cosmic nebulae or fractal geometry. The entire process takes place in absolute silence, with the dominant colors of the painting being extreme black and white, creating a visual atmosphere that is both chaotic and tranquil, full of Zen and a sense of abstract art.", + "seed": 24478, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_7447_seed_24478_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24602587521076202, + "vlm_logprob": -0.24602587521076202, + "vlm_yes_no_margin": 3.5249999880790712, + "vlm_token_credit": -0.24602587521076202, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1854629516601562 + ], + "opposite_logprob_tokens": [ + -0.4354628920555115 + ], + "logprob_reward": -1.1854629516601562, + "yes_no_margin_reward": -0.7500000596046448, + "token_credit_reward": -1.1854629516601562, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15566036105155945, + 0.22022724151611328, + 0.1379876285791397, + 0.06200823560357094, + 0.013296498917043209, + 0.11987859755754471, + 0.09203632175922394, + 0.07680171728134155, + 0.037643518298864365, + 0.0844598188996315 + ], + "credit_map": [ + 0.00641942722722888, + 0.005436891224235296, + 0.05916449427604675, + 0.007434500381350517, + 0.00572167569771409, + 0.005477421451359987, + 0.005545866675674915, + 0.060460109263658524, + 0.0076958900317549706, + 0.0020163736771792173, + 0.09580819308757782, + 0.006664694286882877, + 0.002351408591493964, + 0.004777530208230019, + 0.0037907168734818697, + 0.09712245315313339, + 0.010134191252291203, + 0.0047660949639976025, + 0.04240323603153229, + 0.00527732539921999, + 0.006458452437072992, + 0.03962603583931923, + 0.004060995765030384, + 0.025261320173740387, + 0.008098691701889038, + 0.005340042524039745, + 0.011258437298238277, + 0.006086251698434353, + 0.006067727226763964, + 0.009993062354624271, + 0.0070652742870152, + 0.008098757825791836, + 0.0019886482041329145, + 0.0011578495614230633, + 0.0017682071775197983, + 0.001326433615759015, + 0.0014479331439360976, + 0.0021185213699936867, + 0.0013602887047454715, + 0.0021286197006702423, + 0.01435067132115364, + 0.022161373868584633, + 0.006256974767893553, + 0.0064322855323553085, + 0.03395655006170273, + 0.0204437505453825, + 0.006912977434694767, + 0.009364046156406403, + 0.011353795416653156, + 0.015129257924854755, + 0.003750887466594577, + 0.005483656190335751, + 0.0265168659389019, + 0.01796088181436062, + 0.00566060421988368, + 0.006180385127663612, + 0.0069484286941587925, + 0.008747028186917305, + 0.0042895544320344925, + 0.0031716548837721348, + 0.02026086486876011, + 0.023065149784088135, + 0.005086577031761408, + 0.005232466850429773, + 0.004924315493553877, + 0.005447072442620993, + 0.0035140749532729387, + 0.004223871976137161, + 0.004932767245918512, + 0.004956372547894716, + 0.005057851783931255, + 0.0045871976763010025, + 0.010224824771285057, + 0.019649671390652657, + 0.0067186676897108555, + 0.005287644453346729, + 0.012106116861104965, + 0.010808516293764114, + 0.007186967879533768, + 0.012477420270442963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -1.1854629516601562 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008682594634592533 + ], + "opposite_logprob_tokens": [ + -4.758682727813721 + ], + "logprob_reward": -0.008682594634592533, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008682594634592533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10522248595952988, + 0.5758118629455566, + 0.05953129380941391, + 0.041908979415893555, + 0.05562352016568184, + 0.034476134926080704, + 0.0331667996942997, + 0.017916454002261162, + 0.06214837357401848, + 0.014194125309586525 + ], + "credit_map": [ + 0.005115250125527382, + 0.005165643524378538, + 0.03665550798177719, + 0.006199429742991924, + 0.004472477361559868, + 0.005218311678618193, + 0.005096222274005413, + 0.03729963302612305, + 0.021201178431510925, + 0.004579373635351658, + 0.2513250410556793, + 0.017355764284729958, + 0.0049812947399914265, + 0.01144187431782484, + 0.009730461984872818, + 0.25519680976867676, + 0.004210492130368948, + 0.0023132802452892065, + 0.018706494942307472, + 0.002564159920439124, + 0.0030242677312344313, + 0.015495280735194683, + 0.0018412263598293066, + 0.011376085691154003, + 0.005450798664242029, + 0.006818633060902357, + 0.0051902588456869125, + 0.003439632710069418, + 0.006372516043484211, + 0.00524665042757988, + 0.0037093476857990026, + 0.005681135226041079, + 0.01358901895582676, + 0.0054130395874381065, + 0.004867407958954573, + 0.0027390941977500916, + 0.007274223957210779, + 0.014382562600076199, + 0.0020784903317689896, + 0.0052796779200434685, + 0.0046753063797950745, + 0.005801314488053322, + 0.001434503821656108, + 0.001358048408292234, + 0.010512173175811768, + 0.006641617976129055, + 0.0019869457464665174, + 0.0020662236493080854, + 0.004339286126196384, + 0.005963347852230072, + 0.0014693823177367449, + 0.0019557105842977762, + 0.009188548661768436, + 0.005962464492768049, + 0.0017898855730891228, + 0.00249817268922925, + 0.0022962661460042, + 0.0020073787309229374, + 0.0012383639113977551, + 0.0009743296541273594, + 0.0043080649338662624, + 0.003943698480725288, + 0.0016252360073849559, + 0.0015231141587719321, + 0.009373005479574203, + 0.005146510899066925, + 0.008630125783383846, + 0.005890614353120327, + 0.005771944299340248, + 0.008278731256723404, + 0.008510376326739788, + 0.010547058656811714, + 0.0018176203593611717, + 0.0028628651052713394, + 0.000999802490696311, + 0.0010064663365483284, + 0.001971500227227807, + 0.0023480961099267006, + 0.0012535940622910857, + 0.0019341803854331374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.008682594634592533 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262095481157303 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005262095481157303, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262095481157303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04199514910578728, + 0.4377553462982178, + 0.20766456425189972, + 0.03624441474676132, + 0.03791341930627823, + 0.09159807860851288, + 0.065338134765625, + 0.04795655235648155, + 0.018347160890698433, + 0.015187127515673637 + ], + "credit_map": [ + 0.001840200973674655, + 0.003983275033533573, + 0.010811583139002323, + 0.0033844420686364174, + 0.00325719709508121, + 0.0038251488003879786, + 0.0030275466851890087, + 0.011865753680467606, + 0.01682998426258564, + 0.0034345195163041353, + 0.19114820659160614, + 0.013112716376781464, + 0.0033802331890910864, + 0.008730222471058369, + 0.007271043490618467, + 0.1938484162092209, + 0.010848219506442547, + 0.0041328370571136475, + 0.08165424317121506, + 0.006545695476233959, + 0.004362962674349546, + 0.01030731201171875, + 0.004891580436378717, + 0.0849217027425766, + 0.005463700741529465, + 0.0026069460436701775, + 0.005800746381282806, + 0.003465212183073163, + 0.0034011290408670902, + 0.006242451723664999, + 0.0037057995796203613, + 0.005558430217206478, + 0.006780122872442007, + 0.002721888478845358, + 0.005156606435775757, + 0.003914442379027605, + 0.004574393387883902, + 0.0070481328293681145, + 0.0023441037628799677, + 0.005373730324208736, + 0.013838956132531166, + 0.014605517499148846, + 0.004976235330104828, + 0.004586859606206417, + 0.025427421554923058, + 0.016803022474050522, + 0.005527178291231394, + 0.005832891445606947, + 0.010390223003923893, + 0.009541274979710579, + 0.004999204538762569, + 0.006625643465667963, + 0.01176227442920208, + 0.009785767644643784, + 0.006477906834334135, + 0.005755840800702572, + 0.004452217370271683, + 0.00554134976118803, + 0.0025249014142900705, + 0.002629608381539583, + 0.010436196811497211, + 0.01571182906627655, + 0.0038023660890758038, + 0.002858084859326482, + 0.0024804044514894485, + 0.0026652012020349503, + 0.0019851946271955967, + 0.001869042287580669, + 0.0019904281944036484, + 0.0023123507853597403, + 0.0023057919461280107, + 0.0027387477457523346, + 0.001740691950544715, + 0.0033642796333879232, + 0.0012207027757540345, + 0.0010245006997138262, + 0.0021168093662708998, + 0.0020860154181718826, + 0.0013805202906951308, + 0.0022536073811352253 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 5 + ], + "selected_reward": -0.005262095481157303 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09552668035030365, + 0.4274265170097351, + 0.1338152289390564, + 0.09361694753170013, + 0.07008720934391022, + 0.048367153853178024, + 0.04189720377326012, + 0.03171991929411888, + 0.033348198980093, + 0.024194875732064247 + ], + "credit_map": [ + 0.006337031256407499, + 0.005189100280404091, + 0.02461194060742855, + 0.009663931094110012, + 0.006303762085735798, + 0.008760715834796429, + 0.007745923008769751, + 0.02691427432000637, + 0.016458507627248764, + 0.0034550344571471214, + 0.18668097257614136, + 0.012866610661149025, + 0.003353738458827138, + 0.008698605000972748, + 0.006777354050427675, + 0.18913568556308746, + 0.013135272078216076, + 0.00609351322054863, + 0.023620007559657097, + 0.005492677446454763, + 0.008865606971085072, + 0.0545264333486557, + 0.004696350544691086, + 0.017385359853506088, + 0.010188106447458267, + 0.01272626407444477, + 0.011761145666241646, + 0.00961686298251152, + 0.013791894540190697, + 0.012076308019459248, + 0.010968497022986412, + 0.012487869709730148, + 0.00897525716573, + 0.009364803321659565, + 0.010393764823675156, + 0.003344356082379818, + 0.009860014542937279, + 0.011028536595404148, + 0.004576896782964468, + 0.012543583288788795, + 0.005910114850848913, + 0.008372487500309944, + 0.0030684240628033876, + 0.003023178782314062, + 0.012762109749019146, + 0.009220593608915806, + 0.0030890207272022963, + 0.0029212271329015493, + 0.004952452145516872, + 0.0061493925750255585, + 0.0032778086606413126, + 0.003248985856771469, + 0.008897722698748112, + 0.00669175386428833, + 0.0027661926578730345, + 0.005912897642701864, + 0.003597557544708252, + 0.005838828627020121, + 0.0015694889007136226, + 0.0016364265466108918, + 0.0076860603876411915, + 0.006907380651682615, + 0.0022568916901946068, + 0.0022272856440395117, + 0.004540815483778715, + 0.0043567754328250885, + 0.004296072293072939, + 0.0030428285244852304, + 0.0030894845258444548, + 0.004429878666996956, + 0.00399394053965807, + 0.0055984016507864, + 0.0036595319397747517, + 0.0050756982527673244, + 0.0022980552166700363, + 0.0016226215520873666, + 0.003812768030911684, + 0.0026402988005429506, + 0.0017057149671018124, + 0.0033801887184381485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.012537590228021145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818414404988289 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.01818414404988289, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818414404988289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052090857177972794, + 0.5562747120857239, + 0.1568153202533722, + 0.056417517364025116, + 0.02843424491584301, + 0.038320742547512054, + 0.03491757810115814, + 0.026017839089035988, + 0.02797313965857029, + 0.022737979888916016 + ], + "credit_map": [ + 0.0028750270139425993, + 0.002597798127681017, + 0.01797502115368843, + 0.002670188434422016, + 0.002484310884028673, + 0.0026728305965662003, + 0.0031701920088380575, + 0.017645498737692833, + 0.02084202878177166, + 0.004248897545039654, + 0.24344676733016968, + 0.01686488464474678, + 0.004390920978039503, + 0.01091944333165884, + 0.009320995770394802, + 0.24624082446098328, + 0.007924760691821575, + 0.003456655889749527, + 0.06446069478988647, + 0.005111233331263065, + 0.003088987898081541, + 0.006327150855213404, + 0.003855636576190591, + 0.06259021908044815, + 0.008363114669919014, + 0.008892890065908432, + 0.006240527145564556, + 0.00412426283583045, + 0.00762939453125, + 0.010859846137464046, + 0.004268025979399681, + 0.0060394639149308205, + 0.004713148344308138, + 0.0018594713183119893, + 0.004231632221490145, + 0.0024026352912187576, + 0.002960267011076212, + 0.005959926638752222, + 0.0015853941440582275, + 0.004721773788332939, + 0.006568312179297209, + 0.004902169108390808, + 0.0031287677120417356, + 0.0031616142950952053, + 0.007059652823954821, + 0.006039603613317013, + 0.003230092115700245, + 0.004230533726513386, + 0.004835841711610556, + 0.005678095854818821, + 0.0018604766810312867, + 0.002397803822532296, + 0.008897860534489155, + 0.0061531225219368935, + 0.00247863563708961, + 0.002615748206153512, + 0.0027969595976173878, + 0.003843347541987896, + 0.002252900041639805, + 0.002405749633908272, + 0.0034360710997134447, + 0.005317725706845522, + 0.0036328877322375774, + 0.002332203323021531, + 0.00341933430172503, + 0.004893849603831768, + 0.003159652929753065, + 0.002531912876293063, + 0.003902421798557043, + 0.0034006924834102392, + 0.0031838747672736645, + 0.003481403226032853, + 0.002667826833203435, + 0.004994309041649103, + 0.001803049584850669, + 0.001448436756618321, + 0.0032026399858295918, + 0.0031645228154957294, + 0.0020336362067610025, + 0.0034235618077218533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.01818414404988289 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a rapid-cut montage of macro close-ups, the film captures, with an almost scientific precision, the entire process of a droplet of ink, thick as night, plump and plump, plump black ink plunging into still, pure water. The sequence begins with an extreme close-up shot from directly above, showing the plump, jet-black droplet suspended above the water's surface. It then silently touches the surface, stirring a tiny, perfectly formed ripple. The camera then shifts to a horizontal profile, showing the moment the droplet pierces the surface, sinking like a miniature depth charge, leaving a slender, straight black trail. Then, an even closer close-up captures the body of the droplet's sudden yet graceful unfurling within the water, countless silky black tendrils spreading out in all directions to form a complex, evolving spherical cloud. Illuminated by a bright, even backlight, these inky filaments reveal breathtaking detail, some as dense as smoke, others as thin as gauze, interweaving into magnificent patterns reminiscent of cosmic nebulae or fractal geometry. The entire process takes place in absolute silence, with the dominant colors of the painting being extreme black and white, creating a visual atmosphere that is both chaotic and tranquil, full of Zen and a sense of abstract art.", + "seed": 24479, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_7447_seed_24479_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2240011801943183, + "vlm_logprob": -0.2240011801943183, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.2240011801943183, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0924922227859497 + ], + "opposite_logprob_tokens": [ + -0.4674922525882721 + ], + "logprob_reward": -1.0924922227859497, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0924922227859497, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18817302584648132, + 0.12691187858581543, + 0.030285697430372238, + 0.05050792545080185, + 0.16408489644527435, + 0.036094892770051956, + 0.02333340421319008, + 0.015311377122998238, + 0.1173752173781395, + 0.24792161583900452 + ], + "credit_map": [ + 0.026354949921369553, + 0.05032220110297203, + 0.01507218275219202, + 0.01378319226205349, + 0.01797424629330635, + 0.022166166454553604, + 0.024305101484060287, + 0.018194980919361115, + 0.011181836016476154, + 0.022128799930214882, + 0.02549360692501068, + 0.011958860792219639, + 0.014682874083518982, + 0.013557297177612782, + 0.010723846964538097, + 0.017184754833579063, + 0.003390062367543578, + 0.0030715574976056814, + 0.006829740013927221, + 0.0036450959742069244, + 0.002225189935415983, + 0.004193547647446394, + 0.002082695486024022, + 0.004847810138016939, + 0.006396927870810032, + 0.005700238514691591, + 0.00565055338665843, + 0.006748780608177185, + 0.007157608866691589, + 0.0071431430988013744, + 0.004599026869982481, + 0.00711164390668273, + 0.016298098489642143, + 0.015663042664527893, + 0.03726540133357048, + 0.02347453683614731, + 0.015802739188075066, + 0.01880495436489582, + 0.013107206672430038, + 0.02366892248392105, + 0.0030452441424131393, + 0.005161862354725599, + 0.003779590129852295, + 0.005376641638576984, + 0.006591870449483395, + 0.003295726841315627, + 0.005024268291890621, + 0.003819689154624939, + 0.0022363329771906137, + 0.0016833808040246367, + 0.003462563268840313, + 0.0033061853609979153, + 0.0028598220087587833, + 0.002293359488248825, + 0.004546753596514463, + 0.0029450065921992064, + 0.0016311906510964036, + 0.001255367649719119, + 0.0024463089648634195, + 0.0018163134809583426, + 0.0020096704829484224, + 0.002262543886899948, + 0.002026001922786236, + 0.001863979734480381, + 0.008839820511639118, + 0.009184391237795353, + 0.008594856597483158, + 0.008458910509943962, + 0.026957975700497627, + 0.00810746569186449, + 0.019757704809308052, + 0.027474097907543182, + 0.010551631450653076, + 0.008780235424637794, + 0.015171606093645096, + 0.09984511137008667, + 0.047440920025110245, + 0.00875104684382677, + 0.013905051164329052, + 0.04347602277994156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -1.0924922227859497 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005984009243547916 + ], + "opposite_logprob_tokens": [ + -5.130983829498291 + ], + "logprob_reward": -0.005984009243547916, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005984009243547916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2289193719625473, + 0.12230770289897919, + 0.0777026116847992, + 0.07602909952402115, + 0.09923212975263596, + 0.024655697867274284, + 0.042104676365852356, + 0.03281860053539276, + 0.19359718263149261, + 0.10263293236494064 + ], + "credit_map": [ + 0.035032037645578384, + 0.054694272577762604, + 0.016669446602463722, + 0.017721448093652725, + 0.022143645212054253, + 0.037487246096134186, + 0.021269848570227623, + 0.0239014383405447, + 0.010919884778559208, + 0.019502529874444008, + 0.026857126504182816, + 0.011337114498019218, + 0.0128850182518363, + 0.01187849696725607, + 0.011116241104900837, + 0.01781129091978073, + 0.007981817238032818, + 0.008021948859095573, + 0.019961880519986153, + 0.008824589662253857, + 0.006397140212357044, + 0.008446778170764446, + 0.005769883748143911, + 0.012298577465116978, + 0.011909686028957367, + 0.00677013024687767, + 0.009201694279909134, + 0.007703826762735844, + 0.013131072744727135, + 0.010980784893035889, + 0.006315629929304123, + 0.010016277432441711, + 0.007008629385381937, + 0.00994142983108759, + 0.013811408542096615, + 0.018185516819357872, + 0.015685293823480606, + 0.008330747485160828, + 0.013648024760186672, + 0.012621087022125721, + 0.0022235142532736063, + 0.002416161587461829, + 0.002910769311711192, + 0.003347104648128152, + 0.004668003413826227, + 0.0021554743871092796, + 0.004246316850185394, + 0.002688354579731822, + 0.004914047196507454, + 0.002795629436150193, + 0.007189968600869179, + 0.006411269772797823, + 0.0041412655264139175, + 0.005296272225677967, + 0.005453449673950672, + 0.00590277137234807, + 0.0038171897176653147, + 0.002667295513674617, + 0.005560026038438082, + 0.005131268408149481, + 0.003917546942830086, + 0.0048502059653401375, + 0.0032620758283883333, + 0.003612997243180871, + 0.015836229547858238, + 0.021333614364266396, + 0.01865769922733307, + 0.02489156648516655, + 0.03801910951733589, + 0.015668358653783798, + 0.03625865653157234, + 0.022931961342692375, + 0.006598855368793011, + 0.006396433338522911, + 0.008399834856390953, + 0.034211933612823486, + 0.013774719089269638, + 0.0058096847496926785, + 0.007125070318579674, + 0.020316407084465027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.005984009243547916 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005259842611849308 + ], + "opposite_logprob_tokens": [ + -5.255259990692139 + ], + "logprob_reward": -0.005259842611849308, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005259842611849308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13862942159175873, + 0.1264718919992447, + 0.09141889214515686, + 0.09514922648668289, + 0.23054680228233337, + 0.038857392966747284, + 0.03475387021899223, + 0.02152680791914463, + 0.14534570276737213, + 0.07730001956224442 + ], + "credit_map": [ + 0.020131058990955353, + 0.03178001195192337, + 0.009088334627449512, + 0.01254216581583023, + 0.022931335493922234, + 0.016063617542386055, + 0.011270901188254356, + 0.014821996912360191, + 0.012488744221627712, + 0.018086470663547516, + 0.02626888081431389, + 0.012478912249207497, + 0.01400486845523119, + 0.014626956544816494, + 0.011108173988759518, + 0.017408888787031174, + 0.010821495205163956, + 0.008723314851522446, + 0.02059049718081951, + 0.010824811644852161, + 0.007998690009117126, + 0.010514567606151104, + 0.008337392471730709, + 0.013608123175799847, + 0.008443022146821022, + 0.010186353698372841, + 0.010702595114707947, + 0.012025386095046997, + 0.018237389624118805, + 0.013910898938775063, + 0.009350750595331192, + 0.012292833998799324, + 0.02082897163927555, + 0.02272918075323105, + 0.04189528897404671, + 0.026340873911976814, + 0.03909958526492119, + 0.02627420425415039, + 0.026144379749894142, + 0.02723432332277298, + 0.004470326937735081, + 0.0033329252619296312, + 0.005285143852233887, + 0.006546698044985533, + 0.005236630793660879, + 0.004326078109443188, + 0.004893764853477478, + 0.0047658272087574005, + 0.003815757343545556, + 0.0025557223707437515, + 0.004474969115108252, + 0.004982654005289078, + 0.003930613864213228, + 0.004009174648672342, + 0.006506918929517269, + 0.00447806203737855, + 0.002291039563715458, + 0.0018741877283900976, + 0.002497154287993908, + 0.003539806930348277, + 0.002800556132569909, + 0.0031617500353604555, + 0.003437025472521782, + 0.0019252885831519961, + 0.01015991810709238, + 0.00935280229896307, + 0.012808803468942642, + 0.008554733358323574, + 0.040694911032915115, + 0.009898238815367222, + 0.016174931079149246, + 0.037701357156038284, + 0.009481485933065414, + 0.006834742147475481, + 0.00697612389922142, + 0.01631319336593151, + 0.014028684236109257, + 0.005010769236832857, + 0.00681292824447155, + 0.0118420934304595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.005259842611849308 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628949359059334 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.007628949359059334, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628949359059334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1218826025724411, + 0.10595835000276566, + 0.03294084593653679, + 0.0985625684261322, + 0.11565663665533066, + 0.03314922749996185, + 0.025433512404561043, + 0.044336844235658646, + 0.131948783993721, + 0.2901306450366974 + ], + "credit_map": [ + 0.015678664669394493, + 0.028836844488978386, + 0.007888047955930233, + 0.008580162189900875, + 0.014206252060830593, + 0.010575651191174984, + 0.012079744599759579, + 0.024037232622504234, + 0.011741513386368752, + 0.016440238803625107, + 0.019541658461093903, + 0.010218674317002296, + 0.011737818829715252, + 0.014408828690648079, + 0.007567953318357468, + 0.0143016641959548, + 0.003275712486356497, + 0.003364196512848139, + 0.005833787377923727, + 0.004196173045784235, + 0.00337250716984272, + 0.004314075689762831, + 0.0033108475618064404, + 0.0052735451608896255, + 0.01072120014578104, + 0.008610361255705357, + 0.012869730591773987, + 0.015165332704782486, + 0.013779295608401299, + 0.012223172932863235, + 0.010512806475162506, + 0.014680673368275166, + 0.010642217472195625, + 0.01242803130298853, + 0.02145162783563137, + 0.01240957248955965, + 0.019738774746656418, + 0.013112830929458141, + 0.011780724860727787, + 0.014092858880758286, + 0.0026783160865306854, + 0.0037179363425821066, + 0.0035847260151058435, + 0.005595511756837368, + 0.005218050442636013, + 0.0025970651768147945, + 0.005572376772761345, + 0.004185243044048548, + 0.0033931054640561342, + 0.0018063888419419527, + 0.0033397376537323, + 0.0035591484047472477, + 0.003610016079619527, + 0.0029747160151600838, + 0.003972053062170744, + 0.0027783468831330538, + 0.0055064857006073, + 0.004418636206537485, + 0.004558745305985212, + 0.005490506999194622, + 0.008434297516942024, + 0.004869471304118633, + 0.005741808097809553, + 0.005316894501447678, + 0.009857670404016972, + 0.011164295487105846, + 0.012215458787977695, + 0.00994537677615881, + 0.03448330983519554, + 0.010891682468354702, + 0.013699180446565151, + 0.02969181537628174, + 0.017088131979107857, + 0.018106255680322647, + 0.02007218822836876, + 0.1057576835155487, + 0.07072857022285461, + 0.014096363447606564, + 0.021829035133123398, + 0.02245241031050682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.007628949359059334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640876971185207 + ], + "opposite_logprob_tokens": [ + -4.758640766143799 + ], + "logprob_reward": -0.008640876971185207, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640876971185207, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26820307970046997, + 0.14243623614311218, + 0.09935040026903152, + 0.0559210330247879, + 0.06951776146888733, + 0.052814971655607224, + 0.057940542697906494, + 0.032056599855422974, + 0.18321417272090912, + 0.03854520618915558 + ], + "credit_map": [ + 0.04834984987974167, + 0.04268408194184303, + 0.028531290590763092, + 0.02708585001528263, + 0.028760990127921104, + 0.03169575333595276, + 0.0322723463177681, + 0.028822917491197586, + 0.013928035274147987, + 0.022379620000720024, + 0.029583517462015152, + 0.013634469360113144, + 0.015515581704676151, + 0.013735518790781498, + 0.013559638522565365, + 0.02009984850883484, + 0.01113072782754898, + 0.01077869813889265, + 0.023810481652617455, + 0.01147254928946495, + 0.008433371782302856, + 0.011431748047471046, + 0.007397622335702181, + 0.014895202592015266, + 0.009094307199120522, + 0.00565166212618351, + 0.006630384363234043, + 0.005177222192287445, + 0.009242446161806583, + 0.008046435192227364, + 0.004789563361555338, + 0.007289011962711811, + 0.005276650656014681, + 0.006916272919625044, + 0.01190207526087761, + 0.014108611270785332, + 0.0064535485580563545, + 0.006214819382876158, + 0.008452748879790306, + 0.010193034075200558, + 0.0061502605676651, + 0.0051483577117323875, + 0.007036739960312843, + 0.008832698687911034, + 0.005389044992625713, + 0.0050790137611329556, + 0.008402332663536072, + 0.006776524242013693, + 0.007490434683859348, + 0.003821446793153882, + 0.0083019332960248, + 0.00862284004688263, + 0.006117973010987043, + 0.006584801711142063, + 0.008907311595976353, + 0.008093799464404583, + 0.0047965641133487225, + 0.0030272286385297775, + 0.0049762544222176075, + 0.004103673622012138, + 0.004144930746406317, + 0.004758541472256184, + 0.0030324836261570454, + 0.003216920653358102, + 0.013669912703335285, + 0.015923021361231804, + 0.01482224278151989, + 0.013893211260437965, + 0.04152410477399826, + 0.01312714908272028, + 0.029972275719046593, + 0.040282268077135086, + 0.0032063513062894344, + 0.0038116411305963993, + 0.004766636993736029, + 0.00685312831774354, + 0.002761855721473694, + 0.003516176948323846, + 0.0039054700173437595, + 0.009723945520818233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.008640876971185207 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a rapid-cut montage of macro close-ups, the film captures, with an almost scientific precision, the entire process of a droplet of ink, thick as night, plump and plump, plump black ink plunging into still, pure water. The sequence begins with an extreme close-up shot from directly above, showing the plump, jet-black droplet suspended above the water's surface. It then silently touches the surface, stirring a tiny, perfectly formed ripple. The camera then shifts to a horizontal profile, showing the moment the droplet pierces the surface, sinking like a miniature depth charge, leaving a slender, straight black trail. Then, an even closer close-up captures the body of the droplet's sudden yet graceful unfurling within the water, countless silky black tendrils spreading out in all directions to form a complex, evolving spherical cloud. Illuminated by a bright, even backlight, these inky filaments reveal breathtaking detail, some as dense as smoke, others as thin as gauze, interweaving into magnificent patterns reminiscent of cosmic nebulae or fractal geometry. The entire process takes place in absolute silence, with the dominant colors of the painting being extreme black and white, creating a visual atmosphere that is both chaotic and tranquil, full of Zen and a sense of abstract art.", + "seed": 24480, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_7447_seed_24480_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1309614259749651, + "vlm_logprob": -0.1309614259749651, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.1309614259749651, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.56174635887146 + ], + "opposite_logprob_tokens": [ + -0.93674635887146 + ], + "logprob_reward": -0.56174635887146, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.56174635887146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13652439415454865, + 0.16491548717021942, + 0.29299575090408325, + 0.0436665378510952, + 0.0846126601099968, + 0.10348770022392273, + 0.044914525002241135, + 0.0748891606926918, + 0.03475552797317505, + 0.019238179549574852 + ], + "credit_map": [ + 0.014510367065668106, + 0.007079217117279768, + 0.01929936371743679, + 0.01567046344280243, + 0.00809517316520214, + 0.018759913742542267, + 0.017979271709918976, + 0.035130612552165985, + 0.018360452726483345, + 0.013712423853576183, + 0.027695968747138977, + 0.015192383900284767, + 0.011103379540145397, + 0.03724301978945732, + 0.010927511379122734, + 0.030680358409881592, + 0.09600082784891129, + 0.016634775325655937, + 0.0346236377954483, + 0.019269585609436035, + 0.011550210416316986, + 0.03390716388821602, + 0.01103216502815485, + 0.06997738778591156, + 0.005868194624781609, + 0.005470711272209883, + 0.006495533511042595, + 0.005804987158626318, + 0.0041166795417666435, + 0.005400854628533125, + 0.00430136127397418, + 0.006208213046193123, + 0.011885109357535839, + 0.008317925035953522, + 0.011150694452226162, + 0.009839913807809353, + 0.010976758785545826, + 0.013297059573233128, + 0.009199358522891998, + 0.009945840574800968, + 0.011669869534671307, + 0.017697669565677643, + 0.015186450444161892, + 0.011088922619819641, + 0.008111957460641861, + 0.016098106279969215, + 0.012289668433368206, + 0.011345055885612965, + 0.004449485335499048, + 0.004330052062869072, + 0.008011181838810444, + 0.005176791455596685, + 0.0041510965675115585, + 0.0045204912312328815, + 0.007354475557804108, + 0.006920950952917337, + 0.014103321358561516, + 0.007112157065421343, + 0.008436890318989754, + 0.005696945358067751, + 0.013197683729231358, + 0.010697189718484879, + 0.0065700882114470005, + 0.009074889123439789, + 0.0032604248262941837, + 0.0034866342321038246, + 0.006042510271072388, + 0.0037811887450516224, + 0.003479587845504284, + 0.005529368296265602, + 0.0026436715852469206, + 0.006532141473144293, + 0.0028541195206344128, + 0.0024859076365828514, + 0.002161354059353471, + 0.0014654522528871894, + 0.002348921727389097, + 0.002565772272646427, + 0.0024980378802865744, + 0.0028586138505488634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.56174635887146 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008669831790030003 + ], + "opposite_logprob_tokens": [ + -4.758669853210449 + ], + "logprob_reward": -0.008669831790030003, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008669831790030003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3090987801551819, + 0.2448265701532364, + 0.0989215224981308, + 0.043307963758707047, + 0.06841956079006195, + 0.07490931451320648, + 0.05167633295059204, + 0.041852422058582306, + 0.03207332268357277, + 0.034914229065179825 + ], + "credit_map": [ + 0.04051857069134712, + 0.01277176570147276, + 0.06006849557161331, + 0.023553211241960526, + 0.017561331391334534, + 0.03988264873623848, + 0.03547871857881546, + 0.07926401495933533, + 0.028030049055814743, + 0.015944475308060646, + 0.044857658445835114, + 0.01569880172610283, + 0.015650276094675064, + 0.05889274552464485, + 0.014247220009565353, + 0.05150534585118294, + 0.024834055453538895, + 0.00791175663471222, + 0.01016775518655777, + 0.00886002741754055, + 0.005912122782319784, + 0.013710561208426952, + 0.00604630820453167, + 0.021478937938809395, + 0.004967173561453819, + 0.0045654866844415665, + 0.006268214900046587, + 0.006125130690634251, + 0.004572530742734671, + 0.0068188877776265144, + 0.00438282173126936, + 0.00560772093012929, + 0.0067046633921563625, + 0.006816364359110594, + 0.011719814501702785, + 0.01224349346011877, + 0.0060045961290597916, + 0.008231217041611671, + 0.007165931165218353, + 0.009533477947115898, + 0.008551561273634434, + 0.008646982721984386, + 0.009419006295502186, + 0.012740041129291058, + 0.0048353057354688644, + 0.01350494846701622, + 0.00990808941423893, + 0.00730338366702199, + 0.005471276119351387, + 0.004137734416872263, + 0.011476443149149418, + 0.005592717789113522, + 0.00451304204761982, + 0.005461366381496191, + 0.005028885789215565, + 0.009994867257773876, + 0.009265420027077198, + 0.0033395159989595413, + 0.008524549193680286, + 0.002979980781674385, + 0.006579163484275341, + 0.004498372785747051, + 0.0025782070588320494, + 0.004087213426828384, + 0.003992539364844561, + 0.002070465823635459, + 0.008003076538443565, + 0.0034737740643322468, + 0.002917851787060499, + 0.0045168837532401085, + 0.002616714686155319, + 0.00448201596736908, + 0.0058119529858231544, + 0.006887591443955898, + 0.003762268926948309, + 0.0028436898719519377, + 0.003932845778763294, + 0.004576488398015499, + 0.003603976685553789, + 0.003495412180200219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008669831790030003 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307024925947189 + ], + "opposite_logprob_tokens": [ + -3.1680703163146973 + ], + "logprob_reward": -0.04307024925947189, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307024925947189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14705020189285278, + 0.15143275260925293, + 0.4727788269519806, + 0.031492121517658234, + 0.03858591616153717, + 0.05415133386850357, + 0.05078725889325142, + 0.03286220133304596, + 0.010555466637015343, + 0.010303945280611515 + ], + "credit_map": [ + 0.0235291738063097, + 0.009088877588510513, + 0.022318415343761444, + 0.014616006053984165, + 0.013761583715677261, + 0.01875043474137783, + 0.023113273084163666, + 0.021872449666261673, + 0.018082423135638237, + 0.010144663974642754, + 0.02337801828980446, + 0.02056695520877838, + 0.010949601419270039, + 0.02374253049492836, + 0.015914734452962875, + 0.028653834015130997, + 0.19198012351989746, + 0.015266211703419685, + 0.06312635540962219, + 0.015492970123887062, + 0.011990881524980068, + 0.14084281027317047, + 0.00991381797939539, + 0.024165667593479156, + 0.0034298712853342295, + 0.005518954247236252, + 0.004277804400771856, + 0.004479421768337488, + 0.0037320370320230722, + 0.0037644009571522474, + 0.003705952549353242, + 0.002583684166893363, + 0.0042740232311189175, + 0.004991235677152872, + 0.004233536776155233, + 0.004217781126499176, + 0.006428055930882692, + 0.005268890876322985, + 0.0033164548221975565, + 0.005855944473296404, + 0.0061456006951630116, + 0.008740688674151897, + 0.008090094663202763, + 0.007174298167228699, + 0.0037496767472475767, + 0.009022679179906845, + 0.0061112092807888985, + 0.00511708902195096, + 0.005125876981765032, + 0.005670375656336546, + 0.009655680507421494, + 0.0046600280329585075, + 0.006155852228403091, + 0.006007668096572161, + 0.005856099538505077, + 0.00765567971393466, + 0.005539693403989077, + 0.0022451377008110285, + 0.00887646246701479, + 0.002613081131130457, + 0.0036611107643693686, + 0.003369309473782778, + 0.002347896108403802, + 0.004209511913359165, + 0.000950089655816555, + 0.0007289131172001362, + 0.0012399230618029833, + 0.001075565000064671, + 0.0011750843841582537, + 0.002865660935640335, + 0.0008114348747767508, + 0.0017087963642552495, + 0.0016307022888213396, + 0.0011073254281654954, + 0.001022041542455554, + 0.0008054414065554738, + 0.001262904959730804, + 0.001293162233196199, + 0.001160127460025251, + 0.002022241475060582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.04307024925947189 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369256854057312 + ], + "opposite_logprob_tokens": [ + -3.4086925983428955 + ], + "logprob_reward": -0.03369256854057312, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369256854057312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11365430802106857, + 0.15023070573806763, + 0.36758488416671753, + 0.044838082045316696, + 0.11903145164251328, + 0.07078909128904343, + 0.062022995203733444, + 0.023733582347631454, + 0.028634198009967804, + 0.0194806307554245 + ], + "credit_map": [ + 0.012964537367224693, + 0.005840186029672623, + 0.021263323724269867, + 0.012408931739628315, + 0.008861823938786983, + 0.014369847252964973, + 0.017903216183185577, + 0.02004244178533554, + 0.01915338821709156, + 0.00939544290304184, + 0.02624712698161602, + 0.012207947671413422, + 0.007826455868780613, + 0.035839226096868515, + 0.008508330211043358, + 0.031052788719534874, + 0.12795864045619965, + 0.018650194630026817, + 0.04427780583500862, + 0.022714009508490562, + 0.013073659501969814, + 0.05088885501027107, + 0.012774094007909298, + 0.07724763453006744, + 0.0050235893577337265, + 0.004567524418234825, + 0.006238814443349838, + 0.008588479831814766, + 0.004090140573680401, + 0.004587620031088591, + 0.006543169263750315, + 0.005198745988309383, + 0.014307592995464802, + 0.011511716991662979, + 0.014406335540115833, + 0.01735810562968254, + 0.01658918894827366, + 0.016825709491968155, + 0.013151654042303562, + 0.014881148934364319, + 0.008117920719087124, + 0.013685416430234909, + 0.010148054920136929, + 0.009620408527553082, + 0.007647047284990549, + 0.007840312086045742, + 0.006474227644503117, + 0.007255706004798412, + 0.006439907010644674, + 0.006210456136614084, + 0.009994832798838615, + 0.007747117895632982, + 0.006723771337419748, + 0.008427325636148453, + 0.008009087294340134, + 0.008470498956739902, + 0.004512767773121595, + 0.0020573120564222336, + 0.005706917028874159, + 0.00179780344478786, + 0.0034208314027637243, + 0.0022074601147323847, + 0.0015389457112178206, + 0.0024915465619415045, + 0.0026935015339404345, + 0.0017169115599244833, + 0.003810318186879158, + 0.002327416557818651, + 0.002537696622312069, + 0.006519460119307041, + 0.001827272935770452, + 0.007201622240245342, + 0.0030914866365492344, + 0.002813381375744939, + 0.0022741490975022316, + 0.0018086314667016268, + 0.0023482542019337416, + 0.002389116445556283, + 0.00238060369156301, + 0.0023750076070427895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.03369256854057312 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628121413290501 + ], + "opposite_logprob_tokens": [ + -4.882627964019775 + ], + "logprob_reward": -0.007628121413290501, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628121413290501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12142657488584518, + 0.49960872530937195, + 0.1589369922876358, + 0.013280536979436874, + 0.025046132504940033, + 0.07348334044218063, + 0.037027955055236816, + 0.01630346290767193, + 0.028362460434436798, + 0.026523778215050697 + ], + "credit_map": [ + 0.01641732268035412, + 0.005043629091233015, + 0.03293807432055473, + 0.008816750720143318, + 0.006555706262588501, + 0.015930380672216415, + 0.01241251453757286, + 0.023312196135520935, + 0.05768892168998718, + 0.032730814069509506, + 0.09385523200035095, + 0.0349486768245697, + 0.027210203930735588, + 0.12119129300117493, + 0.027143200859427452, + 0.10484038293361664, + 0.046223174780607224, + 0.007544425781816244, + 0.021454978734254837, + 0.010438929311931133, + 0.005424502305686474, + 0.042804356664419174, + 0.005502148997038603, + 0.019544478505849838, + 0.0016371516976505518, + 0.001759107573889196, + 0.0021541311871260405, + 0.001752388896420598, + 0.001440526219084859, + 0.0015328106237575412, + 0.0016879856120795012, + 0.0013164341216906905, + 0.002884611254557967, + 0.0031632641330361366, + 0.003969206940382719, + 0.002861838089302182, + 0.0040343208238482475, + 0.0024730407167226076, + 0.0028582888189703226, + 0.002801560563966632, + 0.0075120218098163605, + 0.006829462479799986, + 0.015208146534860134, + 0.010494137182831764, + 0.0038179466500878334, + 0.00898519717156887, + 0.006781904026865959, + 0.013854524120688438, + 0.003069991944357753, + 0.003653652500361204, + 0.007010920904576778, + 0.005679284688085318, + 0.005256361328065395, + 0.0032001263462007046, + 0.0044443001970648766, + 0.004713318310678005, + 0.001614041393622756, + 0.0015352616319432855, + 0.002752414671704173, + 0.0011691898107528687, + 0.0023369728587567806, + 0.002303582150489092, + 0.0011613396927714348, + 0.003430661279708147, + 0.002509139245375991, + 0.0013048143591731787, + 0.00736876018345356, + 0.0017601093277335167, + 0.0024182540364563465, + 0.007820723578333855, + 0.0023677186109125614, + 0.002812942722812295, + 0.004023475106805563, + 0.00490687508136034, + 0.0024314955808222294, + 0.0020250692032277584, + 0.003260089084506035, + 0.004080320708453655, + 0.0024262170772999525, + 0.003370236372575164 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.007628121413290501 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a rapid-cut montage of macro close-ups, the film captures, with an almost scientific precision, the entire process of a droplet of ink, thick as night, plump and plump, plump black ink plunging into still, pure water. The sequence begins with an extreme close-up shot from directly above, showing the plump, jet-black droplet suspended above the water's surface. It then silently touches the surface, stirring a tiny, perfectly formed ripple. The camera then shifts to a horizontal profile, showing the moment the droplet pierces the surface, sinking like a miniature depth charge, leaving a slender, straight black trail. Then, an even closer close-up captures the body of the droplet's sudden yet graceful unfurling within the water, countless silky black tendrils spreading out in all directions to form a complex, evolving spherical cloud. Illuminated by a bright, even backlight, these inky filaments reveal breathtaking detail, some as dense as smoke, others as thin as gauze, interweaving into magnificent patterns reminiscent of cosmic nebulae or fractal geometry. The entire process takes place in absolute silence, with the dominant colors of the painting being extreme black and white, creating a visual atmosphere that is both chaotic and tranquil, full of Zen and a sense of abstract art.", + "seed": 24481, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_0/index_7447_seed_24481_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0691634790506214, + "vlm_logprob": -0.0691634790506214, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.0691634790506214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.320298433303833 + ], + "opposite_logprob_tokens": [ + -1.445298433303833 + ], + "logprob_reward": -0.320298433303833, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.320298433303833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031745098531246185, + 0.056039631366729736, + 0.054268572479486465, + 0.468718558549881, + 0.07101355493068695, + 0.14525212347507477, + 0.06958504021167755, + 0.045224178582429886, + 0.030700594186782837, + 0.027452610433101654 + ], + "credit_map": [ + 0.005502159707248211, + 0.0033681532368063927, + 0.005472628399729729, + 0.0033860541880130768, + 0.003801558166742325, + 0.004703680984675884, + 0.002570454264059663, + 0.0029404093511402607, + 0.008037769235670567, + 0.006922894157469273, + 0.007261845748871565, + 0.005152500234544277, + 0.008116786368191242, + 0.009628929197788239, + 0.004733608569949865, + 0.0061852955259382725, + 0.007817110046744347, + 0.005220210179686546, + 0.00788655411452055, + 0.006562930531799793, + 0.0062910085543990135, + 0.010046545416116714, + 0.005306228529661894, + 0.005137983243912458, + 0.07199053466320038, + 0.1902035027742386, + 0.018393537029623985, + 0.018862489610910416, + 0.019346239045262337, + 0.12477583438158035, + 0.012038221582770348, + 0.013108188286423683, + 0.007351428270339966, + 0.0112908985465765, + 0.008496428839862347, + 0.015581903979182243, + 0.005530327558517456, + 0.006546736694872379, + 0.008082806132733822, + 0.00813302118331194, + 0.014377838000655174, + 0.009042215533554554, + 0.02111341431736946, + 0.022788632661104202, + 0.012114258483052254, + 0.012069180607795715, + 0.009527809917926788, + 0.044218774884939194, + 0.00680821668356657, + 0.006912063341587782, + 0.0070961094461381435, + 0.008274316787719727, + 0.011166631244122982, + 0.008447973057627678, + 0.009844165295362473, + 0.011035565286874771, + 0.0035887991543859243, + 0.0031776633113622665, + 0.009678274393081665, + 0.006231993902474642, + 0.003974554128944874, + 0.0039967321790754795, + 0.007169445976614952, + 0.00740671344101429, + 0.002489885315299034, + 0.0029337031301110983, + 0.00440281443297863, + 0.0049479808658361435, + 0.003730025375261903, + 0.002148611703887582, + 0.0058435555547475815, + 0.0042040180414915085, + 0.003189315553754568, + 0.002416711300611496, + 0.0029562772251665592, + 0.0036296930629760027, + 0.0029528560116887093, + 0.0038534863851964474, + 0.004010103642940521, + 0.0044441670179367065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 4 + ], + "selected_reward": -0.320298433303833 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005278460215777159 + ], + "opposite_logprob_tokens": [ + -5.255278587341309 + ], + "logprob_reward": -0.005278460215777159, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005278460215777159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15392470359802246, + 0.04389876127243042, + 0.08471960574388504, + 0.3488040864467621, + 0.07514631003141403, + 0.04221802204847336, + 0.11298557370901108, + 0.037752389907836914, + 0.055647820234298706, + 0.0449027456343174 + ], + "credit_map": [ + 0.02394016832113266, + 0.016955584287643433, + 0.02294626645743847, + 0.016652248799800873, + 0.018146203830838203, + 0.018571853637695312, + 0.023006241768598557, + 0.013706134632229805, + 0.0061409431509673595, + 0.004903629422187805, + 0.007670572027564049, + 0.0046961489133536816, + 0.005022818222641945, + 0.00758858909830451, + 0.003664195304736495, + 0.004211862571537495, + 0.009901310317218304, + 0.010899699293076992, + 0.011889057233929634, + 0.009153584949672222, + 0.008106070570647717, + 0.014907321892678738, + 0.00772863207384944, + 0.012133922427892685, + 0.02401847578585148, + 0.10493014752864838, + 0.01618995890021324, + 0.02119278535246849, + 0.018514977768063545, + 0.1382421851158142, + 0.011408747173845768, + 0.014306786470115185, + 0.009771370328962803, + 0.011653664521872997, + 0.007734423037618399, + 0.015575998462736607, + 0.00575829204171896, + 0.00645756721496582, + 0.0077536809258162975, + 0.010441307909786701, + 0.0058641196228563786, + 0.0034657095093280077, + 0.0037931078113615513, + 0.0096779465675354, + 0.005277327261865139, + 0.003911919891834259, + 0.0048068007454276085, + 0.005421086214482784, + 0.011821864172816277, + 0.013025671243667603, + 0.010928990319371223, + 0.01316162385046482, + 0.019588453695178032, + 0.014242544770240784, + 0.014149025082588196, + 0.016067394986748695, + 0.0028397904243320227, + 0.0026784867513924837, + 0.006625719368457794, + 0.004833918530493975, + 0.0036424128338694572, + 0.002858026186004281, + 0.005510480608791113, + 0.008763551712036133, + 0.0051001436077058315, + 0.004599792417138815, + 0.008760148659348488, + 0.007603791542351246, + 0.005408836994320154, + 0.004851797129958868, + 0.011616069823503494, + 0.007707234937697649, + 0.00623177457600832, + 0.005800517275929451, + 0.00599360466003418, + 0.009237668476998806, + 0.0031609658617526293, + 0.003799332305788994, + 0.005794780794531107, + 0.0048840949311852455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005278460215777159 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639714982360601 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639714982360601, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639714982360601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10465086251497269, + 0.05038246884942055, + 0.05249413102865219, + 0.5581534504890442, + 0.04471781849861145, + 0.07132388651371002, + 0.03118159994482994, + 0.03575338423252106, + 0.027686243876814842, + 0.023656165227293968 + ], + "credit_map": [ + 0.01878315769135952, + 0.010900904424488544, + 0.017477186396718025, + 0.008753502741456032, + 0.017137859016656876, + 0.01097562350332737, + 0.009213295765221119, + 0.011409319005906582, + 0.007767827715724707, + 0.007108308374881744, + 0.006872400175780058, + 0.004974996671080589, + 0.00636301189661026, + 0.007907523773610592, + 0.004163359757512808, + 0.00522503862157464, + 0.008834372274577618, + 0.007069071754813194, + 0.008220086805522442, + 0.006212748121470213, + 0.005184789188206196, + 0.005149793811142445, + 0.005517798010259867, + 0.006305469200015068, + 0.03601500764489174, + 0.19596882164478302, + 0.017824580892920494, + 0.02390388213098049, + 0.029271487146615982, + 0.21733765304088593, + 0.017133181914687157, + 0.020698796957731247, + 0.004166794009506702, + 0.006853112485259771, + 0.004051711410284042, + 0.00674479128792882, + 0.004330309573560953, + 0.007107040379196405, + 0.006703181657940149, + 0.004760872106999159, + 0.005611184984445572, + 0.004473808221518993, + 0.006062611937522888, + 0.02050694264471531, + 0.00442873127758503, + 0.005640676245093346, + 0.005687111988663673, + 0.018912823870778084, + 0.003686283016577363, + 0.003545979969203472, + 0.0038125000428408384, + 0.003576911287382245, + 0.0044489577412605286, + 0.0037903524935245514, + 0.0038561515975743532, + 0.004464462865144014, + 0.002985620405524969, + 0.0030803221743553877, + 0.006893084850162268, + 0.004385374952107668, + 0.003439482068642974, + 0.0028141115326434374, + 0.006683632265776396, + 0.005471750162541866, + 0.0021942362654954195, + 0.0026818218175321817, + 0.004389457404613495, + 0.0031871397513896227, + 0.002873878926038742, + 0.002455061301589012, + 0.005626015830785036, + 0.004278631415218115, + 0.002780201146379113, + 0.0030319977086037397, + 0.003192546311765909, + 0.002416201401501894, + 0.003943500109016895, + 0.002036944730207324, + 0.003207674017176032, + 0.0030470991041511297 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.007639714982360601 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028199690859764814 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.0028199690859764814, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028199690859764814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048149604350328445, + 0.03444278985261917, + 0.10920501500368118, + 0.29385754466056824, + 0.12452512979507446, + 0.10235192626714706, + 0.10634305328130722, + 0.09566891193389893, + 0.03866233676671982, + 0.04679372161626816 + ], + "credit_map": [ + 0.008544780313968658, + 0.005103415809571743, + 0.006504699122160673, + 0.003872360335662961, + 0.00586408656090498, + 0.010328713804483414, + 0.004238656256347895, + 0.0036928888875991106, + 0.005833711940795183, + 0.005690392106771469, + 0.004599840845912695, + 0.002064348431304097, + 0.004051096271723509, + 0.0055322772823274136, + 0.0031161329243332148, + 0.0035549900494515896, + 0.017286017537117004, + 0.010342276655137539, + 0.018086018040776253, + 0.014350233599543571, + 0.009656478650867939, + 0.016444027423858643, + 0.009700224734842777, + 0.013339733704924583, + 0.023594381287693977, + 0.11017441749572754, + 0.00884817074984312, + 0.010750839486718178, + 0.0158756822347641, + 0.10393951833248138, + 0.009261746890842915, + 0.011412766762077808, + 0.015297381207346916, + 0.019046198576688766, + 0.015309812501072884, + 0.021552180871367455, + 0.008607310242950916, + 0.01448499783873558, + 0.012668863870203495, + 0.017558373510837555, + 0.009509306401014328, + 0.009172403253614902, + 0.010182278230786324, + 0.020331382751464844, + 0.008380649611353874, + 0.012147507630288601, + 0.007885005325078964, + 0.024743380025029182, + 0.01657458394765854, + 0.01170619297772646, + 0.013809003867208958, + 0.008795119822025299, + 0.017974670976400375, + 0.013246845453977585, + 0.013972914777696133, + 0.010263722389936447, + 0.007411157246679068, + 0.005641666706651449, + 0.018054865300655365, + 0.012000142596662045, + 0.009708907455205917, + 0.008966268040239811, + 0.016903217881917953, + 0.016982685774564743, + 0.0025941310450434685, + 0.004083557520061731, + 0.0055045015178620815, + 0.00477353623136878, + 0.0051616099663078785, + 0.0032476414926350117, + 0.008165653795003891, + 0.005131701473146677, + 0.005901970434933901, + 0.004502542782574892, + 0.005602303426712751, + 0.007011263631284237, + 0.005994040984660387, + 0.0052819219417870045, + 0.005703836213797331, + 0.006795844063162804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 2 + ], + "selected_reward": -0.0028199690859764814 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video composed of rapid-cut macro close-up shots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the subject a thick, jet-black ink droplet plunging into still, pure water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sequence begin with an extreme close-up shot from directly above showing the droplet suspended above the surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera shift to a horizontal profile view to show the droplet piercing the surface and sinking?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by a bright, even backlight revealing the inky filaments in detail?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04226015508174896, + 0.07661770284175873, + 0.07399101555347443, + 0.2434741109609604, + 0.04416222125291824, + 0.09086041897535324, + 0.12352906167507172, + 0.10289416462182999, + 0.04580722376704216, + 0.15640394389629364 + ], + "credit_map": [ + 0.006128166802227497, + 0.003022467717528343, + 0.009190785698592663, + 0.003153617726638913, + 0.004643373657017946, + 0.006929511204361916, + 0.004039787221699953, + 0.005152442492544651, + 0.011566080152988434, + 0.01021162886172533, + 0.009715567342936993, + 0.005440967157483101, + 0.009612352587282658, + 0.018558770418167114, + 0.0045427423901855946, + 0.006969592999666929, + 0.008897957392036915, + 0.010890222154557705, + 0.0105924466624856, + 0.008291265927255154, + 0.0059253801591694355, + 0.014912046492099762, + 0.005105517338961363, + 0.009376177564263344, + 0.02907755970954895, + 0.05600401759147644, + 0.01007696334272623, + 0.014888038858771324, + 0.013461875729262829, + 0.09891103953123093, + 0.00938087236136198, + 0.011673755012452602, + 0.005392526276409626, + 0.006411758717149496, + 0.004162255674600601, + 0.00708592077717185, + 0.0036943296436220407, + 0.006138541270047426, + 0.004656642209738493, + 0.00662024924531579, + 0.015247787348926067, + 0.007886374369263649, + 0.008803963661193848, + 0.02342762053012848, + 0.009071125648915768, + 0.0071295322850346565, + 0.009189476259052753, + 0.010104537010192871, + 0.010905850678682327, + 0.01412961259484291, + 0.01322921086102724, + 0.015291768126189709, + 0.02123885229229927, + 0.015555793419480324, + 0.015252869576215744, + 0.017925100401043892, + 0.006865350063890219, + 0.006898162886500359, + 0.021500829607248306, + 0.008883249945938587, + 0.008400309830904007, + 0.007920606061816216, + 0.021091653034090996, + 0.021334005519747734, + 0.003922438248991966, + 0.0036071892827749252, + 0.007991462014615536, + 0.005196383222937584, + 0.005131268873810768, + 0.004082581494003534, + 0.007611517794430256, + 0.008264382369816303, + 0.015102505683898926, + 0.016736043617129326, + 0.024735210463404655, + 0.017970269545912743, + 0.021871674805879593, + 0.015618210658431053, + 0.021185139194130898, + 0.02318488247692585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 6 + ], + "selected_reward": -0.009780817665159702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_tensor([186, 186])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_tensor([186, 186])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8b32f25dfc6b8d0eb891538b37eadd8a8a87d188 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_tensor([186, 186])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot, held close to the ground, synchronizes with a crimson, angular, classic 1980s sedan as it speeds relentlessly through a rain-soaked city street, its wheels spinning rapidly, sending streaks of water splattering across the dark asphalt. Neon signs crisscross the towering facades of the street, their massive traditional Chinese characters flashing in saturated red, blue, and green, casting iridescent shadows on the wet pavement and the polished exteriors of the cars, creating a dazzling display of reflections. As the car speeds along, the lights and buildings in the background are stretched into flowing streaks of light by motion blur. The stark contrast between deep blue-black tones and highly saturated neon colors creates a rich, humid, and nostalgic Hong Kong retro feel.", + "seed": 15944, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15944_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02984574609436095, + "vlm_logprob": -0.02984574609436095, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.02984574609436095, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08796022832393646 + ], + "opposite_logprob_tokens": [ + -2.9629602432250977 + ], + "logprob_reward": -0.08796022832393646, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08796022832393646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09854354709386826, + 0.02478802390396595, + 0.024799708276987076, + 0.05861547961831093, + 0.024945301935076714, + 0.04023292288184166, + 0.06527269631624222, + 0.06554657220840454, + 0.07491738349199295, + 0.522338330745697 + ], + "credit_map": [ + 0.01072234008461237, + 0.012405783869326115, + 0.010657601989805698, + 0.015896350145339966, + 0.010552078485488892, + 0.008854285813868046, + 0.01530120987445116, + 0.014153893105685711, + 0.0020300340838730335, + 0.0027004205621778965, + 0.004934489727020264, + 0.00233712256886065, + 0.002308619674295187, + 0.0016739529091864824, + 0.006082163192331791, + 0.002721223281696439, + 0.0022431581746786833, + 0.002128381049260497, + 0.003145880764350295, + 0.00241090077906847, + 0.002971792593598366, + 0.004503253847360611, + 0.004092695657163858, + 0.003303644945845008, + 0.004270156845450401, + 0.005482223816215992, + 0.007221824023872614, + 0.00595951359719038, + 0.009333038702607155, + 0.009626873768866062, + 0.00867045484483242, + 0.00805139634758234, + 0.00202620355412364, + 0.0017112625064328313, + 0.00233309343457222, + 0.002901694504544139, + 0.003068827325478196, + 0.0026523033156991005, + 0.006107127293944359, + 0.004144791513681412, + 0.0045249927788972855, + 0.004121729638427496, + 0.010236775502562523, + 0.006575124803930521, + 0.002294959267601371, + 0.00500783184543252, + 0.003596613882109523, + 0.0038748946972191334, + 0.006028599571436644, + 0.00962773896753788, + 0.010097972117364407, + 0.013206693343818188, + 0.005802278406918049, + 0.004301483277231455, + 0.008231022395193577, + 0.00797690823674202, + 0.006242201663553715, + 0.004770674277096987, + 0.015941325575113297, + 0.008277270942926407, + 0.0049545858055353165, + 0.0026094543281942606, + 0.01822914369404316, + 0.0045219226740300655, + 0.00610402412712574, + 0.005041126627475023, + 0.015345616266131401, + 0.011043324135243893, + 0.006813574116677046, + 0.006518704816699028, + 0.011247238144278526, + 0.012803774327039719, + 0.02532120607793331, + 0.013230805285274982, + 0.006732767913490534, + 0.025541262701153755, + 0.02594994381070137, + 0.22165316343307495, + 0.03979986160993576, + 0.16410931944847107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.08796022832393646 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821235567331314 + ], + "opposite_logprob_tokens": [ + -4.01821231842041 + ], + "logprob_reward": -0.01821235567331314, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821235567331314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08160049468278885, + 0.07879150658845901, + 0.04814368486404419, + 0.05691741779446602, + 0.04596148803830147, + 0.12169395387172699, + 0.040501199662685394, + 0.16578778624534607, + 0.01804809458553791, + 0.3425544500350952 + ], + "credit_map": [ + 0.007847239263355732, + 0.00992133840918541, + 0.0067978971637785435, + 0.014276829548180103, + 0.00832289271056652, + 0.00821217056363821, + 0.01683914102613926, + 0.009382989257574081, + 0.02200058475136757, + 0.006078579928725958, + 0.02635464444756508, + 0.004928409121930599, + 0.003739520674571395, + 0.002880260581150651, + 0.007716722786426544, + 0.00509277731180191, + 0.005054086912423372, + 0.004628163296729326, + 0.007123625371605158, + 0.00538606708869338, + 0.003919954877346754, + 0.008427215740084648, + 0.007223476190119982, + 0.006381093058735132, + 0.007843825966119766, + 0.005756763741374016, + 0.004576615523546934, + 0.005238769575953484, + 0.009102444164454937, + 0.007119810674339533, + 0.009320416487753391, + 0.007958773523569107, + 0.005210001952946186, + 0.0038576454389840364, + 0.004514165688306093, + 0.005342296790331602, + 0.0046887933276593685, + 0.007529717404395342, + 0.008203841745853424, + 0.006615028250962496, + 0.0137118985876441, + 0.012760655023157597, + 0.02874869294464588, + 0.018620304763317108, + 0.007190983276814222, + 0.01535807829350233, + 0.011401408351957798, + 0.013901934027671814, + 0.0032529335003346205, + 0.004494786262512207, + 0.004745075944811106, + 0.005755261052399874, + 0.006539263762533665, + 0.0045541939325630665, + 0.006287387106567621, + 0.004872296936810017, + 0.015042324550449848, + 0.01113143190741539, + 0.036142412573099136, + 0.022506801411509514, + 0.013282006606459618, + 0.007936445996165276, + 0.04668533429503441, + 0.01306102704256773, + 0.0014981597196310759, + 0.0013668036554008722, + 0.0020138039253652096, + 0.0030725044198334217, + 0.0018312163883820176, + 0.0025061285123229027, + 0.00313450931571424, + 0.002624968532472849, + 0.01614880934357643, + 0.008606775663793087, + 0.004716046620160341, + 0.016177088022232056, + 0.017531702294945717, + 0.143878772854805, + 0.02664213441312313, + 0.10885310173034668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.01821235567331314 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865151360630989 + ], + "opposite_logprob_tokens": [ + -4.7586517333984375 + ], + "logprob_reward": -0.00865151360630989, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865151360630989, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07249900698661804, + 0.18696944415569305, + 0.08601070195436478, + 0.07125350087881088, + 0.022734399884939194, + 0.07429001480340958, + 0.05318774655461311, + 0.03573591634631157, + 0.012029657140374184, + 0.38528957962989807 + ], + "credit_map": [ + 0.008672944270074368, + 0.012032481841742992, + 0.0072054313495755196, + 0.009625912643969059, + 0.006446644198149443, + 0.005320442840456963, + 0.010427606292068958, + 0.012767539359629154, + 0.04314827546477318, + 0.01927008666098118, + 0.060119424015283585, + 0.012810665182769299, + 0.009825488552451134, + 0.0069536492228507996, + 0.021959565579891205, + 0.012882270850241184, + 0.010496886447072029, + 0.009883391670882702, + 0.010225224308669567, + 0.01148214004933834, + 0.007158190011978149, + 0.007252204231917858, + 0.01464973483234644, + 0.014862931333482265, + 0.011445578187704086, + 0.00831780955195427, + 0.007697290740907192, + 0.005153102334588766, + 0.012246990576386452, + 0.007897977717220783, + 0.008975917473435402, + 0.009518831968307495, + 0.0029407255351543427, + 0.0020185401663184166, + 0.002865755697712302, + 0.0020564557053148746, + 0.003852148773148656, + 0.0031720453407615423, + 0.0026564919389784336, + 0.0031722369603812695, + 0.012511133216321468, + 0.006019044667482376, + 0.01662028208374977, + 0.008904794231057167, + 0.005172766745090485, + 0.01176045648753643, + 0.005064099095761776, + 0.008237440139055252, + 0.00521667767316103, + 0.007558073382824659, + 0.005419604480266571, + 0.010761147364974022, + 0.0060962047427892685, + 0.004240457899868488, + 0.00737506989389658, + 0.006520509719848633, + 0.0034067498054355383, + 0.002152112079784274, + 0.007928634062409401, + 0.004393428564071655, + 0.0026364168152213097, + 0.0015406383899971843, + 0.010161514393985271, + 0.0035164207220077515, + 0.0009475317783653736, + 0.0008185702026821673, + 0.0018145004287362099, + 0.0017635601107031107, + 0.0014759900514036417, + 0.001702910871244967, + 0.001955486601218581, + 0.0015511068049818277, + 0.01803986355662346, + 0.009613923728466034, + 0.004348088055849075, + 0.017530115321278572, + 0.019325684756040573, + 0.16347619891166687, + 0.029463205486536026, + 0.123492531478405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 2 + ], + "selected_reward": -0.00865151360630989 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007204797584563494 + ], + "opposite_logprob_tokens": [ + -7.250720500946045 + ], + "logprob_reward": -0.0007204797584563494, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007204797584563494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10142862051725388, + 0.043930165469646454, + 0.07178796827793121, + 0.03366086259484291, + 0.030987706035375595, + 0.21873676776885986, + 0.051266297698020935, + 0.046873606741428375, + 0.016812888905405998, + 0.3845151364803314 + ], + "credit_map": [ + 0.007746417075395584, + 0.008255958557128906, + 0.011714422143995762, + 0.0206229779869318, + 0.009690414182841778, + 0.008070743642747402, + 0.015688883140683174, + 0.019638802856206894, + 0.009652530774474144, + 0.004173817578703165, + 0.01257373858243227, + 0.0032144133001565933, + 0.003331759013235569, + 0.0023099714890122414, + 0.00506965909153223, + 0.003604274010285735, + 0.008248026482760906, + 0.007876263931393623, + 0.009707698598504066, + 0.008436976000666618, + 0.005728337913751602, + 0.008005250245332718, + 0.011901999823749065, + 0.011883413419127464, + 0.005743419751524925, + 0.001975905615836382, + 0.0047887456603348255, + 0.003559298347681761, + 0.002921584527939558, + 0.004807896446436644, + 0.004596030805259943, + 0.005267975851893425, + 0.0028802677989006042, + 0.0029891806188970804, + 0.002472864929586649, + 0.0023375931195914745, + 0.004083904903382063, + 0.008533691056072712, + 0.003442036919295788, + 0.004248163662850857, + 0.029907889664173126, + 0.020773151889443398, + 0.06055707484483719, + 0.034349001944065094, + 0.014932344667613506, + 0.018836678937077522, + 0.018039939925074577, + 0.021340662613511086, + 0.0045731570571660995, + 0.00854512583464384, + 0.007349733728915453, + 0.008821528404951096, + 0.0052222879603505135, + 0.0037641669623553753, + 0.006775403395295143, + 0.0062148915603756905, + 0.004282217938452959, + 0.0033233612775802612, + 0.010783974081277847, + 0.005762649700045586, + 0.003583525773137808, + 0.0021956046111881733, + 0.013446693308651447, + 0.003495576558634639, + 0.0013772642705589533, + 0.0012138222809880972, + 0.0026480993255972862, + 0.0022537822369486094, + 0.0023663050960749388, + 0.0014665520284324884, + 0.002899622078984976, + 0.002587441122159362, + 0.01911504939198494, + 0.009785381145775318, + 0.005250881891697645, + 0.01898510381579399, + 0.019390162080526352, + 0.16314929723739624, + 0.029031382873654366, + 0.11980784684419632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.0007204797584563494 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033684153109788895 + ], + "opposite_logprob_tokens": [ + -3.408684253692627 + ], + "logprob_reward": -0.033684153109788895, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033684153109788895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06826484948396683, + 0.02302694320678711, + 0.04805595427751541, + 0.007780773565173149, + 0.026439111679792404, + 0.03272680565714836, + 0.15739016234874725, + 0.021359942853450775, + 0.03045179322361946, + 0.5845036506652832 + ], + "credit_map": [ + 0.006662448868155479, + 0.00964915007352829, + 0.005245627369731665, + 0.015237179584801197, + 0.0043532392010092735, + 0.004689606372267008, + 0.011772148311138153, + 0.010655452497303486, + 0.0044693537056446075, + 0.002480165334418416, + 0.005419161636382341, + 0.0019519184716045856, + 0.001665223389863968, + 0.0014530457556247711, + 0.0036078845150768757, + 0.001980191096663475, + 0.006660193204879761, + 0.0028026883956044912, + 0.007740003988146782, + 0.0069267055951058865, + 0.003906772006303072, + 0.00523576233536005, + 0.0074369111098349094, + 0.007346919272094965, + 0.0008858964429236948, + 0.0007367850630544126, + 0.0009761325200088322, + 0.0009509086376056075, + 0.0008786959806457162, + 0.0009126692893914878, + 0.0013329789508134127, + 0.0011067073792219162, + 0.0030933693051338196, + 0.0014916874933987856, + 0.00483528571203351, + 0.0030600836034864187, + 0.0019702776335179806, + 0.002321746898815036, + 0.0055714985355734825, + 0.004095162730664015, + 0.005698846187442541, + 0.0034933327697217464, + 0.005200104787945747, + 0.005647493992000818, + 0.002096454380080104, + 0.0027299055363982916, + 0.003545238170772791, + 0.004315430298447609, + 0.017880553379654884, + 0.02167835831642151, + 0.024815663695335388, + 0.027519268915057182, + 0.014579996466636658, + 0.010592646896839142, + 0.021529972553253174, + 0.018793705850839615, + 0.001764007261954248, + 0.0017390688881278038, + 0.004666617605835199, + 0.0027868743054568768, + 0.001609155209735036, + 0.0011592012597247958, + 0.005048925522714853, + 0.002586093032732606, + 0.002239904599264264, + 0.0017959617543965578, + 0.00396207720041275, + 0.0032663545571267605, + 0.005289503838866949, + 0.0026382727082818747, + 0.006327612791210413, + 0.00493210693821311, + 0.02729746885597706, + 0.014363228343427181, + 0.006228251848369837, + 0.026834016665816307, + 0.02847752720117569, + 0.24900008738040924, + 0.044306401163339615, + 0.18799662590026855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.033684153109788895 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, held close to the ground, synchronizes with a crimson, angular, classic 1980s sedan as it speeds relentlessly through a rain-soaked city street, its wheels spinning rapidly, sending streaks of water splattering across the dark asphalt. Neon signs crisscross the towering facades of the street, their massive traditional Chinese characters flashing in saturated red, blue, and green, casting iridescent shadows on the wet pavement and the polished exteriors of the cars, creating a dazzling display of reflections. As the car speeds along, the lights and buildings in the background are stretched into flowing streaks of light by motion blur. The stark contrast between deep blue-black tones and highly saturated neon colors creates a rich, humid, and nostalgic Hong Kong retro feel.", + "seed": 15945, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15945_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02101664823712781, + "vlm_logprob": -0.02101664823712781, + "vlm_yes_no_margin": 4.549999952316284, + "vlm_token_credit": -0.02101664823712781, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04372487962245941 + ], + "opposite_logprob_tokens": [ + -3.793724775314331 + ], + "logprob_reward": -0.04372487962245941, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04372487962245941, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1831115484237671, + 0.07237602770328522, + 0.125581294298172, + 0.05757218226790428, + 0.2825189232826233, + 0.033046722412109375, + 0.043619390577077866, + 0.13429245352745056, + 0.03192100673913956, + 0.03596050664782524 + ], + "credit_map": [ + 0.01605362258851528, + 0.024555642157793045, + 0.02033759467303753, + 0.02592715434730053, + 0.032533857971429825, + 0.0259411484003067, + 0.022074826061725616, + 0.015687700361013412, + 0.007073116954416037, + 0.008780387230217457, + 0.015999171882867813, + 0.011963569559156895, + 0.0060169086791574955, + 0.00656081223860383, + 0.008815343491733074, + 0.007166719529777765, + 0.016014691442251205, + 0.013454774394631386, + 0.013842820189893246, + 0.019138555973768234, + 0.012774638831615448, + 0.014356906525790691, + 0.019776633009314537, + 0.016222260892391205, + 0.005028243642300367, + 0.005944628734141588, + 0.006609827280044556, + 0.00920562818646431, + 0.009120379574596882, + 0.008111762814223766, + 0.0073119886219501495, + 0.006239723414182663, + 0.024263877421617508, + 0.010521731339395046, + 0.040391575545072556, + 0.019157780334353447, + 0.020581625401973724, + 0.015957117080688477, + 0.06317394971847534, + 0.08847128599882126, + 0.0027614536229521036, + 0.003519925056025386, + 0.00689640874043107, + 0.004513968247920275, + 0.0019028499955311418, + 0.0026433232706040144, + 0.006801556795835495, + 0.00400723610073328, + 0.006823630537837744, + 0.00573631701990962, + 0.003968374338001013, + 0.009564368054270744, + 0.0026621809229254723, + 0.002606241498142481, + 0.007703683339059353, + 0.004554593935608864, + 0.012323861941695213, + 0.013403612188994884, + 0.013108337298035622, + 0.031850334256887436, + 0.012125234119594097, + 0.013422485440969467, + 0.025258038192987442, + 0.012800540775060654, + 0.003594914684072137, + 0.00432466110214591, + 0.004720886237919331, + 0.007057857699692249, + 0.0022652184125036, + 0.0019529344281181693, + 0.004976754542440176, + 0.003027780679985881, + 0.005990887526422739, + 0.004714040085673332, + 0.0043244012631475925, + 0.005975865293294191, + 0.004187392070889473, + 0.0028776517137885094, + 0.004111012443900108, + 0.0037792560178786516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.04372487962245941 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029824355617165565 + ], + "opposite_logprob_tokens": [ + -3.5298242568969727 + ], + "logprob_reward": -0.029824355617165565, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029824355617165565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17584076523780823, + 0.08829676359891891, + 0.08004651218652725, + 0.07811427861452103, + 0.2928091287612915, + 0.029539350420236588, + 0.033856771886348724, + 0.06740158051252365, + 0.09440947324037552, + 0.0596853531897068 + ], + "credit_map": [ + 0.016288770362734795, + 0.02539151906967163, + 0.01766766794025898, + 0.025317007675766945, + 0.031524933874607086, + 0.021866172552108765, + 0.021001296117901802, + 0.016783395782113075, + 0.008199794217944145, + 0.01001354306936264, + 0.016868937760591507, + 0.0110075194388628, + 0.009337333962321281, + 0.006705403793603182, + 0.012824679724872112, + 0.013339553959667683, + 0.00950505118817091, + 0.010780658572912216, + 0.00932137668132782, + 0.015283800661563873, + 0.006346653215587139, + 0.004496628884226084, + 0.014598218724131584, + 0.009714123792946339, + 0.0070620346814394, + 0.0075041926465928555, + 0.008025023154914379, + 0.013313140720129013, + 0.010997531935572624, + 0.012609132565557957, + 0.010131105780601501, + 0.008472115732729435, + 0.017270362004637718, + 0.011010196059942245, + 0.03863588348031044, + 0.023935506120324135, + 0.020279578864574432, + 0.011448358185589314, + 0.0763481855392456, + 0.09388107061386108, + 0.0027959721628576517, + 0.0018896424444392323, + 0.005733006168156862, + 0.003974082879722118, + 0.0014331615529954433, + 0.0018640176858752966, + 0.006871495861560106, + 0.004977972246706486, + 0.005447566043585539, + 0.004566522315144539, + 0.003514006268233061, + 0.005224494729191065, + 0.0030358347576111555, + 0.0022615829948335886, + 0.005690600723028183, + 0.004116161726415157, + 0.00797960814088583, + 0.009979021735489368, + 0.006001035682857037, + 0.010391788557171822, + 0.01051782350987196, + 0.005199685227125883, + 0.009552269242703915, + 0.007780349347740412, + 0.009895913302898407, + 0.014794008806347847, + 0.01639075018465519, + 0.015414596535265446, + 0.007698297966271639, + 0.007183487992733717, + 0.013571614399552345, + 0.009460804983973503, + 0.0078060864470899105, + 0.010361416265368462, + 0.008269698359072208, + 0.006910500582307577, + 0.005685577169060707, + 0.006794319022446871, + 0.00686425156891346, + 0.006993504241108894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.029824355617165565 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14200738072395325, + 0.054083649069070816, + 0.06616749614477158, + 0.11257600784301758, + 0.3113994896411896, + 0.0420074425637722, + 0.057125721126794815, + 0.13738460838794708, + 0.03735833615064621, + 0.03988982364535332 + ], + "credit_map": [ + 0.014014245942234993, + 0.020742129534482956, + 0.017171842977404594, + 0.02286463975906372, + 0.020172104239463806, + 0.015512967482209206, + 0.01687881536781788, + 0.014650637283921242, + 0.005883479490876198, + 0.007129793055355549, + 0.01096284855157137, + 0.00805003009736538, + 0.005125615280121565, + 0.004188865888863802, + 0.007422646041959524, + 0.0053203715942800045, + 0.011502958834171295, + 0.00793746393173933, + 0.00926755927503109, + 0.007959111593663692, + 0.0075170909985899925, + 0.006571818608790636, + 0.00804157368838787, + 0.007369920611381531, + 0.029677776619791985, + 0.023189473897218704, + 0.011052843183279037, + 0.009457116015255451, + 0.008886353112757206, + 0.007945149205625057, + 0.012161267921328545, + 0.01020602136850357, + 0.033899299800395966, + 0.01493880432099104, + 0.09495431184768677, + 0.050163645297288895, + 0.016539933159947395, + 0.01503235287964344, + 0.03381145000457764, + 0.052059683948755264, + 0.004131217487156391, + 0.0036812995094805956, + 0.0055671981535851955, + 0.004222970921546221, + 0.0023902130778878927, + 0.002284112386405468, + 0.012005883269011974, + 0.007724548690021038, + 0.007540179882198572, + 0.009652132168412209, + 0.005298879463225603, + 0.009500962682068348, + 0.004926763940602541, + 0.0050052921287715435, + 0.00826653279364109, + 0.006934980861842632, + 0.027391351759433746, + 0.018725348636507988, + 0.012814013287425041, + 0.020983735099434853, + 0.014280365779995918, + 0.014290014281868935, + 0.018135935068130493, + 0.010763835161924362, + 0.0055293976329267025, + 0.006800732109695673, + 0.0037910770624876022, + 0.0068654282949864864, + 0.002379945944994688, + 0.0019470113329589367, + 0.006213849410414696, + 0.0038308952935039997, + 0.004863270092755556, + 0.006553204730153084, + 0.0049652415327727795, + 0.007337267976254225, + 0.004809224978089333, + 0.0026169053744524717, + 0.005110400263220072, + 0.003634310094639659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.012544536031782627 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008141063735820353 + ], + "opposite_logprob_tokens": [ + -7.125813961029053 + ], + "logprob_reward": -0.0008141063735820353, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008141063735820353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11511643975973129, + 0.6615644097328186, + 0.014018571935594082, + 0.02783489227294922, + 0.056242186576128006, + 0.01231436338275671, + 0.017668001353740692, + 0.022714704275131226, + 0.04092426970601082, + 0.031602244824171066 + ], + "credit_map": [ + 0.011292420327663422, + 0.017955942079424858, + 0.014097116887569427, + 0.01743525080382824, + 0.017105262726545334, + 0.013922381214797497, + 0.013384630903601646, + 0.00992342084646225, + 0.0742444321513176, + 0.05999261885881424, + 0.1470511257648468, + 0.1054278314113617, + 0.04412044957280159, + 0.05205687880516052, + 0.10053432732820511, + 0.07813668996095657, + 0.0029556050430983305, + 0.0023639914579689503, + 0.001657083979807794, + 0.0016420312458649278, + 0.0012666324619203806, + 0.0014126462629064918, + 0.0014804861275479198, + 0.0012400930281728506, + 0.004297161474823952, + 0.004260637331753969, + 0.003093644278123975, + 0.005767823662608862, + 0.0027196770533919334, + 0.002021474065259099, + 0.0032212764490395784, + 0.002453196793794632, + 0.00627109594643116, + 0.0027596272993832827, + 0.017702264711260796, + 0.00905220489948988, + 0.002779583213850856, + 0.002119713928550482, + 0.01005635503679514, + 0.005501333624124527, + 0.001224555540829897, + 0.0012307295110076666, + 0.0020278107840567827, + 0.001530289649963379, + 0.000765244651120156, + 0.0008029792224988341, + 0.0031470630783587694, + 0.0015856889076530933, + 0.003522503189742565, + 0.0027059565763920546, + 0.0016913654981181026, + 0.0034540025517344475, + 0.0011415655026212335, + 0.0008525774464942515, + 0.0026303816121071577, + 0.001669647521339357, + 0.0033050994388759136, + 0.0033198136370629072, + 0.0017242255853489041, + 0.003876983653753996, + 0.003217854304239154, + 0.002344868378713727, + 0.0031906557269394398, + 0.0017352018039673567, + 0.008564633317291737, + 0.006280518136918545, + 0.005060851573944092, + 0.005716321989893913, + 0.002393037313595414, + 0.0025085348170250654, + 0.006930644158273935, + 0.003469723742455244, + 0.0041223675943911076, + 0.004357536789029837, + 0.004781602881848812, + 0.006352373398840427, + 0.0036588183138519526, + 0.0021591298282146454, + 0.0036917172838002443, + 0.0024786966387182474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0008141063735820353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018175363540649414 + ], + "opposite_logprob_tokens": [ + -4.01817512512207 + ], + "logprob_reward": -0.018175363540649414, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018175363540649414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05238410457968712, + 0.41888290643692017, + 0.05830604210495949, + 0.029667885974049568, + 0.19559787213802338, + 0.04683408513665199, + 0.08584123104810715, + 0.025442583486437798, + 0.03467121347784996, + 0.052372124046087265 + ], + "credit_map": [ + 0.00570492260158062, + 0.007737504318356514, + 0.005635576788336039, + 0.00898214802145958, + 0.0051744915544986725, + 0.004398791119456291, + 0.008616871200501919, + 0.006133800372481346, + 0.04403265565633774, + 0.03716157004237175, + 0.09164050966501236, + 0.06598091870546341, + 0.026855239644646645, + 0.034207504242658615, + 0.06989879161119461, + 0.049105700105428696, + 0.0057479096576571465, + 0.007199558895081282, + 0.008324693888425827, + 0.010857513174414635, + 0.005004473961889744, + 0.005072328727692366, + 0.009903380647301674, + 0.006196180358529091, + 0.0044720848090946674, + 0.004710971377789974, + 0.0033005753066390753, + 0.005131857469677925, + 0.0024901246652007103, + 0.002547550480812788, + 0.004031921736896038, + 0.0029828008264303207, + 0.020851237699389458, + 0.010869630612432957, + 0.058065466582775116, + 0.032407715916633606, + 0.010981759056448936, + 0.007989270612597466, + 0.0350993275642395, + 0.019333461299538612, + 0.004020054824650288, + 0.002269573975354433, + 0.010533259250223637, + 0.0036637266166508198, + 0.0019615550991147757, + 0.001975045073777437, + 0.014870058745145798, + 0.0075408099219202995, + 0.010220436379313469, + 0.006473261397331953, + 0.009213786572217941, + 0.026422269642353058, + 0.00415283627808094, + 0.003969722893089056, + 0.016510674729943275, + 0.00887824036180973, + 0.004541650414466858, + 0.0025673226919025183, + 0.002472503576427698, + 0.004552493803203106, + 0.003530193353071809, + 0.0017352041322737932, + 0.0037113747093826532, + 0.0023318405728787184, + 0.008289653807878494, + 0.004116397816687822, + 0.005349956918507814, + 0.004932741168886423, + 0.0021896392572671175, + 0.0018378796521574259, + 0.005389069207012653, + 0.0025658756494522095, + 0.00862541701644659, + 0.006117418874055147, + 0.006415799260139465, + 0.011126115918159485, + 0.0055287061259150505, + 0.003258025273680687, + 0.0066543505527079105, + 0.004646288696676493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 6 + ], + "selected_reward": -0.018175363540649414 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, held close to the ground, synchronizes with a crimson, angular, classic 1980s sedan as it speeds relentlessly through a rain-soaked city street, its wheels spinning rapidly, sending streaks of water splattering across the dark asphalt. Neon signs crisscross the towering facades of the street, their massive traditional Chinese characters flashing in saturated red, blue, and green, casting iridescent shadows on the wet pavement and the polished exteriors of the cars, creating a dazzling display of reflections. As the car speeds along, the lights and buildings in the background are stretched into flowing streaks of light by motion blur. The stark contrast between deep blue-black tones and highly saturated neon colors creates a rich, humid, and nostalgic Hong Kong retro feel.", + "seed": 15946, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15946_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022444345301482826, + "vlm_logprob": -0.022444345301482826, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.022444345301482826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052485015243291855 + ], + "opposite_logprob_tokens": [ + -3.6774849891662598 + ], + "logprob_reward": -0.052485015243291855, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.052485015243291855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12725110352039337, + 0.04475606232881546, + 0.05143599212169647, + 0.16123338043689728, + 0.08699306845664978, + 0.11678175628185272, + 0.1728651523590088, + 0.08784149587154388, + 0.07055926322937012, + 0.08028274774551392 + ], + "credit_map": [ + 0.01149976346641779, + 0.01112762838602066, + 0.016173478215932846, + 0.01591501198709011, + 0.009878185577690601, + 0.022307250648736954, + 0.020327655598521233, + 0.020022118464112282, + 0.00381612335331738, + 0.003595709567889571, + 0.005822180304676294, + 0.006618589162826538, + 0.0038729447405785322, + 0.008290068246424198, + 0.005771834868937731, + 0.006968609988689423, + 0.003928516060113907, + 0.008908640593290329, + 0.002458093222230673, + 0.00481585506349802, + 0.004758788272738457, + 0.012819821015000343, + 0.007806312292814255, + 0.005939963739365339, + 0.006799210328608751, + 0.00488369120284915, + 0.0059073069132864475, + 0.007977593690156937, + 0.007074616849422455, + 0.007860145531594753, + 0.06412555277347565, + 0.0566052570939064, + 0.008242115378379822, + 0.010181115940213203, + 0.010579610243439674, + 0.015221460722386837, + 0.005340774077922106, + 0.005550140980631113, + 0.015520056709647179, + 0.0163577888160944, + 0.00973554328083992, + 0.010321550071239471, + 0.020366132259368896, + 0.02082422561943531, + 0.01239827647805214, + 0.011256655678153038, + 0.01597745344042778, + 0.01590190827846527, + 0.014289380982518196, + 0.014818282797932625, + 0.025379495695233345, + 0.021280787885189056, + 0.012571650557219982, + 0.014707569032907486, + 0.03362516313791275, + 0.036192819476127625, + 0.006210057530552149, + 0.0031017076689749956, + 0.025020215660333633, + 0.007363420445472002, + 0.003883314784616232, + 0.004067067988216877, + 0.009672978892922401, + 0.02852272428572178, + 0.005399043206125498, + 0.0044615729711949825, + 0.009352047927677631, + 0.014501001685857773, + 0.009540955536067486, + 0.005904654506593943, + 0.009875547140836716, + 0.011524436995387077, + 0.00669518020004034, + 0.007053415756672621, + 0.010503627359867096, + 0.012438713572919369, + 0.010668144561350346, + 0.006153753958642483, + 0.014792599715292454, + 0.01197730377316475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.052485015243291855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012563959695398808 + ], + "opposite_logprob_tokens": [ + -4.387564182281494 + ], + "logprob_reward": -0.012563959695398808, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012563959695398808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1415516585111618, + 0.037458132952451706, + 0.11473550647497177, + 0.08789733797311783, + 0.06917707622051239, + 0.050676144659519196, + 0.08740647882223129, + 0.05095818266272545, + 0.2401394546031952, + 0.1199999749660492 + ], + "credit_map": [ + 0.012356894090771675, + 0.01249956525862217, + 0.013418793678283691, + 0.021521350368857384, + 0.010440802201628685, + 0.028850389644503593, + 0.02045999839901924, + 0.022003885358572006, + 0.003300749696791172, + 0.0031755755189806223, + 0.004280407447367907, + 0.005461306776851416, + 0.0026052254252135754, + 0.006879803724586964, + 0.00632114615291357, + 0.00543392077088356, + 0.006279545370489359, + 0.008582641370594501, + 0.004736375994980335, + 0.010096509009599686, + 0.011649838648736477, + 0.040473271161317825, + 0.01725197397172451, + 0.015665357932448387, + 0.005562205798923969, + 0.006070836912840605, + 0.006337543483823538, + 0.007536000106483698, + 0.00956268236041069, + 0.017181573435664177, + 0.01983354240655899, + 0.015812957659363747, + 0.004073737189173698, + 0.007641100790351629, + 0.004641363397240639, + 0.004984542727470398, + 0.007437107618898153, + 0.008550398983061314, + 0.01515305694192648, + 0.016695767641067505, + 0.006213424261659384, + 0.003919275943189859, + 0.005245002917945385, + 0.008466562256217003, + 0.004965761676430702, + 0.004405911546200514, + 0.007734351325780153, + 0.009725860320031643, + 0.006752061191946268, + 0.009361643344163895, + 0.009257830679416656, + 0.01460499968379736, + 0.008337888866662979, + 0.009352942928671837, + 0.012672241777181625, + 0.017066875472664833, + 0.0036858348175883293, + 0.0035331707913428545, + 0.004979275166988373, + 0.00956017803400755, + 0.00733994971960783, + 0.005905116442590952, + 0.008931189775466919, + 0.007023468613624573, + 0.023116208612918854, + 0.023228133097290993, + 0.04626105725765228, + 0.05032789707183838, + 0.023170866072177887, + 0.01803489401936531, + 0.027751374989748, + 0.028249038383364677, + 0.00840245746076107, + 0.008894721046090126, + 0.02572033368051052, + 0.013753601349890232, + 0.010946458205580711, + 0.012487070634961128, + 0.014846711419522762, + 0.024948623031377792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.012563959695398808 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647613227367401 + ], + "opposite_logprob_tokens": [ + -4.758647441864014 + ], + "logprob_reward": -0.008647613227367401, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647613227367401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11966389417648315, + 0.11588931083679199, + 0.05557972565293312, + 0.1603858470916748, + 0.06863539665937424, + 0.32886239886283875, + 0.01732979342341423, + 0.03113492950797081, + 0.051865216344594955, + 0.050653498619794846 + ], + "credit_map": [ + 0.01661645621061325, + 0.011849182657897472, + 0.012439889833331108, + 0.01562977209687233, + 0.012385137379169464, + 0.022922711446881294, + 0.012694436125457287, + 0.015126303769648075, + 0.010852010920643806, + 0.007609826512634754, + 0.010480034165084362, + 0.01799054816365242, + 0.01226320955902338, + 0.01665160432457924, + 0.01713191531598568, + 0.02291015349328518, + 0.00490484619513154, + 0.007613584864884615, + 0.0025191721506416798, + 0.004688406828790903, + 0.0066067930310964584, + 0.016646919772028923, + 0.005216740071773529, + 0.007383266463875771, + 0.016650645062327385, + 0.012887297198176384, + 0.010021490044891834, + 0.012239798903465271, + 0.009357063099741936, + 0.008553590625524521, + 0.051389411091804504, + 0.0392865389585495, + 0.009730955585837364, + 0.0078598503023386, + 0.005797747988253832, + 0.008204741403460503, + 0.00748751824721694, + 0.007220603991299868, + 0.010924635455012321, + 0.011409343220293522, + 0.023544924333691597, + 0.022355232387781143, + 0.034633513540029526, + 0.046495020389556885, + 0.05260578915476799, + 0.04826914519071579, + 0.04821598157286644, + 0.05274277552962303, + 0.0016049423720687628, + 0.001619927235879004, + 0.002292470308020711, + 0.002337960759177804, + 0.001484470907598734, + 0.001957266591489315, + 0.0027172944974154234, + 0.0033154620323330164, + 0.0018394800135865808, + 0.0013696985552087426, + 0.008719493634998798, + 0.002430650172755122, + 0.002120812190696597, + 0.0018074705731123686, + 0.0036138221621513367, + 0.009233499877154827, + 0.004717733711004257, + 0.004717246163636446, + 0.009688257239758968, + 0.011216009967029095, + 0.005636036396026611, + 0.003887639380991459, + 0.005866352003067732, + 0.0061359405517578125, + 0.0037146888207644224, + 0.0034534146543592215, + 0.006092069670557976, + 0.007158081978559494, + 0.004515766631811857, + 0.004896780010312796, + 0.010516705922782421, + 0.01030599232763052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.008647613227367401 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000439428084064275 + ], + "opposite_logprob_tokens": [ + -7.750439643859863 + ], + "logprob_reward": -0.000439428084064275, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.000439428084064275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28379562497138977, + 0.04998255893588066, + 0.05644536390900612, + 0.04485967010259628, + 0.06488095968961716, + 0.04746026173233986, + 0.10787283629179001, + 0.061737094074487686, + 0.046598345041275024, + 0.2363673448562622 + ], + "credit_map": [ + 0.020988132804632187, + 0.04133950546383858, + 0.03537743166089058, + 0.05093513801693916, + 0.015478434041142464, + 0.017000481486320496, + 0.04140857234597206, + 0.061267927289009094, + 0.004957371857017279, + 0.004591924604028463, + 0.005076141562312841, + 0.006081380881369114, + 0.00389169086702168, + 0.012296889908611774, + 0.005357399117201567, + 0.007729757577180862, + 0.007037147879600525, + 0.010702217929065228, + 0.0023391179274767637, + 0.0035822479985654354, + 0.004716148599982262, + 0.013999305665493011, + 0.006995419505983591, + 0.007073757704347372, + 0.002864842303097248, + 0.002487445715814829, + 0.0023258058354258537, + 0.003075321903452277, + 0.0033616311848163605, + 0.00635033193975687, + 0.013167553581297398, + 0.011226738803088665, + 0.0065043047070503235, + 0.005930835381150246, + 0.00915673840790987, + 0.011502106674015522, + 0.004845506977289915, + 0.004598618019372225, + 0.009447832591831684, + 0.012895009480416775, + 0.003503394778817892, + 0.003923020325601101, + 0.006419478915631771, + 0.007725513074547052, + 0.005586533807218075, + 0.005818248260766268, + 0.006731642410159111, + 0.0077524310909211636, + 0.009025311097502708, + 0.006930192932486534, + 0.016461385414004326, + 0.012148810550570488, + 0.009473786689341068, + 0.008831927552819252, + 0.02684698812663555, + 0.01815442182123661, + 0.0036697094328701496, + 0.0027752562891691923, + 0.014661502093076706, + 0.006052709184587002, + 0.004192374646663666, + 0.004632106516510248, + 0.008335798047482967, + 0.01741763949394226, + 0.004970650188624859, + 0.003754582954570651, + 0.005845468956977129, + 0.009444596245884895, + 0.005012058652937412, + 0.004549527540802956, + 0.005749846808612347, + 0.007271616254001856, + 0.013574457727372646, + 0.015241336077451706, + 0.05921229347586632, + 0.0196619164198637, + 0.011497518979012966, + 0.02501828223466873, + 0.024653330445289612, + 0.06750820577144623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.000439428084064275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038085710257291794 + ], + "opposite_logprob_tokens": [ + -3.288085699081421 + ], + "logprob_reward": -0.038085710257291794, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038085710257291794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05644096061587334, + 0.03606685250997543, + 0.02155095525085926, + 0.03641732409596443, + 0.09130820631980896, + 0.0561758317053318, + 0.2098764181137085, + 0.051533281803131104, + 0.05026720464229584, + 0.3903629779815674 + ], + "credit_map": [ + 0.0065451860427856445, + 0.005665115546435118, + 0.006835778709501028, + 0.008868644014000893, + 0.004846891388297081, + 0.006379376631230116, + 0.008515264838933945, + 0.00878470204770565, + 0.002587580354884267, + 0.002214729320257902, + 0.0034472572151571512, + 0.005358988884836435, + 0.002179752802476287, + 0.007527114357799292, + 0.005367306526750326, + 0.007384123746305704, + 0.0024166316725313663, + 0.003659976413473487, + 0.001367458258755505, + 0.0027093570679426193, + 0.0027822894044220448, + 0.0025155767798423767, + 0.0034757766406983137, + 0.0026238891296088696, + 0.0029903941322118044, + 0.002759265946224332, + 0.0023658957798033953, + 0.0028382516466081142, + 0.0027539420407265425, + 0.002124690217897296, + 0.011395936831831932, + 0.009188948199152946, + 0.007179415784776211, + 0.010196481831371784, + 0.011939956806600094, + 0.014182684943079948, + 0.00727272592484951, + 0.007023667450994253, + 0.014971280470490456, + 0.01854199357330799, + 0.004172082059085369, + 0.005204733926802874, + 0.008616243489086628, + 0.010087920352816582, + 0.0053819334134459496, + 0.007602476980537176, + 0.007756873033940792, + 0.007353567983955145, + 0.016656940802931786, + 0.014920531772077084, + 0.032163605093955994, + 0.0234843622893095, + 0.013339776545763016, + 0.015872573480010033, + 0.0472157821059227, + 0.04622284322977066, + 0.0030147554352879524, + 0.001326205674558878, + 0.01622401736676693, + 0.003703275928273797, + 0.001912695704959333, + 0.0030972338281571865, + 0.0038188262842595577, + 0.018436269834637642, + 0.0046706064604222775, + 0.0036332346498966217, + 0.009470685385167599, + 0.007599380798637867, + 0.003999166656285524, + 0.004364603664726019, + 0.006147571373730898, + 0.01038195937871933, + 0.020400458946824074, + 0.02694142796099186, + 0.0736677497625351, + 0.043584730476140976, + 0.02654220722615719, + 0.03455941751599312, + 0.05066230893135071, + 0.11400467157363892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 4 + ], + "selected_reward": -0.038085710257291794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, held close to the ground, synchronizes with a crimson, angular, classic 1980s sedan as it speeds relentlessly through a rain-soaked city street, its wheels spinning rapidly, sending streaks of water splattering across the dark asphalt. Neon signs crisscross the towering facades of the street, their massive traditional Chinese characters flashing in saturated red, blue, and green, casting iridescent shadows on the wet pavement and the polished exteriors of the cars, creating a dazzling display of reflections. As the car speeds along, the lights and buildings in the background are stretched into flowing streaks of light by motion blur. The stark contrast between deep blue-black tones and highly saturated neon colors creates a rich, humid, and nostalgic Hong Kong retro feel.", + "seed": 15947, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_1/index_186_seed_15947_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033092559105716644, + "vlm_logprob": -0.033092559105716644, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.033092559105716644, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1079164445400238 + ], + "opposite_logprob_tokens": [ + -2.6079163551330566 + ], + "logprob_reward": -0.1079164445400238, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.1079164445400238, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07004781812429428, + 0.04434741660952568, + 0.5168482065200806, + 0.03577293083071709, + 0.06165342405438423, + 0.033175814896821976, + 0.08234633505344391, + 0.07282000780105591, + 0.05303553119301796, + 0.029952503740787506 + ], + "credit_map": [ + 0.008158495649695396, + 0.009112373925745487, + 0.006466845516115427, + 0.01081276498734951, + 0.01142769493162632, + 0.009540040045976639, + 0.006557794287800789, + 0.007971818558871746, + 0.0045646317303180695, + 0.00601210817694664, + 0.0037361658178269863, + 0.005147682968527079, + 0.004201545845717192, + 0.006984919775277376, + 0.007575595285743475, + 0.006124773528426886, + 0.027955390512943268, + 0.17264439165592194, + 0.16310392320156097, + 0.027334317564964294, + 0.022824270650744438, + 0.04714007303118706, + 0.01940818317234516, + 0.036437712609767914, + 0.0035215835087001324, + 0.005536409094929695, + 0.004480902571231127, + 0.006849558558315039, + 0.004021279513835907, + 0.0030533664394170046, + 0.004389497451484203, + 0.00392033951357007, + 0.006086781620979309, + 0.003489499446004629, + 0.004762447439134121, + 0.004987814929336309, + 0.007274348754435778, + 0.010041498579084873, + 0.01432646345347166, + 0.010684575885534286, + 0.002862270688638091, + 0.004480539355427027, + 0.004327276721596718, + 0.005086639430373907, + 0.002127413870766759, + 0.003602544078603387, + 0.006738997530192137, + 0.0039501371793448925, + 0.007481935899704695, + 0.008008565753698349, + 0.008578645996749401, + 0.0175784882158041, + 0.00999755971133709, + 0.008760109543800354, + 0.011056271381676197, + 0.010884766466915607, + 0.009197953157126904, + 0.008579771965742111, + 0.009835581295192242, + 0.008721101097762585, + 0.004861325491219759, + 0.007295053917914629, + 0.011960265226662159, + 0.012368963100016117, + 0.0068006934598088264, + 0.005331610795110464, + 0.0051824492402374744, + 0.008261594921350479, + 0.005045948550105095, + 0.005167949479073286, + 0.00841435045003891, + 0.008830939419567585, + 0.0025153837632387877, + 0.002893673023208976, + 0.003646211000159383, + 0.00535253481939435, + 0.003785479813814163, + 0.0029980114195495844, + 0.004319154657423496, + 0.0044420575723052025 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.1079164445400238 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982921525835991 + ], + "opposite_logprob_tokens": [ + -3.529829263687134 + ], + "logprob_reward": -0.02982921525835991, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982921525835991, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062443140894174576, + 0.06493254750967026, + 0.5286068320274353, + 0.02384793758392334, + 0.02772515080869198, + 0.0032130314502865076, + 0.03379061818122864, + 0.07799649983644485, + 0.1255321353673935, + 0.05191211774945259 + ], + "credit_map": [ + 0.00481499033048749, + 0.006018796470016241, + 0.006128587294369936, + 0.010440005920827389, + 0.009741213172674179, + 0.009159346111118793, + 0.008701300248503685, + 0.007438903674483299, + 0.0048768785782158375, + 0.006100554019212723, + 0.007174962665885687, + 0.006652685813605785, + 0.007320282515138388, + 0.01565573178231716, + 0.009720983915030956, + 0.007430466823279858, + 0.031937479972839355, + 0.13614344596862793, + 0.20416967570781708, + 0.04809047281742096, + 0.025880759581923485, + 0.01985798589885235, + 0.022646550089120865, + 0.03988046944141388, + 0.0016500657657161355, + 0.003502883017063141, + 0.002061400795355439, + 0.004683339037001133, + 0.004232408478856087, + 0.0017088614404201508, + 0.003358351532369852, + 0.002650628099218011, + 0.002455698559060693, + 0.001967787276953459, + 0.0023450094740837812, + 0.003426198847591877, + 0.004315661266446114, + 0.005514355842024088, + 0.003625270677730441, + 0.004075170494616032, + 0.00031672746990807354, + 0.00032627975451759994, + 0.00038390536792576313, + 0.0004204063443467021, + 0.00025638772058300674, + 0.00029646744951605797, + 0.0007312444504350424, + 0.0004816128348466009, + 0.00417548231780529, + 0.002908045193180442, + 0.003539450466632843, + 0.004408876411616802, + 0.0021303424146026373, + 0.0029214138630777597, + 0.005487801972776651, + 0.008219207637012005, + 0.009461098350584507, + 0.007978510111570358, + 0.01066378504037857, + 0.01393404696136713, + 0.006325778551399708, + 0.009047339670360088, + 0.008426421321928501, + 0.012159515172243118, + 0.008088016882538795, + 0.011767081916332245, + 0.006238727830350399, + 0.010691746138036251, + 0.017654458060860634, + 0.01705353893339634, + 0.026449020951986313, + 0.027589550241827965, + 0.0053290436044335365, + 0.005177122075110674, + 0.004558891989290714, + 0.0064470465295016766, + 0.006519454065710306, + 0.0036301985383033752, + 0.00913371704518795, + 0.01111664716154337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.02982921525835991 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00595959834754467 + ], + "opposite_logprob_tokens": [ + -5.130959510803223 + ], + "logprob_reward": -0.00595959834754467, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00595959834754467, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12170325964689255, + 0.11469120532274246, + 0.11819455772638321, + 0.1510511338710785, + 0.10940703004598618, + 0.04273552447557449, + 0.04169035330414772, + 0.10074460506439209, + 0.0716547816991806, + 0.12812753021717072 + ], + "credit_map": [ + 0.015537534840404987, + 0.022305872291326523, + 0.01318841427564621, + 0.017499307170510292, + 0.013303196988999844, + 0.014734962955117226, + 0.013950138352811337, + 0.011183845810592175, + 0.01595858670771122, + 0.01923290081322193, + 0.011028996668756008, + 0.018158672377467155, + 0.012111865915358067, + 0.0160700511187315, + 0.011433265171945095, + 0.010696875862777233, + 0.0072039351798594, + 0.038952238857746124, + 0.021681249141693115, + 0.009049100801348686, + 0.006519597489386797, + 0.018138445913791656, + 0.006721717771142721, + 0.00992827769368887, + 0.015131986699998379, + 0.03686189651489258, + 0.02156023494899273, + 0.02390729822218418, + 0.016436554491519928, + 0.011961322277784348, + 0.011258456856012344, + 0.013933390378952026, + 0.009457672014832497, + 0.005837998818606138, + 0.010132543742656708, + 0.007569146808236837, + 0.030519139021635056, + 0.011372076347470284, + 0.012897046282887459, + 0.02162141166627407, + 0.003701159032061696, + 0.0062300278805196285, + 0.006386670749634504, + 0.007215182296931744, + 0.003088888246566057, + 0.004362656734883785, + 0.0066489363089203835, + 0.005102007184177637, + 0.006185765843838453, + 0.0038131976034492254, + 0.003222829196602106, + 0.010627691633999348, + 0.004381056874990463, + 0.003953409381210804, + 0.005658971611410379, + 0.003847436048090458, + 0.017317984253168106, + 0.011367757804691792, + 0.01301953848451376, + 0.015841349959373474, + 0.007658612448722124, + 0.009686725214123726, + 0.010756370611488819, + 0.015096267685294151, + 0.012028153985738754, + 0.01112350169569254, + 0.005896535236388445, + 0.007829733192920685, + 0.006322929635643959, + 0.008823377080261707, + 0.009101812727749348, + 0.01052874606102705, + 0.009103749878704548, + 0.009687851183116436, + 0.008451088331639767, + 0.012245478108525276, + 0.008414182811975479, + 0.009958861395716667, + 0.030798442661762238, + 0.03946787863969803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.00595959834754467 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011834054021164775 + ], + "opposite_logprob_tokens": [ + -6.75118350982666 + ], + "logprob_reward": -0.0011834054021164775, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011834054021164775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027236072346568108, + 0.0195890162140131, + 0.5659330487251282, + 0.13187448680400848, + 0.06504656374454498, + 0.007267671171575785, + 0.018706610426306725, + 0.11594513803720474, + 0.022244825959205627, + 0.026156626641750336 + ], + "credit_map": [ + 0.0041281841695308685, + 0.003892733948305249, + 0.0021793069317936897, + 0.004384034313261509, + 0.003568403422832489, + 0.0034317418467253447, + 0.003420338500291109, + 0.0022313285153359175, + 0.0019661118276417255, + 0.0022247880697250366, + 0.0018306063720956445, + 0.0024824836291372776, + 0.0017456371570006013, + 0.0036883193533867598, + 0.0030807240400463343, + 0.0025703466963022947, + 0.03246529772877693, + 0.17307084798812866, + 0.20248746871948242, + 0.037328414618968964, + 0.024075740948319435, + 0.03405304625630379, + 0.021750684827566147, + 0.04070153459906578, + 0.010447713546454906, + 0.019017022103071213, + 0.027856217697262764, + 0.03157658502459526, + 0.009049906395375729, + 0.013645860366523266, + 0.007550726179033518, + 0.012730459682643414, + 0.004078974016010761, + 0.0035966881550848484, + 0.005945554468780756, + 0.007744213566184044, + 0.017408622428774834, + 0.006577777210623026, + 0.007224940229207277, + 0.01246979646384716, + 0.0009350157924927771, + 0.0010365522466599941, + 0.000992363435216248, + 0.0009777413215488195, + 0.0004049797134939581, + 0.0008488622261211276, + 0.001121772569604218, + 0.0009503833716735244, + 0.001805438892915845, + 0.0024490244686603546, + 0.002220218302682042, + 0.003192589385434985, + 0.0020325423683971167, + 0.0016318145208060741, + 0.0023895008489489555, + 0.0029854816384613514, + 0.015586724504828453, + 0.01040850579738617, + 0.015040464699268341, + 0.025326989591121674, + 0.006917664781212807, + 0.01012630108743906, + 0.011637463234364986, + 0.020901024341583252, + 0.003918981645256281, + 0.0030482388101518154, + 0.002563332673162222, + 0.0028851984534412622, + 0.0019276116508990526, + 0.002029871568083763, + 0.003108619712293148, + 0.002762972842901945, + 0.0018681016517803073, + 0.002465213416144252, + 0.0018598998431116343, + 0.002732557477429509, + 0.0028713643550872803, + 0.0037004430778324604, + 0.004956968128681183, + 0.0057020788080990314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 7 + ], + "selected_reward": -0.0011834054021164775 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle close to the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a crimson, angular, classic 1980s sedan visible speeding through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are neon signs with traditional Chinese characters flashing in red, blue, and green on the buildings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background showing motion blur with lights and buildings stretched into flowing streaks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene feature a dark asphalt street with rain-soaked conditions and water splattering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574131980538368 + ], + "opposite_logprob_tokens": [ + -3.8955740928649902 + ], + "logprob_reward": -0.020574131980538368, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574131980538368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03905653953552246, + 0.033899009227752686, + 0.2556523382663727, + 0.08853119611740112, + 0.048115409910678864, + 0.022638076916337013, + 0.055432017892599106, + 0.2668781578540802, + 0.11758265644311905, + 0.07221458107233047 + ], + "credit_map": [ + 0.004392861854285002, + 0.005298185627907515, + 0.004496539477258921, + 0.008261255919933319, + 0.004234702792018652, + 0.0042359125800430775, + 0.00457181828096509, + 0.0035652630031108856, + 0.0030223203357309103, + 0.0033230804838240147, + 0.0038557094521820545, + 0.006444945931434631, + 0.003260857192799449, + 0.006212441250681877, + 0.004448912106454372, + 0.0033307436387985945, + 0.013939286582171917, + 0.0823039561510086, + 0.0845101848244667, + 0.014349006116390228, + 0.011105714365839958, + 0.021443845704197884, + 0.009703120216727257, + 0.018297215923666954, + 0.0085483118891716, + 0.01862400770187378, + 0.01579754799604416, + 0.011365045793354511, + 0.011469555087387562, + 0.007700294256210327, + 0.008272037841379642, + 0.006754401605576277, + 0.004316095262765884, + 0.0037064349744468927, + 0.004911959636956453, + 0.005936373025178909, + 0.004368656314909458, + 0.006239338777959347, + 0.011415387503802776, + 0.0072211637161672115, + 0.0025562664959579706, + 0.004516263958066702, + 0.0023192556109279394, + 0.004030600655823946, + 0.0017127863829955459, + 0.0014578072587028146, + 0.003189250361174345, + 0.0028558450285345316, + 0.004694069270044565, + 0.006607904098927975, + 0.006824424024671316, + 0.009108597412705421, + 0.005085413809865713, + 0.0044625564478337765, + 0.007567764725536108, + 0.01108128484338522, + 0.029356835409998894, + 0.02226557768881321, + 0.03886306658387184, + 0.052912637591362, + 0.015544725582003593, + 0.026723042130470276, + 0.03188803046941757, + 0.04932424798607826, + 0.01422998495399952, + 0.012996384873986244, + 0.015449140220880508, + 0.020279476419091225, + 0.009854067116975784, + 0.008587750606238842, + 0.017252793535590172, + 0.018933065235614777, + 0.008091066963970661, + 0.0075720832683146, + 0.0047294325195252895, + 0.006401391699910164, + 0.006158504635095596, + 0.005647093988955021, + 0.01489106472581625, + 0.018723944202065468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 8 + ], + "selected_reward": -0.020574131980538368 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_tensor([9143, 9143])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_tensor([9143, 9143])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..cfdbfddbadddc446b8fe91a1d8b0c72ee731c8a6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_tensor([9143, 9143])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An immersive first-person camera places us in the ornate cockpit of a vast mechanical cog train, gazing out through a vast, curved crystal porthole framed by an intricately carved brass frame. The control panel itself is a work of art, its magical crystal panels shimmering with runes interwoven with countless precisely interlocking brass instrument panels. A core power crystal at its center pulses with a steady and powerful blue light. The train struggles upward along an impossibly thick, towering vine, its every advance accompanied by a deep, rhythmic mechanical throb, as if the entire world trembles slightly under the force. At the edge of our vision, massive brass drive gears can be seen, locked with thunderous force into the crystalline grooves of the vine's surface, emitting golden sparks of mechanics and magic. Outside the window, an otherworldly scene unfolded: countless verdant floating islands dotted a sky shimmering in purple and gold, some glimpsed with elegant spires. Magnificent waterfalls of energy, like liquid light, cascaded down from the edges of lofty islands or through breaks in the clouds, silently washing over the vast expanse. The entire scene, adorned with the sumptuous style of Final Fantasy, seamlessly blended the mechanical aesthetics of steampunk with the magical wonders of high fantasy, creating an epic atmosphere of vastness, grandeur, and endless desire for exploration.", + "seed": 10365, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10365_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009677250136155635, + "vlm_logprob": -0.009677250136155635, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.009677250136155635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016161154955625534 + ], + "opposite_logprob_tokens": [ + -4.39116096496582 + ], + "logprob_reward": -0.016161154955625534, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016161154955625534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06293937563896179, + 0.18087637424468994, + 0.030302265658974648, + 0.08061216026544571, + 0.042630635201931, + 0.055988334119319916, + 0.058636996895074844, + 0.06435657292604446, + 0.38250064849853516, + 0.04115663841366768 + ], + "credit_map": [ + 0.008191793225705624, + 0.007050832733511925, + 0.009523583576083183, + 0.011418684385716915, + 0.005907263141125441, + 0.007483542896807194, + 0.006209636572748423, + 0.0071540349163115025, + 0.02245068922638893, + 0.02433774434030056, + 0.025082038715481758, + 0.029942678287625313, + 0.015819501131772995, + 0.0165574848651886, + 0.030370239168405533, + 0.016315996646881104, + 0.0041158366948366165, + 0.0036956672556698322, + 0.003255991032347083, + 0.0040624914690852165, + 0.002980278106406331, + 0.004831182304769754, + 0.0036583973560482264, + 0.00370241841301322, + 0.006312646437436342, + 0.006715541705489159, + 0.01575297676026821, + 0.012409188784658909, + 0.009193332865834236, + 0.007142412010580301, + 0.009780061431229115, + 0.01330600120127201, + 0.005476084072142839, + 0.008880721405148506, + 0.009874993935227394, + 0.0034615376498550177, + 0.004520413000136614, + 0.0027437403332442045, + 0.0027261055074632168, + 0.004947036970406771, + 0.008679871447384357, + 0.006653317715972662, + 0.0064514195546507835, + 0.007115607615560293, + 0.004246340598911047, + 0.005449446849524975, + 0.006257113069295883, + 0.01113521121442318, + 0.0050499942153692245, + 0.004192253574728966, + 0.010821277275681496, + 0.006551377475261688, + 0.007522230502218008, + 0.007007929030805826, + 0.006533431354910135, + 0.010958501137793064, + 0.01909848488867283, + 0.009997555986046791, + 0.00661426642909646, + 0.005942204035818577, + 0.006921620573848486, + 0.0044439928606152534, + 0.0042749494314193726, + 0.007063488941639662, + 0.1125917136669159, + 0.039124976843595505, + 0.02312380075454712, + 0.01591586321592331, + 0.13661986589431763, + 0.022889871150255203, + 0.012379657477140427, + 0.019854847341775894, + 0.0053666625171899796, + 0.006783125456422567, + 0.004744970705360174, + 0.005444022826850414, + 0.004703521728515625, + 0.004615574609488249, + 0.005012687295675278, + 0.004486067220568657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.016161154955625534 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005036516231484711 + ], + "opposite_logprob_tokens": [ + -7.6255035400390625 + ], + "logprob_reward": -0.0005036516231484711, + "yes_no_margin_reward": 7.625, + "token_credit_reward": -0.0005036516231484711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15712855756282806, + 0.13584129512310028, + 0.08940576761960983, + 0.10425689816474915, + 0.023756032809615135, + 0.07152719795703888, + 0.07719913125038147, + 0.0550880990922451, + 0.23849371075630188, + 0.04730328544974327 + ], + "credit_map": [ + 0.037845976650714874, + 0.01702592894434929, + 0.03424685820937157, + 0.015046670101583004, + 0.010730895213782787, + 0.011346585117280483, + 0.014640637673437595, + 0.01624501869082451, + 0.012732352130115032, + 0.017211856320500374, + 0.021171098574995995, + 0.015644529834389687, + 0.024148477241396904, + 0.017321262508630753, + 0.01326507143676281, + 0.01434665359556675, + 0.016629716381430626, + 0.009669795632362366, + 0.008858265355229378, + 0.012520640157163143, + 0.008665796369314194, + 0.009094025939702988, + 0.012010151520371437, + 0.011957384645938873, + 0.008140486665070057, + 0.008218588307499886, + 0.016623027622699738, + 0.015680521726608276, + 0.00945936981588602, + 0.010991383343935013, + 0.01509685069322586, + 0.020046675577759743, + 0.0032307379879057407, + 0.00393097847700119, + 0.005918137263506651, + 0.0020224114414304495, + 0.0020924240816384554, + 0.0018370923353359103, + 0.0014363847440108657, + 0.003287868807092309, + 0.008758651092648506, + 0.00633463729172945, + 0.006845472380518913, + 0.011222952045500278, + 0.005952239502221346, + 0.0072450824081897736, + 0.01174086332321167, + 0.013427307829260826, + 0.010484190657734871, + 0.007196364924311638, + 0.00924221333116293, + 0.015203229151666164, + 0.005933559034019709, + 0.006035188212990761, + 0.014757907949388027, + 0.008346482180058956, + 0.016241993755102158, + 0.009093253873288631, + 0.005125769414007664, + 0.005036212503910065, + 0.005005933810025454, + 0.004595833830535412, + 0.0042763217352330685, + 0.0057127820327878, + 0.05646165832877159, + 0.027586063370108604, + 0.01792195625603199, + 0.011515848338603973, + 0.07668356597423553, + 0.02375464141368866, + 0.008682445622980595, + 0.015887541696429253, + 0.005267196800559759, + 0.0067153205163776875, + 0.005911368876695633, + 0.00751872081309557, + 0.007949143648147583, + 0.004100232385098934, + 0.005074452608823776, + 0.004766852129250765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.0005036516231484711 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032011240255087614 + ], + "opposite_logprob_tokens": [ + -5.753201007843018 + ], + "logprob_reward": -0.0032011240255087614, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032011240255087614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33440375328063965, + 0.16838999092578888, + 0.0656508058309555, + 0.04227076843380928, + 0.08574605733156204, + 0.07421687245368958, + 0.05835304409265518, + 0.042628198862075806, + 0.060666538774967194, + 0.0676739513874054 + ], + "credit_map": [ + 0.09182316809892654, + 0.03372155874967575, + 0.0927290990948677, + 0.01946597546339035, + 0.01795020140707493, + 0.013534356839954853, + 0.0160884577780962, + 0.049090929329395294, + 0.021718474105000496, + 0.023629793897271156, + 0.020368261262774467, + 0.02285386435687542, + 0.029301641508936882, + 0.01311584934592247, + 0.018900258466601372, + 0.01850185915827751, + 0.007816768251359463, + 0.0068947081454098225, + 0.0064536770805716515, + 0.009204735048115253, + 0.007747585419565439, + 0.008687211200594902, + 0.01219099573791027, + 0.006655124481767416, + 0.0041927495039999485, + 0.004914815071970224, + 0.006317732855677605, + 0.005422831047326326, + 0.0051828715950250626, + 0.005562092177569866, + 0.004959787242114544, + 0.00571789313107729, + 0.010984562337398529, + 0.01692808046936989, + 0.023577159270644188, + 0.006140024401247501, + 0.006024092901498079, + 0.005705112591385841, + 0.00546962209045887, + 0.010917398147284985, + 0.011668188497424126, + 0.007895086891949177, + 0.006848212331533432, + 0.011412950232625008, + 0.006511056330054998, + 0.006379072088748217, + 0.011550229974091053, + 0.011952077969908714, + 0.005885280668735504, + 0.004320490639656782, + 0.005596274044364691, + 0.010463106445968151, + 0.0067610363475978374, + 0.004754894413053989, + 0.008039827458560467, + 0.012532133609056473, + 0.005816908087581396, + 0.003054832573980093, + 0.005814211443066597, + 0.0063238623552024364, + 0.005282620433717966, + 0.004293011035770178, + 0.006499508395791054, + 0.005543242208659649, + 0.009447497315704823, + 0.005429175216704607, + 0.006969162728637457, + 0.0063042426481842995, + 0.012617796659469604, + 0.00747355492785573, + 0.005710424389690161, + 0.006714686751365662, + 0.010272963903844357, + 0.00787475798279047, + 0.009598884731531143, + 0.00680570537224412, + 0.010723010636866093, + 0.0062685394659638405, + 0.00553278811275959, + 0.010597296059131622 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0032011240255087614 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327316626906395 + ], + "opposite_logprob_tokens": [ + -3.773273229598999 + ], + "logprob_reward": -0.02327316626906395, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327316626906395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09644192457199097, + 0.3803524672985077, + 0.04692326486110687, + 0.026462042704224586, + 0.03582284599542618, + 0.04922356456518173, + 0.03430143743753433, + 0.04754527285695076, + 0.09159258753061295, + 0.1913345605134964 + ], + "credit_map": [ + 0.013875514268875122, + 0.010141555219888687, + 0.012457110919058323, + 0.018131833523511887, + 0.00712142838165164, + 0.010916966944932938, + 0.010915776714682579, + 0.01288173720240593, + 0.03496409207582474, + 0.0495629608631134, + 0.057909343391656876, + 0.05297792702913284, + 0.05340362712740898, + 0.03935680538415909, + 0.0467192716896534, + 0.045458439737558365, + 0.006452425383031368, + 0.004648012109100819, + 0.0044459328055381775, + 0.008830418810248375, + 0.003794955089688301, + 0.003773072734475136, + 0.009270710870623589, + 0.0057077365927398205, + 0.002231416292488575, + 0.0030296349432319403, + 0.0037603978998959064, + 0.0037729425821453333, + 0.002585866954177618, + 0.0028147813864052296, + 0.0031536181923002005, + 0.0051133837550878525, + 0.0032098793890327215, + 0.006684382911771536, + 0.010033910162746906, + 0.003937957342714071, + 0.0021808526944369078, + 0.0015718505019322038, + 0.0033879471011459827, + 0.004816063679754734, + 0.009254115633666515, + 0.004734985530376434, + 0.0043076882138848305, + 0.007981883361935616, + 0.0039935726672410965, + 0.002962657483294606, + 0.008374539203941822, + 0.007614120841026306, + 0.003797465469688177, + 0.003962403163313866, + 0.004629846662282944, + 0.003971658181399107, + 0.005199916660785675, + 0.0032357131130993366, + 0.003317513270303607, + 0.006186920218169689, + 0.011105832643806934, + 0.006555001717060804, + 0.004380619619041681, + 0.0061269174329936504, + 0.004135050345212221, + 0.003493927652016282, + 0.00647452287375927, + 0.005273401737213135, + 0.019927775487303734, + 0.012751410715281963, + 0.007747838739305735, + 0.004269576631486416, + 0.027124688029289246, + 0.00833296962082386, + 0.004183086100965738, + 0.0072552417404949665, + 0.04091692715883255, + 0.019355127587914467, + 0.03265810385346413, + 0.0189043078571558, + 0.030251983553171158, + 0.015465239994227886, + 0.01545557752251625, + 0.018327297642827034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.02327316626906395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247153807431459 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247153807431459, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247153807431459, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14286543428897858, + 0.27805113792419434, + 0.1860058754682541, + 0.04999828338623047, + 0.0558386892080307, + 0.04283669963479042, + 0.05822654440999031, + 0.04778419807553291, + 0.06853103637695312, + 0.06986212730407715 + ], + "credit_map": [ + 0.017616169527173042, + 0.010161594487726688, + 0.03505992889404297, + 0.015555605292320251, + 0.010228217579424381, + 0.01479734480381012, + 0.013099992647767067, + 0.02634657919406891, + 0.0315002016723156, + 0.03649333119392395, + 0.02867943048477173, + 0.04657730087637901, + 0.03501689434051514, + 0.028861548751592636, + 0.04428365081548691, + 0.026638761162757874, + 0.019276101142168045, + 0.026242053136229515, + 0.01793866977095604, + 0.026737429201602936, + 0.019851282238960266, + 0.024252912029623985, + 0.028525760397315025, + 0.02318166196346283, + 0.004414609633386135, + 0.006165872327983379, + 0.005752501543611288, + 0.006862223148345947, + 0.005315601825714111, + 0.006656492594629526, + 0.0080804955214262, + 0.0067504895851016045, + 0.006435859948396683, + 0.01704344153404236, + 0.010941273532807827, + 0.0034186383709311485, + 0.005751011893153191, + 0.0038772013504058123, + 0.0037378452252596617, + 0.004633417818695307, + 0.006979456637054682, + 0.0033068815246224403, + 0.007290780544281006, + 0.006812392268329859, + 0.0035025246907025576, + 0.0029942868277430534, + 0.0053488160483539104, + 0.006601560395210981, + 0.0067815231159329414, + 0.008177236653864384, + 0.00837352778762579, + 0.0039643533527851105, + 0.011839020065963268, + 0.008074279874563217, + 0.003098261309787631, + 0.00791834108531475, + 0.009035601280629635, + 0.004553735256195068, + 0.004396966192871332, + 0.0038298035506159067, + 0.00819840282201767, + 0.008766035549342632, + 0.003559359349310398, + 0.005444292910397053, + 0.01701362431049347, + 0.005809175316244364, + 0.0038140288088470697, + 0.0045531075447797775, + 0.021072547882795334, + 0.006575496401637793, + 0.004141072276979685, + 0.005551982671022415, + 0.009197372943162918, + 0.008603201247751713, + 0.012354562059044838, + 0.004646655172109604, + 0.011001823469996452, + 0.006787037942558527, + 0.004084810148924589, + 0.013186662457883358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005247153807431459 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person camera places us in the ornate cockpit of a vast mechanical cog train, gazing out through a vast, curved crystal porthole framed by an intricately carved brass frame. The control panel itself is a work of art, its magical crystal panels shimmering with runes interwoven with countless precisely interlocking brass instrument panels. A core power crystal at its center pulses with a steady and powerful blue light. The train struggles upward along an impossibly thick, towering vine, its every advance accompanied by a deep, rhythmic mechanical throb, as if the entire world trembles slightly under the force. At the edge of our vision, massive brass drive gears can be seen, locked with thunderous force into the crystalline grooves of the vine's surface, emitting golden sparks of mechanics and magic. Outside the window, an otherworldly scene unfolded: countless verdant floating islands dotted a sky shimmering in purple and gold, some glimpsed with elegant spires. Magnificent waterfalls of energy, like liquid light, cascaded down from the edges of lofty islands or through breaks in the clouds, silently washing over the vast expanse. The entire scene, adorned with the sumptuous style of Final Fantasy, seamlessly blended the mechanical aesthetics of steampunk with the magical wonders of high fantasy, creating an epic atmosphere of vastness, grandeur, and endless desire for exploration.", + "seed": 10366, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10366_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0111420804169029, + "vlm_logprob": -0.0111420804169029, + "vlm_yes_no_margin": 5.225, + "vlm_token_credit": -0.0111420804169029, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03960539773106575 + ], + "opposite_logprob_tokens": [ + -3.2896053791046143 + ], + "logprob_reward": -0.03960539773106575, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03960539773106575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26574987173080444, + 0.0460793562233448, + 0.04086063429713249, + 0.11111216247081757, + 0.03496050462126732, + 0.07220236212015152, + 0.08118905127048492, + 0.09499243646860123, + 0.1356152594089508, + 0.1172383576631546 + ], + "credit_map": [ + 0.032793089747428894, + 0.05562610551714897, + 0.032507285475730896, + 0.03169214352965355, + 0.03739220276474953, + 0.027066759765148163, + 0.02456316538155079, + 0.024109134450554848, + 0.008233304135501385, + 0.004846032243221998, + 0.0023292035330086946, + 0.004260431043803692, + 0.01160653680562973, + 0.007621716242283583, + 0.002863968489691615, + 0.004318162798881531, + 0.0060663362964987755, + 0.008149846456944942, + 0.003437008475884795, + 0.004120452329516411, + 0.0047688474878668785, + 0.0066376943141222, + 0.0033303971868008375, + 0.004350053146481514, + 0.009802012704312801, + 0.014940758235752583, + 0.010616679675877094, + 0.019597969949245453, + 0.014706837013363838, + 0.019397063180804253, + 0.010469081811606884, + 0.011581758968532085, + 0.002594795310869813, + 0.005764160770922899, + 0.003164720954373479, + 0.006203516852110624, + 0.005134854465723038, + 0.003369736485183239, + 0.004833875223994255, + 0.003894844325259328, + 0.005679604597389698, + 0.0104767931625247, + 0.006107250228524208, + 0.008779504336416721, + 0.01958688534796238, + 0.008521033450961113, + 0.007532914634793997, + 0.005518373567610979, + 0.011676246300339699, + 0.011115236207842827, + 0.006795985624194145, + 0.009265616536140442, + 0.014398970641195774, + 0.010045871138572693, + 0.006688292138278484, + 0.011202836409211159, + 0.015284991823136806, + 0.008372136391699314, + 0.01014228630810976, + 0.011187487281858921, + 0.013987774960696697, + 0.009409401565790176, + 0.012074440717697144, + 0.01453392393887043, + 0.016975030303001404, + 0.01191090326756239, + 0.026408616453409195, + 0.00955956894904375, + 0.03417782485485077, + 0.010317115113139153, + 0.008866814896464348, + 0.01739937998354435, + 0.016880463808774948, + 0.025278568267822266, + 0.020634621381759644, + 0.012472118251025677, + 0.01122511737048626, + 0.008864630945026875, + 0.00852675549685955, + 0.013356081210076809 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03960539773106575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017184742027893662 + ], + "opposite_logprob_tokens": [ + -6.376718521118164 + ], + "logprob_reward": -0.0017184742027893662, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017184742027893662, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11566411703824997, + 0.074531689286232, + 0.05595950037240982, + 0.036956798285245895, + 0.028346512466669083, + 0.14073577523231506, + 0.061687204986810684, + 0.07067565619945526, + 0.26553773880004883, + 0.1499049812555313 + ], + "credit_map": [ + 0.0161726251244545, + 0.01782827265560627, + 0.013624233193695545, + 0.014564968645572662, + 0.015156855806708336, + 0.015023305080831051, + 0.00937566813081503, + 0.013918201439082623, + 0.0065067666582763195, + 0.007510234136134386, + 0.006978846155107021, + 0.01870311051607132, + 0.013263844884932041, + 0.00815522950142622, + 0.005548578687012196, + 0.007865077815949917, + 0.008590704761445522, + 0.006233901251107454, + 0.0053974599577486515, + 0.005177441984415054, + 0.00670140003785491, + 0.0061770095489919186, + 0.006205725483596325, + 0.011475862003862858, + 0.003532187780365348, + 0.005325333680957556, + 0.008460989221930504, + 0.004474770277738571, + 0.0035590159241110086, + 0.003646639408543706, + 0.003566054394468665, + 0.004391810856759548, + 0.0033626931253820658, + 0.004983804188668728, + 0.0022875305730849504, + 0.0036409208551049232, + 0.004325933754444122, + 0.003470909781754017, + 0.0033689155243337154, + 0.00290580652654171, + 0.009723016060888767, + 0.01915847137570381, + 0.013070913031697273, + 0.02025470696389675, + 0.030861593782901764, + 0.020716968923807144, + 0.01877889595925808, + 0.008171217516064644, + 0.007060976233333349, + 0.00516644399613142, + 0.00559520348906517, + 0.009043355472385883, + 0.009847227483987808, + 0.009228536859154701, + 0.007940969429910183, + 0.007804489694535732, + 0.008699020370841026, + 0.006547354161739349, + 0.0072906347922980785, + 0.011322523467242718, + 0.011617569252848625, + 0.007216870319098234, + 0.008443575352430344, + 0.009538107551634312, + 0.026629390195012093, + 0.024711204692721367, + 0.029568694531917572, + 0.01610289327800274, + 0.07359950989484787, + 0.01860777847468853, + 0.014933889731764793, + 0.06138436496257782, + 0.019170677289366722, + 0.028440019115805626, + 0.022376205772161484, + 0.0205638837069273, + 0.01315375231206417, + 0.014005986973643303, + 0.0133903743699193, + 0.01880408450961113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.0017184742027893662 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716570113785565 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716570113785565, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716570113785565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29224562644958496, + 0.03673767298460007, + 0.025979239493608475, + 0.14069664478302002, + 0.024103518575429916, + 0.10621172189712524, + 0.03528403490781784, + 0.043817758560180664, + 0.21106846630573273, + 0.08385530114173889 + ], + "credit_map": [ + 0.04159921035170555, + 0.05408017337322235, + 0.028733735904097557, + 0.03108089044690132, + 0.04924503341317177, + 0.040330126881599426, + 0.024057334288954735, + 0.0231191273778677, + 0.003303099889308214, + 0.004241407848894596, + 0.003665556199848652, + 0.009365356527268887, + 0.005820699501782656, + 0.0034247993025928736, + 0.0026678815484046936, + 0.004248873330652714, + 0.0037907524965703487, + 0.004044736735522747, + 0.0018957445863634348, + 0.0019863233901560307, + 0.004086398519575596, + 0.0027884223964065313, + 0.002703174250200391, + 0.004683689214289188, + 0.013342633843421936, + 0.018765561282634735, + 0.012202502228319645, + 0.02269521728157997, + 0.020015789195895195, + 0.02523175999522209, + 0.012491581961512566, + 0.015951601788401604, + 0.0022324235178530216, + 0.00478423573076725, + 0.0026673232205212116, + 0.0037858972791582346, + 0.0039809104055166245, + 0.0023474874906241894, + 0.0019024151843041182, + 0.0024028276093304157, + 0.007351241540163755, + 0.015591870993375778, + 0.010086697526276112, + 0.015099362470209599, + 0.023467842489480972, + 0.014458676800131798, + 0.013888397254049778, + 0.006267637480050325, + 0.004670179449021816, + 0.0038166099693626165, + 0.002892063930630684, + 0.003907140344381332, + 0.005877792369574308, + 0.003762322012335062, + 0.0038023453671485186, + 0.0065555814653635025, + 0.006863629911094904, + 0.004892113618552685, + 0.005055862013250589, + 0.003935827873647213, + 0.004763246513903141, + 0.00463600130751729, + 0.007579461205750704, + 0.006091616582125425, + 0.023485751822590828, + 0.017847737297415733, + 0.03312470763921738, + 0.013122052885591984, + 0.06034627929329872, + 0.014770028181374073, + 0.011307609267532825, + 0.037064291536808014, + 0.008837961591780186, + 0.012459495104849339, + 0.00969614926725626, + 0.009607261046767235, + 0.008009588345885277, + 0.0143376924097538, + 0.011566796340048313, + 0.009340357035398483 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.001716570113785565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00594134908169508 + ], + "opposite_logprob_tokens": [ + -5.130941390991211 + ], + "logprob_reward": -0.00594134908169508, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.00594134908169508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12090418487787247, + 0.046645212918519974, + 0.03217620402574539, + 0.03908579424023628, + 0.03218705952167511, + 0.15451142191886902, + 0.03987198695540428, + 0.08216647803783417, + 0.2537554204463959, + 0.19869619607925415 + ], + "credit_map": [ + 0.02030988037586212, + 0.01697813905775547, + 0.012578408233821392, + 0.01614634320139885, + 0.018721425905823708, + 0.01461155153810978, + 0.01172518078237772, + 0.009833262301981449, + 0.007735711056739092, + 0.00634791748598218, + 0.002235732041299343, + 0.004427556414157152, + 0.011533808894455433, + 0.005687691271305084, + 0.00441184313967824, + 0.004264952149242163, + 0.0051482729613780975, + 0.006070945877581835, + 0.0030487419571727514, + 0.0032265842892229557, + 0.005428748205304146, + 0.0037867967039346695, + 0.0032026360277086496, + 0.002263478236272931, + 0.004794493783265352, + 0.007968083955347538, + 0.004120676312595606, + 0.00592027185484767, + 0.004297839477658272, + 0.0047953431494534016, + 0.003734762314707041, + 0.0034543233923614025, + 0.002849783282727003, + 0.0045527988113462925, + 0.003432343015447259, + 0.004026287700980902, + 0.005904033314436674, + 0.0036689003463834524, + 0.003544093342497945, + 0.004208818078041077, + 0.010842260904610157, + 0.02053041197359562, + 0.014658826403319836, + 0.023521998897194862, + 0.03128495812416077, + 0.022769514471292496, + 0.021565942093729973, + 0.009337500669062138, + 0.0064421482384204865, + 0.003995103295892477, + 0.004094364121556282, + 0.0057554892264306545, + 0.005636631976813078, + 0.004636479541659355, + 0.005612657871097326, + 0.0036991138476878405, + 0.009822482243180275, + 0.007375006098300219, + 0.009471578523516655, + 0.010021691210567951, + 0.008815620094537735, + 0.007845967076718807, + 0.018440160900354385, + 0.010373968631029129, + 0.02421732060611248, + 0.02773403562605381, + 0.029798150062561035, + 0.015695901587605476, + 0.09694638848304749, + 0.017052657902240753, + 0.01033663097769022, + 0.031974323093891144, + 0.022320926189422607, + 0.0362253300845623, + 0.027948960661888123, + 0.028400149196386337, + 0.018543563783168793, + 0.02112928405404091, + 0.021313536912202835, + 0.02281445451080799 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.00594134908169508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006728610955178738 + ], + "opposite_logprob_tokens": [ + -5.006728649139404 + ], + "logprob_reward": -0.006728610955178738, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006728610955178738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4465363621711731, + 0.07078351825475693, + 0.04562821239233017, + 0.06743312627077103, + 0.016192683950066566, + 0.013403809629380703, + 0.0750148594379425, + 0.054617393761873245, + 0.15038199722766876, + 0.06000804528594017 + ], + "credit_map": [ + 0.0700230747461319, + 0.06908134371042252, + 0.0417407862842083, + 0.04595543071627617, + 0.07831118255853653, + 0.07016577571630478, + 0.03938954696059227, + 0.03186921402812004, + 0.010432551614940166, + 0.008160828612744808, + 0.0063897427171468735, + 0.014442287385463715, + 0.00969193410128355, + 0.006066448520869017, + 0.0061882538720965385, + 0.009411470964550972, + 0.008329210802912712, + 0.008844934403896332, + 0.004346644505858421, + 0.003605772042647004, + 0.007247649598866701, + 0.007460081018507481, + 0.00243863626383245, + 0.003355280961841345, + 0.006213413551449776, + 0.009497360326349735, + 0.005726542789489031, + 0.014161868952214718, + 0.009095260873436928, + 0.011511188931763172, + 0.005393557250499725, + 0.005833931267261505, + 0.001328705227933824, + 0.001897713402286172, + 0.0018030903302133083, + 0.003123627509921789, + 0.001762889325618744, + 0.0019763305317610502, + 0.002280919114127755, + 0.0020194081589579582, + 0.0012333451304584742, + 0.0022005117498338223, + 0.0011256715515628457, + 0.0016064582159742713, + 0.0025576679036021233, + 0.0020430132281035185, + 0.0011554149677976966, + 0.0014817273477092385, + 0.011680249124765396, + 0.007719319313764572, + 0.007073356769979, + 0.007207897026091814, + 0.013141633942723274, + 0.011713327839970589, + 0.0065071540884673595, + 0.009971927851438522, + 0.00832065474241972, + 0.004880262538790703, + 0.006609808653593063, + 0.005464163608849049, + 0.009671994484961033, + 0.005967036355286837, + 0.005686491262167692, + 0.008016983047127724, + 0.015363248065114021, + 0.02043723687529564, + 0.018694885075092316, + 0.009536636993288994, + 0.05762956663966179, + 0.011049197986721992, + 0.005965298041701317, + 0.011705917306244373, + 0.007608296815305948, + 0.007433051243424416, + 0.00908698607236147, + 0.006570551078766584, + 0.006492530927062035, + 0.008394149132072926, + 0.0068738157860934734, + 0.007548665162175894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.006728610955178738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person camera places us in the ornate cockpit of a vast mechanical cog train, gazing out through a vast, curved crystal porthole framed by an intricately carved brass frame. The control panel itself is a work of art, its magical crystal panels shimmering with runes interwoven with countless precisely interlocking brass instrument panels. A core power crystal at its center pulses with a steady and powerful blue light. The train struggles upward along an impossibly thick, towering vine, its every advance accompanied by a deep, rhythmic mechanical throb, as if the entire world trembles slightly under the force. At the edge of our vision, massive brass drive gears can be seen, locked with thunderous force into the crystalline grooves of the vine's surface, emitting golden sparks of mechanics and magic. Outside the window, an otherworldly scene unfolded: countless verdant floating islands dotted a sky shimmering in purple and gold, some glimpsed with elegant spires. Magnificent waterfalls of energy, like liquid light, cascaded down from the edges of lofty islands or through breaks in the clouds, silently washing over the vast expanse. The entire scene, adorned with the sumptuous style of Final Fantasy, seamlessly blended the mechanical aesthetics of steampunk with the magical wonders of high fantasy, creating an epic atmosphere of vastness, grandeur, and endless desire for exploration.", + "seed": 10367, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10367_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.006421972950920463, + "vlm_logprob": -0.006421972950920463, + "vlm_yes_no_margin": 5.549999952316284, + "vlm_token_credit": -0.006421972950920463, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007285929284989834 + ], + "opposite_logprob_tokens": [ + -5.132286071777344 + ], + "logprob_reward": -0.007285929284989834, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007285929284989834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09532584249973297, + 0.060521453619003296, + 0.1629360169172287, + 0.12158302217721939, + 0.12390462309122086, + 0.16473142802715302, + 0.03863859549164772, + 0.03447341173887253, + 0.16291844844818115, + 0.0349670946598053 + ], + "credit_map": [ + 0.00770546542480588, + 0.015052620321512222, + 0.010542097501456738, + 0.01093872170895338, + 0.011308255605399609, + 0.023970238864421844, + 0.007838805206120014, + 0.007969632744789124, + 0.008810549974441528, + 0.0077282739803195, + 0.007053477689623833, + 0.010518614202737808, + 0.005049781873822212, + 0.006865113042294979, + 0.008832353167235851, + 0.005663290154188871, + 0.014749700203537941, + 0.009599894285202026, + 0.008977705612778664, + 0.011136269196867943, + 0.04144364595413208, + 0.023674022406339645, + 0.014342710375785828, + 0.03901206701993942, + 0.014952296391129494, + 0.016769764944911003, + 0.020057201385498047, + 0.012421440333127975, + 0.016247909516096115, + 0.015127534978091717, + 0.015349237248301506, + 0.010657631792128086, + 0.013613433577120304, + 0.020680246874690056, + 0.01034566294401884, + 0.011409432627260685, + 0.023701196536421776, + 0.019903792068362236, + 0.009373498149216175, + 0.01487736962735653, + 0.014429713599383831, + 0.046480149030685425, + 0.013435068540275097, + 0.010684320703148842, + 0.01929127424955368, + 0.0327855609357357, + 0.014918741770088673, + 0.012706597335636616, + 0.005224361550062895, + 0.004512016661465168, + 0.004576533567160368, + 0.0036334344185888767, + 0.004201312083750963, + 0.005796792916953564, + 0.005492624826729298, + 0.005201521795243025, + 0.00503764720633626, + 0.003664493327960372, + 0.0029844308737665415, + 0.00308562652207911, + 0.00825855415314436, + 0.004568928387016058, + 0.0035870515275746584, + 0.0032866785768419504, + 0.01693853735923767, + 0.016113905236124992, + 0.01800304278731346, + 0.021526001393795013, + 0.027471022680401802, + 0.02890678495168686, + 0.019808582961559296, + 0.014150564558804035, + 0.004933249671012163, + 0.004859300795942545, + 0.004436118993908167, + 0.002407172229140997, + 0.004542801063507795, + 0.0065281870774924755, + 0.0027380292303860188, + 0.004522237461060286 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 8 + ], + "selected_reward": -0.007285929284989834 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015209070406854153 + ], + "opposite_logprob_tokens": [ + -6.501521110534668 + ], + "logprob_reward": -0.0015209070406854153, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015209070406854153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09336380660533905, + 0.03518622741103172, + 0.10426066070795059, + 0.08022548258304596, + 0.052141137421131134, + 0.41788536310195923, + 0.06849252432584763, + 0.045615267008543015, + 0.07508763670921326, + 0.02774188108742237 + ], + "credit_map": [ + 0.007908016443252563, + 0.011127342469990253, + 0.0132988803088665, + 0.010082625783979893, + 0.010981133207678795, + 0.023307396098971367, + 0.007421506103128195, + 0.009236913174390793, + 0.004477728623896837, + 0.0038988927844911814, + 0.0038555895444005728, + 0.005505765322595835, + 0.004757770337164402, + 0.004798842128366232, + 0.004700217861682177, + 0.0031914180144667625, + 0.009242767468094826, + 0.007018920034170151, + 0.006837250664830208, + 0.008798996917903423, + 0.023898225277662277, + 0.012137957848608494, + 0.008859232068061829, + 0.027467304840683937, + 0.008505035191774368, + 0.009006114676594734, + 0.009974537417292595, + 0.014568846672773361, + 0.010467295534908772, + 0.010881676338613033, + 0.009867175482213497, + 0.006954802665859461, + 0.003874420188367367, + 0.00649804575368762, + 0.00470332195982337, + 0.010224083438515663, + 0.00835059117525816, + 0.0058466787450015545, + 0.007191420532763004, + 0.00545257655903697, + 0.03251347690820694, + 0.15667980909347534, + 0.02768048830330372, + 0.02110065333545208, + 0.021409131586551666, + 0.09431105107069016, + 0.039182063192129135, + 0.025008689612150192, + 0.008389700204133987, + 0.008169667795300484, + 0.0050941635854542255, + 0.01125642191618681, + 0.0067734308540821075, + 0.00832380261272192, + 0.015052633360028267, + 0.005432703532278538, + 0.006610256154090166, + 0.004566200077533722, + 0.00712735578417778, + 0.004659231752157211, + 0.006194783374667168, + 0.005120933521538973, + 0.005709169898182154, + 0.0056273373775184155, + 0.010226625949144363, + 0.0056526195257902145, + 0.01013341173529625, + 0.006766322534531355, + 0.01045013777911663, + 0.01650899089872837, + 0.007781993597745895, + 0.0075675412081182, + 0.003363037249073386, + 0.0038912079762667418, + 0.003168628551065922, + 0.0022620565723627806, + 0.004217520821839571, + 0.004458772484213114, + 0.0037173964083194733, + 0.002663260092958808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0015209070406854153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015175743028521538 + ], + "opposite_logprob_tokens": [ + -6.5015177726745605 + ], + "logprob_reward": -0.0015175743028521538, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015175743028521538, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12637057900428772, + 0.06285640597343445, + 0.0854981318116188, + 0.05298808962106705, + 0.0377315953373909, + 0.39282673597335815, + 0.08158627897500992, + 0.04157182574272156, + 0.07858763635158539, + 0.03998270630836487 + ], + "credit_map": [ + 0.011680502444505692, + 0.014437694102525711, + 0.01785869710147381, + 0.017717570066452026, + 0.013457341119647026, + 0.02544981986284256, + 0.012027978897094727, + 0.01374098751693964, + 0.009739051572978497, + 0.006177711300551891, + 0.007648606784641743, + 0.008282405324280262, + 0.007133957929909229, + 0.007483215071260929, + 0.009953780099749565, + 0.0064376783557236195, + 0.006886586081236601, + 0.0030901599675416946, + 0.004530838690698147, + 0.0056381491012871265, + 0.023339012637734413, + 0.012881482020020485, + 0.007581591606140137, + 0.021550321951508522, + 0.006641092710196972, + 0.006688190624117851, + 0.007402725517749786, + 0.00780255114659667, + 0.005701684392988682, + 0.007874403148889542, + 0.0069124470464885235, + 0.003964999690651894, + 0.0034544565714895725, + 0.004505994264036417, + 0.0037491952534765005, + 0.006321638356894255, + 0.006366564426571131, + 0.004148364532738924, + 0.005292249843478203, + 0.003893135115504265, + 0.0300914254039526, + 0.1500546634197235, + 0.026242170482873917, + 0.01950816437602043, + 0.018714500591158867, + 0.08892391622066498, + 0.0356714129447937, + 0.023620499297976494, + 0.011092035099864006, + 0.011208813637495041, + 0.009144920855760574, + 0.009568714536726475, + 0.010801338590681553, + 0.010624202899634838, + 0.010635328479111195, + 0.008510926738381386, + 0.006370254792273045, + 0.004337085410952568, + 0.004506594501435757, + 0.003317844122648239, + 0.009334576316177845, + 0.005667845718562603, + 0.004153268411755562, + 0.003884360659867525, + 0.009263718500733376, + 0.007045872509479523, + 0.007714157458394766, + 0.008834573440253735, + 0.015541546046733856, + 0.015204772353172302, + 0.008349945768713951, + 0.006633056793361902, + 0.005356962326914072, + 0.003671901300549507, + 0.0038474637549370527, + 0.0034163210075348616, + 0.004076503217220306, + 0.008809898979961872, + 0.006546212825924158, + 0.004257448483258486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.0015175743028521538 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613016800954938 + ], + "opposite_logprob_tokens": [ + -5.628612995147705 + ], + "logprob_reward": -0.003613016800954938, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613016800954938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.060599617660045624, + 0.05406341701745987, + 0.036854542791843414, + 0.021507252007722855, + 0.022226309403777122, + 0.5664528012275696, + 0.03413449600338936, + 0.01916027069091797, + 0.06355126947164536, + 0.12144995480775833 + ], + "credit_map": [ + 0.007415541913360357, + 0.008209023624658585, + 0.005234674084931612, + 0.007526093628257513, + 0.006573847960680723, + 0.011513127014040947, + 0.009250146336853504, + 0.004877166356891394, + 0.006216131150722504, + 0.006213683169335127, + 0.004985211417078972, + 0.00735547486692667, + 0.007623524870723486, + 0.008274945430457592, + 0.007865353487432003, + 0.005529098212718964, + 0.0044901613146066666, + 0.0025726810563355684, + 0.003266483312472701, + 0.003985451068729162, + 0.005784290377050638, + 0.0032185399904847145, + 0.003440150525420904, + 0.010096789337694645, + 0.0027034333907067776, + 0.002676028059795499, + 0.0024154288694262505, + 0.0028326159808784723, + 0.0030382396653294563, + 0.0026215764228254557, + 0.0029275729320943356, + 0.0022923583164811134, + 0.0023841990623623133, + 0.0027001239359378815, + 0.0016739539569243789, + 0.0034316021483391523, + 0.003937206231057644, + 0.003391447477042675, + 0.0025948951952159405, + 0.002112884074449539, + 0.043511051684617996, + 0.21576040983200073, + 0.03734287619590759, + 0.02802402526140213, + 0.027651188895106316, + 0.12806159257888794, + 0.052136749029159546, + 0.033964939415454865, + 0.004084707237780094, + 0.0036768470890820026, + 0.0038906114641577005, + 0.003954357001930475, + 0.005479034036397934, + 0.0044683609157800674, + 0.0043430752120912075, + 0.004237504210323095, + 0.0027433924842625856, + 0.0023643283639103174, + 0.0017408878775313497, + 0.0019121229415759444, + 0.003004246624186635, + 0.0027460744604468346, + 0.003162706969305873, + 0.0014865134144201875, + 0.007739547174423933, + 0.006057584658265114, + 0.006927106063812971, + 0.005512857344001532, + 0.009922053664922714, + 0.012274342589080334, + 0.007361105177551508, + 0.00775667792186141, + 0.017106952145695686, + 0.01585712842643261, + 0.006254387553781271, + 0.01070176437497139, + 0.010928967036306858, + 0.03276621550321579, + 0.019658293575048447, + 0.008176264353096485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 8 + ], + "selected_reward": -0.003613016800954938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018172437325119972 + ], + "opposite_logprob_tokens": [ + -4.018172264099121 + ], + "logprob_reward": -0.018172437325119972, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018172437325119972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04543938860297203, + 0.12381496280431747, + 0.041226472705602646, + 0.014083790592849255, + 0.017268702387809753, + 0.628652811050415, + 0.0866309106349945, + 0.014844250865280628, + 0.014570685103535652, + 0.013468004763126373 + ], + "credit_map": [ + 0.007007131353020668, + 0.005718845874071121, + 0.00595835642889142, + 0.005134198348969221, + 0.006176990922540426, + 0.005691743455827236, + 0.0049206227995455265, + 0.004831499885767698, + 0.01730532757937908, + 0.013480392284691334, + 0.01246887631714344, + 0.01903071254491806, + 0.012720907106995583, + 0.014429464936256409, + 0.022291986271739006, + 0.012087306007742882, + 0.002966854954138398, + 0.0014903792180120945, + 0.002357305260375142, + 0.0029609340708702803, + 0.011152667924761772, + 0.006373891141265631, + 0.0038182982243597507, + 0.010106145404279232, + 0.0018122507026419044, + 0.0017022830434143543, + 0.0018944592447951436, + 0.0018278948264196515, + 0.001640492700971663, + 0.0015047710621729493, + 0.0021348625887185335, + 0.0015667770057916641, + 0.002014938509091735, + 0.002595264697447419, + 0.0014835296897217631, + 0.0022916006855666637, + 0.002604252193123102, + 0.0022291340865194798, + 0.0021848923061043024, + 0.001865092315711081, + 0.049237143248319626, + 0.23239298164844513, + 0.04115981236100197, + 0.03227618336677551, + 0.03124854527413845, + 0.14407879114151, + 0.060024432837963104, + 0.03823494911193848, + 0.01112991850823164, + 0.011611166410148144, + 0.011895743198692799, + 0.007890239357948303, + 0.012370259501039982, + 0.011544054374098778, + 0.011109456419944763, + 0.009080073796212673, + 0.002475646324455738, + 0.0020136404782533646, + 0.0014449438313022256, + 0.0014688449446111917, + 0.0027721934020519257, + 0.001431637560017407, + 0.00170558737590909, + 0.0015317588113248348, + 0.0017741969786584377, + 0.001661387155763805, + 0.0014756993623450398, + 0.0015531973913311958, + 0.002557002706453204, + 0.0027841837145388126, + 0.001417214865796268, + 0.00134780362714082, + 0.0017819214845076203, + 0.0014591807266697288, + 0.0016900773625820875, + 0.0010743413586169481, + 0.0016698514809831977, + 0.002545443829149008, + 0.001234953524544835, + 0.0020122353453189135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 6 + ], + "selected_reward": -0.018172437325119972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An immersive first-person camera places us in the ornate cockpit of a vast mechanical cog train, gazing out through a vast, curved crystal porthole framed by an intricately carved brass frame. The control panel itself is a work of art, its magical crystal panels shimmering with runes interwoven with countless precisely interlocking brass instrument panels. A core power crystal at its center pulses with a steady and powerful blue light. The train struggles upward along an impossibly thick, towering vine, its every advance accompanied by a deep, rhythmic mechanical throb, as if the entire world trembles slightly under the force. At the edge of our vision, massive brass drive gears can be seen, locked with thunderous force into the crystalline grooves of the vine's surface, emitting golden sparks of mechanics and magic. Outside the window, an otherworldly scene unfolded: countless verdant floating islands dotted a sky shimmering in purple and gold, some glimpsed with elegant spires. Magnificent waterfalls of energy, like liquid light, cascaded down from the edges of lofty islands or through breaks in the clouds, silently washing over the vast expanse. The entire scene, adorned with the sumptuous style of Final Fantasy, seamlessly blended the mechanical aesthetics of steampunk with the magical wonders of high fantasy, creating an epic atmosphere of vastness, grandeur, and endless desire for exploration.", + "seed": 10368, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_10/index_9143_seed_10368_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016120159439742564, + "vlm_logprob": -0.016120159439742564, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.016120159439742564, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013875334523618221 + ], + "opposite_logprob_tokens": [ + -4.638875484466553 + ], + "logprob_reward": -0.013875334523618221, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.013875334523618221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1319815069437027, + 0.05603370815515518, + 0.0632144957780838, + 0.04959496483206749, + 0.29486456513404846, + 0.10656502097845078, + 0.10595626384019852, + 0.04231354966759682, + 0.063597671687603, + 0.08587832003831863 + ], + "credit_map": [ + 0.011070729233324528, + 0.018688220530748367, + 0.01687202975153923, + 0.02975272946059704, + 0.013772879727184772, + 0.011448648758232594, + 0.016494274139404297, + 0.013881996273994446, + 0.005559199023991823, + 0.007212991360574961, + 0.0068426174111664295, + 0.006226713303476572, + 0.005402485840022564, + 0.006679709535092115, + 0.007115528918802738, + 0.010994459502398968, + 0.006355405785143375, + 0.004167174454778433, + 0.007916274480521679, + 0.013496465049684048, + 0.007000931538641453, + 0.006509267725050449, + 0.011439291760325432, + 0.006329681724309921, + 0.004234146326780319, + 0.003953329287469387, + 0.005252355709671974, + 0.011412948369979858, + 0.004886666778475046, + 0.00287849479354918, + 0.012383950874209404, + 0.004593075253069401, + 0.025323329493403435, + 0.059982769191265106, + 0.0375271737575531, + 0.03816240653395653, + 0.04867717996239662, + 0.027632838115096092, + 0.027260687202215195, + 0.030298177152872086, + 0.013680553995072842, + 0.009435470215976238, + 0.01039293222129345, + 0.0189670417457819, + 0.01852315478026867, + 0.011328454129397869, + 0.011077442206442356, + 0.013159971684217453, + 0.009862084873020649, + 0.00541659165173769, + 0.010066269896924496, + 0.03410198166966438, + 0.004998433869332075, + 0.008455226197838783, + 0.007501403801143169, + 0.025554265826940536, + 0.002285256050527096, + 0.0051252529956400394, + 0.0042329831048846245, + 0.008153781294822693, + 0.01032018568366766, + 0.0023347106762230396, + 0.003481016494333744, + 0.006380367558449507, + 0.006056344602257013, + 0.008620156906545162, + 0.011621152982115746, + 0.005715914070606232, + 0.007650869898498058, + 0.005616859998553991, + 0.005511321127414703, + 0.01280505396425724, + 0.008059597574174404, + 0.008338136598467827, + 0.008343357592821121, + 0.02494162507355213, + 0.011052797548472881, + 0.005058095324784517, + 0.010570229962468147, + 0.00951448269188404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.013875334523618221 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.000929519534111023 + ], + "opposite_logprob_tokens": [ + -7.000929355621338 + ], + "logprob_reward": -0.000929519534111023, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.000929519534111023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10553532838821411, + 0.05039304494857788, + 0.028731128200888634, + 0.06912341713905334, + 0.21906140446662903, + 0.09938070178031921, + 0.14877337217330933, + 0.1897142380475998, + 0.03825710341334343, + 0.051030222326517105 + ], + "credit_map": [ + 0.011690313927829266, + 0.01018703356385231, + 0.014617811888456345, + 0.015891922637820244, + 0.017231807112693787, + 0.010137764737010002, + 0.008809714578092098, + 0.01696896366775036, + 0.005403414368629456, + 0.007501570042222738, + 0.00510691711679101, + 0.0058002476580441, + 0.0073424107395112514, + 0.00543943140655756, + 0.0083819180727005, + 0.0054171388037502766, + 0.0036945415195077658, + 0.004122846759855747, + 0.0033987401984632015, + 0.003268579486757517, + 0.00556702958419919, + 0.0020967889577150345, + 0.004479316994547844, + 0.0021032884251326323, + 0.008373096585273743, + 0.005502238869667053, + 0.008687152527272701, + 0.01626906357705593, + 0.0070609962567687035, + 0.0072697740979492664, + 0.009358757175505161, + 0.006602341774851084, + 0.016629327088594437, + 0.03637680411338806, + 0.02318660169839859, + 0.04001264646649361, + 0.03206075355410576, + 0.02341591753065586, + 0.030966171994805336, + 0.01641320064663887, + 0.01255401223897934, + 0.008546094410121441, + 0.009385168552398682, + 0.014575502835214138, + 0.020312685519456863, + 0.005887944251298904, + 0.010968462564051151, + 0.017150843515992165, + 0.010717874392867088, + 0.006911786273121834, + 0.015148796141147614, + 0.05426929146051407, + 0.0077667031437158585, + 0.0045525128953158855, + 0.010991733521223068, + 0.03841467574238777, + 0.01386373583227396, + 0.008994417265057564, + 0.013529958203434944, + 0.07019304484128952, + 0.011226755566895008, + 0.006269455887377262, + 0.015104087069630623, + 0.05053279921412468, + 0.003605968551710248, + 0.004042363725602627, + 0.008341347798705101, + 0.003405324649065733, + 0.004122683312743902, + 0.002638452686369419, + 0.0042665875516831875, + 0.007834377698600292, + 0.00522506283596158, + 0.003913282882422209, + 0.005144978873431683, + 0.012872775085270405, + 0.006774710491299629, + 0.004752477165311575, + 0.007089290767908096, + 0.005257647018879652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.000929519534111023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641948267817497 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641948267817497, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641948267817497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07861842960119247, + 0.06646739691495895, + 0.04300524294376373, + 0.03656104579567909, + 0.21880075335502625, + 0.0827522799372673, + 0.23073865473270416, + 0.1272103637456894, + 0.05095420032739639, + 0.06489158421754837 + ], + "credit_map": [ + 0.007238405290991068, + 0.013729541562497616, + 0.007893276400864124, + 0.010092109441757202, + 0.01136479526758194, + 0.006827619392424822, + 0.01477955561131239, + 0.006693122908473015, + 0.00855428259819746, + 0.006954772397875786, + 0.006462072022259235, + 0.008778130635619164, + 0.007310306187719107, + 0.009074860252439976, + 0.009971587918698788, + 0.009361383505165577, + 0.004462019074708223, + 0.006106168031692505, + 0.004329611547291279, + 0.005907829850912094, + 0.009745756164193153, + 0.0036775688640773296, + 0.00596860283985734, + 0.0028076886665076017, + 0.0037045904900878668, + 0.0039091031067073345, + 0.005220219027251005, + 0.004571238067001104, + 0.005643714219331741, + 0.0030901976861059666, + 0.005931886378675699, + 0.004490096587687731, + 0.017057131975889206, + 0.04507770389318466, + 0.025679947808384895, + 0.031422603875398636, + 0.03599843010306358, + 0.0218413807451725, + 0.02116597257554531, + 0.020557571202516556, + 0.010493227280676365, + 0.007245433516800404, + 0.010686412453651428, + 0.01010242011398077, + 0.011613772250711918, + 0.00795549526810646, + 0.010092731565237045, + 0.014562786556780338, + 0.016522619873285294, + 0.010600600391626358, + 0.024386856704950333, + 0.08020985126495361, + 0.007318981923162937, + 0.008351084776222706, + 0.01658198982477188, + 0.06676669418811798, + 0.008647669106721878, + 0.011715360917150974, + 0.010707308538258076, + 0.03991873189806938, + 0.02770662121474743, + 0.005514793563634157, + 0.010804183781147003, + 0.01219569705426693, + 0.004613671451807022, + 0.006727900821715593, + 0.009755895473062992, + 0.005726855713874102, + 0.004925006069242954, + 0.004015692975372076, + 0.006574450526386499, + 0.008614730089902878, + 0.0049747563898563385, + 0.0059251743368804455, + 0.007274329662322998, + 0.017641467973589897, + 0.007812080439180136, + 0.00386512721888721, + 0.010482807643711567, + 0.00691584125161171 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.004641948267817497 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012528171762824059 + ], + "opposite_logprob_tokens": [ + -4.387527942657471 + ], + "logprob_reward": -0.012528171762824059, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012528171762824059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08534154295921326, + 0.09407614916563034, + 0.02558630146086216, + 0.042731624096632004, + 0.20122478902339935, + 0.02810981497168541, + 0.2050265073776245, + 0.1832980513572693, + 0.07158675789833069, + 0.06301853805780411 + ], + "credit_map": [ + 0.006467725150287151, + 0.010845650918781757, + 0.010788952931761742, + 0.010818853043019772, + 0.018691085278987885, + 0.007794477511197329, + 0.011982662603259087, + 0.007952116429805756, + 0.007064084522426128, + 0.009006059728562832, + 0.008194302208721638, + 0.013914365321397781, + 0.013618540950119495, + 0.008849355392158031, + 0.023409482091665268, + 0.010019957087934017, + 0.0024403510615229607, + 0.002559519372880459, + 0.0029267685022205114, + 0.004733665380626917, + 0.0045501855202019215, + 0.001494712196290493, + 0.0047090379521250725, + 0.0021720589138567448, + 0.003792739938944578, + 0.0037830574437975883, + 0.005359788425266743, + 0.009684599004685879, + 0.005618480034172535, + 0.0031233292538672686, + 0.0073975068517029285, + 0.003972121048718691, + 0.015636246651411057, + 0.04050002992153168, + 0.0225191842764616, + 0.029851680621504784, + 0.033702220767736435, + 0.0209807176142931, + 0.020290205255150795, + 0.017744485288858414, + 0.004321847576647997, + 0.0021844101138412952, + 0.0038776015862822533, + 0.005601327400654554, + 0.0027914263773709536, + 0.0035716898273676634, + 0.002587592927739024, + 0.00317391543649137, + 0.015165179967880249, + 0.010118305683135986, + 0.020835435017943382, + 0.06841620057821274, + 0.007427781354635954, + 0.008197681047022343, + 0.015310834161937237, + 0.05955507233738899, + 0.0132871363312006, + 0.012837642803788185, + 0.010733175091445446, + 0.0683521255850792, + 0.027697302401065826, + 0.005932230036705732, + 0.0147154675796628, + 0.029742946848273277, + 0.006802230142056942, + 0.009879841469228268, + 0.012661989778280258, + 0.00762164406478405, + 0.008063756860792637, + 0.004099569749087095, + 0.010265777818858624, + 0.012191945686936378, + 0.006132984068244696, + 0.0055356393568217754, + 0.006237714551389217, + 0.011512894183397293, + 0.009217587299644947, + 0.003773503238335252, + 0.01267187763005495, + 0.007936328649520874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 7 + ], + "selected_reward": -0.012528171762824059 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned in a first-person perspective inside an ornate mechanical cockpit?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the view framed by a curved crystal porthole with an intricately carved brass frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the control panel visible with magical crystal panels featuring runes and interlocking brass instrument panels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a core power crystal visible at the center of the panel pulsing with a steady blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible through the window showing verdant floating islands in a purple and gold sky with energy waterfalls?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048625823110342026 + ], + "opposite_logprob_tokens": [ + -3.0486257076263428 + ], + "logprob_reward": -0.048625823110342026, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048625823110342026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10171870142221451, + 0.03295428678393364, + 0.07996469736099243, + 0.03748983517289162, + 0.13982796669006348, + 0.20253829658031464, + 0.138673335313797, + 0.20982593297958374, + 0.0252850204706192, + 0.03172192722558975 + ], + "credit_map": [ + 0.010706182569265366, + 0.010906211100518703, + 0.009993360377848148, + 0.03258112445473671, + 0.010791287757456303, + 0.004925924818962812, + 0.012888779863715172, + 0.008925838395953178, + 0.003945270087569952, + 0.00323937414214015, + 0.002956157550215721, + 0.005163319408893585, + 0.004258261062204838, + 0.0028976881876587868, + 0.0071084992960095406, + 0.0033857165835797787, + 0.009770584292709827, + 0.010747152380645275, + 0.007701218593865633, + 0.00911667849868536, + 0.012232141569256783, + 0.009703542105853558, + 0.01301680039614439, + 0.007676578592509031, + 0.004833015147596598, + 0.00428983848541975, + 0.0037028819788247347, + 0.0054268427193164825, + 0.00530632259324193, + 0.0031300922855734825, + 0.007946365512907505, + 0.0028544773813337088, + 0.011090686544775963, + 0.029327183961868286, + 0.017661739140748978, + 0.01852564327418804, + 0.02352418377995491, + 0.013142953626811504, + 0.01287029217928648, + 0.013685283251106739, + 0.027877770364284515, + 0.017765438184142113, + 0.02247670292854309, + 0.022488193586468697, + 0.028280220925807953, + 0.02380724996328354, + 0.025696801021695137, + 0.03414591774344444, + 0.010371656157076359, + 0.0072702448815107346, + 0.014323459006845951, + 0.05094779282808304, + 0.011479247361421585, + 0.0047705224715173244, + 0.010665273293852806, + 0.02884514071047306, + 0.015601614490151405, + 0.011119029484689236, + 0.012701044790446758, + 0.08017345517873764, + 0.020733751356601715, + 0.006016130559146404, + 0.017122942954301834, + 0.04635796695947647, + 0.0021275714971125126, + 0.003054700791835785, + 0.00563732348382473, + 0.0026396855246275663, + 0.0026109758764505386, + 0.001963021932169795, + 0.0032299868762493134, + 0.004021754954010248, + 0.0025798717979341745, + 0.0024821716360747814, + 0.0035952734760940075, + 0.010676217265427113, + 0.004537587054073811, + 0.0017814668826758862, + 0.0039245616644620895, + 0.0021447769831866026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 4 + ], + "selected_reward": -0.048625823110342026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_tensor([6055, 6055])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_tensor([6055, 6055])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a9ed215532a11019b7099604de4075927c2f6489 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_tensor([6055, 6055])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, tracking, full-body photograph follows a lone male figure as he glides gracefully through a dimly lit vinegar brewery, carrying liquid fluid. Bare-chested, his lean, athletic torso gleams in the warmth, emitting a delicate odor of moisture. His dark, damp hair clings to his forehead, and his expression is filled with pensive concentration. With a sensual and suggestive style, he meanders through the narrow gaps between massive, ancient oak barrels, his hands sliding across the rough, dark wood as he navigates the space. Each movement is slow and deliberate; his torso twists fluidly, his feet, clad in simple dark trousers, press silently against the cool, uneven stone floor. Towering wooden barrels, bound by rusted iron rings, fill the vast space, their surfaces revealing an ancient patina. The air is thick and acrid, dust dancing in the warm, faint light that penetrates the darkness through the tall, grimy windows. This warm lighting casts long, deep shadows that shape the outlines of his body and highlight the rich, dark brown grain of the wood, creating an atmosphere where his patient, controlled dance echoes the slow, silent, and ancient process of fermentation that takes place within the barrels around him.", + "seed": 9338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_6055_seed_9338_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.007024966459721327, + "vlm_logprob": -0.007024966459721327, + "vlm_yes_no_margin": 5.425, + "vlm_token_credit": -0.007024966459721327, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00443486962467432 + ], + "opposite_logprob_tokens": [ + -6.254435062408447 + ], + "logprob_reward": -0.00443486962467432, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00443486962467432, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1940089613199234, + 0.04082232713699341, + 0.08058788627386093, + 0.15470583736896515, + 0.05742437392473221, + 0.0680772066116333, + 0.10918642580509186, + 0.09008044749498367, + 0.1507498323917389, + 0.054356709122657776 + ], + "credit_map": [ + 0.01263744942843914, + 0.02581622265279293, + 0.016344169154763222, + 0.0398477241396904, + 0.01326416153460741, + 0.012606076896190643, + 0.026428213343024254, + 0.04706493765115738, + 0.004495455417782068, + 0.003654328640550375, + 0.00593480933457613, + 0.006438808050006628, + 0.0034178609494119883, + 0.005564605351537466, + 0.005544590298086405, + 0.005771871190518141, + 0.006279051769524813, + 0.008583459071815014, + 0.006046876311302185, + 0.011823180131614208, + 0.0071653644554317, + 0.017754392698407173, + 0.010531283915042877, + 0.01240427978336811, + 0.009749610908329487, + 0.011686250567436218, + 0.007901780307292938, + 0.04982035234570503, + 0.00851032231003046, + 0.008377687074244022, + 0.03445904329419136, + 0.024200791493058205, + 0.00852429773658514, + 0.006870679557323456, + 0.0041962130926549435, + 0.0094748605042696, + 0.003639473579823971, + 0.004159896168857813, + 0.01169492956250906, + 0.008864026516675949, + 0.008062519133090973, + 0.005932352039963007, + 0.00608086446300149, + 0.010606902651488781, + 0.00543726934120059, + 0.007082355674356222, + 0.015616640448570251, + 0.009258303791284561, + 0.007363689597696066, + 0.004021718166768551, + 0.021313441917300224, + 0.012474681250751019, + 0.006642816122621298, + 0.011095044203102589, + 0.03369140625, + 0.012583623640239239, + 0.008420567028224468, + 0.007806740701198578, + 0.006894601508975029, + 0.012512771412730217, + 0.012881684117019176, + 0.013499204069375992, + 0.01643839292228222, + 0.01162648480385542, + 0.012898808345198631, + 0.008440293371677399, + 0.009152309969067574, + 0.025299472734332085, + 0.026260768994688988, + 0.01870407909154892, + 0.03169497102499008, + 0.018299128860235214, + 0.005243885796517134, + 0.003542934078723192, + 0.004890316165983677, + 0.008028130047023296, + 0.007985618896782398, + 0.0058534638956189156, + 0.012001401744782925, + 0.006810956168919802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.00443486962467432 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648086339235306 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.008648086339235306, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648086339235306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21750710904598236, + 0.1257018744945526, + 0.20512500405311584, + 0.026625361293554306, + 0.05934568867087364, + 0.038194309920072556, + 0.09749189019203186, + 0.11185682564973831, + 0.08510660380125046, + 0.03304538130760193 + ], + "credit_map": [ + 0.016570154577493668, + 0.024733802303671837, + 0.016400672495365143, + 0.036897655576467514, + 0.015825267881155014, + 0.030864167958498, + 0.03236621990799904, + 0.0438491515815258, + 0.010967311449348927, + 0.010448435321450233, + 0.011067747138440609, + 0.017644139006733894, + 0.010609273798763752, + 0.03160271421074867, + 0.01604592055082321, + 0.017316322773694992, + 0.014730571769177914, + 0.023311125114560127, + 0.014962938614189625, + 0.027179058641195297, + 0.01999950036406517, + 0.04859580844640732, + 0.02704494446516037, + 0.029301052913069725, + 0.0021366518922150135, + 0.0032496193889528513, + 0.0018269954016432166, + 0.006459183990955353, + 0.001519112614914775, + 0.002640130463987589, + 0.005076820030808449, + 0.0037168473936617374, + 0.007712979335337877, + 0.007245711050927639, + 0.004474112764000893, + 0.013525690883398056, + 0.0030066356994211674, + 0.003580772550776601, + 0.009890606626868248, + 0.009909179992973804, + 0.002163554774597287, + 0.0025151479057967663, + 0.002869098214432597, + 0.005671302322298288, + 0.0032032830640673637, + 0.0038213059306144714, + 0.01167511846870184, + 0.006275500636547804, + 0.012167378328740597, + 0.0046956418082118034, + 0.026941414922475815, + 0.01046678051352501, + 0.005493305157870054, + 0.013756229542195797, + 0.013455684296786785, + 0.010515457950532436, + 0.011121776886284351, + 0.008573844097554684, + 0.007473580073565245, + 0.016277585178613663, + 0.017928725108504295, + 0.017728567123413086, + 0.018909869715571404, + 0.013842880725860596, + 0.008429118432104588, + 0.005301838740706444, + 0.005826244130730629, + 0.013003962114453316, + 0.01334321592003107, + 0.010657859966158867, + 0.01769055426120758, + 0.010853811167180538, + 0.003918980248272419, + 0.0035566664300858974, + 0.003014086512848735, + 0.0039328704588115215, + 0.004275807179510593, + 0.004382729064673185, + 0.005171617027372122, + 0.004792625550180674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.008648086339235306 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016066256910562515 + ], + "opposite_logprob_tokens": [ + -4.141066074371338 + ], + "logprob_reward": -0.016066256910562515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016066256910562515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1121331974864006, + 0.05442407727241516, + 0.13432002067565918, + 0.16344645619392395, + 0.08619372546672821, + 0.043507304042577744, + 0.059010569006204605, + 0.05442507565021515, + 0.2553148567676544, + 0.037224747240543365 + ], + "credit_map": [ + 0.007213179022073746, + 0.008942521177232265, + 0.007384808734059334, + 0.02312208153307438, + 0.005468450952321291, + 0.014704166911542416, + 0.014967577531933784, + 0.030330415815114975, + 0.004462558776140213, + 0.0064432877115905285, + 0.006119418889284134, + 0.011968242935836315, + 0.003700339002534747, + 0.0051802718080580235, + 0.010213883593678474, + 0.006336074322462082, + 0.008893415331840515, + 0.016029952093958855, + 0.007403986062854528, + 0.02494131401181221, + 0.01051146350800991, + 0.024902043864130974, + 0.021516062319278717, + 0.020121777430176735, + 0.00989295169711113, + 0.009733740240335464, + 0.008885796181857586, + 0.051082052290439606, + 0.008844365365803242, + 0.010137105360627174, + 0.03835489973425865, + 0.0265155378729105, + 0.009161080233752728, + 0.008974368683993816, + 0.0062620071694254875, + 0.017856324091553688, + 0.004531609360128641, + 0.005983450915664434, + 0.01960613951086998, + 0.013818750157952309, + 0.005155866965651512, + 0.003874800866469741, + 0.003942027688026428, + 0.0059993695467710495, + 0.002323615364730358, + 0.005835786461830139, + 0.01019711047410965, + 0.006178725510835648, + 0.005223731976002455, + 0.004365038126707077, + 0.009986039251089096, + 0.009465904906392097, + 0.005220821592956781, + 0.00589807890355587, + 0.012119325809180737, + 0.00673162704333663, + 0.005738398525863886, + 0.00496379379183054, + 0.0044535803608596325, + 0.008051681332290173, + 0.0063335890881717205, + 0.006731600966304541, + 0.01187314186245203, + 0.006279291119426489, + 0.018554260954260826, + 0.019113346934318542, + 0.017415640875697136, + 0.03956032171845436, + 0.04446295276284218, + 0.0360511913895607, + 0.04213244467973709, + 0.038024693727493286, + 0.003806669032201171, + 0.0020479392260313034, + 0.0027539634611457586, + 0.006835784297436476, + 0.00413610739633441, + 0.004792103078216314, + 0.007327765692025423, + 0.00552441505715251 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.016066256910562515 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013375869020819664 + ], + "opposite_logprob_tokens": [ + -6.62633752822876 + ], + "logprob_reward": -0.0013375869020819664, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013375869020819664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16106078028678894, + 0.07265636324882507, + 0.07850804924964905, + 0.10709866136312485, + 0.03905974701046944, + 0.08842264860868454, + 0.11656142771244049, + 0.12340086698532104, + 0.11220743507146835, + 0.10102400928735733 + ], + "credit_map": [ + 0.013736976310610771, + 0.01075784396380186, + 0.01993395946919918, + 0.029380615800619125, + 0.008439116179943085, + 0.017456116154789925, + 0.022412383928894997, + 0.03894377127289772, + 0.006896934937685728, + 0.007799321319907904, + 0.0075709703378379345, + 0.011149749159812927, + 0.005472532007843256, + 0.015187463723123074, + 0.008646826259791851, + 0.009932569228112698, + 0.006744080688804388, + 0.010229107923805714, + 0.005257348995655775, + 0.014153072610497475, + 0.007144928444176912, + 0.01417117565870285, + 0.011187256313860416, + 0.009621089324355125, + 0.007223604246973991, + 0.010651949793100357, + 0.0071167307905852795, + 0.032095976173877716, + 0.00632415059953928, + 0.007432879880070686, + 0.02006535790860653, + 0.016188016161322594, + 0.004229732323437929, + 0.0034071586560457945, + 0.0030490104109048843, + 0.009008084423840046, + 0.0019994399044662714, + 0.002703637583181262, + 0.008394367061555386, + 0.006268317345529795, + 0.010693748481571674, + 0.00837321113795042, + 0.008894321508705616, + 0.013143395073711872, + 0.007473023608326912, + 0.009687558747828007, + 0.018888866528868675, + 0.011268530040979385, + 0.010507079772651196, + 0.005541914142668247, + 0.016802458092570305, + 0.01465126033872366, + 0.008720762096345425, + 0.018422899767756462, + 0.02652108483016491, + 0.015393977053463459, + 0.012520324438810349, + 0.009718447923660278, + 0.007759629748761654, + 0.019593508914113045, + 0.0196955818682909, + 0.01730547659099102, + 0.022704612463712692, + 0.014103296212852001, + 0.012020223774015903, + 0.008533208630979061, + 0.00844978354871273, + 0.01332894153892994, + 0.013854055665433407, + 0.018510980531573296, + 0.023224515840411186, + 0.014285742305219173, + 0.008564438670873642, + 0.00595670472830534, + 0.006323660723865032, + 0.01630716770887375, + 0.010720611549913883, + 0.01411699503660202, + 0.021864308044314384, + 0.01717013120651245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.0013375869020819664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046380325220525265 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.0046380325220525265, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046380325220525265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15550632774829865, + 0.07259376347064972, + 0.18693313002586365, + 0.13767419755458832, + 0.04917067661881447, + 0.03626466169953346, + 0.1657731980085373, + 0.06804262846708298, + 0.05190567299723625, + 0.07613580673933029 + ], + "credit_map": [ + 0.012897107750177383, + 0.010479030199348927, + 0.01139655988663435, + 0.02604896016418934, + 0.00809409748762846, + 0.017041221261024475, + 0.024509845301508904, + 0.04503950849175453, + 0.005818936973810196, + 0.0046848938800394535, + 0.006424480117857456, + 0.012958559207618237, + 0.004630073439329863, + 0.017355449497699738, + 0.01105201244354248, + 0.009669349528849125, + 0.013485967181622982, + 0.021281477063894272, + 0.012824400328099728, + 0.029785657301545143, + 0.01883406564593315, + 0.03482969105243683, + 0.031096309423446655, + 0.024795550853013992, + 0.007452106568962336, + 0.010003581643104553, + 0.006388302892446518, + 0.044332507997751236, + 0.008203433826565742, + 0.008612710982561111, + 0.0314863882958889, + 0.021195152774453163, + 0.008824917487800121, + 0.008252114988863468, + 0.003431095974519849, + 0.007733079604804516, + 0.0029696805868297815, + 0.004015396814793348, + 0.006544468924403191, + 0.007399918045848608, + 0.0024411550257354975, + 0.0033806595019996166, + 0.0030430590268224478, + 0.0071908533573150635, + 0.003673666389659047, + 0.003783794119954109, + 0.007662205956876278, + 0.005089265294373035, + 0.018917566165328026, + 0.011143145151436329, + 0.029083868488669395, + 0.01769378036260605, + 0.009153909981250763, + 0.030914347618818283, + 0.032192081212997437, + 0.016674481332302094, + 0.006864021997898817, + 0.005580841097980738, + 0.0067903329618275166, + 0.010678431950509548, + 0.010201268829405308, + 0.009118071757256985, + 0.011285580694675446, + 0.007524069864302874, + 0.005566587671637535, + 0.004880261607468128, + 0.0035842987708747387, + 0.005401504226028919, + 0.00594435166567564, + 0.00704198470339179, + 0.009700276888906956, + 0.009786402806639671, + 0.009737782180309296, + 0.007884880527853966, + 0.006803152617067099, + 0.006855212152004242, + 0.0083438940346241, + 0.00939948856830597, + 0.016399279236793518, + 0.010712112300097942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.0046380325220525265 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tracking, full-body photograph follows a lone male figure as he glides gracefully through a dimly lit vinegar brewery, carrying liquid fluid. Bare-chested, his lean, athletic torso gleams in the warmth, emitting a delicate odor of moisture. His dark, damp hair clings to his forehead, and his expression is filled with pensive concentration. With a sensual and suggestive style, he meanders through the narrow gaps between massive, ancient oak barrels, his hands sliding across the rough, dark wood as he navigates the space. Each movement is slow and deliberate; his torso twists fluidly, his feet, clad in simple dark trousers, press silently against the cool, uneven stone floor. Towering wooden barrels, bound by rusted iron rings, fill the vast space, their surfaces revealing an ancient patina. The air is thick and acrid, dust dancing in the warm, faint light that penetrates the darkness through the tall, grimy windows. This warm lighting casts long, deep shadows that shape the outlines of his body and highlight the rich, dark brown grain of the wood, creating an atmosphere where his patient, controlled dance echoes the slow, silent, and ancient process of fermentation that takes place within the barrels around him.", + "seed": 9339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_6055_seed_9339_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016086292150430382, + "vlm_logprob": -0.016086292150430382, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.016086292150430382, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009373347274959087 + ], + "opposite_logprob_tokens": [ + -5.259373188018799 + ], + "logprob_reward": -0.009373347274959087, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009373347274959087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1870688647031784, + 0.09159030020236969, + 0.048575159162282944, + 0.09936248511075974, + 0.03901040554046631, + 0.05028289929032326, + 0.06258004158735275, + 0.21162037551403046, + 0.1406901329755783, + 0.06921937316656113 + ], + "credit_map": [ + 0.015657464042305946, + 0.022341454401612282, + 0.029145384207367897, + 0.02919871173799038, + 0.019665036350488663, + 0.01468938123434782, + 0.03358502313494682, + 0.022786395624279976, + 0.021725725382566452, + 0.007914670743048191, + 0.01161100622266531, + 0.010423657484352589, + 0.00902572087943554, + 0.005044122226536274, + 0.015398751944303513, + 0.01044665090739727, + 0.005456740036606789, + 0.004850444383919239, + 0.0061387126334011555, + 0.007305623963475227, + 0.006114200688898563, + 0.005241124890744686, + 0.007150274235755205, + 0.006318036001175642, + 0.00827532447874546, + 0.005751531571149826, + 0.013515404425561428, + 0.016864877194166183, + 0.017728831619024277, + 0.007398590911179781, + 0.020784759894013405, + 0.009043164551258087, + 0.004122565034776926, + 0.0036649894900619984, + 0.005776728969067335, + 0.006403711624443531, + 0.0058995699509978294, + 0.002543605864048004, + 0.0052978405728936195, + 0.005301394034177065, + 0.007184825837612152, + 0.006869679782539606, + 0.0063968803733587265, + 0.003910091705620289, + 0.0053436290472745895, + 0.006053214427083731, + 0.00713099492713809, + 0.007393582724034786, + 0.004231440369039774, + 0.0058687711134552956, + 0.008838823065161705, + 0.011451641097664833, + 0.008749951608479023, + 0.007474695332348347, + 0.00919409655034542, + 0.006770624779164791, + 0.011585341766476631, + 0.005245683714747429, + 0.06896714866161346, + 0.018669117242097855, + 0.011064829304814339, + 0.010124043561518192, + 0.014104140922427177, + 0.07186005264520645, + 0.007905959151685238, + 0.012292950414121151, + 0.025969263166189194, + 0.020617058500647545, + 0.015528611838817596, + 0.017437640577554703, + 0.020977959036827087, + 0.01996069774031639, + 0.004922348074615002, + 0.006496886722743511, + 0.010262997820973396, + 0.010266389697790146, + 0.008602244779467583, + 0.005864777136594057, + 0.01309996284544468, + 0.009703760966658592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.009373347274959087 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637963369488716 + ], + "opposite_logprob_tokens": [ + -3.6513795852661133 + ], + "logprob_reward": -0.02637963369488716, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637963369488716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23029503226280212, + 0.05037353187799454, + 0.052471548318862915, + 0.08959724754095078, + 0.07108912616968155, + 0.05061447247862816, + 0.03107292763888836, + 0.07359816133975983, + 0.22763104736804962, + 0.12325692176818848 + ], + "credit_map": [ + 0.024298999458551407, + 0.02601981908082962, + 0.03735451400279999, + 0.030806593596935272, + 0.02672557719051838, + 0.013444903306663036, + 0.042961567640304565, + 0.028683029115200043, + 0.0049935863353312016, + 0.0053698234260082245, + 0.0068139140494167805, + 0.006353589706122875, + 0.0060356189496815205, + 0.0035648043267428875, + 0.010904412716627121, + 0.006337775383144617, + 0.008407019078731537, + 0.005332683678716421, + 0.006187114864587784, + 0.00676922220736742, + 0.007086290046572685, + 0.005577299743890762, + 0.007565182633697987, + 0.00554673234000802, + 0.008415766060352325, + 0.005581429693847895, + 0.011439955793321133, + 0.013688982464373112, + 0.016977360472083092, + 0.006550523452460766, + 0.019953303039073944, + 0.006989917252212763, + 0.00699596619233489, + 0.006243560928851366, + 0.01054002158343792, + 0.00933631882071495, + 0.008415517397224903, + 0.005364837124943733, + 0.014289825223386288, + 0.00990307331085205, + 0.010734343901276588, + 0.006195457652211189, + 0.007824155502021313, + 0.004341789986938238, + 0.003949422389268875, + 0.0033359010703861713, + 0.007078760769218206, + 0.007154636085033417, + 0.004540454130619764, + 0.005066958721727133, + 0.004124051425606012, + 0.0045946030877530575, + 0.0019587529823184013, + 0.002414282178506255, + 0.0040551419369876385, + 0.00431867828592658, + 0.004786900710314512, + 0.003971073776483536, + 0.011163647286593914, + 0.014297787100076675, + 0.009444918483495712, + 0.011688998900353909, + 0.007493982557207346, + 0.010750840418040752, + 0.013647936284542084, + 0.018492726609110832, + 0.04601931944489479, + 0.020725620910525322, + 0.01784118264913559, + 0.034871526062488556, + 0.03447428718209267, + 0.041558440774679184, + 0.009713209234178066, + 0.007007572799921036, + 0.039601802825927734, + 0.007286742329597473, + 0.0066148084588348866, + 0.008759228512644768, + 0.012756284326314926, + 0.031517256051301956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.02637963369488716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809339925646782 + ], + "opposite_logprob_tokens": [ + -3.288093328475952 + ], + "logprob_reward": -0.03809339925646782, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809339925646782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22722995281219482, + 0.09165067970752716, + 0.06395253539085388, + 0.05661914870142937, + 0.07679467648267746, + 0.06529191136360168, + 0.1338481903076172, + 0.06108912080526352, + 0.09521586447954178, + 0.12830790877342224 + ], + "credit_map": [ + 0.015543612651526928, + 0.02032550238072872, + 0.03134062513709068, + 0.03252468258142471, + 0.029116874560713768, + 0.015145900659263134, + 0.05836283415555954, + 0.024869946762919426, + 0.013900830410420895, + 0.008637773804366589, + 0.012010454200208187, + 0.010964713990688324, + 0.009806145913898945, + 0.005868495441973209, + 0.018006274476647377, + 0.012455996125936508, + 0.004803151823580265, + 0.00503673916682601, + 0.009392148815095425, + 0.008696021512150764, + 0.00870837364345789, + 0.006698465906083584, + 0.012893756851553917, + 0.007723882328718901, + 0.007054316345602274, + 0.0041766418144106865, + 0.007246439810842276, + 0.007325753103941679, + 0.00796342920511961, + 0.006597017869353294, + 0.00991940125823021, + 0.006336156744509935, + 0.0041671874932944775, + 0.004344712942838669, + 0.010900327004492283, + 0.011738759465515614, + 0.010201537050306797, + 0.006693398579955101, + 0.0213315412402153, + 0.0074172173626720905, + 0.0047339447773993015, + 0.0039991093799471855, + 0.009507929906249046, + 0.0073421974666416645, + 0.007847778499126434, + 0.004868125542998314, + 0.014652113430202007, + 0.012340717948973179, + 0.007236317731440067, + 0.009798458777368069, + 0.017904061824083328, + 0.02920699119567871, + 0.009825540706515312, + 0.013496694155037403, + 0.017842721194028854, + 0.028537414968013763, + 0.004251864738762379, + 0.0032347794622182846, + 0.008457737974822521, + 0.014289413578808308, + 0.006644005421549082, + 0.006113842129707336, + 0.008767331019043922, + 0.009330150671303272, + 0.006297131069004536, + 0.008488364517688751, + 0.01703571528196335, + 0.010800331830978394, + 0.009684057906270027, + 0.013800915330648422, + 0.015843717381358147, + 0.013265641406178474, + 0.006824236363172531, + 0.012671831995248795, + 0.024120688438415527, + 0.014666053466498852, + 0.015833750367164612, + 0.011971786618232727, + 0.02298104017972946, + 0.01923852600157261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.03809339925646782 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019442004850134254 + ], + "opposite_logprob_tokens": [ + -6.251944065093994 + ], + "logprob_reward": -0.0019442004850134254, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019442004850134254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19915537536144257, + 0.1760501265525818, + 0.04338602349162102, + 0.08603230863809586, + 0.15725769102573395, + 0.06352351605892181, + 0.03029177524149418, + 0.07777353376150131, + 0.0681033656001091, + 0.09842634946107864 + ], + "credit_map": [ + 0.015482371672987938, + 0.025616761296987534, + 0.02764650620520115, + 0.03001682460308075, + 0.01742006652057171, + 0.014667210169136524, + 0.0438418872654438, + 0.024463731795549393, + 0.04932258278131485, + 0.014397384598851204, + 0.023929418995976448, + 0.02153533510863781, + 0.01591988280415535, + 0.00735468789935112, + 0.023667465895414352, + 0.019923362880945206, + 0.003987657837569714, + 0.002786040771752596, + 0.005907805636525154, + 0.006721404381096363, + 0.005368294194340706, + 0.0031671635806560516, + 0.009792545810341835, + 0.005655109416693449, + 0.007652845699340105, + 0.005313711240887642, + 0.011160265654325485, + 0.013401290401816368, + 0.014457658864557743, + 0.006096667610108852, + 0.019737178459763527, + 0.00821269117295742, + 0.008656230755150318, + 0.011382817290723324, + 0.022045211866497993, + 0.02079443819820881, + 0.022018849849700928, + 0.012417818419635296, + 0.04643470048904419, + 0.013507620431482792, + 0.005483036860823631, + 0.00555642182007432, + 0.009561914950609207, + 0.01049947552382946, + 0.0064124129712581635, + 0.003773910691961646, + 0.013512465171515942, + 0.008723879233002663, + 0.002538448665291071, + 0.00222455570474267, + 0.0040983762592077255, + 0.006106528453528881, + 0.0034567106049507856, + 0.0033688873518258333, + 0.0037961143534630537, + 0.004702154081314802, + 0.00502375653013587, + 0.0027863443829119205, + 0.023103389889001846, + 0.0060982368886470795, + 0.0059802173636853695, + 0.005046815145760775, + 0.005895677488297224, + 0.023839091882109642, + 0.005691565107554197, + 0.006767545826733112, + 0.010149240493774414, + 0.007970331236720085, + 0.008553886786103249, + 0.008696124888956547, + 0.010918693616986275, + 0.00935597624629736, + 0.006404906045645475, + 0.00571655435487628, + 0.02241709642112255, + 0.01268758438527584, + 0.010098746046423912, + 0.00832754373550415, + 0.018127940595149994, + 0.014645974151790142 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.0019442004850134254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004640880040824413 + ], + "opposite_logprob_tokens": [ + -5.379641056060791 + ], + "logprob_reward": -0.004640880040824413, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004640880040824413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16896957159042358, + 0.07793140411376953, + 0.1059468537569046, + 0.10142369568347931, + 0.12579642236232758, + 0.08311359584331512, + 0.050498537719249725, + 0.05815599858760834, + 0.1147160604596138, + 0.11344783753156662 + ], + "credit_map": [ + 0.013231818564236164, + 0.017449170351028442, + 0.02423858270049095, + 0.024582941085100174, + 0.02284982055425644, + 0.01036924496293068, + 0.037927448749542236, + 0.018320538103580475, + 0.01481165736913681, + 0.006897398736327887, + 0.009907049126923084, + 0.010699749924242496, + 0.009843984618782997, + 0.004639124032109976, + 0.011122291907668114, + 0.010010146535933018, + 0.011638827621936798, + 0.009414274245500565, + 0.013328760862350464, + 0.019219594076275826, + 0.014642205089330673, + 0.006789426319301128, + 0.01891275681555271, + 0.012001014314591885, + 0.008302578702569008, + 0.0061220768839120865, + 0.012768416665494442, + 0.017530173063278198, + 0.01844102144241333, + 0.008606603369116783, + 0.021854890510439873, + 0.007797929458320141, + 0.00929279811680317, + 0.012695450335741043, + 0.020337820053100586, + 0.016604727134108543, + 0.013788393698632717, + 0.00969534832984209, + 0.024179067462682724, + 0.019202813506126404, + 0.008184600621461868, + 0.007036897819489241, + 0.011722119525074959, + 0.009593501687049866, + 0.008945705369114876, + 0.005273101385682821, + 0.015410413034260273, + 0.016947250813245773, + 0.003579176729544997, + 0.004601003602147102, + 0.00626486586406827, + 0.010545479133725166, + 0.004525335505604744, + 0.004735090304166079, + 0.006878158077597618, + 0.009369428269565105, + 0.004445680417120457, + 0.003059360198676586, + 0.011248160153627396, + 0.0075935726054012775, + 0.006337957456707954, + 0.0055616009049117565, + 0.006535798776894808, + 0.013373871333897114, + 0.006891886703670025, + 0.008420421741902828, + 0.01957928016781807, + 0.011042498983442783, + 0.011506900191307068, + 0.016368744894862175, + 0.024973832070827484, + 0.015932494774460793, + 0.007477802690118551, + 0.010807778686285019, + 0.024724476039409637, + 0.009723453782498837, + 0.012921343557536602, + 0.008797505870461464, + 0.01781761832535267, + 0.021177858114242554 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.004640880040824413 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tracking, full-body photograph follows a lone male figure as he glides gracefully through a dimly lit vinegar brewery, carrying liquid fluid. Bare-chested, his lean, athletic torso gleams in the warmth, emitting a delicate odor of moisture. His dark, damp hair clings to his forehead, and his expression is filled with pensive concentration. With a sensual and suggestive style, he meanders through the narrow gaps between massive, ancient oak barrels, his hands sliding across the rough, dark wood as he navigates the space. Each movement is slow and deliberate; his torso twists fluidly, his feet, clad in simple dark trousers, press silently against the cool, uneven stone floor. Towering wooden barrels, bound by rusted iron rings, fill the vast space, their surfaces revealing an ancient patina. The air is thick and acrid, dust dancing in the warm, faint light that penetrates the darkness through the tall, grimy windows. This warm lighting casts long, deep shadows that shape the outlines of his body and highlight the rich, dark brown grain of the wood, creating an atmosphere where his patient, controlled dance echoes the slow, silent, and ancient process of fermentation that takes place within the barrels around him.", + "seed": 9340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_6055_seed_9340_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012825837335549295, + "vlm_logprob": -0.012825837335549295, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.012825837335549295, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004343362059444189 + ], + "opposite_logprob_tokens": [ + -5.879343509674072 + ], + "logprob_reward": -0.004343362059444189, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004343362059444189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6840386390686035, + 0.043891541659832, + 0.032556675374507904, + 0.04426899179816246, + 0.0236345287412405, + 0.032549697905778885, + 0.015563872642815113, + 0.024755312129855156, + 0.05674733594059944, + 0.04199351742863655 + ], + "credit_map": [ + 0.04092036187648773, + 0.022015688940882683, + 0.2569788694381714, + 0.037390466779470444, + 0.023716852068901062, + 0.028566131368279457, + 0.03260313719511032, + 0.24184700846672058, + 0.0037913864944130182, + 0.003287537954747677, + 0.006594065111130476, + 0.0046857367269694805, + 0.005289438646286726, + 0.0053322031162679195, + 0.005276037380099297, + 0.009635125286877155, + 0.003517524106428027, + 0.003026828169822693, + 0.005465834401547909, + 0.0050487760454416275, + 0.003043064381927252, + 0.00269761448726058, + 0.0038898042403161526, + 0.00586722232401371, + 0.002174686873331666, + 0.0026886032428592443, + 0.00820916797965765, + 0.012924384325742722, + 0.001889886800199747, + 0.002226509852334857, + 0.0063668545335531235, + 0.00778889050707221, + 0.003173036500811577, + 0.0034050727263092995, + 0.002960651181638241, + 0.003809552639722824, + 0.0013506230898201466, + 0.002319039311259985, + 0.0029765539802610874, + 0.003639994189143181, + 0.0031609488651156425, + 0.006245797965675592, + 0.0036896171513944864, + 0.00625932402908802, + 0.0017979119438678026, + 0.002275274135172367, + 0.0037162930238991976, + 0.005404522176831961, + 0.0018504852196201682, + 0.0018234285525977612, + 0.00172988954000175, + 0.002557717263698578, + 0.002212129533290863, + 0.0009603108046576381, + 0.0022158436477184296, + 0.0022140645887702703, + 0.002885574009269476, + 0.002546325558796525, + 0.005344826262444258, + 0.003879227675497532, + 0.001994483405724168, + 0.0023605625610798597, + 0.0018299729563295841, + 0.003914333879947662, + 0.005512406583875418, + 0.008205546997487545, + 0.01252832729369402, + 0.008662227541208267, + 0.004089884459972382, + 0.0026438157074153423, + 0.004864280112087727, + 0.01024083886295557, + 0.005005925428122282, + 0.0028085114900022745, + 0.010891233570873737, + 0.00456315977498889, + 0.0023146101739257574, + 0.0023871492594480515, + 0.0033531715162098408, + 0.010669746436178684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.004343362059444189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081741079688072 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081741079688072, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081741079688072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44850555062294006, + 0.11253372579813004, + 0.11371935158967972, + 0.06944388896226883, + 0.033701639622449875, + 0.0619940347969532, + 0.04918137192726135, + 0.05132352188229561, + 0.025720156729221344, + 0.033876847475767136 + ], + "credit_map": [ + 0.027913063764572144, + 0.02274751476943493, + 0.13573062419891357, + 0.040788620710372925, + 0.026257336139678955, + 0.02772936224937439, + 0.03262011334300041, + 0.1347188651561737, + 0.009225169196724892, + 0.008126320317387581, + 0.014669553376734257, + 0.01725439727306366, + 0.011302785016596317, + 0.010477419942617416, + 0.02538173831999302, + 0.016096346080303192, + 0.014347849413752556, + 0.010793017223477364, + 0.01507805846631527, + 0.017433568835258484, + 0.008750271052122116, + 0.012668576091527939, + 0.01591287925839424, + 0.018735118210315704, + 0.005546040832996368, + 0.005141892470419407, + 0.009274148382246494, + 0.014347842894494534, + 0.007487734779715538, + 0.007429586257785559, + 0.008910661563277245, + 0.011305981315672398, + 0.003336789784952998, + 0.003649103455245495, + 0.003982312045991421, + 0.005789048969745636, + 0.0022409181110560894, + 0.004268793389201164, + 0.004608983639627695, + 0.00582569045946002, + 0.005159484688192606, + 0.006364724133163691, + 0.005381098482757807, + 0.008141597732901573, + 0.007182100787758827, + 0.010426684282720089, + 0.008895245380699635, + 0.010443096049129963, + 0.005043893586844206, + 0.004156775772571564, + 0.005094766151160002, + 0.005528676323592663, + 0.006703721359372139, + 0.0073636723682284355, + 0.008159401826560497, + 0.007130464073270559, + 0.006110661197453737, + 0.005739392712712288, + 0.008747304789721966, + 0.007940012961626053, + 0.0042207748629152775, + 0.006719203665852547, + 0.004347503650933504, + 0.007498664781451225, + 0.0038232796359807253, + 0.0026478830259293318, + 0.005304950289428234, + 0.0027246035169810057, + 0.003259512363001704, + 0.003093779319897294, + 0.0015177425229921937, + 0.0033484057057648897, + 0.004057693760842085, + 0.0047012618742883205, + 0.004372941330075264, + 0.005866263061761856, + 0.003970725927501917, + 0.0036995629779994488, + 0.0038303975015878677, + 0.0033780038356781006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.016081741079688072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038093626499176025 + ], + "opposite_logprob_tokens": [ + -3.2880935668945312 + ], + "logprob_reward": -0.038093626499176025, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038093626499176025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40588194131851196, + 0.08034063875675201, + 0.0671897754073143, + 0.1065962091088295, + 0.046934083104133606, + 0.039979126304388046, + 0.03367471322417259, + 0.08516468107700348, + 0.07617360353469849, + 0.05806521698832512 + ], + "credit_map": [ + 0.02518538571894169, + 0.014311143197119236, + 0.15343758463859558, + 0.022757142782211304, + 0.012373835779726505, + 0.015228508971631527, + 0.01951705478131771, + 0.1430712640285492, + 0.00672296853736043, + 0.005090371705591679, + 0.009238092228770256, + 0.00840846262872219, + 0.01144708227366209, + 0.013572238385677338, + 0.010620063170790672, + 0.01524136308580637, + 0.0044926973059773445, + 0.006097536534070969, + 0.014187686145305634, + 0.009248859249055386, + 0.006159823387861252, + 0.003844392253085971, + 0.005156768951565027, + 0.018002014607191086, + 0.006696432363241911, + 0.007635760586708784, + 0.018538644537329674, + 0.027829203754663467, + 0.005370589904487133, + 0.0056337518617510796, + 0.017258204519748688, + 0.017633618786931038, + 0.006312661338597536, + 0.003760105697438121, + 0.004758049733936787, + 0.006592193618416786, + 0.005655164364725351, + 0.005586456973105669, + 0.006088769529014826, + 0.008180679753422737, + 0.0033066212199628353, + 0.006344108376652002, + 0.0041729467920959, + 0.007092464715242386, + 0.0023394813761115074, + 0.003510491456836462, + 0.006954534910619259, + 0.006258479785174131, + 0.003489617956802249, + 0.0028798438142985106, + 0.0029408480040729046, + 0.006081843748688698, + 0.007405714597553015, + 0.002360847545787692, + 0.004179706797003746, + 0.0043362900614738464, + 0.007458426058292389, + 0.007674572989344597, + 0.021259767934679985, + 0.01141738798469305, + 0.007699148263782263, + 0.007805099245160818, + 0.008825038559734821, + 0.013025235384702682, + 0.009687713347375393, + 0.008468184620141983, + 0.009195884689688683, + 0.013811535201966763, + 0.007149128243327141, + 0.005866674240678549, + 0.008305358700454235, + 0.013689120300114155, + 0.0053792186081409454, + 0.003698742249980569, + 0.01666608266532421, + 0.0052939727902412415, + 0.00449589965865016, + 0.004124952480196953, + 0.004750110674649477, + 0.013656237162649632 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.038093626499176025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015151937259361148 + ], + "opposite_logprob_tokens": [ + -6.5015153884887695 + ], + "logprob_reward": -0.0015151937259361148, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015151937259361148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6070274114608765, + 0.03293652832508087, + 0.029713226482272148, + 0.026224857196211815, + 0.03696364909410477, + 0.03028961829841137, + 0.07140246033668518, + 0.07348906248807907, + 0.07644879072904587, + 0.015504422597587109 + ], + "credit_map": [ + 0.036642033606767654, + 0.029267119243741035, + 0.220014750957489, + 0.035690974444150925, + 0.024301832541823387, + 0.025309229269623756, + 0.031623441725969315, + 0.20417797565460205, + 0.002130262553691864, + 0.0023887273855507374, + 0.004860872868448496, + 0.005129822064191103, + 0.0029689583461731672, + 0.0031135219614952803, + 0.006429202388972044, + 0.0059151602908968925, + 0.002011932199820876, + 0.0014078024541959167, + 0.00646724971011281, + 0.005222734529525042, + 0.001960637280717492, + 0.0013984788674861193, + 0.0034050983376801014, + 0.007839294150471687, + 0.0018701276276260614, + 0.0030694054439663887, + 0.0036920509301126003, + 0.005871813744306564, + 0.0018800856778398156, + 0.002335132332518697, + 0.0035159962717443705, + 0.003990245517343283, + 0.004842856433242559, + 0.0028448826633393764, + 0.003375433385372162, + 0.005094811320304871, + 0.004355688579380512, + 0.004958976060152054, + 0.004814501851797104, + 0.006676497403532267, + 0.00305196619592607, + 0.004661093931645155, + 0.003217687364667654, + 0.0057524992153048515, + 0.0019088307162746787, + 0.002787993522360921, + 0.004464243538677692, + 0.004445304628461599, + 0.009592113085091114, + 0.005728738848119974, + 0.0062494343146681786, + 0.009360423311591148, + 0.01778998412191868, + 0.005269904620945454, + 0.007806732319295406, + 0.009605132974684238, + 0.007299607153981924, + 0.006275827530771494, + 0.015846524387598038, + 0.012007267214357853, + 0.006403143517673016, + 0.007481479085981846, + 0.007048738654702902, + 0.011126477271318436, + 0.008177298121154308, + 0.009985123760998249, + 0.014715039171278477, + 0.012600711546838284, + 0.004128693602979183, + 0.004394881427288055, + 0.006622235756367445, + 0.015824807807803154, + 0.0026650414802134037, + 0.001827864907681942, + 0.002259795553982258, + 0.002252627396956086, + 0.0014486863510683179, + 0.001730056363157928, + 0.001413813908584416, + 0.0019065364031121135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.0015151937259361148 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004095263313502073 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.004095263313502073, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004095263313502073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5927422046661377, + 0.056388553231954575, + 0.05140449106693268, + 0.06269890815019608, + 0.04279778152704239, + 0.02852347306907177, + 0.027025772258639336, + 0.049930427223443985, + 0.07214056700468063, + 0.016347799450159073 + ], + "credit_map": [ + 0.03676903247833252, + 0.01999124325811863, + 0.2189820259809494, + 0.03394518420100212, + 0.024890664964914322, + 0.026255646720528603, + 0.028257496654987335, + 0.20365092158317566, + 0.005069414619356394, + 0.0031220652163028717, + 0.007842532359063625, + 0.006958250422030687, + 0.003759961109608412, + 0.011575481854379177, + 0.0075379335321486, + 0.010522909462451935, + 0.003909170161932707, + 0.002467662561684847, + 0.012500220909714699, + 0.004907034803181887, + 0.004037442617118359, + 0.004285276401787996, + 0.003758542239665985, + 0.015539144165813923, + 0.003699874272570014, + 0.0045444848947227, + 0.009239817969501019, + 0.01606057770550251, + 0.00464472733438015, + 0.00649274094030261, + 0.008166850544512272, + 0.009849836118519306, + 0.004983161110430956, + 0.002947489032521844, + 0.0046425266191363335, + 0.0051870583556592464, + 0.004617375787347555, + 0.007227081339806318, + 0.0062620798125863075, + 0.006931009702384472, + 0.002372360322624445, + 0.003935051150619984, + 0.0018832641653716564, + 0.004121450241655111, + 0.003811984322965145, + 0.004325974266976118, + 0.0036692903377115726, + 0.0044040982611477375, + 0.0033416482619941235, + 0.003446363378316164, + 0.0022027280647307634, + 0.003216885030269623, + 0.004649873357266188, + 0.002484540920704603, + 0.0037293555215001106, + 0.003954378422349691, + 0.0059082829393446445, + 0.004686607513576746, + 0.007272183429449797, + 0.00838342122733593, + 0.003819370875135064, + 0.009164334274828434, + 0.00398163590580225, + 0.006714589893817902, + 0.007577679120004177, + 0.012392316944897175, + 0.006438171025365591, + 0.014686471782624722, + 0.00726571399718523, + 0.005457173101603985, + 0.008732857182621956, + 0.009590184316039085, + 0.0022086461540311575, + 0.0031727382447570562, + 0.0017223050817847252, + 0.00218671141192317, + 0.002131168730556965, + 0.001653750310651958, + 0.001673301449045539, + 0.0015991792315617204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.004095263313502073 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tracking, full-body photograph follows a lone male figure as he glides gracefully through a dimly lit vinegar brewery, carrying liquid fluid. Bare-chested, his lean, athletic torso gleams in the warmth, emitting a delicate odor of moisture. His dark, damp hair clings to his forehead, and his expression is filled with pensive concentration. With a sensual and suggestive style, he meanders through the narrow gaps between massive, ancient oak barrels, his hands sliding across the rough, dark wood as he navigates the space. Each movement is slow and deliberate; his torso twists fluidly, his feet, clad in simple dark trousers, press silently against the cool, uneven stone floor. Towering wooden barrels, bound by rusted iron rings, fill the vast space, their surfaces revealing an ancient patina. The air is thick and acrid, dust dancing in the warm, faint light that penetrates the darkness through the tall, grimy windows. This warm lighting casts long, deep shadows that shape the outlines of his body and highlight the rich, dark brown grain of the wood, creating an atmosphere where his patient, controlled dance echoes the slow, silent, and ancient process of fermentation that takes place within the barrels around him.", + "seed": 9341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_11/index_6055_seed_9341_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.015025698533281683, + "vlm_logprob": -0.015025698533281683, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.015025698533281683, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025795452296733856 + ], + "opposite_logprob_tokens": [ + -3.7757954597473145 + ], + "logprob_reward": -0.025795452296733856, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025795452296733856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13258367776870728, + 0.06628084182739258, + 0.40970805287361145, + 0.03949218988418579, + 0.023091359063982964, + 0.048462219536304474, + 0.14203277230262756, + 0.02909083105623722, + 0.08274398744106293, + 0.02651411108672619 + ], + "credit_map": [ + 0.008376379497349262, + 0.005783630069345236, + 0.04131709411740303, + 0.008720794692635536, + 0.025317879393696785, + 0.00777431670576334, + 0.027371788397431374, + 0.007921801880002022, + 0.005496501456946135, + 0.003972282167524099, + 0.014878977090120316, + 0.005509298760443926, + 0.011530899442732334, + 0.00499299168586731, + 0.014899168163537979, + 0.005000723525881767, + 0.026572557166218758, + 0.02359555847942829, + 0.07324039936065674, + 0.03622709959745407, + 0.050199709832668304, + 0.014890055172145367, + 0.13549329340457916, + 0.04948938637971878, + 0.003213113872334361, + 0.003272322006523609, + 0.00537056103348732, + 0.004455672577023506, + 0.006314770318567753, + 0.0023175638634711504, + 0.010231522843241692, + 0.0043166629038751125, + 0.0033635671716183424, + 0.002027329755946994, + 0.00359125598333776, + 0.0015290046576410532, + 0.0036105068866163492, + 0.003261633450165391, + 0.0033976673148572445, + 0.0023103945422917604, + 0.004313233774155378, + 0.0030071719083935022, + 0.0056622629053890705, + 0.0071310242637991905, + 0.007480158004909754, + 0.005415819119662046, + 0.011183285154402256, + 0.004269260913133621, + 0.011559282429516315, + 0.013951335102319717, + 0.020173730328679085, + 0.01763404719531536, + 0.024478401988744736, + 0.010616973042488098, + 0.03160510212182999, + 0.012013901956379414, + 0.0023425358813256025, + 0.0021965992636978626, + 0.0036573437973856926, + 0.0030029357876628637, + 0.006536895874887705, + 0.0027661104686558247, + 0.005889681167900562, + 0.002698729047551751, + 0.006569003686308861, + 0.004756472073495388, + 0.012930993922054768, + 0.006811592262238264, + 0.01667238399386406, + 0.006147134117782116, + 0.0220868531614542, + 0.00676956120878458, + 0.001968573546037078, + 0.002980631310492754, + 0.0031257409136742353, + 0.003034633817151189, + 0.006129301618784666, + 0.0028922916390001774, + 0.004406777676194906, + 0.001976160565391183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.025795452296733856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029823429882526398 + ], + "opposite_logprob_tokens": [ + -3.5298235416412354 + ], + "logprob_reward": -0.029823429882526398, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029823429882526398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14294593036174774, + 0.15292972326278687, + 0.12655505537986755, + 0.07170157134532928, + 0.04328193888068199, + 0.032352548092603683, + 0.06795252859592438, + 0.0428842194378376, + 0.12877771258354187, + 0.19061878323554993 + ], + "credit_map": [ + 0.012030631303787231, + 0.011109007522463799, + 0.02368028089404106, + 0.014633569866418839, + 0.03007536008954048, + 0.011589247733354568, + 0.02269626222550869, + 0.017131563276052475, + 0.010598449036478996, + 0.008738629519939423, + 0.04119762033224106, + 0.017571887001395226, + 0.02852230705320835, + 0.007902617566287518, + 0.02874547801911831, + 0.009652746841311455, + 0.009863578714430332, + 0.013731948100030422, + 0.020400362089276314, + 0.014688092283904552, + 0.01331422757357359, + 0.007427779957652092, + 0.029831012710928917, + 0.017298057675361633, + 0.008428097702562809, + 0.007298006676137447, + 0.010348580777645111, + 0.007010641973465681, + 0.011141103692352772, + 0.006554822903126478, + 0.013889350928366184, + 0.007030969485640526, + 0.005979783833026886, + 0.003106068354099989, + 0.006563366390764713, + 0.003874463029205799, + 0.005798427388072014, + 0.006460431031882763, + 0.007018938194960356, + 0.004480460193008184, + 0.002924328437075019, + 0.0026905087288469076, + 0.003807506524026394, + 0.004562888760119677, + 0.005808594170957804, + 0.003950705751776695, + 0.005642376374453306, + 0.0029656391125172377, + 0.006505366414785385, + 0.006202463526278734, + 0.00944564025849104, + 0.007200689986348152, + 0.008098534308373928, + 0.007304383907467127, + 0.01601916365325451, + 0.007176283281296492, + 0.00386575935408473, + 0.0039741406217217445, + 0.00571798300370574, + 0.0045231101103127, + 0.008069616742432117, + 0.0052787926979362965, + 0.006761118303984404, + 0.004693698137998581, + 0.0122095737606287, + 0.008734212256968021, + 0.020907245576381683, + 0.014917847700417042, + 0.013323523104190826, + 0.012577389366924763, + 0.029797635972499847, + 0.016310280188918114, + 0.015795476734638214, + 0.017479531466960907, + 0.02607259899377823, + 0.024635812267661095, + 0.034620825201272964, + 0.018281156197190285, + 0.03531245142221451, + 0.018420949578285217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.029823429882526398 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11290176957845688, + 0.05227992311120033, + 0.11066984385251999, + 0.05671929940581322, + 0.06506362557411194, + 0.20562408864498138, + 0.07256151735782623, + 0.12933538854122162, + 0.07944649457931519, + 0.11539801955223083 + ], + "credit_map": [ + 0.009621731005609035, + 0.009202755056321621, + 0.024040671065449715, + 0.009335716255009174, + 0.02402092143893242, + 0.00963589921593666, + 0.01848386414349079, + 0.008560211397707462, + 0.003623310709372163, + 0.0038629055488854647, + 0.00889506097882986, + 0.004519455134868622, + 0.009231848642230034, + 0.006394256837666035, + 0.008632596582174301, + 0.007120491471141577, + 0.00937805324792862, + 0.009261283092200756, + 0.017778409644961357, + 0.012146838940680027, + 0.018318194895982742, + 0.005759797990322113, + 0.029733266681432724, + 0.008294004015624523, + 0.004956558812409639, + 0.0033973532263189554, + 0.008364261128008366, + 0.006121837999671698, + 0.007386221084743738, + 0.005642295349389315, + 0.013684863224625587, + 0.007165907416492701, + 0.012281941249966621, + 0.005518642254173756, + 0.010542651638388634, + 0.004068621899932623, + 0.0075738937593996525, + 0.0070806508883833885, + 0.009123890660703182, + 0.008873331360518932, + 0.016227779909968376, + 0.016767842695116997, + 0.02658720314502716, + 0.03236578404903412, + 0.024446358904242516, + 0.021666059270501137, + 0.04017370566725731, + 0.027389343827962875, + 0.008057760074734688, + 0.0068839238956570625, + 0.010085046291351318, + 0.008903666399419308, + 0.012129453010857105, + 0.006619139574468136, + 0.013475161045789719, + 0.006407369393855333, + 0.009212757460772991, + 0.009219913743436337, + 0.020381730049848557, + 0.014618122018873692, + 0.017538923770189285, + 0.013404804281890392, + 0.029601041227579117, + 0.015358093194663525, + 0.009303011931478977, + 0.00596285006031394, + 0.011877686716616154, + 0.008071640506386757, + 0.011662350036203861, + 0.00443853996694088, + 0.017717378214001656, + 0.010413036681711674, + 0.00908863265067339, + 0.009936991147696972, + 0.015227844938635826, + 0.015798376873135567, + 0.01654241420328617, + 0.010631504468619823, + 0.029276933521032333, + 0.008895321749150753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 9 + ], + "selected_reward": -0.01253841444849968 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001716808183118701 + ], + "opposite_logprob_tokens": [ + -6.376716613769531 + ], + "logprob_reward": -0.001716808183118701, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.001716808183118701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17111589014530182, + 0.15731854736804962, + 0.07689774036407471, + 0.05182425677776337, + 0.10553262382745743, + 0.09071817994117737, + 0.05176237225532532, + 0.05306402966380119, + 0.12700185179710388, + 0.1147644966840744 + ], + "credit_map": [ + 0.014860519208014011, + 0.014658229425549507, + 0.025834625586867332, + 0.01733855903148651, + 0.03219949081540108, + 0.014936486259102821, + 0.025543052703142166, + 0.025744933634996414, + 0.013274011202156544, + 0.009944542311131954, + 0.03996102511882782, + 0.016560684889554977, + 0.02890206128358841, + 0.008748004212975502, + 0.028659891337156296, + 0.011268324218690395, + 0.005949705373495817, + 0.00959505420178175, + 0.012192514725029469, + 0.009859515354037285, + 0.010614975355565548, + 0.004221915267407894, + 0.01668637990951538, + 0.007777678780257702, + 0.0042721061035990715, + 0.0037967448588460684, + 0.007533504161983728, + 0.0056599280796945095, + 0.0069590904749929905, + 0.003173971315845847, + 0.013364254496991634, + 0.007064654491841793, + 0.012164721265435219, + 0.00929531455039978, + 0.015219022519886494, + 0.008573521859943867, + 0.016304995864629745, + 0.012635370716452599, + 0.01949072629213333, + 0.011848955415189266, + 0.008042833767831326, + 0.00937659852206707, + 0.010293535888195038, + 0.01269032247364521, + 0.013126620091497898, + 0.011519533582031727, + 0.016258686780929565, + 0.009410049766302109, + 0.004502320196479559, + 0.0037711099721491337, + 0.007494369521737099, + 0.005816301796585321, + 0.00943969190120697, + 0.004043886438012123, + 0.011401687748730183, + 0.005293002352118492, + 0.0042092325165867805, + 0.004428303334861994, + 0.006784144323319197, + 0.00593156460672617, + 0.010222170501947403, + 0.004723837599158287, + 0.011945049278438091, + 0.004819727037101984, + 0.012093312107026577, + 0.006780312862247229, + 0.019750675186514854, + 0.013669145293533802, + 0.02084384858608246, + 0.008356099016964436, + 0.03375107795000076, + 0.011757389642298222, + 0.010361792519688606, + 0.009903288446366787, + 0.015017561614513397, + 0.01879606582224369, + 0.016754917800426483, + 0.010838879272341728, + 0.02204606495797634, + 0.011045925319194794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.001716808183118701 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a bare-chested male figure with dark, damp hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a dimly lit vinegar brewery filled with massive, ancient oak barrels bound by rusted iron rings?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the camera using a slow, tracking shot to follow the figure's full body as he moves through the narrow gaps between barrels?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure wear simple dark trousers and move with slow, deliberate, fluid motions while his hands slide across the rough wood?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting warm and faint, casting long, deep shadows and illuminating dust dancing in the air?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254387855529785 + ], + "opposite_logprob_tokens": [ + -5.25525426864624 + ], + "logprob_reward": -0.005254387855529785, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254387855529785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11651512235403061, + 0.11756644397974014, + 0.07719555497169495, + 0.06256742030382156, + 0.07048644870519638, + 0.10408554971218109, + 0.0658002644777298, + 0.17596468329429626, + 0.11815506964921951, + 0.0916634276509285 + ], + "credit_map": [ + 0.010538969188928604, + 0.008637896738946438, + 0.019794845953583717, + 0.012800595723092556, + 0.023271897807717323, + 0.010738345794379711, + 0.018540067598223686, + 0.012192502617835999, + 0.01190973911434412, + 0.007765973452478647, + 0.028866225853562355, + 0.010558213107287884, + 0.02246205136179924, + 0.0070283012464642525, + 0.01997445337474346, + 0.009001492522656918, + 0.007105620112270117, + 0.011002243496477604, + 0.012169393710792065, + 0.007837512530386448, + 0.011608741246163845, + 0.004345739260315895, + 0.014254536479711533, + 0.008871769532561302, + 0.006135750096291304, + 0.006872126832604408, + 0.00905520562082529, + 0.006984257139265537, + 0.008718406781554222, + 0.0051573216915130615, + 0.011399920098483562, + 0.00824443157762289, + 0.011779661290347576, + 0.005402976181358099, + 0.0113099105656147, + 0.00550072593614459, + 0.009055285714566708, + 0.008078239858150482, + 0.011585910804569721, + 0.00777373556047678, + 0.008818153291940689, + 0.010337627492845058, + 0.011593853123486042, + 0.014965496957302094, + 0.014594463631510735, + 0.015690796077251434, + 0.0200232844799757, + 0.008061876520514488, + 0.0064588249661028385, + 0.007016669027507305, + 0.008289926685392857, + 0.008121843449771404, + 0.00984905008226633, + 0.007175702601671219, + 0.01277210470288992, + 0.006116146687418222, + 0.017713461071252823, + 0.013289582915604115, + 0.023545775562524796, + 0.014196223579347134, + 0.036031171679496765, + 0.02239380218088627, + 0.02915392629802227, + 0.019640738144516945, + 0.008092970587313175, + 0.006528492085635662, + 0.018761854618787766, + 0.00976782850921154, + 0.020674046128988266, + 0.010725827887654305, + 0.03150555491447449, + 0.012098496779799461, + 0.007266737055033445, + 0.00867085438221693, + 0.01120897475630045, + 0.015142511576414108, + 0.014726275578141212, + 0.008193794637918472, + 0.01559221651405096, + 0.010862067341804504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.005254387855529785 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_tensor([1340, 1340])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_tensor([1340, 1340])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9a30ab19429f37c5159a4f2a015c6826a4e922af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_tensor([1340, 1340])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, low-angle, medium shot captures a young man, reminiscent of Tooo Takaki, with neatly trimmed black hair and a slender build, standing near a rural railway crossing beneath a spectacular meteor shower. The scene is rendered in a style heavily inspired by Van Gogh's \"Starry Night\"; the sky above is a turbulent vortex of deep, swirling blues and fiery yellows, painted with thick, expressive brushstrokes. Bright meteors streak across this celestial canvas, appearing like thick streaks of golden unglazed paint, leaving behind tangible, textured traces. Below, the young man stands in silent awe, his half-profile face looking upward with a look of deep longing. He is at an intersection where two red traffic lights flash rhythmically, casting intermittent crimson pulses across the snow-covered ground. The swirling, textured light from the sky and the flashing signals create a dramatic play, as shadows writhe across the snow, culminating in a scene of intense, emotional beauty and solitary contemplation.", + "seed": 16829, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_1340_seed_16829_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.043618795461952686, + "vlm_logprob": -0.043618795461952686, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.043618795461952686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13350285589694977 + ], + "opposite_logprob_tokens": [ + -2.133502960205078 + ], + "logprob_reward": -0.13350285589694977, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13350285589694977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18345388770103455, + 0.06290578097105026, + 0.020825622603297234, + 0.05721420794725418, + 0.10820382088422775, + 0.11179032176733017, + 0.12852932512760162, + 0.06776713579893112, + 0.07726980745792389, + 0.18204012513160706 + ], + "credit_map": [ + 0.009996470995247364, + 0.016651013866066933, + 0.013044995255768299, + 0.013952421955764294, + 0.014705868437886238, + 0.01068661455065012, + 0.04926912859082222, + 0.05514736473560333, + 0.003799122292548418, + 0.005390538834035397, + 0.005485199391841888, + 0.004977167584002018, + 0.005242439918220043, + 0.005569349974393845, + 0.013494767248630524, + 0.018947191536426544, + 0.0013711295323446393, + 0.0022278467658907175, + 0.0020399666391313076, + 0.0017705747159197927, + 0.0020307928789407015, + 0.0012060977751389146, + 0.005604888778179884, + 0.00457432447001338, + 0.003823943203315139, + 0.005021997727453709, + 0.0033572642132639885, + 0.004353738855570555, + 0.0035274948459118605, + 0.010774764232337475, + 0.021367147564888, + 0.004987853113561869, + 0.00558340921998024, + 0.006105862557888031, + 0.005831252317875624, + 0.004317456390708685, + 0.006333973724395037, + 0.005551812704652548, + 0.044780369848012924, + 0.029699673876166344, + 0.005984379909932613, + 0.00703100860118866, + 0.006408376153558493, + 0.005251116584986448, + 0.006363761145621538, + 0.0064012352377176285, + 0.045932162553071976, + 0.028418276458978653, + 0.006465238519012928, + 0.008328384719789028, + 0.007035693619400263, + 0.0055771502666175365, + 0.006323280278593302, + 0.01436418667435646, + 0.053975045680999756, + 0.02646034210920334, + 0.003367676166817546, + 0.005959846545010805, + 0.00373129197396338, + 0.005946567747741938, + 0.004393568728119135, + 0.004154985770583153, + 0.021853391081094742, + 0.018359802663326263, + 0.0039010930340737104, + 0.006165390368551016, + 0.004609412048012018, + 0.005728092975914478, + 0.0052739279344677925, + 0.0044962503015995026, + 0.026482930406928062, + 0.020612705498933792, + 0.008262014016509056, + 0.014205776154994965, + 0.01033479068428278, + 0.015115220099687576, + 0.009426763281226158, + 0.010356117971241474, + 0.07324756681919098, + 0.04109186679124832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.13350285589694977 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004110934678465128 + ], + "opposite_logprob_tokens": [ + -5.504110813140869 + ], + "logprob_reward": -0.004110934678465128, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004110934678465128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48616257309913635, + 0.07444745302200317, + 0.0510457418859005, + 0.01860336773097515, + 0.1328107863664627, + 0.03055938147008419, + 0.06136944890022278, + 0.07072276622056961, + 0.04101336747407913, + 0.0332651324570179 + ], + "credit_map": [ + 0.02504478394985199, + 0.02564266137778759, + 0.028511423617601395, + 0.02089056372642517, + 0.02643335983157158, + 0.058658745139837265, + 0.19800862669944763, + 0.10297240316867828, + 0.005388784222304821, + 0.005156438797712326, + 0.005236232187598944, + 0.004451772663742304, + 0.005800167564302683, + 0.016149425879120827, + 0.02585085853934288, + 0.006413770839571953, + 0.0031017730943858624, + 0.003462178399786353, + 0.0035801632329821587, + 0.0027216109447181225, + 0.003506575245410204, + 0.002766894642263651, + 0.020838899537920952, + 0.011067645624279976, + 0.0013676591916009784, + 0.0017176612745970488, + 0.0015511628007516265, + 0.001319157425314188, + 0.001260032644495368, + 0.0036389660090208054, + 0.005852216389030218, + 0.0018965089693665504, + 0.008398469537496567, + 0.007523145992308855, + 0.007808021269738674, + 0.006501336116343737, + 0.00719246594235301, + 0.021489687263965607, + 0.057801757007837296, + 0.01609589345753193, + 0.0027439568657428026, + 0.0021149932872503996, + 0.002426216145977378, + 0.0017998082330450416, + 0.0025862865149974823, + 0.00785017665475607, + 0.00483096344396472, + 0.006206977181136608, + 0.002810448408126831, + 0.0032633633818477392, + 0.003405083203688264, + 0.0023784339427948, + 0.003065185621380806, + 0.005085158161818981, + 0.026567114517092705, + 0.014794660732150078, + 0.003399746958166361, + 0.0036016853991895914, + 0.0037189072463661432, + 0.0029194660019129515, + 0.003664315678179264, + 0.006945622153580189, + 0.03104926086962223, + 0.015423756092786789, + 0.001967431278899312, + 0.0025281370617449284, + 0.0023434418253600597, + 0.001887085265479982, + 0.0023705894127488136, + 0.002549721859395504, + 0.016148947179317474, + 0.011218007653951645, + 0.0022869640961289406, + 0.002610900904983282, + 0.0023833473678678274, + 0.002238933928310871, + 0.0020580897107720375, + 0.0074259815737605095, + 0.01149093545973301, + 0.0027699770871549845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.004110934678465128 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632025051862001 + ], + "opposite_logprob_tokens": [ + -4.882632255554199 + ], + "logprob_reward": -0.007632025051862001, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632025051862001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6177030205726624, + 0.04230378568172455, + 0.11114887148141861, + 0.020537234842777252, + 0.04192860797047615, + 0.020978203043341637, + 0.02198195643723011, + 0.0634842962026596, + 0.030937056988477707, + 0.028996942564845085 + ], + "credit_map": [ + 0.03212745860219002, + 0.03167211636900902, + 0.03734181448817253, + 0.028895271942019463, + 0.03420617803931236, + 0.10265146195888519, + 0.24815617501735687, + 0.10265255719423294, + 0.003335868241265416, + 0.0036584422923624516, + 0.00604533776640892, + 0.004399619530886412, + 0.0051902164705097675, + 0.0096810944378376, + 0.0034550412092357874, + 0.006538168992847204, + 0.006902617402374744, + 0.006960385944694281, + 0.009506248869001865, + 0.005928288213908672, + 0.008487758226692677, + 0.006162391509860754, + 0.041810326278209686, + 0.025390855967998505, + 0.0018026670441031456, + 0.0020228701177984476, + 0.0029147847089916468, + 0.0023234798572957516, + 0.002447201404720545, + 0.0016458015888929367, + 0.0041099535301327705, + 0.003270477056503296, + 0.0025027324445545673, + 0.003183509921655059, + 0.003348429687321186, + 0.0025295596569776535, + 0.0032826594542711973, + 0.003931224811822176, + 0.012426587752997875, + 0.010723903775215149, + 0.0016542695229873061, + 0.001450283802114427, + 0.0015132171101868153, + 0.0010581840761005878, + 0.0018659058259800076, + 0.00404698820784688, + 0.0038291672244668007, + 0.005560188554227352, + 0.0011223634937778115, + 0.0015996471047401428, + 0.0018594899447634816, + 0.0017386394320055842, + 0.0019294284284114838, + 0.002855837345123291, + 0.008107383735477924, + 0.0027691680006682873, + 0.0033962016459554434, + 0.003439864609390497, + 0.003959716763347387, + 0.003376783337444067, + 0.0033871771302074194, + 0.009402386844158173, + 0.027896221727132797, + 0.008625943213701248, + 0.0020386669784784317, + 0.002314677694812417, + 0.001888988888822496, + 0.0018136865692213178, + 0.001870633801445365, + 0.003596230410039425, + 0.011807340197265148, + 0.005606833379715681, + 0.0014760317280888557, + 0.002481148112565279, + 0.0021910963114351034, + 0.0019068596884608269, + 0.001738743158057332, + 0.0013195628998801112, + 0.01125261653214693, + 0.006630884017795324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.007632025051862001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043056320399045944 + ], + "opposite_logprob_tokens": [ + -3.1680562496185303 + ], + "logprob_reward": -0.043056320399045944, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043056320399045944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5175873637199402, + 0.036129143089056015, + 0.05986173450946808, + 0.04852980747818947, + 0.1452701836824417, + 0.011342136189341545, + 0.022500721737742424, + 0.043084584176540375, + 0.07790160924196243, + 0.03779270872473717 + ], + "credit_map": [ + 0.027562618255615234, + 0.028856869786977768, + 0.03164716437458992, + 0.025457782670855522, + 0.03162568435072899, + 0.06752944737672806, + 0.20313690602779388, + 0.10177087783813477, + 0.002480413531884551, + 0.005132885649800301, + 0.0053956895135343075, + 0.007658316288143396, + 0.004571631085127592, + 0.0035122486297041178, + 0.0033951629884541035, + 0.003982796333730221, + 0.0036430545151233673, + 0.003950012382119894, + 0.005591128952801228, + 0.0042560771107673645, + 0.0045564803294837475, + 0.004171552136540413, + 0.02394489385187626, + 0.009748537093400955, + 0.0028109520208090544, + 0.002915241289883852, + 0.003196859499439597, + 0.0026595224626362324, + 0.0028650015592575073, + 0.006139236968010664, + 0.021510934457182884, + 0.006432065740227699, + 0.008425688371062279, + 0.0078898835927248, + 0.008180668577551842, + 0.0070843505673110485, + 0.0072174035012722015, + 0.019826030358672142, + 0.06495406478643417, + 0.021692106500267982, + 0.001210142858326435, + 0.0011718121822923422, + 0.0015589662361890078, + 0.001369898789562285, + 0.0010295473039150238, + 0.0016903147334232926, + 0.001937018590979278, + 0.001374435960315168, + 0.0010358670260757208, + 0.0014767084503546357, + 0.0014314728323370218, + 0.0015845716698095202, + 0.0016998635837808251, + 0.0017114065121859312, + 0.00860285758972168, + 0.004957975819706917, + 0.002116777468472719, + 0.0023004249669611454, + 0.0026544041465967894, + 0.0023449938744306564, + 0.0022216213401407003, + 0.005141783971339464, + 0.018718751147389412, + 0.007585830520838499, + 0.003934524022042751, + 0.0043645319528877735, + 0.004325828514993191, + 0.003180090570822358, + 0.004543057177215815, + 0.004086165223270655, + 0.03143628314137459, + 0.022031132131814957, + 0.0017550777411088347, + 0.0023666932247579098, + 0.002094362396746874, + 0.0023140371777117252, + 0.0022237400989979506, + 0.0018720775842666626, + 0.015133584849536419, + 0.010033137165009975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.043056320399045944 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979184128344059 + ], + "opposite_logprob_tokens": [ + -3.529791831970215 + ], + "logprob_reward": -0.02979184128344059, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979184128344059, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23070339858531952, + 0.08131572604179382, + 0.0834684818983078, + 0.054528407752513885, + 0.3313283324241638, + 0.0937308520078659, + 0.028462842106819153, + 0.02440076693892479, + 0.046382565051317215, + 0.025678588077425957 + ], + "credit_map": [ + 0.011618083342909813, + 0.020542526617646217, + 0.01186390407383442, + 0.02026580274105072, + 0.012232305482029915, + 0.015498131513595581, + 0.08130321651697159, + 0.05737943947315216, + 0.0046521443873643875, + 0.009453080594539642, + 0.005727620795369148, + 0.01011406909674406, + 0.005967081990092993, + 0.008873977698385715, + 0.017615647986531258, + 0.018912101164460182, + 0.004966517444700003, + 0.014448280446231365, + 0.005115298088639975, + 0.013487151823937893, + 0.005204753018915653, + 0.006235864944756031, + 0.019726304337382317, + 0.014284314587712288, + 0.0034001972526311874, + 0.005613894201815128, + 0.003996103070676327, + 0.0064529068768024445, + 0.0035346911754459143, + 0.0035562191624194384, + 0.014734924770891666, + 0.013239472173154354, + 0.01670672930777073, + 0.020280182361602783, + 0.020421141758561134, + 0.014825762249529362, + 0.018265455961227417, + 0.01871921867132187, + 0.13968294858932495, + 0.08242688328027725, + 0.0048096925020217896, + 0.007776711136102676, + 0.0054633948020637035, + 0.00967550091445446, + 0.005756116472184658, + 0.004828156903386116, + 0.030867550522089005, + 0.024553731083869934, + 0.0014721059706062078, + 0.0016723587177693844, + 0.0016547184204682708, + 0.0015915125841274858, + 0.001667537959292531, + 0.002371931215748191, + 0.011422810144722462, + 0.006609869655221701, + 0.001453878590837121, + 0.0018379458924755454, + 0.001569187268614769, + 0.001635033288039267, + 0.0016842064214870334, + 0.00330530502833426, + 0.005054624751210213, + 0.007860586047172546, + 0.0035166549496352673, + 0.0057242251932621, + 0.004021736327558756, + 0.006660143844783306, + 0.0038006354589015245, + 0.0059355078265070915, + 0.0059083388186991215, + 0.010815318673849106, + 0.002292147371917963, + 0.003962553106248379, + 0.002802488626912236, + 0.0040550013072788715, + 0.0019318177364766598, + 0.002970480127260089, + 0.0049364385195076466, + 0.002727661281824112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02979184128344059 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a young man, reminiscent of Tooo Takaki, with neatly trimmed black hair and a slender build, standing near a rural railway crossing beneath a spectacular meteor shower. The scene is rendered in a style heavily inspired by Van Gogh's \"Starry Night\"; the sky above is a turbulent vortex of deep, swirling blues and fiery yellows, painted with thick, expressive brushstrokes. Bright meteors streak across this celestial canvas, appearing like thick streaks of golden unglazed paint, leaving behind tangible, textured traces. Below, the young man stands in silent awe, his half-profile face looking upward with a look of deep longing. He is at an intersection where two red traffic lights flash rhythmically, casting intermittent crimson pulses across the snow-covered ground. The swirling, textured light from the sky and the flashing signals create a dramatic play, as shadows writhe across the snow, culminating in a scene of intense, emotional beauty and solitary contemplation.", + "seed": 16830, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_1340_seed_16830_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16844619689509271, + "vlm_logprob": -0.16844619689509271, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.16844619689509271, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7623705267906189 + ], + "opposite_logprob_tokens": [ + -0.6373705267906189 + ], + "logprob_reward": -0.7623705267906189, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7623705267906189, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4036838412284851, + 0.13772016763687134, + 0.08134056627750397, + 0.022808730602264404, + 0.023903407156467438, + 0.03634114935994148, + 0.11281300336122513, + 0.08986972272396088, + 0.035842351615428925, + 0.05567706376314163 + ], + "credit_map": [ + 0.09926706552505493, + 0.0398239903151989, + 0.0602930523455143, + 0.03428789973258972, + 0.0330512709915638, + 0.041146304458379745, + 0.061037857085466385, + 0.03477638214826584, + 0.013618677854537964, + 0.03308560699224472, + 0.009857594035565853, + 0.010625675320625305, + 0.025778910145163536, + 0.013832492753863335, + 0.024725640192627907, + 0.006195555906742811, + 0.006714977324008942, + 0.02223462052643299, + 0.005550116766244173, + 0.0039323922246694565, + 0.005895363166928291, + 0.00710042892023921, + 0.026186136528849602, + 0.0037265215069055557, + 0.002516247099265456, + 0.003302358090877533, + 0.0023639940191060305, + 0.0024809252936393023, + 0.0044649820774793625, + 0.0030317995697259903, + 0.0025841814931482077, + 0.0020642413292080164, + 0.0033168334048241377, + 0.00273851677775383, + 0.0031860859598964453, + 0.0014234185218811035, + 0.003504973603412509, + 0.002561550587415695, + 0.003029448911547661, + 0.004142579156905413, + 0.0034728653263300657, + 0.004746816121041775, + 0.0031223518308252096, + 0.0021208850666880608, + 0.004399912431836128, + 0.006020462140440941, + 0.008902648463845253, + 0.003555207047611475, + 0.019477954134345055, + 0.019684582948684692, + 0.014013134874403477, + 0.008419130928814411, + 0.015336956828832626, + 0.015470830723643303, + 0.011445201002061367, + 0.008965211920440197, + 0.01752743311226368, + 0.010995890013873577, + 0.011289647780358791, + 0.006688334513455629, + 0.006633925251662731, + 0.009733735583722591, + 0.015587691217660904, + 0.011413071304559708, + 0.0040522292256355286, + 0.00571952760219574, + 0.006099905353039503, + 0.0035927100107073784, + 0.004238375462591648, + 0.003472123993560672, + 0.005647613666951656, + 0.0030198676977306604, + 0.003254964714869857, + 0.017626967281103134, + 0.0022130690049380064, + 0.00298282690346241, + 0.0032023927196860313, + 0.004290766082704067, + 0.019833285361528397, + 0.002272791462019086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.7623705267906189 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002837918698787689 + ], + "opposite_logprob_tokens": [ + -5.877838134765625 + ], + "logprob_reward": -0.002837918698787689, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002837918698787689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08700045198202133, + 0.06660990417003632, + 0.04580157995223999, + 0.05607298016548157, + 0.018648196011781693, + 0.02139437198638916, + 0.2810785472393036, + 0.017982205376029015, + 0.36305350065231323, + 0.042358286678791046 + ], + "credit_map": [ + 0.021560389548540115, + 0.010717327706515789, + 0.011194905266165733, + 0.006672097370028496, + 0.010569910518825054, + 0.0090148551389575, + 0.009758606553077698, + 0.007512358948588371, + 0.007315959315747023, + 0.013419180177152157, + 0.005002104211598635, + 0.005152439698576927, + 0.010327340103685856, + 0.006644524168223143, + 0.014562579803168774, + 0.004185775760561228, + 0.005058147478848696, + 0.012227420695126057, + 0.0032630865462124348, + 0.002038459526374936, + 0.0038021430373191833, + 0.003795393044129014, + 0.01352494303137064, + 0.0020919879898428917, + 0.00717960437759757, + 0.007975469343364239, + 0.008004631847143173, + 0.004206015728414059, + 0.005693652201443911, + 0.008424940519034863, + 0.008541171438992023, + 0.006047496572136879, + 0.0016589885344728827, + 0.0027072320226579905, + 0.001998001243919134, + 0.001493020448833704, + 0.004530009813606739, + 0.0024147930089384317, + 0.0021303496323525906, + 0.0017158020054921508, + 0.0022224232088774443, + 0.002820150461047888, + 0.0029055948834866285, + 0.0019544323440641165, + 0.002607045928016305, + 0.0035546466242522, + 0.0035517613869160414, + 0.0017783156363293529, + 0.0493755079805851, + 0.05329612269997597, + 0.033845204859972, + 0.021474042907357216, + 0.038562171161174774, + 0.0364389643073082, + 0.02892516739666462, + 0.019161367788910866, + 0.003945349715650082, + 0.00260778795927763, + 0.002502573886886239, + 0.0012227719416841865, + 0.0020160297863185406, + 0.001965300412848592, + 0.0017028385773301125, + 0.0020195532124489546, + 0.03168926760554314, + 0.09388507902622223, + 0.029955025762319565, + 0.02744310535490513, + 0.04052033647894859, + 0.03257704898715019, + 0.08557243645191193, + 0.02141120284795761, + 0.002350481692701578, + 0.015683405101299286, + 0.0013174375053495169, + 0.0014977848622947931, + 0.0023147440515458584, + 0.0033275343012064695, + 0.014494964852929115, + 0.0013719318667426705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.002837918698787689 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254743155092001 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254743155092001, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254743155092001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26750725507736206, + 0.07760655879974365, + 0.1343931257724762, + 0.1007719412446022, + 0.05697960779070854, + 0.05527540668845177, + 0.15669117867946625, + 0.038555800914764404, + 0.049360502511262894, + 0.06285865604877472 + ], + "credit_map": [ + 0.04342645779252052, + 0.03278519958257675, + 0.03595676273107529, + 0.02130788192152977, + 0.042311348021030426, + 0.030459707602858543, + 0.03444577753543854, + 0.026814086362719536, + 0.006540834903717041, + 0.019198108464479446, + 0.006947478279471397, + 0.006146623287349939, + 0.015720615163445473, + 0.007371969521045685, + 0.011061014607548714, + 0.00461991922929883, + 0.010875286534428596, + 0.038715802133083344, + 0.005199599079787731, + 0.0058153001591563225, + 0.010307636111974716, + 0.009159989655017853, + 0.04874144122004509, + 0.005578073672950268, + 0.014660638757050037, + 0.013524556532502174, + 0.01362560410052538, + 0.007535766810178757, + 0.010825550183653831, + 0.01447408925741911, + 0.015426089987158775, + 0.010699647478759289, + 0.005559119395911694, + 0.0082662682980299, + 0.007315994240343571, + 0.004945884924381971, + 0.010345841757953167, + 0.008061540313065052, + 0.0069148424081504345, + 0.005570120178163052, + 0.005487537942826748, + 0.011300574988126755, + 0.0069327522069215775, + 0.005004070233553648, + 0.008343190886080265, + 0.007744951639324427, + 0.005718685220927, + 0.004743640776723623, + 0.01923554390668869, + 0.033030956983566284, + 0.017044004052877426, + 0.013103945180773735, + 0.02411447837948799, + 0.020294466987252235, + 0.019879188388586044, + 0.009988599456846714, + 0.004567823372781277, + 0.006320476066321135, + 0.003572175744920969, + 0.002895491197705269, + 0.008630769327282906, + 0.00511013250797987, + 0.004972220864146948, + 0.002486712532117963, + 0.004123234655708075, + 0.012389601208269596, + 0.004739583469927311, + 0.004322162363678217, + 0.0075882249511778355, + 0.004781951196491718, + 0.007960541173815727, + 0.0034552032593637705, + 0.003962553571909666, + 0.021363653242588043, + 0.002645377069711685, + 0.002975743729621172, + 0.0036051759961992502, + 0.004560405854135752, + 0.021538805216550827, + 0.002206942066550255 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.005254743155092001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033683691173791885 + ], + "opposite_logprob_tokens": [ + -3.4086837768554688 + ], + "logprob_reward": -0.033683691173791885, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033683691173791885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2078157216310501, + 0.159270241856575, + 0.22423817217350006, + 0.05889586731791496, + 0.03777433931827545, + 0.05096447840332985, + 0.05520101636648178, + 0.023393603041768074, + 0.04228772595524788, + 0.14015887677669525 + ], + "credit_map": [ + 0.05037953332066536, + 0.023411516100168228, + 0.025982286781072617, + 0.013699694536626339, + 0.0190140251070261, + 0.022706329822540283, + 0.03276350721716881, + 0.019858816638588905, + 0.015995178371667862, + 0.041485391557216644, + 0.010661149397492409, + 0.010356667451560497, + 0.021351108327507973, + 0.013946718536317348, + 0.038299478590488434, + 0.007174547761678696, + 0.01530566718429327, + 0.0706203281879425, + 0.009458610787987709, + 0.0086468905210495, + 0.011575991287827492, + 0.016247818246483803, + 0.08386936038732529, + 0.008513500913977623, + 0.0066520655527710915, + 0.011541777290403843, + 0.006352744996547699, + 0.004941159393638372, + 0.009509295225143433, + 0.007813304662704468, + 0.007446243427693844, + 0.004639275372028351, + 0.003384180134162307, + 0.005564655177295208, + 0.003802182152867317, + 0.003112846054136753, + 0.008265421725809574, + 0.006018879823386669, + 0.0045688459649682045, + 0.00305732898414135, + 0.005672140512615442, + 0.010094279423356056, + 0.005297981668263674, + 0.003652365645393729, + 0.006902881432324648, + 0.006808487698435783, + 0.0074583375826478004, + 0.005078006070107222, + 0.0068206083960831165, + 0.009920141659677029, + 0.006822304800152779, + 0.004291336517781019, + 0.00825307797640562, + 0.00795410480350256, + 0.0060045113787055016, + 0.005134932231158018, + 0.005789746530354023, + 0.003140696557238698, + 0.0031158304773271084, + 0.002036410616710782, + 0.0025356856640428305, + 0.0024303279351443052, + 0.002427352126687765, + 0.0019175533670932055, + 0.00457072164863348, + 0.01072047371417284, + 0.0025524485390633345, + 0.0019832560792565346, + 0.0023394771851599216, + 0.0034317849203944206, + 0.014968707226216793, + 0.0017208574572578073, + 0.008937755599617958, + 0.04842272773385048, + 0.007282174192368984, + 0.006737036630511284, + 0.007334583904594183, + 0.010441250167787075, + 0.046115268021821976, + 0.004888094495981932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.033683691173791885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808410465717316 + ], + "opposite_logprob_tokens": [ + -3.288084030151367 + ], + "logprob_reward": -0.03808410465717316, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808410465717316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06548776477575302, + 0.12456517666578293, + 0.06000456213951111, + 0.28046274185180664, + 0.04354695975780487, + 0.049127496778964996, + 0.1706840991973877, + 0.03803299739956856, + 0.13023942708969116, + 0.03784877061843872 + ], + "credit_map": [ + 0.008109706453979015, + 0.006664853543043137, + 0.01229063980281353, + 0.007055599708110094, + 0.006316176615655422, + 0.0066871787421405315, + 0.009653553366661072, + 0.008710060268640518, + 0.010961177758872509, + 0.031841207295656204, + 0.0108341034501791, + 0.007848568260669708, + 0.01298646442592144, + 0.01079152524471283, + 0.03361218050122261, + 0.005689950194209814, + 0.008407645858824253, + 0.0052309236489236355, + 0.009195571765303612, + 0.005188892595469952, + 0.0062917242757976055, + 0.01017780788242817, + 0.010405583307147026, + 0.005106410942971706, + 0.04136715084314346, + 0.038696181029081345, + 0.04380286857485771, + 0.024966884404420853, + 0.02955489791929722, + 0.036613401025533676, + 0.034050263464450836, + 0.03141109272837639, + 0.006207302678376436, + 0.005192355718463659, + 0.006385318003594875, + 0.0049408539198338985, + 0.006315113510936499, + 0.004613687749952078, + 0.006621374282985926, + 0.0032709522638469934, + 0.005839583929628134, + 0.008155585266649723, + 0.006652934942394495, + 0.005099313799291849, + 0.006149648688733578, + 0.005925713572651148, + 0.004802591633051634, + 0.006502128206193447, + 0.02040920965373516, + 0.034444212913513184, + 0.019265951588749886, + 0.014396348968148232, + 0.026559999212622643, + 0.023168791085481644, + 0.02041030302643776, + 0.012029295787215233, + 0.006477787159383297, + 0.0034539364278316498, + 0.008100525476038456, + 0.0034431754611432552, + 0.003118553664535284, + 0.003257444826886058, + 0.006820939015597105, + 0.003360637230798602, + 0.01106527354568243, + 0.03319988027215004, + 0.010575530119240284, + 0.010330225341022015, + 0.016737930476665497, + 0.01240875106304884, + 0.027681361883878708, + 0.008240471594035625, + 0.0026993080973625183, + 0.010847344063222408, + 0.002791945356875658, + 0.002638885984197259, + 0.0023909718729555607, + 0.003961888141930103, + 0.010594993829727173, + 0.0019234297797083855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.03808410465717316 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a young man, reminiscent of Tooo Takaki, with neatly trimmed black hair and a slender build, standing near a rural railway crossing beneath a spectacular meteor shower. The scene is rendered in a style heavily inspired by Van Gogh's \"Starry Night\"; the sky above is a turbulent vortex of deep, swirling blues and fiery yellows, painted with thick, expressive brushstrokes. Bright meteors streak across this celestial canvas, appearing like thick streaks of golden unglazed paint, leaving behind tangible, textured traces. Below, the young man stands in silent awe, his half-profile face looking upward with a look of deep longing. He is at an intersection where two red traffic lights flash rhythmically, casting intermittent crimson pulses across the snow-covered ground. The swirling, textured light from the sky and the flashing signals create a dramatic play, as shadows writhe across the snow, culminating in a scene of intense, emotional beauty and solitary contemplation.", + "seed": 16831, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_1340_seed_16831_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.852367346547544, + "vlm_logprob": -0.852367346547544, + "vlm_yes_no_margin": 1.774999976158142, + "vlm_token_credit": -0.852367346547544, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.018554210662842 + ], + "opposite_logprob_tokens": [ + -0.018554221838712692 + ], + "logprob_reward": -4.018554210662842, + "yes_no_margin_reward": -4.0, + "token_credit_reward": -4.018554210662842, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.16408494114875793, + 0.08560710400342941, + 0.18541428446769714, + 0.044018905609846115, + 0.054828524589538574, + 0.11728575080633163, + 0.08944669365882874, + 0.06576213240623474, + 0.08202023059129715, + 0.11153142154216766 + ], + "credit_map": [ + 0.014977722428739071, + 0.01875179633498192, + 0.02156558446586132, + 0.023850418627262115, + 0.015590930357575417, + 0.011291258968412876, + 0.03225575387477875, + 0.025801483541727066, + 0.008763824589550495, + 0.02154819667339325, + 0.007904695346951485, + 0.01444895751774311, + 0.008515547960996628, + 0.0054554506205022335, + 0.010891847312450409, + 0.008078583516180515, + 0.02426702156662941, + 0.022512570023536682, + 0.020714953541755676, + 0.02163655124604702, + 0.01934969425201416, + 0.02627626433968544, + 0.026800179854035378, + 0.023857038468122482, + 0.0035849222913384438, + 0.006207190919667482, + 0.004587959963828325, + 0.004399046301841736, + 0.004884487017989159, + 0.005651292856782675, + 0.008797196671366692, + 0.005906809587031603, + 0.00421124417334795, + 0.006622092332690954, + 0.004434936214238405, + 0.00599651038646698, + 0.006427699234336615, + 0.010023646056652069, + 0.010656422935426235, + 0.006455973722040653, + 0.012747080996632576, + 0.021512392908334732, + 0.0076024592854082584, + 0.01156316976994276, + 0.009035112336277962, + 0.006851265672594309, + 0.0146146509796381, + 0.03335961699485779, + 0.00925981905311346, + 0.011165270581841469, + 0.00900372862815857, + 0.010922888293862343, + 0.017688225954771042, + 0.008375875651836395, + 0.012156379409134388, + 0.010874503292143345, + 0.005566774867475033, + 0.0066226813942193985, + 0.009334297850728035, + 0.007888362742960453, + 0.008401763625442982, + 0.005447677802294493, + 0.014492908492684364, + 0.00800766795873642, + 0.009981696493923664, + 0.012471726164221764, + 0.01162105891853571, + 0.011208917945623398, + 0.008880961686372757, + 0.006633148528635502, + 0.012133163399994373, + 0.00908955093473196, + 0.005935513414442539, + 0.006618391256779432, + 0.041890811175107956, + 0.005877625197172165, + 0.009490260854363441, + 0.006316884886473417, + 0.005694913677871227, + 0.02970702201128006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -4.018554210662842 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016169482842087746 + ], + "opposite_logprob_tokens": [ + -4.141169548034668 + ], + "logprob_reward": -0.016169482842087746, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016169482842087746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11077363789081573, + 0.07497570663690567, + 0.07293278723955154, + 0.01669856160879135, + 0.04679872468113899, + 0.0968502089381218, + 0.08302205055952072, + 0.05427433177828789, + 0.03322882577776909, + 0.41044512391090393 + ], + "credit_map": [ + 0.011141423135995865, + 0.011145948432385921, + 0.012978477403521538, + 0.013745658099651337, + 0.01062540803104639, + 0.014696464873850346, + 0.01962069422006607, + 0.01681957021355629, + 0.0074549647979438305, + 0.007629318628460169, + 0.005901138763874769, + 0.008826125413179398, + 0.014781180769205093, + 0.011607460677623749, + 0.01144582312554121, + 0.007329702842980623, + 0.009841030463576317, + 0.008441074751317501, + 0.008489780128002167, + 0.009018665179610252, + 0.00826816912740469, + 0.010592928156256676, + 0.009028680622577667, + 0.009252466261386871, + 0.0017532019410282373, + 0.0021744025871157646, + 0.0020375095773488283, + 0.00187695084605366, + 0.0033277280163019896, + 0.0016913062427192926, + 0.0020753368735313416, + 0.0017621266888454556, + 0.003955259453505278, + 0.0059532951563596725, + 0.0037627490237355232, + 0.004602004308253527, + 0.006780027411878109, + 0.01022386085242033, + 0.0067899939604103565, + 0.004731536377221346, + 0.010838352143764496, + 0.01811295934021473, + 0.0062689268961548805, + 0.008395289070904255, + 0.006353235337883234, + 0.005269824527204037, + 0.012140976265072823, + 0.029470646753907204, + 0.007824930362403393, + 0.013186761178076267, + 0.007131706923246384, + 0.008464552462100983, + 0.012033923529088497, + 0.006263444200158119, + 0.008537222631275654, + 0.019579516723752022, + 0.005259210243821144, + 0.005152868572622538, + 0.007375775370746851, + 0.006209676619619131, + 0.00835453998297453, + 0.004662992898374796, + 0.010527213104069233, + 0.006732061039656401, + 0.0030861161649227142, + 0.004139194265007973, + 0.004109206143766642, + 0.004803904332220554, + 0.0036392556503415108, + 0.002424791920930147, + 0.006762768141925335, + 0.004263592418283224, + 0.021430186927318573, + 0.02087290957570076, + 0.16060248017311096, + 0.020754115656018257, + 0.03404882550239563, + 0.022212643176317215, + 0.017819548025727272, + 0.11270444095134735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.016169482842087746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01421962771564722 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.01421962771564722, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01421962771564722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11325683444738388, + 0.08295977860689163, + 0.09681472182273865, + 0.031174445524811745, + 0.06823552399873734, + 0.16967590153217316, + 0.10083037614822388, + 0.08733165264129639, + 0.05820639804005623, + 0.19151437282562256 + ], + "credit_map": [ + 0.010549057275056839, + 0.017041243612766266, + 0.014148334041237831, + 0.016156749799847603, + 0.011677060276269913, + 0.007904189638793468, + 0.01947030797600746, + 0.016309892758727074, + 0.008069331757724285, + 0.008971287868916988, + 0.007801726460456848, + 0.01091302465647459, + 0.01594180054962635, + 0.00954329501837492, + 0.014231537468731403, + 0.007487770635634661, + 0.011335779912769794, + 0.0115360664203763, + 0.012951090931892395, + 0.012457340955734253, + 0.00928378477692604, + 0.011441233567893505, + 0.015261245891451836, + 0.012548183090984821, + 0.0024757012724876404, + 0.004497835878282785, + 0.0031827569473534822, + 0.003147113835439086, + 0.004424705635756254, + 0.004832586273550987, + 0.00507508497685194, + 0.0035386609379202127, + 0.005922621116042137, + 0.00995081290602684, + 0.005587760359048843, + 0.006807224825024605, + 0.010747954249382019, + 0.014409069903194904, + 0.00661648903042078, + 0.008193587884306908, + 0.019019711762666702, + 0.030454108491539955, + 0.012133002281188965, + 0.015043934807181358, + 0.011913876980543137, + 0.009874340146780014, + 0.020219745114445686, + 0.05101718381047249, + 0.009095980785787106, + 0.014172167517244816, + 0.009599851444363594, + 0.011836863122880459, + 0.016891034319996834, + 0.008586033247411251, + 0.011086898855865002, + 0.01956154778599739, + 0.0074403551407158375, + 0.007814288139343262, + 0.011727670207619667, + 0.010629232972860336, + 0.013473632745444775, + 0.008450447581708431, + 0.01659385859966278, + 0.01120216678828001, + 0.004936831071972847, + 0.005244756117463112, + 0.005375280044972897, + 0.010312119498848915, + 0.006172039546072483, + 0.0035326199140399694, + 0.015504787676036358, + 0.007127962075173855, + 0.010454102419316769, + 0.011136583052575588, + 0.0683790072798729, + 0.010631070472300053, + 0.019355323165655136, + 0.012186748906970024, + 0.012622641399502754, + 0.046748898923397064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.01421962771564722 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003860175609589 + ], + "opposite_logprob_tokens": [ + -2.6950385570526123 + ], + "logprob_reward": -0.07003860175609589, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003860175609589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12396014481782913, + 0.12795045971870422, + 0.04440799355506897, + 0.023953193798661232, + 0.027083488181233406, + 0.17115220427513123, + 0.08212948590517044, + 0.09563416242599487, + 0.08030233532190323, + 0.22342656552791595 + ], + "credit_map": [ + 0.011599218472838402, + 0.013057416304945946, + 0.015178768895566463, + 0.015082582831382751, + 0.013139772228896618, + 0.01715002954006195, + 0.022592738270759583, + 0.01615961082279682, + 0.011945325881242752, + 0.025127310305833817, + 0.009523788467049599, + 0.015524864196777344, + 0.028546279296278954, + 0.012796624563634396, + 0.01383867859840393, + 0.01064759399741888, + 0.0038147259037941694, + 0.0043311091139912605, + 0.006647370755672455, + 0.005557895638048649, + 0.004350341856479645, + 0.004787381738424301, + 0.00949853379279375, + 0.005420634988695383, + 0.002029532566666603, + 0.004026713781058788, + 0.0026406170800328255, + 0.0028931067790836096, + 0.003998373635113239, + 0.0019859252497553825, + 0.003705331590026617, + 0.0026735940482467413, + 0.0023810791317373514, + 0.003359061200171709, + 0.0029809100087732077, + 0.003425552975386381, + 0.003862776095047593, + 0.0019032165873795748, + 0.005997910164296627, + 0.0031729803886264563, + 0.017434358596801758, + 0.02855795808136463, + 0.011994278989732265, + 0.01559891365468502, + 0.013404947705566883, + 0.009728234261274338, + 0.02628578245639801, + 0.04814772307872772, + 0.007427645847201347, + 0.010582459159195423, + 0.009391938336193562, + 0.011048154905438423, + 0.014974688179790974, + 0.005809321068227291, + 0.009437101893126965, + 0.01345817930996418, + 0.013148929923772812, + 0.01444259099662304, + 0.010227314196527004, + 0.008979838341474533, + 0.019966963678598404, + 0.007308116648346186, + 0.011565846391022205, + 0.009994562715291977, + 0.007571411784738302, + 0.006551518104970455, + 0.009266003035008907, + 0.014282488264143467, + 0.007945556193590164, + 0.0054242913611233234, + 0.02013806812465191, + 0.009122995659708977, + 0.0116038853302598, + 0.0115915946662426, + 0.08394421637058258, + 0.011839867569506168, + 0.021295977756381035, + 0.013168242760002613, + 0.012479040771722794, + 0.05750372260808945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 1 + ], + "selected_reward": -0.07003860175609589 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285480976104736 + ], + "opposite_logprob_tokens": [ + -2.017854690551758 + ], + "logprob_reward": -0.14285480976104736, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14285480976104736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10165408253669739, + 0.08252491801977158, + 0.12049371749162674, + 0.04808421432971954, + 0.047387249767780304, + 0.10007083415985107, + 0.09338241815567017, + 0.10975346714258194, + 0.10413745045661926, + 0.1925116926431656 + ], + "credit_map": [ + 0.007196483202278614, + 0.013408421538770199, + 0.014196811243891716, + 0.012002288363873959, + 0.011701487004756927, + 0.009036688134074211, + 0.020812956616282463, + 0.013298952020704746, + 0.006848094053566456, + 0.01292972918599844, + 0.007730312179774046, + 0.010041645728051662, + 0.01944059133529663, + 0.007893756031990051, + 0.010644963011145592, + 0.006995827425271273, + 0.012972540222108364, + 0.010788272134959698, + 0.016127701848745346, + 0.013073686510324478, + 0.015026944689452648, + 0.017339246347546577, + 0.02206558547914028, + 0.0130997309461236, + 0.004707960411906242, + 0.008411861024796963, + 0.006104364991188049, + 0.005958448629826307, + 0.006450793240219355, + 0.004760063719004393, + 0.006875697989016771, + 0.004815022926777601, + 0.003928238060325384, + 0.006501571275293827, + 0.0033067306503653526, + 0.005303733050823212, + 0.0066328165121376514, + 0.01026671752333641, + 0.0063198888674378395, + 0.005127553828060627, + 0.009388342499732971, + 0.015076340176165104, + 0.010281740687787533, + 0.00909280776977539, + 0.007870836183428764, + 0.0060050832107663155, + 0.017325466498732567, + 0.025030214339494705, + 0.008917341940104961, + 0.01028039213269949, + 0.010967369191348553, + 0.013552389107644558, + 0.02052607201039791, + 0.007330954074859619, + 0.012651169672608376, + 0.009156730957329273, + 0.015076103620231152, + 0.02059006877243519, + 0.009873542934656143, + 0.00960886012762785, + 0.016838839277625084, + 0.012803817167878151, + 0.011203544214367867, + 0.01375869382172823, + 0.008640853688120842, + 0.009400605224072933, + 0.009626691229641438, + 0.018519822508096695, + 0.010736196301877499, + 0.007032386027276516, + 0.027920333668589592, + 0.012260565534234047, + 0.010100755840539932, + 0.009708845987915993, + 0.0742575153708458, + 0.00978927593678236, + 0.016995582729578018, + 0.011066260747611523, + 0.008729358203709126, + 0.05186409875750542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 7 + ], + "selected_reward": -0.14285480976104736 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, low-angle, medium shot captures a young man, reminiscent of Tooo Takaki, with neatly trimmed black hair and a slender build, standing near a rural railway crossing beneath a spectacular meteor shower. The scene is rendered in a style heavily inspired by Van Gogh's \"Starry Night\"; the sky above is a turbulent vortex of deep, swirling blues and fiery yellows, painted with thick, expressive brushstrokes. Bright meteors streak across this celestial canvas, appearing like thick streaks of golden unglazed paint, leaving behind tangible, textured traces. Below, the young man stands in silent awe, his half-profile face looking upward with a look of deep longing. He is at an intersection where two red traffic lights flash rhythmically, casting intermittent crimson pulses across the snow-covered ground. The swirling, textured light from the sky and the flashing signals create a dramatic play, as shadows writhe across the snow, culminating in a scene of intense, emotional beauty and solitary contemplation.", + "seed": 16832, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_12/index_1340_seed_16832_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.6658247018232941, + "vlm_logprob": -0.6658247018232941, + "vlm_yes_no_margin": 1.7, + "vlm_token_credit": -0.6658247018232941, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.3509409427642822 + ], + "opposite_logprob_tokens": [ + -0.10094093531370163 + ], + "logprob_reward": -2.3509409427642822, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.3509409427642822, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20496365427970886, + 0.18353864550590515, + 0.08100032061338425, + 0.15542258322238922, + 0.1798548847436905, + 0.03316250815987587, + 0.05816426873207092, + 0.02931736782193184, + 0.03207095339894295, + 0.04250481724739075 + ], + "credit_map": [ + 0.026399169117212296, + 0.02909204177558422, + 0.0285012349486351, + 0.027723483741283417, + 0.03415388986468315, + 0.03024141676723957, + 0.014437095262110233, + 0.01441531628370285, + 0.02515474148094654, + 0.02547428011894226, + 0.02842625416815281, + 0.03174709528684616, + 0.022882377728819847, + 0.022875092923641205, + 0.014937580563127995, + 0.012041227892041206, + 0.009452441707253456, + 0.014798949472606182, + 0.010954231023788452, + 0.00984608381986618, + 0.008725035935640335, + 0.009964757598936558, + 0.009990098886191845, + 0.0072687165811657906, + 0.0329301655292511, + 0.04189637303352356, + 0.01519983634352684, + 0.014978349208831787, + 0.014640039764344692, + 0.01265061553567648, + 0.010352236218750477, + 0.012774955481290817, + 0.026516960933804512, + 0.05754678696393967, + 0.010550971142947674, + 0.015222741290926933, + 0.023065069690346718, + 0.02563129924237728, + 0.008053540252149105, + 0.013267525471746922, + 0.004902683198451996, + 0.004290421959012747, + 0.005215934477746487, + 0.004728897009044886, + 0.005371738225221634, + 0.0035768088418990374, + 0.0028530596755445004, + 0.002222963608801365, + 0.009703260846436024, + 0.015645239502191544, + 0.006045638117939234, + 0.00460053700953722, + 0.0059780944138765335, + 0.008861945010721684, + 0.0037893468979746103, + 0.003540204605087638, + 0.003888222621753812, + 0.004857480525970459, + 0.0032646136824041605, + 0.003600521944463253, + 0.0052827936597168446, + 0.00284317578189075, + 0.003162260865792632, + 0.0024182989727705717, + 0.0039433082565665245, + 0.005026557948440313, + 0.004235524218529463, + 0.004711198154836893, + 0.004897895269095898, + 0.0038361097685992718, + 0.002681037876754999, + 0.0027393223717808723, + 0.004327789880335331, + 0.004885535687208176, + 0.005714346654713154, + 0.00685614300891757, + 0.00596616929396987, + 0.005532671697437763, + 0.005400192458182573, + 0.003821968799456954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -2.3509409427642822 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01111341081559658 + ], + "opposite_logprob_tokens": [ + -4.51111364364624 + ], + "logprob_reward": -0.01111341081559658, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01111341081559658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12763020396232605, + 0.12176203727722168, + 0.06004424765706062, + 0.18156751990318298, + 0.15618008375167847, + 0.18831248581409454, + 0.032969266176223755, + 0.05442247912287712, + 0.030352449044585228, + 0.046759288758039474 + ], + "credit_map": [ + 0.015947073698043823, + 0.019549190998077393, + 0.018412472680211067, + 0.017007142305374146, + 0.02164110168814659, + 0.017339518293738365, + 0.00880448892712593, + 0.008929195813834667, + 0.016565369442105293, + 0.016039522364735603, + 0.01858500950038433, + 0.02083948440849781, + 0.01608363352715969, + 0.015178314410150051, + 0.010101567953824997, + 0.008369117975234985, + 0.006138137076050043, + 0.007692547980695963, + 0.008303023874759674, + 0.007663059514015913, + 0.007299130782485008, + 0.008192296139895916, + 0.008473577909171581, + 0.006282468792051077, + 0.03766202554106712, + 0.04659855738282204, + 0.01759307272732258, + 0.01718432642519474, + 0.01640147902071476, + 0.01528913900256157, + 0.01262185350060463, + 0.018217036500573158, + 0.023045441135764122, + 0.05009772256016731, + 0.009351423941552639, + 0.013564225286245346, + 0.01982944831252098, + 0.021943362429738045, + 0.00682748993858695, + 0.011520951986312866, + 0.02252405509352684, + 0.020988231524825096, + 0.03142843768000603, + 0.03485717624425888, + 0.029168345034122467, + 0.020113246515393257, + 0.016646284610033035, + 0.01258669700473547, + 0.005299726501107216, + 0.00810194294899702, + 0.004375901538878679, + 0.002581044565886259, + 0.003528256667777896, + 0.004850826691836119, + 0.002221988281235099, + 0.0020095761865377426, + 0.006724193226546049, + 0.008197786286473274, + 0.006838975008577108, + 0.006773608271032572, + 0.008981326594948769, + 0.007218260318040848, + 0.004810679238289595, + 0.004877642262727022, + 0.0039597139693796635, + 0.005753176752477884, + 0.0032611489295959473, + 0.0043943836353719234, + 0.005098152905702591, + 0.0029855736065655947, + 0.002413647249341011, + 0.002486647106707096, + 0.004939475562423468, + 0.0065569267608225346, + 0.006152171175926924, + 0.007515852339565754, + 0.006313638295978308, + 0.006400931626558304, + 0.005098687019199133, + 0.0037816022522747517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 4 + ], + "selected_reward": -0.01111341081559658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638058625161648 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007638058625161648, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638058625161648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20472373068332672, + 0.11920322477817535, + 0.044935863465070724, + 0.08362554013729095, + 0.138808935880661, + 0.3033545911312103, + 0.016238361597061157, + 0.03059919737279415, + 0.027355413883924484, + 0.03115510195493698 + ], + "credit_map": [ + 0.02726149931550026, + 0.03307399898767471, + 0.029830558225512505, + 0.02733294665813446, + 0.03610397130250931, + 0.025832215324044228, + 0.013140044175088406, + 0.012148484587669373, + 0.01610293984413147, + 0.019837575033307076, + 0.016919013112783432, + 0.01805727742612362, + 0.016663143411278725, + 0.014086678624153137, + 0.009691037237644196, + 0.007845556363463402, + 0.005459609441459179, + 0.007781763561069965, + 0.006587681360542774, + 0.005570333916693926, + 0.0054967668838799, + 0.005185867194086313, + 0.0051986961625516415, + 0.003655147971585393, + 0.01784403994679451, + 0.024214910343289375, + 0.007972410880029202, + 0.00831066258251667, + 0.0077905599027872086, + 0.006520769093185663, + 0.005104718264192343, + 0.005867475643754005, + 0.020246749743819237, + 0.04458176717162132, + 0.008077219128608704, + 0.011603510938584805, + 0.01775205321609974, + 0.020045598968863487, + 0.006462445482611656, + 0.010039578191936016, + 0.03553907945752144, + 0.033984072506427765, + 0.049020566046237946, + 0.058323394507169724, + 0.04501667618751526, + 0.03369016945362091, + 0.026995789259672165, + 0.020784856751561165, + 0.0021447965409606695, + 0.0023681819438934326, + 0.0027779133524745703, + 0.0019933173898607492, + 0.002618866739794612, + 0.0018635508604347706, + 0.0013143614633008838, + 0.0011573733063414693, + 0.004220730625092983, + 0.006370610557496548, + 0.003341089701279998, + 0.0035179387778043747, + 0.005034725181758404, + 0.0032320846803486347, + 0.0023318708408623934, + 0.0025501453783363104, + 0.003507058834657073, + 0.0053859250620007515, + 0.0026248956564813852, + 0.004210426472127438, + 0.004301529377698898, + 0.0031764276791363955, + 0.0020530137699097395, + 0.002096137497574091, + 0.0031843287870287895, + 0.0039038017857819796, + 0.0032357466407120228, + 0.004375948570668697, + 0.005088488571345806, + 0.0034190984442830086, + 0.0045651341788470745, + 0.0033825556747615337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.007638058625161648 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578204274177551 + ], + "opposite_logprob_tokens": [ + -0.6328204274177551 + ], + "logprob_reward": -0.7578204274177551, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578204274177551, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2088409662246704, + 0.09699982404708862, + 0.06611360609531403, + 0.049273185431957245, + 0.0946493148803711, + 0.1425349861383438, + 0.04247405007481575, + 0.1611747294664383, + 0.09554305672645569, + 0.04239623248577118 + ], + "credit_map": [ + 0.028935624286532402, + 0.033609356731176376, + 0.02957003377377987, + 0.025102581828832626, + 0.03534713014960289, + 0.02762625552713871, + 0.014961080625653267, + 0.013688907958567142, + 0.012406997382640839, + 0.012388168834149837, + 0.014883767813444138, + 0.016852684319019318, + 0.013224228285253048, + 0.012125273235142231, + 0.009038019925355911, + 0.006080695893615484, + 0.008969143964350224, + 0.013961326330900192, + 0.00776054197922349, + 0.007783679757267237, + 0.008588014170527458, + 0.0077849868685007095, + 0.006621482316404581, + 0.004644435830414295, + 0.008917318657040596, + 0.005799354985356331, + 0.005805243272334337, + 0.005608294624835253, + 0.005871179047971964, + 0.006347694434225559, + 0.00523767014965415, + 0.005686433985829353, + 0.009179423563182354, + 0.020377691835165024, + 0.007768020965158939, + 0.008334184996783733, + 0.008957506157457829, + 0.015058369375765324, + 0.01020252238959074, + 0.014771600253880024, + 0.018471647053956985, + 0.019645417109131813, + 0.02263207733631134, + 0.02373160421848297, + 0.02111894264817238, + 0.015252563171088696, + 0.012021447531878948, + 0.009661298245191574, + 0.005355116445571184, + 0.003859595861285925, + 0.00889568217098713, + 0.006429968401789665, + 0.005626518279314041, + 0.00546689797192812, + 0.003917268011718988, + 0.0029230047948658466, + 0.022801660001277924, + 0.026145735755562782, + 0.02064662240445614, + 0.015810484066605568, + 0.026736896485090256, + 0.01681090146303177, + 0.016388913616538048, + 0.01583351194858551, + 0.01271417923271656, + 0.011129041202366352, + 0.01913955993950367, + 0.011723726987838745, + 0.011301799677312374, + 0.01014096848666668, + 0.009641158394515514, + 0.009752622805535793, + 0.004412200301885605, + 0.005049722734838724, + 0.004273529630154371, + 0.0066524953581392765, + 0.006977347657084465, + 0.004413587972521782, + 0.006192314904183149, + 0.004425037652254105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.7578204274177551 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young man with neatly trimmed black hair and a slender build?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style heavily inspired by Van Gogh's 'Starry Night' with thick, expressive brushstrokes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the sky depicted as a turbulent vortex of deep, swirling blues and fiery yellows with bright meteors?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set at a rural railway crossing with two red traffic lights flashing rhythmically?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static, low-angle, medium shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20161066949367523 + ], + "opposite_logprob_tokens": [ + -1.7016106843948364 + ], + "logprob_reward": -0.20161066949367523, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20161066949367523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0947132557630539, + 0.08979079127311707, + 0.05144573003053665, + 0.11545775085687637, + 0.154810830950737, + 0.059515632688999176, + 0.17902719974517822, + 0.047568581998348236, + 0.039884764701128006, + 0.16778551042079926 + ], + "credit_map": [ + 0.015491207130253315, + 0.022050650790333748, + 0.010494510643184185, + 0.010926274582743645, + 0.012364531867206097, + 0.009193540550768375, + 0.006597708910703659, + 0.00759483128786087, + 0.014152592048048973, + 0.016856933012604713, + 0.013041548430919647, + 0.013979745097458363, + 0.009848475456237793, + 0.010050613433122635, + 0.0063438271172344685, + 0.0055170520208776, + 0.010584100149571896, + 0.011902001686394215, + 0.005236166995018721, + 0.006192018277943134, + 0.006188559345901012, + 0.004176177550107241, + 0.003274470567703247, + 0.0038922333624213934, + 0.021943505853414536, + 0.022369982674717903, + 0.011618494987487793, + 0.01339359488338232, + 0.009803365916013718, + 0.012159289792180061, + 0.009139450266957283, + 0.015030065551400185, + 0.02210867963731289, + 0.04883629456162453, + 0.01323020737618208, + 0.014099473133683205, + 0.01833675056695938, + 0.01959478110074997, + 0.0069089122116565704, + 0.011695749126374722, + 0.011564424261450768, + 0.011839197017252445, + 0.0064989435486495495, + 0.006754650268703699, + 0.008868473581969738, + 0.00590386800467968, + 0.0037300782278180122, + 0.00435600196942687, + 0.030647210776805878, + 0.05483180657029152, + 0.01586441695690155, + 0.012734455987811089, + 0.01600739359855652, + 0.02710244618356228, + 0.01028449460864067, + 0.011554968543350697, + 0.006295972503721714, + 0.007539804093539715, + 0.00615500146523118, + 0.005160778295248747, + 0.007847062312066555, + 0.005453749559819698, + 0.004684139508754015, + 0.0044320751912891865, + 0.004809040110558271, + 0.005575492978096008, + 0.005506062414497137, + 0.005732933524996042, + 0.006322126369923353, + 0.0039309910498559475, + 0.004228250123560429, + 0.0037798667326569557, + 0.017644453793764114, + 0.025742851197719574, + 0.014343347400426865, + 0.029625630006194115, + 0.025596683844923973, + 0.021539388224482536, + 0.018551982939243317, + 0.014741162769496441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.20161066949367523 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_tensor([6829, 6829])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_tensor([6829, 6829])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..859a37542b979f2ef65b8dd1cfe3e90f6b3485e9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_tensor([6829, 6829])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle tracking shot, rendered in stark, high-contrast black and white, follows a massive, seemingly ancient sturgeon as it feeds slowly and powerfully along a murky riverbed. The camera pans steadily, keeping the creature centered to emphasize its immense size. Its body, a relic of a bygone era, is covered in five distinct rows of sharp bony plates that gleam like armor in the faint filtered light from above. The gritty, realistic style captures the scene's pristine textures; every detail is sharp, from the rough, leathery skin between the shields to its small, dark, impassive eyes. As it glides forward, it deliberately drags its sensitive barbs through the sediment, probing the riverbed for food. This movement stirs up a billowing cloud of silt and fine debris that momentarily obscures some of its form before slowly dissipating into darkness. The desaturated color grading, reduced to pure monochrome, heightens the drama, casting deep, impenetrable shadows that accentuate the pristine textures of both the fish and the riverbed, creating a somber and awe-inspiring atmosphere of ancient existence.", + "seed": 25617, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25617_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.061436330527067186, + "vlm_logprob": -0.061436330527067186, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.061436330527067186, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09260493516921997 + ], + "opposite_logprob_tokens": [ + -2.467604875564575 + ], + "logprob_reward": -0.09260493516921997, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09260493516921997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11663869768381119, + 0.07521321624517441, + 0.059956468641757965, + 0.4654143750667572, + 0.059723738580942154, + 0.0512685552239418, + 0.054517243057489395, + 0.03572733700275421, + 0.021861301735043526, + 0.05967912822961807 + ], + "credit_map": [ + 0.015922099351882935, + 0.01654301956295967, + 0.014651975594460964, + 0.013475620187819004, + 0.01298424880951643, + 0.017406217753887177, + 0.012531504966318607, + 0.013124005869030952, + 0.008512195199728012, + 0.011360825039446354, + 0.009074882604181767, + 0.009820487350225449, + 0.010953892022371292, + 0.012243582867085934, + 0.0066921100951731205, + 0.006555238272994757, + 0.007492315489798784, + 0.007982340641319752, + 0.008959798142313957, + 0.008462007157504559, + 0.0058545987121760845, + 0.008163243532180786, + 0.007492328993976116, + 0.005549828056246042, + 0.05961906909942627, + 0.03316522389650345, + 0.1199369803071022, + 0.12975101172924042, + 0.03769555315375328, + 0.02490219473838806, + 0.032398827373981476, + 0.027945462614297867, + 0.01449075248092413, + 0.007530867587774992, + 0.007619598880410194, + 0.007101469207555056, + 0.005415146239101887, + 0.0056539797224104404, + 0.00708905141800642, + 0.004822863731533289, + 0.009457807056605816, + 0.005972323473542929, + 0.006893308833241463, + 0.00598682276904583, + 0.006753912661224604, + 0.00532904826104641, + 0.006824694573879242, + 0.004050633404403925, + 0.006309249438345432, + 0.005958149675279856, + 0.008149021305143833, + 0.008581609465181828, + 0.0060473158955574036, + 0.00444552069529891, + 0.009793899022042751, + 0.0052324747666716576, + 0.0034171573352068663, + 0.006231399718672037, + 0.0032982209231704473, + 0.00425030617043376, + 0.00417168065905571, + 0.0042242384515702724, + 0.005546808708459139, + 0.004587522242218256, + 0.002278732368722558, + 0.0023323700297623873, + 0.002638706238940358, + 0.0035390742123126984, + 0.0025038428138941526, + 0.0023491433821618557, + 0.0035731263924390078, + 0.0026463030371814966, + 0.004348615184426308, + 0.005112989339977503, + 0.016894305124878883, + 0.017722314223647118, + 0.004620402120053768, + 0.003572627902030945, + 0.00462697260081768, + 0.0027808984741568565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.09260493516921997 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02986404299736023 + ], + "opposite_logprob_tokens": [ + -3.5298640727996826 + ], + "logprob_reward": -0.02986404299736023, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02986404299736023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10477467626333237, + 0.10551589727401733, + 0.05367673560976982, + 0.45247069001197815, + 0.022091690450906754, + 0.021217748522758484, + 0.017778020352125168, + 0.04158902168273926, + 0.03930420055985451, + 0.14158132672309875 + ], + "credit_map": [ + 0.011699952185153961, + 0.015240113250911236, + 0.015124701894819736, + 0.01195185910910368, + 0.011644668877124786, + 0.017298173159360886, + 0.00801018625497818, + 0.013805015943944454, + 0.009062955155968666, + 0.011785674840211868, + 0.014010011218488216, + 0.011815996840596199, + 0.015418601222336292, + 0.016999410465359688, + 0.010343372821807861, + 0.01607987843453884, + 0.006619955878704786, + 0.00408512307330966, + 0.010131175629794598, + 0.01018005795776844, + 0.0045420448295772076, + 0.006974328309297562, + 0.007787768263369799, + 0.003356278408318758, + 0.07777658849954605, + 0.04396175965666771, + 0.10485245287418365, + 0.10672392696142197, + 0.03221961855888367, + 0.02799120731651783, + 0.029910461977124214, + 0.029034653678536415, + 0.0034777710679918528, + 0.003901926102116704, + 0.0023645898327231407, + 0.0027762940153479576, + 0.0019743151497095823, + 0.001871048822067678, + 0.002854245249181986, + 0.002871500561013818, + 0.003057740395888686, + 0.0027689810376614332, + 0.0027788581792265177, + 0.003204992040991783, + 0.00205417163670063, + 0.0020628056954592466, + 0.003336564404889941, + 0.0019536351319402456, + 0.0020244414918124676, + 0.002088950015604496, + 0.002602272666990757, + 0.003135537263005972, + 0.0016609152080491185, + 0.0015003158478066325, + 0.0027129100635647774, + 0.0020526768639683723, + 0.005951315630227327, + 0.0038783354684710503, + 0.0039238762110471725, + 0.006999649573117495, + 0.0038599346298724413, + 0.0056255473755300045, + 0.006645959801971912, + 0.004704401828348637, + 0.004960538353770971, + 0.004001243505626917, + 0.004694248549640179, + 0.0045962040312588215, + 0.003946448676288128, + 0.005880583077669144, + 0.005941904149949551, + 0.005283028818666935, + 0.007236528676003218, + 0.008509023115038872, + 0.05156216397881508, + 0.05243339017033577, + 0.006256808061152697, + 0.004265300463885069, + 0.0076274629682302475, + 0.003690652549266815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 1 + ], + "selected_reward": -0.02986404299736023 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704325556755066 + ], + "opposite_logprob_tokens": [ + -3.4087042808532715 + ], + "logprob_reward": -0.033704325556755066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704325556755066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11556649953126907, + 0.1459777057170868, + 0.061784759163856506, + 0.35083404183387756, + 0.03311847150325775, + 0.06519319862127304, + 0.06812926381826401, + 0.04139658808708191, + 0.055546682327985764, + 0.062452834099531174 + ], + "credit_map": [ + 0.01503234077244997, + 0.017595987766981125, + 0.01313613448292017, + 0.01901855506002903, + 0.01064778957515955, + 0.01842724345624447, + 0.012362603098154068, + 0.009345848113298416, + 0.022935694083571434, + 0.017705075442790985, + 0.022507600486278534, + 0.020813535898923874, + 0.014497076161205769, + 0.021801747381687164, + 0.013785525225102901, + 0.011931448243558407, + 0.005680186673998833, + 0.006372738163918257, + 0.011243701912462711, + 0.015362044796347618, + 0.005411513149738312, + 0.0061647542752325535, + 0.00740809878334403, + 0.004141724668443203, + 0.05612637475132942, + 0.02840506099164486, + 0.08401080965995789, + 0.08914195001125336, + 0.027727052569389343, + 0.020273836329579353, + 0.022633245214819908, + 0.02251570299267769, + 0.005156354047358036, + 0.004986924584954977, + 0.0035012036096304655, + 0.005368730053305626, + 0.0033599003218114376, + 0.0034277006052434444, + 0.004268317483365536, + 0.0030493394006043673, + 0.01342556532472372, + 0.008434481918811798, + 0.010059323161840439, + 0.008087393827736378, + 0.006128035485744476, + 0.005187856033444405, + 0.00797152891755104, + 0.0058990102261304855, + 0.006363152991980314, + 0.009426239877939224, + 0.020741306245326996, + 0.007416712585836649, + 0.004948400426656008, + 0.003380474168807268, + 0.010431413538753986, + 0.005421571899205446, + 0.003979365807026625, + 0.005056573543697596, + 0.004076337907463312, + 0.006929784081876278, + 0.005051179323345423, + 0.005481948610395193, + 0.0054488335736095905, + 0.005372563377022743, + 0.00433583976700902, + 0.004057947546243668, + 0.008388550952076912, + 0.011715955100953579, + 0.006356807891279459, + 0.005990414880216122, + 0.009422875009477139, + 0.00527829397469759, + 0.007823172956705093, + 0.004638817626982927, + 0.01358543150126934, + 0.010909897275269032, + 0.005394131410866976, + 0.007931913249194622, + 0.006932351738214493, + 0.0052371216006577015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.033704325556755066 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11292286962270737 + ], + "opposite_logprob_tokens": [ + -2.2379229068756104 + ], + "logprob_reward": -0.11292286962270737, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11292286962270737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0656454786658287, + 0.04813090339303017, + 0.0731947273015976, + 0.25835976004600525, + 0.0538419745862484, + 0.16950161755084991, + 0.03404051810503006, + 0.13678094744682312, + 0.038123719394207, + 0.12238030880689621 + ], + "credit_map": [ + 0.007416417356580496, + 0.009314698167145252, + 0.006850643083453178, + 0.01122811995446682, + 0.006034426856786013, + 0.010122697800397873, + 0.008162522688508034, + 0.006515953689813614, + 0.00578232342377305, + 0.007394232787191868, + 0.006819070782512426, + 0.005852013826370239, + 0.006419111043214798, + 0.00748707028105855, + 0.0044840313494205475, + 0.003893052227795124, + 0.005940387491136789, + 0.006675212178379297, + 0.012818539515137672, + 0.01478639617562294, + 0.006724913138896227, + 0.0077871521934866905, + 0.012979408726096153, + 0.005482728127390146, + 0.026916222646832466, + 0.01704573817551136, + 0.07093305140733719, + 0.079381063580513, + 0.02054610103368759, + 0.01223970577120781, + 0.01750321127474308, + 0.013794698752462864, + 0.008438224904239178, + 0.006953875534236431, + 0.007062598597258329, + 0.007035593036562204, + 0.004727719351649284, + 0.005026320926845074, + 0.008995012380182743, + 0.005602631717920303, + 0.026462187990546227, + 0.02290816605091095, + 0.018319083377718925, + 0.022829869762063026, + 0.021664699539542198, + 0.017181886360049248, + 0.02577131800353527, + 0.014364414848387241, + 0.003639014670625329, + 0.004742543213069439, + 0.006680617108941078, + 0.004710628651082516, + 0.0038202644791454077, + 0.0029593242798000574, + 0.004127159249037504, + 0.003360966220498085, + 0.01802307553589344, + 0.012790585868060589, + 0.013088861480355263, + 0.023530403152108192, + 0.01732638105750084, + 0.018986687064170837, + 0.018890341743826866, + 0.01414461899548769, + 0.007059973664581776, + 0.00515027018263936, + 0.0037145332898944616, + 0.004130464047193527, + 0.0038857434410601854, + 0.0032600946724414825, + 0.004478116519749165, + 0.006444522645324469, + 0.006595526821911335, + 0.005922097712755203, + 0.044657476246356964, + 0.046476345509290695, + 0.0040297070518136024, + 0.00437725055962801, + 0.006472505163401365, + 0.0038494064938277006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 7 + ], + "selected_reward": -0.11292286962270737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808547928929329 + ], + "opposite_logprob_tokens": [ + -3.288085460662842 + ], + "logprob_reward": -0.03808547928929329, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808547928929329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15362855792045593, + 0.22136026620864868, + 0.06974174827337265, + 0.05735611543059349, + 0.07224714756011963, + 0.14414452016353607, + 0.057899489998817444, + 0.06833765655755997, + 0.07113818824291229, + 0.08414634317159653 + ], + "credit_map": [ + 0.02249826490879059, + 0.017573488876223564, + 0.016131257638335228, + 0.02302793227136135, + 0.01901274546980858, + 0.024511441588401794, + 0.015057100914418697, + 0.01581631414592266, + 0.032098203897476196, + 0.029118366539478302, + 0.03035181574523449, + 0.028918607160449028, + 0.02152884751558304, + 0.03165704011917114, + 0.025774547830224037, + 0.021912813186645508, + 0.008211505599319935, + 0.006712927483022213, + 0.011302432976663113, + 0.01373039186000824, + 0.0068983035162091255, + 0.008039548993110657, + 0.009083125740289688, + 0.00576350511983037, + 0.00646436121314764, + 0.007476507686078548, + 0.014465590938925743, + 0.013041170313954353, + 0.005453362129628658, + 0.0034519443288445473, + 0.003810902824625373, + 0.003192269941791892, + 0.013233647681772709, + 0.010288229212164879, + 0.008690871298313141, + 0.011763941496610641, + 0.005953323096036911, + 0.007164744660258293, + 0.009199547581374645, + 0.005952833686023951, + 0.019895028322935104, + 0.02060794271528721, + 0.018062841147184372, + 0.020985910668969154, + 0.014523324556648731, + 0.011860228143632412, + 0.02100251242518425, + 0.017206717282533646, + 0.005676781293004751, + 0.008004848845303059, + 0.013728104531764984, + 0.007059420924633741, + 0.004550266545265913, + 0.0050369068048894405, + 0.008920672349631786, + 0.004922482650727034, + 0.007785703055560589, + 0.008851704187691212, + 0.007075514178723097, + 0.006509693339467049, + 0.008230968378484249, + 0.008144712075591087, + 0.012004549615085125, + 0.009734801948070526, + 0.010382289066910744, + 0.008024567738175392, + 0.009802662767469883, + 0.010911455377936363, + 0.006320442073047161, + 0.008656835183501244, + 0.008068889379501343, + 0.008971043862402439, + 0.00828948151320219, + 0.0052430774085223675, + 0.021587366238236427, + 0.02443486452102661, + 0.004133416805416346, + 0.007398833986371756, + 0.005812428891658783, + 0.007246857509016991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03808547928929329 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, rendered in stark, high-contrast black and white, follows a massive, seemingly ancient sturgeon as it feeds slowly and powerfully along a murky riverbed. The camera pans steadily, keeping the creature centered to emphasize its immense size. Its body, a relic of a bygone era, is covered in five distinct rows of sharp bony plates that gleam like armor in the faint filtered light from above. The gritty, realistic style captures the scene's pristine textures; every detail is sharp, from the rough, leathery skin between the shields to its small, dark, impassive eyes. As it glides forward, it deliberately drags its sensitive barbs through the sediment, probing the riverbed for food. This movement stirs up a billowing cloud of silt and fine debris that momentarily obscures some of its form before slowly dissipating into darkness. The desaturated color grading, reduced to pure monochrome, heightens the drama, casting deep, impenetrable shadows that accentuate the pristine textures of both the fish and the riverbed, creating a somber and awe-inspiring atmosphere of ancient existence.", + "seed": 25618, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25618_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0933072429150343, + "vlm_logprob": -0.0933072429150343, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.0933072429150343, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21478745341300964 + ], + "opposite_logprob_tokens": [ + -1.714787483215332 + ], + "logprob_reward": -0.21478745341300964, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21478745341300964, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11433283984661102, + 0.027106784284114838, + 0.06745073199272156, + 0.029918286949396133, + 0.20791295170783997, + 0.07286785542964935, + 0.014956112951040268, + 0.07136262953281403, + 0.23775885999202728, + 0.15633293986320496 + ], + "credit_map": [ + 0.008926790207624435, + 0.012281397357583046, + 0.017525959759950638, + 0.01904233545064926, + 0.015597794204950333, + 0.007252596318721771, + 0.017230013385415077, + 0.016475951299071312, + 0.002274754224345088, + 0.00301785534247756, + 0.004146167542785406, + 0.0035184097941964865, + 0.00299327215179801, + 0.0020686436910182238, + 0.0036925142630934715, + 0.0053951675072312355, + 0.0061263139359653, + 0.0022654030472040176, + 0.019552605226635933, + 0.020232880488038063, + 0.0017180248396471143, + 0.0019182859687134624, + 0.006132230162620544, + 0.009504986926913261, + 0.0023461519740521908, + 0.002490543993189931, + 0.00501043489202857, + 0.006414550822228193, + 0.002204513642936945, + 0.0016672234050929546, + 0.005683455616235733, + 0.0041014146991074085, + 0.006760997697710991, + 0.007912847213447094, + 0.028286773711442947, + 0.04750973358750343, + 0.009555289521813393, + 0.00567896943539381, + 0.086077019572258, + 0.01613132655620575, + 0.014763648621737957, + 0.008178233169019222, + 0.012827818281948566, + 0.006209704093635082, + 0.007694834843277931, + 0.004552151542156935, + 0.010503125376999378, + 0.008138339966535568, + 0.0012315106578171253, + 0.0013447984820231795, + 0.002393171191215515, + 0.0033024733420461416, + 0.001345219905488193, + 0.001394730294123292, + 0.0023526246659457684, + 0.0015915854601189494, + 0.006515066605061293, + 0.008050051517784595, + 0.008627871982753277, + 0.011875285767018795, + 0.006157503463327885, + 0.010076997801661491, + 0.01064541656523943, + 0.009414431639015675, + 0.019887369126081467, + 0.022064050659537315, + 0.028437364846467972, + 0.036172445863485336, + 0.01695290021598339, + 0.024488799273967743, + 0.040861792862415314, + 0.04889415577054024, + 0.007646847981959581, + 0.012104230932891369, + 0.020547354593873024, + 0.03044651821255684, + 0.01132263708859682, + 0.008945835754275322, + 0.04853914678096771, + 0.01678038015961647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.21478745341300964 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11308332532644272 + ], + "opposite_logprob_tokens": [ + -2.2380833625793457 + ], + "logprob_reward": -0.11308332532644272, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11308332532644272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1559881865978241, + 0.048045698553323746, + 0.05110359191894531, + 0.3648684322834015, + 0.07757686823606491, + 0.0758010521531105, + 0.05044475197792053, + 0.10831521451473236, + 0.02858109027147293, + 0.0392751507461071 + ], + "credit_map": [ + 0.01076461561024189, + 0.02468183822929859, + 0.017972588539123535, + 0.03074948862195015, + 0.024001458659768105, + 0.008872639387845993, + 0.020800581201910973, + 0.018144972622394562, + 0.0035382574424147606, + 0.005459143780171871, + 0.006207289174199104, + 0.00693173985928297, + 0.008870572783052921, + 0.004406385123729706, + 0.006666205823421478, + 0.005966104567050934, + 0.008080963045358658, + 0.004584034904837608, + 0.007155193015933037, + 0.007946323603391647, + 0.003287676488980651, + 0.002367533976212144, + 0.009174889884889126, + 0.008506976999342442, + 0.011451744474470615, + 0.006088691297918558, + 0.1512618362903595, + 0.022106816992163658, + 0.00812501460313797, + 0.008608449250459671, + 0.14287075400352478, + 0.014355108141899109, + 0.00297835492528975, + 0.003641135059297085, + 0.005297579802572727, + 0.024892618879675865, + 0.0033469225745648146, + 0.002185048535466194, + 0.02991293929517269, + 0.005322272423654795, + 0.012584664858877659, + 0.007821178995072842, + 0.014025253243744373, + 0.006930089555680752, + 0.008621342480182648, + 0.005249395500868559, + 0.010932227596640587, + 0.009636896662414074, + 0.004933352582156658, + 0.004348325543105602, + 0.005609102081507444, + 0.0070562115870416164, + 0.011389090679585934, + 0.003127851290628314, + 0.005849829874932766, + 0.008130986243486404, + 0.007339960895478725, + 0.014804407022893429, + 0.013753756880760193, + 0.019821615889668465, + 0.01129915751516819, + 0.010862912982702255, + 0.014950820244848728, + 0.015482578426599503, + 0.0025236953515559435, + 0.0030133388936519623, + 0.0034156758338212967, + 0.0036770275328308344, + 0.002278406172990799, + 0.0027746371924877167, + 0.005013247486203909, + 0.005885060410946608, + 0.0026130203623324633, + 0.003636818379163742, + 0.0052524046041071415, + 0.006883962545543909, + 0.002766828751191497, + 0.0033900889102369547, + 0.010228227823972702, + 0.0045038009993731976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.11308332532644272 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635827101767063 + ], + "opposite_logprob_tokens": [ + -3.6513583660125732 + ], + "logprob_reward": -0.02635827101767063, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635827101767063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12516719102859497, + 0.08829297125339508, + 0.11407430469989777, + 0.11917071044445038, + 0.16266247630119324, + 0.06447195261716843, + 0.04736682027578354, + 0.10761885344982147, + 0.055169325321912766, + 0.11600537598133087 + ], + "credit_map": [ + 0.01001670677214861, + 0.021184349432587624, + 0.013743950985372066, + 0.01999749429523945, + 0.022415973246097565, + 0.008114292286336422, + 0.01375015638768673, + 0.01594426855444908, + 0.00856644194573164, + 0.01423655916005373, + 0.010090373456478119, + 0.00963969063013792, + 0.018477536737918854, + 0.005619632545858622, + 0.01046302542090416, + 0.011199715547263622, + 0.010869123972952366, + 0.009694482199847698, + 0.02071354351937771, + 0.024255815893411636, + 0.012793959118425846, + 0.006227076053619385, + 0.012230795808136463, + 0.017289509996771812, + 0.004626790061593056, + 0.003699444467201829, + 0.041179969906806946, + 0.009962642565369606, + 0.00398382730782032, + 0.0040775747038424015, + 0.042226292192935944, + 0.009414175525307655, + 0.005707155913114548, + 0.006776120979338884, + 0.010711904615163803, + 0.053068697452545166, + 0.0056125991977751255, + 0.0032997610978782177, + 0.06552506983280182, + 0.011961180716753006, + 0.007974415086209774, + 0.0079823462292552, + 0.012463834136724472, + 0.007718763779848814, + 0.0077659497037529945, + 0.0041283732280135155, + 0.00908756721764803, + 0.00735070463269949, + 0.003457362065091729, + 0.004764172248542309, + 0.007311823312193155, + 0.008217941038310528, + 0.00638640858232975, + 0.002506971126422286, + 0.006989896297454834, + 0.0077322497963905334, + 0.01074378564953804, + 0.015147162601351738, + 0.012494485825300217, + 0.019840963184833527, + 0.009135511703789234, + 0.010954523459076881, + 0.01366338413208723, + 0.0156390480697155, + 0.003870793152600527, + 0.006807237863540649, + 0.006577035877853632, + 0.00961182452738285, + 0.004174044355750084, + 0.004031892865896225, + 0.009201975539326668, + 0.010894526727497578, + 0.009458628483116627, + 0.010195079259574413, + 0.016013195738196373, + 0.01819760911166668, + 0.010445650666952133, + 0.013929398730397224, + 0.022919900715351105, + 0.01484593190252781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.02635827101767063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902386575937271 + ], + "opposite_logprob_tokens": [ + -2.4640238285064697 + ], + "logprob_reward": -0.08902386575937271, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902386575937271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07846906036138535, + 0.05433407053351402, + 0.06861844658851624, + 0.3991222679615021, + 0.0471915639936924, + 0.025612035766243935, + 0.04094612970948219, + 0.019423600286245346, + 0.16788190603256226, + 0.09840094298124313 + ], + "credit_map": [ + 0.005731822457164526, + 0.009667523205280304, + 0.011180108413100243, + 0.013208442367613316, + 0.011262491345405579, + 0.00487174978479743, + 0.011651621200144291, + 0.01089530810713768, + 0.005421056412160397, + 0.00559257110580802, + 0.00756533769890666, + 0.008073108270764351, + 0.004375240299850702, + 0.0032619682606309652, + 0.00907974224537611, + 0.010965043678879738, + 0.006268832832574844, + 0.004280843771994114, + 0.012778491713106632, + 0.01390065811574459, + 0.007269237190485001, + 0.0030499743297696114, + 0.007603451143950224, + 0.013466957025229931, + 0.013374797999858856, + 0.006721566431224346, + 0.16400916874408722, + 0.02330232970416546, + 0.008718477562069893, + 0.009325681254267693, + 0.15743772685527802, + 0.016232529655098915, + 0.0017268072115257382, + 0.0018581915646791458, + 0.01597396284341812, + 0.0071222223341465, + 0.0021827113814651966, + 0.0012136321747675538, + 0.013602695427834988, + 0.003511340357363224, + 0.0020170495845377445, + 0.0021959254518151283, + 0.006986475083976984, + 0.0032872583251446486, + 0.0022833533585071564, + 0.0016197452787309885, + 0.003989503253251314, + 0.0032327244989573956, + 0.003965238109230995, + 0.005028129555284977, + 0.0056382124312222, + 0.004946356639266014, + 0.009061748161911964, + 0.0021345883142203093, + 0.00551890442147851, + 0.004652949515730143, + 0.0017956045921891928, + 0.002613210817798972, + 0.0024468512274324894, + 0.0027416213415563107, + 0.0022790483199059963, + 0.0018967551877722144, + 0.003221507417038083, + 0.002429000334814191, + 0.01348938699811697, + 0.014555355533957481, + 0.020433565601706505, + 0.025277677923440933, + 0.012471972033381462, + 0.01841196045279503, + 0.02902653068304062, + 0.03421545773744583, + 0.00844536628574133, + 0.006944009568542242, + 0.012802023440599442, + 0.018384037539362907, + 0.0072382655926048756, + 0.008978504687547684, + 0.023795176297426224, + 0.011813560500741005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.08902386575937271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283299058675766 + ], + "opposite_logprob_tokens": [ + -3.7732832431793213 + ], + "logprob_reward": -0.023283299058675766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283299058675766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11606601625680923, + 0.08848091959953308, + 0.04432792216539383, + 0.37844279408454895, + 0.07007292658090591, + 0.06746619939804077, + 0.026056980714201927, + 0.09075649082660675, + 0.07048236578702927, + 0.04784735292196274 + ], + "credit_map": [ + 0.010199322365224361, + 0.012082798406481743, + 0.01482009794563055, + 0.019061235710978508, + 0.01375453919172287, + 0.006727447267621756, + 0.01755211129784584, + 0.021868467330932617, + 0.006587891839444637, + 0.011003533378243446, + 0.011131249368190765, + 0.012327639386057854, + 0.017151568084955215, + 0.007753162179142237, + 0.011101365089416504, + 0.01142450887709856, + 0.005910217761993408, + 0.004263109527528286, + 0.006599940825253725, + 0.005999429151415825, + 0.005216704215854406, + 0.0021392195485532284, + 0.006622427608817816, + 0.007576870732009411, + 0.012222850695252419, + 0.007156295236200094, + 0.156440868973732, + 0.025162262842059135, + 0.00857757031917572, + 0.009016687981784344, + 0.1452854871749878, + 0.014580761082470417, + 0.0045468066819012165, + 0.0035033246967941523, + 0.00970040075480938, + 0.020800985395908356, + 0.003966169897466898, + 0.0020975188817828894, + 0.018446937203407288, + 0.007010787259787321, + 0.005144529975950718, + 0.004457093309611082, + 0.020538784563541412, + 0.011205479502677917, + 0.005646625533699989, + 0.0024946925695985556, + 0.011204227805137634, + 0.006774769630283117, + 0.002673954702913761, + 0.0018686369294300675, + 0.003898265305906534, + 0.00543334661051631, + 0.003219969803467393, + 0.0018679248169064522, + 0.0037753921933472157, + 0.0033194907009601593, + 0.007647466845810413, + 0.008313043974339962, + 0.011325297877192497, + 0.014358360320329666, + 0.009718487039208412, + 0.012524339370429516, + 0.01236018631607294, + 0.014509309083223343, + 0.0056938063353300095, + 0.0071104951202869415, + 0.008292745798826218, + 0.011231089942157269, + 0.004690609872341156, + 0.005848433822393417, + 0.011776653118431568, + 0.01583852805197239, + 0.0041001345962285995, + 0.0053544240072369576, + 0.008465450257062912, + 0.00796547345817089, + 0.004133189562708139, + 0.004054671619087458, + 0.008476918563246727, + 0.005297091789543629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.023283299058675766 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, rendered in stark, high-contrast black and white, follows a massive, seemingly ancient sturgeon as it feeds slowly and powerfully along a murky riverbed. The camera pans steadily, keeping the creature centered to emphasize its immense size. Its body, a relic of a bygone era, is covered in five distinct rows of sharp bony plates that gleam like armor in the faint filtered light from above. The gritty, realistic style captures the scene's pristine textures; every detail is sharp, from the rough, leathery skin between the shields to its small, dark, impassive eyes. As it glides forward, it deliberately drags its sensitive barbs through the sediment, probing the riverbed for food. This movement stirs up a billowing cloud of silt and fine debris that momentarily obscures some of its form before slowly dissipating into darkness. The desaturated color grading, reduced to pure monochrome, heightens the drama, casting deep, impenetrable shadows that accentuate the pristine textures of both the fish and the riverbed, creating a somber and awe-inspiring atmosphere of ancient existence.", + "seed": 25619, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25619_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0813782537356019, + "vlm_logprob": -0.0813782537356019, + "vlm_yes_no_margin": 2.825, + "vlm_token_credit": -0.0813782537356019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06612852960824966 + ], + "opposite_logprob_tokens": [ + -2.8161284923553467 + ], + "logprob_reward": -0.06612852960824966, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06612852960824966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13850589096546173, + 0.08903913199901581, + 0.02987673506140709, + 0.12331326305866241, + 0.06865894794464111, + 0.1199449971318245, + 0.05028263479471207, + 0.06585872173309326, + 0.14233459532260895, + 0.17218510806560516 + ], + "credit_map": [ + 0.019054675474762917, + 0.022862698882818222, + 0.015095497481524944, + 0.013395079411566257, + 0.01998816803097725, + 0.02173512801527977, + 0.01337115652859211, + 0.013003495521843433, + 0.007145741954445839, + 0.013128370977938175, + 0.016839904710650444, + 0.01616940088570118, + 0.013831852003932, + 0.006932058371603489, + 0.0067696887999773026, + 0.00822211429476738, + 0.0031154961325228214, + 0.0040603033266961575, + 0.004220748785883188, + 0.006713526789098978, + 0.00543412659317255, + 0.0026444457471370697, + 0.002036233199760318, + 0.0016518555348739028, + 0.019163167104125023, + 0.020662304013967514, + 0.01503349281847477, + 0.017621302977204323, + 0.012533721514046192, + 0.01863013207912445, + 0.011634393595159054, + 0.008034750819206238, + 0.007546281907707453, + 0.006501651834696531, + 0.007584502920508385, + 0.012650027871131897, + 0.008870328776538372, + 0.009133945219218731, + 0.007206043228507042, + 0.009166169911623001, + 0.006175155285745859, + 0.04732873663306236, + 0.04386521503329277, + 0.00486428989097476, + 0.0064349062740802765, + 0.004687842447310686, + 0.0034033695701509714, + 0.0031854743137955666, + 0.007347129285335541, + 0.00805729441344738, + 0.007814022712409496, + 0.005553216673433781, + 0.007064097095280886, + 0.006106751039624214, + 0.003493993077427149, + 0.004846128169447184, + 0.008623654954135418, + 0.008926499634981155, + 0.009850881062448025, + 0.0086299953982234, + 0.006731483153998852, + 0.01050413865596056, + 0.00549536244943738, + 0.007096706423908472, + 0.01653216779232025, + 0.016007138416171074, + 0.016561532393097878, + 0.027759507298469543, + 0.016742408275604248, + 0.01387745700776577, + 0.01869438774883747, + 0.01616000384092331, + 0.01817789301276207, + 0.01621549390256405, + 0.01663096435368061, + 0.023545263335108757, + 0.02197076566517353, + 0.02191348373889923, + 0.021560681983828545, + 0.03217056393623352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.06612852960824966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300165951251984 + ], + "opposite_logprob_tokens": [ + -2.238001585006714 + ], + "logprob_reward": -0.11300165951251984, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300165951251984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292041301727295, + 0.2375382035970688, + 0.09249669313430786, + 0.05474691838026047, + 0.03672761097550392, + 0.06793874502182007, + 0.04016390070319176, + 0.12793023884296417, + 0.0458272248506546, + 0.19371002912521362 + ], + "credit_map": [ + 0.014683588407933712, + 0.01885983534157276, + 0.01905590109527111, + 0.010291569866240025, + 0.008912831544876099, + 0.015421838499605656, + 0.007947332225739956, + 0.00774751789867878, + 0.018440043553709984, + 0.03364383429288864, + 0.05081354081630707, + 0.04601360484957695, + 0.03814540430903435, + 0.01879573054611683, + 0.013636378571391106, + 0.018049679696559906, + 0.012220359407365322, + 0.016251953318715096, + 0.010002213530242443, + 0.014543361961841583, + 0.014629003591835499, + 0.011364390142261982, + 0.007739343214780092, + 0.0057460772804915905, + 0.007065700367093086, + 0.008914791978895664, + 0.006246025674045086, + 0.0077699655666947365, + 0.0074900612235069275, + 0.008445401675999165, + 0.005500108934938908, + 0.003314869711175561, + 0.0033164420165121555, + 0.003910640720278025, + 0.004395095631480217, + 0.0033626712393015623, + 0.005733628291636705, + 0.005476424470543861, + 0.004159450996667147, + 0.006373259238898754, + 0.0041961343958973885, + 0.0266717579215765, + 0.024386391043663025, + 0.002858323510736227, + 0.0031268971506506205, + 0.0024188151583075523, + 0.001833659945987165, + 0.0024467739276587963, + 0.004965073894709349, + 0.006970809772610664, + 0.005482115782797337, + 0.0051066759042441845, + 0.005000413861125708, + 0.004353065975010395, + 0.00527240801602602, + 0.0030133374966681004, + 0.01607762835919857, + 0.013989298604428768, + 0.02412894181907177, + 0.022824443876743317, + 0.013950735330581665, + 0.01697061024606228, + 0.010264411568641663, + 0.009724176488816738, + 0.0051345848478376865, + 0.004632856696844101, + 0.0053908308036625385, + 0.0073121823370456696, + 0.006171340588480234, + 0.006827474106103182, + 0.005275229457765818, + 0.005082728341221809, + 0.01929854415357113, + 0.01680913195014, + 0.022102264687418938, + 0.024643879383802414, + 0.03272445872426033, + 0.02220565266907215, + 0.03697444126009941, + 0.018951673060655594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.11300165951251984 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337083600461483 + ], + "opposite_logprob_tokens": [ + -3.408708333969116 + ], + "logprob_reward": -0.0337083600461483, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337083600461483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06237843632698059, + 0.040765076875686646, + 0.02677132934331894, + 0.01863919571042061, + 0.042942147701978683, + 0.6617538928985596, + 0.04827959090471268, + 0.030028851702809334, + 0.04153180867433548, + 0.026909681037068367 + ], + "credit_map": [ + 0.008745457045733929, + 0.010269473306834698, + 0.008662153966724873, + 0.00682082399725914, + 0.007175749633461237, + 0.01035486813634634, + 0.003983554895967245, + 0.006366348825395107, + 0.0032164761796593666, + 0.0067195710726082325, + 0.008557830937206745, + 0.004404573701322079, + 0.006860574707388878, + 0.005852029658854008, + 0.002538192318752408, + 0.0026158257387578487, + 0.0030970319639891386, + 0.004559145774692297, + 0.0036636032164096832, + 0.0047996328212320805, + 0.003998312633484602, + 0.0026528448797762394, + 0.002114478498697281, + 0.0018862762954086065, + 0.002749415347352624, + 0.002326065208762884, + 0.0025426994543522596, + 0.0028854459524154663, + 0.002824709052219987, + 0.0024040730204433203, + 0.0015999614261090755, + 0.001306826015934348, + 0.0042531671933829784, + 0.006767214275896549, + 0.005082538817077875, + 0.007074302993714809, + 0.005913757719099522, + 0.005240944679826498, + 0.003982815891504288, + 0.0046274010092020035, + 0.03241276741027832, + 0.25646543502807617, + 0.25516244769096375, + 0.024354174733161926, + 0.028381381183862686, + 0.024589110165834427, + 0.01674201712012291, + 0.023646501824259758, + 0.005637377500534058, + 0.005649047438055277, + 0.007521267049014568, + 0.0068717519752681255, + 0.0059148939326405525, + 0.005782696418464184, + 0.006633100099861622, + 0.004269453696906567, + 0.003605803707614541, + 0.0040382929146289825, + 0.00377945345826447, + 0.005285191349685192, + 0.0031207362189888954, + 0.005228748079389334, + 0.002235445426777005, + 0.002735178917646408, + 0.005116655025631189, + 0.005494482349604368, + 0.004990068264305592, + 0.0078091248869895935, + 0.004209564067423344, + 0.0037819920107722282, + 0.005286009982228279, + 0.004843906499445438, + 0.002532342215999961, + 0.0023794081062078476, + 0.0026245799381285906, + 0.003265489125624299, + 0.0038733764085918665, + 0.0040643298998475075, + 0.003421685891225934, + 0.004748465493321419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.0337083600461483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986376583576202 + ], + "opposite_logprob_tokens": [ + -1.8048638105392456 + ], + "logprob_reward": -0.17986376583576202, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986376583576202, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05667087435722351, + 0.09903328865766525, + 0.046753060072660446, + 0.06906212866306305, + 0.028049729764461517, + 0.5301076173782349, + 0.044924937188625336, + 0.04060332104563713, + 0.036337852478027344, + 0.04845717176795006 + ], + "credit_map": [ + 0.007949911057949066, + 0.009703869000077248, + 0.007656294386833906, + 0.007185920607298613, + 0.0075464146211743355, + 0.006405347492545843, + 0.00445821275934577, + 0.0057649072259664536, + 0.008388861082494259, + 0.018251245841383934, + 0.021148528903722763, + 0.012580620124936104, + 0.016355611383914948, + 0.01065804623067379, + 0.005983052775263786, + 0.005667326971888542, + 0.004590632859617472, + 0.0055456990376114845, + 0.00506981136277318, + 0.008264543488621712, + 0.007312135770916939, + 0.007133983541280031, + 0.0045659481547772884, + 0.0042703039944171906, + 0.010165848769247532, + 0.009135578759014606, + 0.007414152845740318, + 0.01007589790970087, + 0.010180460289120674, + 0.00996463093906641, + 0.006679261103272438, + 0.005446299910545349, + 0.004583216272294521, + 0.002723118755966425, + 0.0035163741558790207, + 0.003216365585103631, + 0.0026693488471210003, + 0.0034732758067548275, + 0.002728405874222517, + 0.0051396251656115055, + 0.026386942714452744, + 0.2022790163755417, + 0.20562845468521118, + 0.021471461281180382, + 0.022553307935595512, + 0.01783846691250801, + 0.012984424829483032, + 0.020965557545423508, + 0.0047387112863361835, + 0.005711931269615889, + 0.005536334589123726, + 0.007841351442039013, + 0.006048192735761404, + 0.005082251969724894, + 0.005744077265262604, + 0.004222088027745485, + 0.0051170033402740955, + 0.004198299255222082, + 0.0059691485948860645, + 0.007218104787170887, + 0.004111589398235083, + 0.005859629716724157, + 0.004240167327225208, + 0.0038893793243914843, + 0.004231440369039774, + 0.0033493810333311558, + 0.0038703291211277246, + 0.0056382156908512115, + 0.004413141869008541, + 0.004081187769770622, + 0.005989338271319866, + 0.004764816723763943, + 0.00457953242585063, + 0.004303777124732733, + 0.005172125995159149, + 0.008799489587545395, + 0.005311124958097935, + 0.004758779425173998, + 0.008059552870690823, + 0.007472790312021971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 3 + ], + "selected_reward": -0.17986376583576202 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014188953675329685 + ], + "opposite_logprob_tokens": [ + -4.264188766479492 + ], + "logprob_reward": -0.014188953675329685, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014188953675329685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04033688083291054, + 0.04130030795931816, + 0.04894140735268593, + 0.018207605928182602, + 0.06522705405950546, + 0.7211741209030151, + 0.015777617692947388, + 0.019582783803343773, + 0.013737802393734455, + 0.015714360401034355 + ], + "credit_map": [ + 0.005456557497382164, + 0.0077262865379452705, + 0.007214738056063652, + 0.004441630095243454, + 0.004171460401266813, + 0.004762374330312014, + 0.0032404172234237194, + 0.0033234236761927605, + 0.00392526900395751, + 0.007137349806725979, + 0.009846044704318047, + 0.0059610833413898945, + 0.0045356061309576035, + 0.004000038839876652, + 0.0026825019158422947, + 0.0032124193385243416, + 0.00620977533981204, + 0.0072926925495266914, + 0.0063890996389091015, + 0.00897789653390646, + 0.007029567379504442, + 0.005314325448125601, + 0.004104329273104668, + 0.0036237279418855906, + 0.0021317664068192244, + 0.002464665099978447, + 0.001854713074862957, + 0.0029719441663473845, + 0.002272734884172678, + 0.0029500043019652367, + 0.0021180249750614166, + 0.001443755580112338, + 0.007349970750510693, + 0.007792973425239325, + 0.008175144903361797, + 0.009168625809252262, + 0.007509537506848574, + 0.00857703946530819, + 0.00654161162674427, + 0.01011215802282095, + 0.035747967660427094, + 0.2881462574005127, + 0.28203341364860535, + 0.02489914931356907, + 0.031756483018398285, + 0.022015951573848724, + 0.01584826596081257, + 0.02072673663496971, + 0.0019884626381099224, + 0.0018660224741324782, + 0.0019362285966053605, + 0.001672644866630435, + 0.0026840243954211473, + 0.002130493288859725, + 0.0017568358452990651, + 0.0017429081490263343, + 0.0020790172275155783, + 0.0023009979631751776, + 0.002606486203148961, + 0.003317938419058919, + 0.002358413301408291, + 0.0027814414352178574, + 0.0014056708896532655, + 0.0027328231371939182, + 0.0016552518354728818, + 0.0017576103564351797, + 0.001631497754715383, + 0.0020180705469101667, + 0.0015882570296525955, + 0.0016623837873339653, + 0.001680255401879549, + 0.0017444785917177796, + 0.0016144639812409878, + 0.0012836045352742076, + 0.0020547849126160145, + 0.002744216937571764, + 0.002091684378683567, + 0.0013602483086287975, + 0.002818284323439002, + 0.001747076166793704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.014188953675329685 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle tracking shot, rendered in stark, high-contrast black and white, follows a massive, seemingly ancient sturgeon as it feeds slowly and powerfully along a murky riverbed. The camera pans steadily, keeping the creature centered to emphasize its immense size. Its body, a relic of a bygone era, is covered in five distinct rows of sharp bony plates that gleam like armor in the faint filtered light from above. The gritty, realistic style captures the scene's pristine textures; every detail is sharp, from the rough, leathery skin between the shields to its small, dark, impassive eyes. As it glides forward, it deliberately drags its sensitive barbs through the sediment, probing the riverbed for food. This movement stirs up a billowing cloud of silt and fine debris that momentarily obscures some of its form before slowly dissipating into darkness. The desaturated color grading, reduced to pure monochrome, heightens the drama, casting deep, impenetrable shadows that accentuate the pristine textures of both the fish and the riverbed, creating a somber and awe-inspiring atmosphere of ancient existence.", + "seed": 25620, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_13/index_6829_seed_25620_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05225931853055954, + "vlm_logprob": -0.05225931853055954, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.05225931853055954, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0741397887468338 + ], + "opposite_logprob_tokens": [ + -2.6991398334503174 + ], + "logprob_reward": -0.0741397887468338, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0741397887468338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16774672269821167, + 0.20774592459201813, + 0.05544368177652359, + 0.05722976475954056, + 0.06780370324850082, + 0.06822264194488525, + 0.07784948498010635, + 0.05057234317064285, + 0.12000000476837158, + 0.12738575041294098 + ], + "credit_map": [ + 0.01988997310400009, + 0.021586721763014793, + 0.02211849018931389, + 0.020396003499627113, + 0.016237502917647362, + 0.01973540149629116, + 0.020496224984526634, + 0.027286401018500328, + 0.02923392318189144, + 0.029285335913300514, + 0.02414577826857567, + 0.033406395465135574, + 0.02934611774981022, + 0.010097388178110123, + 0.03020162507891655, + 0.022029349580407143, + 0.004307983908802271, + 0.0035969724413007498, + 0.012646607123315334, + 0.0057939100079238415, + 0.0121809346601367, + 0.0033831957262009382, + 0.005372900515794754, + 0.008161178790032864, + 0.006026681046932936, + 0.006146745290607214, + 0.008189577609300613, + 0.009706070646643639, + 0.005224732216447592, + 0.005937292706221342, + 0.006050677038729191, + 0.009947988204658031, + 0.008155603893101215, + 0.00948942918330431, + 0.011237883940339088, + 0.010016347281634808, + 0.006623629946261644, + 0.005464964080601931, + 0.006300023756921291, + 0.010515823028981686, + 0.006208981852978468, + 0.009221162647008896, + 0.009218611754477024, + 0.00827980600297451, + 0.006238614208996296, + 0.0075257583521306515, + 0.00848389696329832, + 0.013045812025666237, + 0.009117878042161465, + 0.007047838997095823, + 0.011080323718488216, + 0.008166979067027569, + 0.007425862364470959, + 0.011621182784438133, + 0.01232839934527874, + 0.01106102205812931, + 0.003247455693781376, + 0.002880680374801159, + 0.004528319928795099, + 0.003616607515141368, + 0.003959307912737131, + 0.002984369406476617, + 0.014359276741743088, + 0.014996322803199291, + 0.009878849610686302, + 0.007887589745223522, + 0.015510627999901772, + 0.01359289325773716, + 0.010216166265308857, + 0.006641536019742489, + 0.028319239616394043, + 0.027953099459409714, + 0.010203056037425995, + 0.028961872681975365, + 0.014327870681881905, + 0.015172101557254791, + 0.011837300844490528, + 0.012222109362483025, + 0.019037602469325066, + 0.015623828396201134 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.0741397887468338 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029852934181690216 + ], + "opposite_logprob_tokens": [ + -3.529852867126465 + ], + "logprob_reward": -0.029852934181690216, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029852934181690216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07845385372638702, + 0.1120818555355072, + 0.037277646362781525, + 0.13477730751037598, + 0.049990203231573105, + 0.04536498710513115, + 0.048757247626781464, + 0.04957479238510132, + 0.09586425870656967, + 0.34785783290863037 + ], + "credit_map": [ + 0.012312207370996475, + 0.013356286101043224, + 0.01201234757900238, + 0.0065034786239266396, + 0.00678192637860775, + 0.005885034333914518, + 0.009403093717992306, + 0.01219947636127472, + 0.015652736648917198, + 0.016278289258480072, + 0.015453065745532513, + 0.017828091979026794, + 0.009768180549144745, + 0.008080178871750832, + 0.01342355739325285, + 0.015597750432789326, + 0.004975465126335621, + 0.003328117774799466, + 0.007661986164748669, + 0.0035727883223444223, + 0.004436924122273922, + 0.002937508746981621, + 0.0042941393330693245, + 0.0060707153752446175, + 0.017110152170062065, + 0.015050140209496021, + 0.0145941860973835, + 0.025762422010302544, + 0.015589884482324123, + 0.011993153020739555, + 0.013180499896407127, + 0.02149687334895134, + 0.006494357716292143, + 0.009730830788612366, + 0.006943142507225275, + 0.00589456083253026, + 0.0039055279921740294, + 0.003840645542368293, + 0.005055847577750683, + 0.008125290274620056, + 0.00472381291911006, + 0.004492748063057661, + 0.008997095748782158, + 0.006573819555342197, + 0.0037883257027715445, + 0.003748571267351508, + 0.006234660744667053, + 0.0068059517070651054, + 0.005730368196964264, + 0.005212748423218727, + 0.007675455883145332, + 0.0049393754452466965, + 0.004183308687061071, + 0.005960562266409397, + 0.008084828034043312, + 0.006970602553337812, + 0.002717336406931281, + 0.002389373490586877, + 0.00377988675609231, + 0.0033763283863663673, + 0.0030926584731787443, + 0.002390654059126973, + 0.014793764799833298, + 0.01703479140996933, + 0.004377205390483141, + 0.003544050734490156, + 0.006709710229188204, + 0.010181494057178497, + 0.005939668044447899, + 0.0035182414576411247, + 0.032247357070446014, + 0.029346533119678497, + 0.032566823065280914, + 0.033950015902519226, + 0.059388287365436554, + 0.04019153490662575, + 0.03248995542526245, + 0.05080612003803253, + 0.04602360352873802, + 0.05244150012731552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.029852934181690216 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204232022166252 + ], + "opposite_logprob_tokens": [ + -4.264204025268555 + ], + "logprob_reward": -0.014204232022166252, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204232022166252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07742589712142944, + 0.08499540388584137, + 0.07679113000631332, + 0.10336022078990936, + 0.08965322375297546, + 0.048985473811626434, + 0.08747793734073639, + 0.10384032875299454, + 0.16341717541217804, + 0.16405314207077026 + ], + "credit_map": [ + 0.009156756103038788, + 0.007429124787449837, + 0.012607998214662075, + 0.00937039777636528, + 0.009707007557153702, + 0.005508851259946823, + 0.016068261116743088, + 0.007577512878924608, + 0.009762837551534176, + 0.009849530644714832, + 0.008924894966185093, + 0.012793156318366528, + 0.010330216959118843, + 0.010128051042556763, + 0.012603726238012314, + 0.010602994821965694, + 0.007884955033659935, + 0.00528627960011363, + 0.01893594115972519, + 0.006195859517902136, + 0.012540271505713463, + 0.004312736447900534, + 0.010868923738598824, + 0.010766169987618923, + 0.015405451878905296, + 0.015522495843470097, + 0.011489560827612877, + 0.01620231382548809, + 0.012299914844334126, + 0.007691007107496262, + 0.01239708997309208, + 0.012352395802736282, + 0.009468823671340942, + 0.009662454947829247, + 0.014955887570977211, + 0.012802843935787678, + 0.008599454537034035, + 0.0052511729300022125, + 0.007261462975293398, + 0.0216511320322752, + 0.005408063065260649, + 0.005278455559164286, + 0.007567409425973892, + 0.005829001311212778, + 0.004236903507262468, + 0.004991560708731413, + 0.006768238265067339, + 0.008905848488211632, + 0.010584281757473946, + 0.010402340441942215, + 0.014735976234078407, + 0.0073422095738351345, + 0.007905261591076851, + 0.009616431780159473, + 0.01433459110558033, + 0.012556854635477066, + 0.008859643712639809, + 0.012370712123811245, + 0.013021315447986126, + 0.009116517379879951, + 0.010949078015983105, + 0.008284796960651875, + 0.019422713667154312, + 0.021815558895468712, + 0.007090797182172537, + 0.007924946956336498, + 0.011053822003304958, + 0.015531657263636589, + 0.011063579469919205, + 0.009621593169867992, + 0.05235058069229126, + 0.048780206590890884, + 0.01597404293715954, + 0.036730896681547165, + 0.017757510766386986, + 0.023326799273490906, + 0.015963107347488403, + 0.014402315020561218, + 0.019523905590176582, + 0.020374586805701256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.014204232022166252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703420221805573 + ], + "opposite_logprob_tokens": [ + -2.1270341873168945 + ], + "logprob_reward": -0.12703420221805573, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703420221805573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09416476637125015, + 0.07716260850429535, + 0.055717721581459045, + 0.07276872545480728, + 0.12313514947891235, + 0.09376604110002518, + 0.09714820235967636, + 0.1424095183610916, + 0.1286740005016327, + 0.11505322903394699 + ], + "credit_map": [ + 0.01315335463732481, + 0.010446948930621147, + 0.011886436492204666, + 0.008460708893835545, + 0.01231143157929182, + 0.01089266873896122, + 0.013785715214908123, + 0.013227506540715694, + 0.017501268535852432, + 0.008697565644979477, + 0.008911635726690292, + 0.013394651934504509, + 0.005929681472480297, + 0.004178236238658428, + 0.008740906603634357, + 0.009808657690882683, + 0.005474185571074486, + 0.003635361324995756, + 0.005529860500246286, + 0.007010259665548801, + 0.012936622835695744, + 0.004669036716222763, + 0.010910434648394585, + 0.005551961250603199, + 0.008958601392805576, + 0.008822533302009106, + 0.01053618174046278, + 0.009123611263930798, + 0.007364009972661734, + 0.00818302109837532, + 0.006485959980636835, + 0.013294803909957409, + 0.013785244897007942, + 0.015426639467477798, + 0.017808668315410614, + 0.017390171065926552, + 0.012957924976944923, + 0.006907688453793526, + 0.010509997606277466, + 0.028348812833428383, + 0.00903113093227148, + 0.00989033468067646, + 0.0160392913967371, + 0.01384976040571928, + 0.008472580462694168, + 0.005380843300372362, + 0.0156397745013237, + 0.015462330542504787, + 0.011721879243850708, + 0.011593365110456944, + 0.02090797759592533, + 0.010647871531546116, + 0.006908117793500423, + 0.005615964066237211, + 0.014667066745460033, + 0.015085959807038307, + 0.012703156098723412, + 0.01615944877266884, + 0.017225952818989754, + 0.013044502586126328, + 0.016491424292325974, + 0.01644730567932129, + 0.023597871884703636, + 0.026739859953522682, + 0.01611262932419777, + 0.01589886099100113, + 0.01465535070747137, + 0.016242729499936104, + 0.013401483185589314, + 0.021779118105769157, + 0.015312484465539455, + 0.015271341428160667, + 0.01054622046649456, + 0.02510320208966732, + 0.013538223691284657, + 0.016146397218108177, + 0.01213409285992384, + 0.008910568431019783, + 0.01556872483342886, + 0.013105805031955242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 4 + ], + "selected_reward": -0.12703420221805573 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive sturgeon with five distinct rows of sharp bony plates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the video rendered in stark, high-contrast black and white (pure monochrome)?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the sturgeon move slowly and powerfully while dragging its barbs through the sediment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan steadily to keep the creature centered in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene shot from a low angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065435484051704 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065435484051704, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065435484051704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1080615296959877, + 0.05597090348601341, + 0.06457164883613586, + 0.19690540432929993, + 0.16541290283203125, + 0.0346047580242157, + 0.058893173933029175, + 0.09698139876127243, + 0.11068791896104813, + 0.1079103946685791 + ], + "credit_map": [ + 0.015426062978804111, + 0.013140249066054821, + 0.015544440597295761, + 0.01349177397787571, + 0.007620006799697876, + 0.008095773868262768, + 0.016173390671610832, + 0.0185698289424181, + 0.006172912660986185, + 0.008102393709123135, + 0.00773638067767024, + 0.00924008246511221, + 0.004914024379104376, + 0.003694945713505149, + 0.008013454265892506, + 0.008096710778772831, + 0.005809890106320381, + 0.005402673501521349, + 0.012576255016028881, + 0.00667163822799921, + 0.012870686128735542, + 0.005379387177526951, + 0.008106645196676254, + 0.007754472084343433, + 0.024526966735720634, + 0.021249035373330116, + 0.02031772769987583, + 0.037794578820466995, + 0.021137800067663193, + 0.016565129160881042, + 0.018931565806269646, + 0.036382608115673065, + 0.015912072733044624, + 0.019878335297107697, + 0.022729869931936264, + 0.021607253700494766, + 0.016741616651415825, + 0.012916867621243, + 0.020956484600901604, + 0.034670405089855194, + 0.003985168412327766, + 0.0036341801751405, + 0.005423721857368946, + 0.004816056229174137, + 0.0032205062452703714, + 0.0034813103266060352, + 0.004945666063576937, + 0.00509814964607358, + 0.006752615328878164, + 0.006126695778220892, + 0.012062451802194118, + 0.006975786294788122, + 0.005615977570414543, + 0.006120962090790272, + 0.0072502754628658295, + 0.007988406345248222, + 0.007136446423828602, + 0.01004700269550085, + 0.00894909631460905, + 0.006481066811829805, + 0.009832949377596378, + 0.008840176276862621, + 0.02438354305922985, + 0.021311109885573387, + 0.006251197773963213, + 0.00815543346107006, + 0.00794906448572874, + 0.011415504850447178, + 0.010474578477442265, + 0.004433618858456612, + 0.03196696564555168, + 0.03004155308008194, + 0.009407409466803074, + 0.02058006450533867, + 0.011804094538092613, + 0.02210060879588127, + 0.008365223184227943, + 0.009267210029065609, + 0.013916918076574802, + 0.012468867003917694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.016065435484051704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_tensor([6545, 6545])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_tensor([6545, 6545])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3afb10d840a7644513c8b774f84c1e76417d3150 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_tensor([6545, 6545])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping wide shot begins, the camera slowly and steadily panning from left to right across the tranquil surface of a lake at sunrise. A soft mist hovers above the water, capturing the sun's first rays and painting the horizon in a stunning gradient of soft pink, warm orange, and deep purple, all perfectly reflected in the still water. The camera's movement pauses, focusing on a weathered wooden dock where an elderly couple sits. The man, with silver-white hair and friendly wrinkles etched around his eyes, wears a faded plaid flannel shirt over a practical khaki fishing vest. He sits on a simple wooden box, gripping an old-fashioned fishing rod with a steady, practiced hand. His wife sits beside him, their shoulders almost touching, her soft gray hair pulled back in a loose bun, a cozy cream-colored knitted shawl wrapped around her shoulders. She also holds a fishing rod, her posture relaxed and her expression calm. Their lines disappear into the water, their buoys bobbing motionlessly, the only distraction from the perfect reflection. Their shared silence was not empty but fulfilling, a comfortable quiet that spoke of a lifetime together, creating an overall atmosphere of profound peace, nostalgia and tranquility, of unspoken love.", + "seed": 5513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_6545_seed_5513_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.42396475579589604, + "vlm_logprob": -0.42396475579589604, + "vlm_yes_no_margin": 2.200000011920929, + "vlm_token_credit": -0.42396475579589604, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0632383823394775 + ], + "opposite_logprob_tokens": [ + -0.4382384419441223 + ], + "logprob_reward": -1.0632383823394775, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0632383823394775, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.30551302433013916, + 0.1389814168214798, + 0.09794890880584717, + 0.10816524177789688, + 0.06402266770601273, + 0.0692426934838295, + 0.10197850316762924, + 0.04455133154988289, + 0.027228079736232758, + 0.04236810654401779 + ], + "credit_map": [ + 0.014374221675097942, + 0.09335803985595703, + 0.013042078353464603, + 0.01670265756547451, + 0.09209573268890381, + 0.022837430238723755, + 0.022554414346814156, + 0.030548449605703354, + 0.023957932367920876, + 0.017848147079348564, + 0.011817317456007004, + 0.01005737017840147, + 0.01905536651611328, + 0.014251978136599064, + 0.013449965976178646, + 0.028543345630168915, + 0.0139622176066041, + 0.014080597087740898, + 0.011115064844489098, + 0.009329423308372498, + 0.014613316394388676, + 0.011494791135191917, + 0.009701318107545376, + 0.013652181252837181, + 0.010059011168777943, + 0.014680875465273857, + 0.010134019888937473, + 0.009375094436109066, + 0.015032556839287281, + 0.013291668146848679, + 0.011931251734495163, + 0.023660773411393166, + 0.005218686535954475, + 0.006171784829348326, + 0.006047470960766077, + 0.005103345960378647, + 0.010263352654874325, + 0.01181802712380886, + 0.008813276886940002, + 0.010586720891296864, + 0.00602143257856369, + 0.005809559486806393, + 0.006749060470610857, + 0.0042104413732886314, + 0.012634984217584133, + 0.014874696731567383, + 0.008136125281453133, + 0.010806395672261715, + 0.00992676243185997, + 0.010372092016041279, + 0.012737557291984558, + 0.008726446889340878, + 0.015044941566884518, + 0.015040828846395016, + 0.01341240108013153, + 0.016717473044991493, + 0.0038684518076479435, + 0.006590126547962427, + 0.005130686331540346, + 0.004998139105737209, + 0.006021516397595406, + 0.005173163954168558, + 0.006860771682113409, + 0.00590847572311759, + 0.005673885345458984, + 0.0028525274246931076, + 0.0031830123625695705, + 0.0024895803071558475, + 0.003731581149622798, + 0.0038338389713317156, + 0.0028389005456119776, + 0.0026247533969581127, + 0.008072366006672382, + 0.003923249430954456, + 0.005181531421840191, + 0.003466453403234482, + 0.005796280223876238, + 0.00682850880548358, + 0.004423857666552067, + 0.0046758586540818214 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -1.0632383823394775 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9743353128433228 + ], + "opposite_logprob_tokens": [ + -0.47433531284332275 + ], + "logprob_reward": -0.9743353128433228, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9743353128433228, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3417515754699707, + 0.1711084395647049, + 0.06277317553758621, + 0.08592794835567474, + 0.06485750526189804, + 0.06022817641496658, + 0.052683304995298386, + 0.05667325481772423, + 0.043405931442976, + 0.06059066578745842 + ], + "credit_map": [ + 0.02313096821308136, + 0.09551545977592468, + 0.021850710734725, + 0.017053531482815742, + 0.0839449018239975, + 0.02422177605330944, + 0.028280828148126602, + 0.047753434628248215, + 0.029945669695734978, + 0.021221131086349487, + 0.016352156177163124, + 0.02205870859324932, + 0.028004107996821404, + 0.02359134517610073, + 0.014238075353205204, + 0.015697240829467773, + 0.0061257812194526196, + 0.008940850384533405, + 0.006724639795720577, + 0.00577070377767086, + 0.010497622191905975, + 0.009425090625882149, + 0.006756830960512161, + 0.008531657978892326, + 0.014699543826282024, + 0.013305982574820518, + 0.010812356136739254, + 0.009917845949530602, + 0.00996999628841877, + 0.009767575189471245, + 0.008390455506742, + 0.00906419102102518, + 0.005279588047415018, + 0.008143060840666294, + 0.010617699474096298, + 0.006476153153926134, + 0.007984338328242302, + 0.007656799163669348, + 0.011745098978281021, + 0.006954767741262913, + 0.00565644446760416, + 0.006424227263778448, + 0.01023031398653984, + 0.006300070323050022, + 0.007491023745387793, + 0.006151013541966677, + 0.007933562621474266, + 0.010041522793471813, + 0.005234811455011368, + 0.005708522628992796, + 0.008853968232870102, + 0.006075720768421888, + 0.007905196398496628, + 0.006361847743391991, + 0.005785240791738033, + 0.006757998373359442, + 0.00545323034748435, + 0.006892104633152485, + 0.01157087367027998, + 0.0063620517030358315, + 0.007778584491461515, + 0.006657951977103949, + 0.00566884083673358, + 0.006289612501859665, + 0.004834591411054134, + 0.0045163314789533615, + 0.005953798536211252, + 0.004410445224493742, + 0.005240526050329208, + 0.005789807997643948, + 0.006735231261700392, + 0.0059251985512673855, + 0.009506727568805218, + 0.007234928663820028, + 0.005925165954977274, + 0.006063967011868954, + 0.0074963378719985485, + 0.007052325177937746, + 0.009041466750204563, + 0.008269748650491238 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.9743353128433228 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005972514860332012 + ], + "opposite_logprob_tokens": [ + -5.130972385406494 + ], + "logprob_reward": -0.005972514860332012, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005972514860332012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17615434527397156, + 0.14338786900043488, + 0.1177324429154396, + 0.05174540355801582, + 0.09111402183771133, + 0.08319851756095886, + 0.05286256968975067, + 0.14935660362243652, + 0.07157032936811447, + 0.06287789344787598 + ], + "credit_map": [ + 0.01440813671797514, + 0.022006802260875702, + 0.015835845842957497, + 0.01976339891552925, + 0.031334131956100464, + 0.026438908651471138, + 0.018996944651007652, + 0.02737017348408699, + 0.01277864445000887, + 0.012089546769857407, + 0.010832343250513077, + 0.013844684697687626, + 0.02000155858695507, + 0.01476315688341856, + 0.023242902010679245, + 0.03583502024412155, + 0.015865754336118698, + 0.016150930896401405, + 0.01610712520778179, + 0.012451869435608387, + 0.01714986376464367, + 0.013207039795815945, + 0.011809682473540306, + 0.014990173280239105, + 0.0044233850203454494, + 0.004897026345133781, + 0.004351621959358454, + 0.002635515760630369, + 0.008079590275883675, + 0.00993859302252531, + 0.007079096511006355, + 0.010340574197471142, + 0.008039033971726894, + 0.007423010189086199, + 0.007291159592568874, + 0.008981543593108654, + 0.014125769957900047, + 0.0116311926394701, + 0.019319582730531693, + 0.014302728697657585, + 0.008533798158168793, + 0.006912765558809042, + 0.010566266253590584, + 0.007376816589385271, + 0.010952750220894814, + 0.010519832372665405, + 0.012121982872486115, + 0.01621430180966854, + 0.0035578596871346235, + 0.0064046336337924, + 0.0050384062342345715, + 0.0045436532236635685, + 0.007565292064100504, + 0.009729843586683273, + 0.008219597861170769, + 0.007803283631801605, + 0.01613500341773033, + 0.014299517497420311, + 0.019438328221440315, + 0.01213044486939907, + 0.02431703358888626, + 0.02785995788872242, + 0.015874430537223816, + 0.0193018838763237, + 0.013195615261793137, + 0.009958631359040737, + 0.007009370252490044, + 0.0074952831491827965, + 0.00822480209171772, + 0.00790710560977459, + 0.00957583636045456, + 0.008203685283660889, + 0.008360113948583603, + 0.005426991265267134, + 0.007900409400463104, + 0.005920767784118652, + 0.009273604489862919, + 0.011845720931887627, + 0.007324103266000748, + 0.006826186086982489 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.005972514860332012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014215279370546341 + ], + "opposite_logprob_tokens": [ + -4.264215469360352 + ], + "logprob_reward": -0.014215279370546341, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014215279370546341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2297971397638321, + 0.16162990033626556, + 0.07657375186681747, + 0.08839476108551025, + 0.11194740235805511, + 0.10341749340295792, + 0.07861959934234619, + 0.06354986876249313, + 0.022431692108511925, + 0.06363839656114578 + ], + "credit_map": [ + 0.022752143442630768, + 0.05896355211734772, + 0.01831258274614811, + 0.01577877067029476, + 0.06660901755094528, + 0.0166900884360075, + 0.013344939798116684, + 0.017346041277050972, + 0.016730019822716713, + 0.014781671576201916, + 0.013086874037981033, + 0.01547370944172144, + 0.02267472632229328, + 0.01851237565279007, + 0.02432098053395748, + 0.036049533635377884, + 0.013780978508293629, + 0.011608019471168518, + 0.008368886075913906, + 0.007092251908034086, + 0.00995732843875885, + 0.010136708617210388, + 0.006786256562918425, + 0.008843324147164822, + 0.010329139418900013, + 0.012887773104012012, + 0.014035232365131378, + 0.01022642944008112, + 0.010555878281593323, + 0.010262575931847095, + 0.008298865519464016, + 0.011798864230513573, + 0.008334595710039139, + 0.010318235494196415, + 0.009643414989113808, + 0.008746382780373096, + 0.016443749889731407, + 0.01326731313019991, + 0.02334773913025856, + 0.021845974028110504, + 0.007775881327688694, + 0.013168307021260262, + 0.015698267146945, + 0.017246004194021225, + 0.011670438572764397, + 0.012627671472728252, + 0.012708190828561783, + 0.012522735632956028, + 0.00620952108874917, + 0.012974158860743046, + 0.010451843030750751, + 0.011081907898187637, + 0.009505883790552616, + 0.00984709057956934, + 0.010605490766465664, + 0.007943697273731232, + 0.007775136735290289, + 0.005702311638742685, + 0.014695936813950539, + 0.0066847167909145355, + 0.00847956445068121, + 0.006288864649832249, + 0.006566861178725958, + 0.00735647464171052, + 0.0032154633663594723, + 0.004207116086035967, + 0.0019774786196649075, + 0.002787729725241661, + 0.003067568875849247, + 0.0028821693267673254, + 0.0017613543896004558, + 0.002532811602577567, + 0.006002034060657024, + 0.007121080532670021, + 0.008039310574531555, + 0.0077284411527216434, + 0.007399216294288635, + 0.006887012626975775, + 0.012806379236280918, + 0.007654921617358923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.014215279370546341 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06206228956580162 + ], + "opposite_logprob_tokens": [ + -2.8120622634887695 + ], + "logprob_reward": -0.06206228956580162, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06206228956580162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3536786735057831, + 0.1799193024635315, + 0.07177331298589706, + 0.07065913081169128, + 0.043580297380685806, + 0.05913969874382019, + 0.08138830959796906, + 0.059614893049001694, + 0.040169503539800644, + 0.04007689282298088 + ], + "credit_map": [ + 0.012607678771018982, + 0.1122460588812828, + 0.012690049596130848, + 0.021924566477537155, + 0.11541908234357834, + 0.02907157875597477, + 0.023226117715239525, + 0.026493512094020844, + 0.014588414691388607, + 0.015151692554354668, + 0.014363272115588188, + 0.022230317816138268, + 0.02782524935901165, + 0.01928836852312088, + 0.03350270166993141, + 0.03296928107738495, + 0.007451244629919529, + 0.007367931772023439, + 0.00694479513913393, + 0.005154143087565899, + 0.013623223640024662, + 0.008784657344222069, + 0.007611587177962065, + 0.014835727401077747, + 0.014120838604867458, + 0.006806082092225552, + 0.005687739234417677, + 0.005770513787865639, + 0.011084746569395065, + 0.010922913439571857, + 0.007746507879346609, + 0.00851978175342083, + 0.00449795089662075, + 0.004373347386717796, + 0.004755218513309956, + 0.0036860527470707893, + 0.006402095779776573, + 0.006324063055217266, + 0.005198420491069555, + 0.008343146182596684, + 0.005916232708841562, + 0.005475067533552647, + 0.008294930681586266, + 0.005250361282378435, + 0.009130052290856838, + 0.008240975439548492, + 0.006548229604959488, + 0.010283847339451313, + 0.00857067946344614, + 0.009306168183684349, + 0.011198341846466064, + 0.008596032857894897, + 0.010961181484162807, + 0.010751626454293728, + 0.011761447414755821, + 0.010242830961942673, + 0.006644921377301216, + 0.006471884902566671, + 0.009631605818867683, + 0.005010475870221853, + 0.008179932832717896, + 0.007708535064011812, + 0.009024195373058319, + 0.0069433413445949554, + 0.010492381639778614, + 0.004460398107767105, + 0.0042518130503594875, + 0.0027235106099396944, + 0.005024452228099108, + 0.005035751033574343, + 0.004087058361619711, + 0.004094140138477087, + 0.008358467370271683, + 0.007070528343319893, + 0.0038280601147562265, + 0.0038632259238511324, + 0.005287514999508858, + 0.005106284283101559, + 0.002890767529606819, + 0.003672043327242136 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.06206228956580162 + } + ], + "vlm_decoded_outputs": [ + "No", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera slowly and steadily panning from left to right across the tranquil surface of a lake at sunrise. A soft mist hovers above the water, capturing the sun's first rays and painting the horizon in a stunning gradient of soft pink, warm orange, and deep purple, all perfectly reflected in the still water. The camera's movement pauses, focusing on a weathered wooden dock where an elderly couple sits. The man, with silver-white hair and friendly wrinkles etched around his eyes, wears a faded plaid flannel shirt over a practical khaki fishing vest. He sits on a simple wooden box, gripping an old-fashioned fishing rod with a steady, practiced hand. His wife sits beside him, their shoulders almost touching, her soft gray hair pulled back in a loose bun, a cozy cream-colored knitted shawl wrapped around her shoulders. She also holds a fishing rod, her posture relaxed and her expression calm. Their lines disappear into the water, their buoys bobbing motionlessly, the only distraction from the perfect reflection. Their shared silence was not empty but fulfilling, a comfortable quiet that spoke of a lifetime together, creating an overall atmosphere of profound peace, nostalgia and tranquility, of unspoken love.", + "seed": 5514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_6545_seed_5514_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1359644421376288, + "vlm_logprob": -0.1359644421376288, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.1359644421376288, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.35815536975860596 + ], + "opposite_logprob_tokens": [ + -1.233155369758606 + ], + "logprob_reward": -0.35815536975860596, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.35815536975860596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04833824187517166, + 0.03304092586040497, + 0.03130298852920532, + 0.04851141199469566, + 0.019926078617572784, + 0.010905001312494278, + 0.32244768738746643, + 0.056042056530714035, + 0.4233328700065613, + 0.00615268386900425 + ], + "credit_map": [ + 0.00915819313377142, + 0.005531334783881903, + 0.005450753960758448, + 0.0059858583845198154, + 0.005197125021368265, + 0.004385476000607014, + 0.0068288506008684635, + 0.0058006527833640575, + 0.0034306845627725124, + 0.00314593524672091, + 0.004421297460794449, + 0.0035248491913080215, + 0.0035070537123829126, + 0.0038788688834756613, + 0.006216417998075485, + 0.004915819503366947, + 0.0021797940135002136, + 0.0046674832701683044, + 0.0034591585863381624, + 0.0022206478752195835, + 0.003720393870025873, + 0.003893633373081684, + 0.005928968545049429, + 0.00523291015997529, + 0.0031027698423713446, + 0.008191395550966263, + 0.0055246842093765736, + 0.005693725775927305, + 0.005146052222698927, + 0.00485434103757143, + 0.008230892941355705, + 0.007767553441226482, + 0.001925813965499401, + 0.0023920019157230854, + 0.002811466809362173, + 0.0019153824541717768, + 0.002994295908138156, + 0.0030035076197236776, + 0.0019955250900238752, + 0.0028880848549306393, + 0.0013799103908240795, + 0.0013217937666922808, + 0.001109439879655838, + 0.0016989543801173568, + 0.0010967097477987409, + 0.0012160628102719784, + 0.0015858301194384694, + 0.0014963005669414997, + 0.12558133900165558, + 0.00858407374471426, + 0.009552246890962124, + 0.004029514733701944, + 0.009035433642566204, + 0.14408327639102936, + 0.008400604128837585, + 0.013181174173951149, + 0.021254945546388626, + 0.0015089191729202867, + 0.0018617763416841626, + 0.0007971272571012378, + 0.002105355029925704, + 0.025118885561823845, + 0.001400514505803585, + 0.0019945288076996803, + 0.16706441342830658, + 0.009080029092729092, + 0.015158244408667088, + 0.005922229494899511, + 0.009970314800739288, + 0.19157597422599792, + 0.011166165582835674, + 0.0133955217897892, + 0.0008003375260159373, + 0.0005194680416025221, + 0.0008285809308290482, + 0.0007900099153630435, + 0.0006635765894316137, + 0.0013838252052664757, + 0.0004201470874249935, + 0.0007467387476935983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.35815536975860596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211045145988464 + ], + "opposite_logprob_tokens": [ + -1.502110481262207 + ], + "logprob_reward": -0.25211045145988464, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211045145988464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.013980450108647346, + 0.01857888698577881, + 0.013892927207052708, + 0.006219339091330767, + 0.010761955752968788, + 0.009272566996514797, + 0.03809717670083046, + 0.012646219693124294, + 0.8728537559509277, + 0.0036967620253562927 + ], + "credit_map": [ + 0.0026852719020098448, + 0.0017278882442042232, + 0.0015286658890545368, + 0.0015159102622419596, + 0.0016578846843913198, + 0.0015594469150528312, + 0.0017180226277559996, + 0.0015873578377068043, + 0.0009486732888035476, + 0.0012674832250922918, + 0.0013609756715595722, + 0.0012117887381464243, + 0.002161493059247732, + 0.002449027495458722, + 0.0064299143850803375, + 0.0027495299000293016, + 0.0015975182177498937, + 0.0025592283345758915, + 0.002564654452726245, + 0.0014979279367253184, + 0.0014575261157006025, + 0.0011728249955922365, + 0.0015404416481032968, + 0.0015028042253106833, + 0.00033077833359129727, + 0.0008260192698799074, + 0.00047718477435410023, + 0.0006663434905931354, + 0.0006014112732373178, + 0.0007039826014079154, + 0.0012214456219226122, + 0.001392173464410007, + 0.0008750931010581553, + 0.001586133148521185, + 0.001356189139187336, + 0.0011635787086561322, + 0.0013611764879897237, + 0.0015661119250580668, + 0.0010738809360191226, + 0.001779792015440762, + 0.0011947783641517162, + 0.0012822761200368404, + 0.0009023695602081716, + 0.0018110882956534624, + 0.0009877263801172376, + 0.0010692633222788572, + 0.0009549803799018264, + 0.0010700838174670935, + 0.013517496176064014, + 0.0013614341150969267, + 0.002108335494995117, + 0.0010288404300808907, + 0.001339611248113215, + 0.015704454854130745, + 0.0011896934593096375, + 0.0018473074305802584, + 0.004773796070367098, + 0.00036446156445890665, + 0.0004042485379613936, + 0.00026986593729816377, + 0.0004548184806481004, + 0.005597120616585016, + 0.0003414047241676599, + 0.0004405044310260564, + 0.3439912796020508, + 0.01884140446782112, + 0.03178384527564049, + 0.012063294649124146, + 0.020702281966805458, + 0.3951050937175751, + 0.023014763370156288, + 0.027351729571819305, + 0.0005262955673970282, + 0.0003071263199672103, + 0.0004239642294123769, + 0.0007415310828946531, + 0.0004950112197548151, + 0.0006670741713605821, + 0.0002577478298917413, + 0.00027801108080893755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.25211045145988464 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958413239568472 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958413239568472, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958413239568472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012950822710990906, + 0.007189349737018347, + 0.007805784698575735, + 0.011173347011208534, + 0.00722834188491106, + 0.010009003803133965, + 0.026088297367095947, + 0.04596590995788574, + 0.866517961025238, + 0.005071237217634916 + ], + "credit_map": [ + 0.0009606936946511269, + 0.0017089663306251168, + 0.0010006313677877188, + 0.00109980464912951, + 0.0017170764040201902, + 0.0012145527871325612, + 0.00347778364084661, + 0.0017713133711367846, + 0.0008061714470386505, + 0.000794055697042495, + 0.0009529207600280643, + 0.0007630178006365895, + 0.0007514739991165698, + 0.0008863889961503446, + 0.0013307341141626239, + 0.0009045868064276874, + 0.0005928031168878078, + 0.0006717268843203783, + 0.001051974599249661, + 0.0006559425382874906, + 0.0009563409257680178, + 0.0011464221170172095, + 0.0014442343963310122, + 0.001286339364014566, + 0.0009258183417841792, + 0.0019778278656303883, + 0.0012463057646527886, + 0.0009325548307970166, + 0.0012659572530537844, + 0.0014590572100132704, + 0.0016845130594447255, + 0.0016813117545098066, + 0.00059252989012748, + 0.001039534923620522, + 0.0011638959404081106, + 0.0007410137332044542, + 0.0008981323335319757, + 0.0009194579906761646, + 0.0008596935658715665, + 0.001014082576148212, + 0.0010350745869800448, + 0.001818555872887373, + 0.0010464458027854562, + 0.0018120527965947986, + 0.0010274903615936637, + 0.0011465100105851889, + 0.0009401525603607297, + 0.0011827204143628478, + 0.009956025518476963, + 0.0007685417076572776, + 0.0008952324278652668, + 0.00048325685202144086, + 0.0007185571594163775, + 0.011355482041835785, + 0.0007955646142363548, + 0.0011156355030834675, + 0.016699854284524918, + 0.0013918643817305565, + 0.0018369856989011168, + 0.0016703481087461114, + 0.0015797900268808007, + 0.019526474177837372, + 0.001407867530360818, + 0.0018527196953073144, + 0.3412722647190094, + 0.01883799396455288, + 0.03163648396730423, + 0.012022786773741245, + 0.020691346377134323, + 0.39202338457107544, + 0.022817334160208702, + 0.027216272428631783, + 0.000699517666362226, + 0.0004090166767127812, + 0.0007171307224780321, + 0.0010195670183748007, + 0.0005963874864391983, + 0.000871273223310709, + 0.00037736716330982745, + 0.00038097717333585024 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.005958413239568472 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008643004111945629 + ], + "opposite_logprob_tokens": [ + -4.75864315032959 + ], + "logprob_reward": -0.008643004111945629, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008643004111945629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006140407174825668, + 0.006550723221153021, + 0.006218024529516697, + 0.004217995330691338, + 0.00798262283205986, + 0.01830446720123291, + 0.11404041200876236, + 0.11665607243776321, + 0.7141898274421692, + 0.005699448753148317 + ], + "credit_map": [ + 0.0005700053297914565, + 0.0006771849584765732, + 0.0008862625691108406, + 0.0006977366865612566, + 0.0006947257206775248, + 0.0007674755179323256, + 0.0009578700992278755, + 0.0008891460602171719, + 0.0003464991750661284, + 0.00048180564772337675, + 0.0005270170513540506, + 0.0005368019919842482, + 0.0007700411952100694, + 0.0008331535500474274, + 0.00213858624920249, + 0.0009168184478767216, + 0.0005931978230364621, + 0.0006317938677966595, + 0.0014333113795146346, + 0.0006605588714592159, + 0.0006628509145230055, + 0.0007090896833688021, + 0.0007797724683769047, + 0.0007474495214410126, + 0.0004750224179588258, + 0.0010200303513556719, + 0.00046386741450987756, + 0.0003710288438014686, + 0.0004336143028922379, + 0.00040140433702617884, + 0.00040436905692331493, + 0.0006486587808467448, + 0.0006102353800088167, + 0.0009600237826816738, + 0.0012751088943332434, + 0.0007888171239756048, + 0.0010101890657097101, + 0.0011313584400340915, + 0.0010694552911445498, + 0.0011374353198334575, + 0.0017955783987417817, + 0.0030743665993213654, + 0.002097022719681263, + 0.0031244445126503706, + 0.0017571717035025358, + 0.002330603078007698, + 0.001644458039663732, + 0.0024808228481560946, + 0.04444970563054085, + 0.0030548383947461843, + 0.0030967476777732372, + 0.0015016000252217054, + 0.003323034383356571, + 0.05094655975699425, + 0.00295980810187757, + 0.004708119668066502, + 0.04250429570674896, + 0.0034575453028082848, + 0.004379797726869583, + 0.004091246053576469, + 0.0040526799857616425, + 0.049725327640771866, + 0.0036429299507290125, + 0.004802247509360313, + 0.28153344988822937, + 0.015403199009597301, + 0.02587459422647953, + 0.009869400411844254, + 0.016856633126735687, + 0.3235126733779907, + 0.01877536252140999, + 0.02236446924507618, + 0.0009819214465096593, + 0.0006272141472436488, + 0.0005645141936838627, + 0.0005203025066293776, + 0.0010807180078700185, + 0.0011644682381302118, + 0.00036935057141818106, + 0.00039095967076718807 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.008643004111945629 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495497211813927 + ], + "opposite_logprob_tokens": [ + -2.929955005645752 + ], + "logprob_reward": -0.05495497211813927, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495497211813927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036524586379528046, + 0.04106268286705017, + 0.014448031783103943, + 0.011483426205813885, + 0.011745141819119453, + 0.020619677379727364, + 0.18022170662879944, + 0.033531930297613144, + 0.6422460079193115, + 0.008116715587675571 + ], + "credit_map": [ + 0.0023118508979678154, + 0.005516661796718836, + 0.003041383344680071, + 0.0028561337385326624, + 0.0056725009344518185, + 0.006359214894473553, + 0.006082328502088785, + 0.0046845111064612865, + 0.0023298393934965134, + 0.003060187678784132, + 0.003062262898311019, + 0.0030347630381584167, + 0.004489490762352943, + 0.006394470110535622, + 0.011810289695858955, + 0.006881380453705788, + 0.001124874921515584, + 0.001337033580057323, + 0.0011306367814540863, + 0.001031277235597372, + 0.0017510713078081608, + 0.001979082589969039, + 0.003388825571164489, + 0.0027052308432757854, + 0.00075206981273368, + 0.0017829382559284568, + 0.0013014039723202586, + 0.0012749097077175975, + 0.001250980538316071, + 0.0011932679917663336, + 0.0016128551214933395, + 0.002315001329407096, + 0.0010649438481777906, + 0.0012350637698546052, + 0.0016680193366482854, + 0.001408052397891879, + 0.0014522862620651722, + 0.001848435727879405, + 0.0013009555405005813, + 0.0017673851689323783, + 0.002177978167310357, + 0.0036768284626305103, + 0.0025435846764594316, + 0.003164018038660288, + 0.0020412630401551723, + 0.002532805083319545, + 0.001994161633774638, + 0.0024890380445867777, + 0.07007933408021927, + 0.004844012204557657, + 0.0050362651236355305, + 0.002165349666029215, + 0.0051512084901332855, + 0.08063069730997086, + 0.00478361127898097, + 0.007531209383159876, + 0.012125784531235695, + 0.0010628323070704937, + 0.001320033916272223, + 0.001232306705787778, + 0.0011908700689673424, + 0.014165007509291172, + 0.0010315454564988613, + 0.001403547590598464, + 0.25247272849082947, + 0.013946043327450752, + 0.023732805624604225, + 0.009074820205569267, + 0.015295426361262798, + 0.29045024514198303, + 0.017058856785297394, + 0.02021510899066925, + 0.0009829035261645913, + 0.001005765050649643, + 0.001145659713074565, + 0.0012144251959398389, + 0.0010135878110304475, + 0.0014686331851407886, + 0.0005280235200189054, + 0.0007577177020721138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 1 + ], + "selected_reward": -0.05495497211813927 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera slowly and steadily panning from left to right across the tranquil surface of a lake at sunrise. A soft mist hovers above the water, capturing the sun's first rays and painting the horizon in a stunning gradient of soft pink, warm orange, and deep purple, all perfectly reflected in the still water. The camera's movement pauses, focusing on a weathered wooden dock where an elderly couple sits. The man, with silver-white hair and friendly wrinkles etched around his eyes, wears a faded plaid flannel shirt over a practical khaki fishing vest. He sits on a simple wooden box, gripping an old-fashioned fishing rod with a steady, practiced hand. His wife sits beside him, their shoulders almost touching, her soft gray hair pulled back in a loose bun, a cozy cream-colored knitted shawl wrapped around her shoulders. She also holds a fishing rod, her posture relaxed and her expression calm. Their lines disappear into the water, their buoys bobbing motionlessly, the only distraction from the perfect reflection. Their shared silence was not empty but fulfilling, a comfortable quiet that spoke of a lifetime together, creating an overall atmosphere of profound peace, nostalgia and tranquility, of unspoken love.", + "seed": 5515, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_6545_seed_5515_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2991358722560108, + "vlm_logprob": -0.2991358722560108, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.2991358722560108, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6378768682479858 + ], + "opposite_logprob_tokens": [ + -0.7628768682479858 + ], + "logprob_reward": -0.6378768682479858, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6378768682479858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08568612486124039, + 0.06621958315372467, + 0.07179942727088928, + 0.10272076725959778, + 0.3077147305011749, + 0.31513920426368713, + 0.019936148077249527, + 0.013075682334601879, + 0.0072619300335645676, + 0.01044637430459261 + ], + "credit_map": [ + 0.010764263570308685, + 0.01000896468758583, + 0.017519516870379448, + 0.006999575532972813, + 0.00898822396993637, + 0.014746894128620625, + 0.0077129811979830265, + 0.008945703506469727, + 0.007803662214428186, + 0.01108340360224247, + 0.008456437848508358, + 0.0067665050737559795, + 0.010569128207862377, + 0.0075434111058712006, + 0.008256332948803902, + 0.005740700755268335, + 0.00660693971440196, + 0.010908889584243298, + 0.013383972458541393, + 0.009935964830219746, + 0.008703325875103474, + 0.007305827457457781, + 0.008072705008089542, + 0.006881801877170801, + 0.01029125228524208, + 0.013227802701294422, + 0.01470230333507061, + 0.0124126635491848, + 0.0180103350430727, + 0.013024525716900826, + 0.011752358637750149, + 0.009299521334469318, + 0.028257526457309723, + 0.0381176732480526, + 0.03292370215058327, + 0.046279218047857285, + 0.03760891407728195, + 0.04222181811928749, + 0.05053234100341797, + 0.03177354112267494, + 0.05090723931789398, + 0.04753844812512398, + 0.05574238300323486, + 0.035600993782281876, + 0.029697244986891747, + 0.03951795771718025, + 0.03288493677973747, + 0.023250043392181396, + 0.0028816931881010532, + 0.002491237362846732, + 0.004159923642873764, + 0.002735073445364833, + 0.002136499620974064, + 0.002140521537512541, + 0.002345444168895483, + 0.001045755809172988, + 0.0016343052266165614, + 0.0018584891222417355, + 0.0030731267761439085, + 0.0016683605499565601, + 0.0011269146343693137, + 0.0014149207854643464, + 0.0012962808832526207, + 0.0010032843565568328, + 0.0011908103479072452, + 0.0012479263823479414, + 0.0015507229836657643, + 0.00071895302971825, + 0.0007499508210457861, + 0.0008738100877963006, + 0.0006010356009937823, + 0.00032872086740098894, + 0.0012210165150463581, + 0.0015154605498537421, + 0.002589320996776223, + 0.0018819342367351055, + 0.000751239072997123, + 0.0007257787510752678, + 0.001017204369418323, + 0.0007444196380674839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.6378768682479858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578389048576355 + ], + "opposite_logprob_tokens": [ + -0.6328389048576355 + ], + "logprob_reward": -0.7578389048576355, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578389048576355, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.18435291945934296, + 0.05663630738854408, + 0.2056395560503006, + 0.13117419183254242, + 0.18992741405963898, + 0.1531262844800949, + 0.024233384057879448, + 0.02099454030394554, + 0.020762033760547638, + 0.0131533807143569 + ], + "credit_map": [ + 0.042719826102256775, + 0.020525021478533745, + 0.028900010511279106, + 0.00966314785182476, + 0.01827622950077057, + 0.041344624012708664, + 0.011139524169266224, + 0.011784525588154793, + 0.008040884509682655, + 0.007762905675917864, + 0.007358688861131668, + 0.007639551069587469, + 0.007194888778030872, + 0.00814059004187584, + 0.005428252276033163, + 0.005070549435913563, + 0.052449602633714676, + 0.0764981210231781, + 0.01822018437087536, + 0.012249057181179523, + 0.010391277261078358, + 0.016942359507083893, + 0.010500057600438595, + 0.008388880640268326, + 0.021976126357913017, + 0.0341765433549881, + 0.0172768272459507, + 0.013513102196156979, + 0.014383324421942234, + 0.012117529287934303, + 0.01069621928036213, + 0.007034527137875557, + 0.019761184230446815, + 0.02916887030005455, + 0.02189910039305687, + 0.025963081046938896, + 0.02163040265440941, + 0.024441083893179893, + 0.028655653819441795, + 0.0184080321341753, + 0.02320055477321148, + 0.021932588890194893, + 0.026681844145059586, + 0.01976298727095127, + 0.01488738227635622, + 0.017407143488526344, + 0.017759667709469795, + 0.011494111269712448, + 0.0026264162734150887, + 0.0034569143317639828, + 0.004534495994448662, + 0.003936115652322769, + 0.0025657631922513247, + 0.0027715731412172318, + 0.0027695726603269577, + 0.001572535140439868, + 0.0027014086954295635, + 0.003403052920475602, + 0.0040404414758086205, + 0.00289898575283587, + 0.0022319175768643618, + 0.0028719331603497267, + 0.0015357097145169973, + 0.0013110910076647997, + 0.0037439996376633644, + 0.0034066883381456137, + 0.003945508971810341, + 0.0026447055861353874, + 0.0014706209767609835, + 0.0018477042904123664, + 0.002250165678560734, + 0.0014526405138894916, + 0.001930525409989059, + 0.00184995390009135, + 0.002425698796287179, + 0.0015105607453733683, + 0.0017562354914844036, + 0.0016522012883797288, + 0.0011595947435125709, + 0.0008686092915013433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.7578389048576355 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00675359508022666 + ], + "opposite_logprob_tokens": [ + -5.006753444671631 + ], + "logprob_reward": -0.00675359508022666, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00675359508022666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25301802158355713, + 0.17148847877979279, + 0.12341688573360443, + 0.07044053822755814, + 0.14581017196178436, + 0.11345372349023819, + 0.03716818988323212, + 0.027959203347563744, + 0.02268129400908947, + 0.03456345573067665 + ], + "credit_map": [ + 0.05881468579173088, + 0.021065419539809227, + 0.03303594887256622, + 0.01907925494015217, + 0.01840655878186226, + 0.0673568993806839, + 0.016681669279932976, + 0.018577614799141884, + 0.022870279848575592, + 0.03283644840121269, + 0.020223574712872505, + 0.02010626159608364, + 0.02177741751074791, + 0.02013476751744747, + 0.020158806815743446, + 0.013380917720496655, + 0.01760842464864254, + 0.021810257807374, + 0.026803236454725266, + 0.013586724177002907, + 0.014690743759274483, + 0.009281419217586517, + 0.010425273329019547, + 0.009210801683366299, + 0.010987806133925915, + 0.013233195059001446, + 0.009467212483286858, + 0.00823422335088253, + 0.008352985605597496, + 0.007261284627020359, + 0.006900366861373186, + 0.00600346690043807, + 0.016120973974466324, + 0.01810898259282112, + 0.021337782964110374, + 0.023479530587792397, + 0.01643161103129387, + 0.017722411081194878, + 0.01953412964940071, + 0.013074749149382114, + 0.018034445121884346, + 0.018405446782708168, + 0.01899833045899868, + 0.010447109118103981, + 0.012051030993461609, + 0.0159464031457901, + 0.011343901976943016, + 0.008227058686316013, + 0.004490883555263281, + 0.0063773831352591515, + 0.006799963768571615, + 0.00524301640689373, + 0.0034828800708055496, + 0.00431023258715868, + 0.004200956784188747, + 0.0022628721781075, + 0.0030476017855107784, + 0.0033502336591482162, + 0.004305391106754541, + 0.004578376188874245, + 0.0033089653588831425, + 0.0030371316242963076, + 0.004357532598078251, + 0.0019739698618650436, + 0.003737532999366522, + 0.003422039793804288, + 0.004241968970745802, + 0.0032428111881017685, + 0.0022793682292103767, + 0.0020404288079589605, + 0.002397488569840789, + 0.0013196554500609636, + 0.005276838317513466, + 0.006864749826490879, + 0.005575350485742092, + 0.0040757679380476475, + 0.004469582345336676, + 0.0031491615809500217, + 0.003467036411166191, + 0.0016849677776917815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00675359508022666 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420799270272255 + ], + "opposite_logprob_tokens": [ + -4.26420783996582 + ], + "logprob_reward": -0.01420799270272255, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420799270272255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16786116361618042, + 0.0965343713760376, + 0.05273747816681862, + 0.09340377151966095, + 0.23095187544822693, + 0.18101540207862854, + 0.03707418218255043, + 0.05775073915719986, + 0.04067490994930267, + 0.041996102780103683 + ], + "credit_map": [ + 0.024247607216238976, + 0.02181832678616047, + 0.026167213916778564, + 0.01526473369449377, + 0.026468511670827866, + 0.025340644642710686, + 0.014751250855624676, + 0.013802863657474518, + 0.011570265516638756, + 0.017883431166410446, + 0.01247516181319952, + 0.010703218169510365, + 0.015248698182404041, + 0.010936371050775051, + 0.010346646420657635, + 0.007370578125119209, + 0.006872850935906172, + 0.010666840709745884, + 0.00685696629807353, + 0.005063124932348728, + 0.008683759719133377, + 0.0057249669916927814, + 0.005113224498927593, + 0.00375573942437768, + 0.010877293534576893, + 0.013303680345416069, + 0.010610858909785748, + 0.010990536771714687, + 0.013897594064474106, + 0.012342553585767746, + 0.010745877400040627, + 0.010635384358465672, + 0.023500949144363403, + 0.032028939574956894, + 0.0324266143143177, + 0.036502499133348465, + 0.02912604995071888, + 0.027770061045885086, + 0.029076265171170235, + 0.02052050083875656, + 0.029329342767596245, + 0.027987301349639893, + 0.029689667746424675, + 0.018347054719924927, + 0.019021490588784218, + 0.02421385608613491, + 0.018624436110258102, + 0.013802256435155869, + 0.005063262768089771, + 0.007078532129526138, + 0.005502327345311642, + 0.0044561708346009254, + 0.004393992479890585, + 0.004628407768905163, + 0.0035606813617050648, + 0.0023908058647066355, + 0.005909368861466646, + 0.008210609667003155, + 0.009493328630924225, + 0.008655964396893978, + 0.00737919844686985, + 0.007026237435638905, + 0.006808415520936251, + 0.004267612472176552, + 0.0036766519770026207, + 0.005283107981085777, + 0.007048790343105793, + 0.0071353912353515625, + 0.004385269246995449, + 0.005113060586154461, + 0.005157482344657183, + 0.002875156467780471, + 0.005342280492186546, + 0.006235575769096613, + 0.0070089478977024555, + 0.0035547888837754726, + 0.0073958938010036945, + 0.004626513458788395, + 0.004760559182614088, + 0.003071543760597706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.01420799270272255 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0790020003914833 + ], + "opposite_logprob_tokens": [ + -2.5790019035339355 + ], + "logprob_reward": -0.0790020003914833, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0790020003914833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19042515754699707, + 0.1135682538151741, + 0.11766399443149567, + 0.15233035385608673, + 0.08294752240180969, + 0.10044339299201965, + 0.07102856785058975, + 0.09429850429296494, + 0.04697629436850548, + 0.030317997559905052 + ], + "credit_map": [ + 0.019995205104351044, + 0.02573392726480961, + 0.028929779306054115, + 0.016025841236114502, + 0.02795804850757122, + 0.03882838040590286, + 0.019143253564834595, + 0.013810706324875355, + 0.01375556644052267, + 0.01819121651351452, + 0.017247406765818596, + 0.017489269375801086, + 0.01071381289511919, + 0.017001664265990257, + 0.010204941965639591, + 0.008964374661445618, + 0.01663270778954029, + 0.024472128599882126, + 0.020752770826220512, + 0.015991996973752975, + 0.011415808461606503, + 0.008654623292386532, + 0.010269377380609512, + 0.00947456993162632, + 0.017127128317952156, + 0.021415745839476585, + 0.022751303389668465, + 0.017700431868433952, + 0.020284200087189674, + 0.022551128640770912, + 0.019684216007590294, + 0.010816186666488647, + 0.008396441116929054, + 0.010143556632101536, + 0.009306637570261955, + 0.009799507446587086, + 0.008826205506920815, + 0.014874614775180817, + 0.013175352476537228, + 0.008425201289355755, + 0.01684279926121235, + 0.015492110513150692, + 0.018477119505405426, + 0.011051450856029987, + 0.009766464121639729, + 0.012448198162019253, + 0.009692707099020481, + 0.006672534625977278, + 0.008299209177494049, + 0.010649497620761395, + 0.012486393563449383, + 0.009749175049364567, + 0.007662417367100716, + 0.008527384139597416, + 0.00912731233984232, + 0.004527178592979908, + 0.009104078635573387, + 0.013109703548252583, + 0.015690814703702927, + 0.014818784780800343, + 0.009597809053957462, + 0.012462998740375042, + 0.012153672985732555, + 0.007360635790973902, + 0.005597144830971956, + 0.006552319973707199, + 0.008812524378299713, + 0.007959266193211079, + 0.005020194686949253, + 0.003955424763262272, + 0.005844070576131344, + 0.0032353478018194437, + 0.0038037437479943037, + 0.0051102228462696075, + 0.004747761879116297, + 0.003929437603801489, + 0.004575105383992195, + 0.0031752081122249365, + 0.002713192719966173, + 0.002263322938233614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.0790020003914833 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping wide shot begins, the camera slowly and steadily panning from left to right across the tranquil surface of a lake at sunrise. A soft mist hovers above the water, capturing the sun's first rays and painting the horizon in a stunning gradient of soft pink, warm orange, and deep purple, all perfectly reflected in the still water. The camera's movement pauses, focusing on a weathered wooden dock where an elderly couple sits. The man, with silver-white hair and friendly wrinkles etched around his eyes, wears a faded plaid flannel shirt over a practical khaki fishing vest. He sits on a simple wooden box, gripping an old-fashioned fishing rod with a steady, practiced hand. His wife sits beside him, their shoulders almost touching, her soft gray hair pulled back in a loose bun, a cozy cream-colored knitted shawl wrapped around her shoulders. She also holds a fishing rod, her posture relaxed and her expression calm. Their lines disappear into the water, their buoys bobbing motionlessly, the only distraction from the perfect reflection. Their shared silence was not empty but fulfilling, a comfortable quiet that spoke of a lifetime together, creating an overall atmosphere of profound peace, nostalgia and tranquility, of unspoken love.", + "seed": 5516, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_14/index_6545_seed_5516_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046717597544193266, + "vlm_logprob": -0.046717597544193266, + "vlm_yes_no_margin": 3.549999952316284, + "vlm_token_credit": -0.046717597544193266, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019691573455929756 + ], + "opposite_logprob_tokens": [ + -4.144691467285156 + ], + "logprob_reward": -0.019691573455929756, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019691573455929756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9388445615768433, + 0.016207728534936905, + 0.011233112774789333, + 0.003713038517162204, + 0.0033110983204096556, + 0.01830867864191532, + 0.002551690209656954, + 0.001256307354196906, + 0.0012224804377183318, + 0.003351360559463501 + ], + "credit_map": [ + 0.012994307093322277, + 0.09894135594367981, + 0.012084143236279488, + 0.006115548312664032, + 0.44411352276802063, + 0.34662339091300964, + 0.009238885715603828, + 0.008733350783586502, + 0.001591639593243599, + 0.001740075764246285, + 0.0008708779350854456, + 0.0014145325403660536, + 0.002897438360378146, + 0.00290269055403769, + 0.0029354491271078587, + 0.0018550253007560968, + 0.001258881762623787, + 0.001142311841249466, + 0.0018271035514771938, + 0.0014757591998204589, + 0.0014781382633373141, + 0.0014626276679337025, + 0.0013557735364884138, + 0.001232517184689641, + 0.0003383474249858409, + 0.0005024957354180515, + 0.00044243663433007896, + 0.00037705403519794345, + 0.00040920934407040477, + 0.0004989710869267583, + 0.0004643409338314086, + 0.0006801831186749041, + 0.00028576774639077485, + 0.00038163995486684144, + 0.0004333504184614867, + 0.0005122231668792665, + 0.00033695230376906693, + 0.0003408646152820438, + 0.0005095493397675455, + 0.0005107506294734776, + 0.002274814061820507, + 0.002057693200185895, + 0.00303508504293859, + 0.0031072725541889668, + 0.0017994024092331529, + 0.0021303477697074413, + 0.0014123882865533233, + 0.002491674153134227, + 0.00033143319888040423, + 0.00018933524552267045, + 0.00027276057517156005, + 0.00037407525815069675, + 0.0004446670354809612, + 0.00030876719392836094, + 0.00025954062584787607, + 0.00037111129495315254, + 0.00012907915515825152, + 0.00018764578271657228, + 0.00017141138960141689, + 0.00016238454554695636, + 0.0002243493654532358, + 0.00016561200027354062, + 9.438951383344829e-05, + 0.00012143552157795057, + 0.0001544955011922866, + 0.0001274406531592831, + 0.00025832466781139374, + 0.00012960129242856055, + 0.00018112068937625736, + 0.0001524015242466703, + 8.085039735306054e-05, + 0.00013824574125465006, + 0.0004995462368242443, + 0.0003062690666411072, + 0.0006458500283770263, + 0.0004848521784879267, + 0.0003165322996210307, + 0.00040318863466382027, + 0.0002881825785152614, + 0.0004069396818522364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.019691573455929756 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308326914906502 + ], + "opposite_logprob_tokens": [ + -3.1680831909179688 + ], + "logprob_reward": -0.04308326914906502, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308326914906502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.9007712602615356, + 0.014357272535562515, + 0.018063902854919434, + 0.012078669853508472, + 0.023275557905435562, + 0.01621825620532036, + 0.006657919380813837, + 0.002203191863372922, + 0.001777504337951541, + 0.004596504848450422 + ], + "credit_map": [ + 0.01240764930844307, + 0.09493693709373474, + 0.01176078524440527, + 0.0057740905322134495, + 0.42529863119125366, + 0.3326469957828522, + 0.009116560220718384, + 0.008829604834318161, + 0.0016938430489972234, + 0.0017356389435008168, + 0.0011878485092893243, + 0.0021722412202507257, + 0.0018169658724218607, + 0.001801415579393506, + 0.002193751744925976, + 0.0017555673839524388, + 0.0017483182018622756, + 0.0018880044808611274, + 0.002924703760072589, + 0.0028718020766973495, + 0.002477804198861122, + 0.0024190410040318966, + 0.0016345748445019126, + 0.0020996560342609882, + 0.0008866112912073731, + 0.0017982465215027332, + 0.0011855961056426167, + 0.0013084716629236937, + 0.0012941741151735187, + 0.0015735735651105642, + 0.0020281565375626087, + 0.002003840636461973, + 0.0019218864617869258, + 0.0029265235643833876, + 0.0033678202889859676, + 0.0033124787732958794, + 0.0024946797639131546, + 0.0023439109791070223, + 0.0036271095741540194, + 0.003281146753579378, + 0.0018216416938230395, + 0.002057368867099285, + 0.0026629639323800802, + 0.0024971850216388702, + 0.0021747101563960314, + 0.0020095156505703926, + 0.001273059519007802, + 0.0017218113644048572, + 0.0007347743376158178, + 0.000712669687345624, + 0.0009382055141031742, + 0.0010869429679587483, + 0.0010096306214109063, + 0.000764392432756722, + 0.0006628374103456736, + 0.00074846635106951, + 0.00021971283422317356, + 0.0002460382238496095, + 0.00034631314338184893, + 0.0003007168124895543, + 0.0003518813755363226, + 0.00024486545589752495, + 0.0002140953583875671, + 0.00027956857229582965, + 0.0002106891042785719, + 0.00021261100482661277, + 0.00021501035371329635, + 0.00023440140648745, + 0.0002585153852123767, + 0.00020865850092377514, + 0.00022266010637395084, + 0.00021495838882401586, + 0.0006316475919447839, + 0.00038998841773718596, + 0.0009109913371503353, + 0.0005920412950217724, + 0.0005237246514298022, + 0.0005789943388663232, + 0.0004082322120666504, + 0.000560884946025908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.04308326914906502 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7967904210090637, + 0.02719295211136341, + 0.04406877979636192, + 0.016366159543395042, + 0.023801570758223534, + 0.04419244825839996, + 0.010012567043304443, + 0.01520884782075882, + 0.011614952236413956, + 0.010751353576779366 + ], + "credit_map": [ + 0.011292765848338604, + 0.08439885079860687, + 0.010314792394638062, + 0.007942333817481995, + 0.3745204210281372, + 0.29189014434814453, + 0.008294142782688141, + 0.008136927150189877, + 0.0027169915847480297, + 0.0028192405588924885, + 0.002060606610029936, + 0.004017597530037165, + 0.004311331547796726, + 0.004329134710133076, + 0.003924030344933271, + 0.003014018526300788, + 0.0018473081290721893, + 0.002573278034105897, + 0.0036369513254612684, + 0.004918825812637806, + 0.012647769413888454, + 0.012570948339998722, + 0.0027035358361899853, + 0.003170162672176957, + 0.0020774162840098143, + 0.002694233087822795, + 0.001549074542708695, + 0.001763384323567152, + 0.0018383808201178908, + 0.0022595226764678955, + 0.0018491531955078244, + 0.0023349944967776537, + 0.0027622240595519543, + 0.0034889504313468933, + 0.0038056049961596727, + 0.002984063932672143, + 0.0022392594255506992, + 0.0026401884388178587, + 0.0026963206473737955, + 0.0031849590595811605, + 0.005462328437715769, + 0.004288955591619015, + 0.007460554596036673, + 0.007583411410450935, + 0.005183037370443344, + 0.005228039342910051, + 0.0034451792016625404, + 0.005540942307561636, + 0.00112244114279747, + 0.0011099260300397873, + 0.0012226504040881991, + 0.0015719002112746239, + 0.0013709012418985367, + 0.001089831697754562, + 0.0010855701984837651, + 0.0014393458841368556, + 0.0017178626731038094, + 0.0016178643563762307, + 0.0019627504516392946, + 0.0021956139244139194, + 0.0023302887566387653, + 0.0017518597887828946, + 0.0017703075427561998, + 0.0018623006762936711, + 0.0015381249831989408, + 0.0010264954762533307, + 0.0023416040930896997, + 0.0015289988368749619, + 0.001245883060619235, + 0.0013504857197403908, + 0.0011303676292300224, + 0.001452991971746087, + 0.001282107550650835, + 0.0011519964318722486, + 0.0014801444485783577, + 0.0016865599900484085, + 0.0013144606491550803, + 0.0012611685087904334, + 0.00124546408187598, + 0.0013294520322233438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819479651749134 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.01819479651749134, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819479651749134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6449496746063232, + 0.0348985381424427, + 0.04960596561431885, + 0.027592144906520844, + 0.024891002103686333, + 0.028927208855748177, + 0.036295946687459946, + 0.08018758147954941, + 0.03325425460934639, + 0.03939766064286232 + ], + "credit_map": [ + 0.00935844611376524, + 0.0702032744884491, + 0.008571013808250427, + 0.008715679869055748, + 0.30160582065582275, + 0.23239952325820923, + 0.007674172520637512, + 0.006421794183552265, + 0.004199696704745293, + 0.0039682150818407536, + 0.003959068562835455, + 0.0036653587594628334, + 0.004510846454650164, + 0.004599354229867458, + 0.005334517918527126, + 0.004661480896174908, + 0.002514228457584977, + 0.0035381128545850515, + 0.00551706925034523, + 0.006467153318226337, + 0.011913718655705452, + 0.011777547188103199, + 0.003707119496539235, + 0.004171018023043871, + 0.0021960020530968904, + 0.00477451691403985, + 0.0026304172351956367, + 0.0035918690264225006, + 0.0033535179682075977, + 0.0034110378473997116, + 0.00409697787836194, + 0.0035378059837967157, + 0.002208977472037077, + 0.0032849095296114683, + 0.003250890178605914, + 0.003484141081571579, + 0.0030731356237083673, + 0.00269892206415534, + 0.003414619481191039, + 0.003475406439974904, + 0.0023195133544504642, + 0.004503611940890551, + 0.0034266451839357615, + 0.0025458605960011482, + 0.006589486729353666, + 0.004576769191771746, + 0.002133813686668873, + 0.002831509802490473, + 0.003142553847283125, + 0.0036364723928272724, + 0.003518724115565419, + 0.0060034883208572865, + 0.006683371961116791, + 0.0038963397964835167, + 0.00461731618270278, + 0.004797680769115686, + 0.007878917269408703, + 0.01013265922665596, + 0.0082766804844141, + 0.010167288593947887, + 0.014736460521817207, + 0.009385267272591591, + 0.010107351467013359, + 0.009502953849732876, + 0.0035874987952411175, + 0.003772497409954667, + 0.004699091427028179, + 0.004604747518897057, + 0.004716556053608656, + 0.003842610865831375, + 0.004094404634088278, + 0.003936849534511566, + 0.0036874706856906414, + 0.004475702997297049, + 0.004359075333923101, + 0.005456941202282906, + 0.006696832831948996, + 0.0043302131816744804, + 0.005569512024521828, + 0.004821913316845894 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.01819479651749134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a sweeping wide shot of a lake at sunrise?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pan from left to right across the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the horizon painted with a gradient of soft pink, warm orange, and deep purple reflected in the still water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show an elderly couple sitting on a weathered wooden dock, with the man wearing a faded plaid flannel shirt and khaki fishing vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are both the man and woman holding fishing rods with their lines in the water and buoys bobbing motionlessly?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282608032226562 + ], + "opposite_logprob_tokens": [ + -2.0178260803222656 + ], + "logprob_reward": -0.14282608032226562, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14282608032226562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8491464853286743, + 0.013954664580523968, + 0.01322848442941904, + 0.014192254282534122, + 0.014474686235189438, + 0.015293990261852741, + 0.011639799922704697, + 0.019016267731785774, + 0.021672701463103294, + 0.02738068625330925 + ], + "credit_map": [ + 0.011692323721945286, + 0.09047563374042511, + 0.010777194984257221, + 0.005339961498975754, + 0.40116968750953674, + 0.3126710057258606, + 0.008944163098931313, + 0.008076454512774944, + 0.001414048601873219, + 0.0013283235020935535, + 0.0011701168259605765, + 0.0013234317302703857, + 0.002526756376028061, + 0.002666341606527567, + 0.001944560557603836, + 0.0015810861950740218, + 0.0006601039785891771, + 0.0008412533206865191, + 0.000878094753716141, + 0.0012082763714715838, + 0.0037608256097882986, + 0.0036796180065721273, + 0.001232959097251296, + 0.0009673532913438976, + 0.001386525109410286, + 0.0023346696980297565, + 0.0015198851469904184, + 0.0012486095074564219, + 0.0012388416798785329, + 0.0020429580472409725, + 0.001813389710150659, + 0.002607375616207719, + 0.001271669752895832, + 0.0016748307971283793, + 0.0025294797960668802, + 0.0017938101664185524, + 0.0017956632655113935, + 0.0015157575253397226, + 0.0019493034342303872, + 0.0019441721960902214, + 0.0011477823136374354, + 0.0015213759616017342, + 0.002149381674826145, + 0.0022749099880456924, + 0.0027313693426549435, + 0.0018337280489504337, + 0.0012458089040592313, + 0.0023896326310932636, + 0.001142070977948606, + 0.0008751422865316272, + 0.001587232225574553, + 0.0018487039487808943, + 0.0017305511282756925, + 0.0012841031420975924, + 0.0012577768648043275, + 0.0019142191158607602, + 0.0018623850774019957, + 0.0021541810128837824, + 0.0019671148620545864, + 0.0026420343201607466, + 0.0032081040553748608, + 0.0021889160852879286, + 0.002665068255737424, + 0.0023284642957150936, + 0.002312067896127701, + 0.0024617912713438272, + 0.0028707466553896666, + 0.0029468846041709185, + 0.003320209216326475, + 0.0024161087349057198, + 0.0026343963108956814, + 0.002710496773943305, + 0.003082734066992998, + 0.0027345730923116207, + 0.0034099221229553223, + 0.004138469230383635, + 0.003743847832083702, + 0.003118828171864152, + 0.0034187359269708395, + 0.003733575576916337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.14282608032226562 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_tensor([8181, 8181])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_tensor([8181, 8181])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..8779acdfe12d1b9480800a8145615bce9ab2f0f8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_tensor([8181, 8181])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a static, eye-level medium shot that captures the austere aesthetic of a historical drama or fantasy film, two East Asian men with dark, slightly disheveled hair are pressed against a cold, rough, gray stone wall, their torsos bound by thick ropes. Over their clothing, they both wear ornate, armor-like studded leather vests; the one on the left wears a high-collared white tunic, while the one on the right wears simple brown clothing. Despite their predicament, their facial expressions remain deliberately neutral and stoic, their gazes fixed calmly ahead. The scene is steeped in tension, illuminated by a single, low, off-camera light source that casts long, flickering shadows and creates a dramatic black-and-white contrast, highlighting the texture of the stone and the metallic glint of the armor. Heavy, dark wooden beams are barely visible against the oppressive darkness overhead. At the far left of the frame, a third figure wearing a heavy black scapula is visible on his shoulders, suggesting an ominous captive presence and heightened feelings of imprisonment and imminent conflict.", + "seed": 26290, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_8181_seed_26290_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.035890837851911786, + "vlm_logprob": -0.035890837851911786, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.035890837851911786, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06457433849573135 + ], + "opposite_logprob_tokens": [ + -2.8145742416381836 + ], + "logprob_reward": -0.06457433849573135, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06457433849573135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3358975648880005, + 0.09928705543279648, + 0.13782095909118652, + 0.06743737310171127, + 0.08681374043226242, + 0.07550214976072311, + 0.03256755322217941, + 0.06646803021430969, + 0.05143330991268158, + 0.046772345900535583 + ], + "credit_map": [ + 0.03823855146765709, + 0.03706056624650955, + 0.04380137100815773, + 0.033809807151556015, + 0.05198227986693382, + 0.04345206543803215, + 0.04341132566332817, + 0.04414160177111626, + 0.01124807633459568, + 0.010130270384252071, + 0.013174546882510185, + 0.010044630616903305, + 0.014604614116251469, + 0.012468005530536175, + 0.015715070068836212, + 0.011901839636266232, + 0.021752385422587395, + 0.010960270650684834, + 0.020471924915909767, + 0.012443525716662407, + 0.01861882209777832, + 0.019285773858428, + 0.015928007662296295, + 0.018360238522291183, + 0.009415585547685623, + 0.009404311887919903, + 0.011827188543975353, + 0.006278553511947393, + 0.005607826169580221, + 0.009912717156112194, + 0.006117969751358032, + 0.008873219601809978, + 0.014506545849144459, + 0.008661511354148388, + 0.014781360514461994, + 0.007516956888139248, + 0.010117108933627605, + 0.012506721541285515, + 0.00737230246886611, + 0.011351230554282665, + 0.010425196960568428, + 0.00962182879447937, + 0.011653920635581017, + 0.006639352533966303, + 0.005895192269235849, + 0.015396380797028542, + 0.005236943252384663, + 0.010633337311446667, + 0.004589509684592485, + 0.0038546635769307613, + 0.004123465623706579, + 0.002827434567734599, + 0.0032908059656620026, + 0.006096724420785904, + 0.003909478895366192, + 0.00387546862475574, + 0.01085888221859932, + 0.004588145297020674, + 0.01145834568887949, + 0.00577531335875392, + 0.009498345665633678, + 0.008516903035342693, + 0.00681352661922574, + 0.008958566002547741, + 0.008271823637187481, + 0.005851924419403076, + 0.007891624234616756, + 0.003993268124759197, + 0.006182800978422165, + 0.007073653861880302, + 0.006247455254197121, + 0.00592075427994132, + 0.007426399737596512, + 0.004893699195235968, + 0.007107086014002562, + 0.0038206602912396193, + 0.004668566398322582, + 0.00674083549529314, + 0.006379584316164255, + 0.005735516082495451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06457433849573135 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009808794595301151 + ], + "opposite_logprob_tokens": [ + -4.6348090171813965 + ], + "logprob_reward": -0.009808794595301151, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009808794595301151, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32464396953582764, + 0.14961422979831696, + 0.12105325609445572, + 0.049057867377996445, + 0.07440975308418274, + 0.06848811358213425, + 0.05751219391822815, + 0.059996046125888824, + 0.047436390072107315, + 0.047788165509700775 + ], + "credit_map": [ + 0.0406845398247242, + 0.03280385583639145, + 0.04423587769269943, + 0.03584463521838188, + 0.048668138682842255, + 0.04173438996076584, + 0.03732539340853691, + 0.043347135186195374, + 0.018841203302145004, + 0.01866275630891323, + 0.019695013761520386, + 0.01649211160838604, + 0.02030307427048683, + 0.017867641523480415, + 0.016335181891918182, + 0.021417247131466866, + 0.01707279123365879, + 0.009286709129810333, + 0.01416825596243143, + 0.011128900572657585, + 0.01907988265156746, + 0.017906732857227325, + 0.016528362408280373, + 0.01588161662220955, + 0.005923095159232616, + 0.004540205001831055, + 0.0049570775590837, + 0.004363031592220068, + 0.008212312124669552, + 0.008358916267752647, + 0.006060864310711622, + 0.006642367225140333, + 0.011200441047549248, + 0.005874214228242636, + 0.007845893502235413, + 0.005598367657512426, + 0.009087232872843742, + 0.017703287303447723, + 0.007097302004694939, + 0.010003017261624336, + 0.00835941918194294, + 0.00915397610515356, + 0.010114372707903385, + 0.006372067611664534, + 0.007757139392197132, + 0.010642070323228836, + 0.00716158002614975, + 0.008927486836910248, + 0.007887275889515877, + 0.006865010131150484, + 0.006796532776206732, + 0.0056269229389727116, + 0.0063347225077450275, + 0.010452181100845337, + 0.004681391641497612, + 0.008868158794939518, + 0.00932486541569233, + 0.003989714197814465, + 0.00854205060750246, + 0.00443787407130003, + 0.008425009436905384, + 0.010894889011979103, + 0.006260975729674101, + 0.008120669983327389, + 0.006338976789265871, + 0.006433340720832348, + 0.006480028852820396, + 0.004945262800902128, + 0.005357328802347183, + 0.007594672031700611, + 0.004147677682340145, + 0.006139102391898632, + 0.006392645183950663, + 0.004314834251999855, + 0.007057786453515291, + 0.004038393963128328, + 0.006687190383672714, + 0.0077153416350483894, + 0.004778216127306223, + 0.006803758908063173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009808794595301151 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980896644294262 + ], + "opposite_logprob_tokens": [ + -3.52980899810791 + ], + "logprob_reward": -0.02980896644294262, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980896644294262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33546462655067444, + 0.08458010107278824, + 0.06788064539432526, + 0.07954145967960358, + 0.10916494578123093, + 0.1232052743434906, + 0.04515191167593002, + 0.06621634215116501, + 0.04337679222226143, + 0.045417848974466324 + ], + "credit_map": [ + 0.04488798603415489, + 0.040994346141815186, + 0.0378626249730587, + 0.03320593386888504, + 0.04815678671002388, + 0.04278610646724701, + 0.04818795993924141, + 0.039382897317409515, + 0.013587280176579952, + 0.007785655092447996, + 0.013373986817896366, + 0.009319479577243328, + 0.008743553422391415, + 0.01033091451972723, + 0.008851837366819382, + 0.012587389908730984, + 0.010241840966045856, + 0.006613315083086491, + 0.009359942749142647, + 0.006620914675295353, + 0.006562049966305494, + 0.010472544468939304, + 0.009237601421773434, + 0.008772439323365688, + 0.01124881487339735, + 0.006735284347087145, + 0.007106204051524401, + 0.007623860612511635, + 0.01370791345834732, + 0.011234313249588013, + 0.01231716014444828, + 0.009567905217409134, + 0.01717635802924633, + 0.009698180481791496, + 0.016603540629148483, + 0.008717023767530918, + 0.010994166135787964, + 0.021347777917981148, + 0.011283356696367264, + 0.01334453746676445, + 0.016643740236759186, + 0.01612401008605957, + 0.016432275995612144, + 0.011397586204111576, + 0.011051283217966557, + 0.025500984862446785, + 0.009176438674330711, + 0.01687895879149437, + 0.005761648993939161, + 0.0035391023848205805, + 0.004938528873026371, + 0.0036559016443789005, + 0.0073613328859210014, + 0.008904770016670227, + 0.005536714103072882, + 0.005453911609947681, + 0.009280269965529442, + 0.0043601589277386665, + 0.011032219044864178, + 0.005266755819320679, + 0.0071919746696949005, + 0.01383605320006609, + 0.006308425683528185, + 0.008940487168729305, + 0.005553266499191523, + 0.0046021402813494205, + 0.006669862661510706, + 0.0034375854302197695, + 0.00513040367513895, + 0.008218158967792988, + 0.004707491025328636, + 0.005057886242866516, + 0.006188266444951296, + 0.00506320595741272, + 0.008149145171046257, + 0.0034181775990873575, + 0.004350484348833561, + 0.006660895422101021, + 0.0057842424139380455, + 0.00580343184992671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.02980896644294262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1888274848461151, + 0.10372412949800491, + 0.12665526568889618, + 0.063911572098732, + 0.057822663336992264, + 0.2621643543243408, + 0.024760644882917404, + 0.09983779489994049, + 0.029688669368624687, + 0.04260743781924248 + ], + "credit_map": [ + 0.025424785912036896, + 0.02771216444671154, + 0.02876349352300167, + 0.020497532561421394, + 0.01680341176688671, + 0.025660919025540352, + 0.020293405279517174, + 0.023671746253967285, + 0.013690261170268059, + 0.015559764578938484, + 0.014151004143059254, + 0.010306699201464653, + 0.008630341850221157, + 0.019092442467808723, + 0.00854788813740015, + 0.01374572329223156, + 0.024511141702532768, + 0.010750540532171726, + 0.017635567113757133, + 0.009001152589917183, + 0.015368481166660786, + 0.022306356579065323, + 0.01251241099089384, + 0.0145696010440588, + 0.010106096044182777, + 0.007486400194466114, + 0.01023794710636139, + 0.005871335510164499, + 0.0066125704906880856, + 0.00873515848070383, + 0.006544912233948708, + 0.008317151106894016, + 0.0077707041054964066, + 0.0076054842211306095, + 0.009006927721202374, + 0.005130260717123747, + 0.00574796088039875, + 0.009427007287740707, + 0.005451385397464037, + 0.00768292834982276, + 0.04077150300145149, + 0.030895425006747246, + 0.03507275879383087, + 0.019652318209409714, + 0.021785341203212738, + 0.06126348674297333, + 0.017168544232845306, + 0.03555496037006378, + 0.003961875103414059, + 0.0024881712161004543, + 0.004828059580177069, + 0.0021915200632065535, + 0.0028736356180161238, + 0.002469897037371993, + 0.0026829245034605265, + 0.0032645580358803272, + 0.01889977604150772, + 0.0090621467679739, + 0.0098011065274477, + 0.006857362575829029, + 0.010882312431931496, + 0.022470595315098763, + 0.008636589162051678, + 0.01322789303958416, + 0.004381225910037756, + 0.003976806066930294, + 0.003318182425573468, + 0.002489829435944557, + 0.0025282602291554213, + 0.006931559648364782, + 0.002407422289252281, + 0.003655384061858058, + 0.007364432327449322, + 0.004379525780677795, + 0.007542109116911888, + 0.0034653027541935444, + 0.004166752099990845, + 0.006245703902095556, + 0.003418727545067668, + 0.006024880800396204 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.005256166215986013 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000592350959778 + ], + "opposite_logprob_tokens": [ + -2.6950058937072754 + ], + "logprob_reward": -0.07000592350959778, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000592350959778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3142223060131073, + 0.13664066791534424, + 0.07996054738759995, + 0.04698508605360985, + 0.15434499084949493, + 0.06014561280608177, + 0.05002405121922493, + 0.07597102969884872, + 0.03508061170578003, + 0.04662511870265007 + ], + "credit_map": [ + 0.03965761885046959, + 0.03982292488217354, + 0.04577408358454704, + 0.0291590578854084, + 0.038309574127197266, + 0.04745299369096756, + 0.03827411308884621, + 0.035771917551755905, + 0.01996874436736107, + 0.02097720466554165, + 0.020071731880307198, + 0.014542662538588047, + 0.01460171490907669, + 0.016260800883173943, + 0.014038504101336002, + 0.01617930829524994, + 0.010586672462522984, + 0.01128317229449749, + 0.011349053122103214, + 0.008244281634688377, + 0.009413298219442368, + 0.010883156210184097, + 0.007674688007682562, + 0.010526220314204693, + 0.0063387732952833176, + 0.00714251259341836, + 0.004957723896950483, + 0.004243099596351385, + 0.006131417118012905, + 0.005869600921869278, + 0.0057470062747597694, + 0.0065549518913030624, + 0.026527054607868195, + 0.016501929610967636, + 0.02399209514260292, + 0.012426361441612244, + 0.015583835542201996, + 0.025224076583981514, + 0.01079968549311161, + 0.023289944976568222, + 0.011106588877737522, + 0.0040100873447954655, + 0.00975283794105053, + 0.004927579313516617, + 0.007033005356788635, + 0.009258785285055637, + 0.005728966556489468, + 0.008327762596309185, + 0.007034374866634607, + 0.005859294440597296, + 0.007607465144246817, + 0.004368648398667574, + 0.004539465066045523, + 0.007716842461377382, + 0.004814433865249157, + 0.008083527907729149, + 0.011497315019369125, + 0.00840146653354168, + 0.0074173868633806705, + 0.006322116125375032, + 0.010478466749191284, + 0.01471721101552248, + 0.007355305831879377, + 0.009781762957572937, + 0.004263321869075298, + 0.005313693545758724, + 0.0038388443645089865, + 0.0034936298616230488, + 0.003860994940623641, + 0.006337209604680538, + 0.0033030735794454813, + 0.004669840447604656, + 0.005601933225989342, + 0.004962111357599497, + 0.004556084517389536, + 0.004223962314426899, + 0.007311617955565453, + 0.008145713247358799, + 0.005845475010573864, + 0.005978224333375692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.07000592350959778 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, eye-level medium shot that captures the austere aesthetic of a historical drama or fantasy film, two East Asian men with dark, slightly disheveled hair are pressed against a cold, rough, gray stone wall, their torsos bound by thick ropes. Over their clothing, they both wear ornate, armor-like studded leather vests; the one on the left wears a high-collared white tunic, while the one on the right wears simple brown clothing. Despite their predicament, their facial expressions remain deliberately neutral and stoic, their gazes fixed calmly ahead. The scene is steeped in tension, illuminated by a single, low, off-camera light source that casts long, flickering shadows and creates a dramatic black-and-white contrast, highlighting the texture of the stone and the metallic glint of the armor. Heavy, dark wooden beams are barely visible against the oppressive darkness overhead. At the far left of the frame, a third figure wearing a heavy black scapula is visible on his shoulders, suggesting an ominous captive presence and heightened feelings of imprisonment and imminent conflict.", + "seed": 26291, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_8181_seed_26291_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022386549180373548, + "vlm_logprob": -0.022386549180373548, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.022386549180373548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016158340498805046 + ], + "opposite_logprob_tokens": [ + -4.266158103942871 + ], + "logprob_reward": -0.016158340498805046, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.016158340498805046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7725473642349243, + 0.047672338783741, + 0.025622133165597916, + 0.02565377950668335, + 0.05994411185383797, + 0.012873105704784393, + 0.009139454923570156, + 0.015185331925749779, + 0.013298442587256432, + 0.01806393638253212 + ], + "credit_map": [ + 0.047918789088726044, + 0.029436536133289337, + 0.2746138870716095, + 0.028523970395326614, + 0.01711992919445038, + 0.0338224321603775, + 0.02802133746445179, + 0.313090443611145, + 0.0059063127264380455, + 0.008211047388613224, + 0.005433108191937208, + 0.005076418165117502, + 0.005417957436293364, + 0.004749442916363478, + 0.0062055522575974464, + 0.006672493647783995, + 0.0027703153900802135, + 0.002414585556834936, + 0.003236595541238785, + 0.003986442927271128, + 0.0017765738302841783, + 0.0033744843676686287, + 0.0040798247791826725, + 0.003983307629823685, + 0.0025093858130276203, + 0.0031859739683568478, + 0.003276524133980274, + 0.0028798195999115705, + 0.0025766342878341675, + 0.0038637963589280844, + 0.003037802642211318, + 0.004323840606957674, + 0.005773032549768686, + 0.008218628354370594, + 0.007826533168554306, + 0.01344689354300499, + 0.0055693634785711765, + 0.005259080324321985, + 0.007603095844388008, + 0.0062474822625517845, + 0.001172098214738071, + 0.002265495713800192, + 0.0014413559110835195, + 0.0015370320761576295, + 0.0015243013622239232, + 0.0018080342561006546, + 0.0013020222540944815, + 0.001822764752432704, + 0.0013947784900665283, + 0.0009712642640806735, + 0.001187217072583735, + 0.0008502937853336334, + 0.0008558695553801954, + 0.0010895133018493652, + 0.0010399733437225223, + 0.0017505446448922157, + 0.0013180101523175836, + 0.002087780274450779, + 0.001889989827759564, + 0.0016827168874442577, + 0.0025867277290672064, + 0.0017029341543093324, + 0.0016299127601087093, + 0.0022872579284012318, + 0.0013484667288139462, + 0.001238567172549665, + 0.001734023680910468, + 0.002161456272006035, + 0.0009090944658964872, + 0.0020384134259074926, + 0.0014340260531753302, + 0.0024343926925212145, + 0.0024634734727442265, + 0.0024563849437981844, + 0.0021016572136431932, + 0.0032686979975551367, + 0.0013527685077860951, + 0.0017691073007881641, + 0.0021001214627176523, + 0.0025517242029309273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.016158340498805046 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088653467595577 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088653467595577, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088653467595577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7028420567512512, + 0.04282277822494507, + 0.08642512559890747, + 0.02317112497985363, + 0.02590925432741642, + 0.03434114530682564, + 0.017203673720359802, + 0.010701880790293217, + 0.021514371037483215, + 0.03506859019398689 + ], + "credit_map": [ + 0.042799677699804306, + 0.028388774022459984, + 0.24916839599609375, + 0.024739937856793404, + 0.014911421574652195, + 0.03188435733318329, + 0.024485526606440544, + 0.2864639163017273, + 0.005759174469858408, + 0.009482664987444878, + 0.00463433051481843, + 0.004452395252883434, + 0.005602825433015823, + 0.004656639881432056, + 0.0039614480920135975, + 0.004273291677236557, + 0.011764195747673512, + 0.022577136754989624, + 0.009994493797421455, + 0.007736096624284983, + 0.013889007270336151, + 0.007016636431217194, + 0.005705089773982763, + 0.007742458023130894, + 0.003074162406846881, + 0.0036826194263994694, + 0.003551867790520191, + 0.0024844587314873934, + 0.00300692580640316, + 0.0021841267589479685, + 0.002637243829667568, + 0.0025497162714600563, + 0.0036937231197953224, + 0.0044023641385138035, + 0.0031431103125214577, + 0.0024010136257857084, + 0.0035665452014654875, + 0.0025439984165132046, + 0.0028562063816934824, + 0.003302290104329586, + 0.003456499194726348, + 0.0034050154499709606, + 0.004906635731458664, + 0.006534920539706945, + 0.00366999302059412, + 0.003136625047773123, + 0.004244011826813221, + 0.004987443331629038, + 0.002493086038157344, + 0.0037432066164910793, + 0.0017704218626022339, + 0.001473980606533587, + 0.0028810189105570316, + 0.0017206495394930243, + 0.0014655006816610694, + 0.0016558077186346054, + 0.0013857304584234953, + 0.0014693568227812648, + 0.0013690718915313482, + 0.0013802716275677085, + 0.0008426326676271856, + 0.0010609724558889866, + 0.0013023671926930547, + 0.0018914771499112248, + 0.003709220327436924, + 0.0033884625881910324, + 0.0032962241675704718, + 0.0018469465430825949, + 0.0019565701950341463, + 0.002495198743417859, + 0.002450925065204501, + 0.002370819915086031, + 0.004572151694446802, + 0.005224091000854969, + 0.003927299287170172, + 0.005162191577255726, + 0.0037086165975779295, + 0.0030715118627995253, + 0.00524357333779335, + 0.004159150645136833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.011088653467595577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078749783337116 + ], + "opposite_logprob_tokens": [ + -4.511078834533691 + ], + "logprob_reward": -0.011078749783337116, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078749783337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13776445388793945, + 0.0984640046954155, + 0.11531079560518265, + 0.06917460262775421, + 0.23594950139522552, + 0.16507390141487122, + 0.035726647824048996, + 0.04950453341007233, + 0.043977852910757065, + 0.049053773283958435 + ], + "credit_map": [ + 0.025002362206578255, + 0.00978394690901041, + 0.02507411129772663, + 0.012956294231116772, + 0.011623267084360123, + 0.013873009942471981, + 0.01173881534487009, + 0.027712631970643997, + 0.011106756515800953, + 0.01509256474673748, + 0.011143295094370842, + 0.008673379197716713, + 0.011383064091205597, + 0.009881681762635708, + 0.015365335159003735, + 0.015817921608686447, + 0.010829070582985878, + 0.013983890414237976, + 0.013686955906450748, + 0.02260616235435009, + 0.008624632842838764, + 0.010963089764118195, + 0.01628999039530754, + 0.018326997756958008, + 0.011501208879053593, + 0.007536275312304497, + 0.008743682876229286, + 0.007595053408294916, + 0.005421082954853773, + 0.007048605475574732, + 0.008463669568300247, + 0.012865019030869007, + 0.020469028502702713, + 0.032527267932891846, + 0.029229115694761276, + 0.04924362525343895, + 0.03318741172552109, + 0.018402259796857834, + 0.031046900898218155, + 0.021843887865543365, + 0.013367901556193829, + 0.02027261070907116, + 0.021237052977085114, + 0.031218720600008965, + 0.024091819301247597, + 0.013648895546793938, + 0.018520209938287735, + 0.02271667867898941, + 0.005119540728628635, + 0.0035295377019792795, + 0.0043494850397109985, + 0.0034464893396943808, + 0.003777160542085767, + 0.004125252366065979, + 0.003897150745615363, + 0.007482029497623444, + 0.007861682213842869, + 0.006886395625770092, + 0.005654902663081884, + 0.005175648257136345, + 0.006161063443869352, + 0.00572992442175746, + 0.0046648066490888596, + 0.007370105013251305, + 0.006784502882510424, + 0.006399848964065313, + 0.0048030707985162735, + 0.005268719047307968, + 0.00749592948704958, + 0.0038380054756999016, + 0.004143965430557728, + 0.005243810825049877, + 0.007171053905040026, + 0.005931273102760315, + 0.0059570553712546825, + 0.008005146868526936, + 0.003082835581153631, + 0.004546712152659893, + 0.006483415607362986, + 0.007876276038587093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.011078749783337116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036157488357275724 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036157488357275724, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036157488357275724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7828342318534851, + 0.03529378026723862, + 0.021285802125930786, + 0.011570009402930737, + 0.04850026220083237, + 0.021111629903316498, + 0.017524180933833122, + 0.007092118263244629, + 0.02651398442685604, + 0.02827398292720318 + ], + "credit_map": [ + 0.05159223452210426, + 0.030744649469852448, + 0.28285640478134155, + 0.02780274488031864, + 0.016189517453312874, + 0.03269470855593681, + 0.025539400056004524, + 0.31541457772254944, + 0.0037612798623740673, + 0.007531170267611742, + 0.003896595910191536, + 0.003830839181318879, + 0.004873952362686396, + 0.0034322526771575212, + 0.004506760276854038, + 0.0034609301947057247, + 0.003416486782953143, + 0.0027940659783780575, + 0.002959376899525523, + 0.0030992114916443825, + 0.002016248181462288, + 0.0019600042141973972, + 0.0024845947045832872, + 0.002555814804509282, + 0.0018136497819796205, + 0.001557879033498466, + 0.001987804425880313, + 0.0010630310280248523, + 0.001234660274349153, + 0.0013348577776923776, + 0.0011366541730239987, + 0.0014414726756513119, + 0.006084514316171408, + 0.007808566093444824, + 0.005600488279014826, + 0.009154262021183968, + 0.0035738027654588223, + 0.00413154624402523, + 0.005510072689503431, + 0.00663700932636857, + 0.0018605777295306325, + 0.0017631753580644727, + 0.003011440159752965, + 0.003302170429378748, + 0.001443764427676797, + 0.002448588376864791, + 0.0036766626872122288, + 0.003605252131819725, + 0.0018810158362612128, + 0.00309801590628922, + 0.0018558301962912083, + 0.001213566865772009, + 0.0034559189807623625, + 0.0020385354291647673, + 0.002083362312987447, + 0.0018979363376274705, + 0.0010251171188428998, + 0.001132136327214539, + 0.0009101607138291001, + 0.0006382580613717437, + 0.0008576876134611666, + 0.0006877008127048612, + 0.000678500160574913, + 0.00116255774628371, + 0.00296280300244689, + 0.003241034224629402, + 0.003127665026113391, + 0.005371009930968285, + 0.002813156694173813, + 0.0025570597499608994, + 0.003334103850647807, + 0.0031071514822542667, + 0.004247315693646669, + 0.004800587892532349, + 0.0033454466611146927, + 0.002094023395329714, + 0.00494576059281826, + 0.0028483893256634474, + 0.003031408414244652, + 0.002961049322038889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0036157488357275724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999125331640244 + ], + "opposite_logprob_tokens": [ + -2.69499135017395 + ], + "logprob_reward": -0.06999125331640244, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999125331640244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7052558660507202, + 0.0939365029335022, + 0.032358624041080475, + 0.021965045481920242, + 0.05915714427828789, + 0.05058092251420021, + 0.008903849869966507, + 0.00623793201521039, + 0.011191819794476032, + 0.010412255302071571 + ], + "credit_map": [ + 0.0454401820898056, + 0.029488423839211464, + 0.25145041942596436, + 0.026598062366247177, + 0.015361348167061806, + 0.03023519180715084, + 0.025566833093762398, + 0.28111541271209717, + 0.012656258419156075, + 0.023487338796257973, + 0.01105684693902731, + 0.008143681101500988, + 0.01343812607228756, + 0.008103598840534687, + 0.007178010419011116, + 0.00987264234572649, + 0.004903556313365698, + 0.006785827688872814, + 0.0038882384542375803, + 0.004401041194796562, + 0.0033218800090253353, + 0.0030593203846365213, + 0.003282474586740136, + 0.002716288436204195, + 0.002003404311835766, + 0.0037568476982414722, + 0.002600015141069889, + 0.0020442840177565813, + 0.004264165181666613, + 0.0020691470708698034, + 0.002660954836755991, + 0.002566225826740265, + 0.0060798777267336845, + 0.009082190692424774, + 0.007367811165750027, + 0.013098559342324734, + 0.004386554006487131, + 0.0045847236178815365, + 0.0077951871789991856, + 0.006762240547686815, + 0.004224872682243586, + 0.008892561309039593, + 0.005971299950033426, + 0.011256604455411434, + 0.0043443795293569565, + 0.0038222328294068575, + 0.006919242441654205, + 0.005149729549884796, + 0.0012958926381543279, + 0.0009633497102186084, + 0.0015068132197484374, + 0.0008914350182749331, + 0.0008599926368333399, + 0.000986197730526328, + 0.0007667639292776585, + 0.0016334052197635174, + 0.0005256350850686431, + 0.000696737552061677, + 0.0008636440034024417, + 0.0007490920834243298, + 0.0009512963006272912, + 0.0007129740552045405, + 0.0009511729585938156, + 0.000787380151450634, + 0.0013135611079633236, + 0.001757084857672453, + 0.0014521643752232194, + 0.0013063309015706182, + 0.0014640381559729576, + 0.0011047528823837638, + 0.0011292051058262587, + 0.0016646829899400473, + 0.0014493357157334685, + 0.0013657900271937251, + 0.001207808032631874, + 0.0016751751536503434, + 0.000849998090416193, + 0.0011329039698466659, + 0.001128859119489789, + 0.0016023849602788687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.06999125331640244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, eye-level medium shot that captures the austere aesthetic of a historical drama or fantasy film, two East Asian men with dark, slightly disheveled hair are pressed against a cold, rough, gray stone wall, their torsos bound by thick ropes. Over their clothing, they both wear ornate, armor-like studded leather vests; the one on the left wears a high-collared white tunic, while the one on the right wears simple brown clothing. Despite their predicament, their facial expressions remain deliberately neutral and stoic, their gazes fixed calmly ahead. The scene is steeped in tension, illuminated by a single, low, off-camera light source that casts long, flickering shadows and creates a dramatic black-and-white contrast, highlighting the texture of the stone and the metallic glint of the armor. Heavy, dark wooden beams are barely visible against the oppressive darkness overhead. At the far left of the frame, a third figure wearing a heavy black scapula is visible on his shoulders, suggesting an ominous captive presence and heightened feelings of imprisonment and imminent conflict.", + "seed": 26292, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_8181_seed_26292_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.025454786885529757, + "vlm_logprob": -0.025454786885529757, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.025454786885529757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01704402081668377 + ], + "opposite_logprob_tokens": [ + -4.2670440673828125 + ], + "logprob_reward": -0.01704402081668377, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01704402081668377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3133624196052551, + 0.34203094244003296, + 0.07786328345537186, + 0.048338472843170166, + 0.07695353031158447, + 0.024287346750497818, + 0.02473393827676773, + 0.046412207186222076, + 0.02156093902885914, + 0.02445686049759388 + ], + "credit_map": [ + 0.032571885734796524, + 0.025656430050730705, + 0.044629842042922974, + 0.035333484411239624, + 0.0386485680937767, + 0.04719668626785278, + 0.047750603407621384, + 0.04157494381070137, + 0.028027299791574478, + 0.030379051342606544, + 0.04368272423744202, + 0.06978877633810043, + 0.036246996372938156, + 0.035525694489479065, + 0.052578628063201904, + 0.04580177366733551, + 0.007868998683989048, + 0.00681106373667717, + 0.008801648393273354, + 0.014160959050059319, + 0.007697329390794039, + 0.01087610051035881, + 0.012895711697638035, + 0.00875148270279169, + 0.006608850788325071, + 0.00370406499132514, + 0.009485009126365185, + 0.005814188625663519, + 0.0038238833658397198, + 0.0056361425668001175, + 0.007242572959512472, + 0.006023763213306665, + 0.01244026143103838, + 0.009763152338564396, + 0.009585078805685043, + 0.006027138791978359, + 0.007102727424353361, + 0.011813759803771973, + 0.008821536786854267, + 0.011399880982935429, + 0.0022437614388763905, + 0.0027979735750705004, + 0.0024003598373383284, + 0.0034460138995200396, + 0.002956809476017952, + 0.0034706122241914272, + 0.004296918865293264, + 0.002674899296835065, + 0.002538930391892791, + 0.002151533728465438, + 0.0033922807779163122, + 0.002951683010905981, + 0.0026438457425683737, + 0.003733541816473007, + 0.0036068607587367296, + 0.00371526344679296, + 0.007365331519395113, + 0.005563065409660339, + 0.00787577498704195, + 0.003587285755202174, + 0.004730845335870981, + 0.006965744309127331, + 0.004139736760407686, + 0.0061844270676374435, + 0.002095639705657959, + 0.0018172921845689416, + 0.004406249616295099, + 0.0026541308034211397, + 0.0025158843491226435, + 0.00239935633726418, + 0.0026520797982811928, + 0.003020309144631028, + 0.0028106130193918943, + 0.0019917127210646868, + 0.004796478897333145, + 0.002278310712426901, + 0.0020795329473912716, + 0.0034146171528846025, + 0.003556797979399562, + 0.003528801491484046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01704402081668377 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026376383379101753 + ], + "opposite_logprob_tokens": [ + -3.651376485824585 + ], + "logprob_reward": -0.026376383379101753, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026376383379101753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3221435248851776, + 0.1863032728433609, + 0.12009245157241821, + 0.07392086833715439, + 0.09628412872552872, + 0.03285183012485504, + 0.035810500383377075, + 0.04677779972553253, + 0.05142687261104584, + 0.034388747066259384 + ], + "credit_map": [ + 0.03198834881186485, + 0.03173748776316643, + 0.03847431018948555, + 0.03165306895971298, + 0.026848766952753067, + 0.07486989349126816, + 0.037353433668613434, + 0.04921821132302284, + 0.018635915592312813, + 0.01876416616141796, + 0.02183162420988083, + 0.032400596886873245, + 0.019562087953090668, + 0.025568194687366486, + 0.023106636479496956, + 0.026434067636728287, + 0.01287261676043272, + 0.014683429151773453, + 0.012708302587270737, + 0.01162520982325077, + 0.013853967189788818, + 0.01916350983083248, + 0.019331874325871468, + 0.015853533521294594, + 0.013466129079461098, + 0.006912201177328825, + 0.008607177063822746, + 0.006725500337779522, + 0.007942616008222103, + 0.01131125632673502, + 0.007707945071160793, + 0.011248040944337845, + 0.013942868448793888, + 0.012645714916288853, + 0.011499287560582161, + 0.00848469603806734, + 0.009063350036740303, + 0.014759203419089317, + 0.011953845620155334, + 0.013935159891843796, + 0.003887959523126483, + 0.004599331878125668, + 0.0035845974925905466, + 0.0029201540164649487, + 0.003921188414096832, + 0.0043950872495770454, + 0.004452941473573446, + 0.005090571008622646, + 0.004044554196298122, + 0.004946195520460606, + 0.00477744173258543, + 0.0035041107330471277, + 0.003656662069261074, + 0.0048257592134177685, + 0.004358915612101555, + 0.0056968629360198975, + 0.005883600562810898, + 0.005972979590296745, + 0.007492333650588989, + 0.005090360529720783, + 0.00492597883567214, + 0.0055650705471634865, + 0.0051465947180986404, + 0.006700881756842136, + 0.010540962219238281, + 0.0047182003036141396, + 0.008211025968194008, + 0.0031162984669208527, + 0.0043422463349998, + 0.007933665066957474, + 0.0049194092862308025, + 0.007645067758858204, + 0.005379953421652317, + 0.002932419301941991, + 0.00723601458594203, + 0.0024006313178688288, + 0.003612250555306673, + 0.004085773602128029, + 0.004041905514895916, + 0.004699798300862312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026376383379101753 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203527010977268 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203527010977268, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203527010977268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24708488583564758, + 0.2552243769168854, + 0.11055687069892883, + 0.07470101863145828, + 0.07205371558666229, + 0.08361005038022995, + 0.045970939099788666, + 0.023417649790644646, + 0.030954064801335335, + 0.05642645061016083 + ], + "credit_map": [ + 0.021533794701099396, + 0.02992400713264942, + 0.03038855455815792, + 0.024502499029040337, + 0.026696225628256798, + 0.04003380239009857, + 0.03585798665881157, + 0.038148004561662674, + 0.02279122732579708, + 0.02354951947927475, + 0.02989850379526615, + 0.05318346992135048, + 0.02955332025885582, + 0.026567529886960983, + 0.03661458194255829, + 0.03306620940566063, + 0.015522915869951248, + 0.012851690873503685, + 0.011228579096496105, + 0.013664807192981243, + 0.011261338368058205, + 0.01452851016074419, + 0.017627159133553505, + 0.01387186348438263, + 0.007375056855380535, + 0.007051512133330107, + 0.0072457073256373405, + 0.009448504075407982, + 0.008581873029470444, + 0.011036915704607964, + 0.013975059613585472, + 0.0099863912910223, + 0.00840738695114851, + 0.006634544115513563, + 0.010181034915149212, + 0.0075128101743757725, + 0.009221240878105164, + 0.009053287096321583, + 0.01197598222643137, + 0.009067431092262268, + 0.014583170413970947, + 0.008604835718870163, + 0.008296815678477287, + 0.009075397625565529, + 0.007361918687820435, + 0.011561543680727482, + 0.011769982986152172, + 0.012356380000710487, + 0.008933143690228462, + 0.00467384560033679, + 0.006640163250267506, + 0.003905627178028226, + 0.003690137527883053, + 0.0060546621680259705, + 0.005034232512116432, + 0.007039120886474848, + 0.003198671154677868, + 0.0030371483881026506, + 0.003708513220772147, + 0.0019123577512800694, + 0.0023088906891644, + 0.0028961163479834795, + 0.0032714649569243193, + 0.0030844842549413443, + 0.004588596057146788, + 0.0031324655283242464, + 0.004265762399882078, + 0.003953934647142887, + 0.002496500965207815, + 0.004179404582828283, + 0.004772322252392769, + 0.003565076971426606, + 0.006444052327424288, + 0.004603181034326553, + 0.009664146229624748, + 0.004344208166003227, + 0.007286979816854, + 0.00813470222055912, + 0.00691210338845849, + 0.009037076495587826 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.014203527010977268 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36978840827941895, + 0.20917612314224243, + 0.07251866906881332, + 0.05267752334475517, + 0.03560178726911545, + 0.03678330406546593, + 0.09145855158567429, + 0.05683208256959915, + 0.04792540892958641, + 0.027238106355071068 + ], + "credit_map": [ + 0.03602195531129837, + 0.037577297538518906, + 0.055177975445985794, + 0.06132682412862778, + 0.04030320420861244, + 0.04466695338487625, + 0.04506729915738106, + 0.04964691400527954, + 0.018173590302467346, + 0.01887310855090618, + 0.026205629110336304, + 0.042504921555519104, + 0.02236810512840748, + 0.025379447266459465, + 0.028143147006630898, + 0.0275281872600317, + 0.009602909907698631, + 0.009028429165482521, + 0.007752054836601019, + 0.009930727072060108, + 0.008417868986725807, + 0.00754729937762022, + 0.011292453855276108, + 0.008946924470365047, + 0.008208397775888443, + 0.00593251409009099, + 0.00710520101711154, + 0.0058876024559140205, + 0.004921880085021257, + 0.006146912928670645, + 0.006367467809468508, + 0.008107542991638184, + 0.004283220507204533, + 0.0032127967569977045, + 0.005220877006649971, + 0.005276966840028763, + 0.0036086554173380136, + 0.0051158033311367035, + 0.004247166216373444, + 0.004636302590370178, + 0.005655325949192047, + 0.00363146117888391, + 0.0048757558688521385, + 0.0022128226701170206, + 0.005173681303858757, + 0.004572197329252958, + 0.0053391363471746445, + 0.005322922486811876, + 0.010019433684647083, + 0.007758717518299818, + 0.011404565535485744, + 0.012078158557415009, + 0.008957458660006523, + 0.012053134851157665, + 0.01587151736021042, + 0.013315562158823013, + 0.00579181918874383, + 0.00647553289309144, + 0.00877836998552084, + 0.006364085711538792, + 0.007025807164609432, + 0.007322720251977444, + 0.008522720076143742, + 0.006551030557602644, + 0.005465184338390827, + 0.003682582639157772, + 0.0063818409107625484, + 0.007287876680493355, + 0.005516408942639828, + 0.00635592732578516, + 0.007370989769697189, + 0.005864599719643593, + 0.0031585197430104017, + 0.0020584098529070616, + 0.007869565859436989, + 0.0019472520798444748, + 0.002280968241393566, + 0.002776317298412323, + 0.002841440960764885, + 0.004305633250623941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007630960550159216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062019042670726776 + ], + "opposite_logprob_tokens": [ + -2.812019109725952 + ], + "logprob_reward": -0.062019042670726776, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062019042670726776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4092978537082672, + 0.0903111919760704, + 0.2393006980419159, + 0.08568128943443298, + 0.034606412053108215, + 0.03724706545472145, + 0.024865763261914253, + 0.02744527906179428, + 0.0307907871901989, + 0.02045365609228611 + ], + "credit_map": [ + 0.049138911068439484, + 0.05245937407016754, + 0.055354245007038116, + 0.04611949250102043, + 0.036641981452703476, + 0.061470385640859604, + 0.04883744567632675, + 0.0592760294675827, + 0.006767159327864647, + 0.00783578958362341, + 0.01157369464635849, + 0.014021240174770355, + 0.012506766244769096, + 0.012085764668881893, + 0.01446911133825779, + 0.011051672510802746, + 0.03712988644838333, + 0.0358782559633255, + 0.03515733405947685, + 0.02085699327290058, + 0.02252047322690487, + 0.033518362790346146, + 0.018931925296783447, + 0.035307470709085464, + 0.014289284124970436, + 0.01251223124563694, + 0.01094833668321371, + 0.005777616985142231, + 0.009329382330179214, + 0.012506556697189808, + 0.00793491117656231, + 0.01238296926021576, + 0.0050282604061067104, + 0.0030719407368451357, + 0.006369869224727154, + 0.003537731012329459, + 0.0030671211425215006, + 0.0044473689049482346, + 0.004658524878323078, + 0.00442559365183115, + 0.007097392808645964, + 0.004408231936395168, + 0.004177894443273544, + 0.0034888219088315964, + 0.004139615688472986, + 0.004774873610585928, + 0.004351942799985409, + 0.004808295052498579, + 0.004134002607315779, + 0.0024401152040809393, + 0.004147798288613558, + 0.0021435413509607315, + 0.002654917538166046, + 0.003093096660450101, + 0.002927826251834631, + 0.0033244655933231115, + 0.0034837324637919664, + 0.003129656193777919, + 0.0030954398680478334, + 0.003180542029440403, + 0.002914732787758112, + 0.004352705553174019, + 0.003941288683563471, + 0.0033471824135631323, + 0.004815947264432907, + 0.0033230327535420656, + 0.004808138590306044, + 0.0023817310575395823, + 0.0037584479432553053, + 0.004388019908219576, + 0.002995464950799942, + 0.00432000495493412, + 0.0021463290322571993, + 0.0015641641803085804, + 0.004831491969525814, + 0.0019043195061385632, + 0.002041254425421357, + 0.0028255758807063103, + 0.002348286798223853, + 0.002792233368381858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.062019042670726776 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a static, eye-level medium shot that captures the austere aesthetic of a historical drama or fantasy film, two East Asian men with dark, slightly disheveled hair are pressed against a cold, rough, gray stone wall, their torsos bound by thick ropes. Over their clothing, they both wear ornate, armor-like studded leather vests; the one on the left wears a high-collared white tunic, while the one on the right wears simple brown clothing. Despite their predicament, their facial expressions remain deliberately neutral and stoic, their gazes fixed calmly ahead. The scene is steeped in tension, illuminated by a single, low, off-camera light source that casts long, flickering shadows and creates a dramatic black-and-white contrast, highlighting the texture of the stone and the metallic glint of the armor. Heavy, dark wooden beams are barely visible against the oppressive darkness overhead. At the far left of the frame, a third figure wearing a heavy black scapula is visible on his shoulders, suggesting an ominous captive presence and heightened feelings of imprisonment and imminent conflict.", + "seed": 26293, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_15/index_8181_seed_26293_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02366615878418088, + "vlm_logprob": -0.02366615878418088, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.02366615878418088, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015397251583635807 + ], + "opposite_logprob_tokens": [ + -4.265397071838379 + ], + "logprob_reward": -0.015397251583635807, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015397251583635807, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4958357810974121, + 0.09980658441781998, + 0.06918240338563919, + 0.050524137914180756, + 0.05188169702887535, + 0.03215745836496353, + 0.03807023912668228, + 0.03423213213682175, + 0.10409023612737656, + 0.02421933226287365 + ], + "credit_map": [ + 0.09775404632091522, + 0.03680509701371193, + 0.03406081721186638, + 0.025122638791799545, + 0.06116198003292084, + 0.15871697664260864, + 0.022887099534273148, + 0.059327125549316406, + 0.013736385852098465, + 0.009834877215325832, + 0.009230075404047966, + 0.016379913315176964, + 0.012922600843012333, + 0.009973152540624142, + 0.011740862391889095, + 0.01598871685564518, + 0.010747072286903858, + 0.008431264199316502, + 0.008492741733789444, + 0.0051821814849972725, + 0.006510666571557522, + 0.017541587352752686, + 0.004768095910549164, + 0.007508792914450169, + 0.009080461226403713, + 0.0035975289065390825, + 0.00863370019942522, + 0.004296029452234507, + 0.005023561418056488, + 0.009935254231095314, + 0.003458312014117837, + 0.006499290466308594, + 0.008317745290696621, + 0.006552663166075945, + 0.005496927537024021, + 0.00527335237711668, + 0.007288351189345121, + 0.006716467905789614, + 0.005331336986273527, + 0.006904854904860258, + 0.004615393467247486, + 0.004438122268766165, + 0.00461828475818038, + 0.0032047322019934654, + 0.004557744134217501, + 0.003415927290916443, + 0.003888922743499279, + 0.003418330568820238, + 0.0059190536849200726, + 0.004425087012350559, + 0.004100239835679531, + 0.0047320169396698475, + 0.0055291117168962955, + 0.003656411077827215, + 0.004199244547635317, + 0.005509074777364731, + 0.005505732260644436, + 0.003986007068306208, + 0.004680968355387449, + 0.004245605785399675, + 0.004194117151200771, + 0.0023517105728387833, + 0.0031966399401426315, + 0.006071351934224367, + 0.0164299625903368, + 0.014235065318644047, + 0.01639414019882679, + 0.013413169421255589, + 0.01007260475307703, + 0.00866511557251215, + 0.010338672436773777, + 0.0145415049046278, + 0.003574342466890812, + 0.0038805948570370674, + 0.004429822321981192, + 0.0026666445191949606, + 0.0029003925155848265, + 0.0011469911551102996, + 0.002025177236646414, + 0.003595368005335331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.015397251583635807 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005270159337669611 + ], + "opposite_logprob_tokens": [ + -5.255270004272461 + ], + "logprob_reward": -0.005270159337669611, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005270159337669611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.414095014333725, + 0.23269891738891602, + 0.07425310462713242, + 0.06098148599267006, + 0.03850315138697624, + 0.03771138936281204, + 0.035483840852975845, + 0.02739153988659382, + 0.05011590197682381, + 0.028765683993697166 + ], + "credit_map": [ + 0.07911563664674759, + 0.054357241839170456, + 0.03966812416911125, + 0.031211530789732933, + 0.03595087677240372, + 0.10044483840465546, + 0.022871319204568863, + 0.05047542601823807, + 0.031207598745822906, + 0.01928655430674553, + 0.03098534233868122, + 0.02786979265511036, + 0.030802465975284576, + 0.029079735279083252, + 0.02775503881275654, + 0.035712387412786484, + 0.012721396051347256, + 0.013057495467364788, + 0.009982473216950893, + 0.006457825191318989, + 0.005159576423466206, + 0.01341632567346096, + 0.004091363400220871, + 0.0093666547909379, + 0.012207409366965294, + 0.005027702543884516, + 0.00952308252453804, + 0.007365257479250431, + 0.0047805760987102985, + 0.011697719804942608, + 0.004191475920379162, + 0.006188260391354561, + 0.005698261316865683, + 0.007090341299772263, + 0.0035726192872971296, + 0.004461413715034723, + 0.004302568268030882, + 0.004718884825706482, + 0.0046521746553480625, + 0.004006886389106512, + 0.0054862964898347855, + 0.005370459984987974, + 0.007624030113220215, + 0.004002880305051804, + 0.004386396147310734, + 0.0028486321680247784, + 0.0038892102893441916, + 0.004103488754481077, + 0.005836829077452421, + 0.005237293429672718, + 0.007283031474798918, + 0.005186648108065128, + 0.0026910474989563227, + 0.002414607908576727, + 0.00274833757430315, + 0.0040860469453036785, + 0.004495163448154926, + 0.003601576667279005, + 0.002669215900823474, + 0.0036460610572248697, + 0.003551339730620384, + 0.0018702003872022033, + 0.002397487172856927, + 0.005160495638847351, + 0.007708959747105837, + 0.007257780060172081, + 0.007341044023633003, + 0.006282449699938297, + 0.005518341902643442, + 0.004356870893388987, + 0.004766363184899092, + 0.006884091068059206, + 0.0050127506256103516, + 0.005042685195803642, + 0.004303365480154753, + 0.004291398450732231, + 0.002162404591217637, + 0.0013761784648522735, + 0.0031416164711117744, + 0.0034352848306298256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005270159337669611 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077453382313251 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.011077453382313251, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077453382313251, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2725498676300049, + 0.09368574619293213, + 0.11766762286424637, + 0.18618327379226685, + 0.06368333846330643, + 0.05625962093472481, + 0.07943373918533325, + 0.04928145185112953, + 0.056376054883003235, + 0.024879271164536476 + ], + "credit_map": [ + 0.04450701177120209, + 0.03182779997587204, + 0.022706255316734314, + 0.03150923177599907, + 0.04799136146903038, + 0.04513294994831085, + 0.016928941011428833, + 0.03194630146026611, + 0.012630767188966274, + 0.014454696327447891, + 0.00983992125838995, + 0.011428163386881351, + 0.010065427049994469, + 0.014229506254196167, + 0.00649961456656456, + 0.014537647366523743, + 0.01688060164451599, + 0.012479778379201889, + 0.017613641917705536, + 0.011825650930404663, + 0.01358297560364008, + 0.024887748062610626, + 0.007500451058149338, + 0.012896779924631119, + 0.031092151999473572, + 0.019206689670681953, + 0.01847485825419426, + 0.016652313992381096, + 0.015434056520462036, + 0.05605875328183174, + 0.011338695883750916, + 0.017925752326846123, + 0.008761812932789326, + 0.012727583758533001, + 0.00640589464455843, + 0.009545734152197838, + 0.007405062206089497, + 0.006947912275791168, + 0.0054084728471934795, + 0.006480867974460125, + 0.007647228892892599, + 0.007308322470635176, + 0.00787123292684555, + 0.007868489250540733, + 0.0069035557098686695, + 0.005050076637417078, + 0.006512583699077368, + 0.00709813367575407, + 0.012959681451320648, + 0.010541904717683792, + 0.009893812239170074, + 0.007273018825799227, + 0.009659999050199986, + 0.008697516284883022, + 0.007783057633787394, + 0.012624749913811684, + 0.007711251266300678, + 0.007617060095071793, + 0.006243076641112566, + 0.005743726622313261, + 0.006609058938920498, + 0.004280448891222477, + 0.005333136301487684, + 0.005743690300732851, + 0.008826090022921562, + 0.00876834150403738, + 0.009237523190677166, + 0.00731499632820487, + 0.005205110646784306, + 0.003916987683624029, + 0.005774638615548611, + 0.0073323664255440235, + 0.004015608690679073, + 0.004296951927244663, + 0.003346648532897234, + 0.0033911769278347492, + 0.002957660472020507, + 0.0017292738193646073, + 0.0028029615059494972, + 0.002338987775146961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.011077453382313251 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076294224709272385 + ], + "opposite_logprob_tokens": [ + -4.88262939453125 + ], + "logprob_reward": -0.0076294224709272385, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076294224709272385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4177354574203491, + 0.20047228038311005, + 0.04484633728861809, + 0.07161012291908264, + 0.06068963557481766, + 0.04104815796017647, + 0.05910649523139, + 0.03854219615459442, + 0.04340197145938873, + 0.022547313943505287 + ], + "credit_map": [ + 0.06515880674123764, + 0.06092908978462219, + 0.04742339253425598, + 0.03948165848851204, + 0.10077888518571854, + 0.04113353043794632, + 0.02774420566856861, + 0.035085901618003845, + 0.025892751291394234, + 0.01659652777016163, + 0.021889938041567802, + 0.0302863921970129, + 0.02878863923251629, + 0.023231765255331993, + 0.02370993234217167, + 0.03007633425295353, + 0.006315703969448805, + 0.005553556606173515, + 0.0058984775096178055, + 0.004030052572488785, + 0.0037233259063214064, + 0.010464201681315899, + 0.0030255981255322695, + 0.005835420917719603, + 0.012350653298199177, + 0.013040349818766117, + 0.013667949475347996, + 0.006369628012180328, + 0.005816143937408924, + 0.00931358803063631, + 0.0035211334470659494, + 0.007530678063631058, + 0.00930805318057537, + 0.008133961819112301, + 0.008228227496147156, + 0.005112257786095142, + 0.00538285868242383, + 0.013203137554228306, + 0.00498612318187952, + 0.006335015408694744, + 0.006171507760882378, + 0.009138274006545544, + 0.005627655424177647, + 0.004978486802428961, + 0.004412379581481218, + 0.002819795859977603, + 0.002776738489046693, + 0.005123321432620287, + 0.0082109235227108, + 0.006562301889061928, + 0.00755250034853816, + 0.00881484616547823, + 0.008318517357110977, + 0.005044827237725258, + 0.005696577951312065, + 0.00890599749982357, + 0.0064837317913770676, + 0.00534832151606679, + 0.0041182152926921844, + 0.004092889372259378, + 0.005165540147572756, + 0.0029093599878251553, + 0.003562919097021222, + 0.006861219182610512, + 0.00698477728292346, + 0.005278992932289839, + 0.006548572797328234, + 0.005618542432785034, + 0.004311613272875547, + 0.003770538605749607, + 0.004223554395139217, + 0.006665377411991358, + 0.004074451047927141, + 0.00351891340687871, + 0.006059767212718725, + 0.002660970203578472, + 0.001536698080599308, + 0.0012287034187465906, + 0.0015463816234841943, + 0.0019214283674955368 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0076294224709272385 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two East Asian men with dark, slightly disheveled hair visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the two men pressed against a cold, rough, gray stone wall with their torsos bound by thick ropes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man on the left wear a high-collared white tunic and an ornate, armor-like studded leather vest, while the man on the right wears simple brown clothing and a similar vest?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a single, low, off-camera light source casting long, flickering shadows and creating a dramatic black-and-white contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a third figure wearing a heavy black scapula visible at the far left of the frame?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895650714635849 + ], + "opposite_logprob_tokens": [ + -2.5789566040039062 + ], + "logprob_reward": -0.07895650714635849, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895650714635849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32444795966148376, + 0.3383331894874573, + 0.07291807234287262, + 0.04333718493580818, + 0.04215698316693306, + 0.03987183794379234, + 0.03316700831055641, + 0.018978627398610115, + 0.03503900393843651, + 0.051750052720308304 + ], + "credit_map": [ + 0.05845983326435089, + 0.02915962040424347, + 0.0230634193867445, + 0.02262674644589424, + 0.05053463578224182, + 0.08841095864772797, + 0.018380945548415184, + 0.03381183370947838, + 0.04727155715227127, + 0.027796516194939613, + 0.031804971396923065, + 0.0583229698240757, + 0.046678755432367325, + 0.03339127451181412, + 0.047569941729307175, + 0.04549725726246834, + 0.009755803272128105, + 0.014285566285252571, + 0.008940566331148148, + 0.005358729511499405, + 0.005033879540860653, + 0.014056564308702946, + 0.0036755281034857035, + 0.011811449192464352, + 0.007423706818372011, + 0.006056921556591988, + 0.006300183944404125, + 0.0036733122542500496, + 0.0030319038778543472, + 0.008454994298517704, + 0.0023592219222337008, + 0.006036945153027773, + 0.006225949618965387, + 0.01057029515504837, + 0.005995158106088638, + 0.0034343728329986334, + 0.004126487299799919, + 0.004489565268158913, + 0.0028660378884524107, + 0.004449123051017523, + 0.006225955672562122, + 0.005914189852774143, + 0.006891588214784861, + 0.004127513617277145, + 0.004752497188746929, + 0.0029620144050568342, + 0.0045531392097473145, + 0.004444945603609085, + 0.005101350136101246, + 0.0043389503844082355, + 0.0077132307924330235, + 0.004743802361190319, + 0.0031133086886256933, + 0.001973974285647273, + 0.0025027666706591845, + 0.003679627785459161, + 0.002736927242949605, + 0.0020078064408153296, + 0.0026278640143573284, + 0.0018513668328523636, + 0.002640088554471731, + 0.0012812045169994235, + 0.002220403403043747, + 0.003612968372181058, + 0.0053750560618937016, + 0.003958653658628464, + 0.004747385624796152, + 0.00341382366605103, + 0.003996302839368582, + 0.0030418101232498884, + 0.0044728596694767475, + 0.006033118348568678, + 0.006864425726234913, + 0.00811621081084013, + 0.007070702034980059, + 0.007566686253994703, + 0.0065581705421209335, + 0.002864458132535219, + 0.0049653989262878895, + 0.007744009606540203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07895650714635849 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_tensor([6520, 6520])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_tensor([6520, 6520])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8f4150f9c7980ec7082c3ab2505f47d8ad26a8d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_tensor([6520, 6520])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures the surreal presence of a mature giraffe standing in a brightly lit supermarket aisle, its head and long neck jutting out from the top of the frame. The camera then executes a smooth pan to the right, following the animal as it begins to move. The giraffe, its fur a detailed pattern of rich brown polygonal patches separated by cream-colored lines, awkwardly begins to move away from the camera. The movement is both clumsy and restrained; its massive body slowly turns on the polished linen floor, reflecting the harsh, cold-white fluorescent lights overhead. As it completes its turn and begins to walk away, its powerful hind legs and swishy black tail inadvertently sweep across the densely packed metal shelves. This triggers a chaotic cascade as various products—unbranded, brightly colored cereal boxes, cylindrical tin cans, and crumpled snack bags—are sent clattering and tumbling onto the ground. The giraffe continues its steady and deliberate walk down the aisle, its form receding into the distance, seemingly oblivious to the traces of commercial vandalism, creating an atmosphere of startling, deadpan, chaotic comedy.", + "seed": 31807, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31807_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03893469758331776, + "vlm_logprob": -0.03893469758331776, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.03893469758331776, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013303680345416069 + ], + "opposite_logprob_tokens": [ + -4.763303756713867 + ], + "logprob_reward": -0.013303680345416069, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.013303680345416069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33826759457588196, + 0.06312485039234161, + 0.04803020879626274, + 0.24740564823150635, + 0.03984472155570984, + 0.09811834245920181, + 0.03660034015774727, + 0.029435504227876663, + 0.07301327586174011, + 0.02615952305495739 + ], + "credit_map": [ + 0.049899596720933914, + 0.0575350821018219, + 0.03820345178246498, + 0.05511495843529701, + 0.0477786585688591, + 0.039621502161026, + 0.024087609723210335, + 0.026026727631688118, + 0.010195083916187286, + 0.009197002276778221, + 0.005580594297498465, + 0.009758509695529938, + 0.00499682454392314, + 0.010243706405162811, + 0.004951127804815769, + 0.00820199679583311, + 0.005898409988731146, + 0.0053993756882846355, + 0.0069098626263439655, + 0.007603683043271303, + 0.003414801089093089, + 0.00798158161342144, + 0.004128903150558472, + 0.006693587172776461, + 0.031281694769859314, + 0.05106913670897484, + 0.015063295140862465, + 0.0534137599170208, + 0.017870167270302773, + 0.01900572143495083, + 0.0374431274831295, + 0.022258710116147995, + 0.0038440697826445103, + 0.0033109618816524744, + 0.004795592278242111, + 0.005377004388719797, + 0.007556300610303879, + 0.002387087093666196, + 0.0084146773442626, + 0.0041590239852666855, + 0.012710805051028728, + 0.01016837265342474, + 0.013552487827837467, + 0.009734762832522392, + 0.013693121261894703, + 0.013573680073022842, + 0.012089605443179607, + 0.012595493346452713, + 0.0038433074951171875, + 0.004206026904284954, + 0.006262187846004963, + 0.0031991018913686275, + 0.006284150294959545, + 0.0038255308754742146, + 0.003396307583898306, + 0.0055837249383330345, + 0.0018391155172139406, + 0.0027287357952445745, + 0.0024884615559130907, + 0.002608882961794734, + 0.003121588146314025, + 0.0025411497335880995, + 0.008323493413627148, + 0.0057840775698423386, + 0.011780820786952972, + 0.007476191967725754, + 0.008883766829967499, + 0.007131714373826981, + 0.010474181734025478, + 0.009651963599026203, + 0.010628999210894108, + 0.006985633168369532, + 0.0037299245595932007, + 0.002463647397235036, + 0.0031322313006967306, + 0.0014475735370069742, + 0.005277080461382866, + 0.0031217127107083797, + 0.004154617432504892, + 0.0028327314648777246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.013303680345416069 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583592355251312 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583592355251312, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583592355251312, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3009272813796997, + 0.06847424805164337, + 0.036494866013526917, + 0.1446278840303421, + 0.12468156963586807, + 0.09614170342683792, + 0.04764745384454727, + 0.04279223829507828, + 0.09620104730129242, + 0.042011722922325134 + ], + "credit_map": [ + 0.04445084184408188, + 0.050763413310050964, + 0.03570105507969856, + 0.048308853060007095, + 0.042765676975250244, + 0.03560228273272514, + 0.02133973129093647, + 0.021995391696691513, + 0.012315575033426285, + 0.010769160464406013, + 0.006867718417197466, + 0.01682228595018387, + 0.00484170438721776, + 0.006162630394101143, + 0.003948624711483717, + 0.006746543571352959, + 0.005199494305998087, + 0.005009212531149387, + 0.005728941410779953, + 0.0060408031567931175, + 0.0021930045913904905, + 0.004375969525426626, + 0.0031930466648191214, + 0.00475439103320241, + 0.0201942827552557, + 0.016790427267551422, + 0.017893407493829727, + 0.023736990988254547, + 0.010128878057003021, + 0.009790413081645966, + 0.0138773862272501, + 0.032216086983680725, + 0.010936740785837173, + 0.010611162520945072, + 0.014655527658760548, + 0.02912827581167221, + 0.01213737390935421, + 0.010478188283741474, + 0.02233879081904888, + 0.014395501464605331, + 0.014311505481600761, + 0.008059891872107983, + 0.015425615012645721, + 0.011163728311657906, + 0.013432088308036327, + 0.008907032199203968, + 0.01008403766900301, + 0.014757801778614521, + 0.006351112853735685, + 0.008863520808517933, + 0.004305168054997921, + 0.006113917101174593, + 0.008493209257721901, + 0.0038912645541131496, + 0.004659157712012529, + 0.004970098379999399, + 0.003299843054264784, + 0.0048417444340884686, + 0.003898902330547571, + 0.005909827072173357, + 0.0061330548487603664, + 0.004062159918248653, + 0.009497963823378086, + 0.005148737691342831, + 0.011717279441654682, + 0.007889357395470142, + 0.016547953709959984, + 0.008579842746257782, + 0.015395326539874077, + 0.011952048167586327, + 0.014565430581569672, + 0.009553796611726284, + 0.005914188921451569, + 0.004822385963052511, + 0.005449847783893347, + 0.0022826609201729298, + 0.007461277302354574, + 0.005438238848000765, + 0.006071425508707762, + 0.004571693018078804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.020583592355251312 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368980064988136 + ], + "opposite_logprob_tokens": [ + -3.4086897373199463 + ], + "logprob_reward": -0.03368980064988136, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368980064988136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11726796627044678, + 0.11525894701480865, + 0.059991780668497086, + 0.2692253291606903, + 0.08021605014801025, + 0.105148546397686, + 0.04826775938272476, + 0.03946555778384209, + 0.12655235826969147, + 0.03860564902424812 + ], + "credit_map": [ + 0.01528013776987791, + 0.01688935048878193, + 0.010560771450400352, + 0.01748649962246418, + 0.011271992698311806, + 0.02621139958500862, + 0.0074450233951210976, + 0.012122800573706627, + 0.01637728326022625, + 0.020603766664862633, + 0.010434086434543133, + 0.02189141884446144, + 0.008899368345737457, + 0.016541117802262306, + 0.007094917353242636, + 0.013416997157037258, + 0.007213712669909, + 0.006112031638622284, + 0.00756970327347517, + 0.013500147499144077, + 0.0040352679789066315, + 0.00976781826466322, + 0.006174816284328699, + 0.005618291441351175, + 0.03300781175494194, + 0.04237382113933563, + 0.022903291508555412, + 0.03802475333213806, + 0.028409786522388458, + 0.026106007397174835, + 0.027475325390696526, + 0.05092455446720123, + 0.007204325404018164, + 0.0052684904076159, + 0.00826066080480814, + 0.01797688752412796, + 0.012544560246169567, + 0.006591649726033211, + 0.015733329579234123, + 0.006636147387325764, + 0.01128046028316021, + 0.009438867680728436, + 0.012330230325460434, + 0.014818771742284298, + 0.01984686404466629, + 0.011921669356524944, + 0.014541314914822578, + 0.010970369912683964, + 0.005614343099296093, + 0.00619318475946784, + 0.004686322994530201, + 0.007590687833726406, + 0.008908921852707863, + 0.004308138974010944, + 0.004358913749456406, + 0.006607248913496733, + 0.0035557381343096495, + 0.004995902068912983, + 0.002846143441274762, + 0.00473705492913723, + 0.005925450008362532, + 0.0041113970801234245, + 0.0061383782885968685, + 0.007155497092753649, + 0.01151201780885458, + 0.009831132367253304, + 0.016431156545877457, + 0.010285492986440659, + 0.02482178620994091, + 0.015618447214365005, + 0.022717833518981934, + 0.01533450186252594, + 0.006347470451146364, + 0.003230879083275795, + 0.004437823314219713, + 0.0018284064717590809, + 0.007338113617151976, + 0.005589028354734182, + 0.005956963170319796, + 0.003876965958625078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.03368980064988136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809294104576111 + ], + "opposite_logprob_tokens": [ + -3.288092851638794 + ], + "logprob_reward": -0.03809294104576111, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809294104576111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0923842117190361, + 0.1003899946808815, + 0.01879407837986946, + 0.3598005771636963, + 0.0847334936261177, + 0.0779355987906456, + 0.10268672555685043, + 0.07994991540908813, + 0.05974125489592552, + 0.023584144189953804 + ], + "credit_map": [ + 0.01261286810040474, + 0.014411401003599167, + 0.010342005640268326, + 0.013642366975545883, + 0.013464190997183323, + 0.011057491414248943, + 0.007512591779232025, + 0.009341306053102016, + 0.017423225566744804, + 0.0174215417355299, + 0.009281570091843605, + 0.021081602200865746, + 0.0076193129643797874, + 0.00839235633611679, + 0.00713606271892786, + 0.012034323997795582, + 0.0022026565857231617, + 0.0017849300056695938, + 0.002721288939937949, + 0.004132825881242752, + 0.0016629536403343081, + 0.0014851231826469302, + 0.0018159076571464539, + 0.002988393185660243, + 0.040182605385780334, + 0.058282166719436646, + 0.017409799620509148, + 0.11360936611890793, + 0.017672087997198105, + 0.018658360466361046, + 0.0704977959394455, + 0.023488406091928482, + 0.00870174914598465, + 0.006500689312815666, + 0.013069831766188145, + 0.01216254010796547, + 0.009680097922682762, + 0.009190268814563751, + 0.015071507543325424, + 0.01035680714994669, + 0.00987173616886139, + 0.008089420385658741, + 0.011518031358718872, + 0.008498656563460827, + 0.010912362486124039, + 0.008530891500413418, + 0.011265121400356293, + 0.00924938265234232, + 0.019235430285334587, + 0.014200092293322086, + 0.013721595518290997, + 0.010701642371714115, + 0.012081412598490715, + 0.010324913077056408, + 0.012455293908715248, + 0.009966352954506874, + 0.0062622916884720325, + 0.009802478365600109, + 0.007562074810266495, + 0.010135634802281857, + 0.00844286847859621, + 0.005568552762269974, + 0.014699885621666908, + 0.01747613586485386, + 0.0086479177698493, + 0.006265675649046898, + 0.008299492299556732, + 0.006633578799664974, + 0.008241093717515469, + 0.008652185089886189, + 0.006348995491862297, + 0.006652321666479111, + 0.004061313811689615, + 0.0018979215528815985, + 0.002821842208504677, + 0.0010931955184787512, + 0.00412723096087575, + 0.003671148791909218, + 0.0034754241351038218, + 0.0024360688403248787 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.03809294104576111 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08900347352027893 + ], + "opposite_logprob_tokens": [ + -2.464003562927246 + ], + "logprob_reward": -0.08900347352027893, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08900347352027893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15817871689796448, + 0.07258082926273346, + 0.047920484095811844, + 0.1738194078207016, + 0.1578039675951004, + 0.11130791902542114, + 0.1573849469423294, + 0.04276907071471214, + 0.05920548364520073, + 0.01902923919260502 + ], + "credit_map": [ + 0.02146713063120842, + 0.022361138835549355, + 0.017420578747987747, + 0.026182837784290314, + 0.021295616403222084, + 0.028015661984682083, + 0.008226473815739155, + 0.013209273107349873, + 0.009384959936141968, + 0.010796118527650833, + 0.006577353458851576, + 0.013066837564110756, + 0.0061807213351130486, + 0.010639156214892864, + 0.007942218333482742, + 0.007993466220796108, + 0.0060450308956205845, + 0.00545661011710763, + 0.007604358717799187, + 0.00825420767068863, + 0.004582185763865709, + 0.0049206470139324665, + 0.00464959442615509, + 0.0064078522846102715, + 0.014491048641502857, + 0.017193691805005074, + 0.012660814449191093, + 0.039785221219062805, + 0.015705294907093048, + 0.017147209495306015, + 0.023669520393013954, + 0.033166609704494476, + 0.008303641341626644, + 0.011654864065349102, + 0.015122306533157825, + 0.029580919072031975, + 0.03455619886517525, + 0.007498753257095814, + 0.037085629999637604, + 0.014001655392348766, + 0.011846181005239487, + 0.008338211104273796, + 0.01820293813943863, + 0.012997125275433064, + 0.015717830508947372, + 0.00701585179194808, + 0.017035365104675293, + 0.02015441469848156, + 0.014064311049878597, + 0.017333446070551872, + 0.010063264518976212, + 0.014749052003026009, + 0.033110156655311584, + 0.019189398735761642, + 0.010860190726816654, + 0.03801511973142624, + 0.004733914975076914, + 0.005543855018913746, + 0.006195054389536381, + 0.004978150594979525, + 0.007101824507117271, + 0.004622358363121748, + 0.0047435276210308075, + 0.004850386176258326, + 0.009792941622436047, + 0.007462992798537016, + 0.007886701263487339, + 0.00571892224252224, + 0.006399156991392374, + 0.00960929598659277, + 0.0061061629094183445, + 0.006229308433830738, + 0.0026786616072058678, + 0.002244279021397233, + 0.0031460803002119064, + 0.0013663711724802852, + 0.0028345331083983183, + 0.001796610769815743, + 0.0022516995668411255, + 0.002711002714931965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.08900347352027893 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the surreal presence of a mature giraffe standing in a brightly lit supermarket aisle, its head and long neck jutting out from the top of the frame. The camera then executes a smooth pan to the right, following the animal as it begins to move. The giraffe, its fur a detailed pattern of rich brown polygonal patches separated by cream-colored lines, awkwardly begins to move away from the camera. The movement is both clumsy and restrained; its massive body slowly turns on the polished linen floor, reflecting the harsh, cold-white fluorescent lights overhead. As it completes its turn and begins to walk away, its powerful hind legs and swishy black tail inadvertently sweep across the densely packed metal shelves. This triggers a chaotic cascade as various products—unbranded, brightly colored cereal boxes, cylindrical tin cans, and crumpled snack bags—are sent clattering and tumbling onto the ground. The giraffe continues its steady and deliberate walk down the aisle, its form receding into the distance, seemingly oblivious to the traces of commercial vandalism, creating an atmosphere of startling, deadpan, chaotic comedy.", + "seed": 31808, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31808_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03716130182147026, + "vlm_logprob": -0.03716130182147026, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.03716130182147026, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01618168130517006 + ], + "opposite_logprob_tokens": [ + -4.391181468963623 + ], + "logprob_reward": -0.01618168130517006, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01618168130517006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43636348843574524, + 0.023284796625375748, + 0.05791332945227623, + 0.06558272242546082, + 0.017411595210433006, + 0.04714361950755119, + 0.021173784509301186, + 0.22514213621616364, + 0.05232546478509903, + 0.0536591112613678 + ], + "credit_map": [ + 0.0311651062220335, + 0.017937760800123215, + 0.10616733133792877, + 0.035911303013563156, + 0.03413010016083717, + 0.04299388825893402, + 0.04382779449224472, + 0.12423019111156464, + 0.001699040294624865, + 0.002612477168440819, + 0.0026522681582719088, + 0.003604927333071828, + 0.004457639064639807, + 0.002285356167703867, + 0.0036486005410552025, + 0.002324488013982773, + 0.005047271028161049, + 0.008306440897285938, + 0.008651556447148323, + 0.007820488885045052, + 0.006746790837496519, + 0.005858950316905975, + 0.006745560560375452, + 0.008736269548535347, + 0.008018217980861664, + 0.006766477599740028, + 0.006722943391650915, + 0.008212532848119736, + 0.009861558675765991, + 0.0071500930935144424, + 0.01094976905733347, + 0.00790113303810358, + 0.002203766955062747, + 0.0025888399686664343, + 0.002264929935336113, + 0.0020987887401133776, + 0.0019949355628341436, + 0.0021787381265312433, + 0.0018936755368486047, + 0.0021879193373024464, + 0.0062871393747627735, + 0.0037091129925101995, + 0.0049415272660553455, + 0.004581138025969267, + 0.006205108482390642, + 0.005308463238179684, + 0.010361951775848866, + 0.005749179981648922, + 0.0018531159730628133, + 0.0017618368146941066, + 0.002334264572709799, + 0.004633558914065361, + 0.0028119233902543783, + 0.0018641145434230566, + 0.003054544096812606, + 0.0028604259714484215, + 0.00664098747074604, + 0.01032648328691721, + 0.008246093057096004, + 0.07115786522626877, + 0.022576602175831795, + 0.011175540275871754, + 0.017256993800401688, + 0.07776157557964325, + 0.004918341990560293, + 0.008921216242015362, + 0.006232582964003086, + 0.0068106986582279205, + 0.006796963978558779, + 0.007558527402579784, + 0.006411542184650898, + 0.004675592295825481, + 0.005730037111788988, + 0.004709654022008181, + 0.0027159040328115225, + 0.012970549054443836, + 0.004829142242670059, + 0.003588198684155941, + 0.015118577517569065, + 0.0039970469661056995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.01618168130517006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328725904226303 + ], + "opposite_logprob_tokens": [ + -3.773287296295166 + ], + "logprob_reward": -0.02328725904226303, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328725904226303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14987677335739136, + 0.06648688018321991, + 0.07799321413040161, + 0.1042466089129448, + 0.06966699659824371, + 0.08599067479372025, + 0.06695769727230072, + 0.2262195497751236, + 0.07617148011922836, + 0.07639005035161972 + ], + "credit_map": [ + 0.010481330566108227, + 0.008715186268091202, + 0.03464837744832039, + 0.015121706761419773, + 0.01294273417443037, + 0.015205767005681992, + 0.015999820083379745, + 0.036761850118637085, + 0.0033734238240867853, + 0.005284725688397884, + 0.011554109863936901, + 0.00695843668654561, + 0.007062231656163931, + 0.0076776351779699326, + 0.009269981645047665, + 0.015306335873901844, + 0.005884923040866852, + 0.00925933476537466, + 0.013663009740412235, + 0.014047455042600632, + 0.007277423981577158, + 0.006619818974286318, + 0.011766077019274235, + 0.009475169703364372, + 0.01605059579014778, + 0.015093876980245113, + 0.01150521170347929, + 0.0138009712100029, + 0.011029616929590702, + 0.014269532635807991, + 0.009453183971345425, + 0.01304362341761589, + 0.010816680267453194, + 0.008805802091956139, + 0.010185668244957924, + 0.007083214819431305, + 0.006925915367901325, + 0.00906133372336626, + 0.007297532632946968, + 0.009490852244198322, + 0.012609590776264668, + 0.008460252545773983, + 0.010086328722536564, + 0.008971929550170898, + 0.009175021201372147, + 0.011867443099617958, + 0.014956418424844742, + 0.009863689541816711, + 0.006021877750754356, + 0.0047043985687196255, + 0.0065267025493085384, + 0.011085079982876778, + 0.010376367717981339, + 0.007109771948307753, + 0.012462067417800426, + 0.008671431802213192, + 0.008095691911876202, + 0.009134653955698013, + 0.010171465575695038, + 0.0962541326880455, + 0.01014811359345913, + 0.008296383544802666, + 0.019035130739212036, + 0.06508398056030273, + 0.006456070113927126, + 0.006793016567826271, + 0.006631992291659117, + 0.010592777281999588, + 0.013960572890937328, + 0.010035722516477108, + 0.012845274060964584, + 0.00885605439543724, + 0.005568760447204113, + 0.009792475961148739, + 0.006913002580404282, + 0.009704869240522385, + 0.01181295607239008, + 0.00909273698925972, + 0.014046582393348217, + 0.009458668529987335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.02328725904226303 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033691298216581345 + ], + "opposite_logprob_tokens": [ + -3.40869140625 + ], + "logprob_reward": -0.033691298216581345, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033691298216581345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12715935707092285, + 0.05490957200527191, + 0.051516834646463394, + 0.0669422298669815, + 0.03292941674590111, + 0.04635211080312729, + 0.08430136740207672, + 0.2965005934238434, + 0.19604253768920898, + 0.043345991522073746 + ], + "credit_map": [ + 0.010089037008583546, + 0.005392097868025303, + 0.0273018479347229, + 0.01154252141714096, + 0.010844667442142963, + 0.012286270968616009, + 0.012234248220920563, + 0.03746865689754486, + 0.0032814061269164085, + 0.003930509090423584, + 0.006197776645421982, + 0.00610238267108798, + 0.004667055793106556, + 0.005259125493466854, + 0.010319997556507587, + 0.015151314437389374, + 0.005357938352972269, + 0.012016012333333492, + 0.005831777583807707, + 0.008122689090669155, + 0.005738723557442427, + 0.003607877530157566, + 0.005642404779791832, + 0.005199410952627659, + 0.010462300851941109, + 0.007825477048754692, + 0.005642108153551817, + 0.009370110929012299, + 0.006938620936125517, + 0.006740987300872803, + 0.008783468045294285, + 0.011179158464074135, + 0.004740752279758453, + 0.0036576257552951574, + 0.003076828084886074, + 0.003422682173550129, + 0.005034807603806257, + 0.005434220656752586, + 0.003647539997473359, + 0.003914959263056517, + 0.011499418877065182, + 0.005609496496617794, + 0.00485125882551074, + 0.0052147600799798965, + 0.005194008816033602, + 0.004359855316579342, + 0.005469136405736208, + 0.00415417505428195, + 0.007341710850596428, + 0.0056841811165213585, + 0.009243560954928398, + 0.018205083906650543, + 0.012878285720944405, + 0.009021604433655739, + 0.013236083090305328, + 0.008690864779055119, + 0.00756855821236968, + 0.010321510024368763, + 0.014824600890278816, + 0.12029385566711426, + 0.01371793169528246, + 0.009782875888049603, + 0.02441713586449623, + 0.09557413309812546, + 0.01036830060184002, + 0.016713472083210945, + 0.013651493936777115, + 0.031672652810811996, + 0.04481532424688339, + 0.016839273273944855, + 0.038585513830184937, + 0.023396503180265427, + 0.006191015709191561, + 0.0042428309097886086, + 0.002698887838050723, + 0.006597375962883234, + 0.0063168820925056934, + 0.004591725766658783, + 0.007609489839524031, + 0.005097782704979181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.033691298216581345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369152918457985 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.03369152918457985, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369152918457985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1444789320230484, + 0.03592453896999359, + 0.08485815674066544, + 0.13109514117240906, + 0.05823894962668419, + 0.03333284333348274, + 0.08796852827072144, + 0.3056142032146454, + 0.028883619233965874, + 0.08960507810115814 + ], + "credit_map": [ + 0.009212937206029892, + 0.006007506977766752, + 0.036120787262916565, + 0.010786332190036774, + 0.009260045364499092, + 0.013503406196832657, + 0.013594471849501133, + 0.04599342867732048, + 0.0027324187103658915, + 0.0037346764001995325, + 0.0032042423263192177, + 0.00417061522603035, + 0.004076276905834675, + 0.0034276400692760944, + 0.0072310003452003, + 0.007347666192799807, + 0.008296758867800236, + 0.025845129042863846, + 0.00886088889092207, + 0.011216044425964355, + 0.007442782633006573, + 0.004621053580194712, + 0.012276723980903625, + 0.006298769731074572, + 0.023580145090818405, + 0.014678659848868847, + 0.012484748847782612, + 0.016654396429657936, + 0.012573846615850925, + 0.012482375837862492, + 0.01629532314836979, + 0.02234564907848835, + 0.009226301684975624, + 0.007607513107359409, + 0.009125780314207077, + 0.006962111219763756, + 0.00686338497325778, + 0.006653608288615942, + 0.005153838545084, + 0.006646412890404463, + 0.0026137761306017637, + 0.002573553239926696, + 0.004336592275649309, + 0.0052240765653550625, + 0.004794446285814047, + 0.0030664505902677774, + 0.006696605123579502, + 0.004027342889457941, + 0.007083926349878311, + 0.005657849833369255, + 0.010358124040067196, + 0.02007208950817585, + 0.012005586177110672, + 0.008410756476223469, + 0.014431207440793514, + 0.009948987513780594, + 0.006971647031605244, + 0.011558258906006813, + 0.013241195119917393, + 0.12693779170513153, + 0.010418144054710865, + 0.010097639635205269, + 0.02310262620449066, + 0.10328689217567444, + 0.0028735604137182236, + 0.002252961043268442, + 0.002555249258875847, + 0.004305745474994183, + 0.005034653004258871, + 0.002547350712120533, + 0.004709100816398859, + 0.004604997579008341, + 0.013037252239882946, + 0.007057398557662964, + 0.005508530884981155, + 0.019688287749886513, + 0.007095112465322018, + 0.0075781336054205894, + 0.021993208676576614, + 0.007647154852747917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.03369152918457985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895474135875702 + ], + "opposite_logprob_tokens": [ + -2.5789546966552734 + ], + "logprob_reward": -0.07895474135875702, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895474135875702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07698933780193329, + 0.11023910343647003, + 0.05517978593707085, + 0.1303485482931137, + 0.08345327526330948, + 0.11973904073238373, + 0.15909235179424286, + 0.171421617269516, + 0.05466485023498535, + 0.038872141391038895 + ], + "credit_map": [ + 0.00653556315228343, + 0.004015190992504358, + 0.018645890057086945, + 0.006470170337706804, + 0.004851535428315401, + 0.007659749127924442, + 0.008634515106678009, + 0.02017672173678875, + 0.00911919679492712, + 0.010942930355668068, + 0.015173150226473808, + 0.012649924494326115, + 0.007811718620359898, + 0.008568870835006237, + 0.013987070880830288, + 0.03198624402284622, + 0.004416320007294416, + 0.009260175749659538, + 0.0077893720008432865, + 0.006470398046076298, + 0.006044413428753614, + 0.005421813577413559, + 0.009718536399304867, + 0.006058758124709129, + 0.024838510900735855, + 0.018999090418219566, + 0.012831558473408222, + 0.01550974790006876, + 0.010247349739074707, + 0.011978723108768463, + 0.01654578000307083, + 0.019397789612412453, + 0.01476455107331276, + 0.010439502075314522, + 0.011679379269480705, + 0.010223924182355404, + 0.00714788306504488, + 0.010427816770970821, + 0.008372796699404716, + 0.010397421196103096, + 0.022450612857937813, + 0.01483383309096098, + 0.016695227473974228, + 0.017975294962525368, + 0.009961007162928581, + 0.010389142669737339, + 0.016366885975003242, + 0.01106702908873558, + 0.011782576330006123, + 0.01339876838028431, + 0.025869769975543022, + 0.03777429834008217, + 0.018043728545308113, + 0.01438868883997202, + 0.01970871351659298, + 0.018125802278518677, + 0.009380936622619629, + 0.00923101045191288, + 0.01116914488375187, + 0.04694885388016701, + 0.007651268970221281, + 0.01068141683936119, + 0.011084680445492268, + 0.06527432054281235, + 0.004720361437648535, + 0.005445669870823622, + 0.003677397035062313, + 0.006170902866870165, + 0.012814202345907688, + 0.006483017932623625, + 0.008944236673414707, + 0.006409060209989548, + 0.004408219363540411, + 0.003262541489675641, + 0.002128361724317074, + 0.00793365016579628, + 0.005301827564835548, + 0.003135265549644828, + 0.008599159307777882, + 0.004103116225451231 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.07895474135875702 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the surreal presence of a mature giraffe standing in a brightly lit supermarket aisle, its head and long neck jutting out from the top of the frame. The camera then executes a smooth pan to the right, following the animal as it begins to move. The giraffe, its fur a detailed pattern of rich brown polygonal patches separated by cream-colored lines, awkwardly begins to move away from the camera. The movement is both clumsy and restrained; its massive body slowly turns on the polished linen floor, reflecting the harsh, cold-white fluorescent lights overhead. As it completes its turn and begins to walk away, its powerful hind legs and swishy black tail inadvertently sweep across the densely packed metal shelves. This triggers a chaotic cascade as various products—unbranded, brightly colored cereal boxes, cylindrical tin cans, and crumpled snack bags—are sent clattering and tumbling onto the ground. The giraffe continues its steady and deliberate walk down the aisle, its form receding into the distance, seemingly oblivious to the traces of commercial vandalism, creating an atmosphere of startling, deadpan, chaotic comedy.", + "seed": 31809, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31809_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03702926654368639, + "vlm_logprob": -0.03702926654368639, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.03702926654368639, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013585255481302738 + ], + "opposite_logprob_tokens": [ + -4.513585090637207 + ], + "logprob_reward": -0.013585255481302738, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.013585255481302738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17499937117099762, + 0.046023815870285034, + 0.07784789055585861, + 0.060752131044864655, + 0.4499802589416504, + 0.02843306213617325, + 0.026817603036761284, + 0.044314395636320114, + 0.052419569343328476, + 0.03841192275285721 + ], + "credit_map": [ + 0.007200693245977163, + 0.020215952768921852, + 0.01684919185936451, + 0.046784255653619766, + 0.007483832072466612, + 0.009425939060747623, + 0.03584638237953186, + 0.031193122267723083, + 0.005093636456876993, + 0.00521414028480649, + 0.005838209297508001, + 0.00623585470020771, + 0.006523858290165663, + 0.005713643506169319, + 0.005457070656120777, + 0.005947400815784931, + 0.008288180455565453, + 0.004402846563607454, + 0.007483652327209711, + 0.012947425246238708, + 0.011393963359296322, + 0.009220424108207226, + 0.010364795103669167, + 0.013746606186032295, + 0.005929081700742245, + 0.005560725927352905, + 0.004542206879705191, + 0.007921476848423481, + 0.009053848683834076, + 0.010102497413754463, + 0.010255664587020874, + 0.007386628072708845, + 0.019393496215343475, + 0.08224749565124512, + 0.013103931210935116, + 0.01599106565117836, + 0.18131911754608154, + 0.03533794358372688, + 0.05118720978498459, + 0.05140001326799393, + 0.003873379435390234, + 0.002431587316095829, + 0.0027226079255342484, + 0.005082653369754553, + 0.004269187338650227, + 0.002429287414997816, + 0.004917088896036148, + 0.002707271371036768, + 0.0020773918367922306, + 0.0019101989455521107, + 0.0021933126263320446, + 0.003105281852185726, + 0.004548429511487484, + 0.0031483664643019438, + 0.005269829649478197, + 0.004564790520817041, + 0.002443832578137517, + 0.0023726793006062508, + 0.006451712921261787, + 0.0040114023722708225, + 0.010531004518270493, + 0.004775200039148331, + 0.010745452716946602, + 0.002983108162879944, + 0.003915051463991404, + 0.003937331028282642, + 0.005336384754627943, + 0.004924895707517862, + 0.014080104418098927, + 0.007011637091636658, + 0.008253210224211216, + 0.004960951395332813, + 0.005249433219432831, + 0.003364779055118561, + 0.004387029446661472, + 0.007822714745998383, + 0.004189189989119768, + 0.0043485998176038265, + 0.005199441686272621, + 0.003850734792649746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.013585255481302738 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029796700924634933 + ], + "opposite_logprob_tokens": [ + -3.529796600341797 + ], + "logprob_reward": -0.029796700924634933, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029796700924634933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38541528582572937, + 0.045969828963279724, + 0.24499261379241943, + 0.024079259485006332, + 0.13674232363700867, + 0.04499435797333717, + 0.03954951837658882, + 0.026441892609000206, + 0.017489375546574593, + 0.03432551771402359 + ], + "credit_map": [ + 0.00917570386081934, + 0.07925862073898315, + 0.025589946657419205, + 0.1420583873987198, + 0.011721284128725529, + 0.008946489542722702, + 0.08612355589866638, + 0.022541292011737823, + 0.005656992085278034, + 0.009675491601228714, + 0.005539961624890566, + 0.006296911742538214, + 0.006049908697605133, + 0.005328409373760223, + 0.0035683230962604284, + 0.0038538328371942043, + 0.03696209192276001, + 0.0203029103577137, + 0.027389829978346825, + 0.03562634810805321, + 0.030195139348506927, + 0.04398206248879433, + 0.019452832639217377, + 0.031081410124897957, + 0.003380622249096632, + 0.0022164585534483194, + 0.00288213137537241, + 0.0030328312423080206, + 0.002816923428326845, + 0.003700010012835264, + 0.003099293913692236, + 0.0029509919695556164, + 0.006482316181063652, + 0.024637924507260323, + 0.004135994240641594, + 0.0053734928369522095, + 0.05412064492702484, + 0.010483198799192905, + 0.014811194501817226, + 0.016697563230991364, + 0.010770955123007298, + 0.007980525493621826, + 0.0031702942214906216, + 0.003951112739741802, + 0.005353904329240322, + 0.003534836694598198, + 0.005150731187313795, + 0.005082001909613609, + 0.0040429700165987015, + 0.0043752011843025684, + 0.006895532365888357, + 0.0039032516069710255, + 0.005777980200946331, + 0.005280765239149332, + 0.005550473462790251, + 0.0037233440671116114, + 0.0023410881403833628, + 0.003011226886883378, + 0.003962310031056404, + 0.003727436764165759, + 0.003425625851377845, + 0.004362419247627258, + 0.003777372185140848, + 0.0018344172276556492, + 0.0013811994576826692, + 0.001326248748227954, + 0.001909395563416183, + 0.0023706750944256783, + 0.003768069203943014, + 0.002681888174265623, + 0.0022421486210078, + 0.0018097509164363146, + 0.004461158066987991, + 0.003000409109517932, + 0.003684032941237092, + 0.006408765446394682, + 0.004576695617288351, + 0.0040127551183104515, + 0.004978681914508343, + 0.0032030243892222643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.029796700924634933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808467835187912 + ], + "opposite_logprob_tokens": [ + -3.2880847454071045 + ], + "logprob_reward": -0.03808467835187912, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808467835187912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.453809529542923, + 0.06777400523424149, + 0.05260758101940155, + 0.04775814712047577, + 0.12951675057411194, + 0.06592923402786255, + 0.035905174911022186, + 0.039793096482753754, + 0.048980798572301865, + 0.05792570859193802 + ], + "credit_map": [ + 0.012074215337634087, + 0.1099516898393631, + 0.02752738818526268, + 0.17306207120418549, + 0.013206551782786846, + 0.00932535994797945, + 0.08349742740392685, + 0.02516481652855873, + 0.008230676874518394, + 0.010417023673653603, + 0.010307016782462597, + 0.01478387601673603, + 0.005545773543417454, + 0.004329737741500139, + 0.006642660591751337, + 0.0075172362849116325, + 0.006452855654060841, + 0.005932834465056658, + 0.006938522215932608, + 0.00846102274954319, + 0.006031946279108524, + 0.007293075788766146, + 0.005366498604416847, + 0.0061308215372264385, + 0.0061684586107730865, + 0.004053078126162291, + 0.006149179767817259, + 0.004710857290774584, + 0.0073587424121797085, + 0.007205211091786623, + 0.005670547019690275, + 0.006442067679017782, + 0.0058069778606295586, + 0.02163158357143402, + 0.005961222108453512, + 0.004776038229465485, + 0.05293666943907738, + 0.008956233970820904, + 0.012646962888538837, + 0.016801055520772934, + 0.009215254336595535, + 0.005731293000280857, + 0.005651127081364393, + 0.010305970907211304, + 0.010677266865968704, + 0.006169572006911039, + 0.011277408339083195, + 0.0069013419561088085, + 0.0033229535911232233, + 0.0031361933797597885, + 0.004765544552356005, + 0.0032717895228415728, + 0.00657179718837142, + 0.00557480426505208, + 0.005685837008059025, + 0.0035762484185397625, + 0.0027399107348173857, + 0.003387797623872757, + 0.0072135599330067635, + 0.005993279628455639, + 0.006090069655328989, + 0.00529142701998353, + 0.006441861856728792, + 0.0026351853739470243, + 0.0038490639999508858, + 0.0033294400200247765, + 0.004547106102108955, + 0.005333638750016689, + 0.01158520020544529, + 0.006975892931222916, + 0.008852689526975155, + 0.004507764242589474, + 0.004893045872449875, + 0.0041440268978476524, + 0.007306632120162249, + 0.007804763037711382, + 0.009585223160684109, + 0.009649396874010563, + 0.009522289037704468, + 0.0050203269347548485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.03808467835187912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336892269551754 + ], + "opposite_logprob_tokens": [ + -3.408689260482788 + ], + "logprob_reward": -0.0336892269551754, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336892269551754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23155993223190308, + 0.15640679001808167, + 0.09552783519029617, + 0.13031283020973206, + 0.0777973160147667, + 0.061576832085847855, + 0.0677669495344162, + 0.07500878721475601, + 0.0773521140217781, + 0.026690607890486717 + ], + "credit_map": [ + 0.010152556002140045, + 0.05981577932834625, + 0.019346104934811592, + 0.057989317923784256, + 0.011697933077812195, + 0.0106139425188303, + 0.03964944928884506, + 0.02229485660791397, + 0.021104637533426285, + 0.012834032997488976, + 0.02139749564230442, + 0.013652973808348179, + 0.02485242858529091, + 0.0216092299669981, + 0.020666835829615593, + 0.02028915472328663, + 0.009459089487791061, + 0.010017774067819118, + 0.01178805809468031, + 0.01410873606801033, + 0.013989553786814213, + 0.012720463797450066, + 0.010676691308617592, + 0.012767464853823185, + 0.024026945233345032, + 0.012059821747243404, + 0.015611538663506508, + 0.017375607043504715, + 0.014620862901210785, + 0.020565224811434746, + 0.014910288155078888, + 0.011142542585730553, + 0.005270821508020163, + 0.011563523672521114, + 0.003330472158268094, + 0.0048300097696483135, + 0.028196366503834724, + 0.005671457853168249, + 0.009694911539554596, + 0.00923975184559822, + 0.008049539290368557, + 0.00476606423035264, + 0.008937944658100605, + 0.009380429983139038, + 0.008711774833500385, + 0.007448592223227024, + 0.007992771454155445, + 0.006289716809988022, + 0.0061497315764427185, + 0.005187889561057091, + 0.00815653894096613, + 0.0055631306022405624, + 0.008559571579098701, + 0.007585723884403706, + 0.013474523089826107, + 0.013089843094348907, + 0.0033473847433924675, + 0.0056528267450630665, + 0.01067100279033184, + 0.008512932807207108, + 0.01653423346579075, + 0.008896181359887123, + 0.015855854377150536, + 0.005538370925933123, + 0.007517751306295395, + 0.00799486879259348, + 0.010801778174936771, + 0.008817092515528202, + 0.009408405981957912, + 0.014201624318957329, + 0.010873809456825256, + 0.007736780680716038, + 0.0024652944412082434, + 0.002266407711431384, + 0.0034581206273287535, + 0.004576835315674543, + 0.004242110997438431, + 0.003046327270567417, + 0.0043733492493629456, + 0.0022621636744588614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0336892269551754 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999047100543976 + ], + "opposite_logprob_tokens": [ + -2.694990396499634 + ], + "logprob_reward": -0.06999047100543976, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999047100543976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6626380085945129, + 0.05632529407739639, + 0.10781853646039963, + 0.05047065019607544, + 0.023029789328575134, + 0.02068444713950157, + 0.018082736060023308, + 0.020699789747595787, + 0.020750708878040314, + 0.019500039517879486 + ], + "credit_map": [ + 0.010800530202686787, + 0.12924881279468536, + 0.038540925830602646, + 0.22543694078922272, + 0.02313040941953659, + 0.008742783218622208, + 0.17178471386432648, + 0.05495285242795944, + 0.005466051399707794, + 0.00602771807461977, + 0.007171747740358114, + 0.007481792010366917, + 0.008443746715784073, + 0.006881305016577244, + 0.00722428597509861, + 0.007628642953932285, + 0.015727736055850983, + 0.008142892271280289, + 0.01091921329498291, + 0.014732690528035164, + 0.01518566906452179, + 0.0180452112108469, + 0.01042303815484047, + 0.014642083086073399, + 0.007832316681742668, + 0.004354228265583515, + 0.0050934613682329655, + 0.006075426936149597, + 0.006420845165848732, + 0.00787570420652628, + 0.007703916169703007, + 0.005114747676998377, + 0.001752953976392746, + 0.002107605803757906, + 0.0016445186920464039, + 0.0019898563623428345, + 0.007114702835679054, + 0.0018125358037650585, + 0.002723477315157652, + 0.00388413667678833, + 0.0020694502163678408, + 0.001389405457302928, + 0.0016846812795847654, + 0.0020581097342073917, + 0.004976716823875904, + 0.0019370748195797205, + 0.004441014025360346, + 0.002127994317561388, + 0.001554024056531489, + 0.001932057086378336, + 0.00292706536129117, + 0.001821600366383791, + 0.0026258821599185467, + 0.002057103905826807, + 0.00289078033529222, + 0.002274220809340477, + 0.0015788745367899537, + 0.0017462476389482617, + 0.002343760570511222, + 0.0020023416727781296, + 0.00426251906901598, + 0.0032261908054351807, + 0.00418794946745038, + 0.0013519045896828175, + 0.0016040068585425615, + 0.0014642233727499843, + 0.0021979904267936945, + 0.002491278573870659, + 0.003326707985252142, + 0.0041760946623981, + 0.0033478301484137774, + 0.002142573706805706, + 0.002159078372642398, + 0.0017068985616788268, + 0.001926003722473979, + 0.0029267657082527876, + 0.002950507216155529, + 0.002761989366263151, + 0.0032686935737729073, + 0.0018001004355028272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.06999047100543976 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the surreal presence of a mature giraffe standing in a brightly lit supermarket aisle, its head and long neck jutting out from the top of the frame. The camera then executes a smooth pan to the right, following the animal as it begins to move. The giraffe, its fur a detailed pattern of rich brown polygonal patches separated by cream-colored lines, awkwardly begins to move away from the camera. The movement is both clumsy and restrained; its massive body slowly turns on the polished linen floor, reflecting the harsh, cold-white fluorescent lights overhead. As it completes its turn and begins to walk away, its powerful hind legs and swishy black tail inadvertently sweep across the densely packed metal shelves. This triggers a chaotic cascade as various products—unbranded, brightly colored cereal boxes, cylindrical tin cans, and crumpled snack bags—are sent clattering and tumbling onto the ground. The giraffe continues its steady and deliberate walk down the aisle, its form receding into the distance, seemingly oblivious to the traces of commercial vandalism, creating an atmosphere of startling, deadpan, chaotic comedy.", + "seed": 31810, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_2/index_6520_seed_31810_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03569788169115782, + "vlm_logprob": -0.03569788169115782, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.03569788169115782, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009581058286130428 + ], + "opposite_logprob_tokens": [ + -5.134581089019775 + ], + "logprob_reward": -0.009581058286130428, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.009581058286130428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17180785536766052, + 0.07799963653087616, + 0.08327026665210724, + 0.13780656456947327, + 0.036374758929014206, + 0.06093744933605194, + 0.14693273603916168, + 0.11248030513525009, + 0.07001426070928574, + 0.10237617790699005 + ], + "credit_map": [ + 0.01843920908868313, + 0.02694777585566044, + 0.009697985835373402, + 0.03544597700238228, + 0.023579038679599762, + 0.014051657170057297, + 0.0168833639472723, + 0.02676285244524479, + 0.010967188514769077, + 0.010329446755349636, + 0.0088876998052001, + 0.010295317508280277, + 0.009783491492271423, + 0.009969452396035194, + 0.01010623387992382, + 0.007660803385078907, + 0.009529976174235344, + 0.016927450895309448, + 0.007279552984982729, + 0.010893226601183414, + 0.012041360139846802, + 0.006842324044555426, + 0.011694958433508873, + 0.008061415515840054, + 0.017920900136232376, + 0.018777994439005852, + 0.012529225088655949, + 0.0164682287722826, + 0.01722625270485878, + 0.019621631130576134, + 0.01653195358812809, + 0.018730364739894867, + 0.0034472330007702112, + 0.005241947248578072, + 0.004459070041775703, + 0.0030535291880369186, + 0.006660847458988428, + 0.003922806121408939, + 0.0041063264943659306, + 0.005482999142259359, + 0.007983792573213577, + 0.007414722815155983, + 0.0077359662391245365, + 0.005678561516106129, + 0.007742058485746384, + 0.010941801592707634, + 0.0068899220786988735, + 0.0065506258979439735, + 0.019197985529899597, + 0.013529883697628975, + 0.017677582800388336, + 0.021936679258942604, + 0.017448391765356064, + 0.02101120539009571, + 0.02086685039103031, + 0.01526416465640068, + 0.0063970740884542465, + 0.007578641641885042, + 0.01067032478749752, + 0.014334534294903278, + 0.006396911107003689, + 0.01869882643222809, + 0.02354036085307598, + 0.024863624945282936, + 0.008551448583602905, + 0.010972860269248486, + 0.007078059948980808, + 0.011811813339591026, + 0.006295260041952133, + 0.005366113502532244, + 0.013438841328024864, + 0.006499866489320993, + 0.01549497153609991, + 0.007737020496279001, + 0.007672691252082586, + 0.01232858281582594, + 0.00682136882096529, + 0.009395471774041653, + 0.02656537853181362, + 0.016360698267817497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.009581058286130428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183674663305283 + ], + "opposite_logprob_tokens": [ + -4.018183708190918 + ], + "logprob_reward": -0.018183674663305283, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183674663305283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21519094705581665, + 0.07911831885576248, + 0.05313310772180557, + 0.05314243584871292, + 0.06606363505125046, + 0.05371098592877388, + 0.09535655379295349, + 0.1014871746301651, + 0.08559755235910416, + 0.1971992403268814 + ], + "credit_map": [ + 0.040272414684295654, + 0.041661616414785385, + 0.012576496228575706, + 0.02006242237985134, + 0.055459048599004745, + 0.012853416614234447, + 0.010114639066159725, + 0.022190893068909645, + 0.012528084218502045, + 0.012002338655292988, + 0.01182441134005785, + 0.008658732287585735, + 0.010466531850397587, + 0.010337069630622864, + 0.005378807429224253, + 0.007922342978417873, + 0.00742947356775403, + 0.011729148216545582, + 0.00440184073522687, + 0.005826933309435844, + 0.007366103120148182, + 0.006048247218132019, + 0.005931196268647909, + 0.004400167614221573, + 0.00663912994787097, + 0.00653503555804491, + 0.005604034289717674, + 0.009766356088221073, + 0.00841781497001648, + 0.0037936572916805744, + 0.008621128275990486, + 0.0037652812898159027, + 0.007801800966262817, + 0.00896791834384203, + 0.0077299396507442, + 0.005218819715082645, + 0.012366272509098053, + 0.008099455386400223, + 0.007407783064991236, + 0.008471643552184105, + 0.006673004478216171, + 0.006720159202814102, + 0.007185077760368586, + 0.0055738771334290504, + 0.005844515282660723, + 0.007363876793533564, + 0.005099377129226923, + 0.00925110001116991, + 0.012281903997063637, + 0.008186188526451588, + 0.011342868208885193, + 0.01577155478298664, + 0.007329897489398718, + 0.017570171505212784, + 0.014000647701323032, + 0.008873330429196358, + 0.01597663201391697, + 0.015017497353255749, + 0.01733843982219696, + 0.010832159779965878, + 0.008924880996346474, + 0.013132796622812748, + 0.012349498458206654, + 0.00791527796536684, + 0.012059402652084827, + 0.013392978347837925, + 0.01109977439045906, + 0.012108804658055305, + 0.011434424668550491, + 0.0070351492613554, + 0.009996546432375908, + 0.008470476604998112, + 0.011687063612043858, + 0.013118674047291279, + 0.011732117272913456, + 0.02580026537179947, + 0.011678257025778294, + 0.019079338759183884, + 0.062151696532964706, + 0.04195185378193855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.018183674663305283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808605298399925 + ], + "opposite_logprob_tokens": [ + -3.2880859375 + ], + "logprob_reward": -0.03808605298399925, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808605298399925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07680821418762207, + 0.07064934074878693, + 0.041789710521698, + 0.11419854313135147, + 0.047184232622385025, + 0.07957327365875244, + 0.2918475866317749, + 0.12676095962524414, + 0.0966055765748024, + 0.05458258092403412 + ], + "credit_map": [ + 0.01331317238509655, + 0.012105497531592846, + 0.004848360549658537, + 0.009233694523572922, + 0.0172413382679224, + 0.00498397508636117, + 0.006451175082474947, + 0.008630997501313686, + 0.012430231086909771, + 0.00989757850766182, + 0.009303973987698555, + 0.00771807786077261, + 0.008415265940129757, + 0.006805650424212217, + 0.0070625003427267075, + 0.009016052819788456, + 0.0064027137123048306, + 0.006615811493247747, + 0.004600122105330229, + 0.004124271683394909, + 0.00499720498919487, + 0.005775380413979292, + 0.005406368989497423, + 0.0038678350392729044, + 0.018394066020846367, + 0.014210790395736694, + 0.011268225498497486, + 0.017800409346818924, + 0.015946589410305023, + 0.00949086807668209, + 0.015145497396588326, + 0.011942088603973389, + 0.006305002141743898, + 0.005421551410108805, + 0.008282212540507317, + 0.004292161203920841, + 0.006452443078160286, + 0.0032956551294773817, + 0.005172473844140768, + 0.007962730713188648, + 0.007028144784271717, + 0.008295174688100815, + 0.012029037810862064, + 0.006846964359283447, + 0.009883789345622063, + 0.010782381519675255, + 0.01103080902248621, + 0.013676965609192848, + 0.03710537031292915, + 0.02578376978635788, + 0.03443743288516998, + 0.039960555732250214, + 0.033566851168870926, + 0.046186067163944244, + 0.044199138879776, + 0.030608389526605606, + 0.009896272793412209, + 0.01035515870898962, + 0.015572111122310162, + 0.012399001978337765, + 0.010467194952070713, + 0.018526067957282066, + 0.02471897378563881, + 0.024826180189847946, + 0.011061894707381725, + 0.009210600517690182, + 0.010420587845146656, + 0.02007535845041275, + 0.011180886998772621, + 0.00897209718823433, + 0.0134113859385252, + 0.012272756546735764, + 0.005989360623061657, + 0.004627118818461895, + 0.004464345518499613, + 0.007913985289633274, + 0.003389001591131091, + 0.0048923185095191, + 0.015472715720534325, + 0.007833725772798061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.03808605298399925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033687956631183624 + ], + "opposite_logprob_tokens": [ + -3.4086880683898926 + ], + "logprob_reward": -0.033687956631183624, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033687956631183624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14921355247497559, + 0.052773796021938324, + 0.062032535672187805, + 0.10953272879123688, + 0.17832142114639282, + 0.06223936006426811, + 0.1270698606967926, + 0.0750863328576088, + 0.06096527352929115, + 0.12276512384414673 + ], + "credit_map": [ + 0.027725176885724068, + 0.024814514443278313, + 0.008439961820840836, + 0.01970859430730343, + 0.0323500856757164, + 0.008305167779326439, + 0.012813476845622063, + 0.015056573785841465, + 0.008849573321640491, + 0.008121958002448082, + 0.0071634831838309765, + 0.005876844748854637, + 0.007426900789141655, + 0.007242347579449415, + 0.0033486969769001007, + 0.004743989557027817, + 0.011389913968741894, + 0.008499928750097752, + 0.007191254757344723, + 0.007080259267240763, + 0.008969346061348915, + 0.007387982681393623, + 0.00601142318919301, + 0.005502423271536827, + 0.01753643900156021, + 0.014651528559625149, + 0.010595287196338177, + 0.016857856884598732, + 0.014853338710963726, + 0.00957542099058628, + 0.013071740046143532, + 0.012391122989356518, + 0.021999821066856384, + 0.01882658526301384, + 0.04325008764863014, + 0.015318518504500389, + 0.022496432065963745, + 0.01626146025955677, + 0.017561769112944603, + 0.022606750950217247, + 0.007850708439946175, + 0.006795018445700407, + 0.007601388264447451, + 0.005510429851710796, + 0.009587239474058151, + 0.00827937014400959, + 0.00895520020276308, + 0.0076600066386163235, + 0.018905816599726677, + 0.012285055592656136, + 0.01594281569123268, + 0.018428048118948936, + 0.01488306000828743, + 0.01679520681500435, + 0.017572516575455666, + 0.01225734781473875, + 0.00998067669570446, + 0.010446523316204548, + 0.012244857847690582, + 0.006465064827352762, + 0.009105450473725796, + 0.009565320797264576, + 0.009127633646130562, + 0.008150806650519371, + 0.004962899722158909, + 0.0052694897167384624, + 0.0056608691811561584, + 0.011251169256865978, + 0.004435683134943247, + 0.008698135614395142, + 0.010558436624705791, + 0.010128594003617764, + 0.01229399349540472, + 0.007742480840533972, + 0.008882435970008373, + 0.017399100586771965, + 0.006869271397590637, + 0.008782275952398777, + 0.03596826270222664, + 0.02482730895280838 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.033687956631183624 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mature giraffe with rich brown polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a brightly lit supermarket aisle with densely packed metal shelves and a polished linen floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the giraffe perform a clumsy, restrained turn on its massive body before walking away?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do unbranded cereal boxes, cylindrical tin cans, and crumpled snack bags tumble onto the ground after being swept by the giraffe's legs and tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth pan to the right, following the giraffe as it moves away?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0789506658911705 + ], + "opposite_logprob_tokens": [ + -2.5789506435394287 + ], + "logprob_reward": -0.0789506658911705, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.0789506658911705, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12588952481746674, + 0.0416610948741436, + 0.03009830042719841, + 0.07405272871255875, + 0.07856451719999313, + 0.14635030925273895, + 0.2810043394565582, + 0.07831288874149323, + 0.09662012755870819, + 0.04744614660739899 + ], + "credit_map": [ + 0.020244646817445755, + 0.022627737373113632, + 0.0077889165841042995, + 0.013744653202593327, + 0.03025014139711857, + 0.008441406302154064, + 0.007034029811620712, + 0.015757989138364792, + 0.006163312122225761, + 0.006200285162776709, + 0.0038794083520770073, + 0.005483797285705805, + 0.006978568155318499, + 0.00469393702223897, + 0.004326047375798225, + 0.0039357407949864864, + 0.005317977163940668, + 0.003559789387509227, + 0.0032901763916015625, + 0.003370834980159998, + 0.0037062098272144794, + 0.0032851560972630978, + 0.0041310349479317665, + 0.0034371218644082546, + 0.010545969009399414, + 0.008799171075224876, + 0.007175500504672527, + 0.009813808836042881, + 0.009658059105277061, + 0.009530536830425262, + 0.012866425327956676, + 0.005663258023560047, + 0.009041995741426945, + 0.009706070646643639, + 0.011290068738162518, + 0.009186917915940285, + 0.012487032450735569, + 0.007305081468075514, + 0.009375393390655518, + 0.010171957314014435, + 0.015374062582850456, + 0.014127332717180252, + 0.03338940441608429, + 0.010075526311993599, + 0.01201825775206089, + 0.016794292256236076, + 0.014925834722816944, + 0.029645590111613274, + 0.035917531698942184, + 0.02299925871193409, + 0.032765213400125504, + 0.039080142974853516, + 0.02622338943183422, + 0.049341831356287, + 0.04362085461616516, + 0.031056128442287445, + 0.00831456296145916, + 0.0064780632965266705, + 0.011388832703232765, + 0.00799322035163641, + 0.006438952870666981, + 0.012386718764901161, + 0.013074342161417007, + 0.01223819237202406, + 0.011280825361609459, + 0.012207621708512306, + 0.01128771435469389, + 0.01931467466056347, + 0.01216900534927845, + 0.008267303928732872, + 0.010510772466659546, + 0.011582216247916222, + 0.003002371871843934, + 0.00275514111854136, + 0.002595178782939911, + 0.006185419857501984, + 0.0026661595329642296, + 0.004360170103609562, + 0.01457713358104229, + 0.011304570361971855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0789506658911705 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_tensor([639, 639])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_tensor([639, 639])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..50385f8b90d1bb03fb3149b2921629a067951b3e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_tensor([639, 639])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight, extreme close-up, positioned at a slightly elevated angle, remains almost static as it focuses on a polished, antique gyroscope. The precisely machined brass gimbal, radiating a warm golden sheen, catches the light from a single, focused source outside the lens, creating bright, moving highlights on its curved surface. The central flywheel spins at a dizzying pace, becoming a solid, silvery-gray blur that emits a faint, high-frequency hum. As the rotor maintains its rapid speed, the interconnected gimbals begin to precess, rotating and tilting with a slow, deliberate, and hypnotic grace. The entire mechanism glides smoothly across the surface of a dark, rich mahogany table, its smooth surface reflecting the gyroscope's gravity-defying dance in a twisted, shadowed mirror. The interplay of warm, focused light and deep shadows cast on the dark wood creates an atmosphere of focus, contemplation, and almost magical, celebrating the beauty of the moving mechanism.", + "seed": 11460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11460_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.029421449918299914, + "vlm_logprob": -0.029421449918299914, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.029421449918299914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03239701688289642 + ], + "opposite_logprob_tokens": [ + -4.0323967933654785 + ], + "logprob_reward": -0.03239701688289642, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.03239701688289642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3546074628829956, + 0.14702612161636353, + 0.06498776376247406, + 0.09150877594947815, + 0.11714687198400497, + 0.049002449959516525, + 0.050648435950279236, + 0.027470633387565613, + 0.03874604403972626, + 0.058855436742305756 + ], + "credit_map": [ + 0.048924874514341354, + 0.026016641408205032, + 0.05478103831410408, + 0.025141166523098946, + 0.043099287897348404, + 0.04937787726521492, + 0.03811318054795265, + 0.06915339082479477, + 0.03380398824810982, + 0.0105288065969944, + 0.02033178135752678, + 0.011292105540633202, + 0.013555848971009254, + 0.0199155081063509, + 0.013024742715060711, + 0.024573346599936485, + 0.009135096333920956, + 0.0055792080238461494, + 0.010105039924383163, + 0.004205884877592325, + 0.006533351261168718, + 0.008253182284533978, + 0.010021609254181385, + 0.011154398322105408, + 0.012767898850142956, + 0.007689724676311016, + 0.012460160069167614, + 0.00859291385859251, + 0.009702780283987522, + 0.013773624785244465, + 0.012226137332618237, + 0.014295542612671852, + 0.014902973547577858, + 0.00938019622117281, + 0.01560213789343834, + 0.011704277247190475, + 0.014696703292429447, + 0.015244750306010246, + 0.01633422262966633, + 0.019281618297100067, + 0.006617091130465269, + 0.003176060738041997, + 0.007252334151417017, + 0.004050536546856165, + 0.005537847522646189, + 0.007060152944177389, + 0.005062124691903591, + 0.010246305726468563, + 0.00694085331633687, + 0.005723569542169571, + 0.009132486768066883, + 0.004636150784790516, + 0.0041310833767056465, + 0.006766791921108961, + 0.0055595641024410725, + 0.007757936604321003, + 0.003059262642636895, + 0.002456048270687461, + 0.0033372710458934307, + 0.002454628935083747, + 0.002294760663062334, + 0.004351743962615728, + 0.0032076409552246332, + 0.00630927924066782, + 0.005968149285763502, + 0.003605901263654232, + 0.006024807691574097, + 0.004496574401855469, + 0.005208338610827923, + 0.004919111263006926, + 0.00292736804112792, + 0.005595795810222626, + 0.010512659326195717, + 0.003590948646888137, + 0.013692955486476421, + 0.005497802514582872, + 0.007118078880012035, + 0.006706553976982832, + 0.004698794800788164, + 0.007037647999823093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03239701688289642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005276088137179613 + ], + "opposite_logprob_tokens": [ + -5.255276203155518 + ], + "logprob_reward": -0.005276088137179613, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005276088137179613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3157077729701996, + 0.12738636136054993, + 0.06234423816204071, + 0.04356256499886513, + 0.10337135195732117, + 0.04018570855259895, + 0.08694916218519211, + 0.07509896159172058, + 0.09025061875581741, + 0.05514322221279144 + ], + "credit_map": [ + 0.04491612687706947, + 0.024487046524882317, + 0.04770262911915779, + 0.019275955855846405, + 0.036138616502285004, + 0.04962047562003136, + 0.025233738124370575, + 0.0683332234621048, + 0.020013749599456787, + 0.009439211338758469, + 0.020021315664052963, + 0.009422382339835167, + 0.012213188223540783, + 0.019009310752153397, + 0.01026250422000885, + 0.02700471132993698, + 0.011008222587406635, + 0.005129842087626457, + 0.008305970579385757, + 0.004361867904663086, + 0.005773925222456455, + 0.008761027827858925, + 0.007031762041151524, + 0.011971627362072468, + 0.008040797896683216, + 0.0033168953377753496, + 0.00587663846090436, + 0.0050301095470786095, + 0.003458279650658369, + 0.0059425318613648415, + 0.006038260646164417, + 0.005859059281647205, + 0.013263878412544727, + 0.009158826433122158, + 0.015166452154517174, + 0.011236368678510189, + 0.01330698374658823, + 0.012822004035115242, + 0.015102406032383442, + 0.013314448297023773, + 0.005406837444752455, + 0.002771102823317051, + 0.007245596498250961, + 0.0031218298245221376, + 0.0037958198226988316, + 0.005877945106476545, + 0.005352366250008345, + 0.006614219397306442, + 0.008629529736936092, + 0.007224929518997669, + 0.0133562246337533, + 0.008332462050020695, + 0.013610572554171085, + 0.011115611530840397, + 0.011112742125988007, + 0.013567101210355759, + 0.007106119766831398, + 0.004239713307470083, + 0.010440762154757977, + 0.00481760036200285, + 0.006198269780725241, + 0.013397462666034698, + 0.0064780497923493385, + 0.022420991212129593, + 0.01634225994348526, + 0.0066359993070364, + 0.014102255925536156, + 0.007597627118229866, + 0.009225322864949703, + 0.012968719005584717, + 0.005549857392907143, + 0.01782858930528164, + 0.0078003378584980965, + 0.0038891867734491825, + 0.009541154839098454, + 0.004590502008795738, + 0.00760295707732439, + 0.007361557800322771, + 0.005899058189243078, + 0.008458471857011318 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005276088137179613 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02331020124256611 + ], + "opposite_logprob_tokens": [ + -3.7733101844787598 + ], + "logprob_reward": -0.02331020124256611, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02331020124256611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4554956257343292, + 0.1133679524064064, + 0.06587434560060501, + 0.07566681504249573, + 0.08177303522825241, + 0.04402792826294899, + 0.0566321536898613, + 0.04179491102695465, + 0.035974081605672836, + 0.029393108561635017 + ], + "credit_map": [ + 0.11785326898097992, + 0.02349443919956684, + 0.06956635415554047, + 0.023984458297491074, + 0.03828198462724686, + 0.06684572249650955, + 0.035174109041690826, + 0.08029531687498093, + 0.015490192919969559, + 0.0113024627789855, + 0.016162624582648277, + 0.01120208203792572, + 0.012299270369112492, + 0.01592542976140976, + 0.010696029290556908, + 0.02028985694050789, + 0.009350608102977276, + 0.005932642146945, + 0.01056315191090107, + 0.004843603353947401, + 0.00937277264893055, + 0.00923977978527546, + 0.008019073866307735, + 0.008552714250981808, + 0.009039019234478474, + 0.004205189645290375, + 0.010702473111450672, + 0.00498433131724596, + 0.007581524085253477, + 0.016999023035168648, + 0.008108790032565594, + 0.014046471565961838, + 0.011193981394171715, + 0.006041483022272587, + 0.009809071198105812, + 0.005177359562367201, + 0.006335102953016758, + 0.011671734042465687, + 0.012166332453489304, + 0.019377972930669785, + 0.005357966758310795, + 0.004156529437750578, + 0.00598168047145009, + 0.0037014069966971874, + 0.0034864575136452913, + 0.006446296349167824, + 0.006190649699419737, + 0.008706944063305855, + 0.008594861254096031, + 0.0077299936674535275, + 0.007174968253821135, + 0.005428496748209, + 0.00717609329149127, + 0.006801635958254337, + 0.006448580417782068, + 0.007277528755366802, + 0.004650269169360399, + 0.0029736405704170465, + 0.004647743422538042, + 0.002798199187964201, + 0.004771905485540628, + 0.0068770600482821465, + 0.004610047210007906, + 0.010466051287949085, + 0.004435679409652948, + 0.0038632098585367203, + 0.0061184377409517765, + 0.0039391047321259975, + 0.00495991762727499, + 0.004004764836281538, + 0.0033183128107339144, + 0.005334657616913319, + 0.003855873830616474, + 0.0024939808063209057, + 0.003958512097597122, + 0.0027836260851472616, + 0.006448734551668167, + 0.003469911403954029, + 0.0036939112469553947, + 0.002688559703528881 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02331020124256611 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305883124470711 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305883124470711, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305883124470711, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26277032494544983, + 0.10695835947990417, + 0.08940333873033524, + 0.06720587611198425, + 0.13818731904029846, + 0.05133388936519623, + 0.07488413155078888, + 0.09836623072624207, + 0.0866609439253807, + 0.02422957494854927 + ], + "credit_map": [ + 0.053569164127111435, + 0.01408227439969778, + 0.04981605336070061, + 0.019755063578486443, + 0.024455728009343147, + 0.036827705800533295, + 0.01891191676259041, + 0.045352429151535034, + 0.0226926077157259, + 0.007092917338013649, + 0.01975327730178833, + 0.008074590936303139, + 0.0064122737385332584, + 0.015468212775886059, + 0.009589452296495438, + 0.017875030636787415, + 0.010331348516047001, + 0.007335382979363203, + 0.014212171547114849, + 0.006061570718884468, + 0.012616099789738655, + 0.012688796035945415, + 0.013232018798589706, + 0.012925959192216396, + 0.006600173190236092, + 0.004849040415138006, + 0.011042013764381409, + 0.006284239701926708, + 0.006332869175821543, + 0.011702477000653744, + 0.008208049461245537, + 0.012187017127871513, + 0.01581001840531826, + 0.012318509630858898, + 0.020101377740502357, + 0.012974961660802364, + 0.016705619171261787, + 0.017608147114515305, + 0.020737608894705772, + 0.021931063383817673, + 0.005515982396900654, + 0.004768731072545052, + 0.00847737118601799, + 0.004760437645018101, + 0.005686029326170683, + 0.007116371765732765, + 0.00653695547953248, + 0.008472008630633354, + 0.010581201873719692, + 0.009887478314340115, + 0.009329859167337418, + 0.007361337076872587, + 0.008808902464807034, + 0.0095879677683115, + 0.00866143312305212, + 0.0106659522280097, + 0.009031285531818867, + 0.00744539313018322, + 0.011263192631304264, + 0.006954623851925135, + 0.012519175186753273, + 0.016682790592312813, + 0.008194725960493088, + 0.026275038719177246, + 0.009235094301402569, + 0.009640350006520748, + 0.012319251894950867, + 0.008269568905234337, + 0.01204716693609953, + 0.011523773893713951, + 0.008078324608504772, + 0.015547415241599083, + 0.003139899345114827, + 0.002114912262186408, + 0.0035628669429570436, + 0.0021757432259619236, + 0.004026293754577637, + 0.0035089165903627872, + 0.0026578877586871386, + 0.003043054137378931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.04305883124470711 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043065112084150314 + ], + "opposite_logprob_tokens": [ + -3.168065071105957 + ], + "logprob_reward": -0.043065112084150314, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043065112084150314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3259434103965759, + 0.16437990963459015, + 0.08154981583356857, + 0.052153199911117554, + 0.09035041183233261, + 0.06487415730953217, + 0.08432973921298981, + 0.06784936785697937, + 0.04278397560119629, + 0.025786029174923897 + ], + "credit_map": [ + 0.06897789239883423, + 0.02111046016216278, + 0.05051277205348015, + 0.01595122367143631, + 0.027996065095067024, + 0.05143815279006958, + 0.024215741083025932, + 0.06574107706546783, + 0.026551293209195137, + 0.01314620766788721, + 0.025762973353266716, + 0.010192442685365677, + 0.011820775456726551, + 0.02755056507885456, + 0.0170252937823534, + 0.032330356538295746, + 0.009052804671227932, + 0.007258899509906769, + 0.013208966702222824, + 0.006557118147611618, + 0.011051392182707787, + 0.010335796512663364, + 0.009518513455986977, + 0.014566324651241302, + 0.007808321621268988, + 0.0036246636882424355, + 0.00865845289081335, + 0.00444099772721529, + 0.004963770974427462, + 0.00718738092109561, + 0.005410861689597368, + 0.010058755055069923, + 0.011628584936261177, + 0.006670947186648846, + 0.01318572461605072, + 0.006006282288581133, + 0.009333697147667408, + 0.014077331870794296, + 0.011347614228725433, + 0.018100233748555183, + 0.006292583420872688, + 0.004092690069228411, + 0.009793415665626526, + 0.006181025877594948, + 0.003716361243277788, + 0.010540436953306198, + 0.008462119847536087, + 0.015795525163412094, + 0.012021403759717941, + 0.00817266944795847, + 0.01302463561296463, + 0.0069164070300757885, + 0.00715726800262928, + 0.012062673456966877, + 0.009620229713618755, + 0.015354448929429054, + 0.007914019748568535, + 0.004741797689348459, + 0.007544336840510368, + 0.004529409576207399, + 0.008384611457586288, + 0.010993939824402332, + 0.00639527989551425, + 0.0173459742218256, + 0.006709837820380926, + 0.003737455466762185, + 0.008218446746468544, + 0.005055200774222612, + 0.003525908337906003, + 0.004719778895378113, + 0.003941495902836323, + 0.006875853054225445, + 0.004967560991644859, + 0.001831185189075768, + 0.005759391933679581, + 0.0018567235674709082, + 0.002173866843804717, + 0.0034624054096639156, + 0.0021990661043673754, + 0.0035358290188014507 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.043065112084150314 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, extreme close-up, positioned at a slightly elevated angle, remains almost static as it focuses on a polished, antique gyroscope. The precisely machined brass gimbal, radiating a warm golden sheen, catches the light from a single, focused source outside the lens, creating bright, moving highlights on its curved surface. The central flywheel spins at a dizzying pace, becoming a solid, silvery-gray blur that emits a faint, high-frequency hum. As the rotor maintains its rapid speed, the interconnected gimbals begin to precess, rotating and tilting with a slow, deliberate, and hypnotic grace. The entire mechanism glides smoothly across the surface of a dark, rich mahogany table, its smooth surface reflecting the gyroscope's gravity-defying dance in a twisted, shadowed mirror. The interplay of warm, focused light and deep shadows cast on the dark wood creates an atmosphere of focus, contemplation, and almost magical, celebrating the beauty of the moving mechanism.", + "seed": 11461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11461_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03021178636699915, + "vlm_logprob": -0.03021178636699915, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03021178636699915, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07061683386564255 + ], + "opposite_logprob_tokens": [ + -2.8206167221069336 + ], + "logprob_reward": -0.07061683386564255, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07061683386564255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18479064106941223, + 0.12426543235778809, + 0.05721089243888855, + 0.029987536370754242, + 0.05491612106561661, + 0.03174188733100891, + 0.1174892783164978, + 0.05159248784184456, + 0.11960319429636002, + 0.2284025102853775 + ], + "credit_map": [ + 0.019170545041561127, + 0.019599897786974907, + 0.025485165417194366, + 0.025731807574629784, + 0.02493981458246708, + 0.018326401710510254, + 0.030871927738189697, + 0.020665092393755913, + 0.014285861514508724, + 0.014225243590772152, + 0.01585613563656807, + 0.02621813863515854, + 0.01243980135768652, + 0.013068354688584805, + 0.014755230396986008, + 0.01341666467487812, + 0.005636200308799744, + 0.005724041722714901, + 0.008581592701375484, + 0.012485027313232422, + 0.004706296604126692, + 0.0071602268144488335, + 0.0056419833563268185, + 0.00727552454918623, + 0.003008531406521797, + 0.005077878944575787, + 0.0042988914065063, + 0.005829685367643833, + 0.0021849432960152626, + 0.0024289737921208143, + 0.003305430756881833, + 0.003853202098980546, + 0.0068021803162992, + 0.00644815294072032, + 0.008100468665361404, + 0.008531874977052212, + 0.005326246842741966, + 0.006721471901983023, + 0.006049474235624075, + 0.006936250254511833, + 0.003644688054919243, + 0.0038152134511619806, + 0.0049593886360526085, + 0.005734904669225216, + 0.0027672534342855215, + 0.0041952310130000114, + 0.0029040579684078693, + 0.0037211524322628975, + 0.011915119364857674, + 0.02159801311790943, + 0.016752757132053375, + 0.014723623171448708, + 0.011385427787899971, + 0.01825268566608429, + 0.009973980486392975, + 0.012887680903077126, + 0.0061408840119838715, + 0.008513080887496471, + 0.0076662530191242695, + 0.006768316496163607, + 0.0061284746043384075, + 0.006283184979110956, + 0.004578400868922472, + 0.005513891112059355, + 0.01536400243639946, + 0.022188561037182808, + 0.019423091784119606, + 0.015206060372292995, + 0.010495107620954514, + 0.013910098932683468, + 0.009474217891693115, + 0.013542046770453453, + 0.021802201867103577, + 0.024525213986635208, + 0.03457220271229744, + 0.03421482816338539, + 0.027990924194455147, + 0.03507542237639427, + 0.024000387638807297, + 0.026221320033073425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.07061683386564255 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005966234486550093 + ], + "opposite_logprob_tokens": [ + -5.1309661865234375 + ], + "logprob_reward": -0.005966234486550093, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005966234486550093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14410540461540222, + 0.08497531712055206, + 0.08888649940490723, + 0.07373657822608948, + 0.1701248288154602, + 0.05738132447004318, + 0.0633237436413765, + 0.060938432812690735, + 0.17442312836647034, + 0.08210468292236328 + ], + "credit_map": [ + 0.018152965232729912, + 0.02432072162628174, + 0.01993543654680252, + 0.01675475761294365, + 0.02036765217781067, + 0.016252759844064713, + 0.015586299821734428, + 0.012734820134937763, + 0.011143967509269714, + 0.01333601400256157, + 0.01143790502101183, + 0.014669883996248245, + 0.007206595037132502, + 0.009472372941672802, + 0.009743955917656422, + 0.007964626885950565, + 0.010501481592655182, + 0.009393899701535702, + 0.011783521622419357, + 0.014430423267185688, + 0.012303988449275494, + 0.011514285579323769, + 0.010824603028595448, + 0.00813429243862629, + 0.006072604563087225, + 0.01381426490843296, + 0.009955240413546562, + 0.012778526172041893, + 0.007395958062261343, + 0.007722146343439817, + 0.009156460873782635, + 0.0068413810804486275, + 0.016248926520347595, + 0.02709573693573475, + 0.0245585348457098, + 0.023307614028453827, + 0.016921598464250565, + 0.022278716787695885, + 0.015034309588372707, + 0.0246793981641531, + 0.006325606722384691, + 0.009105177596211433, + 0.008255126886069775, + 0.012097762897610664, + 0.005750544834882021, + 0.006113105919212103, + 0.005427198018878698, + 0.0043068029917776585, + 0.0069351112470030785, + 0.009193292818963528, + 0.009524871595203876, + 0.008351590484380722, + 0.007319355383515358, + 0.007702278904616833, + 0.006463628262281418, + 0.007833610288798809, + 0.005570957437157631, + 0.01001039519906044, + 0.008374843746423721, + 0.009234893135726452, + 0.00924003031104803, + 0.004860976245254278, + 0.007252193987369537, + 0.006394145078957081, + 0.020516416057944298, + 0.022791093215346336, + 0.025018764659762383, + 0.026154084131121635, + 0.012075381353497505, + 0.016646558418869972, + 0.014766101725399494, + 0.03645474836230278, + 0.009518874809145927, + 0.011098201386630535, + 0.014519203454256058, + 0.017037061974406242, + 0.005987678654491901, + 0.010656938888132572, + 0.006010271608829498, + 0.007276459131389856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 0 + ], + "selected_reward": -0.005966234486550093 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076450384221971035 + ], + "opposite_logprob_tokens": [ + -4.882645130157471 + ], + "logprob_reward": -0.0076450384221971035, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076450384221971035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23760661482810974, + 0.0783597007393837, + 0.08457502722740173, + 0.05679795891046524, + 0.12973441183567047, + 0.03152182698249817, + 0.10835113376379013, + 0.08072192221879959, + 0.08646406978368759, + 0.10586728900671005 + ], + "credit_map": [ + 0.018534120172262192, + 0.03228062763810158, + 0.03294887766242027, + 0.02701684460043907, + 0.041030947118997574, + 0.02238457091152668, + 0.0396861657500267, + 0.023724447935819626, + 0.007825641892850399, + 0.006331854499876499, + 0.012134160846471786, + 0.016962097957730293, + 0.0065439678728580475, + 0.00626246677711606, + 0.00981107447296381, + 0.012488434091210365, + 0.009617600589990616, + 0.009694257751107216, + 0.012098075821995735, + 0.012229502201080322, + 0.01124365534633398, + 0.009329939261078835, + 0.010477903299033642, + 0.009884089231491089, + 0.006709774024784565, + 0.0062823970802128315, + 0.007927833124995232, + 0.008908417075872421, + 0.006287765689194202, + 0.007873689755797386, + 0.006662303116172552, + 0.006145778112113476, + 0.012681310065090656, + 0.019769810140132904, + 0.01706780306994915, + 0.022008417174220085, + 0.012353087775409222, + 0.015627190470695496, + 0.010599268600344658, + 0.019627531990408897, + 0.0031532212160527706, + 0.006286672782152891, + 0.0040711937472224236, + 0.0037697486113756895, + 0.004122385289520025, + 0.0032380849588662386, + 0.0034885196946561337, + 0.003391999052837491, + 0.013966485857963562, + 0.013317924924194813, + 0.017074933275580406, + 0.02604096382856369, + 0.008357221260666847, + 0.01098736934363842, + 0.007455325685441494, + 0.011150906793773174, + 0.008685242384672165, + 0.012539342045783997, + 0.011744841001927853, + 0.010722445324063301, + 0.01077158935368061, + 0.009166944772005081, + 0.007845885120332241, + 0.009245636872947216, + 0.008090665563941002, + 0.012228022329509258, + 0.013015243224799633, + 0.012479681521654129, + 0.00740864546969533, + 0.00927445013076067, + 0.008564600721001625, + 0.015402758494019508, + 0.012510771863162518, + 0.01673858053982258, + 0.01550181582570076, + 0.01579866372048855, + 0.007901044562458992, + 0.01814626343548298, + 0.007106880657374859, + 0.012163273058831692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0076450384221971035 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01817922666668892 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.01817922666668892, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01817922666668892, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09921129792928696, + 0.13378088176250458, + 0.059490833431482315, + 0.06821530312299728, + 0.16428858041763306, + 0.0486728809773922, + 0.038443323224782944, + 0.23533740639686584, + 0.11540070176124573, + 0.03715880215167999 + ], + "credit_map": [ + 0.008834599517285824, + 0.013437456451356411, + 0.014656883664429188, + 0.016915127635002136, + 0.009848702698946, + 0.011782489717006683, + 0.012271839193999767, + 0.011464199982583523, + 0.01596037857234478, + 0.01086189690977335, + 0.01797645352780819, + 0.025089073926210403, + 0.01225335244089365, + 0.014266600832343102, + 0.014583410695195198, + 0.022789709270000458, + 0.005486928857862949, + 0.006495206151157618, + 0.009273060597479343, + 0.013967783190310001, + 0.006453607697039843, + 0.005440350156277418, + 0.007211096119135618, + 0.00516280485317111, + 0.005728145595639944, + 0.009659158065915108, + 0.009032097645103931, + 0.010697218589484692, + 0.008128797635436058, + 0.008241243660449982, + 0.008665747940540314, + 0.008062897250056267, + 0.015613680705428123, + 0.026382489129900932, + 0.02179623208940029, + 0.023337554186582565, + 0.01590081676840782, + 0.020636573433876038, + 0.015596279874444008, + 0.025024957954883575, + 0.006577002350240946, + 0.00434298999607563, + 0.008150598034262657, + 0.011037679389119148, + 0.0037580139469355345, + 0.004729731939733028, + 0.005482816603034735, + 0.004594048485159874, + 0.004215751308947802, + 0.005803968757390976, + 0.005752831231802702, + 0.005802464671432972, + 0.003834781236946583, + 0.005478099454194307, + 0.0036721413489431143, + 0.003883281722664833, + 0.02585900016129017, + 0.03740084916353226, + 0.033204030245542526, + 0.03055751509964466, + 0.03188861906528473, + 0.026814864948391914, + 0.024512359872460365, + 0.025100158527493477, + 0.011541864834725857, + 0.017225319519639015, + 0.017302462831139565, + 0.014110120013356209, + 0.011454636231064796, + 0.012330758385360241, + 0.009305641986429691, + 0.022129906341433525, + 0.0033391050528734922, + 0.0054051922634243965, + 0.0050629922188818455, + 0.006975415628403425, + 0.003241403726860881, + 0.005854255519807339, + 0.0032340027391910553, + 0.004046434536576271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 1 + ], + "selected_reward": -0.01817922666668892 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048651598393917084 + ], + "opposite_logprob_tokens": [ + -3.048651695251465 + ], + "logprob_reward": -0.048651598393917084, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048651598393917084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22211980819702148, + 0.13976384699344635, + 0.032575491815805435, + 0.021317193284630775, + 0.06132754310965538, + 0.037666093558073044, + 0.15062843263149261, + 0.04844934120774269, + 0.2336023598909378, + 0.052549902349710464 + ], + "credit_map": [ + 0.018065467476844788, + 0.03041798807680607, + 0.03337808698415756, + 0.03867482393980026, + 0.03317062556743622, + 0.02140546590089798, + 0.025340620428323746, + 0.021666720509529114, + 0.014696965925395489, + 0.012731833383440971, + 0.019890664145350456, + 0.025665760040283203, + 0.014667154289782047, + 0.015885327011346817, + 0.018790949136018753, + 0.017435209825634956, + 0.0036498813424259424, + 0.002730749314650893, + 0.00501911761239171, + 0.005442435387521982, + 0.003264788771048188, + 0.004011993296444416, + 0.004075819160789251, + 0.004380705300718546, + 0.0025479572359472513, + 0.002657068893313408, + 0.003065522061660886, + 0.003632187144830823, + 0.0020862948149442673, + 0.0023431505542248487, + 0.002306681592017412, + 0.0026783309876918793, + 0.00746412156149745, + 0.008089481852948666, + 0.008399558253586292, + 0.007489146664738655, + 0.006182740442454815, + 0.00781743973493576, + 0.006538445595651865, + 0.00934660341590643, + 0.0058329738676548, + 0.0044884090311825275, + 0.005780393723398447, + 0.006399817764759064, + 0.002403705846518278, + 0.005005509592592716, + 0.0032262199092656374, + 0.0045290617272257805, + 0.015960421413183212, + 0.023568348959088326, + 0.022151391953229904, + 0.024028535932302475, + 0.012876162305474281, + 0.01964559406042099, + 0.01246440690010786, + 0.019933579489588737, + 0.009054809808731079, + 0.005603970494121313, + 0.007101669907569885, + 0.007928051054477692, + 0.004559643566608429, + 0.004906227812170982, + 0.0037187461275607347, + 0.005576224531978369, + 0.028708763420581818, + 0.04816050082445145, + 0.03739588335156441, + 0.035628288984298706, + 0.018399346619844437, + 0.023888681083917618, + 0.018911277875304222, + 0.022509630769491196, + 0.0062792436219751835, + 0.006852242164313793, + 0.008655041456222534, + 0.009789932519197464, + 0.004328500479459763, + 0.0075902389362454414, + 0.0038202558644115925, + 0.005234444979578257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.048651598393917084 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, extreme close-up, positioned at a slightly elevated angle, remains almost static as it focuses on a polished, antique gyroscope. The precisely machined brass gimbal, radiating a warm golden sheen, catches the light from a single, focused source outside the lens, creating bright, moving highlights on its curved surface. The central flywheel spins at a dizzying pace, becoming a solid, silvery-gray blur that emits a faint, high-frequency hum. As the rotor maintains its rapid speed, the interconnected gimbals begin to precess, rotating and tilting with a slow, deliberate, and hypnotic grace. The entire mechanism glides smoothly across the surface of a dark, rich mahogany table, its smooth surface reflecting the gyroscope's gravity-defying dance in a twisted, shadowed mirror. The interplay of warm, focused light and deep shadows cast on the dark wood creates an atmosphere of focus, contemplation, and almost magical, celebrating the beauty of the moving mechanism.", + "seed": 11462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11462_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019901356380432843, + "vlm_logprob": -0.019901356380432843, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.019901356380432843, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031928230077028275 + ], + "opposite_logprob_tokens": [ + -3.781928300857544 + ], + "logprob_reward": -0.031928230077028275, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.031928230077028275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05087507143616676, + 0.04325072094798088, + 0.008648360148072243, + 0.014092950150370598, + 0.017700975760817528, + 0.5546085238456726, + 0.06500795483589172, + 0.1575814038515091, + 0.029964830726385117, + 0.058269258588552475 + ], + "credit_map": [ + 0.00672485213726759, + 0.005567995831370354, + 0.004109753295779228, + 0.005804827902466059, + 0.00595562718808651, + 0.006728447042405605, + 0.007239483296871185, + 0.008744077757000923, + 0.005422411952167749, + 0.003800880629569292, + 0.004153574351221323, + 0.006622526794672012, + 0.006060383282601833, + 0.005925935227423906, + 0.005920669995248318, + 0.005344333592802286, + 0.0009899402502924204, + 0.001023659366182983, + 0.0009844917804002762, + 0.001090509002096951, + 0.0010394792770966887, + 0.0011798805790022016, + 0.0010155335767194629, + 0.0013248653849586844, + 0.00120260170660913, + 0.0012370769400149584, + 0.0014674313133582473, + 0.0030476744286715984, + 0.0017520004184916615, + 0.0017631981754675508, + 0.0020916347857564688, + 0.0015313304029405117, + 0.0019756562542170286, + 0.0012265032855793834, + 0.0024969179648905993, + 0.002633478492498398, + 0.0022716112434864044, + 0.0021576525177806616, + 0.002203080104663968, + 0.0027360732201486826, + 0.0636039599776268, + 0.04323209077119827, + 0.0901932641863823, + 0.06410010904073715, + 0.043475233018398285, + 0.07578828185796738, + 0.052421897649765015, + 0.12179366499185562, + 0.005242642015218735, + 0.0032942143734544516, + 0.009973661974072456, + 0.007864709012210369, + 0.0057265604846179485, + 0.009099059738218784, + 0.00684711430221796, + 0.016959991306066513, + 0.011227061972022057, + 0.00757346348837018, + 0.02678215317428112, + 0.014964072033762932, + 0.01635458879172802, + 0.02122339978814125, + 0.01713346689939499, + 0.04232318326830864, + 0.0034675400238484144, + 0.002069226698949933, + 0.0051885019056499004, + 0.0027518835850059986, + 0.0032730249222368, + 0.003923266660422087, + 0.0035809874534606934, + 0.00571039691567421, + 0.005292924586683512, + 0.002407911466434598, + 0.01076536439359188, + 0.005362256430089474, + 0.004152282606810331, + 0.008570137433707714, + 0.005906335543841124, + 0.015812037512660027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.031928230077028275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651796538382769 + ], + "opposite_logprob_tokens": [ + -5.37965202331543 + ], + "logprob_reward": -0.004651796538382769, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651796538382769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15320731699466705, + 0.024201558902859688, + 0.013907620683312416, + 0.012806983664631844, + 0.017131466418504715, + 0.39988452196121216, + 0.016986731439828873, + 0.22190797328948975, + 0.10326437652111053, + 0.03670153021812439 + ], + "credit_map": [ + 0.01507782842963934, + 0.013510000891983509, + 0.014054818078875542, + 0.022163255140185356, + 0.01800188422203064, + 0.018387168645858765, + 0.021540265530347824, + 0.03047206625342369, + 0.0033329520374536514, + 0.003153951605781913, + 0.00213346304371953, + 0.0021335086785256863, + 0.003760307328775525, + 0.003579505253583193, + 0.0024913453962653875, + 0.003616523463279009, + 0.001300671836361289, + 0.0010556549532338977, + 0.0013606782304123044, + 0.0012028033379465342, + 0.0014096597442403436, + 0.0021795902866870165, + 0.0014231408713385463, + 0.003975420258939266, + 0.0014970467891544104, + 0.0012534740380942822, + 0.0010886916425079107, + 0.001741680782288313, + 0.001564312377013266, + 0.0017343491781502962, + 0.0020352883730083704, + 0.001892138971015811, + 0.0020178386475890875, + 0.0020964201539754868, + 0.00167610973585397, + 0.002496669301763177, + 0.002285629976540804, + 0.0021143099293112755, + 0.0019399432931095362, + 0.002504543401300907, + 0.042777564376592636, + 0.03524070978164673, + 0.0604194812476635, + 0.05065133050084114, + 0.033423975110054016, + 0.054583411663770676, + 0.04050065204501152, + 0.08228737860918045, + 0.0017078530509024858, + 0.001431054319255054, + 0.0023829012643545866, + 0.0018818447133526206, + 0.0018563501071184874, + 0.002494725864380598, + 0.0020686041098088026, + 0.0031633975449949503, + 0.015706099569797516, + 0.01183327566832304, + 0.03796450048685074, + 0.024004178121685982, + 0.02387787401676178, + 0.02921583689749241, + 0.0244743712246418, + 0.054831817746162415, + 0.011391210369765759, + 0.006166660692542791, + 0.017934825271368027, + 0.009411158040165901, + 0.012080326676368713, + 0.013401884585618973, + 0.012772325426340103, + 0.02010597474873066, + 0.004044403787702322, + 0.004516605753451586, + 0.002821760019287467, + 0.004950047004967928, + 0.004851256962865591, + 0.00493623036891222, + 0.0035037200432270765, + 0.007077502552419901 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.004651796538382769 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255324762314558 + ], + "opposite_logprob_tokens": [ + -4.3875532150268555 + ], + "logprob_reward": -0.01255324762314558, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255324762314558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053207434713840485, + 0.02191021852195263, + 0.013428283855319023, + 0.012334348633885384, + 0.010485414415597916, + 0.6305055618286133, + 0.03618301451206207, + 0.07437508553266525, + 0.12326973676681519, + 0.02430095337331295 + ], + "credit_map": [ + 0.00790367927402258, + 0.004569397307932377, + 0.00781768187880516, + 0.00625789538025856, + 0.0058300611563026905, + 0.006307247094810009, + 0.007160124834626913, + 0.007361340802162886, + 0.0034940794575959444, + 0.00200763251632452, + 0.0023093342315405607, + 0.002768846694380045, + 0.0028764214366674423, + 0.0028884573839604855, + 0.003009034553542733, + 0.002556410152465105, + 0.0014634407125413418, + 0.0012989111710339785, + 0.001447024056687951, + 0.0013275069650262594, + 0.0012228351552039385, + 0.0019912938587367535, + 0.0017537664389237761, + 0.0029235039837658405, + 0.0011622827732935548, + 0.001077420311048627, + 0.00118306465446949, + 0.002074114978313446, + 0.001740907202474773, + 0.001671253819949925, + 0.002079413505271077, + 0.0013458899920806289, + 0.0009987751254811883, + 0.0009366951999254525, + 0.0010960237123072147, + 0.0017796113388612866, + 0.0011657408904284239, + 0.0013053376460447907, + 0.0015853376826271415, + 0.0016178926452994347, + 0.07104479521512985, + 0.051671676337718964, + 0.09916567802429199, + 0.07439499348402023, + 0.05077705159783363, + 0.08574394136667252, + 0.06094011291861534, + 0.13676726818084717, + 0.0029650856740772724, + 0.0016051577404141426, + 0.006400699261575937, + 0.004135369788855314, + 0.002762936055660248, + 0.004916504956781864, + 0.004049800802022219, + 0.009347458370029926, + 0.005183244124054909, + 0.003592939581722021, + 0.011719371192157269, + 0.00987950898706913, + 0.007387577090412378, + 0.010020148940384388, + 0.007917637936770916, + 0.01867465116083622, + 0.01237070094794035, + 0.006973578594624996, + 0.020761540159583092, + 0.01035979762673378, + 0.014036068692803383, + 0.016555655747652054, + 0.016275770962238312, + 0.025936610996723175, + 0.0021497304551303387, + 0.002215823158621788, + 0.003291246248409152, + 0.0027586761862039566, + 0.0022275906521826982, + 0.0032627552282065153, + 0.0027558300644159317, + 0.00563929695636034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.01255324762314558 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020575417205691338 + ], + "opposite_logprob_tokens": [ + -3.895575523376465 + ], + "logprob_reward": -0.020575417205691338, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020575417205691338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07779870927333832, + 0.062094394117593765, + 0.02218032255768776, + 0.044377002865076065, + 0.025388892740011215, + 0.2974768280982971, + 0.10644401609897614, + 0.11646032333374023, + 0.20465023815631866, + 0.04312921687960625 + ], + "credit_map": [ + 0.008800315670669079, + 0.010158251039683819, + 0.00878423172980547, + 0.008953055366873741, + 0.01018533855676651, + 0.00990184023976326, + 0.009179441258311272, + 0.011836244724690914, + 0.007441217079758644, + 0.00611414248123765, + 0.005282184574753046, + 0.008152268826961517, + 0.008374509401619434, + 0.008392480202019215, + 0.010618869215250015, + 0.007718724198639393, + 0.0025147954002022743, + 0.002349280519410968, + 0.001842997153289616, + 0.0027991216629743576, + 0.002540521789342165, + 0.0028753201477229595, + 0.0033007177989929914, + 0.003957568667829037, + 0.004416576586663723, + 0.003351641818881035, + 0.004633326083421707, + 0.011474044993519783, + 0.003919295035302639, + 0.005475348327308893, + 0.005879332311451435, + 0.00522744283080101, + 0.002645624801516533, + 0.0025086987297981977, + 0.002609358401969075, + 0.003518575569614768, + 0.0029061215464025736, + 0.0036593875847756863, + 0.0037488264497369528, + 0.00379230291582644, + 0.034250807017087936, + 0.0259077288210392, + 0.04415363073348999, + 0.03486543893814087, + 0.02552209608256817, + 0.039893738925457, + 0.0318378210067749, + 0.06104559078812599, + 0.008614328689873219, + 0.005260696168988943, + 0.02191196382045746, + 0.010221542790532112, + 0.007768006529659033, + 0.014361992478370667, + 0.014056303538382053, + 0.024249199777841568, + 0.008588903583586216, + 0.007266137283295393, + 0.01785719394683838, + 0.013511361554265022, + 0.012253348715603352, + 0.015652919188141823, + 0.012605254538357258, + 0.028725219890475273, + 0.020663758739829063, + 0.013049270026385784, + 0.033730845898389816, + 0.017523951828479767, + 0.025558244436979294, + 0.02752160094678402, + 0.02604709565639496, + 0.04055548831820488, + 0.004554421175271273, + 0.004962490871548653, + 0.005445280112326145, + 0.004809709265828133, + 0.004207658115774393, + 0.005505204666405916, + 0.004680064972490072, + 0.008964392356574535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.020575417205691338 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979809045791626 + ], + "opposite_logprob_tokens": [ + -3.5297980308532715 + ], + "logprob_reward": -0.02979809045791626, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979809045791626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04587942734360695, + 0.04550713300704956, + 0.017872650176286697, + 0.014764616265892982, + 0.03145084157586098, + 0.43068450689315796, + 0.026710791513323784, + 0.06874038279056549, + 0.17150472104549408, + 0.14688493311405182 + ], + "credit_map": [ + 0.005015929229557514, + 0.0055193910375237465, + 0.004954117815941572, + 0.006088351830840111, + 0.005454327445477247, + 0.0059540714137256145, + 0.006576011888682842, + 0.006317222956568003, + 0.007659010589122772, + 0.003665821859613061, + 0.004676485899835825, + 0.006008507218211889, + 0.004035744816064835, + 0.006295669358223677, + 0.007089246064424515, + 0.006076646503061056, + 0.0028718530666083097, + 0.0011649058433249593, + 0.0022645550779998302, + 0.001823339844122529, + 0.0014460913371294737, + 0.0024086316116154194, + 0.0019685092847794294, + 0.003924763761460781, + 0.001425204798579216, + 0.0010822953190654516, + 0.0019015761790797114, + 0.0017952010966837406, + 0.001579787116497755, + 0.002050595125183463, + 0.0015672175213694572, + 0.003362740855664015, + 0.0030260863713920116, + 0.0027993267867714167, + 0.0038549774326384068, + 0.003709285520017147, + 0.003614590037614107, + 0.0047947014681994915, + 0.0037308072205632925, + 0.005921064410358667, + 0.04923767223954201, + 0.035991422832012177, + 0.0666564553976059, + 0.05068552494049072, + 0.0349225252866745, + 0.05897269397974014, + 0.04184022173285484, + 0.09237799793481827, + 0.0024094104301184416, + 0.0013384644407778978, + 0.005936409346759319, + 0.002657206729054451, + 0.0020285649225115776, + 0.0036417015362530947, + 0.0028762314468622208, + 0.005822801496833563, + 0.004556300118565559, + 0.0032911149319261312, + 0.01278525311499834, + 0.006671388167887926, + 0.0070900581777095795, + 0.009405727498233318, + 0.008418750017881393, + 0.016521792858839035, + 0.016102254390716553, + 0.0087234852835536, + 0.02949303388595581, + 0.014009818434715271, + 0.016274934634566307, + 0.023285456001758575, + 0.02376909740269184, + 0.03984663262963295, + 0.015179050154983997, + 0.00799992959946394, + 0.029714226722717285, + 0.014235815964639187, + 0.009956853464245796, + 0.02031126618385315, + 0.01536710374057293, + 0.03412069007754326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 9 + ], + "selected_reward": -0.02979809045791626 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight, extreme close-up, positioned at a slightly elevated angle, remains almost static as it focuses on a polished, antique gyroscope. The precisely machined brass gimbal, radiating a warm golden sheen, catches the light from a single, focused source outside the lens, creating bright, moving highlights on its curved surface. The central flywheel spins at a dizzying pace, becoming a solid, silvery-gray blur that emits a faint, high-frequency hum. As the rotor maintains its rapid speed, the interconnected gimbals begin to precess, rotating and tilting with a slow, deliberate, and hypnotic grace. The entire mechanism glides smoothly across the surface of a dark, rich mahogany table, its smooth surface reflecting the gyroscope's gravity-defying dance in a twisted, shadowed mirror. The interplay of warm, focused light and deep shadows cast on the dark wood creates an atmosphere of focus, contemplation, and almost magical, celebrating the beauty of the moving mechanism.", + "seed": 11463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_3/index_639_seed_11463_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03417866034433246, + "vlm_logprob": -0.03417866034433246, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.03417866034433246, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02820025570690632 + ], + "opposite_logprob_tokens": [ + -3.903200149536133 + ], + "logprob_reward": -0.02820025570690632, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02820025570690632, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20664598047733307, + 0.17775942385196686, + 0.028179394081234932, + 0.08829156309366226, + 0.09017618745565414, + 0.04461489990353584, + 0.1593865156173706, + 0.08908422291278839, + 0.09588543325662613, + 0.019976304844021797 + ], + "credit_map": [ + 0.032734207808971405, + 0.01868417114019394, + 0.05773620679974556, + 0.01808297261595726, + 0.009244213812053204, + 0.036915358155965805, + 0.014289864338934422, + 0.018958985805511475, + 0.026692068204283714, + 0.014712662436068058, + 0.0488768070936203, + 0.013624397106468678, + 0.0062081534415483475, + 0.03553343564271927, + 0.01013711653649807, + 0.021974781528115273, + 0.002916170284152031, + 0.003455563448369503, + 0.004262383095920086, + 0.004440036602318287, + 0.0022866292856633663, + 0.0022527030669152737, + 0.005890470929443836, + 0.0026754364371299744, + 0.016386039555072784, + 0.007906644605100155, + 0.02297123521566391, + 0.011392898857593536, + 0.005663891788572073, + 0.010873132385313511, + 0.0047750771045684814, + 0.00832263845950365, + 0.014391069300472736, + 0.008244723081588745, + 0.026334915310144424, + 0.00894206203520298, + 0.005166040733456612, + 0.013453339226543903, + 0.0075639947317540646, + 0.0060800425708293915, + 0.007698748726397753, + 0.004732172936201096, + 0.010523542761802673, + 0.0041585383005440235, + 0.0030845843721181154, + 0.005272728856652975, + 0.0035680162254720926, + 0.005576568190008402, + 0.029301365837454796, + 0.013297018595039845, + 0.05550443008542061, + 0.012487559579312801, + 0.003899059258401394, + 0.02830001711845398, + 0.004656539764255285, + 0.011940525844693184, + 0.011681374162435532, + 0.007941343821585178, + 0.022587470710277557, + 0.013313363306224346, + 0.004993514623492956, + 0.013888002373278141, + 0.007429733872413635, + 0.007249421440064907, + 0.014344990253448486, + 0.007952231913805008, + 0.02959500066936016, + 0.008176016621291637, + 0.0037267289590090513, + 0.018683699890971184, + 0.003625306999310851, + 0.009781453758478165, + 0.0020975566003471613, + 0.0020356569439172745, + 0.0032063948456197977, + 0.003856305032968521, + 0.001629662699997425, + 0.002506379736587405, + 0.0028038749005645514, + 0.0018404739676043391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.02820025570690632 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005266483407467604 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.005266483407467604, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005266483407467604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2501194179058075, + 0.2564335763454437, + 0.03854804113507271, + 0.049000367522239685, + 0.027720190584659576, + 0.06080268695950508, + 0.2277728021144867, + 0.047944143414497375, + 0.025817610323429108, + 0.015841158106923103 + ], + "credit_map": [ + 0.040092937648296356, + 0.019950339570641518, + 0.07245218753814697, + 0.02780282497406006, + 0.008237002417445183, + 0.04800266772508621, + 0.010050447657704353, + 0.02353104203939438, + 0.04103672131896019, + 0.02312508411705494, + 0.07167765498161316, + 0.016998661682009697, + 0.0075604673475027084, + 0.04792122542858124, + 0.01238563284277916, + 0.03572814166545868, + 0.0032899274956434965, + 0.00557180680334568, + 0.006478957366198301, + 0.007737008389085531, + 0.0025744822341948748, + 0.00242487620562315, + 0.005987348034977913, + 0.004483634606003761, + 0.00937060359865427, + 0.003620234550908208, + 0.012989308685064316, + 0.006080924533307552, + 0.002996806986629963, + 0.005718765314668417, + 0.002951183822005987, + 0.005272542592138052, + 0.004629520233720541, + 0.0020671251695603132, + 0.007806748151779175, + 0.0033477353863418102, + 0.0013120593503117561, + 0.003963584080338478, + 0.0015316037461161613, + 0.0030618158634752035, + 0.008744834922254086, + 0.006382103078067303, + 0.009747209027409554, + 0.009832127951085567, + 0.005099014844745398, + 0.005369365215301514, + 0.007854153402149677, + 0.007773882243782282, + 0.04055481031537056, + 0.01861143298447132, + 0.07726115733385086, + 0.01850673370063305, + 0.006476713344454765, + 0.03863246366381645, + 0.010536370798945427, + 0.01719312183558941, + 0.0059289182536304, + 0.0040542250499129295, + 0.012465653009712696, + 0.006552516482770443, + 0.002033126074820757, + 0.009065342135727406, + 0.002725128782913089, + 0.005119235720485449, + 0.0032343484926968813, + 0.0025212473701685667, + 0.006195190362632275, + 0.0023104120045900345, + 0.0015261679654940963, + 0.005112423095852137, + 0.0022257836535573006, + 0.0026920398231595755, + 0.0018765188287943602, + 0.0015559118473902345, + 0.002862368943169713, + 0.0032753238920122385, + 0.0013350637163966894, + 0.0010466119274497032, + 0.0019997640047222376, + 0.0018895954126492143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.005266483407467604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371492773294449 + ], + "opposite_logprob_tokens": [ + -3.408715009689331 + ], + "logprob_reward": -0.03371492773294449, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371492773294449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1723846197128296, + 0.1808657944202423, + 0.05586244910955429, + 0.029406243935227394, + 0.062328681349754333, + 0.02446727640926838, + 0.09040766209363937, + 0.04961758479475975, + 0.31373271346092224, + 0.020926928147673607 + ], + "credit_map": [ + 0.025853822007775307, + 0.012024270370602608, + 0.046588409692049026, + 0.0238263551145792, + 0.007589149754494429, + 0.03261002525687218, + 0.008012106642127037, + 0.015880485996603966, + 0.024025335907936096, + 0.01948421448469162, + 0.03585468605160713, + 0.015820790082216263, + 0.009475935250520706, + 0.02310088276863098, + 0.01770309917628765, + 0.03540084883570671, + 0.005057464353740215, + 0.007650918327271938, + 0.006386252120137215, + 0.009940498508512974, + 0.005714535713195801, + 0.004498234484344721, + 0.012219494208693504, + 0.004395049996674061, + 0.004185439087450504, + 0.0033593056723475456, + 0.0054642679169774055, + 0.004195935558527708, + 0.0019677546806633472, + 0.003684219904243946, + 0.002012829761952162, + 0.004536494147032499, + 0.008848316967487335, + 0.005546518601477146, + 0.017407305538654327, + 0.0077013554982841015, + 0.0029290099628269672, + 0.0097238440066576, + 0.005296988878399134, + 0.004875339567661285, + 0.0032361492048949003, + 0.002901887521147728, + 0.0037950545083731413, + 0.00333131430670619, + 0.0025967254769057035, + 0.002416134113445878, + 0.0026896668132394552, + 0.0035003460943698883, + 0.015236882492899895, + 0.007183687295764685, + 0.02892075479030609, + 0.0068998089991509914, + 0.002505728742107749, + 0.017344750463962555, + 0.003882036777213216, + 0.00843400415033102, + 0.007278040051460266, + 0.004646346904337406, + 0.01381194218993187, + 0.004657319746911526, + 0.0024674043525010347, + 0.008795107714831829, + 0.0032588657923042774, + 0.004702556878328323, + 0.048100218176841736, + 0.025196196511387825, + 0.09858973324298859, + 0.02608242817223072, + 0.009472673758864403, + 0.06667543947696686, + 0.008424959145486355, + 0.031191037967801094, + 0.002790854312479496, + 0.0017402594676241279, + 0.004560565110296011, + 0.0035582014825195074, + 0.0014457664219662547, + 0.0023815459571778774, + 0.002467802492901683, + 0.0019819331355392933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.03371492773294449 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369337320327759 + ], + "opposite_logprob_tokens": [ + -3.408693313598633 + ], + "logprob_reward": -0.03369337320327759, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369337320327759, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12920275330543518, + 0.1964343786239624, + 0.0812339335680008, + 0.04000946506857872, + 0.03686798736453056, + 0.07655539363622665, + 0.03172275796532631, + 0.02295740880072117, + 0.37110045552253723, + 0.013915465213358402 + ], + "credit_map": [ + 0.015614300034940243, + 0.01436146255582571, + 0.027372801676392555, + 0.019142525270581245, + 0.010883187875151634, + 0.01842241920530796, + 0.012514217756688595, + 0.01089184358716011, + 0.02954022027552128, + 0.021305926144123077, + 0.04938635230064392, + 0.015257027931511402, + 0.00874793715775013, + 0.03252545744180679, + 0.012048858217895031, + 0.02762259915471077, + 0.007283948827534914, + 0.010289350524544716, + 0.009567088447511196, + 0.013210916891694069, + 0.008412701077759266, + 0.007566310930997133, + 0.01817910373210907, + 0.0067245084792375565, + 0.007264764979481697, + 0.003523551858961582, + 0.010386195965111256, + 0.005592730827629566, + 0.00213425699621439, + 0.004654638934880495, + 0.0022929480765014887, + 0.004160379990935326, + 0.004775036126375198, + 0.0032391261775046587, + 0.009075314737856388, + 0.004847714677453041, + 0.002458398463204503, + 0.004379593767225742, + 0.005282985512167215, + 0.0028098165057599545, + 0.013018074445426464, + 0.008933118544518948, + 0.015140601433813572, + 0.007822077721357346, + 0.005726632662117481, + 0.009342535398900509, + 0.005612834822386503, + 0.010959520936012268, + 0.004193732514977455, + 0.003188887843862176, + 0.007181958761066198, + 0.0047919126227498055, + 0.0015622047940269113, + 0.0045347148552536964, + 0.0027842966374009848, + 0.003485048422589898, + 0.002867833012714982, + 0.0022654656786471605, + 0.004645670764148235, + 0.004353244788944721, + 0.0012476332485675812, + 0.002771221799775958, + 0.0014693925622850657, + 0.0033369455486536026, + 0.05798877775669098, + 0.029947923496365547, + 0.11842525750398636, + 0.030329488217830658, + 0.011434347368776798, + 0.07740567624568939, + 0.009452838450670242, + 0.036116138100624084, + 0.0013188765151426196, + 0.001376278349198401, + 0.0020938704255968332, + 0.002386413048952818, + 0.0013005342334508896, + 0.001598017755895853, + 0.0024721601512283087, + 0.0013693147338926792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.03369337320327759 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight, extreme close-up of a polished antique gyroscope?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera positioned at a slightly elevated angle and remains almost static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central flywheel spinning rapidly enough to appear as a solid silvery-gray blur?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the interconnected gimbals precess, rotating and tilting with a slow, deliberate motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the gyroscope glide across a dark, rich mahogany table that reflects the mechanism?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001826167106628 + ], + "opposite_logprob_tokens": [ + -2.6950182914733887 + ], + "logprob_reward": -0.07001826167106628, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001826167106628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06947014480829239, + 0.0929974839091301, + 0.06434329599142075, + 0.04702412709593773, + 0.1133633628487587, + 0.06222636252641678, + 0.02357468195259571, + 0.028720779344439507, + 0.4792862832546234, + 0.018993400037288666 + ], + "credit_map": [ + 0.009637458249926567, + 0.005721001885831356, + 0.016071943566203117, + 0.010339196771383286, + 0.0028547486290335655, + 0.013171473518013954, + 0.0036525309551507235, + 0.008021801710128784, + 0.012463339604437351, + 0.011597743257880211, + 0.018827883526682854, + 0.007934926077723503, + 0.004698803648352623, + 0.011262071318924427, + 0.008391820825636387, + 0.01782090589404106, + 0.008885371498763561, + 0.006491320673376322, + 0.017614109441637993, + 0.0065288362093269825, + 0.002165565500035882, + 0.011177265085279942, + 0.004119324032217264, + 0.007361510768532753, + 0.007382161915302277, + 0.0036102556623518467, + 0.01154446043074131, + 0.009459885768592358, + 0.0023166814353317022, + 0.0046598645858466625, + 0.0027119270525872707, + 0.005338891409337521, + 0.020997384563088417, + 0.008218568749725819, + 0.03667785972356796, + 0.009687324054539204, + 0.003201077925041318, + 0.020602064207196236, + 0.0045726485550403595, + 0.009406444616615772, + 0.01158824097365141, + 0.005639674142003059, + 0.01406348031014204, + 0.007202669512480497, + 0.0033538758289068937, + 0.007516156882047653, + 0.0037875077687203884, + 0.009074763394892216, + 0.0027306985575705767, + 0.003305447055026889, + 0.0035353631246834993, + 0.004263908602297306, + 0.002196061424911022, + 0.0013377346331253648, + 0.004254774656146765, + 0.001950696692802012, + 0.0034236235078424215, + 0.003097026376053691, + 0.0055161938071250916, + 0.0045934636145830154, + 0.0015773397171869874, + 0.0038858952466398478, + 0.002941121580079198, + 0.0036861172411590815, + 0.07328515499830246, + 0.03798845037817955, + 0.1500360369682312, + 0.03841456398367882, + 0.013746198266744614, + 0.1052529364824295, + 0.011108582839369774, + 0.049454428255558014, + 0.0034700476098805666, + 0.0011296741431578994, + 0.0054628849029541016, + 0.002017741557210684, + 0.0007170456228777766, + 0.0035853025037795305, + 0.0009005741449072957, + 0.0017101304838433862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 1 + ], + "selected_reward": -0.07001826167106628 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_tensor([9468, 9468])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_tensor([9468, 9468])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..df6d8dea3d756ae241fb7ae3b3b865bf33281c7c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_tensor([9468, 9468])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "From a slightly high-angle extreme close-up, the camera remains static, focused on the patient's tightly closed eyelids as they begin to subtly tremble and flutter. As the camera begins a slow, almost imperceptible zoom out to a medium shot, the heavy lids slowly part, revealing hazy, unfocused irises. We see a pale, damp complexion, messy brown hair scattered across a flat white hospital pillow, dressed in a standard-issue pale blue gown. Their heads turn deliberately, feebly, as their disoriented gaze slowly scans the room. The scan begins at the sterile white, acoustic tile ceiling and drifts across the room. Bright, direct sunlight streams in through a large window, casting sharp geometric shapes on the polished linoleum floor and illuminating dust motes dancing in the light. The gaze continues hesitantly, gazing at the metallic sheen of an IV stand dripping with a saline bag, the steady, rhythmic green lines of the silent bedside heart monitor, and finally, the closed, featureless door on the far wall. The entire scene is immersed in a cool, clinical palette of whites and greys, radically interrupted by warm, intrusive sunlight, creating a palpable atmosphere of quiet, sterile disorientation and fragile awakening.", + "seed": 2094, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_9468_seed_2094_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.5395025297999382, + "vlm_logprob": -0.5395025297999382, + "vlm_yes_no_margin": 1.35, + "vlm_token_credit": -0.5395025297999382, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16398978233337402 + ], + "opposite_logprob_tokens": [ + -1.913989782333374 + ], + "logprob_reward": -0.16398978233337402, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16398978233337402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14756788313388824, + 0.07991284132003784, + 0.17314141988754272, + 0.20261305570602417, + 0.061130326241254807, + 0.17979423701763153, + 0.053206365555524826, + 0.03403286635875702, + 0.04374856501817703, + 0.0248524509370327 + ], + "credit_map": [ + 0.030204029753804207, + 0.01376558467745781, + 0.02000771462917328, + 0.013772339560091496, + 0.025265511125326157, + 0.014304803684353828, + 0.014337957836687565, + 0.015909947454929352, + 0.008131027221679688, + 0.007555144373327494, + 0.012705045752227306, + 0.015791969373822212, + 0.0071086459793150425, + 0.010841283947229385, + 0.010913796722888947, + 0.006865927018225193, + 0.028483454138040543, + 0.020552387461066246, + 0.022135071456432343, + 0.022197429090738297, + 0.025655897334218025, + 0.0187099426984787, + 0.018052449449896812, + 0.017354795709252357, + 0.00940600037574768, + 0.013850778341293335, + 0.01434826198965311, + 0.021983670070767403, + 0.057048648595809937, + 0.05326027050614357, + 0.01848572865128517, + 0.014229709282517433, + 0.004546151030808687, + 0.009401602670550346, + 0.00592472217977047, + 0.007878951728343964, + 0.007252105046063662, + 0.007414558902382851, + 0.00794648751616478, + 0.01076575182378292, + 0.021725976839661598, + 0.031666167080402374, + 0.011742278933525085, + 0.023312287405133247, + 0.026815209537744522, + 0.02484464831650257, + 0.023239608854055405, + 0.01644805260002613, + 0.004908924922347069, + 0.011413941159844398, + 0.003747291164472699, + 0.00654961820691824, + 0.006792297121137381, + 0.006696207914501429, + 0.007736148312687874, + 0.005361939314752817, + 0.0038837522733956575, + 0.005294440314173698, + 0.002838700544089079, + 0.004214166197925806, + 0.004086288157850504, + 0.00445198081433773, + 0.0047393240965902805, + 0.0045242151245474815, + 0.005472743883728981, + 0.004285474773496389, + 0.002757775830104947, + 0.004052997101098299, + 0.007266180589795113, + 0.009529110044240952, + 0.00352649949491024, + 0.006857783533632755, + 0.003244420513510704, + 0.002874490339308977, + 0.0019576826598495245, + 0.003386954078450799, + 0.003412992227822542, + 0.0031885995995253325, + 0.0039380136877298355, + 0.002849297598004341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.16398978233337402 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07906566560268402 + ], + "opposite_logprob_tokens": [ + -2.5790655612945557 + ], + "logprob_reward": -0.07906566560268402, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07906566560268402, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1285848766565323, + 0.07153995335102081, + 0.050714150071144104, + 0.055656034499406815, + 0.14037282764911652, + 0.15331050753593445, + 0.1793859750032425, + 0.10152263939380646, + 0.06159881874918938, + 0.05731423944234848 + ], + "credit_map": [ + 0.01989630050957203, + 0.013875485397875309, + 0.021347666159272194, + 0.011826776899397373, + 0.018446698784828186, + 0.01607072353363037, + 0.011590595357120037, + 0.015530625358223915, + 0.008373551070690155, + 0.009012238122522831, + 0.009987146593630314, + 0.012182911857962608, + 0.005935762543231249, + 0.011336104944348335, + 0.008283477276563644, + 0.006428761873394251, + 0.005321775563061237, + 0.006122663151472807, + 0.00858323834836483, + 0.008090865798294544, + 0.0044579291716217995, + 0.004619190469384193, + 0.007749729789793491, + 0.0057687521912157536, + 0.004429709631949663, + 0.00894199963659048, + 0.006221175659447908, + 0.007018015254288912, + 0.00477603217586875, + 0.008503682911396027, + 0.007523356471210718, + 0.008242065086960793, + 0.012382925488054752, + 0.010386320762336254, + 0.00808683317154646, + 0.008753220550715923, + 0.021444270387291908, + 0.023757822811603546, + 0.018882086500525475, + 0.03667934238910675, + 0.011866710148751736, + 0.027416255325078964, + 0.013427789323031902, + 0.026527443900704384, + 0.017208639532327652, + 0.01663607358932495, + 0.026120981201529503, + 0.014106612652540207, + 0.02326042205095291, + 0.020161697641015053, + 0.00861357618123293, + 0.013343483209609985, + 0.03220018371939659, + 0.0324324332177639, + 0.01550270151346922, + 0.03387150168418884, + 0.007342176046222448, + 0.014438704587519169, + 0.01095722895115614, + 0.013186348602175713, + 0.010645180940628052, + 0.012741785496473312, + 0.013765629380941391, + 0.018445579335093498, + 0.010144174098968506, + 0.006642286665737629, + 0.00455894973129034, + 0.007287993095815182, + 0.009102336131036282, + 0.008399154990911484, + 0.007166402414441109, + 0.008297520689666271, + 0.00500013493001461, + 0.004096697084605694, + 0.004313746467232704, + 0.004980612080544233, + 0.007700435817241669, + 0.005261429585516453, + 0.014953396283090115, + 0.01100778765976429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.07906566560268402 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594336092472076 + ], + "opposite_logprob_tokens": [ + -3.895594358444214 + ], + "logprob_reward": -0.020594336092472076, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594336092472076, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1203533411026001, + 0.07085596024990082, + 0.062058400362730026, + 0.1235198900103569, + 0.11610385775566101, + 0.1641974151134491, + 0.05977100133895874, + 0.1242494061589241, + 0.11582072824239731, + 0.04307004436850548 + ], + "credit_map": [ + 0.01912972331047058, + 0.011963771656155586, + 0.01509890053421259, + 0.013200554996728897, + 0.019486991688609123, + 0.01565972901880741, + 0.010815344750881195, + 0.01499833445996046, + 0.006913748104125261, + 0.007774590980261564, + 0.011020566336810589, + 0.011488436721265316, + 0.007089449558407068, + 0.007112324703484774, + 0.009864676743745804, + 0.009592168964445591, + 0.004879429936408997, + 0.006864307913929224, + 0.00900583527982235, + 0.01016553957015276, + 0.007211599033325911, + 0.006345283705741167, + 0.008385918103158474, + 0.009200484491884708, + 0.008716488257050514, + 0.01196468248963356, + 0.01393204927444458, + 0.016972744837403297, + 0.019593285396695137, + 0.026268526911735535, + 0.013201241381466389, + 0.01287087518721819, + 0.012138387188315392, + 0.01695464365184307, + 0.009915247559547424, + 0.013774394989013672, + 0.012668638490140438, + 0.017413103953003883, + 0.012836967594921589, + 0.020402468740940094, + 0.01200849749147892, + 0.027903232723474503, + 0.018467912450432777, + 0.026611868292093277, + 0.020653212442994118, + 0.02547120861709118, + 0.014870038256049156, + 0.01821144111454487, + 0.005803821142762899, + 0.007510293275117874, + 0.0040892865508794785, + 0.005924512166529894, + 0.009784378111362457, + 0.011048387736082077, + 0.005624787881970406, + 0.00998553168028593, + 0.013928026892244816, + 0.017774885520339012, + 0.012166962958872318, + 0.018014565110206604, + 0.015098677016794682, + 0.01636376604437828, + 0.01606220193207264, + 0.014840319752693176, + 0.011161583475768566, + 0.013037987984716892, + 0.007069685496389866, + 0.008855908177793026, + 0.018978463485836983, + 0.026928290724754333, + 0.01283156406134367, + 0.016957245767116547, + 0.005348168779164553, + 0.005757675506174564, + 0.002510741585865617, + 0.0034804586321115494, + 0.0056519159115850925, + 0.005917728878557682, + 0.00803021527826786, + 0.0063731432892382145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 3 + ], + "selected_reward": -0.020594336092472076 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.910383939743042 + ], + "opposite_logprob_tokens": [ + -0.160383939743042 + ], + "logprob_reward": -1.910383939743042, + "yes_no_margin_reward": -1.75, + "token_credit_reward": -1.910383939743042, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12033024430274963, + 0.03463231772184372, + 0.05857748165726662, + 0.18431541323661804, + 0.38249820470809937, + 0.09955821186304092, + 0.018469009548425674, + 0.04976533725857735, + 0.020748738199472427, + 0.031105052679777145 + ], + "credit_map": [ + 0.01503059733659029, + 0.013620106503367424, + 0.017657116055488586, + 0.012072495184838772, + 0.009391814470291138, + 0.019416002556681633, + 0.013216352090239525, + 0.019925745204091072, + 0.0035102565307170153, + 0.003667099867016077, + 0.0063193622045218945, + 0.006076100282371044, + 0.0029439376667141914, + 0.0037083954084664583, + 0.004573920276015997, + 0.003833241295069456, + 0.00694201048463583, + 0.007068573031574488, + 0.009230113588273525, + 0.009441709145903587, + 0.005559764802455902, + 0.005439159460365772, + 0.00888905581086874, + 0.006007086019963026, + 0.01149404514580965, + 0.01614224724471569, + 0.02107209339737892, + 0.028003927320241928, + 0.038710322231054306, + 0.029815759509801865, + 0.02250121347606182, + 0.01657579280436039, + 0.029793528839945793, + 0.027098579332232475, + 0.022194452583789825, + 0.027390042319893837, + 0.06307125836610794, + 0.0777198076248169, + 0.045544449239969254, + 0.08968605846166611, + 0.008237006142735481, + 0.01119838748127222, + 0.008906614035367966, + 0.012081664055585861, + 0.01330120861530304, + 0.018221314996480942, + 0.010788996703922749, + 0.01682300679385662, + 0.0018826194573193789, + 0.0018327247817069292, + 0.0012546690413728356, + 0.002014458877965808, + 0.002673603594303131, + 0.0032607470639050007, + 0.0019169410225003958, + 0.0036332430317997932, + 0.0040995036251842976, + 0.007477113511413336, + 0.004311962518841028, + 0.005958698224276304, + 0.006572858430445194, + 0.010583625175058842, + 0.004722903948277235, + 0.006038661580532789, + 0.002452667336910963, + 0.00245228735730052, + 0.0014094917569309473, + 0.0019035723526030779, + 0.0033598309382796288, + 0.004022559151053429, + 0.0023456162307411432, + 0.0028027116786688566, + 0.0036685059312731028, + 0.0022969769779592752, + 0.0020664851181209087, + 0.003221507417038083, + 0.0044683534651994705, + 0.0030199752654880285, + 0.008245945908129215, + 0.004117298871278763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -1.910383939743042 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5234789252281189 + ], + "opposite_logprob_tokens": [ + -0.8984789252281189 + ], + "logprob_reward": -0.5234789252281189, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5234789252281189, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06432915478944778, + 0.05646928399801254, + 0.0882510393857956, + 0.14274844527244568, + 0.08918122202157974, + 0.09529148042201996, + 0.04301975667476654, + 0.1112353652715683, + 0.20438729226589203, + 0.10508689284324646 + ], + "credit_map": [ + 0.007743110880255699, + 0.006685533560812473, + 0.00903746671974659, + 0.007846909575164318, + 0.006130377296358347, + 0.009482534602284431, + 0.0065376730635762215, + 0.010865549556910992, + 0.0069177341647446156, + 0.007470672018826008, + 0.007110479287803173, + 0.008374705910682678, + 0.004969438537955284, + 0.008055660873651505, + 0.007059850730001926, + 0.006510742008686066, + 0.009459977969527245, + 0.009546741843223572, + 0.014758216217160225, + 0.01589064486324787, + 0.007402343209832907, + 0.008743392303586006, + 0.01437386590987444, + 0.008075853809714317, + 0.010763351805508137, + 0.013616467826068401, + 0.017448097467422485, + 0.020786965265870094, + 0.01248976681381464, + 0.027804644778370857, + 0.016650313511490822, + 0.023188846185803413, + 0.008406182751059532, + 0.007537162397056818, + 0.006183989811688662, + 0.0077000875025987625, + 0.013531739823520184, + 0.016057057306170464, + 0.011089122854173183, + 0.01867588236927986, + 0.00713800685480237, + 0.01627410016953945, + 0.009240343235433102, + 0.016631312668323517, + 0.012301573529839516, + 0.013677646405994892, + 0.00981669407337904, + 0.010211802087724209, + 0.00445916410535574, + 0.005516417324542999, + 0.00260104238986969, + 0.004457000643014908, + 0.005471291486173868, + 0.006440968718379736, + 0.00474020978435874, + 0.009333662688732147, + 0.012407227419316769, + 0.02012251317501068, + 0.008943983353674412, + 0.012819469906389713, + 0.015000551007688046, + 0.018826229497790337, + 0.011612546630203724, + 0.011502847075462341, + 0.0234591793268919, + 0.031030887737870216, + 0.014043084345757961, + 0.0220816470682621, + 0.0293694157153368, + 0.03372344747185707, + 0.0200826283544302, + 0.030597016215324402, + 0.019554996863007545, + 0.018894284963607788, + 0.007085427176207304, + 0.007796050515025854, + 0.01272573508322239, + 0.012783801183104515, + 0.011621939018368721, + 0.014624647796154022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.5234789252281189 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle extreme close-up, the camera remains static, focused on the patient's tightly closed eyelids as they begin to subtly tremble and flutter. As the camera begins a slow, almost imperceptible zoom out to a medium shot, the heavy lids slowly part, revealing hazy, unfocused irises. We see a pale, damp complexion, messy brown hair scattered across a flat white hospital pillow, dressed in a standard-issue pale blue gown. Their heads turn deliberately, feebly, as their disoriented gaze slowly scans the room. The scan begins at the sterile white, acoustic tile ceiling and drifts across the room. Bright, direct sunlight streams in through a large window, casting sharp geometric shapes on the polished linoleum floor and illuminating dust motes dancing in the light. The gaze continues hesitantly, gazing at the metallic sheen of an IV stand dripping with a saline bag, the steady, rhythmic green lines of the silent bedside heart monitor, and finally, the closed, featureless door on the far wall. The entire scene is immersed in a cool, clinical palette of whites and greys, radically interrupted by warm, intrusive sunlight, creating a palpable atmosphere of quiet, sterile disorientation and fragile awakening.", + "seed": 2095, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_9468_seed_2095_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.3319782756268978, + "vlm_logprob": -0.3319782756268978, + "vlm_yes_no_margin": 1.775, + "vlm_token_credit": -0.3319782756268978, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9793680906295776 + ], + "opposite_logprob_tokens": [ + -0.47936809062957764 + ], + "logprob_reward": -0.9793680906295776, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9793680906295776, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19045813381671906, + 0.12292997539043427, + 0.07977815717458725, + 0.17011511325836182, + 0.1596240997314453, + 0.10082820802927017, + 0.030989602208137512, + 0.03027864173054695, + 0.0335659384727478, + 0.08143218606710434 + ], + "credit_map": [ + 0.020694036036729813, + 0.030757978558540344, + 0.01986796036362648, + 0.02860182709991932, + 0.021302444860339165, + 0.020246831700205803, + 0.020724330097436905, + 0.028262723237276077, + 0.007913191802799702, + 0.01346916425973177, + 0.006978132762014866, + 0.012596151791512966, + 0.03637756034731865, + 0.019392317160964012, + 0.011778426356613636, + 0.014425025321543217, + 0.010293602012097836, + 0.011969435960054398, + 0.006051903124898672, + 0.012835255824029446, + 0.008551934733986855, + 0.008760562166571617, + 0.010457386262714863, + 0.010858075693249702, + 0.03036355786025524, + 0.0556403212249279, + 0.006747919600456953, + 0.018200207501649857, + 0.01699870266020298, + 0.02143561653792858, + 0.010213506408035755, + 0.010515277273952961, + 0.03712596371769905, + 0.028793232515454292, + 0.010450045578181744, + 0.017706042155623436, + 0.01565881073474884, + 0.01434912160038948, + 0.020923640578985214, + 0.014617233537137508, + 0.006195486057549715, + 0.008750343695282936, + 0.005102226044982672, + 0.008501659147441387, + 0.025471949949860573, + 0.030338454991579056, + 0.00852620042860508, + 0.00794188678264618, + 0.003405619179829955, + 0.005561720114201307, + 0.002590000629425049, + 0.003999373875558376, + 0.0036821693647652864, + 0.004869857802987099, + 0.003787044668570161, + 0.0030938161071389914, + 0.003233821364119649, + 0.003941433969885111, + 0.002027049195021391, + 0.004051370080560446, + 0.004594149999320507, + 0.0037230493035167456, + 0.005734240170568228, + 0.0029735281132161617, + 0.002870580879971385, + 0.005099013913422823, + 0.002120731398463249, + 0.0018521102610975504, + 0.005079091526567936, + 0.012398874387145042, + 0.0019226170843467116, + 0.0022229182068258524, + 0.0069457548670470715, + 0.007624070160090923, + 0.007787039037793875, + 0.01074427179992199, + 0.010037967003881931, + 0.01187892071902752, + 0.009239370003342628, + 0.017174789682030678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.9793680906295776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020624464377760887 + ], + "opposite_logprob_tokens": [ + -3.8956243991851807 + ], + "logprob_reward": -0.020624464377760887, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020624464377760887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028954001143574715, + 0.10015310347080231, + 0.04024679958820343, + 0.06470213085412979, + 0.08371661603450775, + 0.018569394946098328, + 0.13203510642051697, + 0.31752172112464905, + 0.07281938195228577, + 0.14128176867961884 + ], + "credit_map": [ + 0.0036190724931657314, + 0.0026008530985563993, + 0.002986003179103136, + 0.003558110212907195, + 0.004196477122604847, + 0.0028499194886535406, + 0.0038860137574374676, + 0.0052575524896383286, + 0.009940740652382374, + 0.01785796321928501, + 0.011845353059470654, + 0.011816932819783688, + 0.010353517718613148, + 0.010469495318830013, + 0.012974999845027924, + 0.014894102700054646, + 0.004568855743855238, + 0.003667159704491496, + 0.004425329156219959, + 0.005316146649420261, + 0.004512340296059847, + 0.0043082935735583305, + 0.004919460043311119, + 0.008529214188456535, + 0.0110607435926795, + 0.01821349561214447, + 0.0027597902808338404, + 0.0077895596623420715, + 0.006905996240675449, + 0.007593259215354919, + 0.004898129962384701, + 0.005481157451868057, + 0.01760394684970379, + 0.01330812368541956, + 0.0050199334509670734, + 0.007920296862721443, + 0.012674549594521523, + 0.009242325089871883, + 0.010824792087078094, + 0.0071226502768695354, + 0.0026650652289390564, + 0.0021164421923458576, + 0.0014343240763992071, + 0.00204788357950747, + 0.0019642191473394632, + 0.0017878860235214233, + 0.003483964828774333, + 0.0030696112662553787, + 0.005408973898738623, + 0.00940882507711649, + 0.004906912799924612, + 0.005103766452521086, + 0.0409080944955349, + 0.04913158714771271, + 0.009011860936880112, + 0.008155085146427155, + 0.04937579482793808, + 0.047898098826408386, + 0.01962689496576786, + 0.02999860793352127, + 0.03625652194023132, + 0.06879396736621857, + 0.03701150044798851, + 0.02856033481657505, + 0.009627164341509342, + 0.008279477246105671, + 0.006164575461298227, + 0.006539316847920418, + 0.009134100750088692, + 0.016706526279449463, + 0.00792451947927475, + 0.008443698287010193, + 0.01120656356215477, + 0.016125526279211044, + 0.009792832657694817, + 0.014310427010059357, + 0.017468374222517014, + 0.02330070175230503, + 0.01336774043738842, + 0.035709600895643234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.020624464377760887 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981266938149929 + ], + "opposite_logprob_tokens": [ + -3.5298125743865967 + ], + "logprob_reward": -0.02981266938149929, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981266938149929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06173844635486603, + 0.05512909218668938, + 0.04146480932831764, + 0.08471208810806274, + 0.08257186412811279, + 0.055546559393405914, + 0.033453844487667084, + 0.11939791589975357, + 0.07188922166824341, + 0.39409616589546204 + ], + "credit_map": [ + 0.0056788246147334576, + 0.008357523940503597, + 0.006214343011379242, + 0.009174074977636337, + 0.009520807303488255, + 0.006036388222128153, + 0.007772472687065601, + 0.008984018117189407, + 0.005021837539970875, + 0.007292273920029402, + 0.005327439401298761, + 0.004986881278455257, + 0.011360212229192257, + 0.006703152786940336, + 0.006634339224547148, + 0.007802962325513363, + 0.003737453371286392, + 0.004811490420252085, + 0.003164699301123619, + 0.00805029459297657, + 0.00379467592574656, + 0.0044464292004704475, + 0.006438996642827988, + 0.0070207747630774975, + 0.006319515872746706, + 0.020499946549534798, + 0.005038953386247158, + 0.011970004066824913, + 0.014239059761166573, + 0.012124854139983654, + 0.008488127961754799, + 0.006031641270965338, + 0.01273699663579464, + 0.012007088400423527, + 0.0037650205194950104, + 0.010063918307423592, + 0.01431422121822834, + 0.01056892890483141, + 0.011390081606805325, + 0.007725613657385111, + 0.00626684445887804, + 0.0075179217383265495, + 0.004605303518474102, + 0.007111221551895142, + 0.005665507633239031, + 0.00892269890755415, + 0.0075925616547465324, + 0.007864503189921379, + 0.0031756472308188677, + 0.006281252484768629, + 0.002574484096840024, + 0.0035252326633781195, + 0.004062797874212265, + 0.004915030673146248, + 0.004818785935640335, + 0.004100616555660963, + 0.018412502482533455, + 0.023284489288926125, + 0.009076004847884178, + 0.012479027733206749, + 0.011802343651652336, + 0.022709136828780174, + 0.013553934171795845, + 0.008080490864813328, + 0.010054528713226318, + 0.007368866819888353, + 0.00873028114438057, + 0.008402234874665737, + 0.007538546342402697, + 0.01038314774632454, + 0.008851105347275734, + 0.010560511611402035, + 0.03209277614951134, + 0.04032418876886368, + 0.03210778906941414, + 0.03060515597462654, + 0.050421055406332016, + 0.06340658664703369, + 0.04494059458374977, + 0.10019803792238235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 3 + ], + "selected_reward": -0.02981266938149929 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34873664379119873 + ], + "opposite_logprob_tokens": [ + -1.2237366437911987 + ], + "logprob_reward": -0.34873664379119873, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34873664379119873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07012595236301422, + 0.025599511340260506, + 0.04596415162086487, + 0.05374031513929367, + 0.10199642181396484, + 0.0970953106880188, + 0.02107904851436615, + 0.04186573624610901, + 0.13218404352664948, + 0.4103494882583618 + ], + "credit_map": [ + 0.006130007095634937, + 0.006687177810817957, + 0.006621949840337038, + 0.011550725437700748, + 0.010329545475542545, + 0.0068273283541202545, + 0.00986021850258112, + 0.012119006365537643, + 0.0023146949242800474, + 0.0024037419352680445, + 0.0021609978284686804, + 0.0026903452817350626, + 0.005757996812462807, + 0.0035015062894672155, + 0.003136802464723587, + 0.003633425571024418, + 0.005065356846898794, + 0.005005860701203346, + 0.0030748487915843725, + 0.0077475267462432384, + 0.005788333714008331, + 0.005517196375876665, + 0.0062804436311125755, + 0.0074845836497843266, + 0.009705227799713612, + 0.012106956914067268, + 0.0029792790301144123, + 0.005171510856598616, + 0.007136894389986992, + 0.006009627599269152, + 0.006200066301971674, + 0.004430751781910658, + 0.01126208808273077, + 0.01483179535716772, + 0.005471613258123398, + 0.014867990277707577, + 0.015829037874937057, + 0.01069303136318922, + 0.016076963394880295, + 0.012963896617293358, + 0.0068505387753248215, + 0.008003361523151398, + 0.0033896497916430235, + 0.004191366955637932, + 0.029228854924440384, + 0.03357531875371933, + 0.005669556092470884, + 0.00618665711954236, + 0.0018499282887205482, + 0.0020558040123432875, + 0.0009882401209324598, + 0.0018302283715456724, + 0.004698406904935837, + 0.00578191177919507, + 0.0016838116571307182, + 0.002190718427300453, + 0.009376008994877338, + 0.006955007091164589, + 0.002983538433909416, + 0.004094794392585754, + 0.00401095487177372, + 0.0054932557977736, + 0.005467296577990055, + 0.003484878921881318, + 0.019794082269072533, + 0.01863318495452404, + 0.012536264955997467, + 0.016309604048728943, + 0.014527025632560253, + 0.026451481506228447, + 0.011753205209970474, + 0.012179183773696423, + 0.031003141775727272, + 0.04673802852630615, + 0.03219277784228325, + 0.048339053988456726, + 0.06560097634792328, + 0.061616599559783936, + 0.036794472485780716, + 0.08806443214416504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 4 + ], + "selected_reward": -0.34873664379119873 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813495099544525 + ], + "opposite_logprob_tokens": [ + -1.406349539756775 + ], + "logprob_reward": -0.2813495099544525, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813495099544525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.047754354774951935, + 0.04091738909482956, + 0.0315643846988678, + 0.28055334091186523, + 0.13939696550369263, + 0.07914882898330688, + 0.03425155207514763, + 0.042328253388404846, + 0.08694801479578018, + 0.2171369194984436 + ], + "credit_map": [ + 0.006817150395363569, + 0.0061751240864396095, + 0.0048825787380337715, + 0.0054808869026601315, + 0.0061965626664459705, + 0.005018554162234068, + 0.004973985720425844, + 0.008209514431655407, + 0.003990992903709412, + 0.005827167071402073, + 0.0032231800723820925, + 0.004388997796922922, + 0.008873366750776768, + 0.005505237728357315, + 0.004135667812079191, + 0.00497277919203043, + 0.0033827756997197866, + 0.0039840699173510075, + 0.002214022446423769, + 0.004119582939893007, + 0.004152336623519659, + 0.0038311416283249855, + 0.0044176350347697735, + 0.005462824832648039, + 0.04721013829112053, + 0.08944690227508545, + 0.010702426545321941, + 0.03263721987605095, + 0.024555088952183723, + 0.034105099737644196, + 0.020840460434556007, + 0.021056002005934715, + 0.03271004930138588, + 0.023053564131259918, + 0.008757240138947964, + 0.012837064452469349, + 0.018100261688232422, + 0.013817157596349716, + 0.018769852817058563, + 0.011351760476827621, + 0.005175672005861998, + 0.004818027373403311, + 0.003115729195997119, + 0.0049959635362029076, + 0.024028029292821884, + 0.028392244130373, + 0.005122516769915819, + 0.0035006441175937653, + 0.0025180699303746223, + 0.004205379169434309, + 0.002182576572522521, + 0.003316788701340556, + 0.004494093358516693, + 0.006248906254768372, + 0.00533750606700778, + 0.005948231555521488, + 0.006847976706922054, + 0.006323907990008593, + 0.0025870550889521837, + 0.004137193318456411, + 0.0041736336424946785, + 0.008469210006296635, + 0.005163137800991535, + 0.004626135341823101, + 0.017024772241711617, + 0.010276352055370808, + 0.008135879412293434, + 0.013708915561437607, + 0.009772607125341892, + 0.010783218778669834, + 0.009268909692764282, + 0.007977367378771305, + 0.01854798197746277, + 0.02478104457259178, + 0.017895463854074478, + 0.023417433723807335, + 0.029557520523667336, + 0.027955427765846252, + 0.02043142542243004, + 0.054550617933273315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.2813495099544525 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle extreme close-up, the camera remains static, focused on the patient's tightly closed eyelids as they begin to subtly tremble and flutter. As the camera begins a slow, almost imperceptible zoom out to a medium shot, the heavy lids slowly part, revealing hazy, unfocused irises. We see a pale, damp complexion, messy brown hair scattered across a flat white hospital pillow, dressed in a standard-issue pale blue gown. Their heads turn deliberately, feebly, as their disoriented gaze slowly scans the room. The scan begins at the sterile white, acoustic tile ceiling and drifts across the room. Bright, direct sunlight streams in through a large window, casting sharp geometric shapes on the polished linoleum floor and illuminating dust motes dancing in the light. The gaze continues hesitantly, gazing at the metallic sheen of an IV stand dripping with a saline bag, the steady, rhythmic green lines of the silent bedside heart monitor, and finally, the closed, featureless door on the far wall. The entire scene is immersed in a cool, clinical palette of whites and greys, radically interrupted by warm, intrusive sunlight, creating a palpable atmosphere of quiet, sterile disorientation and fragile awakening.", + "seed": 2096, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_9468_seed_2096_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.4133098041638732, + "vlm_logprob": -0.4133098041638732, + "vlm_yes_no_margin": 2.125, + "vlm_token_credit": -0.4133098041638732, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0456106960773468 + ], + "opposite_logprob_tokens": [ + -3.1706106662750244 + ], + "logprob_reward": -0.0456106960773468, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0456106960773468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30162763595581055, + 0.12022095173597336, + 0.09090904891490936, + 0.10044416785240173, + 0.12989401817321777, + 0.04100705310702324, + 0.06786797195672989, + 0.05024134740233421, + 0.04474335536360741, + 0.05304444208741188 + ], + "credit_map": [ + 0.03077244758605957, + 0.01628100872039795, + 0.03694106638431549, + 0.06659352034330368, + 0.0335998609662056, + 0.037390269339084625, + 0.05185956507921219, + 0.028189899399876595, + 0.008158448152244091, + 0.01547965407371521, + 0.008688494563102722, + 0.024122200906276703, + 0.020069997757673264, + 0.012302550487220287, + 0.01690739020705223, + 0.014492210000753403, + 0.00726827047765255, + 0.007247609551995993, + 0.013102344237267971, + 0.01987621560692787, + 0.010035351850092411, + 0.01002370472997427, + 0.012566211633384228, + 0.010789348743855953, + 0.011026907712221146, + 0.014231599867343903, + 0.00992523692548275, + 0.014621066860854626, + 0.015280458144843578, + 0.011845064349472523, + 0.013713938184082508, + 0.009799894876778126, + 0.00880805216729641, + 0.007996338419616222, + 0.0110147250816226, + 0.0036468966864049435, + 0.024989904835820198, + 0.049655813723802567, + 0.007876063697040081, + 0.01590622402727604, + 0.005356599576771259, + 0.003177260048687458, + 0.0061970981769263744, + 0.004978842567652464, + 0.006416549906134605, + 0.0044503905810415745, + 0.005859503522515297, + 0.004570808727294207, + 0.01039367076009512, + 0.007719780784100294, + 0.008968849666416645, + 0.008854025043547153, + 0.007321298588067293, + 0.007500927895307541, + 0.010602089576423168, + 0.006507336627691984, + 0.004838151857256889, + 0.00754178874194622, + 0.003756539663299918, + 0.004541154019534588, + 0.011382890865206718, + 0.006491614505648613, + 0.007530952338129282, + 0.004158252850174904, + 0.0037989539559930563, + 0.006804774980992079, + 0.005213282536715269, + 0.006883557885885239, + 0.007375248707830906, + 0.006338202394545078, + 0.0037148166447877884, + 0.004614518955349922, + 0.00425190944224596, + 0.009142125025391579, + 0.009796532802283764, + 0.004373827017843723, + 0.00807648804038763, + 0.009461539797484875, + 0.0032979778479784727, + 0.0046440414153039455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0456106960773468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06210307031869888 + ], + "opposite_logprob_tokens": [ + -2.812103033065796 + ], + "logprob_reward": -0.06210307031869888, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06210307031869888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17278563976287842, + 0.20904606580734253, + 0.031114403158426285, + 0.12603653967380524, + 0.0397392176091671, + 0.07321552187204361, + 0.08296258747577667, + 0.06061065196990967, + 0.16573576629161835, + 0.03875355049967766 + ], + "credit_map": [ + 0.009675673209130764, + 0.009836027398705482, + 0.013878190889954567, + 0.05226227268576622, + 0.010645829141139984, + 0.028457850217819214, + 0.03325997292995453, + 0.01476982794702053, + 0.018323294818401337, + 0.014611081220209599, + 0.025803659111261368, + 0.07019868493080139, + 0.02105768956243992, + 0.014178765937685966, + 0.03072318807244301, + 0.014149706810712814, + 0.0041540623642504215, + 0.0027622096240520477, + 0.003688467899337411, + 0.005200643558055162, + 0.0032062537502497435, + 0.0035158211831003428, + 0.003806019201874733, + 0.004780928138643503, + 0.01059004571288824, + 0.012896434403955936, + 0.014707773923873901, + 0.02460230141878128, + 0.013464647345244884, + 0.012070485390722752, + 0.026877736672759056, + 0.010827130638062954, + 0.0037601469084620476, + 0.0035596396774053574, + 0.002371628303080797, + 0.0025045303627848625, + 0.006717886310070753, + 0.01012379303574562, + 0.006117237266153097, + 0.004584358539432287, + 0.007312431465834379, + 0.007341389078646898, + 0.007560316938906908, + 0.012130902148783207, + 0.007776985876262188, + 0.007679220754653215, + 0.0173158161342144, + 0.006098456215113401, + 0.014066374860703945, + 0.009276742115616798, + 0.012402383610606194, + 0.008740718476474285, + 0.010400930419564247, + 0.00921307411044836, + 0.00926714576780796, + 0.009595220908522606, + 0.007227327208966017, + 0.012988298200070858, + 0.005862488877028227, + 0.005711560137569904, + 0.007806057576090097, + 0.007659847382456064, + 0.006528745871037245, + 0.006826332770287991, + 0.013592864386737347, + 0.01907191053032875, + 0.02800299972295761, + 0.01374947838485241, + 0.02217533066868782, + 0.022840453311800957, + 0.022969909012317657, + 0.023332839831709862, + 0.004088539630174637, + 0.004107383545488119, + 0.006546656601130962, + 0.004029369447380304, + 0.0041819168254733086, + 0.006439909804612398, + 0.003175038145855069, + 0.006184738595038652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.06210307031869888 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01108629535883665 + ], + "opposite_logprob_tokens": [ + -4.511086463928223 + ], + "logprob_reward": -0.01108629535883665, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01108629535883665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20153824985027313, + 0.13867367804050446, + 0.07352983951568604, + 0.15048207342624664, + 0.062492094933986664, + 0.08436018228530884, + 0.11137928068637848, + 0.03963862359523773, + 0.06500225514173508, + 0.07290375232696533 + ], + "credit_map": [ + 0.02333468571305275, + 0.017771905288100243, + 0.025517769157886505, + 0.025098668411374092, + 0.03351972624659538, + 0.02096140757203102, + 0.027617814019322395, + 0.027716269716620445, + 0.01468417327851057, + 0.011822208762168884, + 0.015986347571015358, + 0.038542501628398895, + 0.015077942050993443, + 0.009285029955208302, + 0.022750554606318474, + 0.010524926707148552, + 0.004587583709508181, + 0.007694256957620382, + 0.008622284047305584, + 0.011810929514467716, + 0.01472457591444254, + 0.01066640205681324, + 0.008369949646294117, + 0.007053859531879425, + 0.011491172946989536, + 0.01261076983064413, + 0.01794835738837719, + 0.03216812387108803, + 0.017936524003744125, + 0.015159707516431808, + 0.03163391724228859, + 0.01153351180255413, + 0.008769369684159756, + 0.007795743644237518, + 0.0052900961600244045, + 0.004157154355198145, + 0.013269667513668537, + 0.008249294944107533, + 0.007404617965221405, + 0.007556145545095205, + 0.0162409208714962, + 0.01201153639703989, + 0.010163343511521816, + 0.0076911235228180885, + 0.016609573736786842, + 0.006804038304835558, + 0.007262369152158499, + 0.007577280513942242, + 0.011058551259338856, + 0.015269264578819275, + 0.01267462782561779, + 0.010865872725844383, + 0.018297363072633743, + 0.011163766495883465, + 0.01692863367497921, + 0.015121200121939182, + 0.0036469215992838144, + 0.006376395467668772, + 0.0029755535069853067, + 0.002998530864715576, + 0.008655277080833912, + 0.0066091506741940975, + 0.0046870531514286995, + 0.003689741250127554, + 0.005872893147170544, + 0.009963784366846085, + 0.006518792361021042, + 0.005675151944160461, + 0.015144256874918938, + 0.00880188774317503, + 0.00574472127482295, + 0.007280766032636166, + 0.005257331766188145, + 0.011504273861646652, + 0.011560969986021519, + 0.010207007639110088, + 0.01101350225508213, + 0.012272678315639496, + 0.005277207121253014, + 0.005810779985040426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01108629535883665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048774003982544 + ], + "opposite_logprob_tokens": [ + -0.1798774003982544 + ], + "logprob_reward": -1.8048774003982544, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048774003982544, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3084869980812073, + 0.02362630143761635, + 0.06129872798919678, + 0.028464248403906822, + 0.09260440617799759, + 0.06414008885622025, + 0.065528504550457, + 0.08515141904354095, + 0.21033789217472076, + 0.06036144495010376 + ], + "credit_map": [ + 0.025487538427114487, + 0.026438290253281593, + 0.02927093207836151, + 0.04498908296227455, + 0.029256021603941917, + 0.04262640327215195, + 0.0477643646299839, + 0.06265433877706528, + 0.002936669159680605, + 0.002216603606939316, + 0.0025030651595443487, + 0.00529277790337801, + 0.0030292270239442587, + 0.0021329692099243402, + 0.0033188359811902046, + 0.0021961533930152655, + 0.004636948462575674, + 0.005610854830592871, + 0.005919299554079771, + 0.009227894246578217, + 0.013184313662350178, + 0.009078712202608585, + 0.008074581623077393, + 0.005566121079027653, + 0.0026300172321498394, + 0.0034519964829087257, + 0.00338168628513813, + 0.0061083291657269, + 0.003278373507782817, + 0.002851077588275075, + 0.0039677610620856285, + 0.00279500731267035, + 0.009246284142136574, + 0.008620897307991982, + 0.006878893356770277, + 0.00561951007694006, + 0.019156575202941895, + 0.02139369212090969, + 0.013700934126973152, + 0.007987616583704948, + 0.008054893463850021, + 0.006989223416894674, + 0.009280766360461712, + 0.009194249287247658, + 0.0076310038566589355, + 0.006068968679755926, + 0.009792928583920002, + 0.007128056138753891, + 0.00948946550488472, + 0.006315488833934069, + 0.007907883264124393, + 0.004679182544350624, + 0.011436056345701218, + 0.007806362118571997, + 0.00794987939298153, + 0.009944187477231026, + 0.010390778072178364, + 0.01563373953104019, + 0.006650922354310751, + 0.005948417354375124, + 0.02012300305068493, + 0.01223284937441349, + 0.008966506458818913, + 0.005205206573009491, + 0.020798930898308754, + 0.046934135258197784, + 0.01627972163259983, + 0.015461636707186699, + 0.049894534051418304, + 0.0334608256816864, + 0.013875690288841724, + 0.013632412999868393, + 0.0080241858959198, + 0.009379127062857151, + 0.009563097730278969, + 0.008135694079101086, + 0.006932361051440239, + 0.009447867050766945, + 0.004254389554262161, + 0.00462472066283226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -1.8048774003982544 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287155866622925 + ], + "opposite_logprob_tokens": [ + -2.017871618270874 + ], + "logprob_reward": -0.14287155866622925, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14287155866622925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19887635111808777, + 0.05405730754137039, + 0.030044391751289368, + 0.050906792283058167, + 0.13502779603004456, + 0.07729241997003555, + 0.04933829605579376, + 0.1452202945947647, + 0.152430459856987, + 0.10680592060089111 + ], + "credit_map": [ + 0.014785862527787685, + 0.01360207237303257, + 0.019305519759655, + 0.03434361144900322, + 0.021452074870467186, + 0.027817754074931145, + 0.0316486619412899, + 0.03592078387737274, + 0.004274181555956602, + 0.004373869393020868, + 0.005049475934356451, + 0.012279101647436619, + 0.007651256397366524, + 0.004878703970462084, + 0.009227975271642208, + 0.006322743836790323, + 0.004048395901918411, + 0.0028097592294216156, + 0.003969634883105755, + 0.00522204302251339, + 0.004408030770719051, + 0.0033269410487264395, + 0.0037182543892413378, + 0.0025413320399820805, + 0.00864248163998127, + 0.010549641214311123, + 0.003394781844690442, + 0.004584456793963909, + 0.005803112406283617, + 0.006314355880022049, + 0.005956192035228014, + 0.005661772098392248, + 0.008264241740107536, + 0.005450429394841194, + 0.010646452195942402, + 0.004335063509643078, + 0.025997942313551903, + 0.05464721471071243, + 0.0084434375166893, + 0.017243007197976112, + 0.008304416202008724, + 0.008054128848016262, + 0.010519322007894516, + 0.012940065003931522, + 0.008724416606128216, + 0.007898887619376183, + 0.013479229062795639, + 0.007371949497610331, + 0.007153704762458801, + 0.004996204283088446, + 0.006154556293040514, + 0.004151342902332544, + 0.011477578431367874, + 0.006573283113539219, + 0.00546974316239357, + 0.0033618840388953686, + 0.014743922278285027, + 0.029369954019784927, + 0.0098362872377038, + 0.008647083304822445, + 0.03453117609024048, + 0.02293316088616848, + 0.014294679276645184, + 0.01086402591317892, + 0.012819130904972553, + 0.02492506057024002, + 0.017538171261548996, + 0.010080871172249317, + 0.03741777315735817, + 0.023591410368680954, + 0.011405029334127903, + 0.014653009362518787, + 0.011558320373296738, + 0.013886072672903538, + 0.022477244958281517, + 0.012338976375758648, + 0.010508166626095772, + 0.017036359757184982, + 0.007749438751488924, + 0.011251340620219707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.14287155866622925 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "From a slightly high-angle extreme close-up, the camera remains static, focused on the patient's tightly closed eyelids as they begin to subtly tremble and flutter. As the camera begins a slow, almost imperceptible zoom out to a medium shot, the heavy lids slowly part, revealing hazy, unfocused irises. We see a pale, damp complexion, messy brown hair scattered across a flat white hospital pillow, dressed in a standard-issue pale blue gown. Their heads turn deliberately, feebly, as their disoriented gaze slowly scans the room. The scan begins at the sterile white, acoustic tile ceiling and drifts across the room. Bright, direct sunlight streams in through a large window, casting sharp geometric shapes on the polished linoleum floor and illuminating dust motes dancing in the light. The gaze continues hesitantly, gazing at the metallic sheen of an IV stand dripping with a saline bag, the steady, rhythmic green lines of the silent bedside heart monitor, and finally, the closed, featureless door on the far wall. The entire scene is immersed in a cool, clinical palette of whites and greys, radically interrupted by warm, intrusive sunlight, creating a palpable atmosphere of quiet, sterile disorientation and fragile awakening.", + "seed": 2097, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_4/index_9468_seed_2097_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19530561566352844, + "vlm_logprob": -0.19530561566352844, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.19530561566352844, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3153410255908966 + ], + "opposite_logprob_tokens": [ + -1.3153409957885742 + ], + "logprob_reward": -0.3153410255908966, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3153410255908966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13112038373947144, + 0.12846219539642334, + 0.15823888778686523, + 0.05021747574210167, + 0.03984621912240982, + 0.3468759059906006, + 0.06574980914592743, + 0.03457142412662506, + 0.0321778878569603, + 0.012739777565002441 + ], + "credit_map": [ + 0.029455628246068954, + 0.021307798102498055, + 0.017079420387744904, + 0.013275173492729664, + 0.011098973453044891, + 0.013783741742372513, + 0.009598677977919579, + 0.015520983375608921, + 0.020764745771884918, + 0.020509900525212288, + 0.013393665663897991, + 0.015133596956729889, + 0.02589402161538601, + 0.011547602713108063, + 0.01013305876404047, + 0.011085611768066883, + 0.009570867754518986, + 0.01902163028717041, + 0.010599819011986256, + 0.012554110959172249, + 0.03544280678033829, + 0.04409381374716759, + 0.00924775842577219, + 0.01770809106528759, + 0.006658261641860008, + 0.003997249994426966, + 0.0029455148614943027, + 0.006103424821048975, + 0.009794695302844048, + 0.007917850278317928, + 0.006551940459758043, + 0.006248543504625559, + 0.004536440595984459, + 0.004639724735170603, + 0.004557029344141483, + 0.0073877014219760895, + 0.005141595844179392, + 0.0036734442692250013, + 0.0043575456365942955, + 0.005552736576646566, + 0.03333977237343788, + 0.03519468009471893, + 0.04719127342104912, + 0.05633051320910454, + 0.028625041246414185, + 0.04981137812137604, + 0.03539213538169861, + 0.06099114194512367, + 0.004103782121092081, + 0.009393618442118168, + 0.011110587976872921, + 0.00959528237581253, + 0.005862036254256964, + 0.008701181970536709, + 0.006579529959708452, + 0.01040379423648119, + 0.0030713218729943037, + 0.0028680688701570034, + 0.0037529340479522943, + 0.0031270887702703476, + 0.00291212135925889, + 0.008348481729626656, + 0.002796175191178918, + 0.007695235311985016, + 0.004140649922192097, + 0.004391171038150787, + 0.0052249436266720295, + 0.002285789232701063, + 0.0032456754706799984, + 0.005229818634688854, + 0.0021306094713509083, + 0.005529235117137432, + 0.001743600587360561, + 0.001903364434838295, + 0.0025084346998482943, + 0.0017178625566884875, + 0.0015051410300657153, + 0.0008888868032954633, + 0.0008768339175730944, + 0.0015956542920321226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.3153410255908966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03817349299788475 + ], + "opposite_logprob_tokens": [ + -3.2881734371185303 + ], + "logprob_reward": -0.03817349299788475, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03817349299788475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08633219450712204, + 0.04870661720633507, + 0.12033233046531677, + 0.047759752720594406, + 0.05640826001763344, + 0.11384356021881104, + 0.0893159955739975, + 0.12371518462896347, + 0.22827789187431335, + 0.0853082463145256 + ], + "credit_map": [ + 0.014379115775227547, + 0.007948759011924267, + 0.009049028158187866, + 0.008766080252826214, + 0.012522811070084572, + 0.009188070893287659, + 0.008514215238392353, + 0.01596410572528839, + 0.007590085733681917, + 0.005534558091312647, + 0.004013483878225088, + 0.006110925227403641, + 0.008746638894081116, + 0.004105666186660528, + 0.004620363935828209, + 0.007984895259141922, + 0.00653142249211669, + 0.013238740153610706, + 0.00759025989100337, + 0.010111498646438122, + 0.025324247777462006, + 0.0322735421359539, + 0.009622042998671532, + 0.015640558674931526, + 0.005166666116565466, + 0.005371080245822668, + 0.004729564767330885, + 0.005556574556976557, + 0.0053138742223382, + 0.0056366692297160625, + 0.010515989735722542, + 0.005469333380460739, + 0.004800709895789623, + 0.005957362707704306, + 0.005196890793740749, + 0.008697155863046646, + 0.008952394127845764, + 0.005847620777785778, + 0.007228535134345293, + 0.00972758885473013, + 0.01060583908110857, + 0.012876317836344242, + 0.014274869114160538, + 0.021938476711511612, + 0.013010100461542606, + 0.018585611134767532, + 0.010922078974545002, + 0.011630258522927761, + 0.00893988274037838, + 0.01048077829182148, + 0.012179938144981861, + 0.01225066278129816, + 0.007306109182536602, + 0.012111872434616089, + 0.017832055687904358, + 0.00821468885987997, + 0.012592710554599762, + 0.008246172219514847, + 0.013710622675716877, + 0.017815785482525826, + 0.009945617988705635, + 0.02452975884079933, + 0.010806996375322342, + 0.026067513972520828, + 0.042053040117025375, + 0.029524609446525574, + 0.041360948234796524, + 0.025923870503902435, + 0.018732687458395958, + 0.028706427663564682, + 0.014055409468710423, + 0.027920883148908615, + 0.012368024326860905, + 0.015158521942794323, + 0.016211576759815216, + 0.013225553557276726, + 0.008155865594744682, + 0.0058465623296797276, + 0.005305846221745014, + 0.00903629045933485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 2 + ], + "selected_reward": -0.03817349299788475 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867078736424446 + ], + "opposite_logprob_tokens": [ + -3.048670768737793 + ], + "logprob_reward": -0.04867078736424446, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867078736424446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05026470869779587, + 0.07161274552345276, + 0.0548071451485157, + 0.04848231002688408, + 0.03684481233358383, + 0.3566504418849945, + 0.12936022877693176, + 0.08804197609424591, + 0.08875516802072525, + 0.07518047839403152 + ], + "credit_map": [ + 0.007900402881205082, + 0.007728497963398695, + 0.005106070078909397, + 0.0064117214642465115, + 0.005606639198958874, + 0.004946418572217226, + 0.004544943105429411, + 0.008020015433430672, + 0.006685967091470957, + 0.007198076229542494, + 0.004797830246388912, + 0.006482106167823076, + 0.021200140938162804, + 0.008271832019090652, + 0.0053683496080338955, + 0.011608440428972244, + 0.005383613519370556, + 0.011946339160203934, + 0.004310883115977049, + 0.007665169425308704, + 0.008106718771159649, + 0.006240163929760456, + 0.0062104458920657635, + 0.004943813197314739, + 0.007497060112655163, + 0.006454539485275745, + 0.005850706249475479, + 0.0070322295650839806, + 0.005025654099881649, + 0.004250049125403166, + 0.006038938648998737, + 0.006333134602755308, + 0.0039978427812457085, + 0.0035724148619920015, + 0.003679863875731826, + 0.006223604548722506, + 0.0046009947545826435, + 0.004982257727533579, + 0.0048768483102321625, + 0.00491098640486598, + 0.03698136284947395, + 0.027881495654582977, + 0.05302969738841057, + 0.04676678404211998, + 0.027226703241467476, + 0.05743768438696861, + 0.028638530522584915, + 0.07868818938732147, + 0.00898001343011856, + 0.014300203882157803, + 0.020178554579615593, + 0.017568223178386688, + 0.015176638029515743, + 0.018201790750026703, + 0.021405620500445366, + 0.013549181632697582, + 0.0070198397152125835, + 0.0066828359849750996, + 0.009210148826241493, + 0.007209057919681072, + 0.008009485900402069, + 0.024635717272758484, + 0.0072240023873746395, + 0.018050890415906906, + 0.012871944345533848, + 0.011410224251449108, + 0.015140735544264317, + 0.00879445020109415, + 0.007500133942812681, + 0.009731774218380451, + 0.005096506327390671, + 0.018209397792816162, + 0.008020024746656418, + 0.009539734572172165, + 0.013737142086029053, + 0.01047481782734394, + 0.007437754888087511, + 0.010416869074106216, + 0.005578499753028154, + 0.009975636377930641 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.04867078736424446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34872007369995117 + ], + "opposite_logprob_tokens": [ + -1.2237200736999512 + ], + "logprob_reward": -0.34872007369995117, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34872007369995117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05182259902358055, + 0.08174898475408554, + 0.12864600121974945, + 0.058208219707012177, + 0.09754812717437744, + 0.17703759670257568, + 0.15415099263191223, + 0.048045020550489426, + 0.02530389092862606, + 0.17748863995075226 + ], + "credit_map": [ + 0.013184573501348495, + 0.008301992900669575, + 0.006095645483583212, + 0.005327804945409298, + 0.00541442958638072, + 0.004141617100685835, + 0.0034380124416202307, + 0.00591852143406868, + 0.010310227051377296, + 0.010071779601275921, + 0.007009983062744141, + 0.007995531894266605, + 0.018010105937719345, + 0.008000006899237633, + 0.006466694641858339, + 0.013884653337299824, + 0.006460949312895536, + 0.0072795976884663105, + 0.006705996114760637, + 0.008283911272883415, + 0.028075458481907845, + 0.042221929877996445, + 0.0061132460832595825, + 0.023504894226789474, + 0.006897535640746355, + 0.0069056181237101555, + 0.005346393678337336, + 0.0060707940720021725, + 0.010242356918752193, + 0.008013118989765644, + 0.007194188889116049, + 0.007538206409662962, + 0.005147302057594061, + 0.007765431888401508, + 0.005270394030958414, + 0.008407892659306526, + 0.0190139040350914, + 0.01754133403301239, + 0.01060509029775858, + 0.023796770721673965, + 0.01846214197576046, + 0.016219958662986755, + 0.02439146488904953, + 0.02060677483677864, + 0.01611248031258583, + 0.028564494103193283, + 0.01581037975847721, + 0.036869876086711884, + 0.010176338255405426, + 0.015510511584579945, + 0.016658270731568336, + 0.014192922040820122, + 0.03331662714481354, + 0.028434205800294876, + 0.01673927716910839, + 0.019122816622257233, + 0.004123796708881855, + 0.0052541145123541355, + 0.006353107281029224, + 0.005929891485720873, + 0.005861807614564896, + 0.006309852935373783, + 0.0067884111776947975, + 0.007424038834869862, + 0.0035924760159105062, + 0.003675243118777871, + 0.0032467718701809645, + 0.003237801371142268, + 0.0024178521707654, + 0.003381565446034074, + 0.002479100599884987, + 0.003273079404607415, + 0.03377421200275421, + 0.040955182164907455, + 0.02535196766257286, + 0.02367832325398922, + 0.01300846878439188, + 0.009313119575381279, + 0.010505877435207367, + 0.02090146206319332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.34872007369995117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with an extreme close-up of tightly closed eyelids that tremble and flutter before opening?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the patient wearing a pale blue gown and lying on a flat white hospital pillow with messy brown hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start static and then slowly zoom out to a medium shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by bright direct sunlight streaming through a large window, casting sharp geometric shapes on the floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the patient's gaze scan the room, looking at an IV stand, a heart monitor, and a closed door on the far wall?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562269866466522 + ], + "opposite_logprob_tokens": [ + -1.6006226539611816 + ], + "logprob_reward": -0.22562269866466522, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562269866466522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030306074768304825, + 0.05193714052438736, + 0.04746592044830322, + 0.046768464148044586, + 0.03490522503852844, + 0.6127981543540955, + 0.10625120252370834, + 0.025425953790545464, + 0.018256863579154015, + 0.02588501013815403 + ], + "credit_map": [ + 0.005377491936087608, + 0.003323802025988698, + 0.002668464556336403, + 0.0026512001641094685, + 0.0044914172030985355, + 0.0032146815210580826, + 0.0021694961469620466, + 0.006409520749002695, + 0.006689039058983326, + 0.008265553042292595, + 0.00430286256596446, + 0.005158642306923866, + 0.010841786861419678, + 0.004668800160288811, + 0.004502842202782631, + 0.007507616654038429, + 0.002803371986374259, + 0.005887556821107864, + 0.003074405714869499, + 0.003922892268747091, + 0.010194825008511543, + 0.012411702424287796, + 0.003847480285912752, + 0.0053236884996294975, + 0.006990125402808189, + 0.00424837414175272, + 0.0066538709215819836, + 0.006392300128936768, + 0.004364264663308859, + 0.004324279259890318, + 0.009086878970265388, + 0.004708366002887487, + 0.0057181864976882935, + 0.003977203741669655, + 0.0046462020836770535, + 0.006138904951512814, + 0.0035343063063919544, + 0.00285092880949378, + 0.0040551540441811085, + 0.003984338603913784, + 0.06136403605341911, + 0.06606630235910416, + 0.0758473202586174, + 0.10065652430057526, + 0.05450429767370224, + 0.09056410938501358, + 0.06522996723651886, + 0.09856557101011276, + 0.006692206021398306, + 0.00939690601080656, + 0.013342862948775291, + 0.011156082153320312, + 0.016658015549182892, + 0.019532931968569756, + 0.010307421907782555, + 0.01916477456688881, + 0.0022657099179923534, + 0.0020956078078597784, + 0.002850373275578022, + 0.002148775150999427, + 0.001757841557264328, + 0.005579128861427307, + 0.0025224711280316114, + 0.006206047255545855, + 0.0018627921817824244, + 0.002097150543704629, + 0.002161234151571989, + 0.003168690949678421, + 0.002369938651099801, + 0.0020590401254594326, + 0.002233204897493124, + 0.00230481312610209, + 0.004371762741357088, + 0.0050338092260062695, + 0.003923118580132723, + 0.003667124779894948, + 0.0025290593039244413, + 0.001502507133409381, + 0.0018038294510915875, + 0.0030538002029061317 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 1 + ], + "selected_reward": -0.22562269866466522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_tensor([53, 53])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_tensor([53, 53])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..75a7f88fb29a6c2433ba18e7eac79024ff12d0f9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_tensor([53, 53])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "thrilling low-angle tracking shot closely follows a sleek, aerodynamic drone with a matte black carbon fiber shell as it rapidly moves forward. The drone navigates an incredibly narrow, steep-sided red sandstone canyon with astonishing speed and agility. The glaring midday sun streams vertically through the canyon's ceiling, casting deep, intense shadows across the jagged rock faces, creating a stark contrast to the illuminated rock surface. The drone's quadrotor propellers spin at high speed, creating a nearly imperceptible motion blur and emitting a deep, powerful hum. A sharp blue LED indicator lights up on the front of the drone, standing out against the warm background. During a particularly breathtaking maneuver, the drone suddenly banks to the right, its left wing skimming almost flat against the rough, wind-worn rock face, kicking up a cloud of fine sand and dust, each visible particle gleaming in the sunlight. The entire sequence is imbued with a calculated, dangerous beauty, creating an atmosphere of adrenaline, intense focus, and intense excitement.", + "seed": 13989, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_53_seed_13989_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04639285609591752, + "vlm_logprob": -0.04639285609591752, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04639285609591752, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07780192792415619 + ], + "opposite_logprob_tokens": [ + -3.0778019428253174 + ], + "logprob_reward": -0.07780192792415619, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.07780192792415619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19869454205036163, + 0.06739982217550278, + 0.0235983207821846, + 0.047851014882326126, + 0.19303575158119202, + 0.11008259654045105, + 0.09258540719747543, + 0.2113536149263382, + 0.030491728335618973, + 0.024907169863581657 + ], + "credit_map": [ + 0.011745262891054153, + 0.011367014609277248, + 0.015900762751698494, + 0.009631345979869366, + 0.012520081363618374, + 0.06165866553783417, + 0.017323631793260574, + 0.05854777991771698, + 0.003852677531540394, + 0.004169682040810585, + 0.0067609865218400955, + 0.0034516106825321913, + 0.008026273921132088, + 0.018995117396116257, + 0.006811280269175768, + 0.015332191251218319, + 0.003446015063673258, + 0.002635677345097065, + 0.002369088586419821, + 0.002266900148242712, + 0.0030923818703740835, + 0.003163732588291168, + 0.0032225484028458595, + 0.0034019765444099903, + 0.002025432186201215, + 0.004176033660769463, + 0.0019256776431575418, + 0.005239273887127638, + 0.006974366027861834, + 0.007365911267697811, + 0.007215318735688925, + 0.012929003685712814, + 0.01641179248690605, + 0.03046816773712635, + 0.012870865873992443, + 0.015153719112277031, + 0.016318775713443756, + 0.029260288923978806, + 0.026297466829419136, + 0.04625467583537102, + 0.00831248052418232, + 0.006433909758925438, + 0.012644241563975811, + 0.009726841002702713, + 0.015399417839944363, + 0.028824150562286377, + 0.01395050622522831, + 0.014791043475270271, + 0.009203130379319191, + 0.010596392676234245, + 0.011106875725090504, + 0.00926435086876154, + 0.009472979232668877, + 0.010163221508264542, + 0.013712272047996521, + 0.019066184759140015, + 0.026662249118089676, + 0.038958512246608734, + 0.08314362913370132, + 0.017159081995487213, + 0.006179902236908674, + 0.009867757558822632, + 0.008863545022904873, + 0.020518939942121506, + 0.003593501867726445, + 0.0037468564696609974, + 0.00440699839964509, + 0.0049266149289906025, + 0.0032450591679662466, + 0.004176758229732513, + 0.004356171935796738, + 0.00203976733610034, + 0.0034316403325647116, + 0.0031033277045935392, + 0.0032314949203282595, + 0.0031575923785567284, + 0.002307842019945383, + 0.0045547885820269585, + 0.0028908238746225834, + 0.002229660516604781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.07780192792415619 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004404813051224 + ], + "opposite_logprob_tokens": [ + -2.6950440406799316 + ], + "logprob_reward": -0.07004404813051224, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004404813051224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12576329708099365, + 0.13620072603225708, + 0.22260084748268127, + 0.04354448989033699, + 0.08162529766559601, + 0.06866775453090668, + 0.07918892055749893, + 0.11414802074432373, + 0.040982116013765335, + 0.08727847039699554 + ], + "credit_map": [ + 0.010121989995241165, + 0.00940938200801611, + 0.024932969361543655, + 0.0074282013811171055, + 0.013495832681655884, + 0.02048771269619465, + 0.010795509442687035, + 0.029091695323586464, + 0.00858829915523529, + 0.011774986982345581, + 0.011335932649672031, + 0.014074767008423805, + 0.017493143677711487, + 0.0195518396794796, + 0.03249052166938782, + 0.020891236141324043, + 0.021773671731352806, + 0.02734052762389183, + 0.015804065391421318, + 0.023250199854373932, + 0.04332868382334709, + 0.03338250890374184, + 0.029194017872214317, + 0.028527172282338142, + 0.002616289770230651, + 0.004463358782231808, + 0.004129992797970772, + 0.004580237902700901, + 0.007393898908048868, + 0.0055174329318106174, + 0.006382261402904987, + 0.008461019024252892, + 0.004284280817955732, + 0.007441811263561249, + 0.0059333122335374355, + 0.008886109106242657, + 0.012833560816943645, + 0.01223443727940321, + 0.011523266322910786, + 0.01848852075636387, + 0.004339301493018866, + 0.004027543589472771, + 0.00646217493340373, + 0.007294326554983854, + 0.01181080099195242, + 0.010839755646884441, + 0.013249536044895649, + 0.01064431481063366, + 0.007199555169790983, + 0.004993129521608353, + 0.00911241490393877, + 0.011439920403063297, + 0.007996107451617718, + 0.008422784507274628, + 0.013090437278151512, + 0.016934571787714958, + 0.013569359667599201, + 0.01237445417791605, + 0.021526634693145752, + 0.01525677926838398, + 0.012723974883556366, + 0.012470405548810959, + 0.01740860752761364, + 0.008817806839942932, + 0.005704631097614765, + 0.006785471457988024, + 0.004757338669151068, + 0.006161682307720184, + 0.006406147964298725, + 0.0046369051560759544, + 0.004280580673366785, + 0.002249358920380473, + 0.010730298236012459, + 0.009891612455248833, + 0.007443948648869991, + 0.010570156387984753, + 0.01154701504856348, + 0.022536160424351692, + 0.008823977783322334, + 0.00573531212285161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.07004404813051224 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015147175872698426 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015147175872698426, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015147175872698426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0603770911693573, + 0.22538267076015472, + 0.044504377990961075, + 0.02274259179830551, + 0.44691938161849976, + 0.04782378301024437, + 0.02825409546494484, + 0.07659146934747696, + 0.01384770404547453, + 0.033556774258613586 + ], + "credit_map": [ + 0.005505057517439127, + 0.004868345335125923, + 0.011671141721308231, + 0.0053819818422198296, + 0.006182607728987932, + 0.010278006084263325, + 0.005934759974479675, + 0.010555199347436428, + 0.01717093214392662, + 0.0142322713509202, + 0.023418430238962173, + 0.013324710540473461, + 0.022870583459734917, + 0.051619432866573334, + 0.024904415011405945, + 0.057841911911964417, + 0.003216787939891219, + 0.004379539750516415, + 0.003165910253301263, + 0.004364558029919863, + 0.008777279406785965, + 0.008198514580726624, + 0.006134255323559046, + 0.006267537362873554, + 0.0016127254348248243, + 0.002724666381254792, + 0.001846414408646524, + 0.0019722534343600273, + 0.0032857945188879967, + 0.002829375211149454, + 0.002951788017526269, + 0.005519576370716095, + 0.03865877538919449, + 0.06998644024133682, + 0.029857030138373375, + 0.0350780114531517, + 0.03201012685894966, + 0.0695330947637558, + 0.06178919970989227, + 0.11000671982765198, + 0.0049463436007499695, + 0.003258825745433569, + 0.008016214706003666, + 0.005359835457056761, + 0.004110084380954504, + 0.007285906467586756, + 0.006693676579743624, + 0.008152899332344532, + 0.003501051338389516, + 0.003260452998802066, + 0.00353875826112926, + 0.0031072290148586035, + 0.002287530805915594, + 0.002618463011458516, + 0.0038509999867528677, + 0.0060896133072674274, + 0.009559595957398415, + 0.014290633611381054, + 0.029954468831419945, + 0.006402268540114164, + 0.002642330713570118, + 0.003425363916903734, + 0.0030050980858504772, + 0.007311719004064798, + 0.002033510012552142, + 0.002334440825507045, + 0.0018854259978979826, + 0.0016703485744073987, + 0.001855142181739211, + 0.0017098115058615804, + 0.0015886311884969473, + 0.0007703948067501187, + 0.003987014759331942, + 0.0036396535579115152, + 0.0031420965678989887, + 0.004010008182376623, + 0.004366464912891388, + 0.008888023905456066, + 0.0033933152444660664, + 0.0021302017848938704 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.0015147175872698426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058218978345394 + ], + "opposite_logprob_tokens": [ + -3.8955821990966797 + ], + "logprob_reward": -0.02058218978345394, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058218978345394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09071484953165054, + 0.043505147099494934, + 0.040521763265132904, + 0.05706217512488365, + 0.14899855852127075, + 0.30907902121543884, + 0.07291080057621002, + 0.155465766787529, + 0.025565603747963905, + 0.05617636814713478 + ], + "credit_map": [ + 0.005206875037401915, + 0.004924066364765167, + 0.0155404731631279, + 0.005467991344630718, + 0.009463268332183361, + 0.018787644803524017, + 0.007848899811506271, + 0.02347562648355961, + 0.002557074185460806, + 0.003202304942533374, + 0.0035674734972417355, + 0.003506988286972046, + 0.006158754229545593, + 0.00854533165693283, + 0.006448145490139723, + 0.00951907504349947, + 0.004776922054588795, + 0.004575616680085659, + 0.0037085937801748514, + 0.0034022375475615263, + 0.006693006027489901, + 0.005388603080064058, + 0.005907008424401283, + 0.006069773808121681, + 0.003109031356871128, + 0.005407760851085186, + 0.004472332075238228, + 0.004384621512144804, + 0.006686096545308828, + 0.01099774893373251, + 0.007276555523276329, + 0.01472802646458149, + 0.0070901294238865376, + 0.019328584894537926, + 0.009106905199587345, + 0.015403379686176777, + 0.012630920857191086, + 0.026369180530309677, + 0.01975179836153984, + 0.03931765630841255, + 0.02343185432255268, + 0.018604954704642296, + 0.037172842770814896, + 0.03674125671386719, + 0.0536724328994751, + 0.058180782943964005, + 0.045084137469530106, + 0.03619072958827019, + 0.00683685764670372, + 0.004258900415152311, + 0.010657808743417263, + 0.00978820864111185, + 0.006708890199661255, + 0.007296996656805277, + 0.011887498199939728, + 0.015475640073418617, + 0.018406294286251068, + 0.02986246906220913, + 0.06133769825100899, + 0.012600311078131199, + 0.005254507064819336, + 0.006625881418585777, + 0.006808432284742594, + 0.014570169150829315, + 0.003802606835961342, + 0.003373921848833561, + 0.0033392691984772682, + 0.004607084207236767, + 0.0028610515873879194, + 0.003349428065121174, + 0.0025988535489887, + 0.0016333874082192779, + 0.0076511153019964695, + 0.006521408446133137, + 0.005434992723166943, + 0.007556173484772444, + 0.007680382113903761, + 0.011626115068793297, + 0.005746769718825817, + 0.003959408961236477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.02058218978345394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062021397054195404 + ], + "opposite_logprob_tokens": [ + -2.812021493911743 + ], + "logprob_reward": -0.062021397054195404, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062021397054195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31073665618896484, + 0.10607609152793884, + 0.11422263085842133, + 0.047469958662986755, + 0.10912995040416718, + 0.11147257685661316, + 0.04756425321102142, + 0.08383743464946747, + 0.03312808647751808, + 0.03636234626173973 + ], + "credit_map": [ + 0.021030057221651077, + 0.02286803163588047, + 0.04842110723257065, + 0.01668442226946354, + 0.03403157368302345, + 0.05962125584483147, + 0.02784467674791813, + 0.08023553341627121, + 0.007404463365674019, + 0.009246185421943665, + 0.008816827088594437, + 0.011485300958156586, + 0.012642699293792248, + 0.017723437398672104, + 0.019556500017642975, + 0.019200673326849937, + 0.010039564222097397, + 0.012843731790781021, + 0.013115870766341686, + 0.011635364033281803, + 0.017417944967746735, + 0.01517472229897976, + 0.018066655844449997, + 0.015928780660033226, + 0.002806346630677581, + 0.005285460036247969, + 0.004298101644963026, + 0.0061538624577224255, + 0.009075948968529701, + 0.00675670662894845, + 0.0065313540399074554, + 0.006562173832207918, + 0.005253935232758522, + 0.013014771044254303, + 0.005113786086440086, + 0.01262686401605606, + 0.015599005855619907, + 0.015734750777482986, + 0.018346568569540977, + 0.023440269753336906, + 0.00860779732465744, + 0.006258461158722639, + 0.012794310227036476, + 0.011450596153736115, + 0.014154935255646706, + 0.026839599013328552, + 0.01361751276999712, + 0.017749354243278503, + 0.0066823470406234264, + 0.0050089117139577866, + 0.004724789876490831, + 0.006421925500035286, + 0.006474630441516638, + 0.005481337197124958, + 0.0079620610922575, + 0.004808251280337572, + 0.011346353217959404, + 0.010108275339007378, + 0.02095925062894821, + 0.01095812302082777, + 0.008613173849880695, + 0.004880073945969343, + 0.010194788686931133, + 0.006777395494282246, + 0.0040986561216413975, + 0.004717804025858641, + 0.004034502897411585, + 0.004812831990420818, + 0.004795644897967577, + 0.003865808481350541, + 0.004917604848742485, + 0.0018852336797863245, + 0.0047590890899300575, + 0.004569337237626314, + 0.0038650166243314743, + 0.0036235619336366653, + 0.0037550777196884155, + 0.009646506048738956, + 0.0037623809184879065, + 0.002381376689299941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.062021397054195404 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "thrilling low-angle tracking shot closely follows a sleek, aerodynamic drone with a matte black carbon fiber shell as it rapidly moves forward. The drone navigates an incredibly narrow, steep-sided red sandstone canyon with astonishing speed and agility. The glaring midday sun streams vertically through the canyon's ceiling, casting deep, intense shadows across the jagged rock faces, creating a stark contrast to the illuminated rock surface. The drone's quadrotor propellers spin at high speed, creating a nearly imperceptible motion blur and emitting a deep, powerful hum. A sharp blue LED indicator lights up on the front of the drone, standing out against the warm background. During a particularly breathtaking maneuver, the drone suddenly banks to the right, its left wing skimming almost flat against the rough, wind-worn rock face, kicking up a cloud of fine sand and dust, each visible particle gleaming in the sunlight. The entire sequence is imbued with a calculated, dangerous beauty, creating an atmosphere of adrenaline, intense focus, and intense excitement.", + "seed": 13990, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_53_seed_13990_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.046934491582214834, + "vlm_logprob": -0.046934491582214834, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.046934491582214834, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08402406424283981 + ], + "opposite_logprob_tokens": [ + -2.959023952484131 + ], + "logprob_reward": -0.08402406424283981, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.08402406424283981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1974717378616333, + 0.07475343346595764, + 0.02711339108645916, + 0.037007857114076614, + 0.01960878260433674, + 0.34646597504615784, + 0.10492385923862457, + 0.0764329805970192, + 0.09330077469348907, + 0.022921254858374596 + ], + "credit_map": [ + 0.011107305996119976, + 0.004894407000392675, + 0.06932640820741653, + 0.008389499969780445, + 0.00695825507864356, + 0.07394690811634064, + 0.005806159693747759, + 0.01704280637204647, + 0.009710006415843964, + 0.005713251885026693, + 0.011469379998743534, + 0.005999238230288029, + 0.011878003366291523, + 0.013560142368078232, + 0.0076397862285375595, + 0.008783622644841671, + 0.004777212627232075, + 0.003444064175710082, + 0.004072102252393961, + 0.0023043875116854906, + 0.001977545442059636, + 0.0034615916665643454, + 0.0018339762464165688, + 0.005242510698735714, + 0.005354691296815872, + 0.0032073697075247765, + 0.005081926006823778, + 0.0034064087085425854, + 0.005015117581933737, + 0.0064191860146820545, + 0.003639800241217017, + 0.004883357789367437, + 0.0033426701556891203, + 0.0034583124797791243, + 0.0038938350044190884, + 0.0015489611541852355, + 0.0016648145392537117, + 0.0020572857465595007, + 0.0015272144228219986, + 0.002115689916536212, + 0.06367374211549759, + 0.10234323143959045, + 0.03083164244890213, + 0.015622438862919807, + 0.025389250367879868, + 0.04932646080851555, + 0.019284773617982864, + 0.03999442234635353, + 0.014653990976512432, + 0.013013150542974472, + 0.012515719048678875, + 0.00737739959731698, + 0.013834506273269653, + 0.015289354138076305, + 0.006518173031508923, + 0.021721571683883667, + 0.008729169145226479, + 0.007356537040323019, + 0.007867571897804737, + 0.009620464406907558, + 0.007782402914017439, + 0.01294658612459898, + 0.011686380952596664, + 0.010443869046866894, + 0.012895079329609871, + 0.009693993255496025, + 0.008928844705224037, + 0.010337094776332378, + 0.010584886185824871, + 0.015411916188895702, + 0.009447481483221054, + 0.01600148156285286, + 0.00455812131986022, + 0.0032050302252173424, + 0.0029654796235263348, + 0.0021911708172410727, + 0.0025381750892847776, + 0.002953278599306941, + 0.001961346482858062, + 0.0025486533995717764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.08402406424283981 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866567999124527 + ], + "opposite_logprob_tokens": [ + -3.048665761947632 + ], + "logprob_reward": -0.04866567999124527, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866567999124527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4032207429409027, + 0.13342827558517456, + 0.021313359960913658, + 0.013543643057346344, + 0.05690283700823784, + 0.19565914571285248, + 0.0304983239620924, + 0.0354325957596302, + 0.07407385855913162, + 0.035927243530750275 + ], + "credit_map": [ + 0.02363653853535652, + 0.009460506960749626, + 0.13476407527923584, + 0.015031281858682632, + 0.014673324301838875, + 0.15052013099193573, + 0.011294282041490078, + 0.043840594589710236, + 0.01929599791765213, + 0.009330758824944496, + 0.015081080608069897, + 0.011150360107421875, + 0.021900860592722893, + 0.023524854332208633, + 0.01570594310760498, + 0.01743842102587223, + 0.003750840900465846, + 0.0030846279114484787, + 0.002245257841423154, + 0.002851696452125907, + 0.0013158994261175394, + 0.0023379805497825146, + 0.0019455521833151579, + 0.0037815028335899115, + 0.0013954094611108303, + 0.0019115335308015347, + 0.0017366281244903803, + 0.0023426951374858618, + 0.001031211344525218, + 0.0014886755961924791, + 0.0019012222765013576, + 0.0017362663056701422, + 0.00776650570333004, + 0.008039077743887901, + 0.009562473744153976, + 0.00663728779181838, + 0.0028934208676218987, + 0.00844760425388813, + 0.006457500625401735, + 0.007098966743797064, + 0.03164433315396309, + 0.05197562277317047, + 0.01952989585697651, + 0.008890172466635704, + 0.017610900104045868, + 0.029534270986914635, + 0.01720370352268219, + 0.019270215183496475, + 0.003912152256816626, + 0.0031533536966890097, + 0.0037132834549993277, + 0.0029106137808412313, + 0.002443283796310425, + 0.006890768185257912, + 0.0031669707968831062, + 0.004307896364480257, + 0.004358481615781784, + 0.004193832166492939, + 0.003769315080717206, + 0.0033993644174188375, + 0.006048070266842842, + 0.004876853432506323, + 0.0037704857531934977, + 0.005016192328184843, + 0.010164964012801647, + 0.01640532724559307, + 0.01020987518131733, + 0.005352528300136328, + 0.007944812998175621, + 0.008524811826646328, + 0.006619773805141449, + 0.008851759135723114, + 0.006257232744246721, + 0.004329351708292961, + 0.005184773821383715, + 0.004515087231993675, + 0.0038513275794684887, + 0.004935473203659058, + 0.0035424185916781425, + 0.0033115760888904333 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.04866567999124527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021968772634863853 + ], + "opposite_logprob_tokens": [ + -6.127196788787842 + ], + "logprob_reward": -0.0021968772634863853, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021968772634863853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08741326630115509, + 0.020504729822278023, + 0.03695345297455788, + 0.16358619928359985, + 0.18805356323719025, + 0.1932455450296402, + 0.07259967923164368, + 0.03701329603791237, + 0.12964801490306854, + 0.07098221033811569 + ], + "credit_map": [ + 0.00847142655402422, + 0.003642655210569501, + 0.024379108101129532, + 0.005319070070981979, + 0.0037049921229481697, + 0.02170407958328724, + 0.005591291002929211, + 0.014600643888115883, + 0.0035444917157292366, + 0.0016277296235784888, + 0.003244664054363966, + 0.0026944305282086134, + 0.0027450828347355127, + 0.0028552694711834192, + 0.0018403996946290135, + 0.001952660153619945, + 0.007085889112204313, + 0.005965434480458498, + 0.004238891880959272, + 0.0034831520169973373, + 0.003626871155574918, + 0.005233685486018658, + 0.0027062706649303436, + 0.004613255616277456, + 0.02704727090895176, + 0.019789619371294975, + 0.02097918651998043, + 0.014766144566237926, + 0.01708836667239666, + 0.02956525608897209, + 0.013666749000549316, + 0.020683610811829567, + 0.02900613099336624, + 0.024770785123109818, + 0.03389895707368851, + 0.018100827932357788, + 0.010331027209758759, + 0.025837799534201622, + 0.021779300644993782, + 0.024328742176294327, + 0.03197185695171356, + 0.056807514280080795, + 0.018227145075798035, + 0.00913166906684637, + 0.015040124766528606, + 0.02976989559829235, + 0.009922029450535774, + 0.022375306114554405, + 0.009074009954929352, + 0.010314022190868855, + 0.009189112111926079, + 0.005496786907315254, + 0.011938693933188915, + 0.009827140718698502, + 0.00648635346442461, + 0.01027355995029211, + 0.006545853801071644, + 0.00513925077393651, + 0.003356104949489236, + 0.002562428591772914, + 0.005128683988004923, + 0.005166305229067802, + 0.003664495423436165, + 0.005450171418488026, + 0.018409695476293564, + 0.019615257158875465, + 0.014243409037590027, + 0.012277038767933846, + 0.014305406250059605, + 0.017401576042175293, + 0.013983662240207195, + 0.0194119643419981, + 0.01044352538883686, + 0.013509932905435562, + 0.010695171542465687, + 0.0057416209019720554, + 0.009528050199151039, + 0.011228797025978565, + 0.0036770133301615715, + 0.006158097181469202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.0021968772634863853 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029793808236718178 + ], + "opposite_logprob_tokens": [ + -3.5297937393188477 + ], + "logprob_reward": -0.029793808236718178, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029793808236718178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21187150478363037, + 0.19073960185050964, + 0.03170968219637871, + 0.03062519244849682, + 0.047016941010951996, + 0.13918080925941467, + 0.06303051859140396, + 0.14962032437324524, + 0.08147440105676651, + 0.05473095551133156 + ], + "credit_map": [ + 0.01498013362288475, + 0.0054200743325054646, + 0.06822369992733002, + 0.009263441897928715, + 0.007588011212646961, + 0.06951692700386047, + 0.007721470203250647, + 0.029157770797610283, + 0.02770288847386837, + 0.013806154020130634, + 0.023565256968140602, + 0.016257161274552345, + 0.03340739756822586, + 0.030437201261520386, + 0.0240790918469429, + 0.021484464406967163, + 0.005173369310796261, + 0.00501703517511487, + 0.0033930549398064613, + 0.005637950263917446, + 0.0025718987453728914, + 0.003267771564424038, + 0.002841395325958729, + 0.003807206405326724, + 0.0045004249550402164, + 0.0035414777230471373, + 0.004899658728390932, + 0.003456955775618553, + 0.002978748409077525, + 0.004622896667569876, + 0.002719900570809841, + 0.0039051335770636797, + 0.006515185814350843, + 0.004843299742788076, + 0.005774037446826696, + 0.003194873919710517, + 0.003495627548545599, + 0.012460541911423206, + 0.004311803262680769, + 0.006421579513698816, + 0.02548225224018097, + 0.04309951886534691, + 0.011956541799008846, + 0.006008269730955362, + 0.013870005495846272, + 0.017936576157808304, + 0.00667546596378088, + 0.01415220182389021, + 0.008288641460239887, + 0.006532701663672924, + 0.006053145043551922, + 0.007067262660712004, + 0.005461264867335558, + 0.009954026900231838, + 0.007585549261420965, + 0.012087935581803322, + 0.024201491847634315, + 0.020086057484149933, + 0.01419265940785408, + 0.011114043183624744, + 0.022507844492793083, + 0.01991402544081211, + 0.017211783677339554, + 0.02039245329797268, + 0.008400066755712032, + 0.010203015990555286, + 0.008508045226335526, + 0.009826370514929295, + 0.009806038811802864, + 0.01176037359982729, + 0.011687642894685268, + 0.011282853782176971, + 0.00908992625772953, + 0.006415885873138905, + 0.008295265026390553, + 0.0049271974712610245, + 0.006593370344489813, + 0.008098454214632511, + 0.0039032548666000366, + 0.007407607976347208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.029793808236718178 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06999202817678452 + ], + "opposite_logprob_tokens": [ + -2.6949920654296875 + ], + "logprob_reward": -0.06999202817678452, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.06999202817678452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1972542256116867, + 0.22384394705295563, + 0.01761142909526825, + 0.02781630866229534, + 0.0574251189827919, + 0.19757968187332153, + 0.044613711535930634, + 0.100658118724823, + 0.06220462918281555, + 0.07099290192127228 + ], + "credit_map": [ + 0.012478879652917385, + 0.006752378772944212, + 0.07121991366147995, + 0.010642913170158863, + 0.007253008894622326, + 0.06880085170269012, + 0.0050661894492805, + 0.015040060505270958, + 0.028542404994368553, + 0.016350604593753815, + 0.03170425072312355, + 0.018584366887807846, + 0.03225291892886162, + 0.042531516402959824, + 0.024040358141064644, + 0.02983750030398369, + 0.0028514035511761904, + 0.001951546873897314, + 0.0025231034960597754, + 0.0025520308408886194, + 0.0018126521026715636, + 0.00195496017113328, + 0.0016105934046208858, + 0.002355137374252081, + 0.0029913936741650105, + 0.003484805580228567, + 0.003672756487503648, + 0.004495476838201284, + 0.0026612968649715185, + 0.004033567849546671, + 0.0032319221645593643, + 0.003245085244998336, + 0.006708142347633839, + 0.007691744714975357, + 0.009995028376579285, + 0.00855479296296835, + 0.0028033992275595665, + 0.006330468691885471, + 0.009072978049516678, + 0.006268554832786322, + 0.03667505457997322, + 0.0506647564470768, + 0.017170405015349388, + 0.009049823507666588, + 0.016466982662677765, + 0.029581615701317787, + 0.014885745011270046, + 0.02308526635169983, + 0.0042427172884345055, + 0.004448707215487957, + 0.004420354962348938, + 0.005785634741187096, + 0.006121362093836069, + 0.00483900960534811, + 0.00593171501532197, + 0.00882420502603054, + 0.012270542792975903, + 0.009317105636000633, + 0.010069964453577995, + 0.009580027312040329, + 0.011512275785207748, + 0.01811557076871395, + 0.013694046065211296, + 0.016098570078611374, + 0.007729565259069204, + 0.005874706897884607, + 0.00565029913559556, + 0.007942360825836658, + 0.006793087814003229, + 0.009493771009147167, + 0.008033208549022675, + 0.010687618516385555, + 0.011469537392258644, + 0.009337421506643295, + 0.008464342914521694, + 0.007307028863579035, + 0.009450589306652546, + 0.009637574665248394, + 0.007428004872053862, + 0.007898392155766487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.06999202817678452 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "thrilling low-angle tracking shot closely follows a sleek, aerodynamic drone with a matte black carbon fiber shell as it rapidly moves forward. The drone navigates an incredibly narrow, steep-sided red sandstone canyon with astonishing speed and agility. The glaring midday sun streams vertically through the canyon's ceiling, casting deep, intense shadows across the jagged rock faces, creating a stark contrast to the illuminated rock surface. The drone's quadrotor propellers spin at high speed, creating a nearly imperceptible motion blur and emitting a deep, powerful hum. A sharp blue LED indicator lights up on the front of the drone, standing out against the warm background. During a particularly breathtaking maneuver, the drone suddenly banks to the right, its left wing skimming almost flat against the rough, wind-worn rock face, kicking up a cloud of fine sand and dust, each visible particle gleaming in the sunlight. The entire sequence is imbued with a calculated, dangerous beauty, creating an atmosphere of adrenaline, intense focus, and intense excitement.", + "seed": 13991, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_53_seed_13991_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.041814434039406476, + "vlm_logprob": -0.041814434039406476, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.041814434039406476, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05542809143662453 + ], + "opposite_logprob_tokens": [ + -4.4304280281066895 + ], + "logprob_reward": -0.05542809143662453, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.05542809143662453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08589759469032288, + 0.05381381884217262, + 0.06250345706939697, + 0.09637430310249329, + 0.061193905770778656, + 0.07368478924036026, + 0.07443059235811234, + 0.03362271934747696, + 0.30724191665649414, + 0.15123683214187622 + ], + "credit_map": [ + 0.013534968718886375, + 0.015446395613253117, + 0.01570476032793522, + 0.00967155210673809, + 0.008827839978039265, + 0.00959460623562336, + 0.0076380143873393536, + 0.005479465704411268, + 0.012515665963292122, + 0.00866734329611063, + 0.007591091096401215, + 0.007810907904058695, + 0.0031517769675701857, + 0.004276065155863762, + 0.004992389120161533, + 0.004808582831174135, + 0.007977643981575966, + 0.009060739539563656, + 0.007275757379829884, + 0.008503733202815056, + 0.007198922336101532, + 0.008863282389938831, + 0.006023763678967953, + 0.007599621079862118, + 0.012985103763639927, + 0.01604148931801319, + 0.01610400155186653, + 0.0101308049634099, + 0.01359662413597107, + 0.01088233944028616, + 0.007241258397698402, + 0.009392695501446724, + 0.008561867289245129, + 0.008521628566086292, + 0.008687177672982216, + 0.00803252775222063, + 0.006005905102938414, + 0.005650305189192295, + 0.005965894088149071, + 0.009768599644303322, + 0.015438747592270374, + 0.011470911093056202, + 0.01348565611988306, + 0.005775436758995056, + 0.006558306515216827, + 0.006956515833735466, + 0.006281673442572355, + 0.007717558182775974, + 0.009183799847960472, + 0.009084667079150677, + 0.00802994892001152, + 0.005257642362266779, + 0.013293248601257801, + 0.012210584245622158, + 0.007690249942243099, + 0.009680460207164288, + 0.0034165424294769764, + 0.004691985435783863, + 0.003916722256690264, + 0.004398698918521404, + 0.0034087353851646185, + 0.004552650731056929, + 0.004793516360223293, + 0.0044438717886805534, + 0.008472045883536339, + 0.023681355640292168, + 0.005336719565093517, + 0.004627367947250605, + 0.13079822063446045, + 0.018513040617108345, + 0.006071568466722965, + 0.10974160581827164, + 0.007668676320463419, + 0.0074518960900604725, + 0.005475691519677639, + 0.007848845794796944, + 0.011995512060821056, + 0.05528990924358368, + 0.0091028967872262, + 0.04640340805053711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.05542809143662453 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10032714158296585 + ], + "opposite_logprob_tokens": [ + -2.350327253341675 + ], + "logprob_reward": -0.10032714158296585, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10032714158296585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16448986530303955, + 0.043145544826984406, + 0.0788412019610405, + 0.20893697440624237, + 0.06257160007953644, + 0.07933274656534195, + 0.17743061482906342, + 0.03414511680603027, + 0.06884579360485077, + 0.08226057887077332 + ], + "credit_map": [ + 0.03372235223650932, + 0.026235414668917656, + 0.028542522341012955, + 0.018870150670409203, + 0.01339902263134718, + 0.010967222973704338, + 0.015876900404691696, + 0.01687627099454403, + 0.009652060456573963, + 0.006611169781535864, + 0.007628178223967552, + 0.003050591330975294, + 0.0035911148879677057, + 0.0063454522751271725, + 0.0025361357256770134, + 0.003730839816853404, + 0.00722756190225482, + 0.01621333882212639, + 0.015063134953379631, + 0.007408302742987871, + 0.009209587238729, + 0.009069977328181267, + 0.004928596317768097, + 0.00972069799900055, + 0.04201285168528557, + 0.02825935371220112, + 0.025890188291668892, + 0.026919767260551453, + 0.013594950549304485, + 0.01922774314880371, + 0.0168314166367054, + 0.03620067611336708, + 0.0068110814318060875, + 0.007806953974068165, + 0.008128312416374683, + 0.010257948189973831, + 0.005841086618602276, + 0.006314517930150032, + 0.006840166635811329, + 0.01057153008878231, + 0.010381674394011497, + 0.01238629687577486, + 0.030797088518738747, + 0.0037540015764534473, + 0.003140232991427183, + 0.01159712579101324, + 0.002984001999720931, + 0.004292326048016548, + 0.014417486265301704, + 0.020292164757847786, + 0.02291233278810978, + 0.016480831429362297, + 0.020641792565584183, + 0.03667551651597023, + 0.017063090577721596, + 0.02894739620387554, + 0.003085945500060916, + 0.0047170245088636875, + 0.0029987827874720097, + 0.003239966696128249, + 0.0030603979248553514, + 0.005486578214913607, + 0.006366745103150606, + 0.005189673975110054, + 0.004428068641573191, + 0.00689502852037549, + 0.004320942331105471, + 0.0068517220206558704, + 0.013454706408083439, + 0.013671659864485264, + 0.008586007170379162, + 0.01063765399158001, + 0.005199206527322531, + 0.005663849879056215, + 0.003966605290770531, + 0.010959254577755928, + 0.009640812873840332, + 0.019172169268131256, + 0.008954876102507114, + 0.018703801557421684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 0 + ], + "selected_reward": -0.10032714158296585 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010433712741360068 + ], + "opposite_logprob_tokens": [ + -6.876043319702148 + ], + "logprob_reward": -0.0010433712741360068, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010433712741360068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05191368982195854, + 0.033225588500499725, + 0.032119326293468475, + 0.07271748036146164, + 0.015819113701581955, + 0.1663600355386734, + 0.07714249938726425, + 0.01627727597951889, + 0.07452700287103653, + 0.4598980247974396 + ], + "credit_map": [ + 0.008814367465674877, + 0.010422448627650738, + 0.007087016943842173, + 0.006666440982371569, + 0.005602726712822914, + 0.003879052121192217, + 0.0056966496631503105, + 0.0037449891678988934, + 0.007689379155635834, + 0.005617760121822357, + 0.005909926723688841, + 0.0026345259975641966, + 0.0022785228211432695, + 0.003971119876950979, + 0.0024020448327064514, + 0.002722307341173291, + 0.0038578214589506388, + 0.00643266411498189, + 0.009067605249583721, + 0.0026771065313369036, + 0.002835995750501752, + 0.0025788682978600264, + 0.0021136982832103968, + 0.002555564045906067, + 0.014468220993876457, + 0.013331162743270397, + 0.011062512174248695, + 0.0074453940615057945, + 0.006022846791893244, + 0.008379324339330196, + 0.0044755176641047, + 0.007532503921538591, + 0.0021698682103306055, + 0.002349998103454709, + 0.002077596727758646, + 0.0016193435294553638, + 0.002545599825680256, + 0.0025296444073319435, + 0.0010344644542783499, + 0.0014925996074452996, + 0.023662064224481583, + 0.024026568979024887, + 0.060619108378887177, + 0.010023069567978382, + 0.007155901286751032, + 0.024771234020590782, + 0.007097057998180389, + 0.009005030617117882, + 0.008213164284825325, + 0.009679784066975117, + 0.009467965923249722, + 0.0054999650456011295, + 0.01411216426640749, + 0.012430149130523205, + 0.0055358270183205605, + 0.012203476391732693, + 0.0020356308668851852, + 0.0020326755475252867, + 0.0016950422432273626, + 0.0019950619898736477, + 0.002113268245011568, + 0.0024415827356278896, + 0.0019366906490176916, + 0.0020273225381970406, + 0.0021547717042267323, + 0.005805237218737602, + 0.001264914870262146, + 0.0015913554234430194, + 0.03091857023537159, + 0.004406337160617113, + 0.0019512601429596543, + 0.026434551924467087, + 0.011780845001339912, + 0.0103779137134552, + 0.007798179052770138, + 0.02291109412908554, + 0.03373124450445175, + 0.18613781034946442, + 0.02306729182600975, + 0.16409362852573395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.0010433712741360068 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195887371897697 + ], + "opposite_logprob_tokens": [ + -4.264195919036865 + ], + "logprob_reward": -0.014195887371897697, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195887371897697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09968345612287521, + 0.0817674845457077, + 0.13251179456710815, + 0.06440214067697525, + 0.07092159986495972, + 0.2792717516422272, + 0.048832666128873825, + 0.013496209867298603, + 0.1627257615327835, + 0.046387143433094025 + ], + "credit_map": [ + 0.017637664452195168, + 0.015899276360869408, + 0.0194030050188303, + 0.011736097745597363, + 0.010843691416084766, + 0.009942557662725449, + 0.008176382631063461, + 0.006044784095138311, + 0.017847349867224693, + 0.014100547879934311, + 0.011184818111360073, + 0.011832294054329395, + 0.005405970383435488, + 0.007363889366388321, + 0.007357014808803797, + 0.006675601005554199, + 0.021118100732564926, + 0.022058609873056412, + 0.03843606635928154, + 0.008779341354966164, + 0.009731934405863285, + 0.00861027929931879, + 0.010113867931067944, + 0.013663592748343945, + 0.008932643570005894, + 0.010274981148540974, + 0.009782679378986359, + 0.006620226893573999, + 0.00904474314302206, + 0.008297149091959, + 0.004658881109207869, + 0.006790833547711372, + 0.00855198036879301, + 0.009427557699382305, + 0.009302370250225067, + 0.010542492382228374, + 0.0063668969087302685, + 0.0071539911441504955, + 0.009308671578764915, + 0.010267641395330429, + 0.0369701087474823, + 0.03592998906970024, + 0.07968252897262573, + 0.015850404277443886, + 0.036296043545007706, + 0.03700869157910347, + 0.02034926600754261, + 0.017184719443321228, + 0.004626363515853882, + 0.006349480245262384, + 0.00665149325504899, + 0.00626826798543334, + 0.005628049839287996, + 0.0062459236942231655, + 0.0063500734977424145, + 0.006713016889989376, + 0.00111038435716182, + 0.0018269404536113143, + 0.0009628421394154429, + 0.0013471829006448388, + 0.0019169244915246964, + 0.0023381179198622704, + 0.0021096656564623117, + 0.0018841505516320467, + 0.004646265413612127, + 0.012509657070040703, + 0.002923733089119196, + 0.003381890244781971, + 0.06833434849977493, + 0.009833906777203083, + 0.0035357652232050896, + 0.05756019800901413, + 0.0024476510006934404, + 0.002321961335837841, + 0.0016406859504058957, + 0.0038381987251341343, + 0.004000722896307707, + 0.015148961916565895, + 0.003336403053253889, + 0.013652559369802475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.014195887371897697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807767853140831 + ], + "opposite_logprob_tokens": [ + -3.2880775928497314 + ], + "logprob_reward": -0.03807767853140831, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807767853140831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2645498514175415, + 0.045058347284793854, + 0.0536847859621048, + 0.05399778112769127, + 0.030808977782726288, + 0.18914955854415894, + 0.12928980588912964, + 0.02488677203655243, + 0.11434130370616913, + 0.09423284977674484 + ], + "credit_map": [ + 0.04427148401737213, + 0.042997729033231735, + 0.033895574510097504, + 0.03833543509244919, + 0.028199804946780205, + 0.029293153434991837, + 0.02769284136593342, + 0.019863801077008247, + 0.009120505303144455, + 0.008812894113361835, + 0.00823210272938013, + 0.0040908572264015675, + 0.003549139713868499, + 0.0034798786509782076, + 0.004150121472775936, + 0.0036228494718670845, + 0.005040554795414209, + 0.013653997331857681, + 0.014511693269014359, + 0.0037949655670672655, + 0.0043113864958286285, + 0.0034585031680762768, + 0.0037778494879603386, + 0.005135834217071533, + 0.010530347935855389, + 0.007444596849381924, + 0.00812132190912962, + 0.006173156201839447, + 0.006462682504206896, + 0.005535791628062725, + 0.0047876108437776566, + 0.004942267667502165, + 0.003513464704155922, + 0.003771463641896844, + 0.004445572383701801, + 0.006498740520328283, + 0.001832336187362671, + 0.0036966020707041025, + 0.002179768867790699, + 0.004871030803769827, + 0.025095650926232338, + 0.02885233610868454, + 0.0653957650065422, + 0.010725338943302631, + 0.012851853854954243, + 0.02649194933474064, + 0.009918980300426483, + 0.009817678481340408, + 0.009954187087714672, + 0.014665577560663223, + 0.017151515930891037, + 0.017499741166830063, + 0.012355303391814232, + 0.024659644812345505, + 0.01308826357126236, + 0.019915569573640823, + 0.0034430529922246933, + 0.003424344351515174, + 0.0020893269684165716, + 0.0027093796525150537, + 0.0036635934375226498, + 0.0025899556931108236, + 0.0035294045228511095, + 0.003437715582549572, + 0.004239547532051802, + 0.01011408306658268, + 0.0035079147201031446, + 0.006712737958878279, + 0.043774351477622986, + 0.008092335425317287, + 0.004427944775670767, + 0.03347240015864372, + 0.005977134220302105, + 0.004802399780601263, + 0.0037773135118186474, + 0.013371181674301624, + 0.007948442362248898, + 0.02489372342824936, + 0.01039463747292757, + 0.023068012669682503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.03807767853140831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "thrilling low-angle tracking shot closely follows a sleek, aerodynamic drone with a matte black carbon fiber shell as it rapidly moves forward. The drone navigates an incredibly narrow, steep-sided red sandstone canyon with astonishing speed and agility. The glaring midday sun streams vertically through the canyon's ceiling, casting deep, intense shadows across the jagged rock faces, creating a stark contrast to the illuminated rock surface. The drone's quadrotor propellers spin at high speed, creating a nearly imperceptible motion blur and emitting a deep, powerful hum. A sharp blue LED indicator lights up on the front of the drone, standing out against the warm background. During a particularly breathtaking maneuver, the drone suddenly banks to the right, its left wing skimming almost flat against the rough, wind-worn rock face, kicking up a cloud of fine sand and dust, each visible particle gleaming in the sunlight. The entire sequence is imbued with a calculated, dangerous beauty, creating an atmosphere of adrenaline, intense focus, and intense excitement.", + "seed": 13992, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_5/index_53_seed_13992_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0325140691595152, + "vlm_logprob": -0.0325140691595152, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.0325140691595152, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047748252749443054 + ], + "opposite_logprob_tokens": [ + -4.29774808883667 + ], + "logprob_reward": -0.047748252749443054, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.047748252749443054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15875205397605896, + 0.0863410085439682, + 0.06287085264921188, + 0.10371775180101395, + 0.0760573297739029, + 0.06993298977613449, + 0.07622932642698288, + 0.17685148119926453, + 0.11068906635046005, + 0.07855813950300217 + ], + "credit_map": [ + 0.017478175461292267, + 0.014046371914446354, + 0.02316112071275711, + 0.018186863511800766, + 0.012598896399140358, + 0.020998021587729454, + 0.02783074416220188, + 0.02445186674594879, + 0.016032932326197624, + 0.010164733044803143, + 0.009745311923325062, + 0.005111456848680973, + 0.011465107090771198, + 0.01111906673759222, + 0.009280004538595676, + 0.013422396034002304, + 0.004244610667228699, + 0.0047734878025949, + 0.00614775950089097, + 0.0055160257034003735, + 0.009133893996477127, + 0.012058153748512268, + 0.0068473671562969685, + 0.014149554073810577, + 0.011710171587765217, + 0.006276650354266167, + 0.016426319256424904, + 0.006950875744223595, + 0.012646026909351349, + 0.019538333639502525, + 0.011627642437815666, + 0.018541742116212845, + 0.004576447419822216, + 0.007226261310279369, + 0.006696572992950678, + 0.008257304318249226, + 0.013257975690066814, + 0.013145583681762218, + 0.009732304140925407, + 0.013164880685508251, + 0.006537854205816984, + 0.013756532222032547, + 0.0069752526469528675, + 0.004801210947334766, + 0.010474897921085358, + 0.011437560431659222, + 0.005857301410287619, + 0.010092378593981266, + 0.0089631462469697, + 0.009086445905268192, + 0.008872564882040024, + 0.007442074827849865, + 0.011181000620126724, + 0.009912604466080666, + 0.00847555510699749, + 0.012295937165617943, + 0.01944432221353054, + 0.020466197282075882, + 0.020840544253587723, + 0.019279515370726585, + 0.030227256938815117, + 0.027529560029506683, + 0.020183488726615906, + 0.01888059824705124, + 0.013961134478449821, + 0.01046181470155716, + 0.00965013075619936, + 0.011076546274125576, + 0.01216423511505127, + 0.014762918464839458, + 0.023134833201766014, + 0.015477458015084267, + 0.011343494057655334, + 0.009775460697710514, + 0.01234496384859085, + 0.008479372598230839, + 0.011160388588905334, + 0.00826468225568533, + 0.008077207021415234, + 0.009112566709518433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.047748252749443054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497652664780617 + ], + "opposite_logprob_tokens": [ + -2.929976463317871 + ], + "logprob_reward": -0.05497652664780617, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497652664780617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20917704701423645, + 0.05129184573888779, + 0.1060216948390007, + 0.07549645751714706, + 0.054492320865392685, + 0.03839114308357239, + 0.1406756490468979, + 0.19683364033699036, + 0.08946651965379715, + 0.03815365582704544 + ], + "credit_map": [ + 0.0268855057656765, + 0.01794932596385479, + 0.03563659265637398, + 0.01619294472038746, + 0.016478030011057854, + 0.030237486585974693, + 0.024201685562729836, + 0.04159548506140709, + 0.007574340328574181, + 0.005549028981477022, + 0.005376721732318401, + 0.0028052153065800667, + 0.008067076094448566, + 0.006908476818352938, + 0.0039001288823783398, + 0.011110860854387283, + 0.011995709501206875, + 0.00675443559885025, + 0.016046419739723206, + 0.012232622131705284, + 0.011477967724204063, + 0.012868957594037056, + 0.021264664828777313, + 0.013380924239754677, + 0.010642347857356071, + 0.00938478298485279, + 0.008995063602924347, + 0.005032135173678398, + 0.010615411214530468, + 0.01349689718335867, + 0.0066858078353106976, + 0.010644007474184036, + 0.0071032061241567135, + 0.007177151273936033, + 0.0074600218795239925, + 0.006149705033749342, + 0.008800461888313293, + 0.006561623886227608, + 0.005560761783272028, + 0.005679387599229813, + 0.002660101978108287, + 0.0063534206710755825, + 0.004575610626488924, + 0.005134004633873701, + 0.00672962935641408, + 0.004403882659971714, + 0.003835022449493408, + 0.004699470475316048, + 0.014633837155997753, + 0.015734748914837837, + 0.016311440616846085, + 0.012511792592704296, + 0.03663964197039604, + 0.01815912500023842, + 0.010964603163301945, + 0.015720464289188385, + 0.019058581441640854, + 0.032184913754463196, + 0.020730243995785713, + 0.015598688274621964, + 0.029977139085531235, + 0.027474798262119293, + 0.03249845281243324, + 0.019310839474201202, + 0.007095740642398596, + 0.010383108630776405, + 0.006668605841696262, + 0.009772800840437412, + 0.01893499866127968, + 0.01368979923427105, + 0.009493716061115265, + 0.013427749276161194, + 0.003361630719155073, + 0.004375857766717672, + 0.0037639341317117214, + 0.004573195241391659, + 0.004205618519335985, + 0.004389127716422081, + 0.006776994094252586, + 0.006707300432026386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.05497652664780617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011804286623373628 + ], + "opposite_logprob_tokens": [ + -6.751180648803711 + ], + "logprob_reward": -0.0011804286623373628, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011804286623373628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24302887916564941, + 0.037102069705724716, + 0.10496414452791214, + 0.09994413703680038, + 0.052982673048973083, + 0.08384352177381516, + 0.07863326370716095, + 0.1424224078655243, + 0.11409974098205566, + 0.04297921061515808 + ], + "credit_map": [ + 0.019658872857689857, + 0.019509540870785713, + 0.025354065001010895, + 0.01890699751675129, + 0.033274080604314804, + 0.03873253986239433, + 0.038412418216466904, + 0.04918035864830017, + 0.010215122252702713, + 0.004510125610977411, + 0.004697826690971851, + 0.001953687984496355, + 0.0029702959582209587, + 0.004138446878641844, + 0.002639107871800661, + 0.005977456923574209, + 0.012328773736953735, + 0.009101692587137222, + 0.014997153542935848, + 0.011851032264530659, + 0.010768011212348938, + 0.011742550879716873, + 0.02158324420452118, + 0.012591681443154812, + 0.011454356834292412, + 0.015063884668052197, + 0.016627468168735504, + 0.007801617030054331, + 0.012542675249278545, + 0.01298082247376442, + 0.01031846646219492, + 0.013154839165508747, + 0.005879056639969349, + 0.006711540278047323, + 0.007503350730985403, + 0.0044195507653057575, + 0.006121263839304447, + 0.01036590151488781, + 0.0036909026093780994, + 0.008291100151836872, + 0.007061624899506569, + 0.017560262233018875, + 0.006934781093150377, + 0.004984928760677576, + 0.011490157805383205, + 0.016191978007555008, + 0.006960339844226837, + 0.012659446336328983, + 0.0060231066308915615, + 0.008746659383177757, + 0.007651248946785927, + 0.008510840125381947, + 0.023849142715334892, + 0.010787109844386578, + 0.005479581654071808, + 0.007585570216178894, + 0.013950144872069359, + 0.023987270891666412, + 0.012494880706071854, + 0.011463098227977753, + 0.02255001664161682, + 0.019480407238006592, + 0.024799102917313576, + 0.013697492890059948, + 0.013670827262103558, + 0.010633761994540691, + 0.007974828593432903, + 0.012242991477251053, + 0.014455379918217659, + 0.015497246757149696, + 0.023618128150701523, + 0.01600657030940056, + 0.00940559059381485, + 0.005372374784201384, + 0.005676588509231806, + 0.0034829131327569485, + 0.0032905873376876116, + 0.006406901404261589, + 0.003412158228456974, + 0.0059320926666259766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0011804286623373628 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581606775522232 + ], + "opposite_logprob_tokens": [ + -3.8955817222595215 + ], + "logprob_reward": -0.020581606775522232, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581606775522232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20588526129722595, + 0.19643135368824005, + 0.12475217878818512, + 0.0781312957406044, + 0.05460793897509575, + 0.08793390542268753, + 0.06992534548044205, + 0.10578452795743942, + 0.05114244669675827, + 0.025405727326869965 + ], + "credit_map": [ + 0.02053902857005596, + 0.01909475214779377, + 0.019617445766925812, + 0.014229141175746918, + 0.023400690406560898, + 0.0320899672806263, + 0.02680247090756893, + 0.05011177062988281, + 0.0440240278840065, + 0.02450704574584961, + 0.023003948852419853, + 0.009186867624521255, + 0.028400694951415062, + 0.02409329079091549, + 0.010968504473567009, + 0.03224698454141617, + 0.01487341895699501, + 0.017384056001901627, + 0.015114450827240944, + 0.012162801809608936, + 0.013386421836912632, + 0.014718887396156788, + 0.022716717794537544, + 0.014395431615412235, + 0.007888146676123142, + 0.0072648413479328156, + 0.01290622353553772, + 0.0061148954555392265, + 0.00989581923931837, + 0.0132277961820364, + 0.010977187193930149, + 0.009856391698122025, + 0.0028784212190657854, + 0.006183918099850416, + 0.006972645875066519, + 0.010460421442985535, + 0.00822776835411787, + 0.007348442915827036, + 0.006648855749517679, + 0.005887467414140701, + 0.007466057315468788, + 0.014834955334663391, + 0.013693457469344139, + 0.008375992067158222, + 0.012064781039953232, + 0.012855149805545807, + 0.006213899236172438, + 0.012429618276655674, + 0.008234817534685135, + 0.0073917461559176445, + 0.008530988357961178, + 0.009007884189486504, + 0.0117894746363163, + 0.008846997283399105, + 0.00844637956470251, + 0.007677058689296246, + 0.010797669179737568, + 0.015045681037008762, + 0.010266470722854137, + 0.009263194166123867, + 0.018005870282649994, + 0.013933426700532436, + 0.016743943095207214, + 0.011728275567293167, + 0.006413335911929607, + 0.003929477650672197, + 0.003065986093133688, + 0.004457512870430946, + 0.006949849892407656, + 0.00754739111289382, + 0.011521715670824051, + 0.007257176097482443, + 0.004284510854631662, + 0.0030687889084219933, + 0.0024213879369199276, + 0.002883167704567313, + 0.0028299700934439898, + 0.0036320032086223364, + 0.0031573607120662928, + 0.003128539305180311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.020581606775522232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle while tracking the drone?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the drone sleek, aerodynamic, and covered in a matte black carbon fiber shell?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a narrow, steep-sided red sandstone canyon with jagged rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the drone bank to the right with its left wing skimming the rock face and kicking up a cloud of sand and dust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a sharp blue LED indicator visible on the front of the drone?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038083530962467194 + ], + "opposite_logprob_tokens": [ + -3.288083553314209 + ], + "logprob_reward": -0.038083530962467194, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038083530962467194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06498266011476517, + 0.029687078669667244, + 0.04499025642871857, + 0.11735234409570694, + 0.09924746304750443, + 0.07898782193660736, + 0.35438981652259827, + 0.12941598892211914, + 0.054960303008556366, + 0.025986289605498314 + ], + "credit_map": [ + 0.007086554076522589, + 0.006669274065643549, + 0.007489664945751429, + 0.005837154574692249, + 0.0070712389424443245, + 0.009742090478539467, + 0.011230642907321453, + 0.009856041520833969, + 0.006472454871982336, + 0.0034653397742658854, + 0.0036669133696705103, + 0.0023003402166068554, + 0.004752703011035919, + 0.0033286651596426964, + 0.002112631918862462, + 0.0035880308132618666, + 0.004358870442956686, + 0.007962930016219616, + 0.005750918760895729, + 0.00510170916095376, + 0.0067421430721879005, + 0.004944939166307449, + 0.005238029174506664, + 0.004890714306384325, + 0.011090870946645737, + 0.00878322683274746, + 0.02027212269604206, + 0.009663336910307407, + 0.015277271158993244, + 0.01981949619948864, + 0.01840617135167122, + 0.014039849862456322, + 0.01441906951367855, + 0.014647463336586952, + 0.013911191374063492, + 0.010724317282438278, + 0.014899850822985172, + 0.012322481721639633, + 0.0078748669475317, + 0.010448228567838669, + 0.0062308828346431255, + 0.013953554444015026, + 0.007950044237077236, + 0.012258999980986118, + 0.009633313864469528, + 0.01280538272112608, + 0.006466031540185213, + 0.009689614176750183, + 0.022328846156597137, + 0.04040217399597168, + 0.045018747448921204, + 0.059649232774972916, + 0.08549030125141144, + 0.041384048759937286, + 0.026803964748978615, + 0.033312514424324036, + 0.013314133509993553, + 0.02099241130053997, + 0.013094999827444553, + 0.010531624779105186, + 0.020571207627654076, + 0.01806787960231304, + 0.020664529874920845, + 0.012179199606180191, + 0.005020568147301674, + 0.007669225800782442, + 0.005408021155744791, + 0.008044659160077572, + 0.00900919921696186, + 0.005273059941828251, + 0.007821434177458286, + 0.0067141358740627766, + 0.0033297480549663305, + 0.0032128694001585245, + 0.0028065512888133526, + 0.0027144686318933964, + 0.004689953289926052, + 0.003021811367943883, + 0.0034539923071861267, + 0.002756895264610648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 3 + ], + "selected_reward": -0.038083530962467194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_tensor([7957, 7957])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_tensor([7957, 7957])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0fecbaf95a00d2b1d5b284c2accfd1fc4ce56961 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_tensor([7957, 7957])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up begins a slow, graceful orbiting panorama, capturing a magically romantic moment. The camera focuses on the hands of a man and a woman as they slowly and hesitantly reach for each other on a wooden plank winding through an enchanted forest at night. The entire scene is completely illuminated by the gentle, pulsating light of thousands of flickering fireflies, which float lazily in the air like living stars. These tiny floating lanterns cast a warm golden glow, creating a dynamic effect of soft highlights and shifting shadows on skin and grainy wood. The woman's hand is slender and delicate, while the man's is slightly larger and more defined. As the camera continues to gently rotate, creating the feeling that the world is gently spinning around their intimate connection, their fingertips finally make the slightest, most delicate first contact. This single point of connection, held for one beat, achieves a playful and tender anticipation, while the blurred background of the dense forest silhouette and countless distant fireflies creates a profound sense of depth and a deeply romantic, magical atmosphere.", + "seed": 26757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26757_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10347466189414263, + "vlm_logprob": -0.10347466189414263, + "vlm_yes_no_margin": 2.7750000238418577, + "vlm_token_credit": -0.10347466189414263, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07178328931331635 + ], + "opposite_logprob_tokens": [ + -2.8217833042144775 + ], + "logprob_reward": -0.07178328931331635, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07178328931331635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11199082434177399, + 0.13565199077129364, + 0.0448882170021534, + 0.022406762465834618, + 0.045748285949230194, + 0.31076619029045105, + 0.07032347470521927, + 0.16793617606163025, + 0.04838058352470398, + 0.041907526552677155 + ], + "credit_map": [ + 0.0071456111036241055, + 0.01897481270134449, + 0.009304588660597801, + 0.0075843846425414085, + 0.026588452979922295, + 0.014968630857765675, + 0.016343854367733, + 0.011080493219196796, + 0.005749622359871864, + 0.003738927422091365, + 0.003094295971095562, + 0.004244057461619377, + 0.04952070116996765, + 0.01114420872181654, + 0.013286648318171501, + 0.04487352445721626, + 0.004388816654682159, + 0.003189413109794259, + 0.002827725373208523, + 0.0027336026541888714, + 0.004627682268619537, + 0.011892927810549736, + 0.005681431386619806, + 0.009546617045998573, + 0.0014804317615926266, + 0.0018616910092532635, + 0.001279205665923655, + 0.0018510128138586879, + 0.0057025495916605, + 0.00469608511775732, + 0.0025411115493625402, + 0.0029946747235953808, + 0.006614229176193476, + 0.008604446426033974, + 0.004858438391238451, + 0.0024045384488999844, + 0.004654625430703163, + 0.0074172127060592175, + 0.004704515915364027, + 0.00649027805775404, + 0.07028967887163162, + 0.119351826608181, + 0.047847095876932144, + 0.010323104448616505, + 0.019892122596502304, + 0.015021804720163345, + 0.013845725916326046, + 0.014194808900356293, + 0.010801027528941631, + 0.011150415055453777, + 0.008174161426723003, + 0.007912632077932358, + 0.011696002446115017, + 0.006936199497431517, + 0.006610248237848282, + 0.007042787503451109, + 0.020389460027217865, + 0.03957546502351761, + 0.021401315927505493, + 0.011781216599047184, + 0.03269651532173157, + 0.02094227448105812, + 0.010071479715406895, + 0.01107846386730671, + 0.003602064913138747, + 0.0036674051079899073, + 0.014607997611165047, + 0.017067287117242813, + 0.002430184744298458, + 0.002210464794188738, + 0.0021626679226756096, + 0.0026325115468353033, + 0.005741689819842577, + 0.0034088476095348597, + 0.004982439335435629, + 0.012134097516536713, + 0.005983813665807247, + 0.0024650576524436474, + 0.004352507181465626, + 0.0028390735387802124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.07178328931331635 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428188532590866 + ], + "opposite_logprob_tokens": [ + -2.0178189277648926 + ], + "logprob_reward": -0.1428188532590866, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.1428188532590866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04932814836502075, + 0.2046874612569809, + 0.026623474434018135, + 0.013789411634206772, + 0.04602475464344025, + 0.4295370280742645, + 0.019137635827064514, + 0.11300373822450638, + 0.04771028086543083, + 0.05015803128480911 + ], + "credit_map": [ + 0.0031569937709718943, + 0.005504864268004894, + 0.003375802654772997, + 0.00341629097238183, + 0.013948371633887291, + 0.007734130136668682, + 0.006971615366637707, + 0.005220077931880951, + 0.0076920329593122005, + 0.004529663827270269, + 0.00425270851701498, + 0.0050599887035787106, + 0.07825843244791031, + 0.016425814479589462, + 0.015161529183387756, + 0.07330727577209473, + 0.002767214085906744, + 0.0025321717839688063, + 0.00258530187420547, + 0.0020307113882154226, + 0.003052128478884697, + 0.0058100344613194466, + 0.002796997781842947, + 0.005048916209489107, + 0.0017423907993361354, + 0.001945264171808958, + 0.0012973584234714508, + 0.0016687483293935657, + 0.0021958802826702595, + 0.002015920588746667, + 0.0012876751134172082, + 0.0016361738089472055, + 0.006469362881034613, + 0.009622614830732346, + 0.00573769211769104, + 0.0030765929259359837, + 0.005418332759290934, + 0.005490608047693968, + 0.00601185392588377, + 0.0041976976208388805, + 0.1117769107222557, + 0.17043760418891907, + 0.05983791500329971, + 0.01157931424677372, + 0.02697032131254673, + 0.016328994184732437, + 0.01588292606174946, + 0.016723060980439186, + 0.002628323622047901, + 0.00307041616179049, + 0.003137770341709256, + 0.0018698704661801457, + 0.003166138892993331, + 0.0015631093410775065, + 0.001967101590707898, + 0.0017349066911265254, + 0.016467202454805374, + 0.02483043074607849, + 0.012912891805171967, + 0.007635869085788727, + 0.02521243505179882, + 0.010266846045851707, + 0.007314709946513176, + 0.00836334191262722, + 0.006192192900925875, + 0.008961034007370472, + 0.007295586634427309, + 0.009886304847896099, + 0.00631058681756258, + 0.0029425027314573526, + 0.0023078371305018663, + 0.0038142346311360598, + 0.007509646471589804, + 0.005312355235219002, + 0.005559864919632673, + 0.011791834607720375, + 0.005605753511190414, + 0.006876584142446518, + 0.004644380416721106, + 0.0028576115146279335 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.1428188532590866 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632971741259098 + ], + "opposite_logprob_tokens": [ + -4.882633209228516 + ], + "logprob_reward": -0.007632971741259098, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632971741259098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01873444952070713, + 0.5297191143035889, + 0.04260466247797012, + 0.022238286212086678, + 0.015675146132707596, + 0.171042799949646, + 0.05950501188635826, + 0.07773137092590332, + 0.04271399974822998, + 0.020035117864608765 + ], + "credit_map": [ + 0.0014165053144097328, + 0.0036349103320389986, + 0.001353036379441619, + 0.0022355099208652973, + 0.0035122595727443695, + 0.0020903905387967825, + 0.002560787368565798, + 0.0019310496281832457, + 0.017213942483067513, + 0.012178488075733185, + 0.011635310016572475, + 0.012399018742144108, + 0.20282092690467834, + 0.043875377625226974, + 0.038971710950136185, + 0.19062440097332, + 0.00254084006883204, + 0.002768335398286581, + 0.0026706657372415066, + 0.002729055704548955, + 0.006251371931284666, + 0.01143649686127901, + 0.004020363092422485, + 0.010187538340687752, + 0.0020984297152608633, + 0.002344070468097925, + 0.001677389140240848, + 0.001961895264685154, + 0.0036837642546743155, + 0.004486704710870981, + 0.003137249732390046, + 0.0028487832751125097, + 0.004188356455415487, + 0.0020955256186425686, + 0.001561292097903788, + 0.0008733149734325707, + 0.001082044211216271, + 0.003843147773295641, + 0.0008345141541212797, + 0.0011969516053795815, + 0.04222418740391731, + 0.06778603047132492, + 0.025614464655518532, + 0.004554258659482002, + 0.011331554502248764, + 0.006525466684252024, + 0.006330597214400768, + 0.00667625293135643, + 0.008155276998877525, + 0.009088580496609211, + 0.008174103684723377, + 0.006576156709343195, + 0.009836277924478054, + 0.006244645919650793, + 0.00609221076592803, + 0.005337758455425501, + 0.00864941906183958, + 0.01859109289944172, + 0.00980387069284916, + 0.004688770975917578, + 0.016322221606969833, + 0.010235474444925785, + 0.004206350073218346, + 0.005234175827354193, + 0.0035561569966375828, + 0.004843484610319138, + 0.011316949501633644, + 0.01413487084209919, + 0.0028265337459743023, + 0.0025761795695871115, + 0.0015347746666520834, + 0.0019250506302341819, + 0.0029695003759115934, + 0.0019304013112559915, + 0.0020011623855680227, + 0.005047377198934555, + 0.0025577801279723644, + 0.0023697796277701855, + 0.001969774952158332, + 0.0011893447954207659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 7 + ], + "selected_reward": -0.007632971741259098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520826756954193 + ], + "opposite_logprob_tokens": [ + -1.5020827054977417 + ], + "logprob_reward": -0.2520826756954193, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520826756954193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11473733931779861, + 0.2534318268299103, + 0.03299029916524887, + 0.025288650766015053, + 0.06534093618392944, + 0.21206986904144287, + 0.08532153815031052, + 0.0881551206111908, + 0.04526497796177864, + 0.07739941775798798 + ], + "credit_map": [ + 0.007644868455827236, + 0.024164902046322823, + 0.006116747856140137, + 0.009686538018286228, + 0.0263579823076725, + 0.012993527576327324, + 0.016151513904333115, + 0.011621265672147274, + 0.008684319444000721, + 0.014126504771411419, + 0.012320631183683872, + 0.019708236679434776, + 0.07870037108659744, + 0.019784709438681602, + 0.023109279572963715, + 0.0769977942109108, + 0.002592365490272641, + 0.003244617488235235, + 0.001416564453393221, + 0.0031587276607751846, + 0.0041557964868843555, + 0.007226384244859219, + 0.00362783158197999, + 0.007568014785647392, + 0.0028038646560162306, + 0.003479784121736884, + 0.0018404865404590964, + 0.0026705760974437, + 0.004032691475003958, + 0.0034324233420193195, + 0.003926742821931839, + 0.0031020825263112783, + 0.015430099330842495, + 0.006799148861318827, + 0.005292576737701893, + 0.003021493088454008, + 0.0052019041031599045, + 0.016691960394382477, + 0.00642538582906127, + 0.006478363182395697, + 0.031950220465660095, + 0.07909800857305527, + 0.045131269842386246, + 0.006617473438382149, + 0.019582267850637436, + 0.010028800927102566, + 0.009510040283203125, + 0.010151791386306286, + 0.010960670188069344, + 0.012369656004011631, + 0.010140306316316128, + 0.010344487614929676, + 0.01563986949622631, + 0.008783305995166302, + 0.008537973277270794, + 0.008545268326997757, + 0.01062728464603424, + 0.02076045237481594, + 0.010264315642416477, + 0.007185721769928932, + 0.01821216382086277, + 0.009494046680629253, + 0.005940839182585478, + 0.00567029369994998, + 0.006305278278887272, + 0.008539604023098946, + 0.007021612953394651, + 0.006208392791450024, + 0.004150966182351112, + 0.0031526105012744665, + 0.0049947877414524555, + 0.004891726188361645, + 0.010573229752480984, + 0.007607061415910721, + 0.008421788923442364, + 0.01896914839744568, + 0.008903496898710728, + 0.010454719886183739, + 0.007425620220601559, + 0.005044350866228342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.2520826756954193 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043055519461631775 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.043055519461631775, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043055519461631775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044844314455986, + 0.4001895785331726, + 0.056300025433301926, + 0.03776291385293007, + 0.027804726734757423, + 0.11802603304386139, + 0.03524857386946678, + 0.048749543726444244, + 0.04031943902373314, + 0.13111470639705658 + ], + "credit_map": [ + 0.007148698903620243, + 0.014007608406245708, + 0.005775739438831806, + 0.00960377138108015, + 0.02992681972682476, + 0.012615897692739964, + 0.015347621403634548, + 0.010058272629976273, + 0.012565689161419868, + 0.009600594639778137, + 0.0073027294129133224, + 0.010941267013549805, + 0.15512028336524963, + 0.030385268852114677, + 0.030708445236086845, + 0.143565371632576, + 0.00435891468077898, + 0.003503205021843314, + 0.0034116378519684076, + 0.004023249726742506, + 0.0069178370758891106, + 0.015350925736129284, + 0.008250666782259941, + 0.010483591817319393, + 0.0030237806495279074, + 0.00346045452170074, + 0.002762140706181526, + 0.0030365888960659504, + 0.005833873525261879, + 0.009823243133723736, + 0.003590324195101857, + 0.006232511717826128, + 0.0031223781406879425, + 0.0036214643623679876, + 0.0030367157887667418, + 0.00228653522208333, + 0.002202251460403204, + 0.004362138453871012, + 0.002687366446480155, + 0.006485878489911556, + 0.01404313463717699, + 0.02482418343424797, + 0.010512683540582657, + 0.00858575664460659, + 0.012540148571133614, + 0.010379736311733723, + 0.015181261114776134, + 0.02195914275944233, + 0.004982533399015665, + 0.004866382107138634, + 0.003851894987747073, + 0.0032227332703769207, + 0.006129424553364515, + 0.004655381198972464, + 0.0033291741274297237, + 0.004211051389575005, + 0.006760075222700834, + 0.010981476865708828, + 0.0055120112374424934, + 0.003572971560060978, + 0.010909340344369411, + 0.004247073084115982, + 0.003276954870671034, + 0.003489647526293993, + 0.005021955352276564, + 0.008501584641635418, + 0.006790946237742901, + 0.00466357683762908, + 0.004070653580129147, + 0.003619867842644453, + 0.0024933884851634502, + 0.005157470237463713, + 0.020959168672561646, + 0.012646226212382317, + 0.014188776724040508, + 0.034434691071510315, + 0.016076991334557533, + 0.014429162256419659, + 0.011107555590569973, + 0.007272160612046719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.043055519461631775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, graceful orbiting panorama, capturing a magically romantic moment. The camera focuses on the hands of a man and a woman as they slowly and hesitantly reach for each other on a wooden plank winding through an enchanted forest at night. The entire scene is completely illuminated by the gentle, pulsating light of thousands of flickering fireflies, which float lazily in the air like living stars. These tiny floating lanterns cast a warm golden glow, creating a dynamic effect of soft highlights and shifting shadows on skin and grainy wood. The woman's hand is slender and delicate, while the man's is slightly larger and more defined. As the camera continues to gently rotate, creating the feeling that the world is gently spinning around their intimate connection, their fingertips finally make the slightest, most delicate first contact. This single point of connection, held for one beat, achieves a playful and tender anticipation, while the blurred background of the dense forest silhouette and countless distant fireflies creates a profound sense of depth and a deeply romantic, magical atmosphere.", + "seed": 26758, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26758_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07041884418576956, + "vlm_logprob": -0.07041884418576956, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.07041884418576956, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05600761994719505 + ], + "opposite_logprob_tokens": [ + -3.6810076236724854 + ], + "logprob_reward": -0.05600761994719505, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.05600761994719505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14034736156463623, + 0.06444232910871506, + 0.22284866869449615, + 0.033753059804439545, + 0.06495169550180435, + 0.11022200435400009, + 0.1763574331998825, + 0.0424845926463604, + 0.06267561763525009, + 0.08191721886396408 + ], + "credit_map": [ + 0.009376989677548409, + 0.007444567047059536, + 0.02289685048162937, + 0.024666227400302887, + 0.020413709804415703, + 0.02333429828286171, + 0.01953052170574665, + 0.012684193439781666, + 0.006549175828695297, + 0.005555660929530859, + 0.007446660194545984, + 0.013111524283885956, + 0.005325511563569307, + 0.006214491557329893, + 0.013337135314941406, + 0.006902169901877642, + 0.03127053380012512, + 0.020302945747971535, + 0.05390004813671112, + 0.031953465193510056, + 0.014626124873757362, + 0.021236028522253036, + 0.02449483796954155, + 0.025064673274755478, + 0.004751698579639196, + 0.005157009698450565, + 0.0036828897427767515, + 0.0034058529417961836, + 0.004161472897976637, + 0.004659179598093033, + 0.004004457499831915, + 0.003930496517568827, + 0.011118374764919281, + 0.006869752891361713, + 0.011042994447052479, + 0.007871093228459358, + 0.006205160636454821, + 0.007329375948756933, + 0.008647818118333817, + 0.005867125000804663, + 0.009688086807727814, + 0.014128375798463821, + 0.010597523301839828, + 0.015857107937335968, + 0.01419393066316843, + 0.01672685518860817, + 0.014243588782846928, + 0.014786534011363983, + 0.018477408215403557, + 0.028890805318951607, + 0.013662016950547695, + 0.021220486611127853, + 0.040195759385824203, + 0.018791833892464638, + 0.020341364666819572, + 0.014777762815356255, + 0.002853832207620144, + 0.007760328706353903, + 0.011185386218130589, + 0.005277955438941717, + 0.004256824497133493, + 0.004098588600754738, + 0.004285446833819151, + 0.0027662308420985937, + 0.005880129989236593, + 0.008826174773275852, + 0.010486414656043053, + 0.005579336080700159, + 0.004619111772626638, + 0.012770877219736576, + 0.006710350047796965, + 0.0078032249584794044, + 0.0078158313408494, + 0.011840365827083588, + 0.008065827190876007, + 0.009814271703362465, + 0.012549266219139099, + 0.007765994872897863, + 0.013341843150556087, + 0.01072381529957056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 0 + ], + "selected_reward": -0.05600761994719505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866965487599373 + ], + "opposite_logprob_tokens": [ + -3.0486695766448975 + ], + "logprob_reward": -0.04866965487599373, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866965487599373, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09510716050863266, + 0.29972508549690247, + 0.06199250742793083, + 0.039128292351961136, + 0.05313751846551895, + 0.2192428559064865, + 0.06690144538879395, + 0.08215254545211792, + 0.047567520290613174, + 0.035045020282268524 + ], + "credit_map": [ + 0.00886593572795391, + 0.01286016870290041, + 0.010629845783114433, + 0.016231708228588104, + 0.009943311102688313, + 0.015668872743844986, + 0.010706842876970768, + 0.010200479999184608, + 0.04220150038599968, + 0.07895171642303467, + 0.028869815170764923, + 0.03917164355516434, + 0.04616423323750496, + 0.02112705260515213, + 0.02112886682152748, + 0.02211025543510914, + 0.007353303488343954, + 0.0049248612485826015, + 0.01670285128057003, + 0.008360172621905804, + 0.007079749833792448, + 0.0039965370669960976, + 0.008314666338264942, + 0.00526035949587822, + 0.0034982217475771904, + 0.0070171537809073925, + 0.005568417254835367, + 0.0036521591246128082, + 0.0035754311829805374, + 0.005388348363339901, + 0.005484654568135738, + 0.004943904001265764, + 0.009777016937732697, + 0.007350076921284199, + 0.00918112974613905, + 0.005018595140427351, + 0.004419301170855761, + 0.005997705273330212, + 0.006134076975286007, + 0.005259612109512091, + 0.026481620967388153, + 0.023732410743832588, + 0.029075300320982933, + 0.03794575482606888, + 0.02635161578655243, + 0.034928325563669205, + 0.01896798610687256, + 0.021759845316410065, + 0.007063659373670816, + 0.012646852061152458, + 0.006549212150275707, + 0.005885492078959942, + 0.015246004797518253, + 0.007044205907732248, + 0.00735509255900979, + 0.005110921338200569, + 0.005005211569368839, + 0.008999434299767017, + 0.023769397288560867, + 0.0076435948722064495, + 0.01243138313293457, + 0.009735271334648132, + 0.007295586168766022, + 0.007272670045495033, + 0.004640699364244938, + 0.0063131097704172134, + 0.007997551001608372, + 0.005120050627738237, + 0.004529818892478943, + 0.008146288804709911, + 0.005433516576886177, + 0.0053864833898842335, + 0.0038004592061042786, + 0.004090569447726011, + 0.004608131945133209, + 0.00526581984013319, + 0.0037468697410076857, + 0.003651201957836747, + 0.004307137802243233, + 0.005574830807745457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.04866965487599373 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028224652633070946 + ], + "opposite_logprob_tokens": [ + -5.877822399139404 + ], + "logprob_reward": -0.0028224652633070946, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028224652633070946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28450116515159607, + 0.18611659109592438, + 0.11090276390314102, + 0.05668536573648453, + 0.0686846449971199, + 0.08862517774105072, + 0.061726830899715424, + 0.06873741000890732, + 0.03632678836584091, + 0.03769326210021973 + ], + "credit_map": [ + 0.019088340923190117, + 0.020613903179764748, + 0.05108695104718208, + 0.05509048327803612, + 0.035243548452854156, + 0.04580087959766388, + 0.03195788711309433, + 0.025619154796004295, + 0.02089051716029644, + 0.027748698368668556, + 0.024759260937571526, + 0.02967415191233158, + 0.02644750475883484, + 0.013410719111561775, + 0.015250569209456444, + 0.027935171499848366, + 0.012953821569681168, + 0.009979253634810448, + 0.024314913898706436, + 0.015045495703816414, + 0.009895933791995049, + 0.011079373769462109, + 0.011651929467916489, + 0.015982044860720634, + 0.005421566776931286, + 0.009590351954102516, + 0.007046331185847521, + 0.005562251899391413, + 0.007431222591549158, + 0.005706291180104017, + 0.00751537038013339, + 0.008411974646151066, + 0.017649581655859947, + 0.010171676054596901, + 0.01152205653488636, + 0.00402090884745121, + 0.005889405496418476, + 0.005577839910984039, + 0.006571438163518906, + 0.007281737867742777, + 0.008875695988535881, + 0.010442918166518211, + 0.012602887116372585, + 0.019135059788823128, + 0.011669490486383438, + 0.006779758725315332, + 0.009643818251788616, + 0.009475554339587688, + 0.006328624673187733, + 0.011251757852733135, + 0.006459556985646486, + 0.005910768173635006, + 0.01227083895355463, + 0.006558709312230349, + 0.007026406470686197, + 0.005920169409364462, + 0.004554918501526117, + 0.008615399710834026, + 0.01733902283012867, + 0.00719862524420023, + 0.007812672294676304, + 0.010339675471186638, + 0.006662238854914904, + 0.006214860361069441, + 0.0049062143079936504, + 0.006214064545929432, + 0.004887890536338091, + 0.004559254739433527, + 0.0034950023982673883, + 0.004415594041347504, + 0.0037055015563964844, + 0.004143266472965479, + 0.0037846872583031654, + 0.0049189613200724125, + 0.004338925238698721, + 0.004015617072582245, + 0.004063270054757595, + 0.004069375805556774, + 0.0065314993262290955, + 0.0059709264896810055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0028224652633070946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20154090225696564 + ], + "opposite_logprob_tokens": [ + -1.7015409469604492 + ], + "logprob_reward": -0.20154090225696564, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20154090225696564, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23819179832935333, + 0.06286901235580444, + 0.19315649569034576, + 0.08084002882242203, + 0.06190122291445732, + 0.07178071141242981, + 0.10725302994251251, + 0.0694378986954689, + 0.05485201254487038, + 0.05971788987517357 + ], + "credit_map": [ + 0.018232880160212517, + 0.020410915836691856, + 0.03556356206536293, + 0.03135067597031593, + 0.032286111265420914, + 0.047765862196683884, + 0.03143700957298279, + 0.021144766360521317, + 0.006137097254395485, + 0.006643874570727348, + 0.007763087749481201, + 0.014715380035340786, + 0.004677897784858942, + 0.005600019358098507, + 0.00831982959061861, + 0.009011822752654552, + 0.027362320572137833, + 0.017714476212859154, + 0.049963753670454025, + 0.02783852443099022, + 0.012888370081782341, + 0.01715119741857052, + 0.020081156864762306, + 0.02015666291117668, + 0.008458985015749931, + 0.013313409872353077, + 0.007981007918715477, + 0.007431913632899523, + 0.009006760083138943, + 0.012644713744521141, + 0.01240126695483923, + 0.009601965546607971, + 0.007042884360998869, + 0.0069939494132995605, + 0.007915805093944073, + 0.007573016919195652, + 0.007965332828462124, + 0.008406971581280231, + 0.008604167029261589, + 0.0073990891687572, + 0.010364034213125706, + 0.009440668858587742, + 0.006111606955528259, + 0.006425601430237293, + 0.009883199818432331, + 0.011632926762104034, + 0.008990813978016376, + 0.00893185380846262, + 0.007983097806572914, + 0.012161987833678722, + 0.01453248132020235, + 0.022268516942858696, + 0.014368915930390358, + 0.01176617294549942, + 0.011965857818722725, + 0.012205983512103558, + 0.004241380374878645, + 0.010143758729100227, + 0.020115938037633896, + 0.009531102143228054, + 0.007970022968947887, + 0.0064248149283230305, + 0.006979245226830244, + 0.0040316274389624596, + 0.0055258609354496, + 0.008075838908553123, + 0.008072304539382458, + 0.005426920019090176, + 0.008155942894518375, + 0.010799959301948547, + 0.004657829646021128, + 0.004137353505939245, + 0.005765789654105902, + 0.008181664161384106, + 0.00756289716809988, + 0.005829598754644394, + 0.00806322693824768, + 0.005641371943056583, + 0.01012518908828497, + 0.008548147976398468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.20154090225696564 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043053578585386276 + ], + "opposite_logprob_tokens": [ + -3.16805362701416 + ], + "logprob_reward": -0.043053578585386276, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043053578585386276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07750905305147171, + 0.11193609982728958, + 0.1443258374929428, + 0.041873298585414886, + 0.11789635568857193, + 0.17530888319015503, + 0.056518297642469406, + 0.037978921085596085, + 0.04238058254122734, + 0.19427263736724854 + ], + "credit_map": [ + 0.007471229415386915, + 0.011609810404479504, + 0.009748703800141811, + 0.012257132679224014, + 0.008239573799073696, + 0.00985465943813324, + 0.007877705618739128, + 0.010450242087244987, + 0.013074343092739582, + 0.02747996151447296, + 0.011129993014037609, + 0.01466407347470522, + 0.015112658962607384, + 0.009724312461912632, + 0.011737780645489693, + 0.009012976661324501, + 0.018640276044607162, + 0.011817691847682, + 0.04201492294669151, + 0.018278509378433228, + 0.01138992514461279, + 0.013992671854794025, + 0.013452164828777313, + 0.014739684760570526, + 0.004250512458384037, + 0.005135712679475546, + 0.005744137801229954, + 0.005046139471232891, + 0.004703885409981012, + 0.00527530862018466, + 0.005970336962491274, + 0.005747262854129076, + 0.01430873479694128, + 0.010265186429023743, + 0.012507054023444653, + 0.017535751685500145, + 0.012285341508686543, + 0.016611838713288307, + 0.021843452006578445, + 0.012538997456431389, + 0.01611146330833435, + 0.019738050177693367, + 0.023155679926276207, + 0.018594784662127495, + 0.022625505924224854, + 0.036518752574920654, + 0.0174508448690176, + 0.02111380361020565, + 0.005598030518740416, + 0.009066291153430939, + 0.006540681701153517, + 0.006353178061544895, + 0.008322853595018387, + 0.006907118018716574, + 0.006930846255272627, + 0.006799300201237202, + 0.0024436849635094404, + 0.00659608980640769, + 0.01059215422719717, + 0.004614528734236956, + 0.003204162698239088, + 0.0038818069733679295, + 0.0039265453815460205, + 0.0027199469041079283, + 0.004418938420712948, + 0.005225136410444975, + 0.00541636161506176, + 0.006546498276293278, + 0.003672344144433737, + 0.0048163230530917645, + 0.005862420424818993, + 0.006422561127692461, + 0.018810683861374855, + 0.02435186691582203, + 0.021224280819296837, + 0.022426268085837364, + 0.023144012317061424, + 0.022345416247844696, + 0.030212974175810814, + 0.03175714612007141 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.043053578585386276 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, graceful orbiting panorama, capturing a magically romantic moment. The camera focuses on the hands of a man and a woman as they slowly and hesitantly reach for each other on a wooden plank winding through an enchanted forest at night. The entire scene is completely illuminated by the gentle, pulsating light of thousands of flickering fireflies, which float lazily in the air like living stars. These tiny floating lanterns cast a warm golden glow, creating a dynamic effect of soft highlights and shifting shadows on skin and grainy wood. The woman's hand is slender and delicate, while the man's is slightly larger and more defined. As the camera continues to gently rotate, creating the feeling that the world is gently spinning around their intimate connection, their fingertips finally make the slightest, most delicate first contact. This single point of connection, held for one beat, achieves a playful and tender anticipation, while the blurred background of the dense forest silhouette and countless distant fireflies creates a profound sense of depth and a deeply romantic, magical atmosphere.", + "seed": 26759, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26759_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0851573957130313, + "vlm_logprob": -0.0851573957130313, + "vlm_yes_no_margin": 2.949999976158142, + "vlm_token_credit": -0.0851573957130313, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13648493587970734 + ], + "opposite_logprob_tokens": [ + -2.1364848613739014 + ], + "logprob_reward": -0.13648493587970734, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.13648493587970734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727023869752884, + 0.044009774923324585, + 0.10216912627220154, + 0.054691992700099945, + 0.07873059064149857, + 0.11670887470245361, + 0.05820699408650398, + 0.03442376106977463, + 0.03774121403694153, + 0.3006152808666229 + ], + "credit_map": [ + 0.007852070964872837, + 0.00915762223303318, + 0.013653609901666641, + 0.011951427906751633, + 0.007287045009434223, + 0.018234260380268097, + 0.05798257514834404, + 0.04658377170562744, + 0.003082325914874673, + 0.0029189763590693474, + 0.004445602186024189, + 0.007439971435815096, + 0.0070952982641756535, + 0.003668770659714937, + 0.008121288381516933, + 0.007237538695335388, + 0.006868890952318907, + 0.00902608223259449, + 0.011045996099710464, + 0.008149281144142151, + 0.010858409106731415, + 0.007085063494741917, + 0.026408566161990166, + 0.022726833820343018, + 0.00226870970800519, + 0.0031914906576275826, + 0.005703908856958151, + 0.0037082049529999495, + 0.003911938983947039, + 0.003597032278776169, + 0.017360566183924675, + 0.014950139448046684, + 0.01107183750718832, + 0.0076405941508710384, + 0.008487348444759846, + 0.005840558558702469, + 0.007856384851038456, + 0.010178384371101856, + 0.017790384590625763, + 0.009865093976259232, + 0.02378990687429905, + 0.020561812445521355, + 0.012357569299638271, + 0.008856151252985, + 0.008511087857186794, + 0.008761542849242687, + 0.026999730616807938, + 0.006871066987514496, + 0.008601779118180275, + 0.00697994465008378, + 0.010005071759223938, + 0.010454640723764896, + 0.004322749096900225, + 0.005923999007791281, + 0.0059124683029949665, + 0.006006342824548483, + 0.0047030337154865265, + 0.0031665388960391283, + 0.005153000354766846, + 0.004454007372260094, + 0.0032223332673311234, + 0.008062046021223068, + 0.003006212180480361, + 0.0026565908920019865, + 0.0039034902583807707, + 0.004628047347068787, + 0.004484825301915407, + 0.0050902883522212505, + 0.004017250146716833, + 0.006107086315751076, + 0.005637488327920437, + 0.0038727375213056803, + 0.021153319627046585, + 0.009733205661177635, + 0.01103226002305746, + 0.01598723605275154, + 0.03702528402209282, + 0.13435502350330353, + 0.04521865025162697, + 0.026110325008630753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.13648493587970734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905483782291412 + ], + "opposite_logprob_tokens": [ + -2.464054822921753 + ], + "logprob_reward": -0.08905483782291412, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905483782291412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1692085862159729, + 0.03287973627448082, + 0.1129225566983223, + 0.031542956829071045, + 0.03966061398386955, + 0.23201832175254822, + 0.03997749462723732, + 0.03803977742791176, + 0.017314208671450615, + 0.2864357829093933 + ], + "credit_map": [ + 0.009461459703743458, + 0.012604467570781708, + 0.022731972858309746, + 0.01991092972457409, + 0.01946106180548668, + 0.021502861753106117, + 0.03728419169783592, + 0.026251621544361115, + 0.0029830702114850283, + 0.0028464135248214006, + 0.0038652026560157537, + 0.004549799486994743, + 0.0047404286451637745, + 0.003541962942108512, + 0.005876298528164625, + 0.004476557485759258, + 0.00932286772876978, + 0.011922997422516346, + 0.010876444168388844, + 0.011222896166145802, + 0.013325992971658707, + 0.009364545345306396, + 0.026240944862365723, + 0.0206458643078804, + 0.0027382196858525276, + 0.0027529336512088776, + 0.004518427886068821, + 0.004035068675875664, + 0.002289741300046444, + 0.0025076232850551605, + 0.007500565145164728, + 0.005200373474508524, + 0.005537007004022598, + 0.004478381481021643, + 0.005390469916164875, + 0.0027169068343937397, + 0.00358898495323956, + 0.003980232868343592, + 0.009034089744091034, + 0.004934538621455431, + 0.053382016718387604, + 0.030231788754463196, + 0.01895923726260662, + 0.013792626559734344, + 0.013053813017904758, + 0.021310625597834587, + 0.06963858753442764, + 0.011649606749415398, + 0.009115862660109997, + 0.006476554088294506, + 0.0052049593068659306, + 0.004543030634522438, + 0.002186739118769765, + 0.0040596406906843185, + 0.004416319541633129, + 0.003974384628236294, + 0.003982847090810537, + 0.003499649465084076, + 0.005945476237684488, + 0.004186023958027363, + 0.0031655561178922653, + 0.010910860262811184, + 0.003257894655689597, + 0.0030914675444364548, + 0.0019817661959677935, + 0.0010434052674099803, + 0.0018028494669124484, + 0.002351658185943961, + 0.002515851752832532, + 0.002278104890137911, + 0.0026067462749779224, + 0.002733824076130986, + 0.020731836557388306, + 0.010873624123632908, + 0.012057429179549217, + 0.017817920073866844, + 0.033083945512771606, + 0.1239977702498436, + 0.0423586368560791, + 0.025514600798487663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.08905483782291412 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13703079521656036, + 0.06410934031009674, + 0.08592311292886734, + 0.12522949278354645, + 0.052659086883068085, + 0.03544580936431885, + 0.04451151192188263, + 0.033875156193971634, + 0.03382435068488121, + 0.3873913288116455 + ], + "credit_map": [ + 0.006378806196153164, + 0.01102511864155531, + 0.009601349011063576, + 0.009794087149202824, + 0.007574643474072218, + 0.011677732691168785, + 0.04647219926118851, + 0.034506864845752716, + 0.004676950164139271, + 0.006174486596137285, + 0.009972373023629189, + 0.007537661120295525, + 0.007213075645267963, + 0.008022508583962917, + 0.008216342888772488, + 0.01229594461619854, + 0.007637975737452507, + 0.007188891526311636, + 0.007334151305258274, + 0.007500045467168093, + 0.008204282261431217, + 0.006185073405504227, + 0.022472012788057327, + 0.01940068230032921, + 0.005336754489690065, + 0.02082904800772667, + 0.010857737623155117, + 0.010569796897470951, + 0.006230930332094431, + 0.00560857355594635, + 0.04621236398816109, + 0.0195842906832695, + 0.008248946629464626, + 0.005353675223886967, + 0.007663166616111994, + 0.0041251094080507755, + 0.005231162067502737, + 0.0045372359454631805, + 0.010765931569039822, + 0.006733861286193132, + 0.0076118456199765205, + 0.0027994485571980476, + 0.0030608540400862694, + 0.0019455446163192391, + 0.001780940336175263, + 0.004195103421807289, + 0.0120870191603899, + 0.0019650552421808243, + 0.008837404660880566, + 0.00785438995808363, + 0.0062750596553087234, + 0.005113109480589628, + 0.003643463598564267, + 0.004628682509064674, + 0.004578102845698595, + 0.0035812952555716038, + 0.005297267343848944, + 0.0026958389207720757, + 0.004258397966623306, + 0.005002932623028755, + 0.0026552961207926273, + 0.007163768634200096, + 0.0034840998705476522, + 0.003317554946988821, + 0.0022816380951553583, + 0.002318407641723752, + 0.0049029989168047905, + 0.0055014765821397305, + 0.004152266308665276, + 0.00679839076474309, + 0.0035710157826542854, + 0.004298155661672354, + 0.0286552831530571, + 0.011690131388604641, + 0.019181519746780396, + 0.025388848036527634, + 0.04520164430141449, + 0.17074905335903168, + 0.05254107713699341, + 0.03398379683494568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.003195776604115963 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001114636659622 + ], + "opposite_logprob_tokens": [ + -2.6950111389160156 + ], + "logprob_reward": -0.07001114636659622, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001114636659622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1401754766702652, + 0.06677213311195374, + 0.0718708485364914, + 0.11584290117025375, + 0.09294071793556213, + 0.06031811982393265, + 0.07508255541324615, + 0.024629468098282814, + 0.06374216824769974, + 0.2886255979537964 + ], + "credit_map": [ + 0.012247282080352306, + 0.016386833041906357, + 0.018543791025877, + 0.01111669186502695, + 0.0206486564129591, + 0.013958118855953217, + 0.0233987495303154, + 0.023875346407294273, + 0.005422044545412064, + 0.006336078979074955, + 0.009988757781684399, + 0.009715470485389233, + 0.00982179120182991, + 0.006030082702636719, + 0.008597040548920631, + 0.010860865004360676, + 0.006185600999742746, + 0.00735838245600462, + 0.005639348644763231, + 0.005712603684514761, + 0.0073996842838823795, + 0.004174782894551754, + 0.018626485019922256, + 0.01677396334707737, + 0.0049981726333498955, + 0.023601245135068893, + 0.006195007357746363, + 0.008053935132920742, + 0.0053488947451114655, + 0.00553729385137558, + 0.04454690217971802, + 0.017561450600624084, + 0.0170793067663908, + 0.011830943636596203, + 0.016044242307543755, + 0.0075601814314723015, + 0.011280224658548832, + 0.010420230217278004, + 0.010177502408623695, + 0.008548078127205372, + 0.008043908514082432, + 0.008702917024493217, + 0.00884129200130701, + 0.00911423284560442, + 0.005768035538494587, + 0.008362849242985249, + 0.006905328948050737, + 0.004579558968544006, + 0.017760198563337326, + 0.013314984738826752, + 0.007391464896500111, + 0.0076639531180262566, + 0.004978938959538937, + 0.010999016463756561, + 0.007819929160177708, + 0.005154067184776068, + 0.0030654610600322485, + 0.0021580883767455816, + 0.0032440770883113146, + 0.004795434419065714, + 0.0019025136716663837, + 0.0033494967501610518, + 0.0028194142505526543, + 0.0032949838787317276, + 0.005268276669085026, + 0.006598022300750017, + 0.008505592122673988, + 0.008207750506699085, + 0.007120812311768532, + 0.011362438090145588, + 0.008356497623026371, + 0.008322782814502716, + 0.02075236663222313, + 0.009796657599508762, + 0.014658608473837376, + 0.020234838128089905, + 0.03378354012966156, + 0.12592610716819763, + 0.03917491063475609, + 0.024298585951328278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.07001114636659622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704028189182281 + ], + "opposite_logprob_tokens": [ + -2.127040386199951 + ], + "logprob_reward": -0.12704028189182281, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704028189182281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16941514611244202, + 0.04757639020681381, + 0.150176540017128, + 0.10976068675518036, + 0.05941487476229668, + 0.15281134843826294, + 0.1045704185962677, + 0.028659895062446594, + 0.04834657534956932, + 0.12926805019378662 + ], + "credit_map": [ + 0.01834680140018463, + 0.028480494394898415, + 0.03347573056817055, + 0.016149381175637245, + 0.017648834735155106, + 0.013419941067695618, + 0.015419119969010353, + 0.026474855840206146, + 0.003345608478412032, + 0.004290958866477013, + 0.006172510329633951, + 0.006485010962933302, + 0.006231116596609354, + 0.005344943143427372, + 0.008298845030367374, + 0.007407393306493759, + 0.008263194002211094, + 0.010200263001024723, + 0.011943870224058628, + 0.008057321421802044, + 0.01116712111979723, + 0.009231841191649437, + 0.04634523764252663, + 0.04496770724654198, + 0.0055681997910141945, + 0.01712057553231716, + 0.005452289246022701, + 0.004684691317379475, + 0.005287788342684507, + 0.0037258572410792112, + 0.04242991656064987, + 0.025491362437605858, + 0.00907124113291502, + 0.00748656690120697, + 0.007789944298565388, + 0.006538138724863529, + 0.006926830392330885, + 0.007746559102088213, + 0.008137860335409641, + 0.005717737600207329, + 0.03463326022028923, + 0.009260524064302444, + 0.009586063213646412, + 0.009312765672802925, + 0.007688144687563181, + 0.018879452720284462, + 0.05674266442656517, + 0.0067084734328091145, + 0.023323945701122284, + 0.01834321767091751, + 0.011381687596440315, + 0.009632944129407406, + 0.006713042035698891, + 0.012743303552269936, + 0.011584186926484108, + 0.01084810122847557, + 0.004181467462331057, + 0.002460109069943428, + 0.00396741135045886, + 0.004976061638444662, + 0.002445919904857874, + 0.0033800459932535887, + 0.00366052333265543, + 0.0035883570089936256, + 0.00462986808270216, + 0.0037703397683799267, + 0.006255782209336758, + 0.006765842437744141, + 0.00652163103222847, + 0.01052913349121809, + 0.005280032753944397, + 0.004593948367983103, + 0.011534400284290314, + 0.009161664173007011, + 0.0169002003967762, + 0.019408486783504486, + 0.012989739887416363, + 0.03181019797921181, + 0.012934965081512928, + 0.014528384432196617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.12704028189182281 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up begins a slow, graceful orbiting panorama, capturing a magically romantic moment. The camera focuses on the hands of a man and a woman as they slowly and hesitantly reach for each other on a wooden plank winding through an enchanted forest at night. The entire scene is completely illuminated by the gentle, pulsating light of thousands of flickering fireflies, which float lazily in the air like living stars. These tiny floating lanterns cast a warm golden glow, creating a dynamic effect of soft highlights and shifting shadows on skin and grainy wood. The woman's hand is slender and delicate, while the man's is slightly larger and more defined. As the camera continues to gently rotate, creating the feeling that the world is gently spinning around their intimate connection, their fingertips finally make the slightest, most delicate first contact. This single point of connection, held for one beat, achieves a playful and tender anticipation, while the blurred background of the dense forest silhouette and countless distant fireflies creates a profound sense of depth and a deeply romantic, magical atmosphere.", + "seed": 26760, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_6/index_7957_seed_26760_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.14426588155329229, + "vlm_logprob": -0.14426588155329229, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.14426588155329229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.100041463971138 + ], + "opposite_logprob_tokens": [ + -2.475041389465332 + ], + "logprob_reward": -0.100041463971138, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.100041463971138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4204152524471283, + 0.020360831171274185, + 0.05874556303024292, + 0.06635353714227676, + 0.025071226060390472, + 0.08911184966564178, + 0.13466086983680725, + 0.06389782577753067, + 0.07751239836215973, + 0.043870583176612854 + ], + "credit_map": [ + 0.02185788005590439, + 0.014366078190505505, + 0.1641482710838318, + 0.018122384324669838, + 0.024708818644285202, + 0.0902222990989685, + 0.06285889446735382, + 0.024130644276738167, + 0.0021958451252430677, + 0.002397313015535474, + 0.0029076321516186, + 0.003318147035315633, + 0.0023484695702791214, + 0.003040503477677703, + 0.0017635817639529705, + 0.0023893381003290415, + 0.005979103501886129, + 0.004860439337790012, + 0.014344482682645321, + 0.004812939558178186, + 0.003401609370484948, + 0.014605463482439518, + 0.006009718403220177, + 0.004731806460767984, + 0.005572321824729443, + 0.0060432422906160355, + 0.011065232567489147, + 0.005751042161136866, + 0.004738367162644863, + 0.01879092864692211, + 0.006171426270157099, + 0.008220980875194073, + 0.002525686053559184, + 0.003493025666102767, + 0.0038931663148105145, + 0.004110874142497778, + 0.0023740322794765234, + 0.003269450506195426, + 0.0032319920137524605, + 0.002172998618334532, + 0.01218290813267231, + 0.013289660215377808, + 0.013089160434901714, + 0.013520457781851292, + 0.0113136675208807, + 0.010826523415744305, + 0.009073982015252113, + 0.005815485958009958, + 0.013362402096390724, + 0.020455949008464813, + 0.016405506059527397, + 0.01927143894135952, + 0.014681247062981129, + 0.02046942710876465, + 0.016092970967292786, + 0.01392192393541336, + 0.006723879836499691, + 0.012359187006950378, + 0.005232821684330702, + 0.01109201181679964, + 0.007980096153914928, + 0.005428154021501541, + 0.007419349160045385, + 0.007662329822778702, + 0.007178829051554203, + 0.005486557260155678, + 0.009271245449781418, + 0.004459694493561983, + 0.005121610127389431, + 0.020021524280309677, + 0.0150234904140234, + 0.010949449613690376, + 0.0040734559297561646, + 0.0057561700232326984, + 0.004574035294353962, + 0.00596614321693778, + 0.005346192978322506, + 0.005745233036577702, + 0.007241251412779093, + 0.00516810268163681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.100041463971138 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38710999488830566 + ], + "opposite_logprob_tokens": [ + -1.1371099948883057 + ], + "logprob_reward": -0.38710999488830566, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38710999488830566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39468809962272644, + 0.01741994544863701, + 0.11132258176803589, + 0.06855843961238861, + 0.027704467996954918, + 0.10989333689212799, + 0.11854104697704315, + 0.012500214390456676, + 0.08110281080007553, + 0.05826910212635994 + ], + "credit_map": [ + 0.021075664088129997, + 0.013513067737221718, + 0.15656262636184692, + 0.015324716456234455, + 0.024845199659466743, + 0.08117488771677017, + 0.06094404309988022, + 0.02124783582985401, + 0.002611039439216256, + 0.00188416987657547, + 0.002425222657620907, + 0.002546323463320732, + 0.002360887359827757, + 0.0019760990981012583, + 0.0019110237481072545, + 0.0017051779432222247, + 0.010301795788109303, + 0.007395467255264521, + 0.03267735615372658, + 0.006377904210239649, + 0.004466306883841753, + 0.03415894880890846, + 0.007845604792237282, + 0.008099193684756756, + 0.0069837202318012714, + 0.009228530339896679, + 0.008773107081651688, + 0.006527736317366362, + 0.006166277918964624, + 0.015593851916491985, + 0.00555959390476346, + 0.009725615382194519, + 0.0041470094583928585, + 0.002229045145213604, + 0.003860737895593047, + 0.004444410093128681, + 0.0026415111497044563, + 0.004045905079692602, + 0.003951573744416237, + 0.0023842756636440754, + 0.014759560115635395, + 0.01660967618227005, + 0.015191888436675072, + 0.017271297052502632, + 0.014123791828751564, + 0.012936928309500217, + 0.011370092630386353, + 0.007630096282809973, + 0.009868048131465912, + 0.016809336841106415, + 0.018463118001818657, + 0.01724584959447384, + 0.01084698736667633, + 0.015361404046416283, + 0.016682270914316177, + 0.013264021836221218, + 0.0014473703922703862, + 0.00220812507905066, + 0.0008580744033679366, + 0.0014395902398973703, + 0.0018998091109097004, + 0.0009870351059362292, + 0.001437607454136014, + 0.002222600858658552, + 0.004694943781942129, + 0.004883949179202318, + 0.005350841674953699, + 0.003947885241359472, + 0.0036731360014528036, + 0.021499842405319214, + 0.024720661342144012, + 0.012331544421613216, + 0.00626350985839963, + 0.006660232320427895, + 0.007095769979059696, + 0.007287805899977684, + 0.008732276037335396, + 0.005896730814129114, + 0.0072059715166687965, + 0.009126799181103706 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.38710999488830566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597489550709724 + ], + "opposite_logprob_tokens": [ + -3.895597457885742 + ], + "logprob_reward": -0.020597489550709724, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597489550709724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2452281266450882, + 0.020441697910428047, + 0.18175503611564636, + 0.06261701136827469, + 0.017794594168663025, + 0.05950472503900528, + 0.1560317873954773, + 0.053117189556360245, + 0.1501625031232834, + 0.05334730073809624 + ], + "credit_map": [ + 0.013642184436321259, + 0.009933936409652233, + 0.09185541421175003, + 0.013037865981459618, + 0.015476003289222717, + 0.05138779059052467, + 0.035411536693573, + 0.014483400620520115, + 0.0017205624608322978, + 0.002381854923442006, + 0.0028977137990295887, + 0.003776353085413575, + 0.002637941623106599, + 0.0030548698268830776, + 0.0018168933456763625, + 0.0021555095445364714, + 0.015810392796993256, + 0.012625088915228844, + 0.0502437986433506, + 0.010485567152500153, + 0.012167122215032578, + 0.05356576666235924, + 0.013894006609916687, + 0.012963317334651947, + 0.005274682305753231, + 0.005286990199238062, + 0.008492598310112953, + 0.004827567841857672, + 0.00458327354863286, + 0.02001638524234295, + 0.0054271225817501545, + 0.008708395063877106, + 0.002157927257940173, + 0.0016905551310628653, + 0.002431409200653434, + 0.003706398420035839, + 0.001534452079795301, + 0.0019119044300168753, + 0.002776042791083455, + 0.0015859069535508752, + 0.008843319490551949, + 0.010643756948411465, + 0.007854548282921314, + 0.008707083761692047, + 0.007716541178524494, + 0.006794249638915062, + 0.005300077609717846, + 0.003645150223746896, + 0.011245221830904484, + 0.0168357752263546, + 0.020640384405851364, + 0.02354036085307598, + 0.016165001317858696, + 0.024242611601948738, + 0.0256655216217041, + 0.017696907743811607, + 0.005139987450093031, + 0.009652361273765564, + 0.00411625113338232, + 0.009551193565130234, + 0.006202182732522488, + 0.005310984794050455, + 0.0056944009847939014, + 0.0074498336762189865, + 0.009681985713541508, + 0.007219512015581131, + 0.011469317600131035, + 0.008306720294058323, + 0.006368936505168676, + 0.0375809445977211, + 0.04174303635954857, + 0.027792073786258698, + 0.004747894126921892, + 0.0053078108467161655, + 0.006216651294380426, + 0.005757865030318499, + 0.00588260218501091, + 0.004900392144918442, + 0.006492638494819403, + 0.014041452668607235 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.020597489550709724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988397181034088 + ], + "opposite_logprob_tokens": [ + -1.8048839569091797 + ], + "logprob_reward": -0.17988397181034088, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988397181034088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19574005901813507, + 0.03848400339484215, + 0.13954231142997742, + 0.08121411502361298, + 0.048960667103528976, + 0.048420652747154236, + 0.0949016660451889, + 0.02882968634366989, + 0.27443164587020874, + 0.04947519674897194 + ], + "credit_map": [ + 0.011432269588112831, + 0.008685802109539509, + 0.06882913410663605, + 0.015412361361086369, + 0.0137374522164464, + 0.03566115349531174, + 0.030924752354621887, + 0.011057130061089993, + 0.0037915064021945, + 0.005934899672865868, + 0.0056765940971672535, + 0.005667800083756447, + 0.00392242893576622, + 0.004773538559675217, + 0.003379272762686014, + 0.005337960086762905, + 0.01012780237942934, + 0.007806473411619663, + 0.04891156405210495, + 0.007542747538536787, + 0.004122155252844095, + 0.044922031462192535, + 0.006482724566012621, + 0.009626816026866436, + 0.0044441986829042435, + 0.00529894232749939, + 0.011150314472615719, + 0.010140362195670605, + 0.004786861594766378, + 0.02161233499646187, + 0.012204892002046108, + 0.01157621294260025, + 0.01109384372830391, + 0.004595513921231031, + 0.006323043256998062, + 0.007310162298381329, + 0.004262265749275684, + 0.005181232467293739, + 0.005922925658524036, + 0.004271681420505047, + 0.005564154125750065, + 0.006663314066827297, + 0.008118676953017712, + 0.006491502746939659, + 0.005822315346449614, + 0.00416173180565238, + 0.00681639090180397, + 0.004782564006745815, + 0.007912207394838333, + 0.009284106083214283, + 0.007867449894547462, + 0.01709383726119995, + 0.007470902521163225, + 0.008884829469025135, + 0.024113019928336143, + 0.012275313958525658, + 0.003447769908234477, + 0.003753876779228449, + 0.001998627558350563, + 0.004182273056358099, + 0.004394471645355225, + 0.003736938117071986, + 0.0034822907764464617, + 0.0038334366399794817, + 0.01874127797782421, + 0.013581270352005959, + 0.029816193506121635, + 0.014931208454072475, + 0.013560661114752293, + 0.093263640999794, + 0.056554410606622696, + 0.033982980996370316, + 0.004674168303608894, + 0.006072739604860544, + 0.005252808332443237, + 0.006398443598300219, + 0.006135276518762112, + 0.006604457274079323, + 0.007140717934817076, + 0.007196586113423109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.17988397181034088 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the hands of a man and a woman visible reaching for each other on a wooden plank?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night in an enchanted forest with a dense forest silhouette in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene illuminated by thousands of flickering fireflies casting a warm golden glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a slow, graceful orbiting panorama that rotates around the subjects?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the man's and woman's fingertips make contact and hold that connection for a brief moment?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336964875459671 + ], + "opposite_logprob_tokens": [ + -3.408696413040161 + ], + "logprob_reward": -0.0336964875459671, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336964875459671, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3928963840007782, + 0.03676849976181984, + 0.11463000625371933, + 0.06707074493169785, + 0.0373556986451149, + 0.04261002689599991, + 0.17742332816123962, + 0.06406031548976898, + 0.031034229323267937, + 0.03615065664052963 + ], + "credit_map": [ + 0.02108645997941494, + 0.01518960203975439, + 0.15331552922725677, + 0.018931742757558823, + 0.027945170179009438, + 0.0720425695180893, + 0.06535562127828598, + 0.01902974210679531, + 0.007005518302321434, + 0.0027274861931800842, + 0.0066227177157998085, + 0.005816951859742403, + 0.004128336440771818, + 0.003536272095516324, + 0.0034169440623372793, + 0.003514280077069998, + 0.00960987713187933, + 0.006589988712221384, + 0.03508809953927994, + 0.006091485731303692, + 0.0056221154518425465, + 0.03714769333600998, + 0.0061424411833286285, + 0.00833831261843443, + 0.004351100884377956, + 0.005187381058931351, + 0.008903391659259796, + 0.00511506199836731, + 0.0032595410011708736, + 0.021610068157315254, + 0.0074062589555978775, + 0.011237944476306438, + 0.006326727103441954, + 0.003566945670172572, + 0.005011773202568293, + 0.0071403528563678265, + 0.0032768575474619865, + 0.0036638532765209675, + 0.005570534151047468, + 0.0027986587956547737, + 0.00671991566196084, + 0.0069735487923026085, + 0.0059769898653030396, + 0.006314010825008154, + 0.004974169656634331, + 0.0044125220738351345, + 0.004265918396413326, + 0.002972954185679555, + 0.016801126301288605, + 0.03332453966140747, + 0.0157407708466053, + 0.023738304153084755, + 0.019757268950343132, + 0.027025815099477768, + 0.022263137623667717, + 0.018772397190332413, + 0.0064440201967954636, + 0.01049136370420456, + 0.005511576309800148, + 0.011045567691326141, + 0.008259163238108158, + 0.006888899952173233, + 0.007401396986097097, + 0.008018333464860916, + 0.0035668949130922556, + 0.005613413173705339, + 0.0027538840658962727, + 0.002827622462064028, + 0.002486423123627901, + 0.003805925603955984, + 0.007280658930540085, + 0.002699411241337657, + 0.003762471489608288, + 0.003811288857832551, + 0.004444473423063755, + 0.0048764608800411224, + 0.0033931187354028225, + 0.004558060318231583, + 0.0054316832683980465, + 0.005873104091733694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.0336964875459671 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_tensor([9335, 9335])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_tensor([9335, 9335])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e6bd08aa0c09cee9c69aa155d88f55c3cb6aed5e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_tensor([9335, 9335])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium close-up shot, taken from a level angle, focuses on a handsome young scholar seated in an antique-style study. Soft afternoon sunlight slants through the carved wooden window lattice beside him, casting distinct beams of light in the air, within which countless dust particles dance slowly. The scholar's long, jet-black hair is tied back with an elegant wooden hairpin. He wears a moon-white, wide-sleeved, cotton-linen Hanfu (Chinese traditional clothing), with pale blue embroidery adorning the lapels and cuffs. On the rustic desk before him lies an ancient, thread-bound book with yellowed pages, its vertical lines of traditional Chinese ink characters clearly visible. His expression is focused and calm, his slender right index finger carefully, steadily, and deliberately tracing the words across the page, slowly moving from right to left. As his finger moves, his lips part slightly, moving softly, as if silently and reverently reciting ancient verses. The warm light softly illuminates his profile and the precious ancient book, casting mottled light and shadows from the window panes on his clothes and desk. The whole scene is filled with a tranquil, focused and elegant bookish atmosphere, and time seems to have frozen at this moment.", + "seed": 32631, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32631_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02579147811047733, + "vlm_logprob": -0.02579147811047733, + "vlm_yes_no_margin": 3.9999999523162844, + "vlm_token_credit": -0.02579147811047733, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044780973345041275 + ], + "opposite_logprob_tokens": [ + -3.169780969619751 + ], + "logprob_reward": -0.044780973345041275, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044780973345041275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3464632034301758, + 0.15745380520820618, + 0.08116313070058823, + 0.09116441011428833, + 0.06027206778526306, + 0.0793381929397583, + 0.04690762609243393, + 0.05608866736292839, + 0.05421219393610954, + 0.026936668902635574 + ], + "credit_map": [ + 0.025177745148539543, + 0.033172596246004105, + 0.023358164355158806, + 0.037960175424814224, + 0.08680298179388046, + 0.03618105873465538, + 0.03915597125887871, + 0.06465453654527664, + 0.017347177490592003, + 0.017607636749744415, + 0.014998101629316807, + 0.02054194174706936, + 0.031265340745449066, + 0.01825607195496559, + 0.016848281025886536, + 0.020589252933859825, + 0.009785136207938194, + 0.008727829903364182, + 0.009803993627429008, + 0.008795109577476978, + 0.013903350569307804, + 0.008198405615985394, + 0.009000208228826523, + 0.012949095107614994, + 0.013219796121120453, + 0.014177258126437664, + 0.009792989119887352, + 0.014226604253053665, + 0.011425282806158066, + 0.008609551005065441, + 0.00952906347811222, + 0.010183866135776043, + 0.005503894295543432, + 0.005794356111437082, + 0.005881066899746656, + 0.008609176613390446, + 0.011673815548419952, + 0.007328242063522339, + 0.008643759414553642, + 0.006837755907326937, + 0.01522573921829462, + 0.01304428931325674, + 0.009258279576897621, + 0.009288204833865166, + 0.009609727188944817, + 0.00546274334192276, + 0.008990705013275146, + 0.008458497934043407, + 0.006300187204033136, + 0.004947878886014223, + 0.0042647551745176315, + 0.00610081572085619, + 0.008364489302039146, + 0.006064118351787329, + 0.005804162472486496, + 0.005061218049377203, + 0.008850205689668655, + 0.005519624799489975, + 0.007312535308301449, + 0.007764261681586504, + 0.007366606034338474, + 0.007089688442647457, + 0.006371874827891588, + 0.0058138687163591385, + 0.004123741760849953, + 0.004046729765832424, + 0.005175760481506586, + 0.009883622638881207, + 0.009396232664585114, + 0.006453081499785185, + 0.00855688750743866, + 0.006576135754585266, + 0.002204553922638297, + 0.0023801575880497694, + 0.0031646282877773046, + 0.0028150961734354496, + 0.005495263729244471, + 0.005289879161864519, + 0.0027299844659864902, + 0.0028571058064699173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.044780973345041275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819819025695324 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.01819819025695324, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819819025695324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18557709455490112, + 0.21174609661102295, + 0.12778161466121674, + 0.04837917536497116, + 0.049973294138908386, + 0.12229232490062714, + 0.0625210851430893, + 0.07829175889492035, + 0.08200335502624512, + 0.03143415227532387 + ], + "credit_map": [ + 0.016525227576494217, + 0.0174570195376873, + 0.01837332546710968, + 0.02561182714998722, + 0.04331669211387634, + 0.015196716412901878, + 0.016911581158638, + 0.032184723764657974, + 0.017871150746941566, + 0.019467948004603386, + 0.024956369772553444, + 0.02129744552075863, + 0.029128465801477432, + 0.01632057875394821, + 0.03760052099823952, + 0.04510362818837166, + 0.01393456943333149, + 0.012529238127171993, + 0.011924156919121742, + 0.014049612917006016, + 0.02584720216691494, + 0.01740453764796257, + 0.013334233313798904, + 0.01875806599855423, + 0.00511258514598012, + 0.005218783393502235, + 0.006489493418484926, + 0.005625111050903797, + 0.009157788008451462, + 0.0046334825456142426, + 0.006031505297869444, + 0.0061104255728423595, + 0.004252440296113491, + 0.003723383881151676, + 0.007190417964011431, + 0.010025343857705593, + 0.008418355137109756, + 0.004851931240409613, + 0.005725244991481304, + 0.005786178633570671, + 0.028337417170405388, + 0.02010570652782917, + 0.014825884252786636, + 0.01366453431546688, + 0.013092859648168087, + 0.007782270200550556, + 0.01179603859782219, + 0.0126876225695014, + 0.012536829337477684, + 0.011234254576265812, + 0.006100998725742102, + 0.006557786371558905, + 0.005525384098291397, + 0.004620413761585951, + 0.008501001633703709, + 0.00744441756978631, + 0.007293347269296646, + 0.008610048331320286, + 0.006574327126145363, + 0.011256998404860497, + 0.015283831395208836, + 0.010505709797143936, + 0.009997771121561527, + 0.008769731037318707, + 0.007323463913053274, + 0.005697628017514944, + 0.006320101674646139, + 0.013815796002745628, + 0.015463770367205143, + 0.010852662846446037, + 0.012216318398714066, + 0.01031362172216177, + 0.0031129028648138046, + 0.0034239429514855146, + 0.004214719403535128, + 0.004375692922621965, + 0.00493413582444191, + 0.004320341162383556, + 0.0037164087407290936, + 0.0033360098022967577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01819819025695324 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980642206966877 + ], + "opposite_logprob_tokens": [ + -3.52980637550354 + ], + "logprob_reward": -0.02980642206966877, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980642206966877, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1430554836988449, + 0.09866812080144882, + 0.059719186276197433, + 0.10361148416996002, + 0.12011194229125977, + 0.0887303575873375, + 0.047679353505373, + 0.15388688445091248, + 0.1503792107105255, + 0.03415793180465698 + ], + "credit_map": [ + 0.014687367714941502, + 0.013597153127193451, + 0.012232476845383644, + 0.031236670911312103, + 0.02374568022787571, + 0.012181706726551056, + 0.014198591001331806, + 0.021175861358642578, + 0.00810601282864809, + 0.010366297326982021, + 0.014806676656007767, + 0.01119657326489687, + 0.01728152297437191, + 0.006765370257198811, + 0.014652540907263756, + 0.01549314334988594, + 0.0066432286985218525, + 0.007152203936129808, + 0.008942646905779839, + 0.008073489181697369, + 0.008227542974054813, + 0.0048862299881875515, + 0.006644729524850845, + 0.009149127639830112, + 0.024485047906637192, + 0.024471037089824677, + 0.0070932526141405106, + 0.010975885204970837, + 0.01380597148090601, + 0.004743572324514389, + 0.007686160504817963, + 0.010350564494729042, + 0.008157596923410892, + 0.009505813010036945, + 0.01420611422508955, + 0.02208769880235195, + 0.02250649593770504, + 0.014774943701922894, + 0.015065082348883152, + 0.013808207586407661, + 0.014343698509037495, + 0.013914217241108418, + 0.01417771726846695, + 0.011615565046668053, + 0.010031936690211296, + 0.006209795828908682, + 0.009414615109562874, + 0.009022822603583336, + 0.006313536316156387, + 0.004957339260727167, + 0.004459877032786608, + 0.0072111040353775024, + 0.006853756494820118, + 0.0046034278348088264, + 0.007567309774458408, + 0.0057130069471895695, + 0.014074478298425674, + 0.014823737554252148, + 0.012592187151312828, + 0.021955251693725586, + 0.030847294256091118, + 0.02202201448380947, + 0.019728127866983414, + 0.017843810841441154, + 0.014441837556660175, + 0.009383794851601124, + 0.012354608625173569, + 0.027053212746977806, + 0.02737717144191265, + 0.019260715693235397, + 0.021853433921933174, + 0.018654445186257362, + 0.0031995230820029974, + 0.003192990319803357, + 0.003027108730748296, + 0.004109472502022982, + 0.005870521534234285, + 0.005073759239166975, + 0.006013962905853987, + 0.003670600475743413 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02980642206966877 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024873053189367056 + ], + "opposite_logprob_tokens": [ + -6.0024871826171875 + ], + "logprob_reward": -0.0024873053189367056, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024873053189367056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2178627997636795, + 0.20938435196876526, + 0.08455751836299896, + 0.05852559208869934, + 0.0679241344332695, + 0.07403140515089035, + 0.10083562880754471, + 0.0761425569653511, + 0.06999234855175018, + 0.04074366018176079 + ], + "credit_map": [ + 0.03856096416711807, + 0.031896963715553284, + 0.021911248564720154, + 0.02793234772980213, + 0.03125234320759773, + 0.01497956458479166, + 0.023776115849614143, + 0.027553260326385498, + 0.02064289152622223, + 0.022714024409651756, + 0.023373983800411224, + 0.017220696434378624, + 0.01866871491074562, + 0.010504554957151413, + 0.04290524125099182, + 0.05335423722863197, + 0.016104256734251976, + 0.015736311674118042, + 0.006451847497373819, + 0.008276109583675861, + 0.011462504044175148, + 0.009710459969937801, + 0.0072861784137785435, + 0.009529848583042622, + 0.005844730418175459, + 0.005880078300833702, + 0.004760005045682192, + 0.008760164491832256, + 0.01416399609297514, + 0.005869083106517792, + 0.005265556741505861, + 0.007981978356838226, + 0.011562415398657322, + 0.009249637834727764, + 0.008273812010884285, + 0.010310400277376175, + 0.008504875935614109, + 0.006045051850378513, + 0.006620419677346945, + 0.007357522379606962, + 0.006412468384951353, + 0.006434012204408646, + 0.006505925208330154, + 0.010961235500872135, + 0.017089027911424637, + 0.007626251317560673, + 0.009788424707949162, + 0.00921405479311943, + 0.017064159736037254, + 0.015550763346254826, + 0.009862758219242096, + 0.012185189872980118, + 0.01193070225417614, + 0.007839241996407509, + 0.011915885843336582, + 0.014486923813819885, + 0.009374734945595264, + 0.009809110313653946, + 0.007166227325797081, + 0.007939326576888561, + 0.015334728173911572, + 0.009433060884475708, + 0.008734060451388359, + 0.008351302705705166, + 0.005504233296960592, + 0.004891415126621723, + 0.005201100371778011, + 0.011744820512831211, + 0.0128652174025774, + 0.009247565641999245, + 0.011283282190561295, + 0.009254715405404568, + 0.004677397198975086, + 0.005171139724552631, + 0.005040792748332024, + 0.003917692229151726, + 0.0073560201562941074, + 0.006111870054155588, + 0.004062257241457701, + 0.004406491294503212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0024873053189367056 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368449956178665 + ], + "opposite_logprob_tokens": [ + -3.408684492111206 + ], + "logprob_reward": -0.03368449956178665, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368449956178665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18420079350471497, + 0.19396260380744934, + 0.08530737459659576, + 0.07150246202945709, + 0.065467968583107, + 0.172465980052948, + 0.08516068011522293, + 0.05562478303909302, + 0.054747071117162704, + 0.031560249626636505 + ], + "credit_map": [ + 0.02198939025402069, + 0.015768229961395264, + 0.023707939311861992, + 0.026242103427648544, + 0.04012637957930565, + 0.013446890749037266, + 0.015935542061924934, + 0.026984311640262604, + 0.031117264181375504, + 0.033868350088596344, + 0.029988817870616913, + 0.01735256053507328, + 0.025553768500685692, + 0.016492854803800583, + 0.016963524743914604, + 0.022625457495450974, + 0.017691142857074738, + 0.017783861607313156, + 0.007233630865812302, + 0.01136855036020279, + 0.010343337431550026, + 0.005365879274904728, + 0.007424656301736832, + 0.00809631496667862, + 0.010673695243895054, + 0.010497793555259705, + 0.004384391475468874, + 0.0075811538845300674, + 0.01856853812932968, + 0.006405852735042572, + 0.005173452198505402, + 0.008217591792345047, + 0.007554026320576668, + 0.006041062995791435, + 0.004389107692986727, + 0.008393769152462482, + 0.015482068993151188, + 0.007759191561490297, + 0.007761742454022169, + 0.008086995221674442, + 0.024610323831439018, + 0.017218666151165962, + 0.012269177474081516, + 0.020665865391492844, + 0.0382743701338768, + 0.018575014546513557, + 0.02131384052336216, + 0.01953871361911297, + 0.02092767506837845, + 0.016744472086429596, + 0.008709310553967953, + 0.008545869961380959, + 0.0070202588103711605, + 0.005649181082844734, + 0.00829635001718998, + 0.009267573244869709, + 0.005841640755534172, + 0.0066797323524951935, + 0.005013167858123779, + 0.0072844382375478745, + 0.010851657018065453, + 0.00740085169672966, + 0.006313301622867584, + 0.0062399934977293015, + 0.0057505592703819275, + 0.004641261883080006, + 0.004689507652074099, + 0.007771138567477465, + 0.011291664093732834, + 0.006259387359023094, + 0.008214347064495087, + 0.006129206158220768, + 0.0043684071861207485, + 0.004503780510276556, + 0.0032729741651564837, + 0.003297205315902829, + 0.005222764331847429, + 0.003288447391241789, + 0.004245365504175425, + 0.0033613070845603943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.03368449956178665 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, focuses on a handsome young scholar seated in an antique-style study. Soft afternoon sunlight slants through the carved wooden window lattice beside him, casting distinct beams of light in the air, within which countless dust particles dance slowly. The scholar's long, jet-black hair is tied back with an elegant wooden hairpin. He wears a moon-white, wide-sleeved, cotton-linen Hanfu (Chinese traditional clothing), with pale blue embroidery adorning the lapels and cuffs. On the rustic desk before him lies an ancient, thread-bound book with yellowed pages, its vertical lines of traditional Chinese ink characters clearly visible. His expression is focused and calm, his slender right index finger carefully, steadily, and deliberately tracing the words across the page, slowly moving from right to left. As his finger moves, his lips part slightly, moving softly, as if silently and reverently reciting ancient verses. The warm light softly illuminates his profile and the precious ancient book, casting mottled light and shadows from the window panes on his clothes and desk. The whole scene is filled with a tranquil, focused and elegant bookish atmosphere, and time seems to have frozen at this moment.", + "seed": 32632, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32632_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021214643120765687, + "vlm_logprob": -0.021214643120765687, + "vlm_yes_no_margin": 4.175, + "vlm_token_credit": -0.021214643120765687, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020132016390562057 + ], + "opposite_logprob_tokens": [ + -4.020132064819336 + ], + "logprob_reward": -0.020132016390562057, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.020132016390562057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3382399082183838, + 0.12651613354682922, + 0.05199909210205078, + 0.05895650014281273, + 0.07298465818166733, + 0.06608890742063522, + 0.09756582975387573, + 0.03330475836992264, + 0.0853143185377121, + 0.06902986764907837 + ], + "credit_map": [ + 0.020166343078017235, + 0.022227756679058075, + 0.02975357323884964, + 0.06112605705857277, + 0.07036088407039642, + 0.041112493723630905, + 0.05242973193526268, + 0.0410630889236927, + 0.006785984616726637, + 0.008535146713256836, + 0.012037787586450577, + 0.026345277205109596, + 0.019685368984937668, + 0.01353180781006813, + 0.021828696131706238, + 0.01776607148349285, + 0.0029653841629624367, + 0.004223635420203209, + 0.0049814945086836815, + 0.008814442902803421, + 0.008395813405513763, + 0.007391936611384153, + 0.007259252481162548, + 0.007967133074998856, + 0.004341050051152706, + 0.006819220259785652, + 0.010625935159623623, + 0.010285748168826103, + 0.00804839190095663, + 0.005776221863925457, + 0.0076208701357245445, + 0.005439063534140587, + 0.006000260356813669, + 0.009677717462182045, + 0.008486272767186165, + 0.011590314097702503, + 0.007521496154367924, + 0.008812746964395046, + 0.012696556746959686, + 0.008199295960366726, + 0.005333874840289354, + 0.007538619916886091, + 0.007629673928022385, + 0.010538176633417606, + 0.010044951923191547, + 0.005250541493296623, + 0.010883228853344917, + 0.008869839832186699, + 0.004976927302777767, + 0.011025287210941315, + 0.016147350892424583, + 0.01489014457911253, + 0.009345288388431072, + 0.006314323283731937, + 0.02078913524746895, + 0.01407737098634243, + 0.0016989395953714848, + 0.002445896854624152, + 0.0057067167945206165, + 0.005203386768698692, + 0.005177418235689402, + 0.004428918939083815, + 0.0044991387985646725, + 0.004144341219216585, + 0.005161531735211611, + 0.006475407164543867, + 0.006257086526602507, + 0.009029703214764595, + 0.008268744684755802, + 0.01672181859612465, + 0.01253088004887104, + 0.020869147032499313, + 0.004308443982154131, + 0.0053076897747814655, + 0.008635254576802254, + 0.010478449985384941, + 0.007078803144395351, + 0.011300908401608467, + 0.01059962809085846, + 0.011320686899125576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020132016390562057 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072357073426247 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072357073426247, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072357073426247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2306572049856186, + 0.18938356637954712, + 0.10683573782444, + 0.06660109758377075, + 0.060252927243709564, + 0.05611898750066757, + 0.06519721448421478, + 0.03384682163596153, + 0.10712418705224991, + 0.08398230373859406 + ], + "credit_map": [ + 0.011153931729495525, + 0.022420041263103485, + 0.027446236461400986, + 0.04708506539463997, + 0.039353128522634506, + 0.02686048299074173, + 0.027248721569776535, + 0.029089566320180893, + 0.01144944503903389, + 0.014462408609688282, + 0.020251711830496788, + 0.036742646247148514, + 0.030474964529275894, + 0.02762496843934059, + 0.024088239297270775, + 0.02428916096687317, + 0.00759661803022027, + 0.0086398646235466, + 0.010333790443837643, + 0.021176159381866455, + 0.013914469629526138, + 0.01966273970901966, + 0.013245005160570145, + 0.012267075479030609, + 0.0033164529595524073, + 0.007892079651355743, + 0.009826011024415493, + 0.0163462795317173, + 0.009190801531076431, + 0.004159202799201012, + 0.009737818501889706, + 0.006132450886070728, + 0.004392944276332855, + 0.0070116836577653885, + 0.006833030376583338, + 0.010392806492745876, + 0.006851687096059322, + 0.005492121446877718, + 0.012191536836326122, + 0.007087111473083496, + 0.0030818034429103136, + 0.007072627078741789, + 0.0066970838233828545, + 0.007077806629240513, + 0.00793508067727089, + 0.007036246825009584, + 0.010833417065441608, + 0.006384918466210365, + 0.004126964136958122, + 0.007607452105730772, + 0.009742382913827896, + 0.011093770153820515, + 0.006495967041701078, + 0.005454024765640497, + 0.011502518318593502, + 0.009174128994345665, + 0.0025885251816362143, + 0.004772481508553028, + 0.004881210159510374, + 0.007419127505272627, + 0.004124775994569063, + 0.0026657204143702984, + 0.0034629427827894688, + 0.003932032734155655, + 0.007129184436053038, + 0.009193112142384052, + 0.008160468190908432, + 0.008655638433992863, + 0.010497809387743473, + 0.0214780755341053, + 0.01604410819709301, + 0.02596578188240528, + 0.005639962386339903, + 0.008375108242034912, + 0.010481556877493858, + 0.012041711248457432, + 0.007019459269940853, + 0.012630019336938858, + 0.01336634624749422, + 0.01442813128232956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.016072357073426247 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033700864762067795 + ], + "opposite_logprob_tokens": [ + -3.408700942993164 + ], + "logprob_reward": -0.033700864762067795, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033700864762067795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22400449216365814, + 0.11565455049276352, + 0.07967586815357208, + 0.07418598234653473, + 0.06049952656030655, + 0.07409900426864624, + 0.05120040848851204, + 0.05134076997637749, + 0.13240018486976624, + 0.1369391679763794 + ], + "credit_map": [ + 0.014964347705245018, + 0.014062656089663506, + 0.01899326592683792, + 0.03313524276018143, + 0.036505866795778275, + 0.038436539471149445, + 0.03302247077226639, + 0.03488411381840706, + 0.008903192356228828, + 0.00794871337711811, + 0.00996793620288372, + 0.022727061063051224, + 0.016253061592578888, + 0.020315952599048615, + 0.015031879767775536, + 0.014506768435239792, + 0.006644813809543848, + 0.006376801058650017, + 0.00699771661311388, + 0.012777390889823437, + 0.009488705545663834, + 0.017898572608828545, + 0.0091397101059556, + 0.010352167300879955, + 0.003840483259409666, + 0.0082226088270545, + 0.009673722088336945, + 0.016026152297854424, + 0.008606995455920696, + 0.0060823410749435425, + 0.01275548618286848, + 0.00897819735109806, + 0.004181147087365389, + 0.00707633513957262, + 0.006173005793243647, + 0.011824758723378181, + 0.008084415458142757, + 0.005222279112786055, + 0.011281179264187813, + 0.006656411103904247, + 0.00601972034201026, + 0.010863167233765125, + 0.0077308774925768375, + 0.00808221846818924, + 0.010175840929150581, + 0.008540600538253784, + 0.012654674239456654, + 0.010031914338469505, + 0.0024824561551213264, + 0.005073127802461386, + 0.006810434628278017, + 0.006623563822358847, + 0.00612814724445343, + 0.006917859427630901, + 0.009671488776803017, + 0.00749333156272769, + 0.004190999083220959, + 0.00508942361921072, + 0.010191149078309536, + 0.007650529034435749, + 0.0052602640353143215, + 0.004899873863905668, + 0.007779362611472607, + 0.006279175169765949, + 0.009684850461781025, + 0.012972177937626839, + 0.011296087875962257, + 0.01033700630068779, + 0.01581547036767006, + 0.03218267112970352, + 0.015369336120784283, + 0.024742603302001953, + 0.008985849097371101, + 0.010218123905360699, + 0.015479737892746925, + 0.02012571692466736, + 0.012311098165810108, + 0.013720294460654259, + 0.030342213809490204, + 0.02575615979731083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.033700864762067795 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002486472949385643 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.002486472949385643, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002486472949385643, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23795552551746368, + 0.08993709832429886, + 0.053042128682136536, + 0.04446735233068466, + 0.07489314675331116, + 0.12968981266021729, + 0.14971362054347992, + 0.06704884022474289, + 0.066555455327034, + 0.08669697493314743 + ], + "credit_map": [ + 0.0175483338534832, + 0.026554998010396957, + 0.02575128711760044, + 0.046638187021017075, + 0.02772590145468712, + 0.01660994626581669, + 0.0347125269472599, + 0.0424143485724926, + 0.006031082011759281, + 0.01184295117855072, + 0.011974062770605087, + 0.016030743718147278, + 0.015195646323263645, + 0.006151590961962938, + 0.011791493743658066, + 0.010919529013335705, + 0.00425823125988245, + 0.006146810483187437, + 0.006359803024679422, + 0.009551139548420906, + 0.009100542403757572, + 0.004968320019543171, + 0.006763002835214138, + 0.005894277710467577, + 0.0035231271758675575, + 0.005973222199827433, + 0.0038119584787636995, + 0.007591397035866976, + 0.005633384454995394, + 0.006839235778898001, + 0.006475865840911865, + 0.004619162064045668, + 0.005638627335429192, + 0.00872329156845808, + 0.009525056928396225, + 0.009747768752276897, + 0.008668323047459126, + 0.0075493305921554565, + 0.012070298194885254, + 0.012970447540283203, + 0.010551577433943748, + 0.020329652354121208, + 0.019609766080975533, + 0.017875121906399727, + 0.011653807014226913, + 0.02016776241362095, + 0.01391670573502779, + 0.015585421584546566, + 0.01204823236912489, + 0.019109169021248817, + 0.02041887305676937, + 0.01886024884879589, + 0.01529099140316248, + 0.024928057566285133, + 0.019042039290070534, + 0.020016012713313103, + 0.0036657461896538734, + 0.006427163258194923, + 0.008266855962574482, + 0.01563488319516182, + 0.0057323588989675045, + 0.01069401204586029, + 0.008449537679553032, + 0.008178277872502804, + 0.00474770599976182, + 0.005939395632594824, + 0.0065291342325508595, + 0.01028114091604948, + 0.006482157856225967, + 0.01423716265708208, + 0.009156282059848309, + 0.009182475507259369, + 0.007202078588306904, + 0.01555884163826704, + 0.009580610319972038, + 0.009067828767001629, + 0.008871178142726421, + 0.01696181669831276, + 0.011077580973505974, + 0.008377035148441792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.002486472949385643 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368150442838669 + ], + "opposite_logprob_tokens": [ + -3.4086813926696777 + ], + "logprob_reward": -0.03368150442838669, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368150442838669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18589720129966736, + 0.14270338416099548, + 0.0903162881731987, + 0.08810857683420181, + 0.09410558640956879, + 0.10440202057361603, + 0.06817451864480972, + 0.07157169282436371, + 0.07065242528915405, + 0.08406829833984375 + ], + "credit_map": [ + 0.01521745789796114, + 0.018010621890425682, + 0.022975564002990723, + 0.029178636148571968, + 0.03036416321992874, + 0.01755492575466633, + 0.027741599828004837, + 0.02485423907637596, + 0.011993619613349438, + 0.00940091721713543, + 0.014873374253511429, + 0.0182255357503891, + 0.01931091770529747, + 0.027390478178858757, + 0.023682421073317528, + 0.01782611943781376, + 0.005860538221895695, + 0.005116089712828398, + 0.009542264975607395, + 0.01572062261402607, + 0.014066622592508793, + 0.01233771350234747, + 0.01698276214301586, + 0.01068967953324318, + 0.007356122136116028, + 0.010259191505610943, + 0.010050764307379723, + 0.01337582990527153, + 0.011958411894738674, + 0.013494305312633514, + 0.011735038831830025, + 0.009878911077976227, + 0.00671871704980731, + 0.008875280618667603, + 0.014398780651390553, + 0.009719353169202805, + 0.014704830944538116, + 0.012509610503911972, + 0.01414625346660614, + 0.013032762333750725, + 0.008550621569156647, + 0.018873833119869232, + 0.013150184415280819, + 0.013644620776176453, + 0.009585783816874027, + 0.010861900635063648, + 0.01818602718412876, + 0.011549048125743866, + 0.004199539311230183, + 0.005389632191509008, + 0.007908849976956844, + 0.015321132726967335, + 0.00773946987465024, + 0.009080462157726288, + 0.010166079737246037, + 0.008369357325136662, + 0.004945877939462662, + 0.00652651209384203, + 0.012598608620464802, + 0.016923658549785614, + 0.006556835491210222, + 0.008756167255342007, + 0.00874833669513464, + 0.006515693385154009, + 0.005058653187006712, + 0.006464025937020779, + 0.005388092249631882, + 0.007733355741947889, + 0.010067060589790344, + 0.01845303364098072, + 0.008829853497445583, + 0.008658352307975292, + 0.005358458496630192, + 0.011033535934984684, + 0.009433162398636341, + 0.012546523474156857, + 0.009326099418103695, + 0.008538974449038506, + 0.01656605862081051, + 0.011265484616160393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03368150442838669 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, focuses on a handsome young scholar seated in an antique-style study. Soft afternoon sunlight slants through the carved wooden window lattice beside him, casting distinct beams of light in the air, within which countless dust particles dance slowly. The scholar's long, jet-black hair is tied back with an elegant wooden hairpin. He wears a moon-white, wide-sleeved, cotton-linen Hanfu (Chinese traditional clothing), with pale blue embroidery adorning the lapels and cuffs. On the rustic desk before him lies an ancient, thread-bound book with yellowed pages, its vertical lines of traditional Chinese ink characters clearly visible. His expression is focused and calm, his slender right index finger carefully, steadily, and deliberately tracing the words across the page, slowly moving from right to left. As his finger moves, his lips part slightly, moving softly, as if silently and reverently reciting ancient verses. The warm light softly illuminates his profile and the precious ancient book, casting mottled light and shadows from the window panes on his clothes and desk. The whole scene is filled with a tranquil, focused and elegant bookish atmosphere, and time seems to have frozen at this moment.", + "seed": 32633, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32633_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03179111601784825, + "vlm_logprob": -0.03179111601784825, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03179111601784825, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10136403888463974 + ], + "opposite_logprob_tokens": [ + -2.3513641357421875 + ], + "logprob_reward": -0.10136403888463974, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10136403888463974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47631609439849854, + 0.1407313346862793, + 0.07010965794324875, + 0.03146686777472496, + 0.039554666727781296, + 0.030899792909622192, + 0.051369260996580124, + 0.07693705707788467, + 0.020775215700268745, + 0.06184002384543419 + ], + "credit_map": [ + 0.029732130467891693, + 0.040877435356378555, + 0.055387284606695175, + 0.0681455060839653, + 0.06449776142835617, + 0.038790348917245865, + 0.10685385763645172, + 0.07203178852796555, + 0.008819120936095715, + 0.014732932671904564, + 0.015388541854918003, + 0.01892995834350586, + 0.020224183797836304, + 0.008077440783381462, + 0.030975867062807083, + 0.023583311587572098, + 0.006107595283538103, + 0.0061457534320652485, + 0.00895955041050911, + 0.00945371575653553, + 0.007600759156048298, + 0.003348439233377576, + 0.017147812992334366, + 0.01134603563696146, + 0.002739259973168373, + 0.003252904862165451, + 0.004253960680216551, + 0.005826478358358145, + 0.0027893322985619307, + 0.0019235124345868826, + 0.005616214592009783, + 0.005065209232270718, + 0.002893500030040741, + 0.004566646181046963, + 0.004606249742209911, + 0.005666004493832588, + 0.004485901910811663, + 0.003936263266950846, + 0.007317252457141876, + 0.006082852836698294, + 0.0026603133883327246, + 0.003971828613430262, + 0.0039614103734493256, + 0.005705954506993294, + 0.00303949904628098, + 0.0017688092775642872, + 0.0047004735097289085, + 0.005091506522148848, + 0.0038988294545561075, + 0.004346356727182865, + 0.006008056458085775, + 0.010049402713775635, + 0.004683474078774452, + 0.0031858012080192566, + 0.01055869646370411, + 0.008638650178909302, + 0.005586381535977125, + 0.008149519562721252, + 0.009378915652632713, + 0.016155406832695007, + 0.005281089339405298, + 0.00366545794531703, + 0.01754556968808174, + 0.011174725368618965, + 0.00296008144505322, + 0.0022001087199896574, + 0.002328157424926758, + 0.0035299514420330524, + 0.002178533235564828, + 0.0012430232018232346, + 0.0036963559687137604, + 0.0026390061248093843, + 0.00493182148784399, + 0.0064315698109567165, + 0.005755268968641758, + 0.007112405728548765, + 0.007377854082733393, + 0.0038414360024034977, + 0.01603168249130249, + 0.010357990860939026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.10136403888463974 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026362450793385506 + ], + "opposite_logprob_tokens": [ + -3.651362419128418 + ], + "logprob_reward": -0.026362450793385506, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026362450793385506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3404252529144287, + 0.13557197153568268, + 0.10662593692541122, + 0.038779232650995255, + 0.060502566397190094, + 0.03747870773077011, + 0.06471595168113708, + 0.09364340454339981, + 0.060765933245420456, + 0.0614909827709198 + ], + "credit_map": [ + 0.028177840635180473, + 0.026592154055833817, + 0.03193863481283188, + 0.04223558306694031, + 0.04859394580125809, + 0.02197546698153019, + 0.06782825291156769, + 0.07308339327573776, + 0.010451061651110649, + 0.011528678238391876, + 0.01558922789990902, + 0.019604040309786797, + 0.014381828717887402, + 0.00515012675896287, + 0.03130273520946503, + 0.0275642741471529, + 0.007912905886769295, + 0.007506723050028086, + 0.0172208733856678, + 0.01722659356892109, + 0.012240187264978886, + 0.008244950324296951, + 0.02152988314628601, + 0.014743820764124393, + 0.0031966601964086294, + 0.003325649304315448, + 0.004164617974311113, + 0.00742181995883584, + 0.004136503674089909, + 0.003202247666195035, + 0.006539298687130213, + 0.006792433559894562, + 0.005001006182283163, + 0.0055132764391601086, + 0.007986428216099739, + 0.011828224174678326, + 0.006353868171572685, + 0.003949226811528206, + 0.010825132019817829, + 0.009045404382050037, + 0.003281234297901392, + 0.0036480354610830545, + 0.006442251615226269, + 0.005566470790654421, + 0.0037536376621574163, + 0.0021580306347459555, + 0.0058263130486011505, + 0.006802733521908522, + 0.004423264879733324, + 0.005752327386289835, + 0.008129599504172802, + 0.012883885763585567, + 0.005513338837772608, + 0.0035119690001010895, + 0.01483071967959404, + 0.009670846164226532, + 0.007701130583882332, + 0.010483202524483204, + 0.012571328319609165, + 0.01900000311434269, + 0.00669802026823163, + 0.004184550140053034, + 0.01985933445394039, + 0.013145840726792812, + 0.0041064550168812275, + 0.005914624780416489, + 0.00808589719235897, + 0.010843987576663494, + 0.007580362726002932, + 0.004395568743348122, + 0.011715796776115894, + 0.008123238570988178, + 0.006225290708243847, + 0.007739331107586622, + 0.0068962303921580315, + 0.007527886889874935, + 0.009285345673561096, + 0.00445611122995615, + 0.011351093649864197, + 0.008009691722691059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.026362450793385506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0110772168263793 + ], + "opposite_logprob_tokens": [ + -4.511077404022217 + ], + "logprob_reward": -0.0110772168263793, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0110772168263793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3180035650730133, + 0.23616182804107666, + 0.09300342947244644, + 0.05183366313576698, + 0.04401261359453201, + 0.028961023315787315, + 0.0356159433722496, + 0.09461134672164917, + 0.04324205592274666, + 0.05455450713634491 + ], + "credit_map": [ + 0.030088260769844055, + 0.02607046253979206, + 0.039945416152477264, + 0.050665274262428284, + 0.03252836689352989, + 0.019787278026342392, + 0.06818237155675888, + 0.05073613300919533, + 0.02284582518041134, + 0.01608234830200672, + 0.03936317190527916, + 0.05206533521413803, + 0.024772344157099724, + 0.008336828090250492, + 0.040009383112192154, + 0.03268659859895706, + 0.007799226325005293, + 0.00626665772870183, + 0.013339521363377571, + 0.019998539239168167, + 0.008866706863045692, + 0.004401363432407379, + 0.017152342945337296, + 0.015179070644080639, + 0.005371945444494486, + 0.004273524042218924, + 0.005435293540358543, + 0.007901646196842194, + 0.00586417131125927, + 0.002707412466406822, + 0.010461430996656418, + 0.009818242862820625, + 0.005947912111878395, + 0.004313110373914242, + 0.0046373759396374226, + 0.007928513921797276, + 0.004759531933814287, + 0.0031961335334926844, + 0.007294878363609314, + 0.005935154855251312, + 0.0032573090866208076, + 0.0021886008325964212, + 0.003776018274948001, + 0.0061492919921875, + 0.00223420443944633, + 0.0012514840345829725, + 0.005700555630028248, + 0.0044035580940544605, + 0.0032185993622988462, + 0.003139594104140997, + 0.004712698515504599, + 0.006561276037245989, + 0.003058371599763632, + 0.0014301458140835166, + 0.008473951369524002, + 0.005021307151764631, + 0.007163730915635824, + 0.009471110999584198, + 0.011419466696679592, + 0.020489169284701347, + 0.00647115521132946, + 0.004568707663565874, + 0.021269218996167183, + 0.013758794404566288, + 0.003719314467161894, + 0.004290468525141478, + 0.005943424999713898, + 0.011381962336599827, + 0.004313134588301182, + 0.0017972621135413647, + 0.005705364979803562, + 0.006091127637773752, + 0.00500347837805748, + 0.005641634110361338, + 0.007439774461090565, + 0.008990715257823467, + 0.005551142618060112, + 0.0027666515670716763, + 0.01135801151394844, + 0.007803098764270544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0110772168263793 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093007650226355 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093007650226355, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093007650226355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32157912850379944, + 0.12445984780788422, + 0.09532169252634048, + 0.04869275167584419, + 0.05274580046534538, + 0.08715377748012543, + 0.04751509428024292, + 0.05593622475862503, + 0.07853814959526062, + 0.08805757761001587 + ], + "credit_map": [ + 0.03243948519229889, + 0.020691687241196632, + 0.030439401045441628, + 0.0650835931301117, + 0.037020739167928696, + 0.017370058223605156, + 0.05442529916763306, + 0.06410887092351913, + 0.01256072148680687, + 0.009731665253639221, + 0.015386789105832577, + 0.01812250167131424, + 0.01248864270746708, + 0.005678586196154356, + 0.018616756424307823, + 0.03187417611479759, + 0.0061082299798727036, + 0.008070994168519974, + 0.010194971226155758, + 0.02368156984448433, + 0.007466903422027826, + 0.0040828934870660305, + 0.021567296236753464, + 0.014148836955428123, + 0.004652294795960188, + 0.0031538403127342463, + 0.0064574978314340115, + 0.00994991697371006, + 0.004241436719894409, + 0.002581919077783823, + 0.00898967869579792, + 0.008666166104376316, + 0.005479819606989622, + 0.00503346836194396, + 0.006948895752429962, + 0.010126703418791294, + 0.004205285105854273, + 0.0028083366341888905, + 0.010124105028808117, + 0.00801918562501669, + 0.00594743900001049, + 0.01131432130932808, + 0.011421914212405682, + 0.019160570576786995, + 0.007974095642566681, + 0.004388701636344194, + 0.013379625044763088, + 0.013567105866968632, + 0.0035910194274038076, + 0.005785304121673107, + 0.005701000336557627, + 0.009635833092033863, + 0.004607572220265865, + 0.002536279847845435, + 0.006750042550265789, + 0.008908040821552277, + 0.005060662981122732, + 0.005455926526337862, + 0.006774596869945526, + 0.011793483048677444, + 0.004026119597256184, + 0.002655815565958619, + 0.01215897686779499, + 0.00801064632833004, + 0.006021568551659584, + 0.008183371275663376, + 0.008088759146630764, + 0.01676381751894951, + 0.007138619665056467, + 0.004234199412167072, + 0.01622888818383217, + 0.011878919787704945, + 0.006247784476727247, + 0.009862152859568596, + 0.009304841980338097, + 0.011680644936859608, + 0.009379344992339611, + 0.005626338999718428, + 0.01891602948307991, + 0.017040440812706947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.004093007650226355 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058865934610367 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058865934610367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058865934610367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20863056182861328, + 0.13935020565986633, + 0.12786448001861572, + 0.05556774511933327, + 0.05076337605714798, + 0.059838104993104935, + 0.08917585760354996, + 0.13355357944965363, + 0.046220798045396805, + 0.08903533220291138 + ], + "credit_map": [ + 0.010864059440791607, + 0.017445312812924385, + 0.02299921214580536, + 0.03590363264083862, + 0.025784101337194443, + 0.015708087012171745, + 0.04022041708230972, + 0.039705730974674225, + 0.008531320840120316, + 0.0110554788261652, + 0.014447481371462345, + 0.020449209958314896, + 0.015320472419261932, + 0.009438849054276943, + 0.0250179972499609, + 0.03508938476443291, + 0.007203752174973488, + 0.012388094328343868, + 0.01233002357184887, + 0.030745597556233406, + 0.011764773167669773, + 0.005698316730558872, + 0.028486190363764763, + 0.019247733056545258, + 0.004862984176725149, + 0.004545966628938913, + 0.005414376966655254, + 0.010725674219429493, + 0.0056263660080730915, + 0.003841017372906208, + 0.011729639954864979, + 0.008821721188724041, + 0.004233320243656635, + 0.004813786130398512, + 0.007624586578458548, + 0.007787065114825964, + 0.00566741731017828, + 0.00403496203944087, + 0.009573512710630894, + 0.007028729189187288, + 0.004002620931714773, + 0.006309306249022484, + 0.006743414793163538, + 0.010662269778549671, + 0.005377560388296843, + 0.00307649583555758, + 0.010403594002127647, + 0.013262844644486904, + 0.005613813642412424, + 0.009840826503932476, + 0.00911363773047924, + 0.02024798095226288, + 0.007642136886715889, + 0.005602207500487566, + 0.01798621006309986, + 0.013129040598869324, + 0.00896843895316124, + 0.015439805574715137, + 0.014647994190454483, + 0.03193867206573486, + 0.010126895271241665, + 0.00595884257927537, + 0.0269609447568655, + 0.0195119921118021, + 0.004281909205019474, + 0.0053577967919409275, + 0.005824130028486252, + 0.005236609373241663, + 0.007342272438108921, + 0.003560729557648301, + 0.009287158027291298, + 0.005330192390829325, + 0.005408530123531818, + 0.011033737100660801, + 0.01089984830468893, + 0.014318721368908882, + 0.01088106818497181, + 0.0047249267809093, + 0.020091550424695015, + 0.011676955968141556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.016058865934610367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium close-up shot, taken from a level angle, focuses on a handsome young scholar seated in an antique-style study. Soft afternoon sunlight slants through the carved wooden window lattice beside him, casting distinct beams of light in the air, within which countless dust particles dance slowly. The scholar's long, jet-black hair is tied back with an elegant wooden hairpin. He wears a moon-white, wide-sleeved, cotton-linen Hanfu (Chinese traditional clothing), with pale blue embroidery adorning the lapels and cuffs. On the rustic desk before him lies an ancient, thread-bound book with yellowed pages, its vertical lines of traditional Chinese ink characters clearly visible. His expression is focused and calm, his slender right index finger carefully, steadily, and deliberately tracing the words across the page, slowly moving from right to left. As his finger moves, his lips part slightly, moving softly, as if silently and reverently reciting ancient verses. The warm light softly illuminates his profile and the precious ancient book, casting mottled light and shadows from the window panes on his clothes and desk. The whole scene is filled with a tranquil, focused and elegant bookish atmosphere, and time seems to have frozen at this moment.", + "seed": 32634, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_7/index_9335_seed_32634_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05818788390606642, + "vlm_logprob": -0.05818788390606642, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.05818788390606642, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16308584809303284 + ], + "opposite_logprob_tokens": [ + -1.9130858182907104 + ], + "logprob_reward": -0.16308584809303284, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16308584809303284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.803597629070282, + 0.0330883152782917, + 0.019186239689588547, + 0.052529238164424896, + 0.009134774096310139, + 0.012315495871007442, + 0.023803766816854477, + 0.016981011256575584, + 0.018281960859894753, + 0.01108158752322197 + ], + "credit_map": [ + 0.03642302751541138, + 0.08754456043243408, + 0.273571252822876, + 0.2831365168094635, + 0.043393395841121674, + 0.01951596327126026, + 0.03246418759226799, + 0.027548670768737793, + 0.003723610658198595, + 0.005624648183584213, + 0.0048515684902668, + 0.0038301933091133833, + 0.004283435642719269, + 0.002357662422582507, + 0.0053626541048288345, + 0.0030545371118932962, + 0.0019661744590848684, + 0.002880031941458583, + 0.0019621350802481174, + 0.004542392212897539, + 0.002083714585751295, + 0.0013865027576684952, + 0.0026526255533099174, + 0.0017126629827544093, + 0.0029646495822817087, + 0.005513808690011501, + 0.014282401651144028, + 0.01948053203523159, + 0.00301323807798326, + 0.0014437634963542223, + 0.00408547930419445, + 0.0017453614855185151, + 0.001062014838680625, + 0.001266280421987176, + 0.0020308787934482098, + 0.0017229869263246655, + 0.0007832882110960782, + 0.0004997282521799207, + 0.0010696111712604761, + 0.0006999848410487175, + 0.0015078336000442505, + 0.001584690879099071, + 0.001313297892920673, + 0.0026548218447715044, + 0.0015038865385577083, + 0.0010451325215399265, + 0.0015155412256717682, + 0.0011902900878340006, + 0.0022533114533871412, + 0.003490301314741373, + 0.003923024982213974, + 0.004742774181067944, + 0.0032397801987826824, + 0.0016151476884260774, + 0.0021747259888798, + 0.002364698564633727, + 0.0014701070031151175, + 0.002261232351884246, + 0.002954780822619796, + 0.004403229337185621, + 0.0019021021435037255, + 0.0009334942442364991, + 0.0016709055053070188, + 0.0013851592084392905, + 0.0025184822734445333, + 0.00235517555847764, + 0.00498083233833313, + 0.0031953323632478714, + 0.0014910431345924735, + 0.0009100079769268632, + 0.0015241100918501616, + 0.0013069752603769302, + 0.00048516178503632545, + 0.0015431436477229, + 0.0029345399234443903, + 0.00279719359241426, + 0.0007174699567258358, + 0.0008393252501264215, + 0.001063116593286395, + 0.000701635202858597 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.16308584809303284 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043087493628263474 + ], + "opposite_logprob_tokens": [ + -3.1680874824523926 + ], + "logprob_reward": -0.043087493628263474, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043087493628263474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3888894021511078, + 0.036163728684186935, + 0.07585409283638, + 0.03259769454598427, + 0.04003991186618805, + 0.034641340374946594, + 0.11298110336065292, + 0.10599657148122787, + 0.1436169594526291, + 0.02921925112605095 + ], + "credit_map": [ + 0.026087487116456032, + 0.03793536499142647, + 0.10925957560539246, + 0.1342368870973587, + 0.026562632992863655, + 0.013709377497434616, + 0.021733824163675308, + 0.019364241510629654, + 0.00338447536341846, + 0.005006766878068447, + 0.006141075864434242, + 0.004046118352562189, + 0.005758300889283419, + 0.003582573030143976, + 0.004376541823148727, + 0.0038678739219903946, + 0.009384183213114738, + 0.00970620010048151, + 0.008068454451858997, + 0.018293481320142746, + 0.007525499910116196, + 0.005616527050733566, + 0.010543394833803177, + 0.006716346833854914, + 0.0020019032526761293, + 0.004139414522796869, + 0.0037169072311371565, + 0.006987737491726875, + 0.00420119846239686, + 0.0039966213516891, + 0.004864283837378025, + 0.0026896256022155285, + 0.002263867063447833, + 0.005120104644447565, + 0.009953874163329601, + 0.012159754522144794, + 0.003151048207655549, + 0.0014526394661515951, + 0.003554410533979535, + 0.002384207211434841, + 0.004389998968690634, + 0.004517662338912487, + 0.0059460424818098545, + 0.005468986928462982, + 0.0037424610927700996, + 0.0029630730859935284, + 0.00391747010871768, + 0.0036956414114683867, + 0.016195736825466156, + 0.020703734830021858, + 0.01813456229865551, + 0.012628419324755669, + 0.015484358184039593, + 0.008511386811733246, + 0.01160419825464487, + 0.009718694724142551, + 0.009572776965796947, + 0.015668096020817757, + 0.02031274139881134, + 0.023164944723248482, + 0.010698340833187103, + 0.005424943286925554, + 0.012324386276304722, + 0.008830332197248936, + 0.01658070646226406, + 0.018210163339972496, + 0.04064641147851944, + 0.02795037440955639, + 0.011217424646019936, + 0.006475929170846939, + 0.011988276615738869, + 0.010547660291194916, + 0.0016713980585336685, + 0.0030889371410012245, + 0.002262580906972289, + 0.007630109786987305, + 0.003307977458462119, + 0.002430509077385068, + 0.0063969409093260765, + 0.002430796390399337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.043087493628263474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348400861024857 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348400861024857, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348400861024857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6654037833213806, + 0.02005138248205185, + 0.052396129816770554, + 0.13919158279895782, + 0.021318014711141586, + 0.010218555107712746, + 0.03118002600967884, + 0.01583975926041603, + 0.02583666518330574, + 0.018564144149422646 + ], + "credit_map": [ + 0.030367553234100342, + 0.06731466203927994, + 0.21827444434165955, + 0.2461739033460617, + 0.03515985980629921, + 0.015441094525158405, + 0.02818290889263153, + 0.02448931150138378, + 0.0018618031172081828, + 0.0027416066732257605, + 0.0031961321365088224, + 0.0025690223556011915, + 0.00274004228413105, + 0.0022481405176222324, + 0.0026736471336334944, + 0.002020986285060644, + 0.004268001765012741, + 0.007126097567379475, + 0.0076454333029687405, + 0.008316142484545708, + 0.007871599867939949, + 0.004392214585095644, + 0.006809098180383444, + 0.005967543460428715, + 0.007698940113186836, + 0.01429765485227108, + 0.035714276134967804, + 0.049927279353141785, + 0.008874468505382538, + 0.006019409280270338, + 0.0115240179002285, + 0.005135535728186369, + 0.0028293742798268795, + 0.0024344571866095066, + 0.003492262214422226, + 0.005264568608254194, + 0.0024650474078953266, + 0.001453918986953795, + 0.002144615398719907, + 0.0012337699299678206, + 0.0006940191960893571, + 0.00119020976126194, + 0.002101970836520195, + 0.0021712237503379583, + 0.0011875348864123225, + 0.0007139028748497367, + 0.001116273575462401, + 0.0010434198193252087, + 0.003227839944884181, + 0.004812310449779034, + 0.005588304251432419, + 0.0061174724251031876, + 0.003630235092714429, + 0.001722256070934236, + 0.0030861657578498125, + 0.0029954451601952314, + 0.0016143381362780929, + 0.0025651059113442898, + 0.0014060325920581818, + 0.002157645532861352, + 0.0022561971563845873, + 0.0014777097385376692, + 0.002098062075674534, + 0.002264668932184577, + 0.0022026533260941505, + 0.004034505691379309, + 0.007118753623217344, + 0.004122007172554731, + 0.002598126418888569, + 0.0014315114822238684, + 0.0022102235816419125, + 0.0021188841201364994, + 0.0012399303959682584, + 0.001418820582330227, + 0.0026346230879426003, + 0.002632430288940668, + 0.0030147384386509657, + 0.001606775214895606, + 0.004347868729382753, + 0.0016689581098034978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.026348400861024857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978435855358839 + ], + "opposite_logprob_tokens": [ + -4.63478422164917 + ], + "logprob_reward": -0.00978435855358839, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978435855358839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15083527565002441, + 0.07168900966644287, + 0.06960918754339218, + 0.42905473709106445, + 0.021120354533195496, + 0.051671430468559265, + 0.01932978443801403, + 0.012222680263221264, + 0.026791468262672424, + 0.14767605066299438 + ], + "credit_map": [ + 0.007696785964071751, + 0.017023948952555656, + 0.04415374994277954, + 0.03874196857213974, + 0.014583046548068523, + 0.010692586190998554, + 0.007931509055197239, + 0.010011670179665089, + 0.005839974153786898, + 0.007959220558404922, + 0.007462788838893175, + 0.008041306398808956, + 0.011527822352945805, + 0.0170229934155941, + 0.007887349463999271, + 0.005947559140622616, + 0.00453976122662425, + 0.0071621122770011425, + 0.016436316072940826, + 0.023115430027246475, + 0.005479330196976662, + 0.003941292874515057, + 0.005005369894206524, + 0.003929575905203819, + 0.025046315044164658, + 0.043736882507801056, + 0.1103394478559494, + 0.15737472474575043, + 0.026777058839797974, + 0.015540174208581448, + 0.034694939851760864, + 0.015545184724032879, + 0.0022428366355597973, + 0.0021572259720414877, + 0.005044815130531788, + 0.005833131726831198, + 0.0018889779457822442, + 0.0011459039524197578, + 0.0014521999983116984, + 0.0013552630553022027, + 0.0030598260927945375, + 0.006303333677351475, + 0.015658250078558922, + 0.017902199178934097, + 0.0036793635226786137, + 0.0011328430846333504, + 0.0024169047828763723, + 0.0015187058597803116, + 0.0019639097154140472, + 0.0031607693526893854, + 0.0023553387727588415, + 0.0033155870623886585, + 0.003119539935141802, + 0.0014963437570258975, + 0.0018664835952222347, + 0.00205181329511106, + 0.0013943529920652509, + 0.001791371963918209, + 0.0015189365949481726, + 0.001640114700421691, + 0.001710113836452365, + 0.0010535027831792831, + 0.0017437301576137543, + 0.0013705567689612508, + 0.0035803955979645252, + 0.0032986707519739866, + 0.005719906650483608, + 0.003919593524187803, + 0.0028808359056711197, + 0.001670640311203897, + 0.003005903447046876, + 0.0027155226562172174, + 0.008589634671807289, + 0.017446672543883324, + 0.03665979206562042, + 0.03601241484284401, + 0.01291992049664259, + 0.011615601368248463, + 0.016209080815315247, + 0.008222934789955616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.00978435855358839 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a handsome young scholar wearing a moon-white Hanfu with pale blue embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scholar's long jet-black hair tied back with an elegant wooden hairpin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scholar's slender right index finger tracing the words on an ancient thread-bound book with yellowed pages?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene set in an antique-style study with soft afternoon sunlight slanting through a carved wooden window lattice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a static medium close-up taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863331839442253 + ], + "opposite_logprob_tokens": [ + -3.048633337020874 + ], + "logprob_reward": -0.04863331839442253, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863331839442253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.713605523109436, + 0.0188915953040123, + 0.021817578002810478, + 0.048052169382572174, + 0.034656986594200134, + 0.022061754018068314, + 0.025811243802309036, + 0.016169413924217224, + 0.05571983382105827, + 0.043213896453380585 + ], + "credit_map": [ + 0.031659066677093506, + 0.07544755190610886, + 0.23820079863071442, + 0.2610437870025635, + 0.03801022097468376, + 0.015231981873512268, + 0.02896311692893505, + 0.02504904754459858, + 0.0018431282369419932, + 0.0037709209136664867, + 0.0023430560249835253, + 0.002969038439914584, + 0.002125062281265855, + 0.0027454907540231943, + 0.0015362841077148914, + 0.0015586161753162742, + 0.001686228672042489, + 0.0036661012563854456, + 0.0032684546895325184, + 0.002705526305362582, + 0.0022519524209201336, + 0.0016679108375683427, + 0.0036172231193631887, + 0.0029541824478656054, + 0.0036538264248520136, + 0.006492101587355137, + 0.014017466455698013, + 0.011099584400653839, + 0.003561150748282671, + 0.0023781917989253998, + 0.004473073873668909, + 0.002376770367845893, + 0.0062506962567567825, + 0.00589116383343935, + 0.004518516827374697, + 0.0056533170863986015, + 0.004541892558336258, + 0.002819160930812359, + 0.0025979653000831604, + 0.0023842735681682825, + 0.0013180410023778677, + 0.002812508726492524, + 0.006928746122866869, + 0.006714151240885258, + 0.0016045586671680212, + 0.0007596967625431716, + 0.0010467774700373411, + 0.0008772748406045139, + 0.002650610636919737, + 0.0043699839152395725, + 0.004686422646045685, + 0.0047294022515416145, + 0.003145905677229166, + 0.0022078363690525293, + 0.0024078390561044216, + 0.001613244297914207, + 0.0014861305244266987, + 0.0028176710475236177, + 0.0018893862143158913, + 0.0026863357052206993, + 0.0023424250539392233, + 0.001259867218323052, + 0.0021687024272978306, + 0.0015188958495855331, + 0.0056161158718168736, + 0.007266243454068899, + 0.014618614688515663, + 0.012963342480361462, + 0.004477739799767733, + 0.0023723007179796696, + 0.004529997706413269, + 0.003875474212691188, + 0.0039756824262440205, + 0.0038391179405152798, + 0.006553120911121368, + 0.008670436218380928, + 0.0038688876666128635, + 0.004319419153034687, + 0.008601884357631207, + 0.0033853475470095873 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.04863331839442253 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_tensor([1919, 1919])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_tensor([1919, 1919])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b4ee6231fe01f003ba7a1a0412654bc1c1a77468 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_tensor([1919, 1919])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static high-angle shot, as if from an elevated vantage point, gazes down upon an open, weathered hardcover book, spread upon a carpet of lush green moss. The book's dark brown leather cover has cracked and softened with age, its corners gently curled. The exposed pages are a creamy, yellowed parchment, thick and fibrous, covered in dense, elegant calligraphy in faded black ink. The book lies on a bed of fuzzy, damp-looking moss on the forest floor, surrounded by a few scattered, crisp autumn leaves and the dark, gnarled roots of a nearby tree. Dappled sunlight filters through the unseen treetops, casting shifting, warm patterns of light and shadow across the scene. A gentle breeze breathes life into the frame, causing the book's page in the upper right corner to subtly lift and flutter; it rises hesitantly, revealing the pages beneath for a brief moment, before settling back down with a soft, almost imperceptible rustling, creating an atmosphere of tranquility, contemplation, and a touch of magic, of quiet discovery.", + "seed": 23594, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_1919_seed_23594_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05401015556417406, + "vlm_logprob": -0.05401015556417406, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.05401015556417406, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02016192488372326 + ], + "opposite_logprob_tokens": [ + -5.145162105560303 + ], + "logprob_reward": -0.02016192488372326, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.02016192488372326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28106456995010376, + 0.06771613657474518, + 0.08690877258777618, + 0.03664228320121765, + 0.1374289095401764, + 0.11012134701013565, + 0.09264516830444336, + 0.13800296187400818, + 0.02654729038476944, + 0.022922508418560028 + ], + "credit_map": [ + 0.02366969920694828, + 0.05060339719057083, + 0.018139773979783058, + 0.030215324833989143, + 0.07338393479585648, + 0.040314000099897385, + 0.018764574080705643, + 0.025973878800868988, + 0.004409721586853266, + 0.00845095794647932, + 0.005053761415183544, + 0.014764869585633278, + 0.015453743748366833, + 0.00850756000727415, + 0.00562277901917696, + 0.00545274605974555, + 0.0076053855009377, + 0.00971465464681387, + 0.006799873895943165, + 0.009203634224832058, + 0.015825917944312096, + 0.01601882092654705, + 0.01375479158014059, + 0.007985706441104412, + 0.004611433017998934, + 0.004228564910590649, + 0.003278500633314252, + 0.005669517442584038, + 0.003671559737995267, + 0.0032895109616219997, + 0.007910161279141903, + 0.00398304034024477, + 0.01359165646135807, + 0.01934092678129673, + 0.012985510751605034, + 0.020538678392767906, + 0.018816007301211357, + 0.02269611693918705, + 0.016193652525544167, + 0.013266373425722122, + 0.016269225627183914, + 0.01592904143035412, + 0.01080701220780611, + 0.010482427664101124, + 0.015991676598787308, + 0.017538700252771378, + 0.010913425125181675, + 0.01218984555453062, + 0.013069105334579945, + 0.01595970056951046, + 0.01054311916232109, + 0.012144110165536404, + 0.013978935778141022, + 0.008279914036393166, + 0.010553151369094849, + 0.00811714120209217, + 0.015688586980104446, + 0.023193996399641037, + 0.00956078339368105, + 0.017809800803661346, + 0.02913135662674904, + 0.010590578429400921, + 0.020900152623653412, + 0.011127719655632973, + 0.002616363810375333, + 0.002940970007330179, + 0.0032604271546006203, + 0.006435741670429707, + 0.0034012366086244583, + 0.002330166520550847, + 0.003468491602689028, + 0.002093895338475704, + 0.0030194728169590235, + 0.004147404804825783, + 0.002100048353895545, + 0.003575438167899847, + 0.0036769197322428226, + 0.002039822982624173, + 0.0022686668671667576, + 0.0020947367884218693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02016192488372326 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025128712877631187 + ], + "opposite_logprob_tokens": [ + -6.0025129318237305 + ], + "logprob_reward": -0.0025128712877631187, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025128712877631187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28364884853363037, + 0.09060094505548477, + 0.0561530701816082, + 0.05601667985320091, + 0.1817803531885147, + 0.06242837756872177, + 0.06376529484987259, + 0.08649734407663345, + 0.06003817915916443, + 0.059070929884910583 + ], + "credit_map": [ + 0.024249037727713585, + 0.03583814203739166, + 0.017781201750040054, + 0.03984127193689346, + 0.06505628675222397, + 0.054269369691610336, + 0.022098297253251076, + 0.02451523207128048, + 0.007902510464191437, + 0.014414137229323387, + 0.00591467646881938, + 0.014519755728542805, + 0.010503293015062809, + 0.017922816798090935, + 0.01081864908337593, + 0.008605105802416801, + 0.007441550027579069, + 0.008441961370408535, + 0.0070973169058561325, + 0.007703032344579697, + 0.005952861160039902, + 0.008774825371801853, + 0.005317964591085911, + 0.00542356027290225, + 0.0069799404591321945, + 0.007314085029065609, + 0.00582907535135746, + 0.011242068372666836, + 0.006164983846247196, + 0.004273951519280672, + 0.008012750186026096, + 0.0061998250894248486, + 0.02350788190960884, + 0.02872864529490471, + 0.014544118195772171, + 0.030546890571713448, + 0.022939713671803474, + 0.026733826845884323, + 0.013703669421374798, + 0.021075602620840073, + 0.009043464437127113, + 0.007436641026288271, + 0.005101334303617477, + 0.012714856304228306, + 0.007685016840696335, + 0.005693383980542421, + 0.007916586473584175, + 0.006837093736976385, + 0.007576082833111286, + 0.011571206152439117, + 0.007356265559792519, + 0.009216481819748878, + 0.006663487292826176, + 0.006824601907283068, + 0.008964489214122295, + 0.005592685658484697, + 0.008015697821974754, + 0.013680347241461277, + 0.006300021894276142, + 0.012478702701628208, + 0.02039291150867939, + 0.0066315061412751675, + 0.01171216182410717, + 0.0072859907522797585, + 0.008041189052164555, + 0.008906645700335503, + 0.007525262888520956, + 0.011053825728595257, + 0.006364735774695873, + 0.006661747582256794, + 0.006367495749145746, + 0.005117273889482021, + 0.0074339816346764565, + 0.013212152756750584, + 0.00532922800630331, + 0.010061665438115597, + 0.008515161462128162, + 0.004915681667625904, + 0.0045721628703176975, + 0.005030896980315447 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0025128712877631187 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036221628542989492 + ], + "opposite_logprob_tokens": [ + -5.628622055053711 + ], + "logprob_reward": -0.0036221628542989492, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036221628542989492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1900971531867981, + 0.08310706168413162, + 0.11332360655069351, + 0.057306356728076935, + 0.11702454835176468, + 0.19917254149913788, + 0.06614793837070465, + 0.07378920167684555, + 0.05647389590740204, + 0.04355773329734802 + ], + "credit_map": [ + 0.014289447106420994, + 0.0408065989613533, + 0.013027172535657883, + 0.032369449734687805, + 0.031142069026827812, + 0.02786709927022457, + 0.013770109973847866, + 0.016825208440423012, + 0.008865376003086567, + 0.012259910814464092, + 0.007410886697471142, + 0.011170575395226479, + 0.013947846367955208, + 0.011152716353535652, + 0.01012486219406128, + 0.008174880407750607, + 0.014915507286787033, + 0.023701736703515053, + 0.011938363313674927, + 0.014800912700593472, + 0.014948458410799503, + 0.014174510724842548, + 0.009038141928613186, + 0.009805984795093536, + 0.00524258054792881, + 0.0065551516599953175, + 0.006373783573508263, + 0.01042253989726305, + 0.007252952083945274, + 0.0045846495777368546, + 0.009430671110749245, + 0.007444027811288834, + 0.015544856898486614, + 0.020424440503120422, + 0.010759720578789711, + 0.020128775388002396, + 0.012776394374668598, + 0.0144889447838068, + 0.009317141957581043, + 0.013584278523921967, + 0.025829438120126724, + 0.023363130167126656, + 0.01854211464524269, + 0.02470763400197029, + 0.030301783233880997, + 0.03943817317485809, + 0.017084719613194466, + 0.01990554668009281, + 0.008406431414186954, + 0.011996167711913586, + 0.007067282218486071, + 0.009617703035473824, + 0.01065834704786539, + 0.005756690166890621, + 0.007394673302769661, + 0.00525064067915082, + 0.007126410957425833, + 0.01403774879872799, + 0.005365839693695307, + 0.008347120136022568, + 0.017701033502817154, + 0.005262034013867378, + 0.01040156651288271, + 0.00554744154214859, + 0.007255156058818102, + 0.010236070491373539, + 0.00616937130689621, + 0.00990639440715313, + 0.006922369822859764, + 0.005907424725592136, + 0.0055123888887465, + 0.0045647211372852325, + 0.005621955730021, + 0.010378992184996605, + 0.004083465784788132, + 0.00657137343659997, + 0.005231080576777458, + 0.00374716822989285, + 0.004092467948794365, + 0.0038312245160341263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.0036221628542989492 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2790934145450592, + 0.07258728891611099, + 0.16282063722610474, + 0.03594290092587471, + 0.11188380420207977, + 0.12451502680778503, + 0.06001269444823265, + 0.08569515496492386, + 0.021991493180394173, + 0.04545764625072479 + ], + "credit_map": [ + 0.026128577068448067, + 0.03385239094495773, + 0.02212291583418846, + 0.034237831830978394, + 0.055326126515865326, + 0.05770585313439369, + 0.024656306952238083, + 0.025063376873731613, + 0.005039452109485865, + 0.010172940790653229, + 0.005034420173615217, + 0.012157808989286423, + 0.013269979506731033, + 0.008750521577894688, + 0.01128877978771925, + 0.006873375736176968, + 0.01696319878101349, + 0.017201121896505356, + 0.016192298382520676, + 0.021388379856944084, + 0.027073753997683525, + 0.021220708265900612, + 0.026848381385207176, + 0.015932761132717133, + 0.003789902664721012, + 0.006420164834707975, + 0.0036041527055203915, + 0.007191767916083336, + 0.004476162139326334, + 0.0027045707684010267, + 0.0040956218726933, + 0.0036605545319616795, + 0.013449805788695812, + 0.016642557457089424, + 0.009468013420701027, + 0.013991599902510643, + 0.016368715092539787, + 0.009306823834776878, + 0.019211681559681892, + 0.013444595038890839, + 0.01922963373363018, + 0.012805123813450336, + 0.01134126540273428, + 0.01921059563755989, + 0.01823636330664158, + 0.014156839810311794, + 0.018753236159682274, + 0.01078195869922638, + 0.006524847354739904, + 0.010684828273952007, + 0.00764654902741313, + 0.007971490733325481, + 0.007431952282786369, + 0.006507657468318939, + 0.007890126667916775, + 0.005355235654860735, + 0.0078507075086236, + 0.010052754543721676, + 0.00680620688945055, + 0.012927165254950523, + 0.017318278551101685, + 0.006802001968026161, + 0.016610657796263695, + 0.00732737872749567, + 0.002978276228532195, + 0.0036929165944457054, + 0.002579217776656151, + 0.0033527954947203398, + 0.0026027041021734476, + 0.0027554102707654238, + 0.002144459867849946, + 0.0018857114482671022, + 0.006007753778249025, + 0.009863710030913353, + 0.004435739479959011, + 0.007838709279894829, + 0.006722198333591223, + 0.0034235536586493254, + 0.0032075317576527596, + 0.0039584459736943245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01819257251918316 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2255612462759018 + ], + "opposite_logprob_tokens": [ + -1.600561261177063 + ], + "logprob_reward": -0.2255612462759018, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2255612462759018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11631808429956436, + 0.059806425124406815, + 0.08109058439731598, + 0.0381283275783062, + 0.08912118524312973, + 0.4877239763736725, + 0.02222738415002823, + 0.03542541339993477, + 0.03634759038686752, + 0.0338110625743866 + ], + "credit_map": [ + 0.010437805205583572, + 0.021511243656277657, + 0.0071802204474806786, + 0.011990918777883053, + 0.02440713532269001, + 0.02203323133289814, + 0.008673262782394886, + 0.010084262117743492, + 0.0078528868034482, + 0.008269614540040493, + 0.006352761294692755, + 0.005705276504158974, + 0.010876072570681572, + 0.008857452310621738, + 0.005690467543900013, + 0.006201892625540495, + 0.00923593807965517, + 0.011086709797382355, + 0.008119964972138405, + 0.010621738620102406, + 0.012731284834444523, + 0.01365612167865038, + 0.007527354173362255, + 0.008111466653645039, + 0.0029243917670100927, + 0.00645778002217412, + 0.003647506469860673, + 0.009194419719278812, + 0.005936840083450079, + 0.002995012793689966, + 0.00405375799164176, + 0.002918616868555546, + 0.010551336221396923, + 0.017802445217967033, + 0.0081096813082695, + 0.014533501118421555, + 0.012472969479858875, + 0.00839525181800127, + 0.009346853010356426, + 0.007909141480922699, + 0.06428547203540802, + 0.05374382808804512, + 0.04792199656367302, + 0.07217711955308914, + 0.07029693573713303, + 0.09506436437368393, + 0.039266232401132584, + 0.04496803507208824, + 0.002265315968543291, + 0.003993642050772905, + 0.0018430064665153623, + 0.0032147932797670364, + 0.004569533746689558, + 0.0018372569466009736, + 0.0027482989244163036, + 0.0017555352533236146, + 0.004268253222107887, + 0.007535418029874563, + 0.003678275039419532, + 0.005691574886441231, + 0.004745897836983204, + 0.0029722237959504128, + 0.0032584085129201412, + 0.0032753616105765104, + 0.003289120038971305, + 0.00943570677191019, + 0.002849575597792864, + 0.005092116538435221, + 0.00746819656342268, + 0.0025613936595618725, + 0.0034393565729260445, + 0.0022121239453554153, + 0.003975432366132736, + 0.008419915102422237, + 0.0033583384938538074, + 0.005916357971727848, + 0.004182747099548578, + 0.002347835572436452, + 0.0026096291840076447, + 0.0030008063185960054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.2255612462759018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static high-angle shot, as if from an elevated vantage point, gazes down upon an open, weathered hardcover book, spread upon a carpet of lush green moss. The book's dark brown leather cover has cracked and softened with age, its corners gently curled. The exposed pages are a creamy, yellowed parchment, thick and fibrous, covered in dense, elegant calligraphy in faded black ink. The book lies on a bed of fuzzy, damp-looking moss on the forest floor, surrounded by a few scattered, crisp autumn leaves and the dark, gnarled roots of a nearby tree. Dappled sunlight filters through the unseen treetops, casting shifting, warm patterns of light and shadow across the scene. A gentle breeze breathes life into the frame, causing the book's page in the upper right corner to subtly lift and flutter; it rises hesitantly, revealing the pages beneath for a brief moment, before settling back down with a soft, almost imperceptible rustling, creating an atmosphere of tranquility, contemplation, and a touch of magic, of quiet discovery.", + "seed": 23595, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_1919_seed_23595_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04038344044238329, + "vlm_logprob": -0.04038344044238329, + "vlm_yes_no_margin": 4.400000023841858, + "vlm_token_credit": -0.04038344044238329, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02953319624066353 + ], + "opposite_logprob_tokens": [ + -5.779533386230469 + ], + "logprob_reward": -0.02953319624066353, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.02953319624066353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11627401411533356, + 0.10892418771982193, + 0.19574269652366638, + 0.08674370497465134, + 0.18749147653579712, + 0.03421461582183838, + 0.13700982928276062, + 0.025417350232601166, + 0.06264426559209824, + 0.04553781822323799 + ], + "credit_map": [ + 0.01567753590643406, + 0.011176113970577717, + 0.019960816949605942, + 0.014983873814344406, + 0.012643934227526188, + 0.020882654935121536, + 0.010735983029007912, + 0.010213097557425499, + 0.010963576845824718, + 0.012617767788469791, + 0.014073310419917107, + 0.011713502928614616, + 0.012231726199388504, + 0.02156025916337967, + 0.014812166802585125, + 0.010951870121061802, + 0.014953852631151676, + 0.024787120521068573, + 0.01687469147145748, + 0.024427438154816628, + 0.03797256201505661, + 0.031542424112558365, + 0.030969666317105293, + 0.014214946888387203, + 0.006732469890266657, + 0.016251057386398315, + 0.007650352548807859, + 0.00874328427016735, + 0.019231747835874557, + 0.007192963734269142, + 0.013460354879498482, + 0.0074814725667238235, + 0.012887961231172085, + 0.023403124883770943, + 0.007240145467221737, + 0.004834475927054882, + 0.07513618469238281, + 0.01989400014281273, + 0.014103809371590614, + 0.029991772025823593, + 0.003285788930952549, + 0.0032022001687437296, + 0.004176174756139517, + 0.0022411278914660215, + 0.005226619075983763, + 0.009676745161414146, + 0.003151574870571494, + 0.0032543851993978024, + 0.008596600964665413, + 0.04418753832578659, + 0.003823432605713606, + 0.0052567580714821815, + 0.013706968165934086, + 0.012002632953226566, + 0.007653756067156792, + 0.04178215563297272, + 0.0018735723569989204, + 0.003824046114459634, + 0.0015765250427648425, + 0.0033923606388270855, + 0.0029609035700559616, + 0.004283218644559383, + 0.003099378664046526, + 0.004407345317304134, + 0.00494935130700469, + 0.019300177693367004, + 0.003884518053382635, + 0.003608585335314274, + 0.007362255826592445, + 0.005376812536269426, + 0.004098969046026468, + 0.014063593931496143, + 0.004769595805555582, + 0.004952253773808479, + 0.006666796747595072, + 0.004889804869890213, + 0.009504233486950397, + 0.005906916689127684, + 0.0056157661601901054, + 0.0032324513886123896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.02953319624066353 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067657907493412495 + ], + "opposite_logprob_tokens": [ + -5.006765842437744 + ], + "logprob_reward": -0.0067657907493412495, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067657907493412495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10898900777101517, + 0.18528658151626587, + 0.14849959313869476, + 0.04833664372563362, + 0.03668589890003204, + 0.0436415858566761, + 0.04283914342522621, + 0.02592376060783863, + 0.26344719529151917, + 0.09635055065155029 + ], + "credit_map": [ + 0.015867771580815315, + 0.011664272285997868, + 0.01018902100622654, + 0.01572616584599018, + 0.01434335671365261, + 0.012007082812488079, + 0.019030645489692688, + 0.010160688310861588, + 0.024604666978120804, + 0.020877040922641754, + 0.022836238145828247, + 0.02326844446361065, + 0.02125120908021927, + 0.03559022396802902, + 0.021564152091741562, + 0.015294600278139114, + 0.008885272778570652, + 0.022091325372457504, + 0.011053353548049927, + 0.025317508727312088, + 0.0162353515625, + 0.007047554012387991, + 0.04367605596780777, + 0.014193180948495865, + 0.004807499703019857, + 0.00590800354257226, + 0.006485114339739084, + 0.00612858310341835, + 0.009098438546061516, + 0.0043383268639445305, + 0.006906501483172178, + 0.00466417521238327, + 0.0029620446730405092, + 0.006075248587876558, + 0.0030941362492740154, + 0.002195516601204872, + 0.009595225565135479, + 0.005666194949299097, + 0.004111256916075945, + 0.0029862723313272, + 0.0026230481453239918, + 0.006993702147156, + 0.0036837891675531864, + 0.0023722227197140455, + 0.004113386385142803, + 0.004488870035856962, + 0.004296330269426107, + 0.0150702353566885, + 0.0029775011353194714, + 0.011204861104488373, + 0.003176196711137891, + 0.0031623265240341425, + 0.004352486226707697, + 0.005547588691115379, + 0.00340823270380497, + 0.009009945206344128, + 0.0023257588036358356, + 0.004209005739539862, + 0.0031245388090610504, + 0.002997924806550145, + 0.0035861374344676733, + 0.002820234280079603, + 0.0032817909959703684, + 0.00357837020419538, + 0.014314761385321617, + 0.09615463763475418, + 0.008595569059252739, + 0.010019584558904171, + 0.02360033616423607, + 0.019099734723567963, + 0.011378624476492405, + 0.08028394728899002, + 0.00817961897701025, + 0.012063074856996536, + 0.014297634363174438, + 0.010935436002910137, + 0.019940946251153946, + 0.014257130213081837, + 0.010178490541875362, + 0.00649822223931551 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 2 + ], + "selected_reward": -0.0067657907493412495 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004643371794372797 + ], + "opposite_logprob_tokens": [ + -5.379643440246582 + ], + "logprob_reward": -0.004643371794372797, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004643371794372797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13204236328601837, + 0.060628462582826614, + 0.33793649077415466, + 0.05259126052260399, + 0.0320340134203434, + 0.025896264240145683, + 0.01867239736020565, + 0.0237568486481905, + 0.23949019610881805, + 0.07695171982049942 + ], + "credit_map": [ + 0.01712966151535511, + 0.01613519713282585, + 0.013524739071726799, + 0.015606471337378025, + 0.01696188934147358, + 0.017466461285948753, + 0.02392999827861786, + 0.011287938803434372, + 0.0071218558587133884, + 0.007434992119669914, + 0.00639493390917778, + 0.009248444810509682, + 0.007571878377348185, + 0.009942288510501385, + 0.008581272326409817, + 0.004332794342190027, + 0.022992992773652077, + 0.043041959404945374, + 0.020658718422055244, + 0.043429531157016754, + 0.06845542043447495, + 0.043752823024988174, + 0.07007164508104324, + 0.02553337626159191, + 0.0035004892852157354, + 0.0057772002182900906, + 0.0060180616565048695, + 0.007000767160207033, + 0.010213468223810196, + 0.006286948453634977, + 0.0065490747801959515, + 0.007245256099849939, + 0.0025081089697778225, + 0.0040186014957726, + 0.0022994268219918013, + 0.0020586014725267887, + 0.005077896174043417, + 0.00537836691364646, + 0.004035176243633032, + 0.00665783416479826, + 0.00188455858733505, + 0.002905359724536538, + 0.0023235478438436985, + 0.001614572829566896, + 0.003553631715476513, + 0.005615258123725653, + 0.0028997694607824087, + 0.005099564790725708, + 0.001932539977133274, + 0.002326906891539693, + 0.002251547761261463, + 0.0025272693019360304, + 0.002121398923918605, + 0.0028424253687262535, + 0.0029061506502330303, + 0.0017641590675339103, + 0.002061747247353196, + 0.003250510198995471, + 0.0027508079074323177, + 0.0026105644647032022, + 0.0034531557466834784, + 0.002658209530636668, + 0.0034447317011654377, + 0.0035271216183900833, + 0.012870974838733673, + 0.09102405607700348, + 0.006853730417788029, + 0.008388807065784931, + 0.021729251369833946, + 0.016685202717781067, + 0.00940600223839283, + 0.07253218442201614, + 0.006664936430752277, + 0.011457591317594051, + 0.006478570401668549, + 0.005273323506116867, + 0.026788216084241867, + 0.009428337216377258, + 0.006265833508223295, + 0.004594904836267233 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.004643371794372797 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018191752955317497 + ], + "opposite_logprob_tokens": [ + -4.018191814422607 + ], + "logprob_reward": -0.018191752955317497, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018191752955317497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08646953105926514, + 0.19892127811908722, + 0.20575504004955292, + 0.047009825706481934, + 0.04087204113602638, + 0.02822127193212509, + 0.023186106234788895, + 0.010925236158072948, + 0.11122601479291916, + 0.24741366505622864 + ], + "credit_map": [ + 0.01363676693290472, + 0.009301898069679737, + 0.010978040285408497, + 0.008718104101717472, + 0.009395197965204716, + 0.011362116783857346, + 0.013331727124750614, + 0.009745665825903416, + 0.02261248417198658, + 0.02315692976117134, + 0.024571875110268593, + 0.02239924669265747, + 0.022742358967661858, + 0.04134601727128029, + 0.02504718117415905, + 0.017045175656676292, + 0.01474462728947401, + 0.02552986517548561, + 0.017516082152724266, + 0.02918805181980133, + 0.03830329701304436, + 0.031424395740032196, + 0.033895500004291534, + 0.015153195708990097, + 0.003274674993008375, + 0.006204579956829548, + 0.003741490188986063, + 0.004602642264217138, + 0.007911463268101215, + 0.006271603051573038, + 0.005703528877347708, + 0.009299836121499538, + 0.00337984343059361, + 0.005992596037685871, + 0.0021141048055142164, + 0.0020412709563970566, + 0.011449215933680534, + 0.00848485343158245, + 0.004055513069033623, + 0.0033546355552971363, + 0.002278657630085945, + 0.0033866195008158684, + 0.004454542882740498, + 0.0019470403203740716, + 0.0038965418934822083, + 0.004393299110233784, + 0.0027524360921233892, + 0.005112132057547569, + 0.0020155981183052063, + 0.003942538984119892, + 0.0020220940932631493, + 0.0017845623660832644, + 0.0023404399398714304, + 0.005820761900395155, + 0.002346289809793234, + 0.002913818461820483, + 0.001302190124988556, + 0.0010074148885905743, + 0.0013408511877059937, + 0.0012766947038471699, + 0.0015286840498447418, + 0.002026474568992853, + 0.0014383692760020494, + 0.0010045559611171484, + 0.006914341822266579, + 0.038252994418144226, + 0.004366137553006411, + 0.0037393984384834766, + 0.013542675413191319, + 0.007899878546595573, + 0.004145245999097824, + 0.032365329563617706, + 0.02184220775961876, + 0.03708064183592796, + 0.020590173080563545, + 0.016799001023173332, + 0.08750982582569122, + 0.028459466993808746, + 0.0208940077573061, + 0.014238319359719753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.018191752955317497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278309047222137 + ], + "opposite_logprob_tokens": [ + -2.0177831649780273 + ], + "logprob_reward": -0.14278309047222137, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14278309047222137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09450440108776093, + 0.24377848207950592, + 0.10584854334592819, + 0.17515182495117188, + 0.09114334732294083, + 0.056163255125284195, + 0.10073430091142654, + 0.027505049481987953, + 0.07222020626068115, + 0.032950520515441895 + ], + "credit_map": [ + 0.017352323979139328, + 0.008891765028238297, + 0.008846957236528397, + 0.010864939540624619, + 0.011942889541387558, + 0.014668667688965797, + 0.011914861388504505, + 0.010022012516856194, + 0.029228920117020607, + 0.027635671198368073, + 0.03498964011669159, + 0.028576547279953957, + 0.026056399568915367, + 0.04768071323633194, + 0.02696114592254162, + 0.022649457678198814, + 0.006305602379143238, + 0.0173382218927145, + 0.009746633470058441, + 0.024888064712285995, + 0.013255787082016468, + 0.006049067713320255, + 0.02086767926812172, + 0.007397490553557873, + 0.012446166016161442, + 0.033237770199775696, + 0.012254048138856888, + 0.021057650446891785, + 0.04468773305416107, + 0.013016652315855026, + 0.021437352523207664, + 0.01701447181403637, + 0.007149164564907551, + 0.011831997893750668, + 0.00394731480628252, + 0.00293237273581326, + 0.03430260345339775, + 0.011344006285071373, + 0.007149212062358856, + 0.012486682273447514, + 0.004967974498867989, + 0.007622720208019018, + 0.007967502810060978, + 0.004916849546134472, + 0.007638074457645416, + 0.006767105311155319, + 0.00649283779785037, + 0.009790199808776379, + 0.006043099798262119, + 0.03298160433769226, + 0.0035418907646089792, + 0.004691339563578367, + 0.010207145474851131, + 0.007155563682317734, + 0.0058455332182347775, + 0.030268140137195587, + 0.0021225912496447563, + 0.003400270827114582, + 0.0022159817162901163, + 0.0034140031784772873, + 0.003200773848220706, + 0.004663284868001938, + 0.0030079311691224575, + 0.005480215419083834, + 0.005612621083855629, + 0.01650404930114746, + 0.0034592251759022474, + 0.00337356049567461, + 0.016863645985722542, + 0.007147665601223707, + 0.004169199150055647, + 0.015090244822204113, + 0.002962111961096525, + 0.004977277480065823, + 0.0019915469456464052, + 0.002016643527895212, + 0.012604794465005398, + 0.0037604907993227243, + 0.002510106423869729, + 0.0021275547333061695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.14278309047222137 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static high-angle shot, as if from an elevated vantage point, gazes down upon an open, weathered hardcover book, spread upon a carpet of lush green moss. The book's dark brown leather cover has cracked and softened with age, its corners gently curled. The exposed pages are a creamy, yellowed parchment, thick and fibrous, covered in dense, elegant calligraphy in faded black ink. The book lies on a bed of fuzzy, damp-looking moss on the forest floor, surrounded by a few scattered, crisp autumn leaves and the dark, gnarled roots of a nearby tree. Dappled sunlight filters through the unseen treetops, casting shifting, warm patterns of light and shadow across the scene. A gentle breeze breathes life into the frame, causing the book's page in the upper right corner to subtly lift and flutter; it rises hesitantly, revealing the pages beneath for a brief moment, before settling back down with a soft, almost imperceptible rustling, creating an atmosphere of tranquility, contemplation, and a touch of magic, of quiet discovery.", + "seed": 23596, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_1919_seed_23596_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04732112502679229, + "vlm_logprob": -0.04732112502679229, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.04732112502679229, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02072593756020069 + ], + "opposite_logprob_tokens": [ + -5.395725727081299 + ], + "logprob_reward": -0.02072593756020069, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.02072593756020069, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3107313811779022, + 0.06949210166931152, + 0.2582913637161255, + 0.08337199687957764, + 0.051967211067676544, + 0.05001828819513321, + 0.040760524570941925, + 0.061733271926641464, + 0.04697999730706215, + 0.02665381319820881 + ], + "credit_map": [ + 0.05608722195029259, + 0.031636033207178116, + 0.05061500519514084, + 0.03921748697757721, + 0.025684380903840065, + 0.037925027310848236, + 0.023446982726454735, + 0.046119265258312225, + 0.009085429832339287, + 0.01189334411174059, + 0.0052636610344052315, + 0.010890302248299122, + 0.009530276991426945, + 0.010729723609983921, + 0.0042756907641887665, + 0.007823674939572811, + 0.026245223358273506, + 0.02866034395992756, + 0.013300005346536636, + 0.0486418716609478, + 0.02302359975874424, + 0.031056201085448265, + 0.02936519682407379, + 0.057998958975076675, + 0.009573502466082573, + 0.011432626284658909, + 0.005001913756132126, + 0.008631170727312565, + 0.016262531280517578, + 0.009072320535779, + 0.015322436578571796, + 0.008075496181845665, + 0.00747525691986084, + 0.003886534832417965, + 0.004632136318832636, + 0.005787375848740339, + 0.007297074887901545, + 0.006498894654214382, + 0.008441775105893612, + 0.007948166690766811, + 0.006559232249855995, + 0.009325375780463219, + 0.003033257322385907, + 0.005666332319378853, + 0.006789138074964285, + 0.006015541031956673, + 0.005819434765726328, + 0.006809978745877743, + 0.003940504509955645, + 0.010147510096430779, + 0.0027853832580149174, + 0.005131073761731386, + 0.0034271501936018467, + 0.0070575024001300335, + 0.0035127161536365747, + 0.0047586839646101, + 0.003442358458414674, + 0.006531066261231899, + 0.004043687134981155, + 0.021741075441241264, + 0.004781376104801893, + 0.006004402879625559, + 0.005553930066525936, + 0.009635375812649727, + 0.003352011786773801, + 0.004547674208879471, + 0.0024262566585093737, + 0.008463749662041664, + 0.00392018910497427, + 0.013272910378873348, + 0.0029775938019156456, + 0.008019612170755863, + 0.0022771351505070925, + 0.0054002138786017895, + 0.0017903097905218601, + 0.00450263312086463, + 0.0023608659394085407, + 0.0031699344981461763, + 0.001709195552393794, + 0.005443528760224581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.02072593756020069 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008671249262988567 + ], + "opposite_logprob_tokens": [ + -4.758671283721924 + ], + "logprob_reward": -0.008671249262988567, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008671249262988567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12146260589361191, + 0.1321130394935608, + 0.07082750648260117, + 0.0956975519657135, + 0.07566910237073898, + 0.12749022245407104, + 0.07848776131868362, + 0.1762082576751709, + 0.08883324265480042, + 0.03321073576807976 + ], + "credit_map": [ + 0.018538257107138634, + 0.014549589715898037, + 0.01470484770834446, + 0.017553186044096947, + 0.016351722180843353, + 0.013745385222136974, + 0.011014090850949287, + 0.015005535446107388, + 0.013739822432398796, + 0.01497520413249731, + 0.009211658500134945, + 0.04118294268846512, + 0.008979054167866707, + 0.013502378948032856, + 0.006785823032259941, + 0.023736147210001945, + 0.007026459090411663, + 0.017098288983106613, + 0.0027683486696332693, + 0.016551950946450233, + 0.007809002418071032, + 0.006028574891388416, + 0.007496955338865519, + 0.006047923117876053, + 0.010040233843028545, + 0.008747012354433537, + 0.00563851511105895, + 0.011890485882759094, + 0.011680709198117256, + 0.010387582704424858, + 0.011431388556957245, + 0.02588162198662758, + 0.0096540292724967, + 0.006656062789261341, + 0.005743968300521374, + 0.008864374831318855, + 0.011107996106147766, + 0.008828786201775074, + 0.011907621286809444, + 0.012906267307698727, + 0.021177547052502632, + 0.019051706418395042, + 0.009563260711729527, + 0.0125214122235775, + 0.011580553837120533, + 0.016582658514380455, + 0.01921343244612217, + 0.017799651250243187, + 0.009177453815937042, + 0.0133901322260499, + 0.004551141522824764, + 0.00741913728415966, + 0.013753370381891727, + 0.011471396312117577, + 0.007017423398792744, + 0.011707703582942486, + 0.014562835916876793, + 0.028792405501008034, + 0.011312194168567657, + 0.05447385087609291, + 0.016320671886205673, + 0.016765469685196877, + 0.016161009669303894, + 0.01781981624662876, + 0.00602680491283536, + 0.009721996262669563, + 0.005388554651290178, + 0.02452055923640728, + 0.00774808507412672, + 0.020984601229429245, + 0.006950903218239546, + 0.007491734344512224, + 0.0027572130784392357, + 0.00343426875770092, + 0.0021683606319129467, + 0.009648412466049194, + 0.002422845456749201, + 0.0076063042506575584, + 0.0026938789524137974, + 0.0024794507771730423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.008671249262988567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746016908437014 + ], + "opposite_logprob_tokens": [ + -5.0067458152771 + ], + "logprob_reward": -0.006746016908437014, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746016908437014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15145175158977509, + 0.06034139171242714, + 0.36898481845855713, + 0.06786645948886871, + 0.07593555748462677, + 0.05074695870280266, + 0.07317597419023514, + 0.08253690600395203, + 0.039964593946933746, + 0.02899554930627346 + ], + "credit_map": [ + 0.02816217951476574, + 0.016215987503528595, + 0.024371307343244553, + 0.021960875019431114, + 0.015873437747359276, + 0.015329686924815178, + 0.013610602356493473, + 0.015927687287330627, + 0.006702446844428778, + 0.008486421778798103, + 0.004701073281466961, + 0.016732342541217804, + 0.006612482015043497, + 0.006058167200535536, + 0.0033006458543241024, + 0.007747810333967209, + 0.038183070719242096, + 0.04737066477537155, + 0.018633877858519554, + 0.0713447630405426, + 0.03067588247358799, + 0.042340006679296494, + 0.03938015177845955, + 0.0810563936829567, + 0.006678860168904066, + 0.005562575068324804, + 0.0036442324053496122, + 0.00895864050835371, + 0.008811093866825104, + 0.008704389445483685, + 0.007652760483324528, + 0.017853910103440285, + 0.012180143035948277, + 0.005893556401133537, + 0.005975733045488596, + 0.006200701463967562, + 0.006793997250497341, + 0.009281004779040813, + 0.013944671489298344, + 0.015665750950574875, + 0.007660268805921078, + 0.0063051399774849415, + 0.0045568919740617275, + 0.004191313870251179, + 0.005670969840139151, + 0.006989927496761084, + 0.006736979354172945, + 0.008635465055704117, + 0.008507139980793, + 0.007452189922332764, + 0.0039582927711308, + 0.007711062673479319, + 0.006287082564085722, + 0.014495720155537128, + 0.006951826624572277, + 0.017812658101320267, + 0.005679894704371691, + 0.012102614156901836, + 0.0052276975475251675, + 0.026767278090119362, + 0.008706539869308472, + 0.00799898523837328, + 0.00787325855344534, + 0.008180634118616581, + 0.002854209393262863, + 0.004236179869621992, + 0.002708147745579481, + 0.011600734665989876, + 0.0029585135634988546, + 0.009060024283826351, + 0.00348187074996531, + 0.0030649176333099604, + 0.003389436285942793, + 0.006326647009700537, + 0.0022684927098453045, + 0.004192829132080078, + 0.003829573979601264, + 0.003908009268343449, + 0.0019046635134145617, + 0.00317589845508337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.006746016908437014 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020587211474776268 + ], + "opposite_logprob_tokens": [ + -3.895587205886841 + ], + "logprob_reward": -0.020587211474776268, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020587211474776268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4608134925365448, + 0.1118820309638977, + 0.06401000916957855, + 0.03311437740921974, + 0.12992754578590393, + 0.0741080567240715, + 0.06897397339344025, + 0.018836650997400284, + 0.01992008462548256, + 0.018413791432976723 + ], + "credit_map": [ + 0.08995632827281952, + 0.049584466964006424, + 0.06996884196996689, + 0.05925843119621277, + 0.045541536062955856, + 0.05409099534153938, + 0.03863802179694176, + 0.0537748746573925, + 0.013121718540787697, + 0.013737927190959454, + 0.007445963099598885, + 0.027251021936535835, + 0.009945996105670929, + 0.012637862004339695, + 0.005401969887316227, + 0.02233956567943096, + 0.007172355428338051, + 0.011547661386430264, + 0.0030295050237327814, + 0.012439688667654991, + 0.0069421869702637196, + 0.005894817411899567, + 0.006026974879205227, + 0.010956821031868458, + 0.003574477043002844, + 0.004769723396748304, + 0.0017375258030369878, + 0.003112134989351034, + 0.006302599795162678, + 0.00424830662086606, + 0.0035992495249956846, + 0.005770361050963402, + 0.028393451124429703, + 0.010920234024524689, + 0.009538764134049416, + 0.009714685380458832, + 0.010819884948432446, + 0.016138317063450813, + 0.02154470793902874, + 0.02285749837756157, + 0.013103432022035122, + 0.007858376950025558, + 0.0055100033059716225, + 0.005732819903641939, + 0.0049913362599909306, + 0.012612625025212765, + 0.009963284246623516, + 0.014336181804537773, + 0.004671747796237469, + 0.008438554592430592, + 0.0026818811893463135, + 0.0047187707386910915, + 0.010175267234444618, + 0.018821828067302704, + 0.0040765623562037945, + 0.015389357693493366, + 0.0016284254379570484, + 0.0026524695567786694, + 0.0010462719947099686, + 0.003730720840394497, + 0.0013205735012888908, + 0.0033739865757524967, + 0.0014498640084639192, + 0.0036343384999781847, + 0.001648921286687255, + 0.002677739132195711, + 0.0012378052342683077, + 0.0038368061650544405, + 0.0021259505301713943, + 0.004179051611572504, + 0.0018316901987418532, + 0.002382120117545128, + 0.0017246095230802894, + 0.0027203671634197235, + 0.0012485790066421032, + 0.003165902802720666, + 0.0019775775726884604, + 0.0047741299495100975, + 0.0011478947708383203, + 0.0016547319246456027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.020587211474776268 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798752099275589 + ], + "opposite_logprob_tokens": [ + -1.8048752546310425 + ], + "logprob_reward": -0.1798752099275589, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798752099275589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08397030830383301, + 0.08935360610485077, + 0.27797001600265503, + 0.05222311243414879, + 0.03172408789396286, + 0.11920526623725891, + 0.13458746671676636, + 0.07294844835996628, + 0.09397022426128387, + 0.044047459959983826 + ], + "credit_map": [ + 0.01203558873385191, + 0.008295792154967785, + 0.007999606430530548, + 0.011708241887390614, + 0.01061790157109499, + 0.014665340073406696, + 0.0067060841247439384, + 0.011941746808588505, + 0.008055438287556171, + 0.013789013959467411, + 0.004144501872360706, + 0.02379179373383522, + 0.008560429327189922, + 0.011728471145033836, + 0.006691817194223404, + 0.012592143379151821, + 0.027884898707270622, + 0.02956482209265232, + 0.015555227175354958, + 0.05927601084113121, + 0.03198254480957985, + 0.030631784349679947, + 0.03256652504205704, + 0.05050821229815483, + 0.00566613906994462, + 0.006082816980779171, + 0.002410088200122118, + 0.004689416848123074, + 0.008631893433630466, + 0.006067879032343626, + 0.010446597822010517, + 0.008228281512856483, + 0.0037698678206652403, + 0.002406009240075946, + 0.003133571706712246, + 0.003971252124756575, + 0.003606061916798353, + 0.004430610686540604, + 0.004495285451412201, + 0.00591142987832427, + 0.01705104485154152, + 0.015193427912890911, + 0.008483466692268848, + 0.011540000326931477, + 0.009909328073263168, + 0.019116345793008804, + 0.018926318734884262, + 0.018985334783792496, + 0.010882018133997917, + 0.015398481860756874, + 0.005686288699507713, + 0.01475756149739027, + 0.020149771124124527, + 0.03411242365837097, + 0.011091969907283783, + 0.022508956491947174, + 0.004074161406606436, + 0.008558815345168114, + 0.0047119855880737305, + 0.026052918285131454, + 0.005482821259647608, + 0.0104239946231246, + 0.006603527348488569, + 0.007040222175419331, + 0.004561797250062227, + 0.008631519973278046, + 0.0051698037423193455, + 0.03196844831109047, + 0.004909495823085308, + 0.023619620129466057, + 0.008552782237529755, + 0.006556747015565634, + 0.005202156025916338, + 0.009418953210115433, + 0.00345631898380816, + 0.007457743864506483, + 0.006471051834523678, + 0.0038181468844413757, + 0.0035455382894724607, + 0.004677550867199898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.1798752099275589 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static high-angle shot, as if from an elevated vantage point, gazes down upon an open, weathered hardcover book, spread upon a carpet of lush green moss. The book's dark brown leather cover has cracked and softened with age, its corners gently curled. The exposed pages are a creamy, yellowed parchment, thick and fibrous, covered in dense, elegant calligraphy in faded black ink. The book lies on a bed of fuzzy, damp-looking moss on the forest floor, surrounded by a few scattered, crisp autumn leaves and the dark, gnarled roots of a nearby tree. Dappled sunlight filters through the unseen treetops, casting shifting, warm patterns of light and shadow across the scene. A gentle breeze breathes life into the frame, causing the book's page in the upper right corner to subtly lift and flutter; it rises hesitantly, revealing the pages beneath for a brief moment, before settling back down with a soft, almost imperceptible rustling, creating an atmosphere of tranquility, contemplation, and a touch of magic, of quiet discovery.", + "seed": 23597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_8/index_1919_seed_23597_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.0655731105711311, + "vlm_logprob": -0.0655731105711311, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.0655731105711311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.034348756074905396 + ], + "opposite_logprob_tokens": [ + -4.534348964691162 + ], + "logprob_reward": -0.034348756074905396, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.034348756074905396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2326766401529312, + 0.043522439897060394, + 0.03533975034952164, + 0.0511610209941864, + 0.05472507327795029, + 0.09882918000221252, + 0.15897929668426514, + 0.27958688139915466, + 0.028126254677772522, + 0.017053432762622833 + ], + "credit_map": [ + 0.026806777343153954, + 0.05630258470773697, + 0.027421019971370697, + 0.030773013830184937, + 0.018252495676279068, + 0.026819730177521706, + 0.019075609743595123, + 0.027225438505411148, + 0.004426976200193167, + 0.004736450966447592, + 0.004849237389862537, + 0.004056745674461126, + 0.004139121621847153, + 0.007303549908101559, + 0.005951999220997095, + 0.008058357052505016, + 0.0034518020693212748, + 0.002995801158249378, + 0.00495484983548522, + 0.003017040900886059, + 0.003873169654980302, + 0.005724072456359863, + 0.006925302557647228, + 0.004397711716592312, + 0.00527087040245533, + 0.004866499919444323, + 0.006572684273123741, + 0.0040847840718925, + 0.004059506580233574, + 0.010529468767344952, + 0.007318368647247553, + 0.008458847180008888, + 0.007540781982243061, + 0.004111943766474724, + 0.006153506692498922, + 0.007553894072771072, + 0.0036580213345587254, + 0.00845650490373373, + 0.009481203742325306, + 0.007769216317683458, + 0.010188320651650429, + 0.008036275394260883, + 0.008673674426972866, + 0.017810406163334846, + 0.007966604083776474, + 0.010315296240150928, + 0.020732425153255463, + 0.015106188133358955, + 0.023134281858801842, + 0.012989216484129429, + 0.0299699530005455, + 0.016373801976442337, + 0.010109259746968746, + 0.02254445105791092, + 0.014633272774517536, + 0.029225073754787445, + 0.03778279572725296, + 0.015253502875566483, + 0.05930735915899277, + 0.027369214221835136, + 0.016318388283252716, + 0.0617363341152668, + 0.02533291094005108, + 0.036486394703388214, + 0.0033143649343401194, + 0.00275988457724452, + 0.00280479877255857, + 0.0033001648262143135, + 0.0017529433825984597, + 0.00558016961440444, + 0.00372957531362772, + 0.004884356167167425, + 0.0021289896685630083, + 0.0017623186577111483, + 0.002902248175814748, + 0.002242974005639553, + 0.0009673302993178368, + 0.0027749165892601013, + 0.00187811313662678, + 0.002396542811766267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.034348756074905396 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002846239833161235 + ], + "opposite_logprob_tokens": [ + -5.8778462409973145 + ], + "logprob_reward": -0.002846239833161235, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002846239833161235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3121192753314972, + 0.19861634075641632, + 0.04997611045837402, + 0.09418376535177231, + 0.05164065584540367, + 0.0727045089006424, + 0.09540538489818573, + 0.06614181399345398, + 0.024150559678673744, + 0.035061612725257874 + ], + "credit_map": [ + 0.03219299390912056, + 0.06873130798339844, + 0.0237562358379364, + 0.04994898661971092, + 0.022201867774128914, + 0.026663517579436302, + 0.031190788373351097, + 0.05743357166647911, + 0.01927824132144451, + 0.02860955521464348, + 0.018452884629368782, + 0.02496352791786194, + 0.029129892587661743, + 0.016792818903923035, + 0.03572646528482437, + 0.02566295862197876, + 0.0051509723998606205, + 0.004710273817181587, + 0.006350786425173283, + 0.005481679458171129, + 0.005425649229437113, + 0.0058534517884254456, + 0.010484719648957253, + 0.006518578622490168, + 0.011088893748819828, + 0.008361556567251682, + 0.018174543976783752, + 0.009584843181073666, + 0.006905939895659685, + 0.011801722459495068, + 0.014781429432332516, + 0.013484845869243145, + 0.005770283751189709, + 0.004520799033343792, + 0.007567810360342264, + 0.006039086263626814, + 0.0038918382488191128, + 0.0062611219473183155, + 0.008933152072131634, + 0.008656570687890053, + 0.008943229913711548, + 0.0062935189343988895, + 0.005840674042701721, + 0.013169096782803535, + 0.006455594673752785, + 0.006152160465717316, + 0.013291975483298302, + 0.012558257207274437, + 0.01240646280348301, + 0.009279063902795315, + 0.011275217868387699, + 0.015859950333833694, + 0.007410144899040461, + 0.01335330680012703, + 0.00975626427680254, + 0.016064977273344994, + 0.008404134772717953, + 0.0052220881916582584, + 0.009349711239337921, + 0.011180208995938301, + 0.003461395623162389, + 0.009456746280193329, + 0.008872692473232746, + 0.010194837115705013, + 0.0027676799800246954, + 0.0017419001087546349, + 0.005337333772331476, + 0.003214576281607151, + 0.0014403165550902486, + 0.004047892056405544, + 0.002946360968053341, + 0.0026545000728219748, + 0.004663613624870777, + 0.005315511021763086, + 0.002968729007989168, + 0.004303397610783577, + 0.002282150089740753, + 0.005152845289558172, + 0.005990577396005392, + 0.004384791478514671 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.002846239833161235 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005257589276880026 + ], + "opposite_logprob_tokens": [ + -5.255257606506348 + ], + "logprob_reward": -0.005257589276880026, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005257589276880026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4032430648803711, + 0.07512431591749191, + 0.03359111770987511, + 0.029757240787148476, + 0.06396382302045822, + 0.06420817971229553, + 0.16224223375320435, + 0.08700219541788101, + 0.02673383243381977, + 0.05413411557674408 + ], + "credit_map": [ + 0.04240976274013519, + 0.10488169640302658, + 0.022489888593554497, + 0.06480845808982849, + 0.03300326690077782, + 0.026869945228099823, + 0.03595205396413803, + 0.07282792031764984, + 0.010916179046034813, + 0.007593520916998386, + 0.009342328645288944, + 0.006491977721452713, + 0.008816644549369812, + 0.010659981518983841, + 0.01112992875277996, + 0.010173739865422249, + 0.00382999237626791, + 0.0025797677226364613, + 0.0035066967830061913, + 0.004117881413549185, + 0.0033828895539045334, + 0.004367422312498093, + 0.007542773149907589, + 0.004263690672814846, + 0.0029305117204785347, + 0.002737733768299222, + 0.00524270161986351, + 0.0037771800998598337, + 0.0022471663542091846, + 0.004044083878397942, + 0.00443339953199029, + 0.004344456363469362, + 0.007022141944617033, + 0.006238624919205904, + 0.007913059554994106, + 0.008983454667031765, + 0.005173816811293364, + 0.005539997946470976, + 0.012283838354051113, + 0.010808883234858513, + 0.00962617713958025, + 0.005517806392163038, + 0.0052046217024326324, + 0.011062867939472198, + 0.005175258964300156, + 0.00660673389211297, + 0.009858504869043827, + 0.011156200431287289, + 0.022928474470973015, + 0.014859653078019619, + 0.022347157821059227, + 0.024008840322494507, + 0.011718356981873512, + 0.022185297682881355, + 0.01554050948470831, + 0.028653917834162712, + 0.011224917136132717, + 0.00394823495298624, + 0.018742229789495468, + 0.007486409042030573, + 0.003934122156351805, + 0.020671820268034935, + 0.007573118899017572, + 0.013421323150396347, + 0.0028321314603090286, + 0.0017665131017565727, + 0.005183677654713392, + 0.0036489644553512335, + 0.0015376332448795438, + 0.004944100044667721, + 0.0031763825099915266, + 0.0036444258876144886, + 0.006982127670198679, + 0.011296591721475124, + 0.005464424379169941, + 0.005292887333780527, + 0.003820903366431594, + 0.008130663074553013, + 0.006554751191288233, + 0.006591760087758303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.005257589276880026 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004096925724297762 + ], + "opposite_logprob_tokens": [ + -5.504096984863281 + ], + "logprob_reward": -0.004096925724297762, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004096925724297762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5254473686218262, + 0.07071178406476974, + 0.044941943138837814, + 0.07998428493738174, + 0.051238249987363815, + 0.05297117680311203, + 0.06076556444168091, + 0.07154391705989838, + 0.02246725559234619, + 0.019928457215428352 + ], + "credit_map": [ + 0.061099886894226074, + 0.1511247754096985, + 0.03509014472365379, + 0.06861802935600281, + 0.03898097202181816, + 0.04097476601600647, + 0.026549188420176506, + 0.10300961881875992, + 0.007234285119920969, + 0.00836488138884306, + 0.00951390154659748, + 0.007622196339070797, + 0.008164617232978344, + 0.008714119903743267, + 0.010962900705635548, + 0.010134884156286716, + 0.005460982210934162, + 0.003514346433803439, + 0.007001013960689306, + 0.0036619468592107296, + 0.004885287489742041, + 0.006886798422783613, + 0.008110727183520794, + 0.005420839414000511, + 0.010809697210788727, + 0.004293061327189207, + 0.016554972156882286, + 0.005500501487404108, + 0.004742900840938091, + 0.014995653182268143, + 0.010019578970968723, + 0.013067916966974735, + 0.007779484614729881, + 0.0025871607940644026, + 0.010619476437568665, + 0.0036146589554846287, + 0.002315316814929247, + 0.010737662203609943, + 0.004658134188503027, + 0.00892635714262724, + 0.007410774938762188, + 0.003508499823510647, + 0.007508523296564817, + 0.007216674741357565, + 0.0038276314735412598, + 0.007104670628905296, + 0.007589144632220268, + 0.008805257268249989, + 0.009577831253409386, + 0.005719992332160473, + 0.009255748242139816, + 0.005912503227591515, + 0.004251521080732346, + 0.013389580883085728, + 0.004597713705152273, + 0.008060676977038383, + 0.010278896428644657, + 0.003713640384376049, + 0.012709454633295536, + 0.005511975381523371, + 0.0037265943828970194, + 0.020890910178422928, + 0.004436001647263765, + 0.010276442393660545, + 0.0031950566917657852, + 0.0012574068969115615, + 0.004495240282267332, + 0.001701954286545515, + 0.0010550550650805235, + 0.005318756215274334, + 0.0018421463901177049, + 0.003601640462875366, + 0.003288249485194683, + 0.001134560676291585, + 0.003815935691818595, + 0.0015995100839063525, + 0.0012113384436815977, + 0.005026064347475767, + 0.0013730607461184263, + 0.002479739021509886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.004096925724297762 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot taken from a static high-angle perspective looking down?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an open, weathered hardcover book with dark brown cracked leather and curled corners visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the exposed pages creamy yellowed parchment with dense, elegant calligraphy in faded black ink?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the book resting on lush green moss surrounded by scattered autumn leaves and dark gnarled tree roots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does a page in the upper right corner lift and flutter gently due to a breeze before settling back down?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28131604194641113 + ], + "opposite_logprob_tokens": [ + -1.4063160419464111 + ], + "logprob_reward": -0.28131604194641113, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28131604194641113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24478477239608765, + 0.3468521535396576, + 0.0489172488451004, + 0.0364241786301136, + 0.051839109510183334, + 0.0639319196343422, + 0.08217281848192215, + 0.05945444479584694, + 0.04755339026451111, + 0.01806996576488018 + ], + "credit_map": [ + 0.02632053755223751, + 0.04240986704826355, + 0.019697146490216255, + 0.03547830134630203, + 0.018853798508644104, + 0.029964596033096313, + 0.022407203912734985, + 0.04965335130691528, + 0.038005948066711426, + 0.04434170573949814, + 0.03159305825829506, + 0.046476736664772034, + 0.05327601730823517, + 0.02451239712536335, + 0.06189801171422005, + 0.04674828425049782, + 0.005251996219158173, + 0.003049516584724188, + 0.005484552588313818, + 0.004342853091657162, + 0.005669232923537493, + 0.004740830510854721, + 0.011751473881304264, + 0.008626798167824745, + 0.004867078270763159, + 0.002530080499127507, + 0.0050239865668118, + 0.0037870127707719803, + 0.003120721084997058, + 0.004044651053845882, + 0.007208955008536577, + 0.005841696169227362, + 0.005728932097554207, + 0.005007015075534582, + 0.008201880380511284, + 0.008399810642004013, + 0.0041952370665967464, + 0.006484027020633221, + 0.0068154470063745975, + 0.007006764877587557, + 0.009738259017467499, + 0.0053075156174600124, + 0.006596743129193783, + 0.00990553479641676, + 0.005443240981549025, + 0.007336875889450312, + 0.009036396630108356, + 0.010567361488938332, + 0.010257195681333542, + 0.008072390221059322, + 0.011716692708432674, + 0.011907516047358513, + 0.00626376923173666, + 0.0119180241599679, + 0.007609985303133726, + 0.014427248388528824, + 0.007915645837783813, + 0.003506709821522236, + 0.011784081347286701, + 0.004216698929667473, + 0.0031308578327298164, + 0.015758363530039787, + 0.0038576582446694374, + 0.009284434840083122, + 0.005968838464468718, + 0.0036516468971967697, + 0.009972370229661465, + 0.0045828260481357574, + 0.0026385432574898005, + 0.010269632562994957, + 0.004359911195933819, + 0.006109623238444328, + 0.0021148454397916794, + 0.0015699205687269568, + 0.0031555977184325457, + 0.0023529038298875093, + 0.0015577240847051144, + 0.003555868985131383, + 0.0016106227412819862, + 0.002152484841644764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.28131604194641113 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_tensor([8699, 8699])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_tensor([8699, 8699])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..920e5e162f82bedb065592460689f06c0dcda112 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_tensor([8699, 8699])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, long shot, filmed from ground level, captures, with a calm, unadorned perspective, a typical black-and-white police car, its red and blue lights flashing wildly, hurtling down the shoulder of an empty highway at uncontrolled speed towards a solid, weathered, gray concrete pier. At the moment of impact, the front of the car is instantly crushed with astonishing physical precision, the hood folding upward under the immense force, twisting into a heap of scrap metal, while fragments of the front bumper and headlights explode into countless tiny particles and fly in all directions. Then, a thick plume of white steam and black smoke erupts from the ripped-apart engine compartment. The immense momentum causes the rear of the car to briefly lift upward before slamming back to the ground, finally coming to a complete stop with a sharp, grinding sound, a heap of twisted, smoking metal. The entire process takes place under bright but slightly harsh afternoon sunlight. The sharp shadows and high-contrast lighting emphasize the violence and reality of the collision, creating an intense and heart-pounding documentary style.", + "seed": 13956, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_8699_seed_13956_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03561868378892541, + "vlm_logprob": -0.03561868378892541, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.03561868378892541, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09097830206155777 + ], + "opposite_logprob_tokens": [ + -3.2159783840179443 + ], + "logprob_reward": -0.09097830206155777, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.09097830206155777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08895701915025711, + 0.15556949377059937, + 0.03779898211359978, + 0.07079547643661499, + 0.036779675632715225, + 0.1041746512055397, + 0.2339177131652832, + 0.03639018163084984, + 0.20613069832324982, + 0.029486076906323433 + ], + "credit_map": [ + 0.009304055012762547, + 0.016215546056628227, + 0.007382031064480543, + 0.009322955273091793, + 0.016558166593313217, + 0.01456360798329115, + 0.00586765818297863, + 0.009743007831275463, + 0.008456355892121792, + 0.01664993353188038, + 0.013530617579817772, + 0.005423390306532383, + 0.06223904341459274, + 0.03840017691254616, + 0.005416104104369879, + 0.005453886464238167, + 0.0035465604159981012, + 0.006790484301745892, + 0.003979358356446028, + 0.004677306395024061, + 0.0061939856968820095, + 0.005414403509348631, + 0.004316243808716536, + 0.0028806407935917377, + 0.007628134451806545, + 0.010097742080688477, + 0.01562510058283806, + 0.008288483135402203, + 0.008927881717681885, + 0.007293715141713619, + 0.005902362521737814, + 0.007032057736068964, + 0.004393588751554489, + 0.004850226920098066, + 0.004260842688381672, + 0.003384237177670002, + 0.007645761594176292, + 0.005920357536524534, + 0.003443214576691389, + 0.002881450578570366, + 0.008316544815897942, + 0.01351119577884674, + 0.013721337541937828, + 0.00746346078813076, + 0.007024270482361317, + 0.02078535407781601, + 0.00955152790993452, + 0.023800969123840332, + 0.01633596234023571, + 0.03727341070771217, + 0.04131577908992767, + 0.009718760848045349, + 0.022345656529068947, + 0.07488564401865005, + 0.01300144661217928, + 0.019041072577238083, + 0.005245019216090441, + 0.004181263968348503, + 0.0039007856976240873, + 0.003980828449130058, + 0.007765182759612799, + 0.004059820901602507, + 0.0034838421270251274, + 0.00377344130538404, + 0.01740834303200245, + 0.010900421068072319, + 0.056413594633340836, + 0.007368938066065311, + 0.013241633772850037, + 0.07990498095750809, + 0.014359789900481701, + 0.006533009931445122, + 0.004618191625922918, + 0.0025085608940571547, + 0.007143120747059584, + 0.0024703575763851404, + 0.003622419899329543, + 0.004890776239335537, + 0.002112311078235507, + 0.002120340010151267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.09097830206155777 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029841016978025436 + ], + "opposite_logprob_tokens": [ + -3.5298409461975098 + ], + "logprob_reward": -0.029841016978025436, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029841016978025436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10453709959983826, + 0.055106159299612045, + 0.16039559245109558, + 0.24324630200862885, + 0.02284531109035015, + 0.021610669791698456, + 0.2982046604156494, + 0.022392524406313896, + 0.04931629076600075, + 0.0223454087972641 + ], + "credit_map": [ + 0.00955486111342907, + 0.026973558589816093, + 0.00829279888421297, + 0.012923290953040123, + 0.01721186749637127, + 0.01273727510124445, + 0.009277825243771076, + 0.00756564037874341, + 0.003812426934018731, + 0.01053551398217678, + 0.00954762939363718, + 0.0051370845176279545, + 0.010420184582471848, + 0.006388192530721426, + 0.005717853549867868, + 0.0035472798626869917, + 0.013692144304513931, + 0.023852800950407982, + 0.0380808562040329, + 0.009869587607681751, + 0.01674206554889679, + 0.035185590386390686, + 0.01316544134169817, + 0.00980712752789259, + 0.02179841324687004, + 0.03518220782279968, + 0.026717590168118477, + 0.021026451140642166, + 0.07134511321783066, + 0.023987552151083946, + 0.0248824805021286, + 0.018306508660316467, + 0.0023062750697135925, + 0.004368417896330357, + 0.0031327109318226576, + 0.002271609613671899, + 0.003420890774577856, + 0.0026576430536806583, + 0.002705818507820368, + 0.0019819475710392, + 0.002330881543457508, + 0.00276794470846653, + 0.001991687109693885, + 0.0019419303862378001, + 0.0017732030246406794, + 0.004851603880524635, + 0.0019942650105804205, + 0.003959156572818756, + 0.017345111817121506, + 0.04673968255519867, + 0.06068217009305954, + 0.00922362133860588, + 0.022644901648163795, + 0.10914722830057144, + 0.014769220724701881, + 0.017652736976742744, + 0.002680211327970028, + 0.002275709994137287, + 0.002410956658422947, + 0.002597227692604065, + 0.004233095794916153, + 0.00267860502935946, + 0.0025998863857239485, + 0.0029168338514864445, + 0.004800636786967516, + 0.002644002204760909, + 0.012522522360086441, + 0.002142395358532667, + 0.004118201322853565, + 0.01757434755563736, + 0.0035205120220780373, + 0.001993676647543907, + 0.0030290870927274227, + 0.001199871301651001, + 0.0061491248197853565, + 0.001919273054227233, + 0.0016568148275837302, + 0.005028273910284042, + 0.0015443508746102452, + 0.0018186154775321484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.029841016978025436 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288074880838394 + ], + "opposite_logprob_tokens": [ + -3.7732880115509033 + ], + "logprob_reward": -0.023288074880838394, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288074880838394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09757427871227264, + 0.10493893176317215, + 0.13480344414710999, + 0.054048214107751846, + 0.06519532203674316, + 0.13077670335769653, + 0.22906392812728882, + 0.038069017231464386, + 0.12351128458976746, + 0.02201891504228115 + ], + "credit_map": [ + 0.010524137876927853, + 0.02398163452744484, + 0.007635799236595631, + 0.011424033902585506, + 0.014838599599897861, + 0.012376897037029266, + 0.00717051001265645, + 0.0096226641908288, + 0.006393635645508766, + 0.01316207367926836, + 0.0072892410680651665, + 0.003661027178168297, + 0.034432049840688705, + 0.029736444354057312, + 0.00468941405415535, + 0.0055750515311956406, + 0.012611362151801586, + 0.024124452844262123, + 0.029208820313215256, + 0.007482531946152449, + 0.011690440587699413, + 0.03438738361001015, + 0.00804140791296959, + 0.007257039193063974, + 0.00559312105178833, + 0.006926541682332754, + 0.006325793452560902, + 0.00500832311809063, + 0.01396210491657257, + 0.0048464094288647175, + 0.00786575023084879, + 0.0035201688297092915, + 0.00782824493944645, + 0.008803678676486015, + 0.010919813066720963, + 0.003983699716627598, + 0.01046515442430973, + 0.012301310896873474, + 0.0050352648831903934, + 0.005858157761394978, + 0.013105169869959354, + 0.017960887402296066, + 0.013585991226136684, + 0.010723307728767395, + 0.013562064617872238, + 0.02779332362115383, + 0.01297129038721323, + 0.021074673160910606, + 0.011497885920107365, + 0.03950445353984833, + 0.05211995542049408, + 0.005752959288656712, + 0.011676310561597347, + 0.09062907844781876, + 0.008783827535808086, + 0.009099451825022697, + 0.003892352106049657, + 0.004058490507304668, + 0.0038197843823581934, + 0.004532761871814728, + 0.008289682678878307, + 0.004231202881783247, + 0.004333832301199436, + 0.0049109142273664474, + 0.010845091193914413, + 0.005500673316419125, + 0.03603508323431015, + 0.004854434169828892, + 0.0075177159160375595, + 0.045711930841207504, + 0.00867853406816721, + 0.004367824178189039, + 0.0030333297327160835, + 0.0016831723041832447, + 0.005500450264662504, + 0.001894167740829289, + 0.0027991505339741707, + 0.003687398275360465, + 0.0015210338169708848, + 0.0019002126064151525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.023288074880838394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639596704393625 + ], + "opposite_logprob_tokens": [ + -4.882639408111572 + ], + "logprob_reward": -0.007639596704393625, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639596704393625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05555839091539383, + 0.13729414343833923, + 0.08272268623113632, + 0.04686560109257698, + 0.03963887318968773, + 0.09398134052753448, + 0.34602347016334534, + 0.10584680736064911, + 0.07483883202075958, + 0.017229849472641945 + ], + "credit_map": [ + 0.006633278448134661, + 0.00944406446069479, + 0.004626056179404259, + 0.007729633245617151, + 0.008964591659605503, + 0.009056995622813702, + 0.004263540729880333, + 0.004840231500566006, + 0.006179785821586847, + 0.010141958482563496, + 0.01950976252555847, + 0.0039000362157821655, + 0.05449620634317398, + 0.03255905210971832, + 0.0041648102924227715, + 0.00634252792224288, + 0.007478384766727686, + 0.012631074525415897, + 0.018650084733963013, + 0.004921863786876202, + 0.010679660364985466, + 0.01533360406756401, + 0.007628156337887049, + 0.0053998506627976894, + 0.005414361599832773, + 0.007201441563665867, + 0.010064384900033474, + 0.004361190367490053, + 0.007413034792989492, + 0.003517305478453636, + 0.0038210179191082716, + 0.0050728642381727695, + 0.004033968783915043, + 0.0069693271070718765, + 0.004813557490706444, + 0.0033299019560217857, + 0.007151979021728039, + 0.004992765840142965, + 0.004690325353294611, + 0.0036570485681295395, + 0.008361946791410446, + 0.01549137756228447, + 0.011302717961370945, + 0.00654489966109395, + 0.01155687216669321, + 0.017748897895216942, + 0.0065462891943752766, + 0.016428345814347267, + 0.019692840054631233, + 0.05255237594246864, + 0.07449158281087875, + 0.009929248131811619, + 0.026717547327280045, + 0.12966758012771606, + 0.015312678180634975, + 0.017659621313214302, + 0.011705308221280575, + 0.011199661530554295, + 0.009912504814565182, + 0.01307961530983448, + 0.02341429702937603, + 0.011012482456862926, + 0.012575840577483177, + 0.012947103008627892, + 0.006496740039438009, + 0.00376878515817225, + 0.019750088453292847, + 0.0030952871311455965, + 0.00571505818516016, + 0.027451734989881516, + 0.005369760096073151, + 0.0031913728453218937, + 0.0032096458598971367, + 0.0010461814235895872, + 0.0032691906671971083, + 0.0016894923755899072, + 0.001738591818138957, + 0.0030380375683307648, + 0.0014884703559800982, + 0.0017502395203337073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.007639596704393625 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634642831981182 + ], + "opposite_logprob_tokens": [ + -3.651346445083618 + ], + "logprob_reward": -0.02634642831981182, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634642831981182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0833660289645195, + 0.10615280270576477, + 0.035638123750686646, + 0.03364530950784683, + 0.04376089200377464, + 0.02496456168591976, + 0.45436641573905945, + 0.031015530228614807, + 0.1503206044435501, + 0.03676967695355415 + ], + "credit_map": [ + 0.01096316333860159, + 0.013268081471323967, + 0.008133829571306705, + 0.0062547652050852776, + 0.017723912373185158, + 0.009640103206038475, + 0.008036850951611996, + 0.00934532843530178, + 0.006443648599088192, + 0.011519331485033035, + 0.017027147114276886, + 0.00415556225925684, + 0.039536215364933014, + 0.018761420622467995, + 0.004206186626106501, + 0.004503295291215181, + 0.0031620454974472523, + 0.006761032156646252, + 0.003644958371296525, + 0.0017745262011885643, + 0.008667011745274067, + 0.0060567473992705345, + 0.0037257664371281862, + 0.0018460405990481377, + 0.0040240040980279446, + 0.0046682353131473064, + 0.008693789131939411, + 0.0031140679493546486, + 0.004231152590364218, + 0.002385841216892004, + 0.002861856482923031, + 0.0036663662176579237, + 0.006187691353261471, + 0.007156710140407085, + 0.007320489268749952, + 0.002657755510881543, + 0.006543254014104605, + 0.00591841246932745, + 0.003709734883159399, + 0.004266843199729919, + 0.003134458791464567, + 0.0043642958626151085, + 0.0024785115383565426, + 0.0017045222921296954, + 0.0027642056811600924, + 0.004861104767769575, + 0.0018337881192564964, + 0.003823675215244293, + 0.025673644617199898, + 0.0771925300359726, + 0.10023295134305954, + 0.013336061500012875, + 0.017745696008205414, + 0.17280815541744232, + 0.01707427017390728, + 0.030303115025162697, + 0.005314120557159185, + 0.005663474556058645, + 0.004705645143985748, + 0.0022229179739952087, + 0.0031715519726276398, + 0.00505679240450263, + 0.0019958324264734983, + 0.002885199151933193, + 0.013007516972720623, + 0.006897483952343464, + 0.04374000057578087, + 0.005692303646355867, + 0.009006030857563019, + 0.055187612771987915, + 0.010381988249719143, + 0.00640766741707921, + 0.005633704364299774, + 0.0019796197302639484, + 0.00797309447079897, + 0.0028260149993002415, + 0.003133438527584076, + 0.009593144990503788, + 0.0026476620696485043, + 0.0029829975683242083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.02634642831981182 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot, filmed from ground level, captures, with a calm, unadorned perspective, a typical black-and-white police car, its red and blue lights flashing wildly, hurtling down the shoulder of an empty highway at uncontrolled speed towards a solid, weathered, gray concrete pier. At the moment of impact, the front of the car is instantly crushed with astonishing physical precision, the hood folding upward under the immense force, twisting into a heap of scrap metal, while fragments of the front bumper and headlights explode into countless tiny particles and fly in all directions. Then, a thick plume of white steam and black smoke erupts from the ripped-apart engine compartment. The immense momentum causes the rear of the car to briefly lift upward before slamming back to the ground, finally coming to a complete stop with a sharp, grinding sound, a heap of twisted, smoking metal. The entire process takes place under bright but slightly harsh afternoon sunlight. The sharp shadows and high-contrast lighting emphasize the violence and reality of the collision, creating an intense and heart-pounding documentary style.", + "seed": 13957, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_8699_seed_13957_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10188178420066833, + "vlm_logprob": -0.10188178420066833, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.10188178420066833, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31480228900909424 + ], + "opposite_logprob_tokens": [ + -1.8148022890090942 + ], + "logprob_reward": -0.31480228900909424, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.31480228900909424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10241478681564331, + 0.06309553980827332, + 0.09117097407579422, + 0.17773310840129852, + 0.13668882846832275, + 0.09241390973329544, + 0.04588153213262558, + 0.04949741065502167, + 0.06068105250597, + 0.18042291700839996 + ], + "credit_map": [ + 0.021500181406736374, + 0.01450654212385416, + 0.013104096986353397, + 0.008417046628892422, + 0.011981785297393799, + 0.010682658292353153, + 0.009497754275798798, + 0.012724710628390312, + 0.005889780353754759, + 0.00607586232945323, + 0.012231560423970222, + 0.006231759209185839, + 0.006167483050376177, + 0.006943690590560436, + 0.008770540356636047, + 0.010784859769046307, + 0.00909296702593565, + 0.013291813433170319, + 0.013327313587069511, + 0.005599997006356716, + 0.011318960227072239, + 0.014149483293294907, + 0.010534564964473248, + 0.013855867087841034, + 0.026477057486772537, + 0.02084190398454666, + 0.04371993616223335, + 0.011835913173854351, + 0.011256585828959942, + 0.014959756284952164, + 0.02014288306236267, + 0.0284990556538105, + 0.017616866156458855, + 0.017914706841111183, + 0.01649494096636772, + 0.013672403991222382, + 0.033993955701589584, + 0.012242354452610016, + 0.011170705780386925, + 0.013582892715930939, + 0.016598278656601906, + 0.014180943369865417, + 0.013781599700450897, + 0.011689932085573673, + 0.01118883304297924, + 0.01147057581692934, + 0.005545060615986586, + 0.00795868132263422, + 0.008041700348258018, + 0.009443648159503937, + 0.005625880789011717, + 0.006364933215081692, + 0.005430073477327824, + 0.0033017992973327637, + 0.0037677420768886805, + 0.0039057491812855005, + 0.006792823318392038, + 0.005082463379949331, + 0.005100434646010399, + 0.008779969997704029, + 0.005602973513305187, + 0.005490564275532961, + 0.0058113629929721355, + 0.006836816668510437, + 0.010984922759234905, + 0.00931417103856802, + 0.006475830916315317, + 0.00674293702468276, + 0.008531040512025356, + 0.004815370310097933, + 0.006799872498959303, + 0.00701690511777997, + 0.027118252590298653, + 0.0441150963306427, + 0.02220311015844345, + 0.01804562285542488, + 0.01977406069636345, + 0.015132361091673374, + 0.012735901400446892, + 0.021298490464687347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 4 + ], + "selected_reward": -0.31480228900909424 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907337486743927 + ], + "opposite_logprob_tokens": [ + -2.464073419570923 + ], + "logprob_reward": -0.08907337486743927, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907337486743927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17666518688201904, + 0.272797167301178, + 0.0811869204044342, + 0.058082424104213715, + 0.06141125410795212, + 0.0262881051748991, + 0.09503209590911865, + 0.05706070363521576, + 0.11650679260492325, + 0.054969385266304016 + ], + "credit_map": [ + 0.022521113976836205, + 0.03318845108151436, + 0.019119691103696823, + 0.017056116834282875, + 0.017866259440779686, + 0.015192813239991665, + 0.019361330196261406, + 0.03235940262675285, + 0.03926025703549385, + 0.03958216682076454, + 0.031207818537950516, + 0.015460954047739506, + 0.028365686535835266, + 0.017621053382754326, + 0.044718991965055466, + 0.05658026039600372, + 0.008124209940433502, + 0.02185324765741825, + 0.00870039314031601, + 0.005752235651016235, + 0.008177737705409527, + 0.010746963322162628, + 0.008522470481693745, + 0.009309658780694008, + 0.007744755130261183, + 0.007419551257044077, + 0.01012595184147358, + 0.0040738824754953384, + 0.00728720473125577, + 0.008382626809179783, + 0.006044481415301561, + 0.007003966253250837, + 0.009250042960047722, + 0.010807300917804241, + 0.007581142708659172, + 0.006499640643596649, + 0.012381350621581078, + 0.005129484459757805, + 0.004289916716516018, + 0.0054723783396184444, + 0.0032308856025338173, + 0.0039938222616910934, + 0.0037887636572122574, + 0.002688742009922862, + 0.0031198894139379263, + 0.004052494186908007, + 0.0025706682354211807, + 0.002842840040102601, + 0.01746288500726223, + 0.013838255777955055, + 0.010900739580392838, + 0.01352191623300314, + 0.011998731642961502, + 0.01081795059144497, + 0.008220059797167778, + 0.008271557278931141, + 0.0056299204006791115, + 0.005947265308350325, + 0.0055598109029233456, + 0.006581746973097324, + 0.010973083786666393, + 0.007953330874443054, + 0.007237146608531475, + 0.007178397383540869, + 0.024267228320240974, + 0.023988982662558556, + 0.012384580448269844, + 0.011085713282227516, + 0.013503196649253368, + 0.0105125168338418, + 0.009947157464921474, + 0.010817415080964565, + 0.007719305343925953, + 0.011541235260665417, + 0.007362499367445707, + 0.006171871442347765, + 0.006745936349034309, + 0.006040652748197317, + 0.004134501796215773, + 0.005253382958471775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.08907337486743927 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369740769267082 + ], + "opposite_logprob_tokens": [ + -3.4086973667144775 + ], + "logprob_reward": -0.03369740769267082, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369740769267082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17517471313476562, + 0.08790966868400574, + 0.05103409290313721, + 0.06849386543035507, + 0.28410086035728455, + 0.07324419170618057, + 0.06425595283508301, + 0.05336389318108559, + 0.05086016654968262, + 0.09156260639429092 + ], + "credit_map": [ + 0.0195886492729187, + 0.01929076574742794, + 0.026442475616931915, + 0.0217727143317461, + 0.021503442898392677, + 0.015167545527219772, + 0.017792273312807083, + 0.033616844564676285, + 0.006519358605146408, + 0.005467954091727734, + 0.009100345894694328, + 0.00601839367300272, + 0.012923318892717361, + 0.014272323809564114, + 0.01592661812901497, + 0.017681356519460678, + 0.005183739122003317, + 0.00800873339176178, + 0.006891374010592699, + 0.004386496264487505, + 0.006427221931517124, + 0.0074325064197182655, + 0.005381138063967228, + 0.007322880905121565, + 0.008316623978316784, + 0.007965694181621075, + 0.01564502716064453, + 0.0047378260642290115, + 0.005856105126440525, + 0.006840602494776249, + 0.007000409997999668, + 0.012131579220294952, + 0.03984149917960167, + 0.04261573404073715, + 0.04235658422112465, + 0.01955021731555462, + 0.06892376393079758, + 0.02728104032576084, + 0.01815752685070038, + 0.025374485179781914, + 0.015385318547487259, + 0.01218022033572197, + 0.011500637046992779, + 0.00786644872277975, + 0.009309002198278904, + 0.006340996362268925, + 0.005004228558391333, + 0.005657334811985493, + 0.014166681095957756, + 0.010923970490694046, + 0.00923721119761467, + 0.008486644364893436, + 0.008382945321500301, + 0.004415750969201326, + 0.003962572198361158, + 0.004680176265537739, + 0.011760558001697063, + 0.008069929666817188, + 0.006069486029446125, + 0.0048388224095106125, + 0.00732463551685214, + 0.005804873071610928, + 0.004630849696695805, + 0.004864740651100874, + 0.010052385739982128, + 0.010069038718938828, + 0.005291717126965523, + 0.004744985140860081, + 0.005876616574823856, + 0.005517521873116493, + 0.004451361950486898, + 0.004856542218476534, + 0.021279986947774887, + 0.017753921449184418, + 0.012949079275131226, + 0.007725602947175503, + 0.007108410354703665, + 0.008995379321277142, + 0.0065927449613809586, + 0.009157483465969563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.03369740769267082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0097922682762146 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.0097922682762146, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.0097922682762146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1995859444141388, + 0.09870204329490662, + 0.04913174733519554, + 0.09256003051996231, + 0.08416930586099625, + 0.06420633941888809, + 0.13482649624347687, + 0.09712240844964981, + 0.10946482419967651, + 0.07023094594478607 + ], + "credit_map": [ + 0.03715750575065613, + 0.0332721471786499, + 0.027083830907940865, + 0.01295419316738844, + 0.01957445964217186, + 0.021394694223999977, + 0.016370395198464394, + 0.03177869692444801, + 0.015375884249806404, + 0.009579864330589771, + 0.01990155316889286, + 0.00984338391572237, + 0.0075803413055837154, + 0.010810031555593014, + 0.012389566749334335, + 0.013221418485045433, + 0.0076686916872859, + 0.004666694439947605, + 0.0077255964279174805, + 0.004922843538224697, + 0.005363489966839552, + 0.004286059644073248, + 0.005775753874331713, + 0.008722616359591484, + 0.0120753925293684, + 0.012146076187491417, + 0.0226754080504179, + 0.00685657374560833, + 0.00842355191707611, + 0.006544021423906088, + 0.01061263307929039, + 0.013226361945271492, + 0.012596734799444675, + 0.012301354669034481, + 0.016311032697558403, + 0.00837110634893179, + 0.010723584331572056, + 0.007243417203426361, + 0.007086632307618856, + 0.009535442106425762, + 0.011511363089084625, + 0.010973560623824596, + 0.010063727386295795, + 0.0061705224215984344, + 0.008821660652756691, + 0.005362422205507755, + 0.005066780373454094, + 0.006236300803720951, + 0.026667393743991852, + 0.027850516140460968, + 0.017682697623968124, + 0.01954353228211403, + 0.013556646183133125, + 0.01266003493219614, + 0.007204871624708176, + 0.009660794399678707, + 0.014255381189286709, + 0.010117162019014359, + 0.0066164699383080006, + 0.012464449740946293, + 0.00823380146175623, + 0.009710736572742462, + 0.013897919096052647, + 0.021826479583978653, + 0.02594696171581745, + 0.018342502415180206, + 0.011524813249707222, + 0.010684742592275143, + 0.012405090034008026, + 0.009740250185132027, + 0.009639902040362358, + 0.011180558241903782, + 0.014751674607396126, + 0.01068757101893425, + 0.007807225454598665, + 0.006180905736982822, + 0.005673326551914215, + 0.0074307238683104515, + 0.006075041834264994, + 0.011624475009739399 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.0097922682762146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062043581157922745 + ], + "opposite_logprob_tokens": [ + -2.8120436668395996 + ], + "logprob_reward": -0.062043581157922745, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062043581157922745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12990406155586243, + 0.1057109609246254, + 0.0806145891547203, + 0.06716664880514145, + 0.18132293224334717, + 0.07034442573785782, + 0.08011248707771301, + 0.07651660591363907, + 0.059357039630413055, + 0.14895027875900269 + ], + "credit_map": [ + 0.03385374695062637, + 0.016000857576727867, + 0.018254945054650307, + 0.009002536535263062, + 0.015367993153631687, + 0.01208434347063303, + 0.010761537589132786, + 0.014578104950487614, + 0.01227387972176075, + 0.0093185193836689, + 0.017497047781944275, + 0.007796304766088724, + 0.016337081789970398, + 0.007442235015332699, + 0.01799447275698185, + 0.017051422968506813, + 0.011697574518620968, + 0.007572485134005547, + 0.009393938817083836, + 0.005034538451582193, + 0.015574973076581955, + 0.013128958642482758, + 0.008857782930135727, + 0.009354340843856335, + 0.008118296973407269, + 0.009150144644081593, + 0.01331896148622036, + 0.005148346070200205, + 0.006419713143259287, + 0.009344748221337795, + 0.006870290730148554, + 0.008796148002147675, + 0.024170853197574615, + 0.027234267443418503, + 0.021497298032045364, + 0.011730475351214409, + 0.043661728501319885, + 0.025354627519845963, + 0.011853717267513275, + 0.0158199705183506, + 0.014300324022769928, + 0.011970574967563152, + 0.010764139704406261, + 0.005984966177493334, + 0.007922306656837463, + 0.00759283360093832, + 0.00551935750991106, + 0.006289916578680277, + 0.016337763518095016, + 0.01531524769961834, + 0.010294894687831402, + 0.011962863616645336, + 0.008182908408343792, + 0.006123249419033527, + 0.00571645051240921, + 0.006179105490446091, + 0.013082792051136494, + 0.009401426650583744, + 0.00916786678135395, + 0.016107559204101562, + 0.008986574597656727, + 0.005984731018543243, + 0.008303046226501465, + 0.005482611712068319, + 0.009812777861952782, + 0.00896051712334156, + 0.006082254461944103, + 0.006749716121703386, + 0.007529042195528746, + 0.005788595415651798, + 0.006449821405112743, + 0.007984315045177937, + 0.03623462840914726, + 0.02412663772702217, + 0.019772475585341454, + 0.013419493101537228, + 0.011256003752350807, + 0.016409430652856827, + 0.011989755555987358, + 0.0157418604940176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.062043581157922745 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot, filmed from ground level, captures, with a calm, unadorned perspective, a typical black-and-white police car, its red and blue lights flashing wildly, hurtling down the shoulder of an empty highway at uncontrolled speed towards a solid, weathered, gray concrete pier. At the moment of impact, the front of the car is instantly crushed with astonishing physical precision, the hood folding upward under the immense force, twisting into a heap of scrap metal, while fragments of the front bumper and headlights explode into countless tiny particles and fly in all directions. Then, a thick plume of white steam and black smoke erupts from the ripped-apart engine compartment. The immense momentum causes the rear of the car to briefly lift upward before slamming back to the ground, finally coming to a complete stop with a sharp, grinding sound, a heap of twisted, smoking metal. The entire process takes place under bright but slightly harsh afternoon sunlight. The sharp shadows and high-contrast lighting emphasize the violence and reality of the collision, creating an intense and heart-pounding documentary style.", + "seed": 13958, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_8699_seed_13958_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05260619791224599, + "vlm_logprob": -0.05260619791224599, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05260619791224599, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1742957979440689 + ], + "opposite_logprob_tokens": [ + -2.5492959022521973 + ], + "logprob_reward": -0.1742957979440689, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1742957979440689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11653703451156616, + 0.09922682493925095, + 0.21557040512561798, + 0.1756211221218109, + 0.0904291570186615, + 0.06147079914808273, + 0.06844254583120346, + 0.04778050258755684, + 0.07053637504577637, + 0.054385315626859665 + ], + "credit_map": [ + 0.010250011458992958, + 0.012354192323982716, + 0.019711071625351906, + 0.01756151393055916, + 0.010691300965845585, + 0.020097138360142708, + 0.01408784743398428, + 0.011783942580223083, + 0.00973391905426979, + 0.008442642167210579, + 0.021522101014852524, + 0.01593714952468872, + 0.00783903244882822, + 0.009615177288651466, + 0.0140072675421834, + 0.0121295265853405, + 0.01875423826277256, + 0.0227352287620306, + 0.029996957629919052, + 0.03344305604696274, + 0.03400333970785141, + 0.020319711416959763, + 0.0342097282409668, + 0.02210812270641327, + 0.01176410261541605, + 0.03532366454601288, + 0.01635625772178173, + 0.027011483907699585, + 0.02674025669693947, + 0.020253369584679604, + 0.021142378449440002, + 0.017029592767357826, + 0.00963747501373291, + 0.012857726775109768, + 0.014416293241083622, + 0.009679192677140236, + 0.013798404484987259, + 0.012067301198840141, + 0.008800712414085865, + 0.009172036312520504, + 0.008132473565638065, + 0.008457313291728497, + 0.012104224413633347, + 0.007968122139573097, + 0.004906912334263325, + 0.007113358937203884, + 0.005242704413831234, + 0.007545679807662964, + 0.013217258267104626, + 0.009042859077453613, + 0.011636762879788876, + 0.010996612720191479, + 0.005847586318850517, + 0.0057982089929282665, + 0.0071189142763614655, + 0.004784335847944021, + 0.00540007371455431, + 0.0069611878134310246, + 0.008651819080114365, + 0.006498744711279869, + 0.003947441931813955, + 0.008273516781628132, + 0.004304408095777035, + 0.0037433048710227013, + 0.007957265712320805, + 0.013394162990152836, + 0.00875880103558302, + 0.010505037382245064, + 0.007906511425971985, + 0.009634535759687424, + 0.005444606766104698, + 0.006935447920113802, + 0.009198243729770184, + 0.008927436545491219, + 0.00698832655325532, + 0.007993795908987522, + 0.007042369339615107, + 0.0051164510659873486, + 0.004164793062955141, + 0.004953900817781687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.1742957979440689 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020618975162506104 + ], + "opposite_logprob_tokens": [ + -3.8956189155578613 + ], + "logprob_reward": -0.020618975162506104, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020618975162506104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0982721820473671, + 0.03443995863199234, + 0.49385690689086914, + 0.13354483246803284, + 0.05356733873486519, + 0.03739233314990997, + 0.02882305346429348, + 0.06086830049753189, + 0.02945803664624691, + 0.02977704629302025 + ], + "credit_map": [ + 0.008653046563267708, + 0.008829048834741116, + 0.02001115120947361, + 0.016301436349749565, + 0.009791389107704163, + 0.012996582314372063, + 0.011076939292252064, + 0.010612592101097107, + 0.001932745217345655, + 0.0030016452074050903, + 0.00623583747074008, + 0.005949960555881262, + 0.0032656281255185604, + 0.0031381386797875166, + 0.004488660953938961, + 0.006427343934774399, + 0.04001965746283531, + 0.044854000210762024, + 0.06762686371803284, + 0.08383911848068237, + 0.08463741093873978, + 0.048717569559812546, + 0.07680849730968475, + 0.04735380783677101, + 0.009883672930300236, + 0.02356128580868244, + 0.011803359724581242, + 0.020261622965335846, + 0.024173235520720482, + 0.01605326309800148, + 0.015188354067504406, + 0.012620024383068085, + 0.003938730806112289, + 0.008357846178114414, + 0.005821372382342815, + 0.007001094054430723, + 0.01060783676803112, + 0.0057729026302695274, + 0.005074207670986652, + 0.006993347778916359, + 0.0036184480413794518, + 0.0048580835573375225, + 0.0052566854283213615, + 0.005456325598061085, + 0.006063979584723711, + 0.0036088370252400637, + 0.003476005746051669, + 0.00505397142842412, + 0.002800136571750045, + 0.004959799349308014, + 0.004458382725715637, + 0.004681715741753578, + 0.0028201499953866005, + 0.0036141842138022184, + 0.0026920700911432505, + 0.0027966145426034927, + 0.004894947167485952, + 0.009802475571632385, + 0.011070044711232185, + 0.00870642252266407, + 0.00568932481110096, + 0.01076763216406107, + 0.00469455448910594, + 0.005242903251200914, + 0.0031311442144215107, + 0.005010875407606363, + 0.0037614936009049416, + 0.0037310044281184673, + 0.003910393454134464, + 0.0048688980750739574, + 0.0025539840571582317, + 0.002490243874490261, + 0.005543525330722332, + 0.003749280469492078, + 0.004040921106934547, + 0.0044447509571909904, + 0.005003056954592466, + 0.003553644986823201, + 0.0018841690616682172, + 0.0015576989389955997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.020618975162506104 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200941659510136 + ], + "opposite_logprob_tokens": [ + -4.2642011642456055 + ], + "logprob_reward": -0.014200941659510136, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200941659510136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10544409602880478, + 0.10635413229465485, + 0.07471141219139099, + 0.18822215497493744, + 0.12264017760753632, + 0.022562095895409584, + 0.08718486875295639, + 0.16418515145778656, + 0.05946192517876625, + 0.06923399865627289 + ], + "credit_map": [ + 0.007949805818498135, + 0.010030262172222137, + 0.015321943908929825, + 0.012873605825006962, + 0.012103497050702572, + 0.024596063420176506, + 0.012850490398705006, + 0.009718420915305614, + 0.0071274284273386, + 0.013189638033509254, + 0.01906273327767849, + 0.011007905006408691, + 0.013450107537209988, + 0.018387233838438988, + 0.010151064023375511, + 0.013978020288050175, + 0.006746796425431967, + 0.00880676880478859, + 0.009685933589935303, + 0.01091291569173336, + 0.011857577599585056, + 0.008186106570065022, + 0.011306068859994411, + 0.007209241855889559, + 0.010977324098348618, + 0.034796398133039474, + 0.01638680323958397, + 0.021726103499531746, + 0.04037294536828995, + 0.027230607345700264, + 0.019617291167378426, + 0.017114682123064995, + 0.007454622536897659, + 0.016623185947537422, + 0.01548354048281908, + 0.018191715702414513, + 0.026648979634046555, + 0.01371045783162117, + 0.011108501814305782, + 0.013419163413345814, + 0.0023435589391738176, + 0.004636242985725403, + 0.004462930839508772, + 0.0023694175761193037, + 0.002523693023249507, + 0.00212447764351964, + 0.001640153001062572, + 0.0024616210721433163, + 0.0201489869505167, + 0.014793623238801956, + 0.014816906303167343, + 0.009460516273975372, + 0.0074409861117601395, + 0.008001267910003662, + 0.006128458771854639, + 0.0063941217958927155, + 0.01601826213300228, + 0.033023759722709656, + 0.02885952778160572, + 0.01740623451769352, + 0.019008485600352287, + 0.023698896169662476, + 0.010229311883449554, + 0.015940682962536812, + 0.0070425961166620255, + 0.008095444180071354, + 0.006664109416306019, + 0.008212235756218433, + 0.007273871451616287, + 0.010031254962086678, + 0.005157071631401777, + 0.0069853393360972404, + 0.01156893465667963, + 0.010686231777071953, + 0.00962215568870306, + 0.00686299055814743, + 0.01278014201670885, + 0.00827956572175026, + 0.005353128537535667, + 0.004080846905708313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.014200941659510136 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005261858459562063 + ], + "opposite_logprob_tokens": [ + -5.2552618980407715 + ], + "logprob_reward": -0.005261858459562063, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005261858459562063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11054815351963043, + 0.05974729359149933, + 0.22437921166419983, + 0.09497164934873581, + 0.16241733729839325, + 0.04550115391612053, + 0.08156777173280716, + 0.0866071879863739, + 0.08610734343528748, + 0.04815290868282318 + ], + "credit_map": [ + 0.012246196158230305, + 0.011169922538101673, + 0.01636219397187233, + 0.01131251361221075, + 0.01352723315358162, + 0.024575769901275635, + 0.010502966120839119, + 0.0108513617888093, + 0.005508264526724815, + 0.00540159409865737, + 0.013743909075856209, + 0.007832985371351242, + 0.005236754659563303, + 0.0066181146539747715, + 0.0076789106242358685, + 0.007726758718490601, + 0.01860121637582779, + 0.02183631621301174, + 0.02966206707060337, + 0.03912413865327835, + 0.03852703422307968, + 0.024682309478521347, + 0.031311605125665665, + 0.020634518936276436, + 0.0049397605471313, + 0.015630075708031654, + 0.012242177501320839, + 0.014736881479620934, + 0.012170210480690002, + 0.01010194979608059, + 0.015012267976999283, + 0.010138321667909622, + 0.010746745392680168, + 0.024652758613228798, + 0.019048478454351425, + 0.026830701157450676, + 0.028954166918992996, + 0.018506284803152084, + 0.0161980502307415, + 0.0174801554530859, + 0.00408427556976676, + 0.004503085743635893, + 0.011444022879004478, + 0.00450323848053813, + 0.004566884599626064, + 0.00506151607260108, + 0.0032595088705420494, + 0.0080786207690835, + 0.012322457507252693, + 0.010067174211144447, + 0.010930897668004036, + 0.012699351646006107, + 0.007778136059641838, + 0.011889209970831871, + 0.006729287561029196, + 0.009151259437203407, + 0.007421472575515509, + 0.011557810939848423, + 0.011845585890114307, + 0.01952141895890236, + 0.009866812266409397, + 0.008057380095124245, + 0.007971005514264107, + 0.010365698486566544, + 0.01010315865278244, + 0.011571584269404411, + 0.010107788257300854, + 0.011761079542338848, + 0.010744184255599976, + 0.015944352373480797, + 0.00687575014308095, + 0.00899945106357336, + 0.007060244679450989, + 0.00786813348531723, + 0.005497163627296686, + 0.004710455425083637, + 0.01092542428523302, + 0.005045380908995867, + 0.003729749470949173, + 0.003316357033327222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.005261858459562063 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865341633558273 + ], + "opposite_logprob_tokens": [ + -3.0486533641815186 + ], + "logprob_reward": -0.04865341633558273, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865341633558273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16955147683620453, + 0.0528116337954998, + 0.07631746679544449, + 0.06776337325572968, + 0.10802420228719711, + 0.04466328024864197, + 0.16317324340343475, + 0.08727841079235077, + 0.12876516580581665, + 0.10165172070264816 + ], + "credit_map": [ + 0.013521123677492142, + 0.017091400921344757, + 0.039154063910245895, + 0.016712939366698265, + 0.014812661334872246, + 0.03924977779388428, + 0.01109051052480936, + 0.017918989062309265, + 0.004394185729324818, + 0.0059522041119635105, + 0.008803592063486576, + 0.007114557549357414, + 0.007477380335330963, + 0.007284002378582954, + 0.005631412845104933, + 0.006154300644993782, + 0.005282630678266287, + 0.005609757266938686, + 0.00924182590097189, + 0.010702046565711498, + 0.015687553212046623, + 0.01599915511906147, + 0.007254446391016245, + 0.006540058646351099, + 0.0045769247226417065, + 0.010484998114407063, + 0.006508356891572475, + 0.010491457767784595, + 0.008996429853141308, + 0.01070482563227415, + 0.009627544321119785, + 0.006372838746756315, + 0.006541901733726263, + 0.008976063691079617, + 0.023000769317150116, + 0.009506206028163433, + 0.01822425238788128, + 0.013370725326240063, + 0.01133279874920845, + 0.01707148738205433, + 0.007135603576898575, + 0.005398804321885109, + 0.007047245744615793, + 0.004349970258772373, + 0.006959906779229641, + 0.00530711654573679, + 0.003601459553465247, + 0.004863171838223934, + 0.04165540635585785, + 0.026372648775577545, + 0.027372298762202263, + 0.01502974983304739, + 0.015572167932987213, + 0.015555568970739841, + 0.010132190771400928, + 0.011483209207654, + 0.007645192556083202, + 0.01323223952203989, + 0.013503543101251125, + 0.02031802199780941, + 0.0075800009071826935, + 0.009127326309680939, + 0.007915833033621311, + 0.007956252433359623, + 0.018603315576910973, + 0.025442328304052353, + 0.01455923542380333, + 0.017360063269734383, + 0.01597265899181366, + 0.019342396408319473, + 0.007809693459421396, + 0.009675472974777222, + 0.015671715140342712, + 0.01228414662182331, + 0.017832649871706963, + 0.01519063487648964, + 0.011606824584305286, + 0.013690588064491749, + 0.006338659208267927, + 0.00903649814426899 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.04865341633558273 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot, filmed from ground level, captures, with a calm, unadorned perspective, a typical black-and-white police car, its red and blue lights flashing wildly, hurtling down the shoulder of an empty highway at uncontrolled speed towards a solid, weathered, gray concrete pier. At the moment of impact, the front of the car is instantly crushed with astonishing physical precision, the hood folding upward under the immense force, twisting into a heap of scrap metal, while fragments of the front bumper and headlights explode into countless tiny particles and fly in all directions. Then, a thick plume of white steam and black smoke erupts from the ripped-apart engine compartment. The immense momentum causes the rear of the car to briefly lift upward before slamming back to the ground, finally coming to a complete stop with a sharp, grinding sound, a heap of twisted, smoking metal. The entire process takes place under bright but slightly harsh afternoon sunlight. The sharp shadows and high-contrast lighting emphasize the violence and reality of the collision, creating an intense and heart-pounding documentary style.", + "seed": 13959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000007/group_9/index_8699_seed_13959_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1610224585980177, + "vlm_logprob": -0.1610224585980177, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.1610224585980177, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.630426824092865 + ], + "opposite_logprob_tokens": [ + -0.880426824092865 + ], + "logprob_reward": -0.630426824092865, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.630426824092865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13338904082775116, + 0.36818477511405945, + 0.07372191548347473, + 0.08945176750421524, + 0.07543622702360153, + 0.03058725595474243, + 0.049522045999765396, + 0.0918378233909607, + 0.03557765856385231, + 0.05229150503873825 + ], + "credit_map": [ + 0.011037541553378105, + 0.014091246761381626, + 0.01709679886698723, + 0.023351650685071945, + 0.013579793274402618, + 0.020169483497738838, + 0.017781129106879234, + 0.01628139615058899, + 0.01590059883892536, + 0.023797471076250076, + 0.037126339972019196, + 0.02944418229162693, + 0.039808668196201324, + 0.0691894143819809, + 0.030571220442652702, + 0.12234685570001602, + 0.0055710007436573505, + 0.011613565497100353, + 0.008628996089100838, + 0.015201842412352562, + 0.0098832156509161, + 0.008441448211669922, + 0.006962059997022152, + 0.007419790141284466, + 0.007994577288627625, + 0.009313341230154037, + 0.01273477915674448, + 0.01187297236174345, + 0.013001397252082825, + 0.01932714320719242, + 0.008565838448703289, + 0.006641724146902561, + 0.008815676905214787, + 0.01165797095745802, + 0.013843284919857979, + 0.010405691340565681, + 0.008017824031412601, + 0.008743793703615665, + 0.008812902495265007, + 0.00513908825814724, + 0.004133866634219885, + 0.004087382461875677, + 0.0047039068304002285, + 0.006149051710963249, + 0.002564638387411833, + 0.0023844949901103973, + 0.003386643948033452, + 0.0031772712245583534, + 0.004410125780850649, + 0.005855328403413296, + 0.006571038160473108, + 0.005763858091086149, + 0.005474532954394817, + 0.006101719103753567, + 0.008974232710897923, + 0.0063712093979120255, + 0.010317958891391754, + 0.010980664752423763, + 0.009591766633093357, + 0.008722108788788319, + 0.011746852658689022, + 0.011277086101472378, + 0.01716979220509529, + 0.012031595222651958, + 0.0036113925743848085, + 0.0047205532900989056, + 0.004855659324675798, + 0.00734412157908082, + 0.004098646342754364, + 0.0035597370006144047, + 0.004904836881905794, + 0.002482712734490633, + 0.007404920179396868, + 0.006850046571344137, + 0.00518501503393054, + 0.009949089959263802, + 0.006423245649784803, + 0.006477544084191322, + 0.003804860170930624, + 0.0061967819929122925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.630426824092865 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018229445442557335 + ], + "opposite_logprob_tokens": [ + -4.0182294845581055 + ], + "logprob_reward": -0.018229445442557335, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018229445442557335, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10770730674266815, + 0.19851863384246826, + 0.08311005681753159, + 0.07148025184869766, + 0.1048261970281601, + 0.08052550256252289, + 0.12991446256637573, + 0.06770920008420944, + 0.09373900294303894, + 0.06246935576200485 + ], + "credit_map": [ + 0.00833627674728632, + 0.010972094722092152, + 0.012922873720526695, + 0.020409971475601196, + 0.008571635000407696, + 0.015847953036427498, + 0.017516426742076874, + 0.013130090199410915, + 0.007455796469002962, + 0.010705027729272842, + 0.013958762399852276, + 0.012710416689515114, + 0.019576555117964745, + 0.04809829220175743, + 0.009630898013710976, + 0.07638289034366608, + 0.005222624633461237, + 0.009747416712343693, + 0.005485424306243658, + 0.015645893290638924, + 0.01060295570641756, + 0.01672670803964138, + 0.01117324735969305, + 0.008505788631737232, + 0.005110999569296837, + 0.0073522222228348255, + 0.009106618352234364, + 0.008690264075994492, + 0.009366152808070183, + 0.015405286103487015, + 0.008167680352926254, + 0.008281026966869831, + 0.007460406981408596, + 0.014280715957283974, + 0.011824463494122028, + 0.013338632881641388, + 0.018617944791913033, + 0.017520036548376083, + 0.012830208986997604, + 0.008953791111707687, + 0.00892110075801611, + 0.0093165198341012, + 0.008369485847651958, + 0.012342536821961403, + 0.009192105382680893, + 0.007458524778485298, + 0.012843534350395203, + 0.01208170223981142, + 0.010350221768021584, + 0.014243056066334248, + 0.01672588847577572, + 0.015045881271362305, + 0.011115385219454765, + 0.016903888434171677, + 0.027717458084225655, + 0.01781269907951355, + 0.007246415596455336, + 0.007676922250539064, + 0.0065205637365579605, + 0.008503618650138378, + 0.00781463086605072, + 0.007204538211226463, + 0.014240843243896961, + 0.00850167777389288, + 0.008255485445261002, + 0.010639574378728867, + 0.006683992221951485, + 0.015923084691166878, + 0.017852460965514183, + 0.008540486916899681, + 0.018689027056097984, + 0.007154899649322033, + 0.006269547622650862, + 0.007436026353389025, + 0.007614600472152233, + 0.008576037362217903, + 0.009877054020762444, + 0.009169913828372955, + 0.004472174681723118, + 0.00905400700867176 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.018229445442557335 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369740769267082 + ], + "opposite_logprob_tokens": [ + -3.4086973667144775 + ], + "logprob_reward": -0.03369740769267082, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369740769267082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0691114291548729, + 0.12018196284770966, + 0.057054564356803894, + 0.0781516581773758, + 0.05221160501241684, + 0.043833691626787186, + 0.021221527829766273, + 0.03962096944451332, + 0.4553488492965698, + 0.06326370686292648 + ], + "credit_map": [ + 0.005292295943945646, + 0.01036323606967926, + 0.007603360805660486, + 0.008735359646379948, + 0.012131060473620892, + 0.010610883124172688, + 0.007684466894716024, + 0.00669076619669795, + 0.005707006901502609, + 0.00810888409614563, + 0.012054133228957653, + 0.008536258712410927, + 0.00970580242574215, + 0.029342902824282646, + 0.0074197081848979, + 0.03930726274847984, + 0.005612743087112904, + 0.008833901956677437, + 0.007962671108543873, + 0.008609199896454811, + 0.007644078694283962, + 0.007571166846901178, + 0.005480754654854536, + 0.005340051371604204, + 0.006720408331602812, + 0.008103462867438793, + 0.007666562218219042, + 0.00851608719676733, + 0.010230005718767643, + 0.015212898142635822, + 0.013394632376730442, + 0.008307605050504208, + 0.00626032892614603, + 0.007337674498558044, + 0.007541871163994074, + 0.009716966189444065, + 0.004628442227840424, + 0.006925031542778015, + 0.005707258824259043, + 0.004094033036381006, + 0.005578265525400639, + 0.005984592717140913, + 0.006032673642039299, + 0.007816432043910027, + 0.0038831597194075584, + 0.004072846379131079, + 0.005815944168716669, + 0.004649781621992588, + 0.002321337815374136, + 0.002703443868085742, + 0.003335274988785386, + 0.002480424242094159, + 0.0020311372354626656, + 0.002025321824476123, + 0.003897062735632062, + 0.0024275234900414944, + 0.003739770734682679, + 0.004442778881639242, + 0.0038861564826220274, + 0.005425720941275358, + 0.004389260429888964, + 0.004035635851323605, + 0.007161410059779882, + 0.006540238857269287, + 0.036554474383592606, + 0.05557142570614815, + 0.028941353783011436, + 0.061599306762218475, + 0.10579630732536316, + 0.03543306514620781, + 0.10137733072042465, + 0.030075622722506523, + 0.008400027640163898, + 0.008096745237708092, + 0.008066195994615555, + 0.008686088025569916, + 0.007789856754243374, + 0.010679581202566624, + 0.00398329459130764, + 0.007561916951090097 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.03369740769267082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792622178792953 + ], + "opposite_logprob_tokens": [ + -4.634792804718018 + ], + "logprob_reward": -0.009792622178792953, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792622178792953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10496719181537628, + 0.3593791127204895, + 0.044216338545084, + 0.05235232040286064, + 0.04262830317020416, + 0.04032154008746147, + 0.045190248638391495, + 0.03229014575481415, + 0.19249477982521057, + 0.08616001158952713 + ], + "credit_map": [ + 0.009178634732961655, + 0.015555750578641891, + 0.013289861381053925, + 0.01885288953781128, + 0.008655163459479809, + 0.01522967778146267, + 0.012475785799324512, + 0.011729425750672817, + 0.012983918190002441, + 0.016307786107063293, + 0.018311843276023865, + 0.01901240088045597, + 0.03811507299542427, + 0.09283285588026047, + 0.012955423444509506, + 0.14885979890823364, + 0.004647294990718365, + 0.008212677203118801, + 0.0059400261379778385, + 0.007430497091263533, + 0.0055331336334347725, + 0.004714362323284149, + 0.0036152752581983805, + 0.004123068414628506, + 0.005279352888464928, + 0.006567937321960926, + 0.013148222118616104, + 0.006082900799810886, + 0.005695280618965626, + 0.00555761344730854, + 0.005521733779460192, + 0.0044992719776928425, + 0.0050354828126728535, + 0.006664438638836145, + 0.008189836516976357, + 0.007601422257721424, + 0.003929018974304199, + 0.004053929354995489, + 0.004822009243071079, + 0.0023321614135056734, + 0.005431866738945246, + 0.0053456127643585205, + 0.006203657016158104, + 0.008186880499124527, + 0.0035934692714363337, + 0.0031697656959295273, + 0.00402854336425662, + 0.004361743573099375, + 0.003690112382173538, + 0.005745965987443924, + 0.005750732496380806, + 0.0073445942252874374, + 0.005160108674317598, + 0.004677081946283579, + 0.007622709032148123, + 0.0051989369094371796, + 0.0033658002503216267, + 0.003852853551506996, + 0.0037990726996213198, + 0.003922359086573124, + 0.004661853425204754, + 0.003324211109429598, + 0.005118506960570812, + 0.004245485179126263, + 0.016819922253489494, + 0.02441490814089775, + 0.013691139407455921, + 0.01901303231716156, + 0.04967835918068886, + 0.01524943858385086, + 0.03976697847247124, + 0.013860988430678844, + 0.011287469416856766, + 0.01108533889055252, + 0.01275501400232315, + 0.016516396775841713, + 0.010863656178116798, + 0.009877912700176239, + 0.007652392145246267, + 0.006121835205703974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.009792622178792953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene filmed from a static, ground-level long shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a black-and-white police car with flashing red and blue lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the car crash into a solid, weathered, gray concrete pier?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the impact show the front of the car crushing, the hood folding upward, and fragments exploding into particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene lit by bright, slightly harsh afternoon sunlight creating sharp shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296599358320236 + ], + "opposite_logprob_tokens": [ + -2.2379660606384277 + ], + "logprob_reward": -0.11296599358320236, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296599358320236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17415103316307068, + 0.2979448139667511, + 0.09431742876768112, + 0.0461343452334404, + 0.11168166249990463, + 0.047929976135492325, + 0.058291833847761154, + 0.06733734160661697, + 0.048302777111530304, + 0.05390877276659012 + ], + "credit_map": [ + 0.01840166002511978, + 0.02363278716802597, + 0.02125558629631996, + 0.027402006089687347, + 0.019840853288769722, + 0.02424878254532814, + 0.018117137253284454, + 0.0212522242218256, + 0.013625109568238258, + 0.025195203721523285, + 0.02109869197010994, + 0.017026077955961227, + 0.031781867146492004, + 0.0670645609498024, + 0.01311564538627863, + 0.10903766751289368, + 0.008990872651338577, + 0.016648342832922935, + 0.011012103408575058, + 0.01372001226991415, + 0.01281851064413786, + 0.01188606210052967, + 0.009742491878569126, + 0.009499040432274342, + 0.0054540811106562614, + 0.004587913863360882, + 0.00669310474768281, + 0.008317505940794945, + 0.004604756832122803, + 0.005352979060262442, + 0.00524468207731843, + 0.005879322998225689, + 0.013686630874872208, + 0.017198162153363228, + 0.02176079899072647, + 0.022465841844677925, + 0.006718269549310207, + 0.009670708328485489, + 0.012962831184267998, + 0.007218421436846256, + 0.005441323854029179, + 0.007685433141887188, + 0.007787346374243498, + 0.008572892285883427, + 0.003827183274552226, + 0.00386068900115788, + 0.0060026622377336025, + 0.004752448294311762, + 0.0073575847782194614, + 0.00925352331250906, + 0.014002695679664612, + 0.00903412140905857, + 0.0037656710483133793, + 0.003263310994952917, + 0.006565493997186422, + 0.005049435887485743, + 0.008730627596378326, + 0.008307406678795815, + 0.007257393561303616, + 0.008791790343821049, + 0.007243017200380564, + 0.007805295754224062, + 0.009958305396139622, + 0.009243513457477093, + 0.005587633233517408, + 0.005068326834589243, + 0.0039201765321195126, + 0.007280523888766766, + 0.007712531369179487, + 0.004895450081676245, + 0.009444544091820717, + 0.004393593408167362, + 0.007301864214241505, + 0.006684092339128256, + 0.0065712048672139645, + 0.009481594897806644, + 0.006619056686758995, + 0.007069780491292477, + 0.0044769104570150375, + 0.005704275798052549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.11296599358320236 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_tensor([3943, 3943])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_tensor([3943, 3943])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4ed93f6f15f4ca1e3607e2b211128432636d8897 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_tensor([3943, 3943])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, extreme macro shot delves into an intimate group of luminous pearls, huddled against a dark, non-reflective surface that absorbs all light. The camera maintains tight, unwavering focus, revealing each pearl's subtle flaws and unique characteristics. Their surfaces are rendered in surreal detail, not just white but an intricate tapestry of iridescent pearls, shimmering with an ethereal glow that seems to emanate from within. A single, sharp key light from a high side angle sculpts their spherical forms, creating brilliant, almost blown-out highlights at their peaks and casting deep, softly-edged shadows where they press against one another. The visual narrative unfolds in a rapid sequence of crisp, high-contrast photographic frames, each staccato cut revealing a subtle shift in light, causing subtle overtones of pink, lavender, and mint green to shimmer and ripple across the pearls' smooth surfaces. This rhythmic, stop-motion-like progression creates a mesmerizing and hypnotic atmosphere, celebrating the objects' sheer tactile beauty with a focused, almost reverent approach.", + "seed": 33139, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_3943_seed_33139_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3520846668630838, + "vlm_logprob": -0.3520846668630838, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.3520846668630838, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004706257954239845 + ], + "opposite_logprob_tokens": [ + -6.504706382751465 + ], + "logprob_reward": -0.004706257954239845, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.004706257954239845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13225549459457397, + 0.16890163719654083, + 0.1836780160665512, + 0.04604025557637215, + 0.1400785893201828, + 0.10782753676176071, + 0.03481251001358032, + 0.03910726681351662, + 0.042780663818120956, + 0.10451807081699371 + ], + "credit_map": [ + 0.012020045891404152, + 0.01305727194994688, + 0.03882161155343056, + 0.014905373565852642, + 0.022611362859606743, + 0.006146484520286322, + 0.015011025592684746, + 0.009682311676442623, + 0.014980686828494072, + 0.02332231216132641, + 0.02383180521428585, + 0.03835400938987732, + 0.021718705072999, + 0.01283089816570282, + 0.017836173996329308, + 0.01602703332901001, + 0.011864240281283855, + 0.054657772183418274, + 0.05199664086103439, + 0.017642658203840256, + 0.016669122502207756, + 0.009568016976118088, + 0.011911542154848576, + 0.009368023835122585, + 0.004260598681867123, + 0.004218115471303463, + 0.008169549517333508, + 0.006556966342031956, + 0.005223384592682123, + 0.0030360950622707605, + 0.009000190533697605, + 0.005575352814048529, + 0.01667439192533493, + 0.01621244288980961, + 0.039363108575344086, + 0.016335729509592056, + 0.012169373221695423, + 0.015000108629465103, + 0.014269748702645302, + 0.010053668171167374, + 0.011821752414107323, + 0.013087231665849686, + 0.017966480925679207, + 0.024511151015758514, + 0.010850160382688046, + 0.012401885353028774, + 0.010930459015071392, + 0.006258411332964897, + 0.003738295752555132, + 0.0034996813628822565, + 0.006329377647489309, + 0.0058084288612008095, + 0.005017005372792482, + 0.0031651360914111137, + 0.0048372782766819, + 0.0024173029232770205, + 0.003181089647114277, + 0.004949560388922691, + 0.006549858022481203, + 0.006100434344261885, + 0.005116064101457596, + 0.003966903313994408, + 0.005157317966222763, + 0.004086037632077932, + 0.004791095852851868, + 0.004360273014754057, + 0.008517037145793438, + 0.007010060828179121, + 0.0044808657839894295, + 0.0040310765616595745, + 0.005214707460254431, + 0.004375547636300325, + 0.007569687906652689, + 0.010878034867346287, + 0.010399938561022282, + 0.013852991163730621, + 0.028454577550292015, + 0.007768298499286175, + 0.017613064497709274, + 0.007981471717357635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 4 + ], + "selected_reward": -0.004706257954239845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4288897216320038 + ], + "opposite_logprob_tokens": [ + -1.0538897514343262 + ], + "logprob_reward": -0.4288897216320038, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4288897216320038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18545307219028473, + 0.06614022701978683, + 0.05680493265390396, + 0.05574776232242584, + 0.3438901901245117, + 0.051519520580768585, + 0.04686189070343971, + 0.07860460132360458, + 0.051582805812358856, + 0.06339498609304428 + ], + "credit_map": [ + 0.04113181307911873, + 0.028564315289258957, + 0.01930495910346508, + 0.021335456520318985, + 0.025219431146979332, + 0.012278188019990921, + 0.018236685544252396, + 0.019382234662771225, + 0.006357461679726839, + 0.007007434032857418, + 0.00939867366105318, + 0.012760289013385773, + 0.010474596172571182, + 0.006449731066823006, + 0.006188513245433569, + 0.007503530476242304, + 0.003433853853493929, + 0.014226919040083885, + 0.014734157361090183, + 0.004682381171733141, + 0.006476290989667177, + 0.0034536353778094053, + 0.00474106939509511, + 0.005056624300777912, + 0.0060322643257677555, + 0.005385562311857939, + 0.008353705517947674, + 0.007407120428979397, + 0.00900423526763916, + 0.004456688184291124, + 0.009144201874732971, + 0.005963984411209822, + 0.042386580258607864, + 0.04285440593957901, + 0.10086829960346222, + 0.03708413988351822, + 0.02997441031038761, + 0.03435530140995979, + 0.029599320143461227, + 0.026767723262310028, + 0.005959260743111372, + 0.00646542850881815, + 0.007550036069005728, + 0.00934402272105217, + 0.006172895431518555, + 0.007193176541477442, + 0.005626916419714689, + 0.0032077839132398367, + 0.004501451272517443, + 0.004919050261378288, + 0.006399726029485464, + 0.00786686036735773, + 0.006678332574665546, + 0.002809974830597639, + 0.010028096847236156, + 0.0036584006156772375, + 0.006500618066638708, + 0.009055794216692448, + 0.011115290224552155, + 0.01659679040312767, + 0.00941856112331152, + 0.00738506019115448, + 0.011536399833858013, + 0.006996092852205038, + 0.005642932374030352, + 0.004398769699037075, + 0.009822837077081203, + 0.009995169937610626, + 0.005874310154467821, + 0.004510422237217426, + 0.006526951678097248, + 0.004811413120478392, + 0.004078249912708998, + 0.006591931451112032, + 0.00667541055008769, + 0.007675561122596264, + 0.01457761600613594, + 0.00485139200463891, + 0.012500480748713017, + 0.006444346625357866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.4288897216320038 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261864185333252 + ], + "opposite_logprob_tokens": [ + -0.5761864185333252 + ], + "logprob_reward": -0.8261864185333252, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261864185333252, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1701619178056717, + 0.14857672154903412, + 0.17281655967235565, + 0.03473471850156784, + 0.1275823414325714, + 0.061243269592523575, + 0.023234397172927856, + 0.028247332200407982, + 0.16450431942939758, + 0.06889839470386505 + ], + "credit_map": [ + 0.013347949832677841, + 0.030975477769970894, + 0.055109187960624695, + 0.018839778378605843, + 0.019831866025924683, + 0.006759230047464371, + 0.014086293987929821, + 0.011212133802473545, + 0.01055409386754036, + 0.03950512781739235, + 0.02428480237722397, + 0.024390868842601776, + 0.016587132588028908, + 0.009138072840869427, + 0.012739802710711956, + 0.011376813985407352, + 0.008267123252153397, + 0.04937170445919037, + 0.05585086718201637, + 0.013192448765039444, + 0.013041069731116295, + 0.0068564941175282, + 0.014742501080036163, + 0.011494341306388378, + 0.005203206557780504, + 0.003648296231403947, + 0.004183141514658928, + 0.007421115413308144, + 0.004184165969491005, + 0.001970987068489194, + 0.004774688743054867, + 0.003349117934703827, + 0.020370805636048317, + 0.010784756392240524, + 0.030899781733751297, + 0.01893245242536068, + 0.008366559632122517, + 0.009032986126840115, + 0.013585791923105717, + 0.01560919638723135, + 0.007003554608672857, + 0.006909817457199097, + 0.010660977102816105, + 0.011877787299454212, + 0.006839165464043617, + 0.007643207907676697, + 0.007019486743956804, + 0.003289273241534829, + 0.0026819307822734118, + 0.002535543404519558, + 0.0028225404676049948, + 0.004200043622404337, + 0.002859627828001976, + 0.0019761305302381516, + 0.0035477045457810163, + 0.0026108764577656984, + 0.002745729172602296, + 0.0037710757460445166, + 0.003716518869623542, + 0.004871696699410677, + 0.0037483847700059414, + 0.0027570906095206738, + 0.004066120367497206, + 0.002570715267211199, + 0.015477592125535011, + 0.018286019563674927, + 0.024149131029844284, + 0.02155952900648117, + 0.026049498468637466, + 0.018906263634562492, + 0.028360093012452126, + 0.011716188862919807, + 0.004426134284585714, + 0.004768034443259239, + 0.007176555227488279, + 0.008580761030316353, + 0.014520217664539814, + 0.004595709498971701, + 0.017498275265097618, + 0.0073327128775417805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.8261864185333252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354439556598663 + ], + "opposite_logprob_tokens": [ + -3.6513545513153076 + ], + "logprob_reward": -0.026354439556598663, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354439556598663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11460170894861221, + 0.09310054779052734, + 0.3608089089393616, + 0.031921304762363434, + 0.08066712319850922, + 0.052747298032045364, + 0.013036093674600124, + 0.02440248429775238, + 0.03647827357053757, + 0.19223619997501373 + ], + "credit_map": [ + 0.015816232189536095, + 0.02155463397502899, + 0.030425697565078735, + 0.008513144217431545, + 0.012283200398087502, + 0.004513219930231571, + 0.011995146051049232, + 0.009500443004071712, + 0.011258263140916824, + 0.014210113324224949, + 0.012196063064038754, + 0.020028557628393173, + 0.010564153082668781, + 0.00647148210555315, + 0.01038284506648779, + 0.007989073172211647, + 0.016467325389385223, + 0.11393003165721893, + 0.1165853887796402, + 0.027674591168761253, + 0.028381112962961197, + 0.013823765330016613, + 0.021242788061499596, + 0.02270393632352352, + 0.0039385403506457806, + 0.0037140187341719866, + 0.004191090818494558, + 0.004858496133238077, + 0.003918266389518976, + 0.003217601217329502, + 0.004913699813187122, + 0.0031695901416242123, + 0.009536062367260456, + 0.00896663498133421, + 0.01567644253373146, + 0.013190923258662224, + 0.007114842999726534, + 0.007117942441254854, + 0.011379795148968697, + 0.007684482727199793, + 0.005862279795110226, + 0.006989032495766878, + 0.009286115877330303, + 0.00895735714584589, + 0.005661449395120144, + 0.005801673047244549, + 0.007033588830381632, + 0.003155806567519903, + 0.0017068448942154646, + 0.0015275916084647179, + 0.001391058904118836, + 0.001634271815419197, + 0.0018447202164679766, + 0.0010339344153180718, + 0.0025447215884923935, + 0.0013529506977647543, + 0.0023385784588754177, + 0.004022813402116299, + 0.002675320254638791, + 0.005344730336219072, + 0.0025516734458506107, + 0.0029756997246295214, + 0.0023402185179293156, + 0.0021534517873078585, + 0.003968113102018833, + 0.004265200346708298, + 0.006200837437063456, + 0.0041445898823440075, + 0.0057949707843363285, + 0.0033894346561282873, + 0.00606480473652482, + 0.002650324022397399, + 0.012532451190054417, + 0.019410211592912674, + 0.021828021854162216, + 0.024930274114012718, + 0.026935629546642303, + 0.013747583143413067, + 0.04807242378592491, + 0.02477959729731083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.026354439556598663 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742864966392517 + ], + "opposite_logprob_tokens": [ + -0.9742864966392517 + ], + "logprob_reward": -0.4742864966392517, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742864966392517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2263752967119217, + 0.2241159826517105, + 0.1303972601890564, + 0.026807937771081924, + 0.025234706699848175, + 0.041577354073524475, + 0.03531372919678688, + 0.0351608507335186, + 0.21546322107315063, + 0.03955370560288429 + ], + "credit_map": [ + 0.0157502219080925, + 0.04279033839702606, + 0.07008662074804306, + 0.022323157638311386, + 0.029287628829479218, + 0.011297996155917645, + 0.015845950692892075, + 0.018993349745869637, + 0.018831539899110794, + 0.037254393100738525, + 0.0376119427382946, + 0.04640965908765793, + 0.027543993666768074, + 0.01324066985398531, + 0.022283047437667847, + 0.020940721035003662, + 0.007224664557725191, + 0.04024196416139603, + 0.04643293842673302, + 0.006988928187638521, + 0.008888489566743374, + 0.0051558418199419975, + 0.007888847030699253, + 0.007575568277388811, + 0.004890326876193285, + 0.002942855004221201, + 0.003404782386496663, + 0.0034196083433926105, + 0.002890804084017873, + 0.0026548474561423063, + 0.003336745547130704, + 0.0032679664436727762, + 0.00270244968123734, + 0.003257251577451825, + 0.005080383270978928, + 0.003572809277102351, + 0.002286758739501238, + 0.0025021228939294815, + 0.0037590134888887405, + 0.0020739161409437656, + 0.003909483551979065, + 0.0031161538790911436, + 0.010658609680831432, + 0.008273882791399956, + 0.005771327298134565, + 0.0037689104210585356, + 0.0033238325268030167, + 0.0027551488019526005, + 0.0038340918254107237, + 0.003702629590407014, + 0.004480300005525351, + 0.004459483548998833, + 0.006380993407219648, + 0.0029203782323747873, + 0.0057372753508389, + 0.003798572113737464, + 0.002888215472921729, + 0.004182196222245693, + 0.004723359830677509, + 0.0057653700932860374, + 0.005640460643917322, + 0.0034143878147006035, + 0.004976590164005756, + 0.0035702669993042946, + 0.027769705280661583, + 0.01772751472890377, + 0.046697523444890976, + 0.03809158131480217, + 0.020822182297706604, + 0.022989418357610703, + 0.026191862300038338, + 0.015173407271504402, + 0.003866414539515972, + 0.005976846441626549, + 0.0045031532645225525, + 0.006476700305938721, + 0.004978251177817583, + 0.00375408586114645, + 0.006317516788840294, + 0.003680729540064931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.4742864966392517 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "No", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, extreme macro shot delves into an intimate group of luminous pearls, huddled against a dark, non-reflective surface that absorbs all light. The camera maintains tight, unwavering focus, revealing each pearl's subtle flaws and unique characteristics. Their surfaces are rendered in surreal detail, not just white but an intricate tapestry of iridescent pearls, shimmering with an ethereal glow that seems to emanate from within. A single, sharp key light from a high side angle sculpts their spherical forms, creating brilliant, almost blown-out highlights at their peaks and casting deep, softly-edged shadows where they press against one another. The visual narrative unfolds in a rapid sequence of crisp, high-contrast photographic frames, each staccato cut revealing a subtle shift in light, causing subtle overtones of pink, lavender, and mint green to shimmer and ripple across the pearls' smooth surfaces. This rhythmic, stop-motion-like progression creates a mesmerizing and hypnotic atmosphere, celebrating the objects' sheer tactile beauty with a focused, almost reverent approach.", + "seed": 33140, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_3943_seed_33140_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.9731247995048762, + "vlm_logprob": -0.9731247995048762, + "vlm_yes_no_margin": 0.75, + "vlm_token_credit": -0.9731247995048762, + "vlm_yes_ratio": 0.4, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015726706013083458 + ], + "opposite_logprob_tokens": [ + -4.515726566314697 + ], + "logprob_reward": -0.015726706013083458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.015726706013083458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10618773102760315, + 0.05440820753574371, + 0.049822498112916946, + 0.1449129432439804, + 0.037428345531225204, + 0.015503659844398499, + 0.09617707133293152, + 0.35598599910736084, + 0.1001545861363411, + 0.03941887989640236 + ], + "credit_map": [ + 0.01185683161020279, + 0.011743803508579731, + 0.011487283743917942, + 0.020825644955039024, + 0.008691588416695595, + 0.0108095183968544, + 0.013179564848542213, + 0.017593497410416603, + 0.010579558089375496, + 0.005413838662207127, + 0.006423885468393564, + 0.009258249774575233, + 0.005485231522470713, + 0.003619159571826458, + 0.007099708542227745, + 0.006528580095618963, + 0.002933579497039318, + 0.0024360883980989456, + 0.015170513652265072, + 0.014516551047563553, + 0.003881594631820917, + 0.0015093432739377022, + 0.005657648667693138, + 0.0037171824369579554, + 0.015391794964671135, + 0.011413459666073322, + 0.019071945920586586, + 0.026798943057656288, + 0.010320430621504784, + 0.02118632383644581, + 0.016662536188960075, + 0.02406751736998558, + 0.0024016171228140593, + 0.0025645755231380463, + 0.006769384723156691, + 0.007444542367011309, + 0.002730855019763112, + 0.005074908956885338, + 0.004774731118232012, + 0.005667729768902063, + 0.001288220169954002, + 0.001181079656817019, + 0.002411419525742531, + 0.0025931005366146564, + 0.0013819466112181544, + 0.001367853139527142, + 0.0033115018159151077, + 0.001968540484085679, + 0.006757277995347977, + 0.00676828483119607, + 0.01118393149226904, + 0.015899335965514183, + 0.013232514262199402, + 0.008477688767015934, + 0.02052929624915123, + 0.013328758999705315, + 0.011257761158049107, + 0.01335756666958332, + 0.025978850200772285, + 0.1584187000989914, + 0.019677238538861275, + 0.012303011491894722, + 0.06381607800722122, + 0.05117683485150337, + 0.0065650031901896, + 0.008608539588749409, + 0.022001683712005615, + 0.015162411145865917, + 0.008855355903506279, + 0.014719710685312748, + 0.013476960361003876, + 0.01076493039727211, + 0.0043715196661651134, + 0.002246054820716381, + 0.0070076193660497665, + 0.006783024873584509, + 0.004700489342212677, + 0.003334960900247097, + 0.0037928475067019463, + 0.007182365749031305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.015726706013083458 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.695005416870117 + ], + "opposite_logprob_tokens": [ + -0.07000547647476196 + ], + "logprob_reward": -2.695005416870117, + "yes_no_margin_reward": -2.625, + "token_credit_reward": -2.695005416870117, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.09092560410499573, + 0.08890704810619354, + 0.12511847913265228, + 0.08457431197166443, + 0.06464225053787231, + 0.028289947658777237, + 0.09361164271831512, + 0.3080984950065613, + 0.03547288104891777, + 0.08035928010940552 + ], + "credit_map": [ + 0.00973364058881998, + 0.007326433900743723, + 0.0110038286074996, + 0.015802601352334023, + 0.008220450952649117, + 0.01013488881289959, + 0.012449607253074646, + 0.016254156827926636, + 0.013054804876446724, + 0.008426819927990437, + 0.010977723635733128, + 0.012990375980734825, + 0.007621476426720619, + 0.0074618165381252766, + 0.011796482838690281, + 0.016577549278736115, + 0.011249152943491936, + 0.007943440228700638, + 0.028882160782814026, + 0.03262710943818092, + 0.014878037385642529, + 0.008852889761328697, + 0.012179862707853317, + 0.008505836129188538, + 0.006540860515087843, + 0.004895536694675684, + 0.007265158463269472, + 0.006301203742623329, + 0.0047748195938766, + 0.005196773447096348, + 0.026331370696425438, + 0.02326859161257744, + 0.004290103912353516, + 0.0044931937009096146, + 0.011389052495360374, + 0.01295334380120039, + 0.004538813140243292, + 0.006145017221570015, + 0.01167990267276764, + 0.009152824990451336, + 0.00170916598290205, + 0.001828487729653716, + 0.0032742584589868784, + 0.004148708190768957, + 0.0021725494880229235, + 0.0023515173234045506, + 0.008747520856559277, + 0.004057740792632103, + 0.006585431285202503, + 0.005844829138368368, + 0.011375602334737778, + 0.014294314198195934, + 0.012731894850730896, + 0.007490560412406921, + 0.020798029378056526, + 0.014490979723632336, + 0.01006918866187334, + 0.010567477904260159, + 0.019083334133028984, + 0.13691836595535278, + 0.01579195447266102, + 0.011202539317309856, + 0.06273193657398224, + 0.04173373058438301, + 0.002793567255139351, + 0.004697217606008053, + 0.005755057092756033, + 0.004241279326379299, + 0.003466458525508642, + 0.005069765727967024, + 0.004940102808177471, + 0.004509433172643185, + 0.008046797476708889, + 0.006642748136073351, + 0.010947681963443756, + 0.010378632694482803, + 0.009770807810127735, + 0.012608022429049015, + 0.010586867108941078, + 0.011377736926078796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -2.695005416870117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9745762348175049 + ], + "opposite_logprob_tokens": [ + -0.4745762348175049 + ], + "logprob_reward": -0.9745762348175049, + "yes_no_margin_reward": -0.5, + "token_credit_reward": -0.9745762348175049, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1636306196451187, + 0.05401720106601715, + 0.11269492655992508, + 0.05584044009447098, + 0.149790421128273, + 0.02198600396513939, + 0.09418045729398727, + 0.15090352296829224, + 0.11146432906389236, + 0.0854921042919159 + ], + "credit_map": [ + 0.015047091990709305, + 0.012400052510201931, + 0.013571321032941341, + 0.04319831728935242, + 0.01133060920983553, + 0.013990765437483788, + 0.024066345766186714, + 0.030026111751794815, + 0.00742994574829936, + 0.007044176571071148, + 0.008514942601323128, + 0.006612923461943865, + 0.006153435911983252, + 0.005889417137950659, + 0.005259460769593716, + 0.007112897001206875, + 0.013832141645252705, + 0.009943396784365177, + 0.019073830917477608, + 0.01296291034668684, + 0.017235923558473587, + 0.012410392053425312, + 0.013477219268679619, + 0.013759108260273933, + 0.003992557525634766, + 0.004482552409172058, + 0.006746573373675346, + 0.014258150942623615, + 0.004426977131515741, + 0.004526550881564617, + 0.010151699185371399, + 0.007255373522639275, + 0.008468307554721832, + 0.008488972671329975, + 0.015076270326972008, + 0.03329019993543625, + 0.007719129789620638, + 0.021548178046941757, + 0.04312826693058014, + 0.012071114033460617, + 0.0017731365514919162, + 0.00193775724619627, + 0.0018971894169226289, + 0.005299283657222986, + 0.001938889967277646, + 0.002351870760321617, + 0.0038127643056213856, + 0.0029751111287623644, + 0.006626455578953028, + 0.005617436487227678, + 0.01491885632276535, + 0.01658349111676216, + 0.011018948629498482, + 0.008507298305630684, + 0.01881794072687626, + 0.012090031057596207, + 0.0066519626416265965, + 0.006203989963978529, + 0.013608590699732304, + 0.0646272599697113, + 0.010430943220853806, + 0.0058648730628192425, + 0.01977904886007309, + 0.02373686246573925, + 0.013426098972558975, + 0.012115756049752235, + 0.020739540457725525, + 0.016077224165201187, + 0.010252737440168858, + 0.016471320763230324, + 0.010857908055186272, + 0.011523745022714138, + 0.008389675989747047, + 0.008413629606366158, + 0.012312847189605236, + 0.011495023034512997, + 0.014389327727258205, + 0.013986854813992977, + 0.007332733366638422, + 0.009172013029456139 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.9745762348175049 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04314937815070152 + ], + "opposite_logprob_tokens": [ + -3.168149471282959 + ], + "logprob_reward": -0.04314937815070152, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04314937815070152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24619275331497192, + 0.09814297407865524, + 0.08117491751909256, + 0.03992846980690956, + 0.06740850955247879, + 0.030982784926891327, + 0.06624115258455276, + 0.2348155677318573, + 0.09858513623476028, + 0.03652775287628174 + ], + "credit_map": [ + 0.018961910158395767, + 0.017119497060775757, + 0.01494656689465046, + 0.07753488421440125, + 0.014570290222764015, + 0.015055842697620392, + 0.05463457107543945, + 0.03336917236447334, + 0.007874077185988426, + 0.006895502097904682, + 0.01111698430031538, + 0.021847277879714966, + 0.010163161903619766, + 0.007477594073861837, + 0.01609734632074833, + 0.01667102426290512, + 0.004074780270457268, + 0.004966116044670343, + 0.02493240125477314, + 0.01985253021121025, + 0.005104222800582647, + 0.004977205768227577, + 0.008671860210597515, + 0.008595787920057774, + 0.004751373082399368, + 0.002781928051263094, + 0.004482925403863192, + 0.0028162712696939707, + 0.0031250864267349243, + 0.003639510367065668, + 0.008856350556015968, + 0.009475021623075008, + 0.0051185511983931065, + 0.005209070630371571, + 0.009833522140979767, + 0.0073719448409974575, + 0.00678462628275156, + 0.008017869666218758, + 0.01571110635995865, + 0.00936181377619505, + 0.0026593240909278393, + 0.002906401176005602, + 0.004391919821500778, + 0.006643233355134726, + 0.0033996300771832466, + 0.0031252142507582903, + 0.004398623015731573, + 0.0034584372770041227, + 0.004962834995239973, + 0.004584548994898796, + 0.008929301984608173, + 0.011483374983072281, + 0.008989902213215828, + 0.005927982274442911, + 0.014134123921394348, + 0.007229081820696592, + 0.007582911755889654, + 0.008463500998914242, + 0.01573372446000576, + 0.10458088666200638, + 0.012696555815637112, + 0.008169947192072868, + 0.044752418994903564, + 0.032835617661476135, + 0.00841197744011879, + 0.009501973167061806, + 0.02115863747894764, + 0.014096635393798351, + 0.010974434204399586, + 0.009531556628644466, + 0.017224574461579323, + 0.007685346063226461, + 0.002854224992915988, + 0.0033992459066212177, + 0.006017500534653664, + 0.005201476626098156, + 0.00410565547645092, + 0.0039794715121388435, + 0.0037862113676965237, + 0.007183963432908058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.04314937815070152 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371662616729736 + ], + "opposite_logprob_tokens": [ + -0.38716626167297363 + ], + "logprob_reward": -1.1371662616729736, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371662616729736, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1320318579673767, + 0.05772896483540535, + 0.06635169684886932, + 0.06243954226374626, + 0.07030285894870758, + 0.03442602604627609, + 0.07940889894962311, + 0.25338155031204224, + 0.1287563592195511, + 0.11517220735549927 + ], + "credit_map": [ + 0.014048048295080662, + 0.012635481543838978, + 0.01237547118216753, + 0.026254775002598763, + 0.011371636763215065, + 0.01542449090629816, + 0.01673637516796589, + 0.023185566067695618, + 0.00877122487872839, + 0.007899105548858643, + 0.006473167799413204, + 0.006649431772530079, + 0.007035757880657911, + 0.005639015696942806, + 0.007160907611250877, + 0.008100352250039577, + 0.005366010591387749, + 0.00585926603525877, + 0.01388078648597002, + 0.01411505974829197, + 0.009749863296747208, + 0.0052535925060510635, + 0.006787527818232775, + 0.005339590832591057, + 0.004340945743024349, + 0.003604464465752244, + 0.005857154261320829, + 0.020071983337402344, + 0.0029483221005648375, + 0.004294991958886385, + 0.015510797500610352, + 0.0058108847588300705, + 0.006686759646981955, + 0.005932605359703302, + 0.00802451279014349, + 0.008382664062082767, + 0.0057289935648441315, + 0.008483699522912502, + 0.016159074380993843, + 0.010904552415013313, + 0.0026690056547522545, + 0.0023456374183297157, + 0.0035039393696933985, + 0.009016592055559158, + 0.002422379795461893, + 0.0031400180887430906, + 0.005842125508934259, + 0.005486328154802322, + 0.004863767884671688, + 0.004690123256295919, + 0.01333319116383791, + 0.012356624007225037, + 0.009544073604047298, + 0.006623978726565838, + 0.015786105766892433, + 0.012211035937070847, + 0.008234831504523754, + 0.008571029640734196, + 0.016147760674357414, + 0.11306515336036682, + 0.013604438863694668, + 0.00930841639637947, + 0.04968797415494919, + 0.034761942923069, + 0.01209723949432373, + 0.012189209461212158, + 0.02695058099925518, + 0.02116890251636505, + 0.012044963426887989, + 0.019022271037101746, + 0.013049453496932983, + 0.012233740650117397, + 0.013498477637767792, + 0.007443050388246775, + 0.018156811594963074, + 0.015216018073260784, + 0.014867809601128101, + 0.014540253207087517, + 0.011490277014672756, + 0.019959505647420883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -1.1371662616729736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "No", + "Yes", + "No" + ] + }, + { + "prompt": "static, extreme macro shot delves into an intimate group of luminous pearls, huddled against a dark, non-reflective surface that absorbs all light. The camera maintains tight, unwavering focus, revealing each pearl's subtle flaws and unique characteristics. Their surfaces are rendered in surreal detail, not just white but an intricate tapestry of iridescent pearls, shimmering with an ethereal glow that seems to emanate from within. A single, sharp key light from a high side angle sculpts their spherical forms, creating brilliant, almost blown-out highlights at their peaks and casting deep, softly-edged shadows where they press against one another. The visual narrative unfolds in a rapid sequence of crisp, high-contrast photographic frames, each staccato cut revealing a subtle shift in light, causing subtle overtones of pink, lavender, and mint green to shimmer and ripple across the pearls' smooth surfaces. This rhythmic, stop-motion-like progression creates a mesmerizing and hypnotic atmosphere, celebrating the objects' sheer tactile beauty with a focused, almost reverent approach.", + "seed": 33141, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_3943_seed_33141_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.5796438896562904, + "vlm_logprob": -0.5796438896562904, + "vlm_yes_no_margin": 2.125000023841858, + "vlm_token_credit": -0.5796438896562904, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031201986130326986 + ], + "opposite_logprob_tokens": [ + -7.378120422363281 + ], + "logprob_reward": -0.0031201986130326986, + "yes_no_margin_reward": 7.375, + "token_credit_reward": -0.0031201986130326986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1263890415430069, + 0.09286561608314514, + 0.11118634045124054, + 0.051409680396318436, + 0.3226909637451172, + 0.041893232613801956, + 0.06974136829376221, + 0.0470777302980423, + 0.08147548139095306, + 0.055270563811063766 + ], + "credit_map": [ + 0.012106213718652725, + 0.020157184451818466, + 0.028336305171251297, + 0.01062006875872612, + 0.011862183921039104, + 0.012650706805288792, + 0.016154592856764793, + 0.014501787722110748, + 0.009657416492700577, + 0.016999663785099983, + 0.011730053462088108, + 0.007212264463305473, + 0.01002624724060297, + 0.011971901170909405, + 0.014837952330708504, + 0.010430127382278442, + 0.005278123077005148, + 0.02134210243821144, + 0.006685967091470957, + 0.006747158709913492, + 0.038131266832351685, + 0.02388325147330761, + 0.005293205380439758, + 0.0038252684753388166, + 0.0061152176931500435, + 0.007557389326393604, + 0.007594741880893707, + 0.0077278269454836845, + 0.0055148242972791195, + 0.007368346210569143, + 0.004962167236953974, + 0.004569167736917734, + 0.03685355931520462, + 0.054099589586257935, + 0.043005552142858505, + 0.0344429537653923, + 0.054632026702165604, + 0.04834805056452751, + 0.027825402095913887, + 0.02348385378718376, + 0.004778625443577766, + 0.006729532033205032, + 0.004147164523601532, + 0.005267478991299868, + 0.005654029082506895, + 0.005825802218168974, + 0.004279120825231075, + 0.005211481358855963, + 0.008332672528922558, + 0.011116307228803635, + 0.00875275582075119, + 0.0072625381872057915, + 0.010755820199847221, + 0.006682874169200659, + 0.008748792111873627, + 0.00808961782604456, + 0.004811897873878479, + 0.007800625171512365, + 0.006703983061015606, + 0.004848030861467123, + 0.005740821827203035, + 0.00486206216737628, + 0.005837063770741224, + 0.00647324975579977, + 0.006632405333220959, + 0.014539196155965328, + 0.007663662079721689, + 0.007744917646050453, + 0.011359942145645618, + 0.0068939123302698135, + 0.01780015416443348, + 0.008841292932629585, + 0.004397088196128607, + 0.005758604500442743, + 0.005217126104980707, + 0.0066301762126386166, + 0.009921052493155003, + 0.007920818403363228, + 0.0050054313614964485, + 0.010420269332826138 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.0031201986130326986 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.2235898971557617 + ], + "opposite_logprob_tokens": [ + -0.3485899567604065 + ], + "logprob_reward": -1.2235898971557617, + "yes_no_margin_reward": -0.8749999403953552, + "token_credit_reward": -1.2235898971557617, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10437240451574326, + 0.06772694736719131, + 0.22126507759094238, + 0.09824240952730179, + 0.35869085788726807, + 0.018321195617318153, + 0.014371771365404129, + 0.01858394965529442, + 0.01850193738937378, + 0.07992351800203323 + ], + "credit_map": [ + 0.009602128528058529, + 0.0161936953663826, + 0.012958843261003494, + 0.009051206521689892, + 0.012887317687273026, + 0.020507391542196274, + 0.01053904090076685, + 0.012632768601179123, + 0.0056444075889885426, + 0.014832921326160431, + 0.00976163987070322, + 0.005712305661290884, + 0.009780538268387318, + 0.008431186899542809, + 0.008689346723258495, + 0.004874594043940306, + 0.013039292767643929, + 0.04051794856786728, + 0.01676393859088421, + 0.011635405011475086, + 0.06908635050058365, + 0.044628292322158813, + 0.014378590509295464, + 0.011215226724743843, + 0.009600609540939331, + 0.010451234877109528, + 0.010087547823786736, + 0.015968812629580498, + 0.009622475132346153, + 0.011937498115003109, + 0.02102283015847206, + 0.00955138634890318, + 0.04436219483613968, + 0.05757974460721016, + 0.04999653995037079, + 0.04128691554069519, + 0.060970209538936615, + 0.04908871650695801, + 0.02879985421895981, + 0.026606645435094833, + 0.0021456656977534294, + 0.0022127244155853987, + 0.0024519343860447407, + 0.002513763727620244, + 0.002362022642046213, + 0.0019584065303206444, + 0.002715734066441655, + 0.0019609415903687477, + 0.0012843835866078734, + 0.0024505381006747484, + 0.0017752728890627623, + 0.0017514127539470792, + 0.0016911171842366457, + 0.001477963523939252, + 0.001842009136453271, + 0.0020990718621760607, + 0.0016455299919471145, + 0.0023077994119375944, + 0.002661439124494791, + 0.0023161189164966345, + 0.002396202879026532, + 0.0018459502607584, + 0.003272078465670347, + 0.0021388274617493153, + 0.001573423738591373, + 0.002327251248061657, + 0.002320569707080722, + 0.00251387944445014, + 0.002415949711576104, + 0.0011394559405744076, + 0.004334890283644199, + 0.0018765139393508434, + 0.009190211072564125, + 0.012323289178311825, + 0.010538099333643913, + 0.010073634795844555, + 0.009565768763422966, + 0.00800306349992752, + 0.00846587773412466, + 0.01176355592906475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -1.2235898971557617 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233768224716187 + ], + "opposite_logprob_tokens": [ + -0.8983768224716187 + ], + "logprob_reward": -0.5233768224716187, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233768224716187, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07899689674377441, + 0.044920872896909714, + 0.0948416069149971, + 0.033674001693725586, + 0.6171753406524658, + 0.027601752430200577, + 0.019290447235107422, + 0.015249130316078663, + 0.026384182274341583, + 0.04186579957604408 + ], + "credit_map": [ + 0.007713342551141977, + 0.014711040072143078, + 0.013935381546616554, + 0.0063921487890183926, + 0.01136802788823843, + 0.010801808908581734, + 0.007312744855880737, + 0.006762400735169649, + 0.0052488683722913265, + 0.006762019824236631, + 0.007921443320810795, + 0.0032532266341149807, + 0.0058530899696052074, + 0.005177083890885115, + 0.006290995515882969, + 0.00441414350643754, + 0.006731237284839153, + 0.015783941373229027, + 0.008820145390927792, + 0.007407172583043575, + 0.022425338625907898, + 0.019608113914728165, + 0.00815780833363533, + 0.005907843820750713, + 0.0028002196922898293, + 0.004137151408940554, + 0.0035210223868489265, + 0.0060250586830079556, + 0.0032852606382220984, + 0.0036095473915338516, + 0.00609919847920537, + 0.004196542780846357, + 0.07184692472219467, + 0.09585078805685043, + 0.08331726491451263, + 0.08745277673006058, + 0.09436590224504471, + 0.08206520974636078, + 0.052350714802742004, + 0.04992573335766792, + 0.002827313030138612, + 0.003240132937207818, + 0.0031053451821208, + 0.0055853682570159435, + 0.0043001011945307255, + 0.001647544326260686, + 0.005071691237390041, + 0.0018242556834593415, + 0.003446299349889159, + 0.002068398054689169, + 0.0017158165574073792, + 0.0022013604175299406, + 0.001916050328873098, + 0.0012877328554168344, + 0.0024035987444221973, + 0.0042511895298957825, + 0.001519187237136066, + 0.0027197988238185644, + 0.0018670354038476944, + 0.0017880998784676194, + 0.001694125821813941, + 0.0016726661706343293, + 0.0016359834698960185, + 0.002352232113480568, + 0.002497945912182331, + 0.004405489191412926, + 0.0034285886213183403, + 0.0030579857993870974, + 0.002977404510602355, + 0.0021663368679583073, + 0.004814155399799347, + 0.003036273643374443, + 0.004230851773172617, + 0.005676309112459421, + 0.00654353154823184, + 0.005354658234864473, + 0.005985952913761139, + 0.0034477447625249624, + 0.004688888322561979, + 0.005937859881669283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.5233768224716187 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011084762401878834 + ], + "opposite_logprob_tokens": [ + -4.51108455657959 + ], + "logprob_reward": -0.011084762401878834, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011084762401878834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11803053319454193, + 0.1904018223285675, + 0.0823301151394844, + 0.18328413367271423, + 0.10865000635385513, + 0.027622036635875702, + 0.10302208364009857, + 0.07372651994228363, + 0.041293635964393616, + 0.07163908332586288 + ], + "credit_map": [ + 0.008736511692404747, + 0.018968552350997925, + 0.020344264805316925, + 0.00878893118351698, + 0.021287184208631516, + 0.015045572072267532, + 0.01249598991125822, + 0.012363524176180363, + 0.0122832627967, + 0.026192430406808853, + 0.02923663705587387, + 0.01609458215534687, + 0.026789873838424683, + 0.022064516320824623, + 0.042158666998147964, + 0.015581846237182617, + 0.005841071717441082, + 0.02121913619339466, + 0.006202561780810356, + 0.005066205281764269, + 0.026050901040434837, + 0.007883676327764988, + 0.005672377068549395, + 0.004394175950437784, + 0.01646397076547146, + 0.023071879521012306, + 0.02258795127272606, + 0.03581972047686577, + 0.017831774428486824, + 0.024769045412540436, + 0.023709500208497047, + 0.019030291587114334, + 0.008512305095791817, + 0.01370995957404375, + 0.015060554258525372, + 0.014796500094234943, + 0.018981292843818665, + 0.020390165969729424, + 0.009411671198904514, + 0.007787557318806648, + 0.0026352412533015013, + 0.004694574978202581, + 0.0031490842811763287, + 0.003337155096232891, + 0.003614627756178379, + 0.003481715451925993, + 0.003802977269515395, + 0.002906657988205552, + 0.01688375324010849, + 0.014005310833454132, + 0.012584700249135494, + 0.013579714111983776, + 0.016443174332380295, + 0.008604303002357483, + 0.011962693184614182, + 0.008958433754742146, + 0.007853932678699493, + 0.012069996446371078, + 0.008547346107661724, + 0.008769218809902668, + 0.010607792064547539, + 0.006362090352922678, + 0.00925995409488678, + 0.010256193578243256, + 0.003142153611406684, + 0.006935284473001957, + 0.004133693408221006, + 0.0040951925329864025, + 0.005203582812100649, + 0.00394385727122426, + 0.009647471830248833, + 0.004192397929728031, + 0.0063122534193098545, + 0.008429711684584618, + 0.011154085397720337, + 0.010204803198575974, + 0.010239369235932827, + 0.008346319198608398, + 0.0078330272808671, + 0.00911951344460249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.011084762401878834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370477676391602 + ], + "opposite_logprob_tokens": [ + -0.38704782724380493 + ], + "logprob_reward": -1.1370477676391602, + "yes_no_margin_reward": -0.7499999403953552, + "token_credit_reward": -1.1370477676391602, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10272902995347977, + 0.0514020137488842, + 0.18873070180416107, + 0.0675857737660408, + 0.20996879041194916, + 0.05285056307911873, + 0.07928623259067535, + 0.048110537230968475, + 0.05788198485970497, + 0.14145438373088837 + ], + "credit_map": [ + 0.00900011695921421, + 0.01796663925051689, + 0.01334060076624155, + 0.008207695558667183, + 0.011866630986332893, + 0.014093248173594475, + 0.01552682276815176, + 0.012727278284728527, + 0.004528489429503679, + 0.008430391550064087, + 0.0089958431199193, + 0.005526281427592039, + 0.005038960836827755, + 0.005587370600551367, + 0.008229875937104225, + 0.0050647989846765995, + 0.013105655089020729, + 0.037201810628175735, + 0.015219993889331818, + 0.00902431458234787, + 0.056491781026124954, + 0.03901337459683418, + 0.009641118347644806, + 0.009032671339809895, + 0.008992007933557034, + 0.0077550578862428665, + 0.007245308719575405, + 0.00992008950561285, + 0.007378978654742241, + 0.010761170648038387, + 0.009285702370107174, + 0.00624746223911643, + 0.022834915667772293, + 0.02658272534608841, + 0.02974928915500641, + 0.03639204427599907, + 0.03200500085949898, + 0.02377571538090706, + 0.02082229033112526, + 0.017806805670261383, + 0.006036953069269657, + 0.007339489180594683, + 0.005195399280637503, + 0.009473762474954128, + 0.00794089026749134, + 0.004226539749652147, + 0.0067714499309659, + 0.005866076331585646, + 0.007446453906595707, + 0.012639329768717289, + 0.01001488696783781, + 0.007513546850532293, + 0.014506774954497814, + 0.008154118433594704, + 0.011221186257898808, + 0.007789944764226675, + 0.005452766083180904, + 0.005874072201550007, + 0.00571299297735095, + 0.004507438745349646, + 0.005909450352191925, + 0.005742181092500687, + 0.007636775262653828, + 0.0072748600505292416, + 0.004565931390970945, + 0.009825879707932472, + 0.00673639914020896, + 0.006621400360018015, + 0.006672265473753214, + 0.005157084204256535, + 0.012921354733407497, + 0.0053816670551896095, + 0.011220094747841358, + 0.01741984859108925, + 0.02464112639427185, + 0.015369965694844723, + 0.022005029022693634, + 0.013621312566101551, + 0.01636645942926407, + 0.020810550078749657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -1.1370477676391602 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "static, extreme macro shot delves into an intimate group of luminous pearls, huddled against a dark, non-reflective surface that absorbs all light. The camera maintains tight, unwavering focus, revealing each pearl's subtle flaws and unique characteristics. Their surfaces are rendered in surreal detail, not just white but an intricate tapestry of iridescent pearls, shimmering with an ethereal glow that seems to emanate from within. A single, sharp key light from a high side angle sculpts their spherical forms, creating brilliant, almost blown-out highlights at their peaks and casting deep, softly-edged shadows where they press against one another. The visual narrative unfolds in a rapid sequence of crisp, high-contrast photographic frames, each staccato cut revealing a subtle shift in light, causing subtle overtones of pink, lavender, and mint green to shimmer and ripple across the pearls' smooth surfaces. This rhythmic, stop-motion-like progression creates a mesmerizing and hypnotic atmosphere, celebrating the objects' sheer tactile beauty with a focused, almost reverent approach.", + "seed": 33142, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_0/index_3943_seed_33142_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.6991540662012994, + "vlm_logprob": -0.6991540662012994, + "vlm_yes_no_margin": 1.425, + "vlm_token_credit": -0.6991540662012994, + "vlm_yes_ratio": 0.6, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00423451466485858 + ], + "opposite_logprob_tokens": [ + -5.754234313964844 + ], + "logprob_reward": -0.00423451466485858, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00423451466485858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42655491828918457, + 0.09076067060232162, + 0.06315277516841888, + 0.046346064656972885, + 0.03995823487639427, + 0.061109721660614014, + 0.05666632205247879, + 0.06882575154304504, + 0.10627312958240509, + 0.040352437645196915 + ], + "credit_map": [ + 0.047980453819036484, + 0.06283632665872574, + 0.056304384022951126, + 0.05238962173461914, + 0.07009478658437729, + 0.03867146000266075, + 0.04536788538098335, + 0.05290998890995979, + 0.008044221438467503, + 0.013668238185346127, + 0.013393767178058624, + 0.009595701470971107, + 0.014304046519100666, + 0.011664562858641148, + 0.01111691165715456, + 0.008973225019872189, + 0.0057555874809622765, + 0.009324489161372185, + 0.007907077670097351, + 0.011143664829432964, + 0.006074333097785711, + 0.0054949684999883175, + 0.009612489491701126, + 0.007840168662369251, + 0.006868432741612196, + 0.006460605189204216, + 0.006815917789936066, + 0.003693142905831337, + 0.005254495423287153, + 0.007582083344459534, + 0.004942431580275297, + 0.004728959873318672, + 0.0036387694999575615, + 0.005975243169814348, + 0.006661882624030113, + 0.005266758147627115, + 0.004459046758711338, + 0.0041098869405686855, + 0.005405896808952093, + 0.004440752323716879, + 0.005782487336546183, + 0.007245199754834175, + 0.009543676860630512, + 0.010999053716659546, + 0.00653991149738431, + 0.008810436353087425, + 0.007322215009480715, + 0.00486674252897501, + 0.006089031230658293, + 0.008329322561621666, + 0.006631332915276289, + 0.004871279001235962, + 0.00739137502387166, + 0.004374320153146982, + 0.01567293517291546, + 0.0033067252952605486, + 0.007237424608319998, + 0.012387022376060486, + 0.00842774473130703, + 0.006731660105288029, + 0.009038195013999939, + 0.0095087680965662, + 0.010269773192703724, + 0.005225165747106075, + 0.007895495742559433, + 0.014858751557767391, + 0.018983790650963783, + 0.01273177471011877, + 0.02332284487783909, + 0.007714947685599327, + 0.01232189405709505, + 0.008443632163107395, + 0.003680472495034337, + 0.004881301429122686, + 0.006055147852748632, + 0.00552195543423295, + 0.00474281283095479, + 0.005064267665147781, + 0.0049439347349107265, + 0.005462546832859516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00423451466485858 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.1270103454589844 + ], + "opposite_logprob_tokens": [ + -0.1270102560520172 + ], + "logprob_reward": -2.1270103454589844, + "yes_no_margin_reward": -2.0, + "token_credit_reward": -2.1270103454589844, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.242886021733284, + 0.0927245244383812, + 0.09634443372488022, + 0.06441085040569305, + 0.06209735572338104, + 0.04113443195819855, + 0.12021248787641525, + 0.08278095722198486, + 0.116784006357193, + 0.0806250274181366 + ], + "credit_map": [ + 0.022572891786694527, + 0.033177971839904785, + 0.023462513461709023, + 0.029517028480768204, + 0.033365365117788315, + 0.024745622649788857, + 0.029064372181892395, + 0.046980250626802444, + 0.007330854423344135, + 0.010687202215194702, + 0.010178886353969574, + 0.00968015193939209, + 0.019713468849658966, + 0.016520533710718155, + 0.011110496707260609, + 0.0075029185973107815, + 0.007073229178786278, + 0.014947705902159214, + 0.010830865241587162, + 0.016984814777970314, + 0.009606604464352131, + 0.012083311565220356, + 0.014754511415958405, + 0.010063388384878635, + 0.011155156418681145, + 0.008924420922994614, + 0.007018785458058119, + 0.006314785219728947, + 0.007175242528319359, + 0.00894907396286726, + 0.0074156797491014, + 0.007457703351974487, + 0.0046125613152980804, + 0.007531802169978619, + 0.010082164779305458, + 0.008885454386472702, + 0.009057492949068546, + 0.007249326910823584, + 0.007872411049902439, + 0.006806135177612305, + 0.005224211607128382, + 0.004053136333823204, + 0.007032286841422319, + 0.007071805186569691, + 0.006737405434250832, + 0.0037964191287755966, + 0.004671633243560791, + 0.002547531621530652, + 0.00875167641788721, + 0.0135719645768404, + 0.009411330334842205, + 0.008014480583369732, + 0.02807949297130108, + 0.009033172391355038, + 0.035936206579208374, + 0.007414149586111307, + 0.00907381996512413, + 0.012496686540544033, + 0.007614876609295607, + 0.007185404654592276, + 0.012266981415450573, + 0.011728925630450249, + 0.0143444137647748, + 0.0080698411911726, + 0.009910338558256626, + 0.013645722530782223, + 0.010360765270888805, + 0.0110782440751791, + 0.019386153668165207, + 0.017791956663131714, + 0.017576752230525017, + 0.017034068703651428, + 0.009824094362556934, + 0.01261238381266594, + 0.013633550144731998, + 0.009289582259953022, + 0.010287299752235413, + 0.007673758547753096, + 0.008454429917037487, + 0.008849918842315674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -2.1270103454589844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6329761743545532 + ], + "opposite_logprob_tokens": [ + -0.7579761743545532 + ], + "logprob_reward": -0.6329761743545532, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6329761743545532, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25355058908462524, + 0.049954451620578766, + 0.11673226207494736, + 0.05600748583674431, + 0.06321088969707489, + 0.04438287019729614, + 0.10452862083911896, + 0.07371162623167038, + 0.12565717101097107, + 0.11226404458284378 + ], + "credit_map": [ + 0.02682957611978054, + 0.03250172361731529, + 0.025489594787359238, + 0.037329383194446564, + 0.04729732871055603, + 0.027538836002349854, + 0.02729799970984459, + 0.02926613576710224, + 0.00353889144025743, + 0.005577019415795803, + 0.008838113397359848, + 0.004205001052469015, + 0.014009884558618069, + 0.004049580544233322, + 0.006612712051719427, + 0.003123248228803277, + 0.0088178850710392, + 0.025556489825248718, + 0.010281543247401714, + 0.021888909861445427, + 0.010022880509495735, + 0.01668793149292469, + 0.014365199953317642, + 0.009111431427299976, + 0.00925367046147585, + 0.006553704850375652, + 0.007664200384169817, + 0.0045993453823029995, + 0.006322733126580715, + 0.008177456445991993, + 0.006918248254805803, + 0.006518127396702766, + 0.0063933138735592365, + 0.009140457026660442, + 0.012625074945390224, + 0.009220167994499207, + 0.0097417077049613, + 0.004889564588665962, + 0.005793809425085783, + 0.00540680019184947, + 0.004745727404952049, + 0.005116191226989031, + 0.005586809478700161, + 0.010137807577848434, + 0.00480559142306447, + 0.005869007203727961, + 0.004864577203989029, + 0.0032571570482105017, + 0.011787359602749348, + 0.01717519201338291, + 0.011478941887617111, + 0.010678277350962162, + 0.012563441880047321, + 0.009786506183445454, + 0.0242730975151062, + 0.006785799283534288, + 0.011222412809729576, + 0.011007198132574558, + 0.007520412560552359, + 0.007540312595665455, + 0.009243149310350418, + 0.009700682945549488, + 0.010765713639557362, + 0.006711749825626612, + 0.01633349061012268, + 0.01714949682354927, + 0.014679204672574997, + 0.014704247005283833, + 0.013422158546745777, + 0.011894939467310905, + 0.022810393944382668, + 0.01466323807835579, + 0.013312065042555332, + 0.014784934930503368, + 0.015244622714817524, + 0.018455402925610542, + 0.0115799680352211, + 0.013835234567523003, + 0.013202287256717682, + 0.011849526315927505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.6329761743545532 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814492002129555 + ], + "opposite_logprob_tokens": [ + -3.288144826889038 + ], + "logprob_reward": -0.03814492002129555, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814492002129555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40074020624160767, + 0.13006168603897095, + 0.0939803496003151, + 0.04197104275226593, + 0.03213312849402428, + 0.04774365574121475, + 0.0633872002363205, + 0.031529758125543594, + 0.09985852241516113, + 0.05859445780515671 + ], + "credit_map": [ + 0.03959465026855469, + 0.059675782918930054, + 0.0497029647231102, + 0.05495770648121834, + 0.06425712257623672, + 0.03748691827058792, + 0.04252384230494499, + 0.05254119634628296, + 0.007828601635992527, + 0.017190199345350266, + 0.017823942005634308, + 0.00995064340531826, + 0.03382636606693268, + 0.01678638346493244, + 0.01830735243856907, + 0.00834820605814457, + 0.00715655367821455, + 0.017822539433836937, + 0.009150329045951366, + 0.015430117957293987, + 0.008353842422366142, + 0.018441643565893173, + 0.011613406240940094, + 0.006011915858834982, + 0.0055960495956242085, + 0.005332717206329107, + 0.004442562349140644, + 0.004426208790391684, + 0.0042944373562932014, + 0.006352949421852827, + 0.005790111143141985, + 0.005736005958169699, + 0.002651154063642025, + 0.004323423840105534, + 0.004223987460136414, + 0.004934345837682486, + 0.004821759648621082, + 0.0037859901785850525, + 0.004258256405591965, + 0.003134208731353283, + 0.004411519039422274, + 0.005559375509619713, + 0.0059063248336315155, + 0.012711730785667896, + 0.007135269697755575, + 0.004212948959320784, + 0.005626447964459658, + 0.0021800354588776827, + 0.005564083810895681, + 0.0072541795670986176, + 0.004712873604148626, + 0.004454527981579304, + 0.016404157504439354, + 0.004437985830008984, + 0.016233740374445915, + 0.004325657617300749, + 0.004256228916347027, + 0.004813920706510544, + 0.0033819505479186773, + 0.002941971644759178, + 0.004128258675336838, + 0.003663125680759549, + 0.005486573092639446, + 0.002857728162780404, + 0.009256636723876, + 0.015434300526976585, + 0.014356871135532856, + 0.013930886052548885, + 0.011064974591135979, + 0.010048340074717999, + 0.010169530287384987, + 0.015596983022987843, + 0.005574654322117567, + 0.008237564004957676, + 0.0080574257299304, + 0.007951393723487854, + 0.005594682414084673, + 0.009260331280529499, + 0.0069250380620360374, + 0.0069933668710291386 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.03814492002129555 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are the main subjects a group of luminous pearls visible in the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background a dark, non-reflective surface that absorbs light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a single sharp key light coming from a high side angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the pearls display shifting overtones of pink, lavender, and mint green?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the visual style a rapid sequence of crisp, high-contrast photographic frames?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6934043765068054 + ], + "opposite_logprob_tokens": [ + -0.6934043765068054 + ], + "logprob_reward": -0.6934043765068054, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6934043765068054, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.24557550251483917, + 0.05683044716715813, + 0.07326860725879669, + 0.11725788563489914, + 0.048530541360378265, + 0.05404617264866829, + 0.04000747948884964, + 0.09357574582099915, + 0.13940154016017914, + 0.1315060704946518 + ], + "credit_map": [ + 0.027287129312753677, + 0.033860135823488235, + 0.025293778628110886, + 0.029792601242661476, + 0.05998886004090309, + 0.021004222333431244, + 0.021122870966792107, + 0.02722591534256935, + 0.006079599726945162, + 0.00865185633301735, + 0.007370821200311184, + 0.007791437674313784, + 0.006571957841515541, + 0.007046302780508995, + 0.00852448120713234, + 0.004793989006429911, + 0.005357989110052586, + 0.008913890458643436, + 0.006842165719717741, + 0.014529571868479252, + 0.00620646309107542, + 0.009970191866159439, + 0.012323837727308273, + 0.00912450347095728, + 0.02135550230741501, + 0.013172206468880177, + 0.018782515078783035, + 0.00814578216522932, + 0.01433236338198185, + 0.01701752468943596, + 0.01058070920407772, + 0.013871275819838047, + 0.004747831728309393, + 0.006004003807902336, + 0.007999826222658157, + 0.006713227368891239, + 0.008250446058809757, + 0.00557217001914978, + 0.004803979769349098, + 0.004439056385308504, + 0.007972330786287785, + 0.007136869709938765, + 0.006507785525172949, + 0.0113819045946002, + 0.007754167541861534, + 0.004911832045763731, + 0.005287169944494963, + 0.0030941120348870754, + 0.0046590412966907024, + 0.00602900143712759, + 0.003265939187258482, + 0.005152998957782984, + 0.006111641880124807, + 0.0055425819009542465, + 0.00495749618858099, + 0.004288780502974987, + 0.010430528782308102, + 0.015448974445462227, + 0.010626380331814289, + 0.008740559220314026, + 0.01681688241660595, + 0.009834113530814648, + 0.015301867388188839, + 0.006376441568136215, + 0.012367455288767815, + 0.01537454966455698, + 0.01714024506509304, + 0.014214645139873028, + 0.019249897450208664, + 0.017257824540138245, + 0.02185015007853508, + 0.021946769207715988, + 0.014396882615983486, + 0.018410904332995415, + 0.017864849418401718, + 0.02029380574822426, + 0.016446538269519806, + 0.013870155438780785, + 0.015281250700354576, + 0.014941678382456303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.6934043765068054 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_tensor([2247, 2247])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_tensor([2247, 2247])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0e64a8d0d11466223dd54ad04c7ff40033b0707b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_tensor([2247, 2247])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a powerful, live-action shot, the camera pans at breakneck speed in a tight 360-degree pan. At the center of the frame is a muscular, classically sculpted king of the sea. Bare-chested, his bronzed skin gleams under the spotlights overhead; his long, wet, dark hair clings to his broad shoulders and back. He's at the apex of a movement, his arms flexing with muscle as he thrusts a trident, radiating a brilliant cyan-blue light, into a massive block of black basalt in the center of the arena. The moment the trident's tip impacts the rock, a dazzling shower of sparks and rubble erupts, while a wave of energy whips the shallow puddle beneath his feet into a spectacular, outward-spreading white spray. As the camera rotates at high speed, we can see water droplets condensing into crystal beads in the air, reflecting the light of the trident, and his resolute face, his eyes are focused and fierce, as if declaring the victory of an epic duel. The whole scene is filled with primitive, magnificent and highly impactful epic atmosphere under the cold light and shadow.", + "seed": 4959, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4959_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29321686904877425, + "vlm_logprob": -0.29321686904877425, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.29321686904877425, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02014545165002346 + ], + "opposite_logprob_tokens": [ + -4.020145416259766 + ], + "logprob_reward": -0.02014545165002346, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02014545165002346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09501699358224869, + 0.032924484461545944, + 0.0641099289059639, + 0.05842523276805878, + 0.09248937666416168, + 0.058006756007671356, + 0.12355013936758041, + 0.0826852023601532, + 0.3121716380119324, + 0.08062022924423218 + ], + "credit_map": [ + 0.012244011275470257, + 0.012801853008568287, + 0.009268454276025295, + 0.01556929387152195, + 0.011223862878978252, + 0.011688729748129845, + 0.011405318044126034, + 0.010815469548106194, + 0.003558242227882147, + 0.0050339410081505775, + 0.003235205542296171, + 0.0030764795374125242, + 0.005996455904096365, + 0.0047205002047121525, + 0.0038080746307969093, + 0.0034955847077071667, + 0.007979998365044594, + 0.00868772342801094, + 0.006857564672827721, + 0.011200707405805588, + 0.008108142763376236, + 0.006833267863839865, + 0.008494020439684391, + 0.0059485044330358505, + 0.005137584172189236, + 0.0064789485186338425, + 0.008424035273492336, + 0.011259233579039574, + 0.008745343424379826, + 0.008318708278238773, + 0.005437762476503849, + 0.004623617976903915, + 0.011865483596920967, + 0.008574816398322582, + 0.023332156240940094, + 0.00855009350925684, + 0.0062130168080329895, + 0.016016028821468353, + 0.005612119100987911, + 0.012325667776167393, + 0.004865066614001989, + 0.010569359175860882, + 0.005170049145817757, + 0.007834376767277718, + 0.007861051708459854, + 0.005871085915714502, + 0.0061072236858308315, + 0.009728544391691685, + 0.013904248364269733, + 0.017275618389248848, + 0.009411385282874107, + 0.020183412358164787, + 0.01688789203763008, + 0.013950977474451065, + 0.018079683184623718, + 0.013856922276318073, + 0.012902259826660156, + 0.011758954264223576, + 0.009467247873544693, + 0.008286632597446442, + 0.011753801256418228, + 0.01002461276948452, + 0.009594305418431759, + 0.008897388353943825, + 0.04126439988613129, + 0.04671059548854828, + 0.03808579966425896, + 0.02293679304420948, + 0.04988425597548485, + 0.06709127873182297, + 0.022988062351942062, + 0.023210473358631134, + 0.008652926422655582, + 0.01198226772248745, + 0.006850047502666712, + 0.009285824373364449, + 0.015396364964544773, + 0.010902401991188526, + 0.00815791916102171, + 0.009392480365931988 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.02014545165002346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794392623007298 + ], + "opposite_logprob_tokens": [ + -4.634794235229492 + ], + "logprob_reward": -0.009794392623007298, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794392623007298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1299981325864792, + 0.048651594668626785, + 0.118108369410038, + 0.20579496026039124, + 0.12392272800207138, + 0.0813598483800888, + 0.07690547406673431, + 0.03294463828206062, + 0.07280639559030533, + 0.10950784385204315 + ], + "credit_map": [ + 0.015762390568852425, + 0.018530214205384254, + 0.009247750043869019, + 0.025507107377052307, + 0.017884202301502228, + 0.015105324797332287, + 0.01448383741080761, + 0.013477322645485401, + 0.004942628089338541, + 0.007289647124707699, + 0.006358424201607704, + 0.005644547753036022, + 0.008404900319874287, + 0.005324136931449175, + 0.005304776597768068, + 0.005382536444813013, + 0.012175522744655609, + 0.009781214408576488, + 0.012612156569957733, + 0.009340612217783928, + 0.019576625898480415, + 0.01769394613802433, + 0.018547631800174713, + 0.01838066056370735, + 0.017293620854616165, + 0.02321046218276024, + 0.025635870173573494, + 0.047182511538267136, + 0.024285417050123215, + 0.026806890964508057, + 0.022436263039708138, + 0.018943913280963898, + 0.017945095896720886, + 0.015551760792732239, + 0.026581216603517532, + 0.014822443947196007, + 0.00977591797709465, + 0.01751963049173355, + 0.00875153113156557, + 0.012975123710930347, + 0.0070174322463572025, + 0.017749859020113945, + 0.009083469398319721, + 0.0087975412607193, + 0.01367419958114624, + 0.009025480598211288, + 0.00743842450901866, + 0.00857344176620245, + 0.007025442086160183, + 0.013586864806711674, + 0.006590208970010281, + 0.013333884999155998, + 0.009466076269745827, + 0.009736687876284122, + 0.011178757064044476, + 0.005987555719912052, + 0.004299674183130264, + 0.005625524092465639, + 0.003254014765843749, + 0.003037768881767988, + 0.005739418789744377, + 0.003063421929255128, + 0.004586342256516218, + 0.0033384712878614664, + 0.012696667574346066, + 0.0128669748082757, + 0.0076829819008708, + 0.006708499509841204, + 0.0114052202552557, + 0.008514480665326118, + 0.007111544720828533, + 0.005820030812174082, + 0.008905243128538132, + 0.021485377103090286, + 0.007992461323738098, + 0.013384446501731873, + 0.019356319680809975, + 0.016324739903211594, + 0.01585409976541996, + 0.006205152720212936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.009794392623007298 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033697985112667084 + ], + "opposite_logprob_tokens": [ + -3.408698081970215 + ], + "logprob_reward": -0.033697985112667084, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033697985112667084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22820095717906952, + 0.09748117625713348, + 0.09051274508237839, + 0.1240866631269455, + 0.10625676810741425, + 0.06881336867809296, + 0.08833727985620499, + 0.03996284678578377, + 0.08382172137498856, + 0.07252644747495651 + ], + "credit_map": [ + 0.02495669201016426, + 0.02243746630847454, + 0.03533013164997101, + 0.01802937686443329, + 0.019391445443034172, + 0.02624817192554474, + 0.04645916447043419, + 0.03534851223230362, + 0.009228313341736794, + 0.015217069536447525, + 0.01282932423055172, + 0.010312799364328384, + 0.015239132568240166, + 0.009276910685002804, + 0.012797345407307148, + 0.01258029229938984, + 0.009833923541009426, + 0.008030449040234089, + 0.009370118379592896, + 0.008963228203356266, + 0.016382696107029915, + 0.011980624869465828, + 0.012502777390182018, + 0.013448935002088547, + 0.010528656654059887, + 0.015629496425390244, + 0.016089146956801414, + 0.02388322912156582, + 0.0182590764015913, + 0.015019991435110569, + 0.01368117704987526, + 0.010995890945196152, + 0.01647394709289074, + 0.010319915600121021, + 0.0212349072098732, + 0.010289783589541912, + 0.008757532574236393, + 0.014278126880526543, + 0.005819523241370916, + 0.019083039835095406, + 0.006304590497165918, + 0.012330221943557262, + 0.006615129299461842, + 0.006000223103910685, + 0.011086305603384972, + 0.008307404816150665, + 0.007461049128323793, + 0.01070844754576683, + 0.009844991378486156, + 0.015580456703901291, + 0.0073214066214859486, + 0.009435052052140236, + 0.012461019679903984, + 0.01169334165751934, + 0.01204712875187397, + 0.009953883476555347, + 0.0059535205364227295, + 0.0072446661069989204, + 0.0029113609343767166, + 0.003968758042901754, + 0.0060454742051661015, + 0.005181867629289627, + 0.0057722111232578754, + 0.0028849905356764793, + 0.013530164957046509, + 0.01289973221719265, + 0.0111498748883605, + 0.00750342570245266, + 0.015683691948652267, + 0.009962433949112892, + 0.0062051499262452126, + 0.006887256633490324, + 0.009421342983841896, + 0.01103643886744976, + 0.008837349712848663, + 0.009696750901639462, + 0.01025380752980709, + 0.009806192480027676, + 0.006499298848211765, + 0.006975269876420498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.033697985112667084 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134069442749023 + ], + "opposite_logprob_tokens": [ + -0.31340691447257996 + ], + "logprob_reward": -1.3134069442749023, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134069442749023, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11313384026288986, + 0.05809343606233597, + 0.10939733684062958, + 0.277634859085083, + 0.04880933463573456, + 0.13837434351444244, + 0.11707013100385666, + 0.036385126411914825, + 0.040025293827056885, + 0.061076316982507706 + ], + "credit_map": [ + 0.013902420178055763, + 0.01901034265756607, + 0.012123167514801025, + 0.014909287914633751, + 0.01361575536429882, + 0.012989716604351997, + 0.013649864122271538, + 0.0129332784563303, + 0.004849272780120373, + 0.008012966252863407, + 0.006656190846115351, + 0.007046772167086601, + 0.009105376899242401, + 0.00768074719235301, + 0.006937697995454073, + 0.007804409135133028, + 0.012787319719791412, + 0.01451173797249794, + 0.010969887487590313, + 0.024915844202041626, + 0.010705978609621525, + 0.010830359533429146, + 0.013648787513375282, + 0.01102741900831461, + 0.024550041183829308, + 0.04027481749653816, + 0.040059644728899, + 0.06929898262023926, + 0.029051190242171288, + 0.03854680061340332, + 0.015957074239850044, + 0.019896283745765686, + 0.007084008771926165, + 0.007017557974904776, + 0.009181715548038483, + 0.004962458740919828, + 0.005605787504464388, + 0.005965204443782568, + 0.0041050901636481285, + 0.0048875049687922, + 0.010341149754822254, + 0.02970830164849758, + 0.014657585881650448, + 0.017054857686161995, + 0.01866302639245987, + 0.015421335585415363, + 0.014805778861045837, + 0.01772230863571167, + 0.013680233620107174, + 0.015876704826951027, + 0.009039283730089664, + 0.020572658628225327, + 0.011905036866664886, + 0.012964760884642601, + 0.021754195913672447, + 0.011277252808213234, + 0.0047046951949596405, + 0.004326223861426115, + 0.003857780946418643, + 0.003558521391823888, + 0.006136434152722359, + 0.005528971552848816, + 0.004470417741686106, + 0.0038020831998437643, + 0.004989348351955414, + 0.006344597786664963, + 0.004673717077821493, + 0.004477326292544603, + 0.006538225803524256, + 0.005983514245599508, + 0.0033043145667761564, + 0.003714244347065687, + 0.00926788616925478, + 0.009240835905075073, + 0.005754971411079168, + 0.008580800145864487, + 0.007567274384200573, + 0.008051728829741478, + 0.005505772773176432, + 0.007107045967131853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -1.3134069442749023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903957158327103 + ], + "opposite_logprob_tokens": [ + -2.4640395641326904 + ], + "logprob_reward": -0.08903957158327103, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903957158327103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2236417680978775, + 0.05337011441588402, + 0.1819378286600113, + 0.10368848592042923, + 0.05459319055080414, + 0.11270543187856674, + 0.10418827831745148, + 0.03874078392982483, + 0.041231513023376465, + 0.0859026163816452 + ], + "credit_map": [ + 0.0390060730278492, + 0.029917635023593903, + 0.01882144622504711, + 0.02725851535797119, + 0.03717081621289253, + 0.026027092710137367, + 0.024675264954566956, + 0.020764939486980438, + 0.005494860466569662, + 0.00844066496938467, + 0.005383581388741732, + 0.006875923834741116, + 0.006849254947155714, + 0.008702735416591167, + 0.004304240923374891, + 0.00731885340064764, + 0.016345594078302383, + 0.01626564748585224, + 0.01581636071205139, + 0.020305823534727097, + 0.031966082751750946, + 0.025716107338666916, + 0.024587515741586685, + 0.030934683978557587, + 0.009736387990415096, + 0.01344031747430563, + 0.011247873306274414, + 0.02720336802303791, + 0.012852364219725132, + 0.012813732959330082, + 0.008738378994166851, + 0.007656058296561241, + 0.00814084243029356, + 0.004967991728335619, + 0.00837766658514738, + 0.004534484352916479, + 0.007529871538281441, + 0.0073769595474004745, + 0.003968960605561733, + 0.009696414694190025, + 0.008684656582772732, + 0.01851525716483593, + 0.010812508873641491, + 0.014407671056687832, + 0.018872031942009926, + 0.012141269631683826, + 0.014311444945633411, + 0.014960593543946743, + 0.010130197741091251, + 0.013725368306040764, + 0.010926205664873123, + 0.018546191975474358, + 0.011586535722017288, + 0.011762813664972782, + 0.017729543149471283, + 0.009781417436897755, + 0.006675343960523605, + 0.007429929915815592, + 0.0029918618965893984, + 0.003812919370830059, + 0.006441722624003887, + 0.003236324293538928, + 0.004913808777928352, + 0.0032388740219175816, + 0.0062240781262516975, + 0.005543578416109085, + 0.006143407896161079, + 0.00400267168879509, + 0.0057113622315227985, + 0.0061223614029586315, + 0.003939952701330185, + 0.0035441056825220585, + 0.010453836061060429, + 0.010015655308961868, + 0.015896087512373924, + 0.010928948409855366, + 0.009005198255181313, + 0.00965594407171011, + 0.007181948982179165, + 0.012764997780323029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.08903957158327103 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a powerful, live-action shot, the camera pans at breakneck speed in a tight 360-degree pan. At the center of the frame is a muscular, classically sculpted king of the sea. Bare-chested, his bronzed skin gleams under the spotlights overhead; his long, wet, dark hair clings to his broad shoulders and back. He's at the apex of a movement, his arms flexing with muscle as he thrusts a trident, radiating a brilliant cyan-blue light, into a massive block of black basalt in the center of the arena. The moment the trident's tip impacts the rock, a dazzling shower of sparks and rubble erupts, while a wave of energy whips the shallow puddle beneath his feet into a spectacular, outward-spreading white spray. As the camera rotates at high speed, we can see water droplets condensing into crystal beads in the air, reflecting the light of the trident, and his resolute face, his eyes are focused and fierce, as if declaring the victory of an epic duel. The whole scene is filled with primitive, magnificent and highly impactful epic atmosphere under the cold light and shadow.", + "seed": 4960, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4960_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.38668858632445335, + "vlm_logprob": -0.38668858632445335, + "vlm_yes_no_margin": 2.5, + "vlm_token_credit": -0.38668858632445335, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03590403124690056 + ], + "opposite_logprob_tokens": [ + -3.4109039306640625 + ], + "logprob_reward": -0.03590403124690056, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03590403124690056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2204260230064392, + 0.09677071869373322, + 0.056002434343099594, + 0.22830012440681458, + 0.03769716992974281, + 0.027230428531765938, + 0.11498727649450302, + 0.06634163856506348, + 0.06211528927087784, + 0.09012892097234726 + ], + "credit_map": [ + 0.05035156384110451, + 0.014528238214552402, + 0.03450582176446915, + 0.01586749218404293, + 0.014337034896016121, + 0.0635475367307663, + 0.01579783856868744, + 0.011490504257380962, + 0.013614209368824959, + 0.017301321029663086, + 0.006479970645159483, + 0.01147484965622425, + 0.011705429293215275, + 0.01149973925203085, + 0.013601535931229591, + 0.011093664914369583, + 0.006987066939473152, + 0.0061940220184624195, + 0.005584580823779106, + 0.0046845898032188416, + 0.004986045882105827, + 0.014895536936819553, + 0.00622037285938859, + 0.006450216751545668, + 0.048420872539281845, + 0.02670777030289173, + 0.012596688233315945, + 0.02896580472588539, + 0.016678249463438988, + 0.06315038353204727, + 0.01693965308368206, + 0.014840694144368172, + 0.006117942277342081, + 0.004093702882528305, + 0.0027366310823708773, + 0.00443705590441823, + 0.005132253281772137, + 0.007519010454416275, + 0.003817466087639332, + 0.003843106096610427, + 0.004042347427457571, + 0.00427428400143981, + 0.002800900721922517, + 0.0038711673114448786, + 0.0027203878853470087, + 0.0035000594798475504, + 0.0033007063902914524, + 0.0027205762453377247, + 0.0074722603894770145, + 0.012974916025996208, + 0.012433568947017193, + 0.014152836985886097, + 0.021578844636678696, + 0.015011841431260109, + 0.021065741777420044, + 0.010297260247170925, + 0.010660079307854176, + 0.007222872693091631, + 0.008046271279454231, + 0.006100026424974203, + 0.009110122919082642, + 0.008553958497941494, + 0.011202706955373287, + 0.0054456009529531, + 0.007578691467642784, + 0.010484130121767521, + 0.009142527356743813, + 0.003889603540301323, + 0.009120733477175236, + 0.011160917580127716, + 0.004598537925630808, + 0.006140145473182201, + 0.013245250098407269, + 0.014019896276295185, + 0.012394900433719158, + 0.009084081277251244, + 0.011175595223903656, + 0.009528721682727337, + 0.010181725025177002, + 0.010498753748834133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.03590403124690056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214927330613136 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014214927330613136, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214927330613136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49288421869277954, + 0.08900725841522217, + 0.0554467998445034, + 0.07181023806333542, + 0.04701804369688034, + 0.034686096012592316, + 0.058963317424058914, + 0.03327495604753494, + 0.06520848721265793, + 0.05170057341456413 + ], + "credit_map": [ + 0.11284095793962479, + 0.04085070639848709, + 0.09009921550750732, + 0.021405933424830437, + 0.03505883365869522, + 0.1533072143793106, + 0.01900184527039528, + 0.020319519564509392, + 0.015012838877737522, + 0.012665958143770695, + 0.005728038493543863, + 0.010454890318214893, + 0.010971978306770325, + 0.011690127663314342, + 0.00963518675416708, + 0.012848244048655033, + 0.0075702848844230175, + 0.007347476202994585, + 0.0044401357881724834, + 0.006749621592462063, + 0.004783271346241236, + 0.010938715189695358, + 0.008312906138598919, + 0.005304391495883465, + 0.01144501380622387, + 0.011127383448183537, + 0.007859952747821808, + 0.010423851199448109, + 0.006875501945614815, + 0.013266152702271938, + 0.006049880757927895, + 0.004762505646795034, + 0.009575331583619118, + 0.004106026142835617, + 0.0034628501161932945, + 0.004750661086291075, + 0.006264363415539265, + 0.007007434032857418, + 0.004451033193618059, + 0.007400347851216793, + 0.003513782052323222, + 0.003782985033467412, + 0.004289662465453148, + 0.007541014812886715, + 0.0033892053179442883, + 0.003852806519716978, + 0.006181764416396618, + 0.0021348807495087385, + 0.00559203140437603, + 0.008444077335298061, + 0.006835269741714001, + 0.007012004964053631, + 0.011536137200891972, + 0.0067160846665501595, + 0.007928571663796902, + 0.00489914370700717, + 0.003362514777109027, + 0.0036836506333202124, + 0.0028845679480582476, + 0.003030637977644801, + 0.008321906439960003, + 0.004287284333258867, + 0.005270057823508978, + 0.0024343361146748066, + 0.00934786256402731, + 0.007941258139908314, + 0.0062830024398863316, + 0.004083144012838602, + 0.007950756698846817, + 0.017422981560230255, + 0.00566739309579134, + 0.006512098014354706, + 0.006774458568543196, + 0.00843295268714428, + 0.007513065822422504, + 0.005076826084405184, + 0.008364335633814335, + 0.005233680829405785, + 0.004890883341431618, + 0.0054143802262842655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014214927330613136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980063669383526 + ], + "opposite_logprob_tokens": [ + -3.5298006534576416 + ], + "logprob_reward": -0.02980063669383526, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980063669383526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12227894365787506, + 0.22339710593223572, + 0.2149074226617813, + 0.05606456845998764, + 0.20788732171058655, + 0.01987481489777565, + 0.02416904643177986, + 0.030770592391490936, + 0.042655378580093384, + 0.0579947866499424 + ], + "credit_map": [ + 0.02467961236834526, + 0.014637229032814503, + 0.009348499588668346, + 0.010068581439554691, + 0.017906595021486282, + 0.025334598496556282, + 0.011257757432758808, + 0.009046069346368313, + 0.03547362610697746, + 0.037661824375391006, + 0.016521230340003967, + 0.027958907186985016, + 0.02405862882733345, + 0.028205621987581253, + 0.030810263007879257, + 0.022707009688019753, + 0.029546793550252914, + 0.029577048495411873, + 0.029341258108615875, + 0.021594548597931862, + 0.013919815421104431, + 0.049609504640102386, + 0.01568133942782879, + 0.02563711628317833, + 0.011368916369974613, + 0.00848578941076994, + 0.0033411490730941296, + 0.006397998426109552, + 0.00499594584107399, + 0.011875580996274948, + 0.004213530570268631, + 0.005385659635066986, + 0.038108646869659424, + 0.018772782757878304, + 0.014616620726883411, + 0.019905153661966324, + 0.033043403178453445, + 0.0352499820291996, + 0.020474748685956, + 0.02771599031984806, + 0.0023676254786551, + 0.002433003159239888, + 0.0030140704475343227, + 0.00311900838278234, + 0.0026751048862934113, + 0.002224700292572379, + 0.0021125846542418003, + 0.0019287178292870522, + 0.0022574712056666613, + 0.0032042209059000015, + 0.0027699973434209824, + 0.0025283112190663815, + 0.004047033376991749, + 0.003438851796090603, + 0.0034203394316136837, + 0.002502821385860443, + 0.0031468605156987906, + 0.003139603417366743, + 0.003501148894429207, + 0.0023319730535149574, + 0.00672942353412509, + 0.004858957603573799, + 0.004217016510665417, + 0.002845608629286289, + 0.011114166118204594, + 0.0063073765486478806, + 0.0034326983150094748, + 0.002194367814809084, + 0.005789430346339941, + 0.007508124224841595, + 0.0032111676409840584, + 0.0030980503652244806, + 0.0075714001432061195, + 0.009214901365339756, + 0.01002434641122818, + 0.005515221040695906, + 0.007541632279753685, + 0.006188075989484787, + 0.005364143289625645, + 0.0065750665962696075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.02980063669383526 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.8048490285873413 + ], + "opposite_logprob_tokens": [ + -0.1798490285873413 + ], + "logprob_reward": -1.8048490285873413, + "yes_no_margin_reward": -1.625, + "token_credit_reward": -1.8048490285873413, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2714252769947052, + 0.08103498071432114, + 0.06139172613620758, + 0.23175135254859924, + 0.06046309322118759, + 0.03826018050312996, + 0.035925883799791336, + 0.1496271938085556, + 0.036667678505182266, + 0.03345266729593277 + ], + "credit_map": [ + 0.06126416474580765, + 0.01992899551987648, + 0.05685989931225777, + 0.01340364571660757, + 0.013418781571090221, + 0.08611615747213364, + 0.009896378964185715, + 0.010537235997617245, + 0.014310992322862148, + 0.009794538840651512, + 0.006942473817616701, + 0.009639261290431023, + 0.01022682711482048, + 0.01100176852196455, + 0.010379275307059288, + 0.00873984582722187, + 0.01086939312517643, + 0.006701059639453888, + 0.005944882519543171, + 0.00793483853340149, + 0.005864690989255905, + 0.01209037285298109, + 0.007174558937549591, + 0.004811930004507303, + 0.06780795753002167, + 0.02094871923327446, + 0.011684623546898365, + 0.014677656814455986, + 0.013576164841651917, + 0.08039937168359756, + 0.010545537807047367, + 0.012111340649425983, + 0.01265104953199625, + 0.006710011046379805, + 0.004327184520661831, + 0.004890588112175465, + 0.006557703483849764, + 0.010275342501699924, + 0.0063070566393435, + 0.00874415785074234, + 0.004735203925520182, + 0.0054036471992731094, + 0.003024836303666234, + 0.007017396856099367, + 0.0034020820166915655, + 0.004693376831710339, + 0.006784205790609121, + 0.0031994280871003866, + 0.003667716169729829, + 0.005489802453666925, + 0.0036190946120768785, + 0.003654246451333165, + 0.005887266248464584, + 0.005760099273175001, + 0.004859038162976503, + 0.002988621359691024, + 0.020945316180586815, + 0.015835048630833626, + 0.015268990769982338, + 0.014961679466068745, + 0.03024633415043354, + 0.02280774526298046, + 0.020016031339764595, + 0.009546048939228058, + 0.0034603625535964966, + 0.00468075554817915, + 0.004235456697642803, + 0.003916212823241949, + 0.005087343975901604, + 0.005747634917497635, + 0.004675978794693947, + 0.0048639364540576935, + 0.0050219022668898106, + 0.006139043718576431, + 0.004873265046626329, + 0.003166933311149478, + 0.004641348496079445, + 0.0032444060780107975, + 0.003196531441062689, + 0.0031692381016910076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -1.8048490285873413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867430776357651 + ], + "opposite_logprob_tokens": [ + -3.0486743450164795 + ], + "logprob_reward": -0.04867430776357651, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867430776357651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.165270134806633, + 0.05033333972096443, + 0.041427724063396454, + 0.0962597131729126, + 0.1354209929704666, + 0.12979534268379211, + 0.04719135910272598, + 0.09640979021787643, + 0.16234448552131653, + 0.07554706931114197 + ], + "credit_map": [ + 0.032579291611909866, + 0.014515144750475883, + 0.02156667597591877, + 0.01730869896709919, + 0.009740947745740414, + 0.034537531435489655, + 0.01777113601565361, + 0.017250707373023033, + 0.008212961256504059, + 0.006992252543568611, + 0.0036022053100168705, + 0.005455325357615948, + 0.00542073417454958, + 0.006164648570120335, + 0.0072088055312633514, + 0.007276405114680529, + 0.005627933423966169, + 0.007067010272294283, + 0.0032942399848252535, + 0.006369760259985924, + 0.0034272661432623863, + 0.005710614379495382, + 0.004889643285423517, + 0.005041253753006458, + 0.02495463192462921, + 0.008154611103236675, + 0.008624890819191933, + 0.006266679614782333, + 0.0053525385446846485, + 0.033510662615299225, + 0.004311512690037489, + 0.005084188189357519, + 0.02224310301244259, + 0.012056691572070122, + 0.012580631300807, + 0.011099855415523052, + 0.01811419054865837, + 0.033809810876846313, + 0.012705960310995579, + 0.012810752727091312, + 0.014671633951365948, + 0.017613714560866356, + 0.01725395955145359, + 0.020041076466441154, + 0.01609751209616661, + 0.014271480031311512, + 0.01739850826561451, + 0.012447468005120754, + 0.005005029030144215, + 0.005799105856567621, + 0.00762241380289197, + 0.005548772867769003, + 0.008266042917966843, + 0.0069357589818537235, + 0.0046329074539244175, + 0.003381328424438834, + 0.01566988416016102, + 0.010940734297037125, + 0.008434665389358997, + 0.00880978349596262, + 0.01809719391167164, + 0.016695862635970116, + 0.01274875458329916, + 0.005012921057641506, + 0.044516660273075104, + 0.029633022844791412, + 0.016855327412486076, + 0.007408013101667166, + 0.012685410678386688, + 0.029963050037622452, + 0.008318914100527763, + 0.012964095920324326, + 0.009288324043154716, + 0.010343503206968307, + 0.006542035844177008, + 0.01106302160769701, + 0.011303446255624294, + 0.008777606301009655, + 0.008214153349399567, + 0.010014981031417847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.04867430776357651 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a powerful, live-action shot, the camera pans at breakneck speed in a tight 360-degree pan. At the center of the frame is a muscular, classically sculpted king of the sea. Bare-chested, his bronzed skin gleams under the spotlights overhead; his long, wet, dark hair clings to his broad shoulders and back. He's at the apex of a movement, his arms flexing with muscle as he thrusts a trident, radiating a brilliant cyan-blue light, into a massive block of black basalt in the center of the arena. The moment the trident's tip impacts the rock, a dazzling shower of sparks and rubble erupts, while a wave of energy whips the shallow puddle beneath his feet into a spectacular, outward-spreading white spray. As the camera rotates at high speed, we can see water droplets condensing into crystal beads in the air, reflecting the light of the trident, and his resolute face, his eyes are focused and fierce, as if declaring the victory of an epic duel. The whole scene is filled with primitive, magnificent and highly impactful epic atmosphere under the cold light and shadow.", + "seed": 4961, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4961_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.3835961684584618, + "vlm_logprob": -0.3835961684584618, + "vlm_yes_no_margin": 2.15, + "vlm_token_credit": -0.3835961684584618, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019909657537937164 + ], + "opposite_logprob_tokens": [ + -4.019909858703613 + ], + "logprob_reward": -0.019909657537937164, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.019909657537937164, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09635709971189499, + 0.07455254346132278, + 0.08290310204029083, + 0.16680528223514557, + 0.057122088968753815, + 0.16910871863365173, + 0.05341681092977524, + 0.1019609272480011, + 0.11425703018903732, + 0.08351635932922363 + ], + "credit_map": [ + 0.010928181931376457, + 0.011126349680125713, + 0.013174237683415413, + 0.01664232648909092, + 0.0081255491822958, + 0.008774652145802975, + 0.012036671862006187, + 0.01554912980645895, + 0.008215930312871933, + 0.007427397649735212, + 0.009526213631033897, + 0.011896304786205292, + 0.009780229069292545, + 0.008205297403037548, + 0.010557514615356922, + 0.008943663910031319, + 0.01461600512266159, + 0.007730519864708185, + 0.010443764738738537, + 0.009198064915835857, + 0.01085208635777235, + 0.01106738019734621, + 0.011308957822620869, + 0.00768632348626852, + 0.016171682626008987, + 0.02829168550670147, + 0.017562981694936752, + 0.015698488801717758, + 0.042103130370378494, + 0.019508546218276024, + 0.01667456328868866, + 0.010794197209179401, + 0.007008939981460571, + 0.007840525358915329, + 0.008855168707668781, + 0.005951641593128443, + 0.0049074562266469, + 0.007573329843580723, + 0.007943890988826752, + 0.0070411404594779015, + 0.04520513117313385, + 0.01820203848183155, + 0.0167741347104311, + 0.02376573719084263, + 0.018709659576416016, + 0.01405711006373167, + 0.01825925149023533, + 0.014135659672319889, + 0.006132671143859625, + 0.010170945897698402, + 0.005268204025924206, + 0.006090257782489061, + 0.009317036718130112, + 0.005614034365862608, + 0.005797798279672861, + 0.00502586318179965, + 0.011581662110984325, + 0.014795929193496704, + 0.011988277547061443, + 0.011984367854893208, + 0.014492612332105637, + 0.011591953225433826, + 0.013534888625144958, + 0.011991235427558422, + 0.00989660993218422, + 0.01540665328502655, + 0.009055578149855137, + 0.011558296158909798, + 0.020288757979869843, + 0.01091969758272171, + 0.02115258388221264, + 0.01597885601222515, + 0.006515153218060732, + 0.009325917810201645, + 0.007223730906844139, + 0.007952289655804634, + 0.010214035399258137, + 0.018347620964050293, + 0.010271873325109482, + 0.01366573665291071 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.019909657537937164 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304030299186707 + ], + "opposite_logprob_tokens": [ + -3.773303985595703 + ], + "logprob_reward": -0.023304030299186707, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304030299186707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09275981038808823, + 0.07299408316612244, + 0.11414281278848648, + 0.06717772781848907, + 0.07097841799259186, + 0.1719920039176941, + 0.1312401443719864, + 0.10468348115682602, + 0.12069883197546005, + 0.05333273857831955 + ], + "credit_map": [ + 0.010199896059930325, + 0.011523928493261337, + 0.008878412656486034, + 0.01326633058488369, + 0.017776519060134888, + 0.011456090956926346, + 0.009781024418771267, + 0.009877600707113743, + 0.0116584999486804, + 0.009582743979990482, + 0.008253691717982292, + 0.010031333193182945, + 0.009519980289041996, + 0.006813339423388243, + 0.00939213577657938, + 0.007742349989712238, + 0.018224474042654037, + 0.01266616489738226, + 0.01200019009411335, + 0.020002106204628944, + 0.011642890982329845, + 0.01456957496702671, + 0.013705904595553875, + 0.011331499554216862, + 0.007710691075772047, + 0.00892872828990221, + 0.006739150267094374, + 0.009134446270763874, + 0.013249198906123638, + 0.007827860303223133, + 0.008108876645565033, + 0.0054787760600447655, + 0.006298983469605446, + 0.008498664014041424, + 0.009314201772212982, + 0.0065725152380764484, + 0.006026790477335453, + 0.021005932241678238, + 0.006055113859474659, + 0.007206206675618887, + 0.03653584420681, + 0.01845654658973217, + 0.018880702555179596, + 0.021236147731542587, + 0.019748549908399582, + 0.019197430461645126, + 0.024186646565794945, + 0.013750139623880386, + 0.015109576284885406, + 0.02023407816886902, + 0.01352972537279129, + 0.01125701330602169, + 0.02586291916668415, + 0.023006465286016464, + 0.01294270996004343, + 0.009297640062868595, + 0.012106478214263916, + 0.012533202767372131, + 0.011330511420965195, + 0.013055805116891861, + 0.01450691744685173, + 0.011754626408219337, + 0.016215011477470398, + 0.013180926442146301, + 0.013211715035140514, + 0.016603592783212662, + 0.010980588383972645, + 0.014103299006819725, + 0.021299604326486588, + 0.013188505545258522, + 0.01825803890824318, + 0.013053474016487598, + 0.006116641219705343, + 0.006102803163230419, + 0.0042876084335148335, + 0.006590350531041622, + 0.0066947913728654385, + 0.009870943613350391, + 0.00678300904110074, + 0.006886588875204325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.023304030299186707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865148663520813 + ], + "opposite_logprob_tokens": [ + -3.0486514568328857 + ], + "logprob_reward": -0.04865148663520813, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865148663520813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07195362448692322, + 0.06365067511796951, + 0.08260748535394669, + 0.09876123815774918, + 0.10526000708341599, + 0.07751017808914185, + 0.08074036985635757, + 0.27288731932640076, + 0.10250027477741241, + 0.04412882775068283 + ], + "credit_map": [ + 0.007701662369072437, + 0.00787028856575489, + 0.006148225627839565, + 0.007798359729349613, + 0.01612752117216587, + 0.01006244495511055, + 0.008489659987390041, + 0.0077554574236273766, + 0.0133030591532588, + 0.009638484567403793, + 0.005942162591964006, + 0.009539489634335041, + 0.00589642534032464, + 0.005081122741103172, + 0.0071115330792963505, + 0.007138400804251432, + 0.00670109037309885, + 0.006247656885534525, + 0.006469746120274067, + 0.018260233104228973, + 0.009504023939371109, + 0.01099202036857605, + 0.012287653051316738, + 0.012145058251917362, + 0.011761381290853024, + 0.017419153824448586, + 0.010259195230901241, + 0.012349442578852177, + 0.015102033503353596, + 0.010968396440148354, + 0.01341310515999794, + 0.007488529197871685, + 0.009655828587710857, + 0.010175853036344051, + 0.019832052290439606, + 0.01060382928699255, + 0.008493293076753616, + 0.0264944639056921, + 0.0103931725025177, + 0.009611513465642929, + 0.01989147625863552, + 0.008511089719831944, + 0.007767166942358017, + 0.01300005055963993, + 0.008485909551382065, + 0.006435939110815525, + 0.006686046719551086, + 0.006732496432960033, + 0.009350361302495003, + 0.00761968782171607, + 0.01547300349920988, + 0.008832686580717564, + 0.009034961462020874, + 0.010411042720079422, + 0.008681114763021469, + 0.011337516829371452, + 0.029301932081580162, + 0.03128553926944733, + 0.03170056268572807, + 0.036623723804950714, + 0.033445440232753754, + 0.033194661140441895, + 0.03577035665512085, + 0.04156510531902313, + 0.010803124867379665, + 0.013616442680358887, + 0.008910086005926132, + 0.012214144691824913, + 0.018305018544197083, + 0.011869319714605808, + 0.014754809439182281, + 0.012027326971292496, + 0.0036177809815853834, + 0.0034393633250147104, + 0.004314877558499575, + 0.0049074008129537106, + 0.004213427659124136, + 0.010610939003527164, + 0.005474372301250696, + 0.00755066704005003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.04865148663520813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.600520372390747 + ], + "opposite_logprob_tokens": [ + -0.2255203276872635 + ], + "logprob_reward": -1.600520372390747, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.600520372390747, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13391520082950592, + 0.08789961785078049, + 0.05692537501454353, + 0.23152893781661987, + 0.12411894649267197, + 0.09874247014522552, + 0.09780630469322205, + 0.03606489300727844, + 0.08713995665311813, + 0.045858316123485565 + ], + "credit_map": [ + 0.014485388062894344, + 0.015426895581185818, + 0.01594557613134384, + 0.024274177849292755, + 0.01974600926041603, + 0.014307328499853611, + 0.015960246324539185, + 0.013769581913948059, + 0.011973723769187927, + 0.01118968240916729, + 0.015079583041369915, + 0.011929872445762157, + 0.011747987940907478, + 0.00763006042689085, + 0.01008009072393179, + 0.008268618956208229, + 0.008228314109146595, + 0.005049463361501694, + 0.007758231367915869, + 0.010696986690163612, + 0.006423308979719877, + 0.007026240229606628, + 0.006101650185883045, + 0.005641177296638489, + 0.027391821146011353, + 0.04103465750813484, + 0.026366710662841797, + 0.026963913813233376, + 0.03745898976922035, + 0.026247311383485794, + 0.02858656644821167, + 0.017478983849287033, + 0.015887565910816193, + 0.01812138967216015, + 0.023124461993575096, + 0.014983157627284527, + 0.012246145866811275, + 0.011490951292216778, + 0.019286509603261948, + 0.008978764526546001, + 0.02387789823114872, + 0.01073417253792286, + 0.014233599416911602, + 0.009847095236182213, + 0.011056696064770222, + 0.0072095803916454315, + 0.011681361123919487, + 0.010102064348757267, + 0.011290473863482475, + 0.011931609362363815, + 0.011492935940623283, + 0.018902447074651718, + 0.012680684216320515, + 0.011388265527784824, + 0.010381580330431461, + 0.009738308377563953, + 0.005262257065623999, + 0.004730979911983013, + 0.0034162334632128477, + 0.004070846363902092, + 0.004926344845443964, + 0.003878069808706641, + 0.005018018186092377, + 0.0047621410340070724, + 0.008919216692447662, + 0.008837252855300903, + 0.007490647491067648, + 0.009409179911017418, + 0.012959732674062252, + 0.008434249088168144, + 0.014825418591499329, + 0.016264259815216064, + 0.003869691863656044, + 0.004345935769379139, + 0.004410243593156338, + 0.004391644150018692, + 0.005494601093232632, + 0.0078001744113862514, + 0.006797187961637974, + 0.008748829364776611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -1.600520372390747 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559529542922974 + ], + "opposite_logprob_tokens": [ + -1.600595235824585 + ], + "logprob_reward": -0.22559529542922974, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559529542922974, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11720216274261475, + 0.07623617351055145, + 0.07078906148672104, + 0.04988180473446846, + 0.1288478970527649, + 0.09509755671024323, + 0.1599653959274292, + 0.08666841685771942, + 0.08164052665233612, + 0.13367100059986115 + ], + "credit_map": [ + 0.01195610873401165, + 0.012746671214699745, + 0.013312491588294506, + 0.022632580250501633, + 0.01647942326962948, + 0.011230653151869774, + 0.014443755149841309, + 0.014400480315089226, + 0.010595309548079967, + 0.008142251521348953, + 0.008451254107058048, + 0.012726722285151482, + 0.008025896735489368, + 0.01102025993168354, + 0.009002232924103737, + 0.008272246457636356, + 0.011478221975266933, + 0.00571327144280076, + 0.0079392921179533, + 0.009534364566206932, + 0.008305440656840801, + 0.008940077386796474, + 0.00954408198595047, + 0.009334312751889229, + 0.008001338690519333, + 0.0077439481392502785, + 0.005883392412215471, + 0.0044054677709937096, + 0.007992727681994438, + 0.005553919356316328, + 0.0061546675860881805, + 0.004146346356719732, + 0.009635469876229763, + 0.013766066171228886, + 0.015896914526820183, + 0.016569403931498528, + 0.014516577124595642, + 0.027905508875846863, + 0.016486693173646927, + 0.014071252197027206, + 0.01803020015358925, + 0.01017751544713974, + 0.015041078440845013, + 0.012253976427018642, + 0.012028730474412441, + 0.007761497050523758, + 0.012897436507046223, + 0.006907121744006872, + 0.019510572776198387, + 0.01326274499297142, + 0.02563905343413353, + 0.015588147565722466, + 0.02407231740653515, + 0.03122333064675331, + 0.011778067797422409, + 0.018891170620918274, + 0.00992548931390047, + 0.009593049995601177, + 0.012714020907878876, + 0.010105195455253124, + 0.011365214362740517, + 0.008174855262041092, + 0.012753826566040516, + 0.012036764994263649, + 0.008317060768604279, + 0.010955830104649067, + 0.00813940167427063, + 0.011010052636265755, + 0.011387275531888008, + 0.008326621726155281, + 0.012180565856397152, + 0.0113237164914608, + 0.008065379224717617, + 0.009117710404098034, + 0.012013447470963001, + 0.012065539136528969, + 0.0135792912915349, + 0.030933687463402748, + 0.018876567482948303, + 0.02901938371360302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.22559529542922974 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "In a powerful, live-action shot, the camera pans at breakneck speed in a tight 360-degree pan. At the center of the frame is a muscular, classically sculpted king of the sea. Bare-chested, his bronzed skin gleams under the spotlights overhead; his long, wet, dark hair clings to his broad shoulders and back. He's at the apex of a movement, his arms flexing with muscle as he thrusts a trident, radiating a brilliant cyan-blue light, into a massive block of black basalt in the center of the arena. The moment the trident's tip impacts the rock, a dazzling shower of sparks and rubble erupts, while a wave of energy whips the shallow puddle beneath his feet into a spectacular, outward-spreading white spray. As the camera rotates at high speed, we can see water droplets condensing into crystal beads in the air, reflecting the light of the trident, and his resolute face, his eyes are focused and fierce, as if declaring the victory of an epic duel. The whole scene is filled with primitive, magnificent and highly impactful epic atmosphere under the cold light and shadow.", + "seed": 4962, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_1/index_2247_seed_4962_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.34807060225866737, + "vlm_logprob": -0.34807060225866737, + "vlm_yes_no_margin": 2.85, + "vlm_token_credit": -0.34807060225866737, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.039814140647649765 + ], + "opposite_logprob_tokens": [ + -3.2898142337799072 + ], + "logprob_reward": -0.039814140647649765, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.039814140647649765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20857775211334229, + 0.08226417005062103, + 0.05631479620933533, + 0.08286236226558685, + 0.19838134944438934, + 0.06679639220237732, + 0.10008889436721802, + 0.08804561197757721, + 0.07049628347158432, + 0.046172402799129486 + ], + "credit_map": [ + 0.04541631415486336, + 0.025869866833090782, + 0.024474741891026497, + 0.019854171201586723, + 0.016471756622195244, + 0.024261314421892166, + 0.019996201619505882, + 0.03223337233066559, + 0.006064633373171091, + 0.006448072846978903, + 0.010637792758643627, + 0.010921875014901161, + 0.008663993328809738, + 0.01015238743275404, + 0.014945297501981258, + 0.014430113136768341, + 0.005595558788627386, + 0.008962901309132576, + 0.005172905046492815, + 0.010342703200876713, + 0.006019327323883772, + 0.004102454986423254, + 0.008734374307096004, + 0.0073845707811415195, + 0.007639499381184578, + 0.010745007544755936, + 0.008922042325139046, + 0.01136693637818098, + 0.01385855209082365, + 0.009950494393706322, + 0.010347986593842506, + 0.010031846351921558, + 0.022907918319106102, + 0.026291847229003906, + 0.014914627186954021, + 0.017076825723052025, + 0.02827964350581169, + 0.04044746980071068, + 0.027776233851909637, + 0.020686781033873558, + 0.005306712817400694, + 0.00740403775125742, + 0.009765369817614555, + 0.009952966123819351, + 0.007390846032649279, + 0.009179610759019852, + 0.00681667635217309, + 0.010980173945426941, + 0.01218600757420063, + 0.012803878635168076, + 0.017435217276215553, + 0.007996641099452972, + 0.012482909485697746, + 0.0166772473603487, + 0.012832729145884514, + 0.0076742698438465595, + 0.008922835811972618, + 0.00891084037721157, + 0.00732925022020936, + 0.015232293866574764, + 0.013836676254868507, + 0.009166239760816097, + 0.015687186270952225, + 0.008960291743278503, + 0.006947239860892296, + 0.010639533400535583, + 0.006523227319121361, + 0.010093777440488338, + 0.009723256342113018, + 0.007675061002373695, + 0.008484767749905586, + 0.010409420356154442, + 0.005478710867464542, + 0.0038981614634394646, + 0.004126852843910456, + 0.004697798285633326, + 0.005471597891300917, + 0.01325166691094637, + 0.004939830396324396, + 0.004307783208787441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.039814140647649765 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028350658249109983 + ], + "opposite_logprob_tokens": [ + -5.877835273742676 + ], + "logprob_reward": -0.0028350658249109983, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028350658249109983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27362969517707825, + 0.09808360040187836, + 0.08315200358629227, + 0.0800073966383934, + 0.04692167416214943, + 0.07352413237094879, + 0.07259392738342285, + 0.17620117962360382, + 0.07110818475484848, + 0.02477814070880413 + ], + "credit_map": [ + 0.05047256872057915, + 0.034478772431612015, + 0.03144213557243347, + 0.042421650141477585, + 0.024783845990896225, + 0.023291639983654022, + 0.030536262318491936, + 0.03620283678174019, + 0.006737865041941404, + 0.00931607373058796, + 0.010209864005446434, + 0.013614777475595474, + 0.014205196872353554, + 0.01368014421314001, + 0.017642879858613014, + 0.012676802463829517, + 0.005727911368012428, + 0.012248506769537926, + 0.009039627388119698, + 0.01143592968583107, + 0.01683470420539379, + 0.010594972409307957, + 0.009756047278642654, + 0.007514301687479019, + 0.006309814285486937, + 0.012046130374073982, + 0.008598090149462223, + 0.008782822638750076, + 0.015211666002869606, + 0.008199523203074932, + 0.010600931942462921, + 0.010258425027132034, + 0.003133648307994008, + 0.007100472692400217, + 0.0042406790889799595, + 0.005278593860566616, + 0.007878804579377174, + 0.005168193019926548, + 0.008246056735515594, + 0.005875225644558668, + 0.005899501033127308, + 0.011111640371382236, + 0.007472700905054808, + 0.009453791193664074, + 0.011567668989300728, + 0.007610760163515806, + 0.010738711804151535, + 0.009669356979429722, + 0.008497967384755611, + 0.007271385286003351, + 0.011116052977740765, + 0.0051332623697817326, + 0.008018452674150467, + 0.016331158578395844, + 0.009945901110768318, + 0.0062797474674880505, + 0.022567182779312134, + 0.04077715426683426, + 0.0186087004840374, + 0.02469373680651188, + 0.01850186474621296, + 0.023201854899525642, + 0.013278832659125328, + 0.014571861363947392, + 0.0065443310886621475, + 0.008909451775252819, + 0.006269136909395456, + 0.011037865653634071, + 0.011396965011954308, + 0.007689961697906256, + 0.009398071095347404, + 0.009862395003437996, + 0.003661870490759611, + 0.0028885521460324526, + 0.002126803621649742, + 0.0023788942489773035, + 0.002621921245008707, + 0.004474555142223835, + 0.004045187961310148, + 0.002580357016995549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0028350658249109983 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187422305345535 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187422305345535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187422305345535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1523965299129486, + 0.3039935231208801, + 0.07195062190294266, + 0.09217094630002975, + 0.04941020533442497, + 0.042468730360269547, + 0.05829533562064171, + 0.08271394670009613, + 0.05079624056816101, + 0.09580392390489578 + ], + "credit_map": [ + 0.01661308854818344, + 0.017629211768507957, + 0.016511622816324234, + 0.024795997887849808, + 0.019169500097632408, + 0.02141137607395649, + 0.01768677309155464, + 0.018578948453068733, + 0.02151016891002655, + 0.03576720878481865, + 0.031999848783016205, + 0.05199078470468521, + 0.038753021508455276, + 0.032938748598098755, + 0.06152397766709328, + 0.02950974553823471, + 0.0061194696463644505, + 0.008756309747695923, + 0.006476717535406351, + 0.012786957435309887, + 0.010151563212275505, + 0.006641070824116468, + 0.012948600575327873, + 0.00806993618607521, + 0.007569623179733753, + 0.01232023723423481, + 0.009688274003565311, + 0.01185366790741682, + 0.0165958721190691, + 0.009218461811542511, + 0.012475916184484959, + 0.012448892928659916, + 0.004256642423570156, + 0.006779355462640524, + 0.004769996274262667, + 0.005617393646389246, + 0.009591192938387394, + 0.006071073468774557, + 0.007422842551022768, + 0.004901709500700235, + 0.004158944822847843, + 0.006526041775941849, + 0.003915166482329369, + 0.005168812815099955, + 0.006583955138921738, + 0.006171878427267075, + 0.0049707526341080666, + 0.00497317872941494, + 0.008210767060518265, + 0.009036194533109665, + 0.008299036882817745, + 0.005761653184890747, + 0.0077437362633645535, + 0.008191780187189579, + 0.0069142417050898075, + 0.00413792859762907, + 0.011416036635637283, + 0.020433273166418076, + 0.008651437237858772, + 0.011244233697652817, + 0.007611365523189306, + 0.010882272385060787, + 0.0062250737100839615, + 0.0062502529472112656, + 0.0039617568254470825, + 0.005385173950344324, + 0.004773053806275129, + 0.010362754575908184, + 0.007313500624150038, + 0.004762143362313509, + 0.006496907211840153, + 0.007740952540189028, + 0.008987683802843094, + 0.009739052504301071, + 0.008690489456057549, + 0.013772410340607166, + 0.010610218159854412, + 0.018722932785749435, + 0.014922011643648148, + 0.010359120555222034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.018187422305345535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6005284786224365 + ], + "opposite_logprob_tokens": [ + -0.2255285084247589 + ], + "logprob_reward": -1.6005284786224365, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6005284786224365, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20361199975013733, + 0.2163306027650833, + 0.04141991585493088, + 0.0731055736541748, + 0.05214255675673485, + 0.07996030896902084, + 0.16690132021903992, + 0.07168478518724442, + 0.055361662060022354, + 0.03948131576180458 + ], + "credit_map": [ + 0.017238294705748558, + 0.02040798030793667, + 0.02522214688360691, + 0.023049088194966316, + 0.027904387563467026, + 0.03424505889415741, + 0.025021325796842575, + 0.03052368201315403, + 0.01752442680299282, + 0.022542906925082207, + 0.027979066595435143, + 0.028793057426810265, + 0.019405843690037727, + 0.027394821867346764, + 0.0355302058160305, + 0.0371602438390255, + 0.003692813217639923, + 0.00731792813166976, + 0.004706544801592827, + 0.005639746785163879, + 0.004450541455298662, + 0.005118999630212784, + 0.0045397523790597916, + 0.005953585729002953, + 0.00605453597381711, + 0.010723615065217018, + 0.0073966458439826965, + 0.006865797098726034, + 0.012496044859290123, + 0.008808826096355915, + 0.012012883089482784, + 0.008747207932174206, + 0.0040265885181725025, + 0.006057363469153643, + 0.005278274416923523, + 0.006085319444537163, + 0.009037203155457973, + 0.006757193710654974, + 0.007674947381019592, + 0.007225656881928444, + 0.005638777744024992, + 0.013049525208771229, + 0.00923537090420723, + 0.0083498265594244, + 0.011138899251818657, + 0.01026248186826706, + 0.01133561972528696, + 0.01094979327172041, + 0.019603393971920013, + 0.016934234648942947, + 0.030600758269429207, + 0.009958513081073761, + 0.02089514769613743, + 0.033183783292770386, + 0.02326132170855999, + 0.012464150786399841, + 0.009538299404084682, + 0.016070302575826645, + 0.010472599416971207, + 0.009844895452260971, + 0.007157689426094294, + 0.008173651061952114, + 0.005889390595257282, + 0.004537947475910187, + 0.005854637362062931, + 0.005812083370983601, + 0.004486325196921825, + 0.007424802519381046, + 0.008563258685171604, + 0.00667567690834403, + 0.009914476424455643, + 0.006630394607782364, + 0.0036686547100543976, + 0.004821981303393841, + 0.003978438209742308, + 0.005849077366292477, + 0.004802686162292957, + 0.0035491983871906996, + 0.007405179087072611, + 0.005406096111983061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -1.6005284786224365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the central subject a muscular, bare-chested king of the sea with bronzed skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character thrusting a trident that radiates a brilliant cyan-blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the trident impact a massive block of black basalt, causing sparks and rubble to erupt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform a tight 360-degree pan at breakneck speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a white spray of water spreading outward from a shallow puddle beneath the character's feet?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898790389299393 + ], + "opposite_logprob_tokens": [ + -2.5789878368377686 + ], + "logprob_reward": -0.07898790389299393, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898790389299393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2584232687950134, + 0.23384317755699158, + 0.038866203278303146, + 0.05991421639919281, + 0.10847847163677216, + 0.059694405645132065, + 0.07981736212968826, + 0.05478513613343239, + 0.036535151302814484, + 0.06964261829853058 + ], + "credit_map": [ + 0.032287415117025375, + 0.028198737651109695, + 0.02979533188045025, + 0.03035612963140011, + 0.031871337443590164, + 0.035974495112895966, + 0.03467371687293053, + 0.03526610508561134, + 0.017988018691539764, + 0.023682497441768646, + 0.02557450719177723, + 0.03671709820628166, + 0.0317048504948616, + 0.0336216576397419, + 0.03431784734129906, + 0.03023671545088291, + 0.003164540510624647, + 0.005041354335844517, + 0.003740813583135605, + 0.0067311483435332775, + 0.005731501616537571, + 0.0037566914688795805, + 0.006090232171118259, + 0.004609918687492609, + 0.005648836959153414, + 0.006535204127430916, + 0.005863287020474672, + 0.009160454384982586, + 0.007746872957795858, + 0.00714096799492836, + 0.010143839754164219, + 0.0076747494749724865, + 0.01105614472180605, + 0.014791158959269524, + 0.010678657330572605, + 0.010771958157420158, + 0.012518472038209438, + 0.016881676390767097, + 0.018208438530564308, + 0.013571963645517826, + 0.004442628938704729, + 0.007907979190349579, + 0.007266323547810316, + 0.005634425673633814, + 0.008862479589879513, + 0.00800127349793911, + 0.00848398543894291, + 0.009095307439565659, + 0.009869995526969433, + 0.006908949930220842, + 0.013916080817580223, + 0.0054341871291399, + 0.008782310411334038, + 0.01750851795077324, + 0.011289490386843681, + 0.006107828579843044, + 0.005840901285409927, + 0.007503157947212458, + 0.007090854458510876, + 0.007429440971463919, + 0.007628186140209436, + 0.0061729406006634235, + 0.007868933491408825, + 0.005250721704214811, + 0.004833992104977369, + 0.00560804083943367, + 0.0029762231279164553, + 0.0049109007231891155, + 0.004743826109915972, + 0.00475456053391099, + 0.004828880541026592, + 0.0038787280209362507, + 0.006588033866137266, + 0.007097065448760986, + 0.006965615786612034, + 0.013758591376245022, + 0.004876752384006977, + 0.006654093042016029, + 0.012560969218611717, + 0.011141500435769558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07898790389299393 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_tensor([3532, 3532])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_tensor([3532, 3532])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5ea5580c0f0cfdfc44dc3269b36ef06ce2201e96 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_tensor([3532, 3532])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins with a dynamic close-up, capturing an unexpected scene at an Olympic-level indoor ice climbing arena: Sanji, dressed in a perfectly tailored black single-breasted suit, his signature short blond hair meticulously combed back, gleaming softly under the intense, cold white light cast by the arena's ceiling spotlights. With the precision and power of a ballerina, he swings his right arm. The metal tip of his professional ice axe traces a swift arc through the air, and with a sharp click, it drives decisively and deeply into the vertical, shimmering, artificial ice wall before him. At the moment of impact, countless tiny, shimmering ice shards explode toward the camera like diamonds, each one radiating a dazzling light in the strong sunlight. The frame is now sharply focused on these dancing ice crystals. Then, the camera executes a swift rack focus, the blurred ice shards instantly receding into the background. Focus is fixed sharply on his slightly raised face, capturing the steely, determined eyes beneath his signature curled brows. In that instant, the screen flashes rapidly, as if overloaded with energy, a flash of white light amplifying the moment's impact. The entire image, in its stark black, white, and blue tones, creates a surreal atmosphere, a unique blend of ultimate elegance and the intensity of extreme sports.", + "seed": 17053, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17053_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05652140276506543, + "vlm_logprob": -0.05652140276506543, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05652140276506543, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13507327437400818 + ], + "opposite_logprob_tokens": [ + -2.260073184967041 + ], + "logprob_reward": -0.13507327437400818, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.13507327437400818, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10138298571109772, + 0.07409094274044037, + 0.29882070422172546, + 0.0760052353143692, + 0.07365764677524567, + 0.06648535281419754, + 0.06662175804376602, + 0.0755661353468895, + 0.07968150079250336, + 0.08768772333860397 + ], + "credit_map": [ + 0.010494866408407688, + 0.011665347032248974, + 0.009266093373298645, + 0.008287142962217331, + 0.013713039457798004, + 0.019398214295506477, + 0.014726114459335804, + 0.013832172378897667, + 0.00560016231611371, + 0.007245491724461317, + 0.007123569957911968, + 0.008688102476298809, + 0.008990567177534103, + 0.01701146550476551, + 0.008561499416828156, + 0.010870084166526794, + 0.007879391312599182, + 0.005203052423894405, + 0.006377768237143755, + 0.005655791610479355, + 0.04843003675341606, + 0.08593308180570602, + 0.1265731155872345, + 0.012768457643687725, + 0.0061296867206692696, + 0.007435327395796776, + 0.006476618815213442, + 0.008963216096162796, + 0.0076739895157516, + 0.014697766862809658, + 0.01081467978656292, + 0.013813941739499569, + 0.009146107360720634, + 0.00911283865571022, + 0.009347719140350819, + 0.00897117331624031, + 0.010679925791919231, + 0.008615581318736076, + 0.00963732972741127, + 0.008146975189447403, + 0.004623836372047663, + 0.004559915978461504, + 0.004131597466766834, + 0.004415773320943117, + 0.02377144992351532, + 0.010498707182705402, + 0.00912064965814352, + 0.005363422445952892, + 0.006115269381552935, + 0.0072815571911633015, + 0.006948502268642187, + 0.006996870040893555, + 0.014725346118211746, + 0.009588356129825115, + 0.008048203773796558, + 0.006917655002325773, + 0.007004824932664633, + 0.007323405705392361, + 0.010060571134090424, + 0.009880192577838898, + 0.013391991145908833, + 0.010321609675884247, + 0.010401000268757343, + 0.007182534784078598, + 0.011041385121643543, + 0.009986345656216145, + 0.004939379170536995, + 0.00680432328954339, + 0.015245277434587479, + 0.009299096651375294, + 0.012793743051588535, + 0.009571947157382965, + 0.00812002457678318, + 0.011741151101887226, + 0.0031413540709763765, + 0.00871826522052288, + 0.013073588721454144, + 0.017054157331585884, + 0.008690042421221733, + 0.01714913919568062 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.13507327437400818 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062106654047966 + ], + "opposite_logprob_tokens": [ + -2.8121066093444824 + ], + "logprob_reward": -0.062106654047966, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062106654047966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13468365371227264, + 0.15325519442558289, + 0.3404875695705414, + 0.17601467669010162, + 0.07108744233846664, + 0.026048194617033005, + 0.02965216152369976, + 0.020296791568398476, + 0.02764897793531418, + 0.020825326442718506 + ], + "credit_map": [ + 0.012829985469579697, + 0.012424821965396404, + 0.012583634816110134, + 0.009715934284031391, + 0.02792542800307274, + 0.02088671550154686, + 0.02081453613936901, + 0.017502613365650177, + 0.008279277011752129, + 0.012791967019438744, + 0.014251833781599998, + 0.016844918951392174, + 0.021055489778518677, + 0.03946838527917862, + 0.02448943816125393, + 0.016073886305093765, + 0.01026403158903122, + 0.005215090233832598, + 0.006677794270217419, + 0.005254449788480997, + 0.13823877274990082, + 0.026768729090690613, + 0.13265426456928253, + 0.015414456836879253, + 0.016184957697987556, + 0.018754884600639343, + 0.022597819566726685, + 0.02017817460000515, + 0.02857501246035099, + 0.03409586474299431, + 0.012421567924320698, + 0.023206396028399467, + 0.011078046634793282, + 0.012760365381836891, + 0.00810268148779869, + 0.007072758860886097, + 0.009105219505727291, + 0.009208842180669308, + 0.0056870924308896065, + 0.008072445169091225, + 0.002327057532966137, + 0.0021523265168070793, + 0.0019142822129651904, + 0.001582679687999189, + 0.008374367840588093, + 0.003953256644308567, + 0.004111863672733307, + 0.0016323613235726953, + 0.0025437860749661922, + 0.003162924200296402, + 0.002669546753168106, + 0.0035144260618835688, + 0.006220689974725246, + 0.004370945505797863, + 0.004065260756760836, + 0.003104585222899914, + 0.002285453025251627, + 0.0025048593524843454, + 0.0028852091636508703, + 0.003434207057580352, + 0.002796337241306901, + 0.0021165485959500074, + 0.0024436672683805227, + 0.0018305110279470682, + 0.004883667454123497, + 0.0027967237401753664, + 0.002892445307224989, + 0.0029501523822546005, + 0.00437778839841485, + 0.0033426897134631872, + 0.003012183355167508, + 0.00339332758449018, + 0.001997758401557803, + 0.0019832225516438484, + 0.0007623438723385334, + 0.0027819639071822166, + 0.0018682486843317747, + 0.004587343893945217, + 0.001341113355010748, + 0.005503332708030939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.062106654047966 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004860043525696 + ], + "opposite_logprob_tokens": [ + -2.6950485706329346 + ], + "logprob_reward": -0.07004860043525696, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004860043525696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06971234828233719, + 0.030444275587797165, + 0.8060348033905029, + 0.01880822516977787, + 0.010513242334127426, + 0.01751575991511345, + 0.016974536702036858, + 0.007725566625595093, + 0.014376942068338394, + 0.007894298061728477 + ], + "credit_map": [ + 0.00829231645911932, + 0.00767589220777154, + 0.005761795677244663, + 0.005075330380350351, + 0.012902957387268543, + 0.01127773616462946, + 0.012403246015310287, + 0.006323077715933323, + 0.0019087359542027116, + 0.0030533766839653254, + 0.0025574478786438704, + 0.0030200406908988953, + 0.004514434840530157, + 0.006489221006631851, + 0.0063292342238128185, + 0.0025717834942042828, + 0.007997008971869946, + 0.008375316858291626, + 0.013560895808041096, + 0.006990425754338503, + 0.3782292902469635, + 0.22900128364562988, + 0.14093081653118134, + 0.020949712023139, + 0.001908687874674797, + 0.0021034979727119207, + 0.0026779528707265854, + 0.003199532860890031, + 0.002054347191005945, + 0.0025829137302935123, + 0.0017336445162072778, + 0.0025476478040218353, + 0.0010909470729529858, + 0.000978936324827373, + 0.0009759836830198765, + 0.0013237829552963376, + 0.0018441574648022652, + 0.0014616043772548437, + 0.0015377710806205869, + 0.001300059724599123, + 0.001623885938897729, + 0.0019718341063708067, + 0.001656386535614729, + 0.001367722637951374, + 0.0043905735947191715, + 0.002200578572228551, + 0.0030671211425215006, + 0.0012376576196402311, + 0.002023535082116723, + 0.0025272604543715715, + 0.001720051746815443, + 0.0017049676971510053, + 0.002880432875826955, + 0.0024543043691664934, + 0.0022851552348583937, + 0.0013788307551294565, + 0.0006511642714031041, + 0.0008248391095548868, + 0.00110564602073282, + 0.0008417440112680197, + 0.001523544080555439, + 0.0011029975721612573, + 0.0009206953109242022, + 0.000754936074372381, + 0.001693287049420178, + 0.0018181673949584365, + 0.0013748575001955032, + 0.001673599355854094, + 0.002186995930969715, + 0.001654077204875648, + 0.0022723209112882614, + 0.0017036378849297762, + 0.0009271649178117514, + 0.0006637450424022973, + 0.00035864440724253654, + 0.0009904892649501562, + 0.0005521883140318096, + 0.0017338560428470373, + 0.0008841990493237972, + 0.0017840109067037702 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.07004860043525696 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0689154863357544, + 0.05007277801632881, + 0.6442906260490417, + 0.02259712852537632, + 0.03410986065864563, + 0.024363301694393158, + 0.028565876185894012, + 0.051826827228069305, + 0.05156063660979271, + 0.023697521537542343 + ], + "credit_map": [ + 0.004696047864854336, + 0.005336804781109095, + 0.004329514689743519, + 0.006011064164340496, + 0.015571561641991138, + 0.01132205780595541, + 0.016157396137714386, + 0.005491031799465418, + 0.003928212448954582, + 0.005135785788297653, + 0.005917034111917019, + 0.005179751664400101, + 0.009309304878115654, + 0.007042127661406994, + 0.007394108921289444, + 0.006166447885334492, + 0.010971156880259514, + 0.00905860960483551, + 0.011546272784471512, + 0.007765563204884529, + 0.29387155175209045, + 0.09072861820459366, + 0.20169134438037872, + 0.018657464534044266, + 0.00227166386321187, + 0.002388786757364869, + 0.0021727473940700293, + 0.003284655511379242, + 0.0032234115060418844, + 0.003903236472979188, + 0.002194040222093463, + 0.0031585858669131994, + 0.005246391054242849, + 0.007394240703433752, + 0.0028150456491857767, + 0.0025048512034118176, + 0.005161208566278219, + 0.004639331251382828, + 0.0026486190035939217, + 0.0037001732271164656, + 0.002747397171333432, + 0.0030849548056721687, + 0.0024347479920834303, + 0.0022113025188446045, + 0.004663344006985426, + 0.0028517628088593483, + 0.004819187801331282, + 0.00155060482211411, + 0.0023707523941993713, + 0.001641872338950634, + 0.0019240448018535972, + 0.0027686194516718388, + 0.008470711298286915, + 0.004752891603857279, + 0.0034695283975452185, + 0.0031674548517912626, + 0.005116921383887529, + 0.0037692703772336245, + 0.007845007814466953, + 0.007503856904804707, + 0.008968116715550423, + 0.006638866383582354, + 0.007357879541814327, + 0.004626901354640722, + 0.006645574700087309, + 0.005302739329636097, + 0.004927098751068115, + 0.005953173153102398, + 0.008724719285964966, + 0.005804712418466806, + 0.008794263005256653, + 0.005408354569226503, + 0.003663617419078946, + 0.002618337282910943, + 0.0018411398632451892, + 0.001741505227982998, + 0.0022917164023965597, + 0.003818647237494588, + 0.003758444683626294, + 0.00396411307156086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.008642058819532394 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006736426148563623 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.006736426148563623, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006736426148563623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08505531400442123, + 0.05546937882900238, + 0.5159488916397095, + 0.03337117284536362, + 0.0362292155623436, + 0.08183486014604568, + 0.05993010848760605, + 0.03253324329853058, + 0.05438590422272682, + 0.04524195194244385 + ], + "credit_map": [ + 0.006043741945177317, + 0.0064231716096401215, + 0.004521089140325785, + 0.006071368232369423, + 0.013148325495421886, + 0.02090451866388321, + 0.015374688431620598, + 0.012568404898047447, + 0.003987670876085758, + 0.005236215889453888, + 0.005509187933057547, + 0.0058354963548481464, + 0.008119913749396801, + 0.01196117140352726, + 0.007477210368961096, + 0.007342511788010597, + 0.008409774862229824, + 0.006887253373861313, + 0.00882516149431467, + 0.005015947390347719, + 0.2376594841480255, + 0.0974532961845398, + 0.13707078993320465, + 0.01462717168033123, + 0.002007419243454933, + 0.00286498898640275, + 0.003239183221012354, + 0.005006715655326843, + 0.004732797387987375, + 0.005286029540002346, + 0.00624079629778862, + 0.003993244841694832, + 0.0033686074893921614, + 0.0068465471267700195, + 0.0026936777867376804, + 0.003628958249464631, + 0.007374683395028114, + 0.0048814634792506695, + 0.004072684794664383, + 0.0033625951036810875, + 0.005645350553095341, + 0.006374660413712263, + 0.004649563692510128, + 0.005131393205374479, + 0.02980124019086361, + 0.013297387398779392, + 0.011617500334978104, + 0.005317769479006529, + 0.0038006382528692484, + 0.005013215355575085, + 0.004881450440734625, + 0.006148626562207937, + 0.015407165512442589, + 0.011327816173434258, + 0.007679687347263098, + 0.005671507213264704, + 0.0028145923279225826, + 0.0035678029526025057, + 0.003929763566702604, + 0.0032872932497411966, + 0.006108337547630072, + 0.004738511983305216, + 0.0047834208235144615, + 0.003303518984466791, + 0.004845831077545881, + 0.005520636681467295, + 0.0036910693161189556, + 0.004527532495558262, + 0.012925899587571621, + 0.006988080218434334, + 0.009879302233457565, + 0.006007555406540632, + 0.0057789902202785015, + 0.004502809606492519, + 0.0029620674904435873, + 0.0032313053961843252, + 0.003870444605126977, + 0.009287926368415356, + 0.006567064207047224, + 0.009041343815624714 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.006736426148563623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a dynamic close-up, capturing an unexpected scene at an Olympic-level indoor ice climbing arena: Sanji, dressed in a perfectly tailored black single-breasted suit, his signature short blond hair meticulously combed back, gleaming softly under the intense, cold white light cast by the arena's ceiling spotlights. With the precision and power of a ballerina, he swings his right arm. The metal tip of his professional ice axe traces a swift arc through the air, and with a sharp click, it drives decisively and deeply into the vertical, shimmering, artificial ice wall before him. At the moment of impact, countless tiny, shimmering ice shards explode toward the camera like diamonds, each one radiating a dazzling light in the strong sunlight. The frame is now sharply focused on these dancing ice crystals. Then, the camera executes a swift rack focus, the blurred ice shards instantly receding into the background. Focus is fixed sharply on his slightly raised face, capturing the steely, determined eyes beneath his signature curled brows. In that instant, the screen flashes rapidly, as if overloaded with energy, a flash of white light amplifying the moment's impact. The entire image, in its stark black, white, and blue tones, creates a surreal atmosphere, a unique blend of ultimate elegance and the intensity of extreme sports.", + "seed": 17054, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17054_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.13971534734591842, + "vlm_logprob": -0.13971534734591842, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.13971534734591842, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.29760709404945374 + ], + "opposite_logprob_tokens": [ + -1.4226070642471313 + ], + "logprob_reward": -0.29760709404945374, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.29760709404945374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.011260507628321648, + 0.005157098174095154, + 0.0063853575848042965, + 0.015028778463602066, + 0.00634384760633111, + 0.011412753723561764, + 0.013219719752669334, + 0.9232074618339539, + 0.004583288449794054, + 0.0034012964460998774 + ], + "credit_map": [ + 0.0015532474499195814, + 0.0020340606570243835, + 0.0018134762067347765, + 0.00071807048516348, + 0.002771463943645358, + 0.0007497288752347231, + 0.0006334870122373104, + 0.00098697142675519, + 0.0008079242543317378, + 0.0009531081886962056, + 0.0006864856695756316, + 0.000968265114352107, + 0.0006543117924593389, + 0.00033673419966362417, + 0.00046435469994321465, + 0.00028591349837370217, + 0.0006573737482540309, + 0.001903911354020238, + 0.0018395694205537438, + 0.0004595117934513837, + 0.0004926584661006927, + 0.0003386415774002671, + 0.0003092717961408198, + 0.000384419021429494, + 0.0018972110701724887, + 0.0024491839576512575, + 0.0015880119754001498, + 0.0018883513985201716, + 0.0038038629572838545, + 0.0007179466192610562, + 0.0018713237950578332, + 0.0008128852350637317, + 0.000782729301135987, + 0.0011533602373674512, + 0.000937030476052314, + 0.00044589597382582724, + 0.0012730127200484276, + 0.0006501950556412339, + 0.0005390082369558513, + 0.0005626145284622908, + 0.001837060321122408, + 0.002085622865706682, + 0.001609809696674347, + 0.0009634417365305126, + 0.002435343572869897, + 0.0006404048181138933, + 0.0012971851974725723, + 0.0005438836524263024, + 0.0016475330339744687, + 0.0026565692387521267, + 0.00148411572445184, + 0.0015165625372901559, + 0.0021257263142615557, + 0.0017706066137179732, + 0.0010179984383285046, + 0.0010006062220782042, + 0.29835620522499084, + 0.053489092737436295, + 0.08153820037841797, + 0.01762998290359974, + 0.031558889895677567, + 0.3714398741722107, + 0.033665817230939865, + 0.03552928939461708, + 0.000497194065246731, + 0.000599205493927002, + 0.0005363755044527352, + 0.0005416263593360782, + 0.0005996051477268338, + 0.0007213196950033307, + 0.00044913371675647795, + 0.0006388281472027302, + 0.00032078527146950364, + 0.0006532068946398795, + 0.0004839196044486016, + 0.0003793995419982821, + 0.0004048985429108143, + 0.00031065961229614913, + 0.0005043919081799686, + 0.000344034779118374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.29760709404945374 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28137704730033875 + ], + "opposite_logprob_tokens": [ + -1.4063770771026611 + ], + "logprob_reward": -0.28137704730033875, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28137704730033875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021064165979623795, + 0.0030719172209501266, + 0.016804946586489677, + 0.013551536947488785, + 0.006734875962138176, + 0.03768976777791977, + 0.00741439126431942, + 0.874631404876709, + 0.011214389465749264, + 0.007822548970580101 + ], + "credit_map": [ + 0.0031591178849339485, + 0.005027334671467543, + 0.0036187192890793085, + 0.0015883763553574681, + 0.0028875337447971106, + 0.0017724150093272328, + 0.001007209881208837, + 0.0020034639164805412, + 0.0004195394867565483, + 0.00044856584281660616, + 0.00040979936602525413, + 0.0005213721306063235, + 0.00036117053241468966, + 0.00034213875187560916, + 0.0002545648312661797, + 0.0003147668030578643, + 0.004658310674130917, + 0.003789022332057357, + 0.003171333344653249, + 0.001461689011193812, + 0.0012773836497217417, + 0.000924118619877845, + 0.0006828905316069722, + 0.0008402004605159163, + 0.001680591027252376, + 0.001998847583308816, + 0.0014408996794372797, + 0.0015412654029205441, + 0.0034120085183531046, + 0.0009183952352032065, + 0.0016802483005449176, + 0.0008792835869826376, + 0.0009105346980504692, + 0.0011911765905097127, + 0.0012931829551234841, + 0.00046410883078351617, + 0.0010781108867377043, + 0.0005979318520985544, + 0.0005947579629719257, + 0.0006050730589777231, + 0.005911544896662235, + 0.005241937469691038, + 0.008419075980782509, + 0.0026267252396792173, + 0.0074874055571854115, + 0.0024661479983478785, + 0.003653845516964793, + 0.001883086166344583, + 0.001015136600472033, + 0.0011601054575294256, + 0.0011919812532141805, + 0.0007017920725047588, + 0.0011055396171286702, + 0.001241278019733727, + 0.0006292950129136443, + 0.00036926474422216415, + 0.28227245807647705, + 0.050395604223012924, + 0.077501080930233, + 0.01677740179002285, + 0.02996833249926567, + 0.3520662784576416, + 0.03187904879450798, + 0.03377126529812813, + 0.001123241032473743, + 0.0017539352411404252, + 0.0014089535688981414, + 0.0012490703957155347, + 0.0011829780414700508, + 0.0017919500824064016, + 0.0010905327508226037, + 0.0016137297498062253, + 0.0006916261627338827, + 0.0012808364117518067, + 0.001143821282312274, + 0.0006115156575106084, + 0.0010640280088409781, + 0.0010167418513447046, + 0.0009735297062434256, + 0.0010404506465420127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.28137704730033875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003340408205986 + ], + "opposite_logprob_tokens": [ + -2.3503339290618896 + ], + "logprob_reward": -0.1003340408205986, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003340408205986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006141688674688339, + 0.006072669755667448, + 0.007198132108896971, + 0.0032693392131477594, + 0.0016472331481054425, + 0.008903718553483486, + 0.005975617561489344, + 0.9540874361991882, + 0.0030328857246786356, + 0.0036712992005050182 + ], + "credit_map": [ + 0.0010593303013592958, + 0.0009306142455898225, + 0.0009078167495317757, + 0.0005558494594879448, + 0.0008983614388853312, + 0.0004779445589520037, + 0.0004605086287483573, + 0.0008512635831721127, + 0.0006899072322994471, + 0.0008457517833448946, + 0.0006049245712347329, + 0.0008460211683996022, + 0.0010911200661212206, + 0.0006466041668318212, + 0.0008606973569840193, + 0.0004876429447904229, + 0.0010717797558754683, + 0.0013401391915977001, + 0.0014307363890111446, + 0.0005307656829245389, + 0.0008239261223934591, + 0.000731343578081578, + 0.0005493845092132688, + 0.0007200564141385257, + 0.0007319857249967754, + 0.0008746953681111336, + 0.00024142571783158928, + 0.00023714991402812302, + 0.0004704563179984689, + 0.00018294871551916003, + 0.0003687510616146028, + 0.00016192632028833032, + 0.000242720459937118, + 0.0002454131899867207, + 0.0002356150944251567, + 0.0001355668209725991, + 0.0003063417971134186, + 0.00015564035857096314, + 0.0001777604193193838, + 0.00014817500778008252, + 0.00146652408875525, + 0.0015424886951223016, + 0.0013868504902347922, + 0.0006880737491883337, + 0.002043515909463167, + 0.0004854606231674552, + 0.0009038568823598325, + 0.0003869483480229974, + 0.0008162405574694276, + 0.0009717380744405091, + 0.0006784378201700747, + 0.0005414875922724605, + 0.001109420438297093, + 0.0007452183053828776, + 0.0005749636329710484, + 0.0005381114897318184, + 0.30816200375556946, + 0.05508176237344742, + 0.08419840782880783, + 0.01830853521823883, + 0.032773055136203766, + 0.3838377892971039, + 0.0348534993827343, + 0.03687240183353424, + 0.0004191659390926361, + 0.00031861814204603434, + 0.0005809510475955904, + 0.00020625957404263318, + 0.0002307882095919922, + 0.0005339375929906964, + 0.0002428176230750978, + 0.0005003472906537354, + 0.0003091899852734059, + 0.0005873896297998726, + 0.0005029101157560945, + 0.0007588986773043871, + 0.00033675075974315405, + 0.00020902564574498683, + 0.0004922557272948325, + 0.0004748785577248782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.1003340408205986 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160687193274498 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.0160687193274498, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160687193274498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045526016503572464, + 0.039734989404678345, + 0.0497296042740345, + 0.12770631909370422, + 0.045831505209207535, + 0.027228889986872673, + 0.017717817798256874, + 0.5477482080459595, + 0.03606706112623215, + 0.06270962953567505 + ], + "credit_map": [ + 0.0065784091129899025, + 0.00611306494101882, + 0.006941501051187515, + 0.0029250280931591988, + 0.013075103051960468, + 0.003526566317304969, + 0.002345690270885825, + 0.004020648542791605, + 0.0051507847383618355, + 0.006398309022188187, + 0.004581986926496029, + 0.004622113890945911, + 0.006042104214429855, + 0.004834363702684641, + 0.004323133267462254, + 0.0037821857258677483, + 0.0125811155885458, + 0.013383790850639343, + 0.00648852065205574, + 0.004066535271704197, + 0.004764836747199297, + 0.0034562298096716404, + 0.0022934440057724714, + 0.002695126924663782, + 0.01637238822877407, + 0.02200964279472828, + 0.010971824638545513, + 0.011569730006158352, + 0.03388448804616928, + 0.009976048022508621, + 0.015699509531259537, + 0.007222686428576708, + 0.0056797293946146965, + 0.007772278506308794, + 0.006152940448373556, + 0.0044400799088180065, + 0.008432486094534397, + 0.00457134610041976, + 0.004763402976095676, + 0.00401923805475235, + 0.004164260346442461, + 0.00403367867693305, + 0.007013795431703329, + 0.0020732867997139692, + 0.004211602732539177, + 0.002028187969699502, + 0.002062544459477067, + 0.0016415291465818882, + 0.0025544455274939537, + 0.0029283110052347183, + 0.002424046164378524, + 0.0022116105537861586, + 0.003245776519179344, + 0.0019401750760152936, + 0.001465965062379837, + 0.0009474873659200966, + 0.1779518574476242, + 0.03179458901286125, + 0.048861656337976456, + 0.010480355471372604, + 0.01843043975532055, + 0.21932227909564972, + 0.019856467843055725, + 0.02105049043893814, + 0.0060327365063130856, + 0.003178320126608014, + 0.009835404343903065, + 0.003366296412423253, + 0.00309739843942225, + 0.0050493283197283745, + 0.00249332282692194, + 0.0030142473988234997, + 0.00541027681902051, + 0.010330269113183022, + 0.010715777985751629, + 0.006843134760856628, + 0.004678856581449509, + 0.010426858440041542, + 0.006098905578255653, + 0.008205545134842396 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 9 + ], + "selected_reward": -0.0160687193274498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031898352317512035 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.0031898352317512035, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031898352317512035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.007626587990671396, + 0.003299364121630788, + 0.004175660666078329, + 0.007096572779119015, + 0.00691589992493391, + 0.004378415644168854, + 0.012649673968553543, + 0.9306409358978271, + 0.006100321654230356, + 0.017116444185376167 + ], + "credit_map": [ + 0.0011054225033149123, + 0.0019514033338055015, + 0.0012791281333193183, + 0.0005136775434948504, + 0.0008954430813901126, + 0.0005987606709823012, + 0.00038912906893529, + 0.0008936250233091414, + 0.00045714035513810813, + 0.0005638014408759773, + 0.00039145248592831194, + 0.0005377993802540004, + 0.0005045738653279841, + 0.00023893345496617258, + 0.0003810132038779557, + 0.00022465044457931072, + 0.000995697802864015, + 0.0010437406599521637, + 0.00043958512833341956, + 0.000332509953295812, + 0.00035901166847907007, + 0.0003809173940680921, + 0.00027364245033822954, + 0.00035055651096627116, + 0.0008772179717198014, + 0.001203522551804781, + 0.0006626836839132011, + 0.000762897776439786, + 0.0019411901012063026, + 0.00035929892328567803, + 0.0009232442243956029, + 0.00036651871050707996, + 0.000830296368803829, + 0.0011281644692644477, + 0.0010865102522075176, + 0.0005357107729651034, + 0.0009384571458213031, + 0.0009957081638276577, + 0.0006274700863286853, + 0.0007735839462839067, + 0.0007064088713377714, + 0.000769791950006038, + 0.0006820134585723281, + 0.0003298575757071376, + 0.0007220009574666619, + 0.0003723353729583323, + 0.0004996631178073585, + 0.00029634503880515695, + 0.0016624655108898878, + 0.0023883820977061987, + 0.001724658883176744, + 0.001569232321344316, + 0.002229897305369377, + 0.001417771796695888, + 0.0009380863630212843, + 0.0007191816112026572, + 0.30050843954086304, + 0.053729452192783356, + 0.08239423483610153, + 0.01796778291463852, + 0.032026633620262146, + 0.3740902841091156, + 0.03396763652563095, + 0.03595660999417305, + 0.0011092531494796276, + 0.0006674987962469459, + 0.0016500807832926512, + 0.0003200440260116011, + 0.000550776778254658, + 0.000679086078889668, + 0.0006550959078595042, + 0.0004684866580646485, + 0.0017546515446156263, + 0.003086959943175316, + 0.00259858975186944, + 0.0014195689000189304, + 0.0018576806178316474, + 0.0020714732818305492, + 0.0026313099078834057, + 0.0016962140798568726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0031898352317512035 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a dynamic close-up, capturing an unexpected scene at an Olympic-level indoor ice climbing arena: Sanji, dressed in a perfectly tailored black single-breasted suit, his signature short blond hair meticulously combed back, gleaming softly under the intense, cold white light cast by the arena's ceiling spotlights. With the precision and power of a ballerina, he swings his right arm. The metal tip of his professional ice axe traces a swift arc through the air, and with a sharp click, it drives decisively and deeply into the vertical, shimmering, artificial ice wall before him. At the moment of impact, countless tiny, shimmering ice shards explode toward the camera like diamonds, each one radiating a dazzling light in the strong sunlight. The frame is now sharply focused on these dancing ice crystals. Then, the camera executes a swift rack focus, the blurred ice shards instantly receding into the background. Focus is fixed sharply on his slightly raised face, capturing the steely, determined eyes beneath his signature curled brows. In that instant, the screen flashes rapidly, as if overloaded with energy, a flash of white light amplifying the moment's impact. The entire image, in its stark black, white, and blue tones, creates a surreal atmosphere, a unique blend of ultimate elegance and the intensity of extreme sports.", + "seed": 17055, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17055_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0891368580982089, + "vlm_logprob": -0.0891368580982089, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.0891368580982089, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17966796457767487 + ], + "opposite_logprob_tokens": [ + -1.9296679496765137 + ], + "logprob_reward": -0.17966796457767487, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.17966796457767487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.164781853556633, + 0.09286383539438248, + 0.026719754561781883, + 0.027598213404417038, + 0.03550899773836136, + 0.1259036660194397, + 0.15525197982788086, + 0.13949061930179596, + 0.09783831983804703, + 0.13404275476932526 + ], + "credit_map": [ + 0.015878139063715935, + 0.014215361326932907, + 0.00952569767832756, + 0.014687609858810902, + 0.020035145804286003, + 0.04014391824603081, + 0.017072021961212158, + 0.033223956823349, + 0.007182104978710413, + 0.010554881766438484, + 0.004633956588804722, + 0.003848777851089835, + 0.011653684079647064, + 0.028360020369291306, + 0.009168215095996857, + 0.017462190240621567, + 0.0021763755939900875, + 0.00320443045347929, + 0.0024455608800053596, + 0.0034618477802723646, + 0.004490765742957592, + 0.0040689934976398945, + 0.003340106224641204, + 0.00353167369030416, + 0.002591364551335573, + 0.005042394157499075, + 0.0035471487790346146, + 0.0029735160060226917, + 0.0029211232904344797, + 0.0030682911165058613, + 0.002754510845988989, + 0.004699863959103823, + 0.0050786626525223255, + 0.01004942785948515, + 0.004744702950119972, + 0.0026936677750200033, + 0.0057019735686481, + 0.002537087770178914, + 0.002110666362568736, + 0.0025928111281245947, + 0.007913504727184772, + 0.03186542168259621, + 0.015309741720557213, + 0.01364877074956894, + 0.03125372529029846, + 0.009123219177126884, + 0.010483749210834503, + 0.006305526942014694, + 0.011852972209453583, + 0.024244092404842377, + 0.012919862754642963, + 0.020918557420372963, + 0.03497064486145973, + 0.018709097057580948, + 0.01701364479959011, + 0.01462310180068016, + 0.01984984427690506, + 0.01861790008842945, + 0.016240915283560753, + 0.017262093722820282, + 0.012965763919055462, + 0.019593359902501106, + 0.019823389127850533, + 0.015137355774641037, + 0.010053861886262894, + 0.01262999139726162, + 0.008239454589784145, + 0.008391361683607101, + 0.015533577650785446, + 0.017045317217707634, + 0.015306172892451286, + 0.010638578794896603, + 0.012664726935327053, + 0.027863210067152977, + 0.007443126291036606, + 0.02056671492755413, + 0.021733514964580536, + 0.016463803127408028, + 0.0149813462048769, + 0.01232630480080843 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.17966796457767487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22570279240608215 + ], + "opposite_logprob_tokens": [ + -1.6007027626037598 + ], + "logprob_reward": -0.22570279240608215, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22570279240608215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14070197939872742, + 0.045639295130968094, + 0.0863458514213562, + 0.0425126813352108, + 0.1324429214000702, + 0.05793292075395584, + 0.16603845357894897, + 0.046731315553188324, + 0.254670649766922, + 0.026984035968780518 + ], + "credit_map": [ + 0.010838816873729229, + 0.015005108900368214, + 0.006108499132096767, + 0.006892693229019642, + 0.023281080648303032, + 0.023465828970074654, + 0.02273941971361637, + 0.03237052261829376, + 0.0027446227613836527, + 0.0055926088243722916, + 0.0018212390132248402, + 0.00168980925809592, + 0.008679404854774475, + 0.008947712369263172, + 0.00709279952570796, + 0.00907109770923853, + 0.004620118532329798, + 0.007498323451727629, + 0.005953456740826368, + 0.005463247187435627, + 0.02038384974002838, + 0.016222143545746803, + 0.010951622389256954, + 0.015253077261149883, + 0.004471748601645231, + 0.005738500040024519, + 0.005078048445284367, + 0.003935322165489197, + 0.005781564395874739, + 0.005086055025458336, + 0.004021030850708485, + 0.008400407619774342, + 0.010078908875584602, + 0.026671260595321655, + 0.02320215106010437, + 0.011943729594349861, + 0.026902247220277786, + 0.013395353220403194, + 0.009749618358910084, + 0.010499631986021996, + 0.0042666103690862656, + 0.014435229822993279, + 0.006312964949756861, + 0.006256023421883583, + 0.014597522094845772, + 0.004138575866818428, + 0.005010726395994425, + 0.0029152596835047007, + 0.014982063323259354, + 0.02857886254787445, + 0.013344586826860905, + 0.012803197838366032, + 0.0569830946624279, + 0.01605084538459778, + 0.013691743835806847, + 0.009604036808013916, + 0.006329298485070467, + 0.005426310934126377, + 0.004368157591670752, + 0.00788917951285839, + 0.004689469467848539, + 0.00705215847119689, + 0.005863511003553867, + 0.005113224498927593, + 0.015077036805450916, + 0.016217978671193123, + 0.010124501772224903, + 0.012951619923114777, + 0.05952663719654083, + 0.05873974785208702, + 0.05407490208745003, + 0.027958199381828308, + 0.0023826148826628923, + 0.0044531566090881824, + 0.001285435981117189, + 0.0039009610190987587, + 0.0042741186916828156, + 0.004389934241771698, + 0.0028054662980139256, + 0.0034923471976071596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.22570279240608215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020600641146302223 + ], + "opposite_logprob_tokens": [ + -3.8956005573272705 + ], + "logprob_reward": -0.020600641146302223, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020600641146302223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10802555084228516, + 0.18724611401557922, + 0.0568164698779583, + 0.05537280812859535, + 0.07915674149990082, + 0.08766605705022812, + 0.1718878448009491, + 0.0769721046090126, + 0.09390074014663696, + 0.08295554667711258 + ], + "credit_map": [ + 0.007193252444267273, + 0.012403707951307297, + 0.00635466119274497, + 0.004937783349305391, + 0.01784457638859749, + 0.017444312572479248, + 0.01980244740843773, + 0.022044816985726357, + 0.013675425201654434, + 0.02393212355673313, + 0.008401389233767986, + 0.007729292381554842, + 0.03226730227470398, + 0.035024672746658325, + 0.02632232941687107, + 0.03989357128739357, + 0.0050346399657428265, + 0.006467149127274752, + 0.005615039728581905, + 0.005849928129464388, + 0.012013299390673637, + 0.008349607698619366, + 0.006732959300279617, + 0.006753846071660519, + 0.005424373783171177, + 0.008715755306184292, + 0.008640371263027191, + 0.005399156827479601, + 0.007093602791428566, + 0.006715418770909309, + 0.005360228009521961, + 0.008023897185921669, + 0.0063162436708807945, + 0.008618459105491638, + 0.010553543455898762, + 0.007378275506198406, + 0.018573731184005737, + 0.00880555622279644, + 0.009486589580774307, + 0.009424342773854733, + 0.005863395985215902, + 0.015869932249188423, + 0.01241967175155878, + 0.013350842520594597, + 0.01950543187558651, + 0.008280145935714245, + 0.007414595223963261, + 0.004962041974067688, + 0.013248506002128124, + 0.030039429664611816, + 0.01192285493016243, + 0.016661610454320908, + 0.053987786173820496, + 0.018136050552129745, + 0.016085462644696236, + 0.011806154623627663, + 0.008929459378123283, + 0.009957983158528805, + 0.008665554225444794, + 0.010150670073926449, + 0.006845872383564711, + 0.010369041003286839, + 0.012238916009664536, + 0.009814612567424774, + 0.005539951846003532, + 0.006636408157646656, + 0.004685757216066122, + 0.005840204656124115, + 0.02409888617694378, + 0.019473299384117126, + 0.018392549827694893, + 0.00923367589712143, + 0.010208013467490673, + 0.015428067184984684, + 0.0042988755740225315, + 0.009790739044547081, + 0.01028517447412014, + 0.012591914273798466, + 0.0060371533036231995, + 0.014315605163574219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.020600641146302223 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01107780635356903 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.01107780635356903, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01107780635356903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08448164910078049, + 0.09140592813491821, + 0.04068379104137421, + 0.03963353857398033, + 0.07435828447341919, + 0.2023683488368988, + 0.18478098511695862, + 0.17651790380477905, + 0.05159003287553787, + 0.05417952314019203 + ], + "credit_map": [ + 0.0072298841550946236, + 0.00862080417573452, + 0.0031761135905981064, + 0.004594907164573669, + 0.010353952646255493, + 0.017981797456741333, + 0.01238301582634449, + 0.020141176879405975, + 0.004463678225874901, + 0.014269358478486538, + 0.003358968999236822, + 0.004753931425511837, + 0.017397569492459297, + 0.019763875752687454, + 0.015575280413031578, + 0.01182327326387167, + 0.0033850481268018484, + 0.005497165024280548, + 0.002705160528421402, + 0.0043862988241016865, + 0.00738245714455843, + 0.0064945886842906475, + 0.005107247270643711, + 0.005725831724703312, + 0.0036595240235328674, + 0.005470495205372572, + 0.007042557001113892, + 0.004014861304312944, + 0.003921442665159702, + 0.004954241216182709, + 0.0056284088641405106, + 0.004942008759826422, + 0.008426040410995483, + 0.016132129356265068, + 0.011270326562225819, + 0.00541573716327548, + 0.01299977209419012, + 0.008729969151318073, + 0.004180742893368006, + 0.007203570567071438, + 0.013145367614924908, + 0.055608902126550674, + 0.02165900357067585, + 0.018788499757647514, + 0.053463179618120193, + 0.014166978187859058, + 0.016003694385290146, + 0.009532729163765907, + 0.017758868634700775, + 0.02614474669098854, + 0.01907251589000225, + 0.025093210861086845, + 0.0310968067497015, + 0.023718727752566338, + 0.02185097150504589, + 0.020045151934027672, + 0.021904604509472847, + 0.02257986180484295, + 0.018444232642650604, + 0.026485687121748924, + 0.017582183703780174, + 0.02461051568388939, + 0.026115193963050842, + 0.018795639276504517, + 0.005250297486782074, + 0.006689455360174179, + 0.003443598048761487, + 0.003655589185655117, + 0.010698557831346989, + 0.00760618457570672, + 0.009429484605789185, + 0.004816870670765638, + 0.0041397688910365105, + 0.012445337139070034, + 0.0020552920177578926, + 0.005566959269344807, + 0.00988946482539177, + 0.0069624269381165504, + 0.006131380796432495, + 0.006988896988332272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.01107780635356903 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635086007416248 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635086007416248, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635086007416248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05573006719350815, + 0.05841619148850441, + 0.026901375502347946, + 0.06641897559165955, + 0.06922027468681335, + 0.14051203429698944, + 0.19068574905395508, + 0.13859416544437408, + 0.10644228011369705, + 0.14707887172698975 + ], + "credit_map": [ + 0.00900212675333023, + 0.006017843261361122, + 0.002883095061406493, + 0.003654477884992957, + 0.009260217659175396, + 0.008076689206063747, + 0.0069254618138074875, + 0.009910156019032001, + 0.0051493290811777115, + 0.006319267675280571, + 0.002657183911651373, + 0.004594298545271158, + 0.008032044395804405, + 0.015430991537868977, + 0.006408803164958954, + 0.009824272245168686, + 0.001955057494342327, + 0.0031593292951583862, + 0.0019706636667251587, + 0.002490101382136345, + 0.006171792279928923, + 0.004374421201646328, + 0.003667096607387066, + 0.003112913342192769, + 0.006119104567915201, + 0.008664530701935291, + 0.010151748545467854, + 0.006514106877148151, + 0.00863329041749239, + 0.009319176897406578, + 0.005491030868142843, + 0.011525988578796387, + 0.009869548492133617, + 0.01872960478067398, + 0.009911930188536644, + 0.003654635278508067, + 0.013457109220325947, + 0.005803389009088278, + 0.0034899027086794376, + 0.004304149653762579, + 0.009966230019927025, + 0.027706876397132874, + 0.019311215728521347, + 0.018271449953317642, + 0.032233305275440216, + 0.011757849715650082, + 0.013208405114710331, + 0.008056698366999626, + 0.019302943721413612, + 0.02819627709686756, + 0.019162597134709358, + 0.014045752584934235, + 0.05883822590112686, + 0.0195420291274786, + 0.016596093773841858, + 0.015001832507550716, + 0.017448658123612404, + 0.018105652183294296, + 0.014327986165881157, + 0.0159470122307539, + 0.01315246894955635, + 0.022138988599181175, + 0.021851984784007072, + 0.015621436759829521, + 0.009983427822589874, + 0.013505813665688038, + 0.005162757355719805, + 0.007753497455269098, + 0.019551312550902367, + 0.01884079910814762, + 0.021422697231173515, + 0.010221971198916435, + 0.012228811159729958, + 0.028568293899297714, + 0.005919466260820627, + 0.015463335439562798, + 0.02097621001303196, + 0.02314290963113308, + 0.013625695370137691, + 0.02715415135025978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.008635086007416248 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a dynamic close-up, capturing an unexpected scene at an Olympic-level indoor ice climbing arena: Sanji, dressed in a perfectly tailored black single-breasted suit, his signature short blond hair meticulously combed back, gleaming softly under the intense, cold white light cast by the arena's ceiling spotlights. With the precision and power of a ballerina, he swings his right arm. The metal tip of his professional ice axe traces a swift arc through the air, and with a sharp click, it drives decisively and deeply into the vertical, shimmering, artificial ice wall before him. At the moment of impact, countless tiny, shimmering ice shards explode toward the camera like diamonds, each one radiating a dazzling light in the strong sunlight. The frame is now sharply focused on these dancing ice crystals. Then, the camera executes a swift rack focus, the blurred ice shards instantly receding into the background. Focus is fixed sharply on his slightly raised face, capturing the steely, determined eyes beneath his signature curled brows. In that instant, the screen flashes rapidly, as if overloaded with energy, a flash of white light amplifying the moment's impact. The entire image, in its stark black, white, and blue tones, creates a surreal atmosphere, a unique blend of ultimate elegance and the intensity of extreme sports.", + "seed": 17056, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_10/index_3532_seed_17056_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04843007996678352, + "vlm_logprob": -0.04843007996678352, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04843007996678352, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.173772931098938 + ], + "opposite_logprob_tokens": [ + -1.923772931098938 + ], + "logprob_reward": -0.173772931098938, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.173772931098938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13186146318912506, + 0.1642332226037979, + 0.21112988889217377, + 0.05914675071835518, + 0.0816456601023674, + 0.05665616691112518, + 0.13061870634555817, + 0.0497087761759758, + 0.06133665889501572, + 0.05366269871592522 + ], + "credit_map": [ + 0.009487750940024853, + 0.01872277818620205, + 0.009248549118638039, + 0.009294564835727215, + 0.026602143421769142, + 0.02852702885866165, + 0.017414752393960953, + 0.012563895434141159, + 0.01569671928882599, + 0.012918609194457531, + 0.010504689998924732, + 0.011438698507845402, + 0.03309902548789978, + 0.02890399657189846, + 0.03539252281188965, + 0.01627897284924984, + 0.009332315064966679, + 0.008497466333210468, + 0.010095655918121338, + 0.007732951547950506, + 0.020110808312892914, + 0.07344106584787369, + 0.01504792645573616, + 0.06687170267105103, + 0.007971393875777721, + 0.0072119710966944695, + 0.0070647201500833035, + 0.006231098901480436, + 0.008262556977570057, + 0.007617034018039703, + 0.005847490858286619, + 0.008940489031374454, + 0.011210208758711815, + 0.01221091765910387, + 0.020533006638288498, + 0.010338653810322285, + 0.010470480658113956, + 0.004408517852425575, + 0.007865042425692081, + 0.0046088299714028835, + 0.005784442648291588, + 0.00839981622993946, + 0.0075599015690386295, + 0.008553697727620602, + 0.006467464845627546, + 0.0073514217510819435, + 0.006254331674426794, + 0.006285092327743769, + 0.012876009568572044, + 0.02024644985795021, + 0.017977923154830933, + 0.025197671726346016, + 0.01664249412715435, + 0.012151351198554039, + 0.01633431576192379, + 0.00919248815625906, + 0.004550717771053314, + 0.006922935135662556, + 0.005566352512687445, + 0.00721852807328105, + 0.005175032652914524, + 0.005755930207669735, + 0.008738470263779163, + 0.005780808627605438, + 0.006213693413883448, + 0.005353440530598164, + 0.009552905336022377, + 0.008232081308960915, + 0.0101750697940588, + 0.0072732167318463326, + 0.008682139217853546, + 0.00585411349311471, + 0.006121376529335976, + 0.007078071124851704, + 0.0062463716603815556, + 0.007087466772645712, + 0.009221740067005157, + 0.0032441788353025913, + 0.008679280057549477, + 0.005984210409224033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.173772931098938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02982458658516407 + ], + "opposite_logprob_tokens": [ + -3.5298244953155518 + ], + "logprob_reward": -0.02982458658516407, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02982458658516407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10789437592029572, + 0.12894238531589508, + 0.08115635812282562, + 0.14802582561969757, + 0.09518507122993469, + 0.1086849719285965, + 0.04840466380119324, + 0.14585387706756592, + 0.08617804944515228, + 0.04967447370290756 + ], + "credit_map": [ + 0.008067024871706963, + 0.010572674684226513, + 0.009811258874833584, + 0.014251030050218105, + 0.019385281950235367, + 0.018925128504633904, + 0.016594601795077324, + 0.01028736587613821, + 0.018495982512831688, + 0.019330576062202454, + 0.010401593521237373, + 0.009896081872284412, + 0.020141957327723503, + 0.018793854862451553, + 0.017040342092514038, + 0.014841991476714611, + 0.00523808877915144, + 0.003440410364419222, + 0.005912729073315859, + 0.003750209929421544, + 0.011742349714040756, + 0.02158244512975216, + 0.011936233378946781, + 0.01755388453602791, + 0.02472914196550846, + 0.01673441007733345, + 0.01265040785074234, + 0.012255271896719933, + 0.016665970906615257, + 0.022763261571526527, + 0.01705184392631054, + 0.02517550252377987, + 0.012529871426522732, + 0.01520558726042509, + 0.013443893752992153, + 0.00862092524766922, + 0.01766996644437313, + 0.010169324465095997, + 0.011213727295398712, + 0.006331761833280325, + 0.013851706869900227, + 0.010403590276837349, + 0.0131528964266181, + 0.012491855770349503, + 0.022112727165222168, + 0.01629089005291462, + 0.011644260957837105, + 0.008737029507756233, + 0.004999976139515638, + 0.004879182670265436, + 0.007643392775207758, + 0.004607091657817364, + 0.0060832626186311245, + 0.004834472201764584, + 0.005403180606663227, + 0.009954102337360382, + 0.008375362493097782, + 0.010645284317433834, + 0.00925503671169281, + 0.019680626690387726, + 0.012852446176111698, + 0.033828817307949066, + 0.03259788826107979, + 0.018618397414684296, + 0.00951302982866764, + 0.007985449396073818, + 0.00969571154564619, + 0.009444857016205788, + 0.015286913141608238, + 0.006473447196185589, + 0.019360948354005814, + 0.008417678065598011, + 0.007328883744776249, + 0.0035739089362323284, + 0.006384801119565964, + 0.005255464464426041, + 0.008350105956196785, + 0.0027485210448503494, + 0.011465080082416534, + 0.004567703232169151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.02982458658516407 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02330007031559944 + ], + "opposite_logprob_tokens": [ + -3.7733001708984375 + ], + "logprob_reward": -0.02330007031559944, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02330007031559944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14047382771968842, + 0.2110886126756668, + 0.10160694271326065, + 0.063432477414608, + 0.10845240205526352, + 0.05950184166431427, + 0.07708492875099182, + 0.11042893677949905, + 0.04612612724304199, + 0.08180393278598785 + ], + "credit_map": [ + 0.008984231390058994, + 0.01938752643764019, + 0.009802966378629208, + 0.011880925856530666, + 0.032700106501579285, + 0.024661799892783165, + 0.020223846659064293, + 0.01283240132033825, + 0.016428595408797264, + 0.018750648945569992, + 0.014054068364202976, + 0.019722426310181618, + 0.04358304664492607, + 0.026846520602703094, + 0.04780038818717003, + 0.0239028949290514, + 0.005386820062994957, + 0.004696834832429886, + 0.006095889024436474, + 0.005921287927776575, + 0.013468642719089985, + 0.0314662903547287, + 0.007497703190892935, + 0.027073465287685394, + 0.006858335342258215, + 0.0062135374173521996, + 0.007755032274872065, + 0.01018164400011301, + 0.007840264588594437, + 0.007704396732151508, + 0.009829471819102764, + 0.007049790117889643, + 0.013069012202322483, + 0.014441460371017456, + 0.013621014542877674, + 0.02039227820932865, + 0.017791714519262314, + 0.006948630791157484, + 0.014519395306706429, + 0.007668884005397558, + 0.0061050658114254475, + 0.007234972435981035, + 0.006375104188919067, + 0.009843999519944191, + 0.007138851564377546, + 0.007179108913987875, + 0.009371946565806866, + 0.006252788007259369, + 0.00550862168893218, + 0.011580218560993671, + 0.00949755497276783, + 0.01720939762890339, + 0.009993809275329113, + 0.007077650632709265, + 0.010366593487560749, + 0.005851073190569878, + 0.005290734115988016, + 0.008417497389018536, + 0.004852753598242998, + 0.018669884651899338, + 0.010731251910328865, + 0.012110409326851368, + 0.03721144422888756, + 0.013144944794476032, + 0.005500925704836845, + 0.00432234350591898, + 0.006653775926679373, + 0.0038441894575953484, + 0.007745443843305111, + 0.004709142260253429, + 0.008001684211194515, + 0.005348618142306805, + 0.007462857756763697, + 0.010156744159758091, + 0.010545235127210617, + 0.010494574904441833, + 0.013226985931396484, + 0.006342349108308554, + 0.013649559579789639, + 0.009925615973770618 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.02330007031559944 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007628594525158405 + ], + "opposite_logprob_tokens": [ + -4.882628440856934 + ], + "logprob_reward": -0.007628594525158405, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007628594525158405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08226461708545685, + 0.11632426828145981, + 0.23204292356967926, + 0.08519550412893295, + 0.14212988317012787, + 0.0909685418009758, + 0.07207431644201279, + 0.06779203563928604, + 0.03449021652340889, + 0.07671767473220825 + ], + "credit_map": [ + 0.006004017312079668, + 0.007299880962818861, + 0.009154066443443298, + 0.005879045929759741, + 0.020777013152837753, + 0.016309775412082672, + 0.009258471429347992, + 0.007582355290651321, + 0.01593012735247612, + 0.010595247149467468, + 0.012651787139475346, + 0.009074758738279343, + 0.019934555515646935, + 0.021045662462711334, + 0.014110618270933628, + 0.012981517240405083, + 0.012403546832501888, + 0.015641910955309868, + 0.013638634234666824, + 0.007863079197704792, + 0.03202630579471588, + 0.06310690939426422, + 0.023966297507286072, + 0.06339625269174576, + 0.011355815455317497, + 0.009786977432668209, + 0.008127759210765362, + 0.008420919999480247, + 0.01214584056288004, + 0.012410481460392475, + 0.01063492987304926, + 0.012312781065702438, + 0.016576504334807396, + 0.018948640674352646, + 0.01841343566775322, + 0.02570069581270218, + 0.021473100408911705, + 0.013410437852144241, + 0.017830105498433113, + 0.009776960127055645, + 0.010112842544913292, + 0.010047350078821182, + 0.00984708871692419, + 0.009467832744121552, + 0.016203973442316055, + 0.016930285841226578, + 0.011915958486497402, + 0.006443215999752283, + 0.00925643090158701, + 0.011603377759456635, + 0.008218985982239246, + 0.00987849198281765, + 0.006778253708034754, + 0.007269131485372782, + 0.007381153758615255, + 0.011688489466905594, + 0.004423747304826975, + 0.00811200961470604, + 0.0040633948519825935, + 0.009094072505831718, + 0.008170022629201412, + 0.011627237312495708, + 0.014779210090637207, + 0.007522343657910824, + 0.0035626450553536415, + 0.003103113267570734, + 0.0032711594831198454, + 0.0044949534349143505, + 0.006868399679660797, + 0.002966729924082756, + 0.005336453206837177, + 0.004886763170361519, + 0.010825357399880886, + 0.007179871201515198, + 0.007031155284494162, + 0.00837237760424614, + 0.01588798686861992, + 0.003640575334429741, + 0.0176360122859478, + 0.006144342944025993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.007628594525158405 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a character named Sanji wearing a black single-breasted suit with short blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting an indoor ice climbing arena with a vertical artificial ice wall?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character swing a professional ice axe to drive it into the ice wall, causing ice shards to explode?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera execute a swift rack focus from the exploding ice shards to the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final shot a close-up of the character's face showing steely eyes and curled brows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076242173090577126 + ], + "opposite_logprob_tokens": [ + -4.88262414932251 + ], + "logprob_reward": -0.0076242173090577126, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076242173090577126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1001536175608635, + 0.06710035353899002, + 0.07987908273935318, + 0.05496080592274666, + 0.06219838559627533, + 0.08104643225669861, + 0.13346783816814423, + 0.16290801763534546, + 0.12279734760522842, + 0.13548819720745087 + ], + "credit_map": [ + 0.007110784761607647, + 0.014784620143473148, + 0.007795743178576231, + 0.009494495578110218, + 0.019602563232183456, + 0.019069908186793327, + 0.014453326351940632, + 0.007842163555324078, + 0.006586065981537104, + 0.007868465036153793, + 0.005150409881025553, + 0.005277291871607304, + 0.014293879270553589, + 0.01083446852862835, + 0.009043162688612938, + 0.008046608418226242, + 0.0032960062380880117, + 0.006133547984063625, + 0.0039057116955518723, + 0.0029414736200124025, + 0.012947797775268555, + 0.018140636384487152, + 0.010874630883336067, + 0.021639268845319748, + 0.009784722700715065, + 0.008056930266320705, + 0.006224333308637142, + 0.00492847291752696, + 0.006529692560434341, + 0.00726350536569953, + 0.004952142480760813, + 0.007221000734716654, + 0.006341905798763037, + 0.009397150948643684, + 0.008607296273112297, + 0.00860282126814127, + 0.010976656340062618, + 0.006586800329387188, + 0.006617709528654814, + 0.0050680371932685375, + 0.008945932611823082, + 0.009198448620736599, + 0.013892855495214462, + 0.012974428944289684, + 0.00983678549528122, + 0.010144981555640697, + 0.010997450910508633, + 0.005055542103946209, + 0.013476172462105751, + 0.017693480476737022, + 0.019833238795399666, + 0.02331405133008957, + 0.015393652953207493, + 0.014710888266563416, + 0.016972122713923454, + 0.012074222788214684, + 0.008195143193006516, + 0.012526505626738071, + 0.009493040852248669, + 0.03674343228340149, + 0.014326800592243671, + 0.014160675927996635, + 0.053575560450553894, + 0.013886839151382446, + 0.015673359856009483, + 0.011585704982280731, + 0.02256547100841999, + 0.01082431711256504, + 0.01649671234190464, + 0.010175218805670738, + 0.019564861431717873, + 0.015911700204014778, + 0.014457420445978642, + 0.014866768382489681, + 0.012827605940401554, + 0.015102111734449863, + 0.02961825393140316, + 0.008774630725383759, + 0.028174256905913353, + 0.01166712585836649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 6 + ], + "selected_reward": -0.0076242173090577126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_tensor([8562, 8562])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_tensor([8562, 8562])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a08228c5472cc9ae071931cabc5ca1f6bfee3c0b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_tensor([8562, 8562])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The camera begins with a striking overhead shot, focusing on a lion in its prime, standing majestically on the edge of a rugged, rocky cliff. Its thick, dark brown and golden mane, like a crown of kings, sways gently in the evening breeze. Its body, covered in short, sandy-brown fur, is defined by its muscularity and strength, its limbs firmly planted on the rock. Facing the sky, it leans forward slightly, its maw suddenly opening in a silent yet earth-shattering roar. Its sharp, gleaming white canines and nostrils flaring with effort are clearly visible. Its chest heaves violently, a stream of condensed hot air erupts from its mouth, and its eyes are filled with sharpness and majesty. The backdrop is the vast African savannah, the sky stained in a magnificent gradation of orange, red, and deep purple by the setting sun. Warm, dramatic golden sidelights outline the lion's majestic silhouette, casting a long shadow behind it. As this shocking scene unfolds, the camera begins to rise steadily and slowly, gradually incorporating more of the sky and the distant horizon into the picture, further emphasizing the lion's loneliness and supreme status in the vast world. The whole picture is full of primitive, magnificent and epic atmosphere of absolute dominance.", + "seed": 20160, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_8562_seed_20160_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13899722155183553, + "vlm_logprob": -0.13899722155183553, + "vlm_yes_no_margin": 2.674999976158142, + "vlm_token_credit": -0.13899722155183553, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017814982682466507 + ], + "opposite_logprob_tokens": [ + -4.517815113067627 + ], + "logprob_reward": -0.017814982682466507, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017814982682466507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2357587218284607, + 0.03570881485939026, + 0.058872997760772705, + 0.07929173856973648, + 0.03459456190466881, + 0.018552394583821297, + 0.019330015406012535, + 0.02144065871834755, + 0.10753268748521805, + 0.3889175057411194 + ], + "credit_map": [ + 0.02115125209093094, + 0.008910473436117172, + 0.013834510929882526, + 0.013846684247255325, + 0.031024010851979256, + 0.08818479627370834, + 0.014682440087199211, + 0.04412451758980751, + 0.004484835546463728, + 0.004902165848761797, + 0.0036131690721958876, + 0.0045247757807374, + 0.004721707198768854, + 0.0035177061799913645, + 0.004663269501179457, + 0.005281179212033749, + 0.004044216126203537, + 0.006164643447846174, + 0.0071861413307487965, + 0.006594231352210045, + 0.009676932357251644, + 0.005419973284006119, + 0.013198690488934517, + 0.006588163319975138, + 0.00818838831037283, + 0.010046734474599361, + 0.007860655896365643, + 0.009680517017841339, + 0.01327925268560648, + 0.006921784020960331, + 0.011829549446702003, + 0.011484850198030472, + 0.0032913913019001484, + 0.003774033859372139, + 0.005085748620331287, + 0.0045071933418512344, + 0.00605436135083437, + 0.003382255556061864, + 0.004805636592209339, + 0.003693938720971346, + 0.0016187317669391632, + 0.00254940171726048, + 0.002029132330790162, + 0.0032561030238866806, + 0.0033306560944765806, + 0.002012804616242647, + 0.0016885255463421345, + 0.002067037159577012, + 0.0023808039259165525, + 0.0036471723578870296, + 0.0022371860686689615, + 0.001840711571276188, + 0.0034814723767340183, + 0.0018819533288478851, + 0.0021874315571039915, + 0.0016732817748561502, + 0.00285403267480433, + 0.0023639616556465626, + 0.0021098325960338116, + 0.002715750364586711, + 0.00408084224909544, + 0.003105435287579894, + 0.001967379590496421, + 0.0022434222046285868, + 0.01616915501654148, + 0.003662385744974017, + 0.004214676562696695, + 0.003734639147296548, + 0.004597669467329979, + 0.03984004631638527, + 0.005779578350484371, + 0.029534529894590378, + 0.06041489541530609, + 0.008136512711644173, + 0.013868619687855244, + 0.006991813890635967, + 0.012257936410605907, + 0.15739156305789948, + 0.0154325682669878, + 0.11442357301712036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.017814982682466507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560548782348633 + ], + "opposite_logprob_tokens": [ + -1.6006054878234863 + ], + "logprob_reward": -0.22560548782348633, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560548782348633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0885498970746994, + 0.05511440709233284, + 0.04613765701651573, + 0.2459179162979126, + 0.2437518686056137, + 0.012989458627998829, + 0.008914333768188953, + 0.01341056078672409, + 0.05319269374012947, + 0.2320212423801422 + ], + "credit_map": [ + 0.009247832000255585, + 0.005863053258508444, + 0.006860516034066677, + 0.0058425781317055225, + 0.01035127229988575, + 0.027205953374505043, + 0.00853704009205103, + 0.01464165560901165, + 0.008371236734092236, + 0.006228650454431772, + 0.00674467533826828, + 0.004812998231500387, + 0.007988300174474716, + 0.00628620246425271, + 0.006953921634703875, + 0.007728420197963715, + 0.005710314027965069, + 0.004372882191091776, + 0.00493222801014781, + 0.004939130041748285, + 0.006289576645940542, + 0.004123196471482515, + 0.009681159630417824, + 0.0060891699977219105, + 0.019374672323465347, + 0.020669294521212578, + 0.019483555108308792, + 0.029298920184373856, + 0.02591422200202942, + 0.024066992104053497, + 0.07158041000366211, + 0.03552985563874245, + 0.036633942276239395, + 0.016354035586118698, + 0.03166685998439789, + 0.014923266135156155, + 0.02822064608335495, + 0.029962560161948204, + 0.024335786700248718, + 0.061654772609472275, + 0.0016716009704396129, + 0.0014296460431069136, + 0.0019972878508269787, + 0.001870020991191268, + 0.0013783667236566544, + 0.0016438207821920514, + 0.0013107158010825515, + 0.001687999116256833, + 0.001180989551357925, + 0.000980899203568697, + 0.0015298881335183978, + 0.001302859978750348, + 0.0014470316236838698, + 0.000927593617234379, + 0.0006892012315802276, + 0.00085587048670277, + 0.0018705122638493776, + 0.001156720332801342, + 0.0014893088955432177, + 0.0011378417257219553, + 0.0017025513807311654, + 0.002532282145693898, + 0.0012577672023326159, + 0.002263576490804553, + 0.006903068628162146, + 0.004166742786765099, + 0.002834862796589732, + 0.002680485835298896, + 0.0035926029086112976, + 0.016085844486951828, + 0.004168609622865915, + 0.012760479003190994, + 0.03594109043478966, + 0.004988294094800949, + 0.008685540407896042, + 0.003252737456932664, + 0.005756688304245472, + 0.09446169435977936, + 0.009084849618375301, + 0.06985032558441162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.22560548782348633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705309689044952 + ], + "opposite_logprob_tokens": [ + -2.1270530223846436 + ], + "logprob_reward": -0.12705309689044952, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12705309689044952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13760767877101898, + 0.11790713667869568, + 0.26856833696365356, + 0.08378682285547256, + 0.07945532351732254, + 0.03782118484377861, + 0.021584829315543175, + 0.06387881934642792, + 0.03253796324133873, + 0.15685179829597473 + ], + "credit_map": [ + 0.019218042492866516, + 0.015653027221560478, + 0.015417784452438354, + 0.018351934850215912, + 0.014723498374223709, + 0.017034193500876427, + 0.020322881639003754, + 0.016886325553059578, + 0.010121095925569534, + 0.014226925559341908, + 0.013276428915560246, + 0.013934599235653877, + 0.018350206315517426, + 0.013377232477068901, + 0.02058786153793335, + 0.014032792299985886, + 0.02105369232594967, + 0.02860403247177601, + 0.02798624336719513, + 0.040743838995695114, + 0.04751009866595268, + 0.026476910337805748, + 0.04566188156604767, + 0.030531687662005424, + 0.006550135090947151, + 0.012303462252020836, + 0.008905473165214062, + 0.010187088511884212, + 0.01924549601972103, + 0.007556935306638479, + 0.010747154243290424, + 0.008291084319353104, + 0.00883709080517292, + 0.008593075908720493, + 0.014279821887612343, + 0.010789746418595314, + 0.009642213582992554, + 0.007697635795921087, + 0.007927619852125645, + 0.011688131839036942, + 0.004598051775246859, + 0.0036871270276606083, + 0.006094856653362513, + 0.006383795291185379, + 0.005317924078553915, + 0.0036942155566066504, + 0.003751874202862382, + 0.004293343052268028, + 0.0023887131828814745, + 0.0038014801684767008, + 0.0023148651234805584, + 0.002439802745357156, + 0.004046192858368158, + 0.00207443512044847, + 0.0023922654800117016, + 0.0021270769648253918, + 0.007303926628082991, + 0.008100724779069424, + 0.006032600067555904, + 0.00782701838761568, + 0.013451946899294853, + 0.008255211636424065, + 0.006833234336227179, + 0.006074158940464258, + 0.005943329073488712, + 0.0023146886378526688, + 0.003994612954556942, + 0.002553089987486601, + 0.0025376516859978437, + 0.008686685934662819, + 0.002408545697107911, + 0.004099363461136818, + 0.024013066664338112, + 0.0049301511608064175, + 0.006284062750637531, + 0.0075529227033257484, + 0.007041800767183304, + 0.05863075703382492, + 0.00768536189571023, + 0.04071370139718056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.12705309689044952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073916219174862 + ], + "opposite_logprob_tokens": [ + -4.511074066162109 + ], + "logprob_reward": -0.011073916219174862, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073916219174862, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29247960448265076, + 0.0358460433781147, + 0.06494013965129852, + 0.05972862243652344, + 0.10492175817489624, + 0.020785732194781303, + 0.006529372651129961, + 0.016893386840820312, + 0.016064723953604698, + 0.38181060552597046 + ], + "credit_map": [ + 0.026066578924655914, + 0.009299256838858128, + 0.008719300851225853, + 0.010703613050282001, + 0.04288068413734436, + 0.12319258600473404, + 0.013387246988713741, + 0.05823035165667534, + 0.004246735945343971, + 0.004261984024196863, + 0.003795187920331955, + 0.0029162047430872917, + 0.006260336842387915, + 0.003873081412166357, + 0.005590986460447311, + 0.0049015250988304615, + 0.0055127558298408985, + 0.006822315976023674, + 0.007161173038184643, + 0.009645077399909496, + 0.009560583159327507, + 0.005802984815090895, + 0.013002411462366581, + 0.007432838436216116, + 0.007266496308147907, + 0.006130052264779806, + 0.005304621532559395, + 0.005779181607067585, + 0.009503399953246117, + 0.005923830438405275, + 0.009901797398924828, + 0.00991924200206995, + 0.016472309827804565, + 0.008046207018196583, + 0.012278754264116287, + 0.008716181851923466, + 0.01431179791688919, + 0.011685588397085667, + 0.011244162917137146, + 0.022166753187775612, + 0.0023107260931283236, + 0.0016724251909181476, + 0.003176967380568385, + 0.002185248536989093, + 0.0020998474210500717, + 0.0038557597436010838, + 0.0017453820910304785, + 0.0037393749225884676, + 0.0009165494702756405, + 0.0007078754715621471, + 0.0008326327078975737, + 0.0007470324635505676, + 0.001008623163215816, + 0.000814650091342628, + 0.0007088009733706713, + 0.0007932077278383076, + 0.0028013053815811872, + 0.001683493028394878, + 0.0021586562506854534, + 0.0020859038922935724, + 0.002593688666820526, + 0.0023765715304762125, + 0.0014936171937733889, + 0.0017001514788717031, + 0.0018083886243402958, + 0.0016393710393458605, + 0.001546899788081646, + 0.0017118446994572878, + 0.0029926325660198927, + 0.0021465513855218887, + 0.0018276199698448181, + 0.0023914158809930086, + 0.06222028657793999, + 0.007781025022268295, + 0.014119602739810944, + 0.005053823348134756, + 0.00924971979111433, + 0.15822413563728333, + 0.01420015562325716, + 0.11096185445785522 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.011073916219174862 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31343862414360046 + ], + "opposite_logprob_tokens": [ + -1.3134386539459229 + ], + "logprob_reward": -0.31343862414360046, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31343862414360046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2771073281764984, + 0.12407639622688293, + 0.0630224272608757, + 0.061586156487464905, + 0.08199417591094971, + 0.047612447291612625, + 0.07295568287372589, + 0.07051514834165573, + 0.17803870141506195, + 0.023091552779078484 + ], + "credit_map": [ + 0.02713830955326557, + 0.014483037404716015, + 0.019063754007220268, + 0.018357936292886734, + 0.033708374947309494, + 0.08002153038978577, + 0.03679875284433365, + 0.047535646706819534, + 0.008673866279423237, + 0.0188816599547863, + 0.019501226022839546, + 0.01646706461906433, + 0.020932605490088463, + 0.012610947713255882, + 0.013523354195058346, + 0.013485671021044254, + 0.006761998403817415, + 0.00700193177908659, + 0.00905151292681694, + 0.00642782635986805, + 0.01158418320119381, + 0.006807783152908087, + 0.009685048833489418, + 0.005702144932001829, + 0.007008381653577089, + 0.007662511430680752, + 0.005060703027993441, + 0.007201971486210823, + 0.011217051185667515, + 0.006589742377400398, + 0.007513480260968208, + 0.009332315064966679, + 0.01039009541273117, + 0.008025111630558968, + 0.007679034490138292, + 0.007325527723878622, + 0.014329899102449417, + 0.009333891794085503, + 0.007551576476544142, + 0.017359044402837753, + 0.006267201621085405, + 0.006121496204286814, + 0.005753190256655216, + 0.007533262018114328, + 0.006719288416206837, + 0.004728381056338549, + 0.004681790713220835, + 0.0058078402653336525, + 0.008171960711479187, + 0.011892218142747879, + 0.008255247958004475, + 0.008244548924267292, + 0.013653360307216644, + 0.007256124634295702, + 0.00751773314550519, + 0.007964493706822395, + 0.009861843660473824, + 0.006008830852806568, + 0.008390243165194988, + 0.006118436809629202, + 0.009308434091508389, + 0.014673599041998386, + 0.005710729397833347, + 0.01044302899390459, + 0.028948497027158737, + 0.01001078262925148, + 0.012725620530545712, + 0.010327812284231186, + 0.012495200149714947, + 0.054698359221220016, + 0.01212749257683754, + 0.03670494630932808, + 0.004803371150046587, + 0.0012901769950985909, + 0.003584962571039796, + 0.0019332007504999638, + 0.0024496635887771845, + 0.004069621674716473, + 0.0023801387287676334, + 0.0025804161559790373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.31343862414360046 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a striking overhead shot, focusing on a lion in its prime, standing majestically on the edge of a rugged, rocky cliff. Its thick, dark brown and golden mane, like a crown of kings, sways gently in the evening breeze. Its body, covered in short, sandy-brown fur, is defined by its muscularity and strength, its limbs firmly planted on the rock. Facing the sky, it leans forward slightly, its maw suddenly opening in a silent yet earth-shattering roar. Its sharp, gleaming white canines and nostrils flaring with effort are clearly visible. Its chest heaves violently, a stream of condensed hot air erupts from its mouth, and its eyes are filled with sharpness and majesty. The backdrop is the vast African savannah, the sky stained in a magnificent gradation of orange, red, and deep purple by the setting sun. Warm, dramatic golden sidelights outline the lion's majestic silhouette, casting a long shadow behind it. As this shocking scene unfolds, the camera begins to rise steadily and slowly, gradually incorporating more of the sky and the distant horizon into the picture, further emphasizing the lion's loneliness and supreme status in the vast world. The whole picture is full of primitive, magnificent and epic atmosphere of absolute dominance.", + "seed": 20161, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_8562_seed_20161_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10224443040788174, + "vlm_logprob": -0.10224443040788174, + "vlm_yes_no_margin": 2.950000011920929, + "vlm_token_credit": -0.10224443040788174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01704554446041584 + ], + "opposite_logprob_tokens": [ + -4.517045497894287 + ], + "logprob_reward": -0.01704554446041584, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01704554446041584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1345464438199997, + 0.15491214394569397, + 0.16043677926063538, + 0.09200182557106018, + 0.12425243854522705, + 0.020515335723757744, + 0.04518170282244682, + 0.09844590723514557, + 0.07336366921663284, + 0.09634380787611008 + ], + "credit_map": [ + 0.02035948634147644, + 0.017842799425125122, + 0.011215867474675179, + 0.013698946684598923, + 0.02365502342581749, + 0.013472900725901127, + 0.018911372870206833, + 0.015390019863843918, + 0.0324382483959198, + 0.041727013885974884, + 0.011416221037507057, + 0.01917530782520771, + 0.008375083096325397, + 0.01551868487149477, + 0.011660778895020485, + 0.014600791968405247, + 0.02008231170475483, + 0.01933799311518669, + 0.016543786972761154, + 0.020272739231586456, + 0.022891875356435776, + 0.022075137123465538, + 0.01587728224694729, + 0.0233556367456913, + 0.01833689957857132, + 0.012918018735945225, + 0.011081038042902946, + 0.01674097403883934, + 0.009064807556569576, + 0.007117391563951969, + 0.008243365213274956, + 0.008499321527779102, + 0.018467681482434273, + 0.01884358935058117, + 0.01129787415266037, + 0.014132155105471611, + 0.020376287400722504, + 0.018830986693501472, + 0.00994558073580265, + 0.012358266860246658, + 0.0024713336024433374, + 0.0027647893875837326, + 0.00261731562204659, + 0.0028668877203017473, + 0.002668920438736677, + 0.002248512813821435, + 0.0028857230208814144, + 0.0019918496254831553, + 0.007095914799720049, + 0.0055968728847801685, + 0.008678646758198738, + 0.0072457012720406055, + 0.00466783344745636, + 0.004863137379288673, + 0.003910392988473177, + 0.0031231979373842478, + 0.014416169375181198, + 0.01814289018511772, + 0.010763258673250675, + 0.012250258587300777, + 0.013356215320527554, + 0.012123691849410534, + 0.009635485708713531, + 0.007757934276014566, + 0.008571031503379345, + 0.007063116878271103, + 0.01233502198010683, + 0.015040426515042782, + 0.007573414593935013, + 0.008305968716740608, + 0.010417249985039234, + 0.004057438112795353, + 0.014634340070188046, + 0.015808289870619774, + 0.013296262361109257, + 0.016468431800603867, + 0.011040489189326763, + 0.00937195960432291, + 0.00935555249452591, + 0.006368468981236219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.01704554446041584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07004271447658539 + ], + "opposite_logprob_tokens": [ + -2.695042610168457 + ], + "logprob_reward": -0.07004271447658539, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07004271447658539, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2823769748210907, + 0.037995852530002594, + 0.04725922271609306, + 0.05161725729703903, + 0.06122554466128349, + 0.054497476667165756, + 0.024775544181466103, + 0.09018648415803909, + 0.09076214581727982, + 0.2593035399913788 + ], + "credit_map": [ + 0.03394124284386635, + 0.04103926196694374, + 0.03697816655039787, + 0.03647374361753464, + 0.05075441300868988, + 0.03278607875108719, + 0.028911689296364784, + 0.021492354571819305, + 0.004120763391256332, + 0.004369854927062988, + 0.0035480454098433256, + 0.004795530810952187, + 0.0044723618775606155, + 0.00661720335483551, + 0.004433563444763422, + 0.005638524889945984, + 0.006704594008624554, + 0.0068672774359583855, + 0.0062986016273498535, + 0.006278745364397764, + 0.0051688156090676785, + 0.005253246054053307, + 0.00456174137070775, + 0.006126196589320898, + 0.010779435746371746, + 0.008303373120725155, + 0.0058066342025995255, + 0.007922932505607605, + 0.004597872495651245, + 0.006876053754240274, + 0.0034379276912659407, + 0.0038930189330130816, + 0.011122580617666245, + 0.010755173861980438, + 0.007595131639391184, + 0.007897550240159035, + 0.006594021338969469, + 0.008052143268287182, + 0.0038991679903119802, + 0.005309766158461571, + 0.009666694328188896, + 0.008367234840989113, + 0.010099890641868114, + 0.008121011778712273, + 0.0040849982760846615, + 0.005200102925300598, + 0.0026869620196521282, + 0.006270576734095812, + 0.0038605902809649706, + 0.005059039685875177, + 0.0030226181261241436, + 0.0033472145441919565, + 0.002644470427185297, + 0.003143594367429614, + 0.0019335700199007988, + 0.001764445099979639, + 0.011129667051136494, + 0.01630527898669243, + 0.01122831366956234, + 0.01236544456332922, + 0.008575708605349064, + 0.010002249851822853, + 0.00927723292261362, + 0.011302580125629902, + 0.01301854383200407, + 0.01890057697892189, + 0.010131855495274067, + 0.011037764139473438, + 0.008989022113382816, + 0.011046980507671833, + 0.009453552775084972, + 0.008183840662240982, + 0.03793508559465408, + 0.04753990098834038, + 0.03740961104631424, + 0.041156575083732605, + 0.027500350028276443, + 0.032604608684778214, + 0.019386474043130875, + 0.015770914033055305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.07004271447658539 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495057255029678 + ], + "opposite_logprob_tokens": [ + -2.929950475692749 + ], + "logprob_reward": -0.05495057255029678, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495057255029678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1306983232498169, + 0.1466246396303177, + 0.2445041984319687, + 0.10373225808143616, + 0.04785620793700218, + 0.03440307453274727, + 0.04258766025304794, + 0.17508402466773987, + 0.031362343579530716, + 0.0431472510099411 + ], + "credit_map": [ + 0.031283795833587646, + 0.03156420961022377, + 0.01019351463764906, + 0.012135263532400131, + 0.015700150281190872, + 0.010982700623571873, + 0.010447918437421322, + 0.008390773087739944, + 0.01898907870054245, + 0.014435088261961937, + 0.014574646949768066, + 0.01990135945379734, + 0.01877496764063835, + 0.020999789237976074, + 0.02123001404106617, + 0.017719699069857597, + 0.03203791379928589, + 0.02939414419233799, + 0.024824541062116623, + 0.035746652632951736, + 0.03304741904139519, + 0.03319340571761131, + 0.020274672657251358, + 0.03598548099398613, + 0.011947627179324627, + 0.012216837145388126, + 0.012080724351108074, + 0.01768471486866474, + 0.01033432874828577, + 0.015081535093486309, + 0.010125978849828243, + 0.014260527677834034, + 0.006006184499710798, + 0.005574224516749382, + 0.005680696107447147, + 0.00599296111613512, + 0.007151264231652021, + 0.00772433914244175, + 0.006073927041143179, + 0.0036526129115372896, + 0.00412251939997077, + 0.004379290621727705, + 0.006237530615180731, + 0.00510491942986846, + 0.004346441477537155, + 0.0043517970480024815, + 0.0033900386188179255, + 0.002470539417117834, + 0.005539323203265667, + 0.004739409312605858, + 0.007670932915061712, + 0.006678670644760132, + 0.00542276818305254, + 0.005082448944449425, + 0.003781783627346158, + 0.003672325750812888, + 0.02475295588374138, + 0.031169047579169273, + 0.018790913745760918, + 0.020382048562169075, + 0.020444458350539207, + 0.026351207867264748, + 0.01859445497393608, + 0.014598947018384933, + 0.00390990125015378, + 0.005321416538208723, + 0.003844362683594227, + 0.004200776573270559, + 0.003927063196897507, + 0.003885958343744278, + 0.0032313198316842318, + 0.0030415502842515707, + 0.006321795284748077, + 0.007415663450956345, + 0.0046999710611999035, + 0.006476819980889559, + 0.005327331367880106, + 0.005699352361261845, + 0.0035864689853042364, + 0.0036198520101606846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.05495057255029678 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577985793352127 + ], + "opposite_logprob_tokens": [ + -3.895577907562256 + ], + "logprob_reward": -0.020577985793352127, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577985793352127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20156463980674744, + 0.05909741297364235, + 0.14026615023612976, + 0.0681183934211731, + 0.07419933378696442, + 0.047076981514692307, + 0.06256632506847382, + 0.09916427731513977, + 0.05454014241695404, + 0.19340640306472778 + ], + "credit_map": [ + 0.037145473062992096, + 0.03686146438121796, + 0.01651858352124691, + 0.019423015415668488, + 0.03126468136906624, + 0.019390283152461052, + 0.022593040019273758, + 0.018368089571595192, + 0.007661933079361916, + 0.009572065435349941, + 0.005272664595395327, + 0.007529933005571365, + 0.006510948296636343, + 0.006169433239847422, + 0.008728963322937489, + 0.0076514738611876965, + 0.021783242002129555, + 0.016721006482839584, + 0.014204207807779312, + 0.019514087587594986, + 0.016696685925126076, + 0.018494123592972755, + 0.012454533949494362, + 0.020398272201418877, + 0.01670003868639469, + 0.011516454629600048, + 0.007431601174175739, + 0.010633350349962711, + 0.005621595773845911, + 0.00564892403781414, + 0.005697321146726608, + 0.004869107622653246, + 0.012175112962722778, + 0.011804018169641495, + 0.0077669695019721985, + 0.009034319780766964, + 0.01086343452334404, + 0.010389619506895542, + 0.0046559651382267475, + 0.007509891875088215, + 0.007764074020087719, + 0.005619489587843418, + 0.00722786458209157, + 0.008970032446086407, + 0.005842522252351046, + 0.004326929338276386, + 0.004178173840045929, + 0.0031478938180953264, + 0.008350903168320656, + 0.010645986534655094, + 0.008736046962440014, + 0.008475455455482006, + 0.007388796657323837, + 0.007190861273556948, + 0.00655987486243248, + 0.0052184006199240685, + 0.012638990767300129, + 0.014518804848194122, + 0.014898969791829586, + 0.01722515933215618, + 0.008618759922683239, + 0.0099971704185009, + 0.011366414837539196, + 0.009900003671646118, + 0.006707468070089817, + 0.00817115232348442, + 0.007821738719940186, + 0.008577045984566212, + 0.004900380503386259, + 0.007198705803602934, + 0.006884153466671705, + 0.004279498476535082, + 0.027959229424595833, + 0.036553867161273956, + 0.026318132877349854, + 0.030332978814840317, + 0.021153073757886887, + 0.02521255984902382, + 0.014493039809167385, + 0.011383503675460815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.020577985793352127 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34860533475875854 + ], + "opposite_logprob_tokens": [ + -1.2236053943634033 + ], + "logprob_reward": -0.34860533475875854, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.34860533475875854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25402840971946716, + 0.08902741223573685, + 0.2229522466659546, + 0.059056274592876434, + 0.11517923325300217, + 0.04628978297114372, + 0.08970575779676437, + 0.05233415216207504, + 0.020932363346219063, + 0.05049436166882515 + ], + "credit_map": [ + 0.058877572417259216, + 0.06179514154791832, + 0.014852730557322502, + 0.01840892992913723, + 0.040880829095840454, + 0.019863348454236984, + 0.022720934823155403, + 0.016628921031951904, + 0.016188545152544975, + 0.017975468188524246, + 0.0067344107665121555, + 0.011378692463040352, + 0.006914013065397739, + 0.011206841096282005, + 0.009177025407552719, + 0.009452417492866516, + 0.029554061591625214, + 0.030032839626073837, + 0.02102518640458584, + 0.02967679500579834, + 0.031172454357147217, + 0.03142255172133446, + 0.019133539870381355, + 0.030934829264879227, + 0.007178353611379862, + 0.009052105247974396, + 0.007780066225677729, + 0.00935277994722128, + 0.006916467100381851, + 0.007811919320374727, + 0.004870619624853134, + 0.006093961652368307, + 0.01877991296350956, + 0.020886769518256187, + 0.012784936465322971, + 0.0170629620552063, + 0.014274735003709793, + 0.014004820957779884, + 0.006508592050522566, + 0.010876494459807873, + 0.006985684856772423, + 0.00861149188131094, + 0.0043906401842832565, + 0.005203647073358297, + 0.006599937565624714, + 0.005755106918513775, + 0.004636718425899744, + 0.004106554668396711, + 0.014519665390253067, + 0.013448871672153473, + 0.012538241222500801, + 0.011128812097012997, + 0.011295374482870102, + 0.009740117006003857, + 0.006822172086685896, + 0.010212501510977745, + 0.0085690813139081, + 0.009264674969017506, + 0.007069460581988096, + 0.007254777476191521, + 0.004511785693466663, + 0.004990205634385347, + 0.006452109664678574, + 0.004222051706165075, + 0.0023032764438539743, + 0.0035310275852680206, + 0.0023740206379443407, + 0.0026374931912869215, + 0.0030853799544274807, + 0.0028240117244422436, + 0.00265418435446918, + 0.0015229691052809358, + 0.0066214026883244514, + 0.006423326674848795, + 0.008060664869844913, + 0.01044179406017065, + 0.0051205949857831, + 0.004011780489236116, + 0.006194009445607662, + 0.003620789386332035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.34860533475875854 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a striking overhead shot, focusing on a lion in its prime, standing majestically on the edge of a rugged, rocky cliff. Its thick, dark brown and golden mane, like a crown of kings, sways gently in the evening breeze. Its body, covered in short, sandy-brown fur, is defined by its muscularity and strength, its limbs firmly planted on the rock. Facing the sky, it leans forward slightly, its maw suddenly opening in a silent yet earth-shattering roar. Its sharp, gleaming white canines and nostrils flaring with effort are clearly visible. Its chest heaves violently, a stream of condensed hot air erupts from its mouth, and its eyes are filled with sharpness and majesty. The backdrop is the vast African savannah, the sky stained in a magnificent gradation of orange, red, and deep purple by the setting sun. Warm, dramatic golden sidelights outline the lion's majestic silhouette, casting a long shadow behind it. As this shocking scene unfolds, the camera begins to rise steadily and slowly, gradually incorporating more of the sky and the distant horizon into the picture, further emphasizing the lion's loneliness and supreme status in the vast world. The whole picture is full of primitive, magnificent and epic atmosphere of absolute dominance.", + "seed": 20162, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_8562_seed_20162_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19578445758670568, + "vlm_logprob": -0.19578445758670568, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.19578445758670568, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018430527299642563 + ], + "opposite_logprob_tokens": [ + -4.768430709838867 + ], + "logprob_reward": -0.018430527299642563, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018430527299642563, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12155608087778091, + 0.0327395536005497, + 0.07199510186910629, + 0.029326213523745537, + 0.04838273301720619, + 0.1739850640296936, + 0.11118772625923157, + 0.15525107085704803, + 0.18646672368049622, + 0.0691097229719162 + ], + "credit_map": [ + 0.014200803823769093, + 0.020385805517435074, + 0.021276813000440598, + 0.017845554277300835, + 0.01442658994346857, + 0.00796575378626585, + 0.012283126823604107, + 0.013171643950045109, + 0.0034451938699930906, + 0.004979270976036787, + 0.0057424018159508705, + 0.005240883678197861, + 0.004016137681901455, + 0.002137113129720092, + 0.0037904444616287947, + 0.0033881082199513912, + 0.007497712504118681, + 0.008278070017695427, + 0.01552713941782713, + 0.010790255852043629, + 0.008751251734793186, + 0.005515804514288902, + 0.007183095905929804, + 0.008451763540506363, + 0.0035249649081379175, + 0.002836870262399316, + 0.006805081851780415, + 0.005744937341660261, + 0.0024796780198812485, + 0.001393006881698966, + 0.0033051392529159784, + 0.0032365324441343546, + 0.006002988666296005, + 0.005406222306191921, + 0.007919250056147575, + 0.0069176857359707355, + 0.00543065182864666, + 0.003548700362443924, + 0.006038421764969826, + 0.007118812296539545, + 0.02276843972504139, + 0.02180580608546734, + 0.02263224497437477, + 0.024450408294796944, + 0.021080605685710907, + 0.013017842546105385, + 0.023535026237368584, + 0.024694694206118584, + 0.013257352635264397, + 0.009408405981957912, + 0.015027574263513088, + 0.01670871302485466, + 0.01598436012864113, + 0.006649153772741556, + 0.01524541899561882, + 0.018906760960817337, + 0.019691690802574158, + 0.007497772108763456, + 0.05272620543837547, + 0.029671519994735718, + 0.011862602084875107, + 0.013405839912593365, + 0.011600798927247524, + 0.008794641122221947, + 0.017686743289232254, + 0.011842217296361923, + 0.04489751532673836, + 0.04270361363887787, + 0.01720310002565384, + 0.02354550175368786, + 0.01476910151541233, + 0.013818934559822083, + 0.008747737854719162, + 0.009735822677612305, + 0.009301923215389252, + 0.008254124782979488, + 0.010833170264959335, + 0.0075239501893520355, + 0.007966205477714539, + 0.00674678198993206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 7 + ], + "selected_reward": -0.018430527299642563 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42894506454467773 + ], + "opposite_logprob_tokens": [ + -1.0539450645446777 + ], + "logprob_reward": -0.42894506454467773, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42894506454467773, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23264141380786896, + 0.07370420545339584, + 0.043989624828100204, + 0.041322339326143265, + 0.034801725298166275, + 0.04826263710856438, + 0.04133336618542671, + 0.09876281768083572, + 0.33030763268470764, + 0.05487425625324249 + ], + "credit_map": [ + 0.027030806988477707, + 0.016618145629763603, + 0.07511252909898758, + 0.054022494703531265, + 0.014606024138629436, + 0.012543605640530586, + 0.011735906824469566, + 0.020971886813640594, + 0.008836254477500916, + 0.004892663564532995, + 0.02533286064863205, + 0.01582612469792366, + 0.0051208618097007275, + 0.003366350196301937, + 0.005320198833942413, + 0.005008886568248272, + 0.004836118780076504, + 0.003656348679214716, + 0.013008931651711464, + 0.008701358921825886, + 0.0037178227212280035, + 0.0030726324766874313, + 0.0030915397219359875, + 0.003904870245605707, + 0.005029826425015926, + 0.0027927469927817583, + 0.01367950439453125, + 0.008000541478395462, + 0.0031446930952370167, + 0.0021110328380018473, + 0.002469367114827037, + 0.004094622563570738, + 0.00463186576962471, + 0.004526595119386911, + 0.008300963789224625, + 0.005180447828024626, + 0.002872117795050144, + 0.002475943649187684, + 0.003408190095797181, + 0.003405597759410739, + 0.006465950049459934, + 0.00715361163020134, + 0.008356219157576561, + 0.0074178907088935375, + 0.005610022693872452, + 0.004096613731235266, + 0.004930655471980572, + 0.004231668077409267, + 0.005419409833848476, + 0.003970764111727476, + 0.007972551509737968, + 0.007017887197434902, + 0.006487792823463678, + 0.0035329898819327354, + 0.0033050428610295057, + 0.00362692354246974, + 0.008687465451657772, + 0.006799666676670313, + 0.030881892889738083, + 0.01189784798771143, + 0.0034014484845101833, + 0.029675163328647614, + 0.0029193521477282047, + 0.0044999755918979645, + 0.02704690396785736, + 0.020923743024468422, + 0.08505845069885254, + 0.06952821463346481, + 0.04079103097319603, + 0.03837664797902107, + 0.022428471595048904, + 0.02615414932370186, + 0.00695407809689641, + 0.006412220653146505, + 0.0070805796422064304, + 0.006221736781299114, + 0.008100675418972969, + 0.005140013061463833, + 0.00825723260641098, + 0.006707713473588228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.42894506454467773 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048657845705747604 + ], + "opposite_logprob_tokens": [ + -3.0486578941345215 + ], + "logprob_reward": -0.048657845705747604, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048657845705747604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05690186843276024, + 0.03765636309981346, + 0.04891333729028702, + 0.04042162746191025, + 0.03468189388513565, + 0.1368602216243744, + 0.06186489760875702, + 0.03502863273024559, + 0.326945036649704, + 0.2207261025905609 + ], + "credit_map": [ + 0.006281370297074318, + 0.008802602998912334, + 0.011857409030199051, + 0.008237192407250404, + 0.006508846301585436, + 0.004232518840581179, + 0.005878245458006859, + 0.005103689152747393, + 0.003991756588220596, + 0.005640655290335417, + 0.005194648634642363, + 0.004839453846216202, + 0.00511931162327528, + 0.0036396507639437914, + 0.0054783569648861885, + 0.0037525345105677843, + 0.005427720956504345, + 0.007582411170005798, + 0.007122992072254419, + 0.005967809818685055, + 0.006040312349796295, + 0.0060111153870821, + 0.0051404875703155994, + 0.00562049588188529, + 0.004598390311002731, + 0.007784760091453791, + 0.004755477886646986, + 0.005460883025079966, + 0.002733750967308879, + 0.0026379982009530067, + 0.004392759408801794, + 0.008057610131800175, + 0.003984925802797079, + 0.005883981008082628, + 0.006424540188163519, + 0.004586780443787575, + 0.003657076507806778, + 0.0020971661433577538, + 0.004505170974880457, + 0.0035422546789050102, + 0.01801370643079281, + 0.01576373353600502, + 0.015623615123331547, + 0.018537618219852448, + 0.017132604494690895, + 0.011214149184525013, + 0.021388405933976173, + 0.01918639801442623, + 0.007802834268659353, + 0.006305041257292032, + 0.010049371048808098, + 0.009618562646210194, + 0.011649960651993752, + 0.005459857173264027, + 0.006316415499895811, + 0.004662861116230488, + 0.004351899027824402, + 0.003661140101030469, + 0.0053396401926875114, + 0.005944000091403723, + 0.004106281790882349, + 0.004446037113666534, + 0.003045019693672657, + 0.004134615417569876, + 0.020319387316703796, + 0.024289283901453018, + 0.07287373393774033, + 0.07623329758644104, + 0.03363032639026642, + 0.038906730711460114, + 0.028427472338080406, + 0.03226481378078461, + 0.02036084607243538, + 0.03325110673904419, + 0.030288591980934143, + 0.03346675634384155, + 0.0340333916246891, + 0.018940811976790428, + 0.029506754130125046, + 0.020877856761217117 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.048657845705747604 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863614957779646 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863614957779646, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863614957779646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08693360537290573, + 0.06621568650007248, + 0.04254867136478424, + 0.033118803054094315, + 0.023824265226721764, + 0.05853864923119545, + 0.06834226846694946, + 0.2216263860464096, + 0.3078650236129761, + 0.0909866914153099 + ], + "credit_map": [ + 0.008797026239335537, + 0.0071672797203063965, + 0.02819833718240261, + 0.018051186576485634, + 0.0069562094286084175, + 0.005320681259036064, + 0.00443543354049325, + 0.008007455617189407, + 0.007146277464926243, + 0.006123557686805725, + 0.020250149071216583, + 0.0123025206848979, + 0.004854958038777113, + 0.0039520603604614735, + 0.004896144848316908, + 0.006690019275993109, + 0.005118072498589754, + 0.005448139272630215, + 0.011916029267013073, + 0.006295790895819664, + 0.003004599828273058, + 0.0033906472381204367, + 0.0030109272338449955, + 0.004364462569355965, + 0.0033128985669463873, + 0.0036188336089253426, + 0.005201715975999832, + 0.0054264068603515625, + 0.0026285992935299873, + 0.00205200701020658, + 0.003233938477933407, + 0.00764440419152379, + 0.002695737173780799, + 0.0029072973411530256, + 0.006674184463918209, + 0.0036880928091704845, + 0.00202625198289752, + 0.0016862768679857254, + 0.0020412837620824575, + 0.0021051394287496805, + 0.013950703665614128, + 0.01012247335165739, + 0.007853467017412186, + 0.007377495523542166, + 0.005128967575728893, + 0.005081966053694487, + 0.0050696018151938915, + 0.00395397562533617, + 0.008776434697210789, + 0.006739508826285601, + 0.012605083175003529, + 0.018419433385133743, + 0.0076675512827932835, + 0.005703179631382227, + 0.004032399505376816, + 0.004398683086037636, + 0.021966006606817245, + 0.01449431013315916, + 0.06036629155278206, + 0.030477028340101242, + 0.006108283065259457, + 0.07402579486370087, + 0.006081708241254091, + 0.008106951601803303, + 0.03343801200389862, + 0.01535536628216505, + 0.09225516021251678, + 0.07780091464519501, + 0.015575209632515907, + 0.050204213708639145, + 0.011770662851631641, + 0.011465484276413918, + 0.009058607742190361, + 0.015697717666625977, + 0.012587408535182476, + 0.014286047779023647, + 0.013134542852640152, + 0.00797831267118454, + 0.01022281777113676, + 0.00802123174071312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.00863614957779646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47425270080566406 + ], + "opposite_logprob_tokens": [ + -0.9742527008056641 + ], + "logprob_reward": -0.47425270080566406, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47425270080566406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057424161583185196, + 0.04905496910214424, + 0.015284035354852676, + 0.040006961673498154, + 0.0360792800784111, + 0.056767188012599945, + 0.03794976323843002, + 0.5775293111801147, + 0.08800703287124634, + 0.04189726710319519 + ], + "credit_map": [ + 0.007468244060873985, + 0.00807751901447773, + 0.010765806771814823, + 0.0075162313878536224, + 0.008649956434965134, + 0.005102850031107664, + 0.004371358081698418, + 0.005472193937748671, + 0.005956373177468777, + 0.006039437837898731, + 0.008786591701209545, + 0.006582380272448063, + 0.008022471331059933, + 0.00502501567825675, + 0.00396103598177433, + 0.004681659396737814, + 0.001432496472261846, + 0.0026111893821507692, + 0.002568059368059039, + 0.001969856210052967, + 0.0018250608118250966, + 0.0009669574210420251, + 0.0017036189092323184, + 0.0022067974787205458, + 0.004364506807178259, + 0.006634937599301338, + 0.00553647568449378, + 0.0050465939566493034, + 0.004797334782779217, + 0.0027247851248830557, + 0.004931746516376734, + 0.005970581900328398, + 0.004726797342300415, + 0.004188466817140579, + 0.006985639221966267, + 0.004391418769955635, + 0.0031472144182771444, + 0.0027655453886836767, + 0.005396978929638863, + 0.004477219190448523, + 0.009259229525923729, + 0.007117951288819313, + 0.008767087012529373, + 0.008891490288078785, + 0.0065086656250059605, + 0.0031424611806869507, + 0.00705968402326107, + 0.006020617671310902, + 0.004083320964127779, + 0.004052262287586927, + 0.004575691651552916, + 0.005499762948602438, + 0.007431346457451582, + 0.004019931424409151, + 0.004140234552323818, + 0.004147213418036699, + 0.0687912330031395, + 0.042823661118745804, + 0.15499965846538544, + 0.040940236300230026, + 0.01689114049077034, + 0.2209443897008896, + 0.012833886779844761, + 0.019305115565657616, + 0.013055644929409027, + 0.00904148817062378, + 0.009268855676054955, + 0.008810040540993214, + 0.014990490861237049, + 0.010799414478242397, + 0.00875230971723795, + 0.013288787566125393, + 0.006853538099676371, + 0.005676803644746542, + 0.005193824414163828, + 0.00598538713529706, + 0.005632504820823669, + 0.005589837674051523, + 0.004075942095369101, + 0.0028894280549138784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.47425270080566406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The camera begins with a striking overhead shot, focusing on a lion in its prime, standing majestically on the edge of a rugged, rocky cliff. Its thick, dark brown and golden mane, like a crown of kings, sways gently in the evening breeze. Its body, covered in short, sandy-brown fur, is defined by its muscularity and strength, its limbs firmly planted on the rock. Facing the sky, it leans forward slightly, its maw suddenly opening in a silent yet earth-shattering roar. Its sharp, gleaming white canines and nostrils flaring with effort are clearly visible. Its chest heaves violently, a stream of condensed hot air erupts from its mouth, and its eyes are filled with sharpness and majesty. The backdrop is the vast African savannah, the sky stained in a magnificent gradation of orange, red, and deep purple by the setting sun. Warm, dramatic golden sidelights outline the lion's majestic silhouette, casting a long shadow behind it. As this shocking scene unfolds, the camera begins to rise steadily and slowly, gradually incorporating more of the sky and the distant horizon into the picture, further emphasizing the lion's loneliness and supreme status in the vast world. The whole picture is full of primitive, magnificent and epic atmosphere of absolute dominance.", + "seed": 20163, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_11/index_8562_seed_20163_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11275981105864048, + "vlm_logprob": -0.11275981105864048, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.11275981105864048, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022014880552887917 + ], + "opposite_logprob_tokens": [ + -4.14701509475708 + ], + "logprob_reward": -0.022014880552887917, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.022014880552887917, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21507707238197327, + 0.054190631955862045, + 0.02117238938808441, + 0.08012678474187851, + 0.04579963907599449, + 0.1321292370557785, + 0.06828434765338898, + 0.14303429424762726, + 0.11792194098234177, + 0.12226361036300659 + ], + "credit_map": [ + 0.03057878650724888, + 0.019769422709941864, + 0.033064741641283035, + 0.025635819882154465, + 0.053883522748947144, + 0.019764209166169167, + 0.015424532815814018, + 0.016956038773059845, + 0.005472636315971613, + 0.005419125314801931, + 0.007004953920841217, + 0.010739754885435104, + 0.005078808404505253, + 0.004328298382461071, + 0.00804569385945797, + 0.008101362735033035, + 0.0017725059296935797, + 0.0020644953474402428, + 0.002598579740151763, + 0.0032022371888160706, + 0.002601028885692358, + 0.0021099578589200974, + 0.003571216482669115, + 0.003252368653193116, + 0.008072208613157272, + 0.009125896729528904, + 0.011350063607096672, + 0.011453372426331043, + 0.012380264699459076, + 0.006947780959308147, + 0.009674168191850185, + 0.011123024858534336, + 0.006870649289339781, + 0.0059731705114245415, + 0.0064690448343753815, + 0.0055381120182573795, + 0.004912732169032097, + 0.007082049734890461, + 0.004321793559938669, + 0.0046320888213813305, + 0.014756862074136734, + 0.008761445991694927, + 0.03283527493476868, + 0.014569820836186409, + 0.009104994125664234, + 0.034229572862386703, + 0.008485355414450169, + 0.009385921061038971, + 0.012906966730952263, + 0.006605576258152723, + 0.006432871334254742, + 0.007336904760450125, + 0.007701059803366661, + 0.008248534984886646, + 0.006928339600563049, + 0.012124096043407917, + 0.018470047041773796, + 0.016816062852740288, + 0.009960702620446682, + 0.014230945147573948, + 0.01996891014277935, + 0.021150853484869003, + 0.018388669937849045, + 0.024048108607530594, + 0.01255473680794239, + 0.010067431256175041, + 0.021387742832303047, + 0.01451859250664711, + 0.01507558673620224, + 0.01925363950431347, + 0.007409083656966686, + 0.017655132338404655, + 0.011722691357135773, + 0.010179305449128151, + 0.023607831448316574, + 0.01601838693022728, + 0.017762809991836548, + 0.014954165555536747, + 0.012518787756562233, + 0.015499632805585861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.022014880552887917 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296024173498154 + ], + "opposite_logprob_tokens": [ + -2.2379603385925293 + ], + "logprob_reward": -0.11296024173498154, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296024173498154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4896887540817261, + 0.03473172336816788, + 0.02219979278743267, + 0.021145157516002655, + 0.041572924703359604, + 0.02348269522190094, + 0.09121707081794739, + 0.24664555490016937, + 0.011985411867499352, + 0.017330916598439217 + ], + "credit_map": [ + 0.08739769458770752, + 0.04580474644899368, + 0.10844892263412476, + 0.07737497985363007, + 0.059646882116794586, + 0.042586538940668106, + 0.034581974148750305, + 0.03384704142808914, + 0.004281525034457445, + 0.002831135643646121, + 0.004162193741649389, + 0.007508855778723955, + 0.003074867883697152, + 0.002892706310376525, + 0.003326370380818844, + 0.0066540720872581005, + 0.0025166228879243135, + 0.001718949293717742, + 0.0036840420216321945, + 0.004377895034849644, + 0.0027586412616074085, + 0.0016546305269002914, + 0.002835276536643505, + 0.0026537366211414337, + 0.004881040193140507, + 0.0023691519163548946, + 0.0031371715012937784, + 0.0034019392915070057, + 0.0016938529442995787, + 0.0016879432369023561, + 0.0022442126646637917, + 0.0017298486782237887, + 0.005401054862886667, + 0.004881163127720356, + 0.004514859523624182, + 0.0037015643902122974, + 0.00445312075316906, + 0.006879361346364021, + 0.00613537710160017, + 0.005606425926089287, + 0.0031966606620699167, + 0.003191816620528698, + 0.002303070155903697, + 0.002728355349972844, + 0.0034156653564423323, + 0.002649446716532111, + 0.002645818516612053, + 0.0033518641721457243, + 0.013999764807522297, + 0.00929277203977108, + 0.010120903141796589, + 0.010146168060600758, + 0.010728327557444572, + 0.012047644704580307, + 0.010782329365611076, + 0.01409917138516903, + 0.03278335928916931, + 0.02893858589231968, + 0.01958215795457363, + 0.02598467841744423, + 0.03575371205806732, + 0.03421948105096817, + 0.03062470629811287, + 0.03875888139009476, + 0.0015670633874833584, + 0.001254921080544591, + 0.0017748801037669182, + 0.0018347763689234853, + 0.0014109067851677537, + 0.0016047080280259252, + 0.000996139133349061, + 0.0015420179115608335, + 0.002583327004685998, + 0.0015441784635186195, + 0.0024452980142086744, + 0.0025989997666329145, + 0.0023306733928620815, + 0.0018386084120720625, + 0.0014544050209224224, + 0.0025354286190122366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.11296024173498154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204257160425186 + ], + "opposite_logprob_tokens": [ + -2.812042474746704 + ], + "logprob_reward": -0.06204257160425186, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204257160425186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08892656862735748, + 0.08400191366672516, + 0.0597711056470871, + 0.04990826174616814, + 0.14841920137405396, + 0.10013154149055481, + 0.16974006593227386, + 0.08347538858652115, + 0.13459047675132751, + 0.08103551715612411 + ], + "credit_map": [ + 0.014506639912724495, + 0.008014040067791939, + 0.011488519608974457, + 0.01375578809529543, + 0.014532790519297123, + 0.008509645238518715, + 0.007123867981135845, + 0.01099527720361948, + 0.01022280752658844, + 0.006057906895875931, + 0.01160823181271553, + 0.014379343949258327, + 0.01203795988112688, + 0.00988789927214384, + 0.01084823813289404, + 0.008959515020251274, + 0.007086614146828651, + 0.004990360699594021, + 0.006929016672074795, + 0.009953009895980358, + 0.011038084514439106, + 0.005011304281651974, + 0.008121971040964127, + 0.006640740670263767, + 0.006224583834409714, + 0.004945169668644667, + 0.006425756029784679, + 0.008233095519244671, + 0.006820825394243002, + 0.004248777404427528, + 0.0058369082398712635, + 0.007173140998929739, + 0.03291906416416168, + 0.014996667392551899, + 0.022526998072862625, + 0.01563752070069313, + 0.01139772217720747, + 0.026706933975219727, + 0.01100923027843237, + 0.01322505995631218, + 0.010561253875494003, + 0.010061394423246384, + 0.021618874743580818, + 0.009578571654856205, + 0.0069219027645885944, + 0.025648778304457664, + 0.007282634265720844, + 0.00845812913030386, + 0.03319874405860901, + 0.016118569299578667, + 0.016430186107754707, + 0.015264422632753849, + 0.020170986652374268, + 0.02122075855731964, + 0.01622978411614895, + 0.031106604263186455, + 0.018849030137062073, + 0.008402490988373756, + 0.008336794562637806, + 0.007807116024196148, + 0.010482348501682281, + 0.011603697203099728, + 0.007237239275127649, + 0.010756667703390121, + 0.012581986375153065, + 0.01504370104521513, + 0.02349928766489029, + 0.018326029181480408, + 0.016762226819992065, + 0.01794213056564331, + 0.008525081910192966, + 0.02191002480685711, + 0.008496856316924095, + 0.007042743265628815, + 0.01776266284286976, + 0.014600618742406368, + 0.00897278729826212, + 0.008045464754104614, + 0.0038389672990888357, + 0.01227541919797659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.06204257160425186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018183089792728424 + ], + "opposite_logprob_tokens": [ + -4.01818323135376 + ], + "logprob_reward": -0.018183089792728424, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018183089792728424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3283339738845825, + 0.09703630208969116, + 0.03820069134235382, + 0.0381193645298481, + 0.092423215508461, + 0.07844211161136627, + 0.05600665137171745, + 0.05126822367310524, + 0.12688302993774414, + 0.09328647702932358 + ], + "credit_map": [ + 0.061989299952983856, + 0.027336103841662407, + 0.07109083235263824, + 0.0623713843524456, + 0.030906865373253822, + 0.03177487850189209, + 0.01816047914326191, + 0.024704109877347946, + 0.021639438346028328, + 0.009429319761693478, + 0.015621802769601345, + 0.015070010907948017, + 0.007292727008461952, + 0.009099459275603294, + 0.00861757155507803, + 0.01026596687734127, + 0.00795187707990408, + 0.0031224244739860296, + 0.005289414897561073, + 0.007171017583459616, + 0.0029626048635691404, + 0.004584657959640026, + 0.004314632620662451, + 0.0028040611650794744, + 0.007116385735571384, + 0.00421228026971221, + 0.004774439614266157, + 0.005499572493135929, + 0.0035915086045861244, + 0.004257493186742067, + 0.004543690476566553, + 0.004123989026993513, + 0.018693584948778152, + 0.009518093429505825, + 0.01198169868439436, + 0.012536443769931793, + 0.009846936911344528, + 0.011325040832161903, + 0.010121412575244904, + 0.00840000156313181, + 0.011469678021967411, + 0.0034527068492025137, + 0.021770235151052475, + 0.009078399278223515, + 0.002528959885239601, + 0.023940522223711014, + 0.002173882909119129, + 0.004027718678116798, + 0.012729051522910595, + 0.005108917597681284, + 0.00655306177213788, + 0.005717090331017971, + 0.005393337458372116, + 0.008168742060661316, + 0.0037912968546152115, + 0.008545146323740482, + 0.010662021115422249, + 0.004233062267303467, + 0.006441238336265087, + 0.005655802320688963, + 0.005820478778332472, + 0.009308707900345325, + 0.0035959489177912474, + 0.005550959147512913, + 0.013565476052463055, + 0.010303681716322899, + 0.03134314715862274, + 0.0271095372736454, + 0.008122263476252556, + 0.019906558096408844, + 0.006576170213520527, + 0.00995619222521782, + 0.011098743416368961, + 0.0074719698168337345, + 0.026557376608252525, + 0.023496655747294426, + 0.005059029906988144, + 0.011323017999529839, + 0.002979945857077837, + 0.005299730692058802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.018183089792728424 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lion with a thick dark brown and golden mane?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lion standing on the edge of a rugged, rocky cliff?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the lion open its mouth in a roar with visible white canines and flaring nostrils?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vast African savannah with an orange, red, and deep purple sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with an overhead shot and then rise steadily to include more sky?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34859827160835266 + ], + "opposite_logprob_tokens": [ + -1.2235982418060303 + ], + "logprob_reward": -0.34859827160835266, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34859827160835266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42541348934173584, + 0.05753675848245621, + 0.04515513777732849, + 0.01858769915997982, + 0.0919388085603714, + 0.11798111349344254, + 0.03968401625752449, + 0.10170020163059235, + 0.030957810580730438, + 0.07104501128196716 + ], + "credit_map": [ + 0.06506714224815369, + 0.04203911870718002, + 0.08450624346733093, + 0.06636693328619003, + 0.06881488859653473, + 0.03943312540650368, + 0.02670786716043949, + 0.03247816488146782, + 0.006359077990055084, + 0.005909343250095844, + 0.007128579542040825, + 0.008504828438162804, + 0.010286517441272736, + 0.004035304766148329, + 0.008223481476306915, + 0.007089630234986544, + 0.00771141005679965, + 0.0030564756598323584, + 0.006597852800041437, + 0.008413443341851234, + 0.003824508748948574, + 0.004350976552814245, + 0.0063244360499084, + 0.004876036662608385, + 0.003012784058228135, + 0.002171358559280634, + 0.002280834363773465, + 0.0025286762975156307, + 0.0019298320403322577, + 0.0022034819703549147, + 0.002725932514294982, + 0.0017347981920465827, + 0.014569289050996304, + 0.01028869766741991, + 0.012806696817278862, + 0.01416451670229435, + 0.011266588233411312, + 0.00814608670771122, + 0.011588411405682564, + 0.009108527563512325, + 0.01614702306687832, + 0.009453408420085907, + 0.025789357721805573, + 0.0127399368211627, + 0.007256025914102793, + 0.026497606188058853, + 0.010030395351350307, + 0.010067365132272243, + 0.01058050338178873, + 0.003660537302494049, + 0.004586143884807825, + 0.003818885423243046, + 0.003767762565985322, + 0.004551696591079235, + 0.0037923597265034914, + 0.004926125518977642, + 0.02243771217763424, + 0.010668279603123665, + 0.008496738038957119, + 0.010836737230420113, + 0.008581613190472126, + 0.014766630716621876, + 0.011555531993508339, + 0.014356957748532295, + 0.006398772355169058, + 0.0021713990718126297, + 0.004795706365257502, + 0.00473090261220932, + 0.0031155182514339685, + 0.0036827418953180313, + 0.0022411399986594915, + 0.003821630496531725, + 0.01005804818123579, + 0.006055021192878485, + 0.007452181074768305, + 0.010598035529255867, + 0.006854821462184191, + 0.006913192104548216, + 0.003948417957872152, + 0.019165288656949997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 7 + ], + "selected_reward": -0.34859827160835266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_tensor([6966, 6966])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_tensor([6966, 6966])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..730dfc0b2364486006a87991d0da44ac4d7457b9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_tensor([6966, 6966])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a tightly composed close-up shot, embodying the aesthetics of the ultimate martial arts hero, a princess reacts in a split second. The camera captures her from a slightly upward angle, emphasizing her extraordinary presence. Her hair is meticulously styled in a classically elegant updo, secured by a simple yet delicate jade hairpin. A few strands of wind-blown black hair cling to her smooth forehead. Her clear, phoenix-like eyes now pierce with a sharp gaze, revealing an unyielding will and unwavering composure in the face of danger. Her face is beautiful, yet her expression is solemn, her lips tightly pursed. She wears a tailored, deep purple silk ensemble, its lapels and cuffs embroidered with intricate silver cloud patterns, shimmering against the hazy moonlight. As an invisible, deadly attack strikes, she swiftly raises her left arm across her chest. From her fair palm, a pure, blazing blue flame erupts, quickly congealing into a translucent, circular shield rimmed with energy runes. When the attack slammed into the center of the shield, the entire scene shook violently. The blue flame shield became the only light source, dyeing her resolute face and the surrounding air with a mysterious and powerful blue halo. Circles of visible energy aftermath violently vibrated and spread out from the impact point, blowing her clothes and hair, creating a tense atmosphere of power and beauty intertwined, full of epic feeling.", + "seed": 29911, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_6966_seed_29911_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009110232256352902, + "vlm_logprob": -0.009110232256352902, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.009110232256352902, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007481059525161982 + ], + "opposite_logprob_tokens": [ + -5.132481098175049 + ], + "logprob_reward": -0.007481059525161982, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.007481059525161982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1132092997431755, + 0.14767517149448395, + 0.08120591193437576, + 0.04604383185505867, + 0.03857188671827316, + 0.10183385014533997, + 0.05344533175230026, + 0.29500749707221985, + 0.05722620338201523, + 0.06578106433153152 + ], + "credit_map": [ + 0.014334941282868385, + 0.009273935109376907, + 0.02387956716120243, + 0.013378893956542015, + 0.012625083327293396, + 0.017688404768705368, + 0.014584879390895367, + 0.007443604525178671, + 0.02575799450278282, + 0.010157087817788124, + 0.03374033421278, + 0.017665181308984756, + 0.014068800024688244, + 0.019659236073493958, + 0.017208609730005264, + 0.009417924098670483, + 0.010384085588157177, + 0.0072973258793354034, + 0.012253804132342339, + 0.012912455014884472, + 0.010414492338895798, + 0.007744851056486368, + 0.010493917390704155, + 0.009704970754683018, + 0.004163248464465141, + 0.0036977073177695274, + 0.009764672257006168, + 0.012262586504220963, + 0.003798716003075242, + 0.005255019757896662, + 0.003524854313582182, + 0.003577029798179865, + 0.004506735596805811, + 0.0036245007067918777, + 0.005506853573024273, + 0.004680874291807413, + 0.004492834210395813, + 0.004808529745787382, + 0.005028962157666683, + 0.0059225973673164845, + 0.0088812792673707, + 0.009724696166813374, + 0.008103554137051105, + 0.017728891223669052, + 0.013679510913789272, + 0.01731259375810623, + 0.009174768812954426, + 0.01722855120897293, + 0.00486174738034606, + 0.007000376004725695, + 0.0066195824183523655, + 0.006850923411548138, + 0.011291506700217724, + 0.004542120732367039, + 0.006211760453879833, + 0.006067316047847271, + 0.036334194242954254, + 0.01713990606367588, + 0.015590238384902477, + 0.010420153848826885, + 0.019234763458371162, + 0.1148882508277893, + 0.023270945996046066, + 0.05812905356287956, + 0.006046494003385305, + 0.005511310882866383, + 0.003550103399902582, + 0.008482443168759346, + 0.009686201810836792, + 0.010143554769456387, + 0.007876625284552574, + 0.00592946819961071, + 0.005773778539150953, + 0.008836719207465649, + 0.005452869459986687, + 0.012939888052642345, + 0.009076341986656189, + 0.007151398342102766, + 0.008840367197990417, + 0.007709703408181667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 0 + ], + "selected_reward": -0.007481059525161982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02062387950718403 + ], + "opposite_logprob_tokens": [ + -3.8956239223480225 + ], + "logprob_reward": -0.02062387950718403, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02062387950718403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16877016425132751, + 0.17992264032363892, + 0.06623613834381104, + 0.03379308059811592, + 0.05881775543093681, + 0.18480029702186584, + 0.19704177975654602, + 0.060077715665102005, + 0.02506501041352749, + 0.025475408881902695 + ], + "credit_map": [ + 0.015432702377438545, + 0.013961292803287506, + 0.03359289467334747, + 0.019978433847427368, + 0.019724030047655106, + 0.028956687077879906, + 0.022876065224409103, + 0.014248074032366276, + 0.01961587183177471, + 0.015495345927774906, + 0.024578392505645752, + 0.021547997370362282, + 0.03120874986052513, + 0.027669843286275864, + 0.02469274401664734, + 0.015113708563148975, + 0.007162628695368767, + 0.004767053294926882, + 0.014501030556857586, + 0.011728715151548386, + 0.005302254110574722, + 0.009711544029414654, + 0.006588430143892765, + 0.006474483758211136, + 0.0033886998426169157, + 0.003926753997802734, + 0.005530062597244978, + 0.0040821111761033535, + 0.003982524387538433, + 0.004958703648298979, + 0.0034549881238490343, + 0.004469235427677631, + 0.009124384261667728, + 0.004233871586620808, + 0.012226877734065056, + 0.004998383112251759, + 0.0053719510324299335, + 0.007530195638537407, + 0.005623471923172474, + 0.009708615951240063, + 0.012667344883084297, + 0.015221543610095978, + 0.01837613247334957, + 0.021130777895450592, + 0.020594650879502296, + 0.04456629976630211, + 0.012592492625117302, + 0.0396510474383831, + 0.024607686325907707, + 0.009567816741764545, + 0.011719465255737305, + 0.010438187047839165, + 0.016107484698295593, + 0.03189856559038162, + 0.07187016308307648, + 0.020832419395446777, + 0.005564906168729067, + 0.003895931411534548, + 0.0069997264072299, + 0.0037072759587317705, + 0.005422006826847792, + 0.018693676218390465, + 0.0046024383045732975, + 0.011191758327186108, + 0.0029411083087325096, + 0.003334097331389785, + 0.002181325340643525, + 0.003674907609820366, + 0.004285934381186962, + 0.0023466963320970535, + 0.003265108447521925, + 0.0030358340591192245, + 0.0025868769735097885, + 0.00297765270806849, + 0.003684484865516424, + 0.0026561494451016188, + 0.004083584528416395, + 0.0038818237371742725, + 0.0018282545497640967, + 0.0037765828892588615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.02062387950718403 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088889092206955 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088889092206955, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088889092206955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08300264179706573, + 0.04236133024096489, + 0.02241058647632599, + 0.019707737490534782, + 0.046350590884685516, + 0.5453451871871948, + 0.07644496858119965, + 0.08792032301425934, + 0.03858166188001633, + 0.037874985486269 + ], + "credit_map": [ + 0.007542163133621216, + 0.006773248314857483, + 0.010513837449252605, + 0.008883189409971237, + 0.012631162069737911, + 0.01134035736322403, + 0.0181868989020586, + 0.007131791673600674, + 0.006692662835121155, + 0.003723914036527276, + 0.007280781865119934, + 0.006030579097568989, + 0.003771501826122403, + 0.00695014139637351, + 0.004438774194568396, + 0.003472973359748721, + 0.0018678223714232445, + 0.0024685333482921124, + 0.002771027386188507, + 0.004199646413326263, + 0.003012615954503417, + 0.0022342437878251076, + 0.0030203356873244047, + 0.0028363624587655067, + 0.0015607727691531181, + 0.0016914469888433814, + 0.0035815013106912374, + 0.004636699333786964, + 0.0017974049551412463, + 0.0024066960904747248, + 0.0018099668668583035, + 0.002223248826339841, + 0.006292811129242182, + 0.002292630495503545, + 0.010499843396246433, + 0.004261913243681192, + 0.00392179423943162, + 0.0062032160349190235, + 0.004479047376662493, + 0.008399332873523235, + 0.03554641827940941, + 0.04366130009293556, + 0.04200437664985657, + 0.07575304061174393, + 0.07172536849975586, + 0.12389560043811798, + 0.045358698815107346, + 0.10740038007497787, + 0.009048332460224628, + 0.005093262530863285, + 0.004653020761907101, + 0.004853231832385063, + 0.00835792999714613, + 0.011060205288231373, + 0.024749405682086945, + 0.008629575371742249, + 0.008269701153039932, + 0.00533490814268589, + 0.007096142042428255, + 0.006892669480293989, + 0.010006961412727833, + 0.016291556879878044, + 0.013149503618478775, + 0.020878875628113747, + 0.002994927577674389, + 0.003707326017320156, + 0.0030413425993174314, + 0.0054132672958076, + 0.0074720801785588264, + 0.004117144271731377, + 0.005559454672038555, + 0.0062761190347373486, + 0.0037455542478710413, + 0.005148143973201513, + 0.0033485263120383024, + 0.006797442678362131, + 0.0047110156156122684, + 0.005154274869710207, + 0.004434056114405394, + 0.004535974934697151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.011088889092206955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017171651124954224 + ], + "opposite_logprob_tokens": [ + -6.3767170906066895 + ], + "logprob_reward": -0.0017171651124954224, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017171651124954224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04694988951086998, + 0.03716547414660454, + 0.01254276279360056, + 0.03884267807006836, + 0.03020850382745266, + 0.27607887983322144, + 0.2642340660095215, + 0.11487235128879547, + 0.15068428218364716, + 0.02842109464108944 + ], + "credit_map": [ + 0.0036186398938298225, + 0.004029182251542807, + 0.006681627128273249, + 0.006841101683676243, + 0.0076063526794314384, + 0.004541249014437199, + 0.01001073606312275, + 0.003620999865233898, + 0.004142573103308678, + 0.0020822195801883936, + 0.008656268008053303, + 0.003188967937603593, + 0.003672196064144373, + 0.009504684247076511, + 0.0036038649268448353, + 0.0023147009778767824, + 0.0015247029950842261, + 0.0010091583244502544, + 0.0016725625609979033, + 0.0015087815700098872, + 0.0016590639716014266, + 0.0023024959955364466, + 0.0013478192267939448, + 0.001518177567049861, + 0.004514422733336687, + 0.003504407126456499, + 0.007513722404837608, + 0.007280788384377956, + 0.00279557635076344, + 0.004831626545637846, + 0.0031866212375462055, + 0.005215509328991175, + 0.003482641652226448, + 0.002317641396075487, + 0.0053757731802761555, + 0.0029018023051321507, + 0.0020449652802199125, + 0.004625089466571808, + 0.002652578055858612, + 0.006808013189584017, + 0.02033398486673832, + 0.023496752604842186, + 0.024313779547810555, + 0.03346284106373787, + 0.03469519317150116, + 0.062345489859580994, + 0.02324744127690792, + 0.054183393716812134, + 0.03581046685576439, + 0.015326798893511295, + 0.015605381689965725, + 0.014987590722739697, + 0.021901581436395645, + 0.03681662306189537, + 0.09380903095006943, + 0.02997659705579281, + 0.013564691878855228, + 0.011416507884860039, + 0.019675498828291893, + 0.012377046048641205, + 0.018033171072602272, + 0.011408372782170773, + 0.01566155068576336, + 0.012735511176288128, + 0.016031650826334953, + 0.016614744439721107, + 0.007941841147840023, + 0.016917552798986435, + 0.0345444530248642, + 0.030737394466996193, + 0.011952636763453484, + 0.01594400405883789, + 0.00248429412022233, + 0.003335892688483, + 0.003696937346830964, + 0.004425563849508762, + 0.004216117784380913, + 0.0026585401501506567, + 0.003174838377162814, + 0.004428909160196781 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 8 + ], + "selected_reward": -0.0017171651124954224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0849430188536644, + 0.05496498942375183, + 0.026022497564554214, + 0.049421463161706924, + 0.02380952797830105, + 0.30612194538116455, + 0.21996639668941498, + 0.10184314101934433, + 0.06641695648431778, + 0.06649010628461838 + ], + "credit_map": [ + 0.0076892864890396595, + 0.006079887039959431, + 0.014191018417477608, + 0.011001941747963428, + 0.010250307619571686, + 0.015805363655090332, + 0.013240199536085129, + 0.006685005035251379, + 0.007311910390853882, + 0.004335941281169653, + 0.008165459148585796, + 0.004777308087795973, + 0.006986961700022221, + 0.012983446940779686, + 0.0059205167926847935, + 0.004483439493924379, + 0.002429947955533862, + 0.0030838134698569775, + 0.0026112685445696115, + 0.003684597322717309, + 0.004394500982016325, + 0.0027525441255420446, + 0.003899670671671629, + 0.0031661514658480883, + 0.0060828449204564095, + 0.003386217635124922, + 0.009999136440455914, + 0.008835476823151112, + 0.0034041316248476505, + 0.007365086581557989, + 0.0040331436321139336, + 0.006315422244369984, + 0.003290679072961211, + 0.0017475408967584372, + 0.00481838546693325, + 0.0014833287568762898, + 0.0017883713589981198, + 0.0033435451332479715, + 0.0026554178912192583, + 0.004682257305830717, + 0.01883186399936676, + 0.02242625504732132, + 0.024438168853521347, + 0.03946767374873161, + 0.034705743193626404, + 0.0774056687951088, + 0.019702579826116562, + 0.06914396584033966, + 0.031359098851680756, + 0.013751531951129436, + 0.012211072258651257, + 0.011798521503806114, + 0.020261188969016075, + 0.027759669348597527, + 0.07715655863285065, + 0.025668742135167122, + 0.009275024756789207, + 0.007184404414147139, + 0.008406272158026695, + 0.0073217665776610374, + 0.012399760074913502, + 0.02882908657193184, + 0.011515258811414242, + 0.016911551356315613, + 0.005721190478652716, + 0.006947857793420553, + 0.005397180560976267, + 0.010519660077989101, + 0.008474822156131268, + 0.010932032018899918, + 0.008015702478587627, + 0.010408512316644192, + 0.006910639349371195, + 0.009677618741989136, + 0.00560803385451436, + 0.009093829430639744, + 0.011179960332810879, + 0.007370795588940382, + 0.010094229131937027, + 0.0065549928694963455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.00464016804471612 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tightly composed close-up shot, embodying the aesthetics of the ultimate martial arts hero, a princess reacts in a split second. The camera captures her from a slightly upward angle, emphasizing her extraordinary presence. Her hair is meticulously styled in a classically elegant updo, secured by a simple yet delicate jade hairpin. A few strands of wind-blown black hair cling to her smooth forehead. Her clear, phoenix-like eyes now pierce with a sharp gaze, revealing an unyielding will and unwavering composure in the face of danger. Her face is beautiful, yet her expression is solemn, her lips tightly pursed. She wears a tailored, deep purple silk ensemble, its lapels and cuffs embroidered with intricate silver cloud patterns, shimmering against the hazy moonlight. As an invisible, deadly attack strikes, she swiftly raises her left arm across her chest. From her fair palm, a pure, blazing blue flame erupts, quickly congealing into a translucent, circular shield rimmed with energy runes. When the attack slammed into the center of the shield, the entire scene shook violently. The blue flame shield became the only light source, dyeing her resolute face and the surrounding air with a mysterious and powerful blue halo. Circles of visible energy aftermath violently vibrated and spread out from the impact point, blowing her clothes and hair, creating a tense atmosphere of power and beauty intertwined, full of epic feeling.", + "seed": 29912, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_6966_seed_29912_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.012761111347936093, + "vlm_logprob": -0.012761111347936093, + "vlm_yes_no_margin": 4.749999952316284, + "vlm_token_credit": -0.012761111347936093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007932942360639572 + ], + "opposite_logprob_tokens": [ + -5.007933139801025 + ], + "logprob_reward": -0.007932942360639572, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007932942360639572, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07934701442718506, + 0.3273133635520935, + 0.07964911311864853, + 0.0693802535533905, + 0.03619800880551338, + 0.07099415361881256, + 0.06443698704242706, + 0.08281640708446503, + 0.1140654906630516, + 0.07579924166202545 + ], + "credit_map": [ + 0.013542959466576576, + 0.005272332578897476, + 0.012457686476409435, + 0.009168057702481747, + 0.008446340449154377, + 0.008756408467888832, + 0.010626272298395634, + 0.011076963506639004, + 0.0926017165184021, + 0.014289976097643375, + 0.13267844915390015, + 0.03320648521184921, + 0.01676665060222149, + 0.00848294235765934, + 0.021209916099905968, + 0.008077219128608704, + 0.010572051629424095, + 0.007091164588928223, + 0.017796751111745834, + 0.016825411468744278, + 0.007293990813195705, + 0.0058736843056976795, + 0.00990154780447483, + 0.004294506274163723, + 0.008284117095172405, + 0.006223004776984453, + 0.006563048809766769, + 0.011835950426757336, + 0.009837680496275425, + 0.006062883418053389, + 0.008333813399076462, + 0.012239756993949413, + 0.005448773968964815, + 0.0037327834870666265, + 0.003736743237823248, + 0.0042048911564052105, + 0.006017714273184538, + 0.004149849992245436, + 0.004587570670992136, + 0.004319681786000729, + 0.008366475813090801, + 0.007969147525727749, + 0.007998976856470108, + 0.005463736597448587, + 0.009851126000285149, + 0.00992036797106266, + 0.013895767740905285, + 0.0075285546481609344, + 0.006648216862231493, + 0.005709976889193058, + 0.014866570942103863, + 0.008393279276788235, + 0.00806623324751854, + 0.006558223161846399, + 0.007559241261333227, + 0.006635239813476801, + 0.011440507136285305, + 0.009816507808864117, + 0.01109646912664175, + 0.009398792870342731, + 0.011136273853480816, + 0.006855604704469442, + 0.012498239986598492, + 0.01057401578873396, + 0.0247197262942791, + 0.006331809796392918, + 0.01145220547914505, + 0.007697915192693472, + 0.006830744910985231, + 0.01051833014935255, + 0.005393501836806536, + 0.041121259331703186, + 0.006863834336400032, + 0.007411842234432697, + 0.010646326467394829, + 0.012541615404188633, + 0.009589523077011108, + 0.008317576721310616, + 0.01082547102123499, + 0.009603051468729973 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.007932942360639572 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02637963369488716 + ], + "opposite_logprob_tokens": [ + -3.6513795852661133 + ], + "logprob_reward": -0.02637963369488716, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02637963369488716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05975557863712311, + 0.1373574286699295, + 0.1176278218626976, + 0.19520264863967896, + 0.07196937501430511, + 0.0682491660118103, + 0.046175938099622726, + 0.16048744320869446, + 0.08939359337091446, + 0.05378105491399765 + ], + "credit_map": [ + 0.007797927595674992, + 0.004904971923679113, + 0.011577890254557133, + 0.010012158192694187, + 0.0065788766369223595, + 0.00458540441468358, + 0.009849491529166698, + 0.0044488562270998955, + 0.02682948112487793, + 0.009930068626999855, + 0.03762708604335785, + 0.016879471018910408, + 0.015545062720775604, + 0.008605566807091236, + 0.01647667959332466, + 0.005464010871946812, + 0.01616494543850422, + 0.011776573956012726, + 0.02296694926917553, + 0.024286141619086266, + 0.014397535473108292, + 0.008958070538938046, + 0.011069954372942448, + 0.008007650263607502, + 0.03357408568263054, + 0.013402002863585949, + 0.03767651319503784, + 0.05881087854504585, + 0.013004756532609463, + 0.007668037433177233, + 0.014253147877752781, + 0.016813213005661964, + 0.009558984078466892, + 0.00547787407413125, + 0.012259473092854023, + 0.01195505727082491, + 0.008006765507161617, + 0.004138966090977192, + 0.00775380851700902, + 0.012818444520235062, + 0.009340955875813961, + 0.005930119659751654, + 0.013141782023012638, + 0.01076910924166441, + 0.00641200365498662, + 0.004599565640091896, + 0.009456918574869633, + 0.008598718792200089, + 0.0041155265644192696, + 0.004771315027028322, + 0.007662900723516941, + 0.006916652899235487, + 0.007141573820263147, + 0.0033165733329951763, + 0.007944222539663315, + 0.004307171329855919, + 0.05209702253341675, + 0.006404379848390818, + 0.06329521536827087, + 0.012323940172791481, + 0.005612345412373543, + 0.00790584459900856, + 0.005100823473185301, + 0.0077478704042732716, + 0.019204044714570045, + 0.00580661790445447, + 0.007431782316416502, + 0.00542808510363102, + 0.01060941070318222, + 0.0073252771981060505, + 0.006946326699107885, + 0.026642052456736565, + 0.005197849590331316, + 0.004837004467844963, + 0.010066845454275608, + 0.006708329543471336, + 0.007439904380589724, + 0.005134806036949158, + 0.008233894594013691, + 0.006162422243505716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.02637963369488716 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792503900825977 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792503900825977, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792503900825977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051947612315416336, + 0.21605680882930756, + 0.09075864404439926, + 0.02663486823439598, + 0.04492177441716194, + 0.08334506303071976, + 0.12690213322639465, + 0.19362083077430725, + 0.11249646544456482, + 0.053315840661525726 + ], + "credit_map": [ + 0.007874912582337856, + 0.004202902317047119, + 0.009144607931375504, + 0.0066658128052949905, + 0.005106356926262379, + 0.004877408035099506, + 0.007668499834835529, + 0.006407112814486027, + 0.059113480150699615, + 0.009814455173909664, + 0.08454997092485428, + 0.023237451910972595, + 0.010432525537908077, + 0.008376097306609154, + 0.013032758608460426, + 0.007500060833990574, + 0.013700948096811771, + 0.009064477868378162, + 0.012654443271458149, + 0.015023721382021904, + 0.010524608194828033, + 0.010689690709114075, + 0.010954965837299824, + 0.008145790547132492, + 0.004511014558374882, + 0.0027494672685861588, + 0.0051613859832286835, + 0.004991517867892981, + 0.003166475100442767, + 0.0017386164981871843, + 0.0023814390879124403, + 0.0019349494250491261, + 0.006453969050198793, + 0.004015395883470774, + 0.009060641750693321, + 0.0074613322503864765, + 0.0048748888075351715, + 0.0048579382710158825, + 0.0038912734016776085, + 0.004306334536522627, + 0.010287545621395111, + 0.006515829358249903, + 0.018356328830122948, + 0.01275727991014719, + 0.009909210726618767, + 0.006060403306037188, + 0.01042167004197836, + 0.009036793373525143, + 0.015288491733372211, + 0.010305359959602356, + 0.03581145778298378, + 0.020651867613196373, + 0.0114029785618186, + 0.009536100551486015, + 0.012590015307068825, + 0.011315863579511642, + 0.05166773498058319, + 0.009606518782675266, + 0.06793991476297379, + 0.01699497178196907, + 0.013759981840848923, + 0.012129570357501507, + 0.010010707192122936, + 0.011511426419019699, + 0.0219222754240036, + 0.006947664078325033, + 0.015386413782835007, + 0.0077168927527964115, + 0.011070618405938148, + 0.008697064593434334, + 0.008203317411243916, + 0.03255221247673035, + 0.005402665585279465, + 0.005606096237897873, + 0.00772192282602191, + 0.008318683132529259, + 0.007506807334721088, + 0.005875238683074713, + 0.007108703721314669, + 0.005775725468993187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.009792503900825977 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015132891712710261 + ], + "opposite_logprob_tokens": [ + -6.501513481140137 + ], + "logprob_reward": -0.0015132891712710261, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015132891712710261, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02404741756618023, + 0.17950385808944702, + 0.03645605593919754, + 0.017419707030057907, + 0.02400495857000351, + 0.013605617918074131, + 0.11911354959011078, + 0.4569607079029083, + 0.08043363690376282, + 0.04845450073480606 + ], + "credit_map": [ + 0.003105384297668934, + 0.0025192387402057648, + 0.0048422846011817455, + 0.003301282413303852, + 0.0032471446320414543, + 0.0016340612201020122, + 0.003961237147450447, + 0.001436783466488123, + 0.057480331510305405, + 0.007997360080480576, + 0.0798906534910202, + 0.016919715330004692, + 0.004013120196759701, + 0.0036770657170563936, + 0.005295052193105221, + 0.004230547696352005, + 0.006365280598402023, + 0.0029277396388351917, + 0.007220289204269648, + 0.01014478225260973, + 0.0030170155223459005, + 0.001335785840637982, + 0.003473585471510887, + 0.001971578225493431, + 0.002210862934589386, + 0.0013144343392923474, + 0.002593911485746503, + 0.002887896029278636, + 0.0024056052789092064, + 0.00140133046079427, + 0.002616752637550235, + 0.0019889126997441053, + 0.003481560153886676, + 0.0018973530968651175, + 0.005606943275779486, + 0.0035942187532782555, + 0.0024432179052382708, + 0.0015377020463347435, + 0.0024594382848590612, + 0.0029845230747014284, + 0.0014689727686345577, + 0.0008832569583319128, + 0.0014547724276781082, + 0.001961233327165246, + 0.0015451550716534257, + 0.0010673890355974436, + 0.0021435897797346115, + 0.0030812486074864864, + 0.014220654964447021, + 0.010615604929625988, + 0.030219092965126038, + 0.020313145592808723, + 0.011883536353707314, + 0.00781123386695981, + 0.014158736914396286, + 0.009891539812088013, + 0.15195469558238983, + 0.017848143354058266, + 0.1839478462934494, + 0.033096808940172195, + 0.01715289056301117, + 0.02263379655778408, + 0.013161661103367805, + 0.017164871096611023, + 0.014076132327318192, + 0.00505022332072258, + 0.008197121322154999, + 0.006652405485510826, + 0.00951080210506916, + 0.0067837839014828205, + 0.006002521142363548, + 0.02416064590215683, + 0.005269818473607302, + 0.004777852445840836, + 0.0053395796567201614, + 0.006619445979595184, + 0.008878227323293686, + 0.004578340798616409, + 0.0049233101308345795, + 0.008067925460636616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.0015132891712710261 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187187612056732 + ], + "opposite_logprob_tokens": [ + -4.018187046051025 + ], + "logprob_reward": -0.018187187612056732, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018187187612056732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02274526096880436, + 0.054104410111904144, + 0.04045749455690384, + 0.0735802873969078, + 0.017601894214749336, + 0.01634693704545498, + 0.10258840769529343, + 0.5060548186302185, + 0.0886390432715416, + 0.07788150757551193 + ], + "credit_map": [ + 0.00377638335339725, + 0.001379710272885859, + 0.0054825409315526485, + 0.003449077717959881, + 0.0023018864449113607, + 0.001130006741732359, + 0.0035272426903247833, + 0.0016984118847176433, + 0.010104474611580372, + 0.00416378490626812, + 0.01354091428220272, + 0.007344595156610012, + 0.007125312462449074, + 0.003193913260474801, + 0.006308927666395903, + 0.002322488697245717, + 0.006054833065718412, + 0.00447511300444603, + 0.005796754267066717, + 0.006908119656145573, + 0.005655097309499979, + 0.0025500778574496508, + 0.007060282398015261, + 0.001957218162715435, + 0.015209757722914219, + 0.0044832369312644005, + 0.023289650678634644, + 0.014283750206232071, + 0.003323672106489539, + 0.003091772785410285, + 0.00527216587215662, + 0.004626276902854443, + 0.002437960123643279, + 0.001842417404986918, + 0.003575156908482313, + 0.0026195538230240345, + 0.0017491996986791492, + 0.001192142954096198, + 0.002241631504148245, + 0.0019438310991972685, + 0.002162380376830697, + 0.0013043760554865003, + 0.0034368312917649746, + 0.0021214839071035385, + 0.002241429639980197, + 0.0013059271732345223, + 0.0018913678359240294, + 0.0018831397173926234, + 0.014223395846784115, + 0.008642914704978466, + 0.028050988912582397, + 0.016793861985206604, + 0.008122964762151241, + 0.00719322357326746, + 0.010170212015509605, + 0.009390845894813538, + 0.1642555445432663, + 0.01769682765007019, + 0.20330345630645752, + 0.03704819083213806, + 0.017127444967627525, + 0.02564087323844433, + 0.016515545547008514, + 0.024466894567012787, + 0.014264475554227829, + 0.005045723170042038, + 0.010267293080687523, + 0.0051581175066530704, + 0.010618259198963642, + 0.00861612893640995, + 0.005784954875707626, + 0.028884097933769226, + 0.008542051538825035, + 0.007994725368916988, + 0.01335285697132349, + 0.010794823989272118, + 0.010465739294886589, + 0.008839046582579613, + 0.009469760581851006, + 0.00842250231653452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.018187187612056732 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tightly composed close-up shot, embodying the aesthetics of the ultimate martial arts hero, a princess reacts in a split second. The camera captures her from a slightly upward angle, emphasizing her extraordinary presence. Her hair is meticulously styled in a classically elegant updo, secured by a simple yet delicate jade hairpin. A few strands of wind-blown black hair cling to her smooth forehead. Her clear, phoenix-like eyes now pierce with a sharp gaze, revealing an unyielding will and unwavering composure in the face of danger. Her face is beautiful, yet her expression is solemn, her lips tightly pursed. She wears a tailored, deep purple silk ensemble, its lapels and cuffs embroidered with intricate silver cloud patterns, shimmering against the hazy moonlight. As an invisible, deadly attack strikes, she swiftly raises her left arm across her chest. From her fair palm, a pure, blazing blue flame erupts, quickly congealing into a translucent, circular shield rimmed with energy runes. When the attack slammed into the center of the shield, the entire scene shook violently. The blue flame shield became the only light source, dyeing her resolute face and the surrounding air with a mysterious and powerful blue halo. Circles of visible energy aftermath violently vibrated and spread out from the impact point, blowing her clothes and hair, creating a tense atmosphere of power and beauty intertwined, full of epic feeling.", + "seed": 29913, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_6966_seed_29913_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010679548978805542, + "vlm_logprob": -0.010679548978805542, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.010679548978805542, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009993401356041431 + ], + "opposite_logprob_tokens": [ + -4.759993553161621 + ], + "logprob_reward": -0.009993401356041431, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009993401356041431, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.044031351804733276, + 0.07305735349655151, + 0.5907467007637024, + 0.06723333150148392, + 0.03749791905283928, + 0.05647584795951843, + 0.022436408326029778, + 0.02371358312666416, + 0.041867323219776154, + 0.04294019564986229 + ], + "credit_map": [ + 0.00459221750497818, + 0.004291431978344917, + 0.004356159828603268, + 0.005096865817904472, + 0.006215206813067198, + 0.007223331835120916, + 0.006196265108883381, + 0.00605987012386322, + 0.00888702180236578, + 0.009934545494616032, + 0.0085292998701334, + 0.006753240246325731, + 0.010541759431362152, + 0.011650852859020233, + 0.00902259536087513, + 0.007738037034869194, + 0.06951644271612167, + 0.02818405069410801, + 0.1212962195277214, + 0.0747603252530098, + 0.03055824339389801, + 0.061481133103370667, + 0.07992425560951233, + 0.12502603232860565, + 0.008635029196739197, + 0.006388512905687094, + 0.005773099604994059, + 0.008686583489179611, + 0.006238230038434267, + 0.007636857684701681, + 0.01814797706902027, + 0.0057270340621471405, + 0.005641030613332987, + 0.0067629157565534115, + 0.005571865942329168, + 0.0032188049517571926, + 0.0037410666700452566, + 0.004867101088166237, + 0.004530380945652723, + 0.003164751222357154, + 0.006942893844097853, + 0.006059017963707447, + 0.005940757691860199, + 0.007585483603179455, + 0.005251223221421242, + 0.008226122707128525, + 0.009569717571139336, + 0.0069006322883069515, + 0.0025271368212997913, + 0.0033849726896733046, + 0.002179094823077321, + 0.0032945997081696987, + 0.003264518454670906, + 0.002867939416319132, + 0.00316677731461823, + 0.0017513680504634976, + 0.0026578903198242188, + 0.004290055483579636, + 0.0026600502897053957, + 0.00309752463363111, + 0.0033539184369146824, + 0.0031528521794825792, + 0.001945520518347621, + 0.0025557735934853554, + 0.004460802301764488, + 0.006996260490268469, + 0.0028861595783382654, + 0.006118220277130604, + 0.004625589121133089, + 0.004462701268494129, + 0.006252385675907135, + 0.006065205205231905, + 0.007832854986190796, + 0.0051107569597661495, + 0.005538797006011009, + 0.004153871908783913, + 0.003658868605270982, + 0.008526547811925411, + 0.0041720131412148476, + 0.003946484997868538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.009993401356041431 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01611153595149517 + ], + "opposite_logprob_tokens": [ + -4.141111373901367 + ], + "logprob_reward": -0.01611153595149517, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01611153595149517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053431689739227295, + 0.03613520786166191, + 0.652893602848053, + 0.04191512241959572, + 0.04034710302948952, + 0.04780452325940132, + 0.03339909762144089, + 0.02169918827712536, + 0.021772298961877823, + 0.050602152943611145 + ], + "credit_map": [ + 0.004828886594623327, + 0.006661959923803806, + 0.0057195196859538555, + 0.00735869538038969, + 0.008621105924248695, + 0.004941517487168312, + 0.008503688499331474, + 0.006796316709369421, + 0.004967516288161278, + 0.003989144694060087, + 0.00660885451361537, + 0.004571118857711554, + 0.0032376369927078485, + 0.004723835736513138, + 0.003994189202785492, + 0.004042909946292639, + 0.07818865031003952, + 0.031039820984005928, + 0.13944916427135468, + 0.07672671228647232, + 0.03354004770517349, + 0.07119693607091904, + 0.07615651190280914, + 0.1465957760810852, + 0.005496903322637081, + 0.0038285423070192337, + 0.004259343259036541, + 0.005251007154583931, + 0.004200231749564409, + 0.004444903694093227, + 0.010759423486888409, + 0.0036747632548213005, + 0.00493151368573308, + 0.007431086618453264, + 0.005451648961752653, + 0.0042143817991018295, + 0.005320696625858545, + 0.004461853299289942, + 0.004052915144711733, + 0.004483007825911045, + 0.0056722527369856834, + 0.008540084585547447, + 0.004607239738106728, + 0.007137189153581858, + 0.004145107697695494, + 0.006522461771965027, + 0.007180124055594206, + 0.004000060725957155, + 0.0036143241450190544, + 0.0072102793492376804, + 0.004365080967545509, + 0.004884856287389994, + 0.0035312981344759464, + 0.0036784198600798845, + 0.003289798041805625, + 0.0028250431641936302, + 0.0027612242847681046, + 0.0024428197648376226, + 0.002563013229519129, + 0.0025841675233095884, + 0.0018847895553335547, + 0.004632194992154837, + 0.0016136345220729709, + 0.0032173420768231153, + 0.0021173045970499516, + 0.003524255473166704, + 0.0018925451440736651, + 0.0034067153465002775, + 0.002766579855233431, + 0.0019186391727998853, + 0.004240044392645359, + 0.0019062150968238711, + 0.008302799426019192, + 0.0071828290820121765, + 0.005345988553017378, + 0.005251760594546795, + 0.0046966541558504105, + 0.0076247286051511765, + 0.006804570555686951, + 0.005392823833972216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.01611153595149517 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551246210932732 + ], + "opposite_logprob_tokens": [ + -4.387551307678223 + ], + "logprob_reward": -0.012551246210932732, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551246210932732, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09511128813028336, + 0.09325896948575974, + 0.4003921151161194, + 0.06919433176517487, + 0.07868088781833649, + 0.05297238752245903, + 0.025527460500597954, + 0.03777509182691574, + 0.03874308988451958, + 0.10834433883428574 + ], + "credit_map": [ + 0.009605268947780132, + 0.013792145997285843, + 0.011109190061688423, + 0.011886687017977238, + 0.01587316207587719, + 0.0129720913246274, + 0.010762615129351616, + 0.009110131300985813, + 0.010908044874668121, + 0.011272995732724667, + 0.010526887141168118, + 0.011150741949677467, + 0.012158717028796673, + 0.01652262732386589, + 0.010141800157725811, + 0.010577158071100712, + 0.04819884151220322, + 0.01836855709552765, + 0.08043485134840012, + 0.04788893461227417, + 0.02203845977783203, + 0.04634413123130798, + 0.0517144612967968, + 0.08540387451648712, + 0.00820203684270382, + 0.006025085225701332, + 0.009477338753640652, + 0.009435467422008514, + 0.006327655166387558, + 0.006999088451266289, + 0.016413066536188126, + 0.006314597558230162, + 0.01052738819271326, + 0.005835939664393663, + 0.011846719309687614, + 0.008402870036661625, + 0.009427525103092194, + 0.010291961021721363, + 0.01310470700263977, + 0.009243778884410858, + 0.007408087141811848, + 0.006551312282681465, + 0.006451655179262161, + 0.005883111152797937, + 0.006637231912463903, + 0.006033277604728937, + 0.008856674656271935, + 0.005151042714715004, + 0.0030603324994444847, + 0.0023137046955525875, + 0.003537182230502367, + 0.003432608675211668, + 0.0029277894645929337, + 0.0031749859917908907, + 0.0032828273251652718, + 0.0037980300839990377, + 0.004815542139112949, + 0.00393506558611989, + 0.004618640057742596, + 0.004884746391326189, + 0.004179443698376417, + 0.006746374536305666, + 0.003125210525467992, + 0.005470068193972111, + 0.0035124102141708136, + 0.006642324849963188, + 0.0034748248290270567, + 0.007368730381131172, + 0.0033578453585505486, + 0.002999961143359542, + 0.00640399195253849, + 0.004983000922948122, + 0.015022477135062218, + 0.008229049853980541, + 0.009960480965673923, + 0.010742752812802792, + 0.007300182245671749, + 0.011845055036246777, + 0.030018027871847153, + 0.015226300805807114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.012551246210932732 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021998509764671326 + ], + "opposite_logprob_tokens": [ + -6.127199649810791 + ], + "logprob_reward": -0.0021998509764671326, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021998509764671326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05004741623997688, + 0.02148236706852913, + 0.3976442217826843, + 0.0729491263628006, + 0.15797023475170135, + 0.055607471615076065, + 0.04263510927557945, + 0.05795837938785553, + 0.09376201778650284, + 0.049943696707487106 + ], + "credit_map": [ + 0.004307233262807131, + 0.005220284219831228, + 0.007017339114099741, + 0.005802719853818417, + 0.007612253539264202, + 0.004784951917827129, + 0.007604238111525774, + 0.007698392495512962, + 0.002342588035389781, + 0.003487038891762495, + 0.002704276703298092, + 0.003301086137071252, + 0.0027158986777067184, + 0.002087569097056985, + 0.002468082122504711, + 0.0023758255410939455, + 0.0473460853099823, + 0.018739724531769753, + 0.08403605222702026, + 0.045768532902002335, + 0.02045346423983574, + 0.04242682829499245, + 0.048955436795949936, + 0.08991806954145432, + 0.009090338833630085, + 0.01117844507098198, + 0.008388870395720005, + 0.010512682609260082, + 0.006916928105056286, + 0.008210630156099796, + 0.01042805053293705, + 0.008223176933825016, + 0.016586270183324814, + 0.018276996910572052, + 0.02400221675634384, + 0.01651826500892639, + 0.020919842645525932, + 0.020299460738897324, + 0.01859031617641449, + 0.022776853293180466, + 0.005760464817285538, + 0.006879139691591263, + 0.004803257528692484, + 0.006493276450783014, + 0.0077260746620595455, + 0.007862681522965431, + 0.010510914959013462, + 0.005571658723056316, + 0.0045089940540492535, + 0.006565894931554794, + 0.0052693379111588, + 0.004891172982752323, + 0.005488939117640257, + 0.005534799303859472, + 0.00593836372718215, + 0.004437603056430817, + 0.006824633106589317, + 0.0069109536707401276, + 0.005101997405290604, + 0.0063191675581038, + 0.009251298382878304, + 0.006863342598080635, + 0.007984384894371033, + 0.008702592924237251, + 0.011703597381711006, + 0.012331251055002213, + 0.008298204280436039, + 0.010554956272244453, + 0.010406666435301304, + 0.008018291555345058, + 0.01967564783990383, + 0.012773390859365463, + 0.007992659695446491, + 0.004439756274223328, + 0.00527625298127532, + 0.004666340537369251, + 0.0033117085695266724, + 0.0063663446344435215, + 0.011904955841600895, + 0.005985673516988754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.0021998509764671326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541710399091244 + ], + "opposite_logprob_tokens": [ + -4.387541770935059 + ], + "logprob_reward": -0.012541710399091244, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541710399091244, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04141182824969292, + 0.03066321276128292, + 0.28832104802131653, + 0.038170747458934784, + 0.11634291708469391, + 0.07515323162078857, + 0.06368827819824219, + 0.21130229532718658, + 0.07677728682756424, + 0.05816909298300743 + ], + "credit_map": [ + 0.004284752998501062, + 0.006176236551254988, + 0.0056475140154361725, + 0.0033105118200182915, + 0.007265937048941851, + 0.006160041783004999, + 0.004733058158308268, + 0.0038337791338562965, + 0.0031295293010771275, + 0.005535257514566183, + 0.003924857825040817, + 0.004834209568798542, + 0.0032384733203798532, + 0.0031686052680015564, + 0.003740707878023386, + 0.003091573715209961, + 0.03340913727879524, + 0.014486303552985191, + 0.06010625883936882, + 0.03275579214096069, + 0.014160175807774067, + 0.030243517830967903, + 0.04013918712735176, + 0.06302067637443542, + 0.004903074353933334, + 0.005559965502470732, + 0.004783084616065025, + 0.004830765537917614, + 0.00379379210062325, + 0.004356639925390482, + 0.006129146087914705, + 0.003814281662926078, + 0.007663874886929989, + 0.015530379489064217, + 0.010166746564209461, + 0.015484336763620377, + 0.014514053240418434, + 0.01667398400604725, + 0.02658722922205925, + 0.009722321294248104, + 0.010797076858580112, + 0.010335300117731094, + 0.010838834568858147, + 0.012981808744370937, + 0.006596495397388935, + 0.007996817119419575, + 0.007555846590548754, + 0.008051054552197456, + 0.007693254388868809, + 0.008539373055100441, + 0.008276202715933323, + 0.006380896549671888, + 0.00882619246840477, + 0.008436406031250954, + 0.006595451384782791, + 0.008940503932535648, + 0.023327890783548355, + 0.025451872497797012, + 0.023287924006581306, + 0.021523522213101387, + 0.029036084190011024, + 0.027449799701571465, + 0.027244051918387413, + 0.033981163054704666, + 0.00899062305688858, + 0.009893004782497883, + 0.00613820506259799, + 0.011661243624985218, + 0.009171241894364357, + 0.00721837067976594, + 0.014672604389488697, + 0.009031998924911022, + 0.009288459084928036, + 0.006581134162843227, + 0.0059244041331112385, + 0.006826841272413731, + 0.005645093508064747, + 0.009903059341013432, + 0.008454612456262112, + 0.0055454932153224945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 4 + ], + "selected_reward": -0.012541710399091244 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a tightly composed close-up shot, embodying the aesthetics of the ultimate martial arts hero, a princess reacts in a split second. The camera captures her from a slightly upward angle, emphasizing her extraordinary presence. Her hair is meticulously styled in a classically elegant updo, secured by a simple yet delicate jade hairpin. A few strands of wind-blown black hair cling to her smooth forehead. Her clear, phoenix-like eyes now pierce with a sharp gaze, revealing an unyielding will and unwavering composure in the face of danger. Her face is beautiful, yet her expression is solemn, her lips tightly pursed. She wears a tailored, deep purple silk ensemble, its lapels and cuffs embroidered with intricate silver cloud patterns, shimmering against the hazy moonlight. As an invisible, deadly attack strikes, she swiftly raises her left arm across her chest. From her fair palm, a pure, blazing blue flame erupts, quickly congealing into a translucent, circular shield rimmed with energy runes. When the attack slammed into the center of the shield, the entire scene shook violently. The blue flame shield became the only light source, dyeing her resolute face and the surrounding air with a mysterious and powerful blue halo. Circles of visible energy aftermath violently vibrated and spread out from the impact point, blowing her clothes and hair, creating a tense atmosphere of power and beauty intertwined, full of epic feeling.", + "seed": 29914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_12/index_6966_seed_29914_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.007640821137465536, + "vlm_logprob": -0.007640821137465536, + "vlm_yes_no_margin": 5.2, + "vlm_token_credit": -0.007640821137465536, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006164584308862686 + ], + "opposite_logprob_tokens": [ + -5.38116455078125 + ], + "logprob_reward": -0.006164584308862686, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.006164584308862686, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10025269538164139, + 0.12282779812812805, + 0.057063281536102295, + 0.24913418292999268, + 0.08340977877378464, + 0.11495733261108398, + 0.04869897663593292, + 0.08337660878896713, + 0.08107440173625946, + 0.059204958379268646 + ], + "credit_map": [ + 0.014807750470936298, + 0.005645460914820433, + 0.017352785915136337, + 0.010935919359326363, + 0.007378163281828165, + 0.014144218526780605, + 0.010426347143948078, + 0.019562041386961937, + 0.021151069551706314, + 0.008037927560508251, + 0.017772993072867393, + 0.01789802871644497, + 0.009558453224599361, + 0.0171163622289896, + 0.013509659096598625, + 0.017783306539058685, + 0.008247234858572483, + 0.005328478757292032, + 0.007278421428054571, + 0.007244917098432779, + 0.005106080323457718, + 0.006950573995709419, + 0.008042445406317711, + 0.008865129202604294, + 0.03339623287320137, + 0.011334187351167202, + 0.07275617867708206, + 0.018716175109148026, + 0.012041790410876274, + 0.04085627943277359, + 0.022618787363171577, + 0.0374145545065403, + 0.007218706887215376, + 0.0088325971737504, + 0.005025891587138176, + 0.013742653653025627, + 0.012495324946939945, + 0.010363385081291199, + 0.013081599958240986, + 0.012649615295231342, + 0.015346983447670937, + 0.007696166634559631, + 0.01840399205684662, + 0.009794950485229492, + 0.014203127473592758, + 0.0173232089728117, + 0.013238751329481602, + 0.018950149416923523, + 0.005689483135938644, + 0.00285940314643085, + 0.003477566409856081, + 0.005303301382809877, + 0.008367854170501232, + 0.006422586273401976, + 0.010713538154959679, + 0.0058652483858168125, + 0.013891909271478653, + 0.006359532941132784, + 0.01648627407848835, + 0.00891016237437725, + 0.008165370672941208, + 0.011382096447050571, + 0.008219989016652107, + 0.00996127724647522, + 0.009418196976184845, + 0.00654342956840992, + 0.00969689805060625, + 0.0074547333642840385, + 0.016815127804875374, + 0.01174917072057724, + 0.009044068865478039, + 0.010352772660553455, + 0.00563201354816556, + 0.005955107510089874, + 0.007577843964099884, + 0.008640523999929428, + 0.008251582272350788, + 0.009069310501217842, + 0.007159565109759569, + 0.00691900635138154 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.006164584308862686 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012568079866468906 + ], + "opposite_logprob_tokens": [ + -4.38756799697876 + ], + "logprob_reward": -0.012568079866468906, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012568079866468906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19137516617774963, + 0.17773965001106262, + 0.10092703253030777, + 0.13262106478214264, + 0.10899711400270462, + 0.05715242400765419, + 0.03812972456216812, + 0.07441194355487823, + 0.037043243646621704, + 0.08160262554883957 + ], + "credit_map": [ + 0.026834560558199883, + 0.017142845317721367, + 0.03249464929103851, + 0.02396245114505291, + 0.01753055490553379, + 0.023857546970248222, + 0.022989125922322273, + 0.026563439518213272, + 0.030775224789977074, + 0.011027715168893337, + 0.03574807196855545, + 0.0245184488594532, + 0.013218349777162075, + 0.022022215649485588, + 0.019188489764928818, + 0.02124112658202648, + 0.015006211586296558, + 0.008309919387102127, + 0.011568845249712467, + 0.014703474007546902, + 0.012034700252115726, + 0.012211333960294724, + 0.01052280142903328, + 0.01656973920762539, + 0.01834237575531006, + 0.005251014139503241, + 0.034977786242961884, + 0.010230976156890392, + 0.00524714682251215, + 0.023644596338272095, + 0.016536632552742958, + 0.018390526995062828, + 0.012270570732653141, + 0.0071547990664839745, + 0.00990760512650013, + 0.013177202083170414, + 0.009494521655142307, + 0.01612161099910736, + 0.01608324982225895, + 0.024787558242678642, + 0.0059897941537201405, + 0.004822571761906147, + 0.013418919406831264, + 0.006494713481515646, + 0.006035217083990574, + 0.008304553106427193, + 0.005943441763520241, + 0.006143213715404272, + 0.0050238859839737415, + 0.0030068764463067055, + 0.006113674025982618, + 0.0050417473539710045, + 0.006459956988692284, + 0.004272722639143467, + 0.004621646832674742, + 0.003589215688407421, + 0.01012768317013979, + 0.007207464426755905, + 0.007345359772443771, + 0.006846082396805286, + 0.010852648876607418, + 0.010188538581132889, + 0.012346038594841957, + 0.009498129598796368, + 0.003976888954639435, + 0.0033675841987133026, + 0.0038109614979475737, + 0.004911254160106182, + 0.005741413217037916, + 0.00490432046353817, + 0.005924148019403219, + 0.004406672902405262, + 0.009967065416276455, + 0.007106703240424395, + 0.012401491403579712, + 0.010481172241270542, + 0.011025811545550823, + 0.01024370826780796, + 0.010121924802660942, + 0.010254749096930027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.012568079866468906 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008647022768855095 + ], + "opposite_logprob_tokens": [ + -4.7586469650268555 + ], + "logprob_reward": -0.008647022768855095, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008647022768855095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13007356226444244, + 0.07735929638147354, + 0.06460919231176376, + 0.1834402084350586, + 0.07949056476354599, + 0.09208939224481583, + 0.20141901075839996, + 0.06326020509004593, + 0.032730527222156525, + 0.07552807778120041 + ], + "credit_map": [ + 0.015348974615335464, + 0.011654707603156567, + 0.023210112005472183, + 0.01675674319267273, + 0.010565362870693207, + 0.015835946425795555, + 0.01813824102282524, + 0.018563464283943176, + 0.01277428213506937, + 0.005213355645537376, + 0.014823953621089458, + 0.007544072810560465, + 0.004369321744889021, + 0.011783694848418236, + 0.008307289332151413, + 0.012543325312435627, + 0.008113933727145195, + 0.005945554934442043, + 0.005497538484632969, + 0.007353790104389191, + 0.011480641551315784, + 0.0075003947131335735, + 0.008798502385616302, + 0.009918827563524246, + 0.027670076116919518, + 0.010402083396911621, + 0.035582538694143295, + 0.01606745645403862, + 0.008961732499301434, + 0.029218779876828194, + 0.019938673824071884, + 0.03559885546565056, + 0.008989375084638596, + 0.005557196214795113, + 0.0060678874142467976, + 0.019710374996066093, + 0.007099709939211607, + 0.01099770050495863, + 0.010104510933160782, + 0.010963804088532925, + 0.009276830591261387, + 0.007749328389763832, + 0.012435113079845905, + 0.00900090578943491, + 0.013129188679158688, + 0.013009173795580864, + 0.011166978627443314, + 0.016321862116456032, + 0.019643863663077354, + 0.015035616233944893, + 0.017200153321027756, + 0.024574976414442062, + 0.038998719304800034, + 0.025676168501377106, + 0.04073185473680496, + 0.0195576474070549, + 0.006433994043618441, + 0.006626125890761614, + 0.006976693402975798, + 0.008453681133687496, + 0.010416621342301369, + 0.008593630976974964, + 0.008830569684505463, + 0.006928886752575636, + 0.004101642873138189, + 0.002762238960713148, + 0.003575726877897978, + 0.004713105037808418, + 0.004423860926181078, + 0.0045946864411234856, + 0.0041229878552258015, + 0.004436274990439415, + 0.008780497126281261, + 0.007200242485851049, + 0.009299091063439846, + 0.008706795983016491, + 0.011245738714933395, + 0.009541775099933147, + 0.010625802911818027, + 0.010128128342330456 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.008647022768855095 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001042656716890633 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.001042656716890633, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.001042656716890633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06390257924795151, + 0.06384406983852386, + 0.06609432399272919, + 0.09841914474964142, + 0.13059687614440918, + 0.11075477302074432, + 0.2007673680782318, + 0.06598644703626633, + 0.14707805216312408, + 0.052556362003088 + ], + "credit_map": [ + 0.009233739227056503, + 0.004643022082746029, + 0.01072787307202816, + 0.005473562050610781, + 0.006162825506180525, + 0.008345565758645535, + 0.006931740790605545, + 0.012384247034788132, + 0.008172070607542992, + 0.007704863790422678, + 0.007666322402656078, + 0.008537922985851765, + 0.006241374649107456, + 0.008010663092136383, + 0.009022895246744156, + 0.00848795473575592, + 0.010867591015994549, + 0.004710718058049679, + 0.008245160803198814, + 0.004957209341228008, + 0.003637544112280011, + 0.010011924430727959, + 0.008488625288009644, + 0.015175550244748592, + 0.012379162944853306, + 0.005232928320765495, + 0.023070281371474266, + 0.008989331312477589, + 0.005377208348363638, + 0.01446614135056734, + 0.017657360062003136, + 0.011246726848185062, + 0.0184571985155344, + 0.009987886995077133, + 0.013674468733370304, + 0.031000688672065735, + 0.013458731584250927, + 0.01748599112033844, + 0.014406319707632065, + 0.012125595472753048, + 0.017004836350679398, + 0.008064000867307186, + 0.017084728926420212, + 0.010187518782913685, + 0.016878802329301834, + 0.015406962484121323, + 0.015272018499672413, + 0.01085589174181223, + 0.022384537383913994, + 0.013108443468809128, + 0.022816892713308334, + 0.01803683675825596, + 0.04015560820698738, + 0.02801189012825489, + 0.03412605822086334, + 0.022127101197838783, + 0.0090408343821764, + 0.006575131323188543, + 0.012938289903104305, + 0.006975554395467043, + 0.008313770405948162, + 0.008679629303514957, + 0.006832797545939684, + 0.006630441639572382, + 0.015155528672039509, + 0.015155883505940437, + 0.014002171345055103, + 0.021745430305600166, + 0.023395227268338203, + 0.02037489041686058, + 0.02147442288696766, + 0.01577448472380638, + 0.0055626556277275085, + 0.003963011782616377, + 0.007279818877577782, + 0.006221255287528038, + 0.007995808497071266, + 0.007437021005898714, + 0.007109468337148428, + 0.006987324450165033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.001042656716890633 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tightly composed close-up of a princess?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera angle slightly upward, and does the character wear a deep purple silk ensemble with silver cloud embroidery?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character have black hair styled in an updo with a jade hairpin, and does she raise her left arm to form a blue flame shield?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene show a violent impact where the blue shield becomes the primary light source, casting a blue halo on the character's face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there visible energy circles vibrating and spreading from the impact point while the character's clothes and hair are blown by the force?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09625142812728882, + 0.0736771896481514, + 0.040427885949611664, + 0.06928705424070358, + 0.14413097500801086, + 0.18136337399482727, + 0.1748037189245224, + 0.07014486938714981, + 0.03532976284623146, + 0.11458372324705124 + ], + "credit_map": [ + 0.012070823460817337, + 0.008363754488527775, + 0.01726156286895275, + 0.011306182481348515, + 0.007683371659368277, + 0.012652129866182804, + 0.011005992069840431, + 0.015907615423202515, + 0.014235825277864933, + 0.008079751394689083, + 0.013797411695122719, + 0.007344045676290989, + 0.004287411458790302, + 0.01122625358402729, + 0.006781928241252899, + 0.007924559526145458, + 0.006934610661119223, + 0.0027546619530767202, + 0.005192380864173174, + 0.00351884076371789, + 0.003946793265640736, + 0.005879141855984926, + 0.004440783988684416, + 0.007760670967400074, + 0.007880989462137222, + 0.004079104866832495, + 0.013620587065815926, + 0.006318508181720972, + 0.004164368845522404, + 0.009817605838179588, + 0.012793991714715958, + 0.01061189454048872, + 0.016882551833987236, + 0.010881083086133003, + 0.009191678836941719, + 0.025636959820985794, + 0.01525172684341669, + 0.01918535679578781, + 0.02572796121239662, + 0.021373648196458817, + 0.019944846630096436, + 0.014220164157450199, + 0.028806716203689575, + 0.015850892290472984, + 0.019828129559755325, + 0.027811627835035324, + 0.019684579223394394, + 0.03521641716361046, + 0.021290894597768784, + 0.012320230714976788, + 0.018901027739048004, + 0.023750746622681618, + 0.03185088559985161, + 0.023746296763420105, + 0.02728564850986004, + 0.01565798930823803, + 0.006741629447788, + 0.006035939324647188, + 0.006215133238583803, + 0.007625368889421225, + 0.01066454965621233, + 0.009792222641408443, + 0.010828706435859203, + 0.012241325341165066, + 0.003712540725246072, + 0.003414818784222007, + 0.0035615945234894753, + 0.004756580572575331, + 0.006053693126887083, + 0.004663648549467325, + 0.005242085549980402, + 0.003924803342670202, + 0.015590173192322254, + 0.007768114097416401, + 0.02426169626414776, + 0.014874307438731194, + 0.013964494690299034, + 0.013641719706356525, + 0.012061567977070808, + 0.012421649880707264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.009781762026250362 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_tensor([7625, 7625])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_tensor([7625, 7625])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..284336217c5e6529787952297e16a9afe777a448 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_tensor([7625, 7625])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a scene reminiscent of a child's drawing, replete with childish brushstrokes and a flattened perspective, the camera begins with a dynamic, upward-looking angle, following a plump piglet, painted all over in vibrant pink, with simple black polka-dot eyes and a spiral tail. From the top of a vast haystack constructed from rough, thick yellow crayon lines, it glides rapidly downward, a trail of white, dynamic lines representing its speed in its wake. As it reaches the edge of the haystack, it leaps with all its might, performing a fluid, dramatic mid-air flip, its four tiny hooves stomping joyfully in the air. At the apex of its flip, the camera begins a rapid-fire cycle of focus shifts and jump cuts: one moment a close-up of the piglet in mid-air, the next abruptly cutting to the lower reaches of the haystack, where an audience of animals, also composed of simple geometric shapes—a wide-eyed cow, a clucking hen, and a fluffy sheep—raise their heads and open their mouths, cheering in the purest of ways. The rapid-fire editing is accompanied by a violent flickering of the screen, as if from a bright white flashbulb, each flicker fading the saturated colors of the frame into white. The entire scene is illuminated by strings of giant, brightly colored spheres suspended against a deep blue night sky, casting a warm, dappled glow and creating a party atmosphere that is both chaotic and joyful, brimming with childlike innocence and energy.", + "seed": 31622, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31622_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07763414485380053, + "vlm_logprob": -0.07763414485380053, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.07763414485380053, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002869419287890196 + ], + "opposite_logprob_tokens": [ + -6.252869606018066 + ], + "logprob_reward": -0.002869419287890196, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.002869419287890196, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13480240106582642, + 0.0889294371008873, + 0.1184990182518959, + 0.06298281997442245, + 0.0702168270945549, + 0.0931372418999672, + 0.06311418861150742, + 0.11989478766918182, + 0.20143146812915802, + 0.04699181392788887 + ], + "credit_map": [ + 0.019735926762223244, + 0.018571168184280396, + 0.016858592629432678, + 0.019781289622187614, + 0.013274087570607662, + 0.02300063706934452, + 0.009411279112100601, + 0.014169408939778805, + 0.01249311026185751, + 0.01126578263938427, + 0.009274184703826904, + 0.013579734601080418, + 0.013402365148067474, + 0.013543955981731415, + 0.007994710467755795, + 0.007375591434538364, + 0.004935096018016338, + 0.008757473900914192, + 0.021702377125620842, + 0.015552892349660397, + 0.010432973504066467, + 0.005714474245905876, + 0.0244011040776968, + 0.027002621442079544, + 0.004601570311933756, + 0.010870974510908127, + 0.013411888852715492, + 0.007288777735084295, + 0.0071341125294566154, + 0.005918907001614571, + 0.007710101082921028, + 0.006046484224498272, + 0.007650140672922134, + 0.013739596121013165, + 0.0091254822909832, + 0.006282927934080362, + 0.009390785358846188, + 0.00682896189391613, + 0.005570318084210157, + 0.011628608219325542, + 0.013295146636664867, + 0.010484492406249046, + 0.012273913249373436, + 0.01885111816227436, + 0.01000642403960228, + 0.008198566734790802, + 0.012085177935659885, + 0.007942394353449345, + 0.006279083900153637, + 0.015096288174390793, + 0.005759172607213259, + 0.006161761004477739, + 0.00848438497632742, + 0.00840062741190195, + 0.005283340811729431, + 0.007649530190974474, + 0.012298714369535446, + 0.04400114342570305, + 0.009299623779952526, + 0.009898685850203037, + 0.013577647507190704, + 0.012258094735443592, + 0.005965008400380611, + 0.012595868669450283, + 0.02354300580918789, + 0.06061369925737381, + 0.011587781831622124, + 0.014235066249966621, + 0.012764180079102516, + 0.05668327584862709, + 0.006994133815169334, + 0.015010323375463486, + 0.005953447427600622, + 0.007229385431855917, + 0.006122015416622162, + 0.0053215837106108665, + 0.005209307186305523, + 0.006262295413762331, + 0.005935098044574261, + 0.004958678036928177 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.002869419287890196 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896454632282257 + ], + "opposite_logprob_tokens": [ + -2.5789644718170166 + ], + "logprob_reward": -0.07896454632282257, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896454632282257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.191188782453537, + 0.06720688194036484, + 0.1456003040075302, + 0.10199864953756332, + 0.07016599178314209, + 0.1233772486448288, + 0.090976282954216, + 0.08669862151145935, + 0.059670355170965195, + 0.06311686336994171 + ], + "credit_map": [ + 0.02914542518556118, + 0.024053996428847313, + 0.02592805027961731, + 0.024176115170121193, + 0.016040127724409103, + 0.03533797338604927, + 0.01377050206065178, + 0.022736605256795883, + 0.00805404968559742, + 0.008404615335166454, + 0.013981879688799381, + 0.008911612443625927, + 0.005993491038680077, + 0.007132813334465027, + 0.006114891730248928, + 0.008613534271717072, + 0.008393704891204834, + 0.01254831813275814, + 0.022705689072608948, + 0.0230268482118845, + 0.015061961486935616, + 0.00799369253218174, + 0.018931569531559944, + 0.03693852946162224, + 0.010554387234151363, + 0.018926968798041344, + 0.010982080362737179, + 0.011215278878808022, + 0.016398319974541664, + 0.011930161155760288, + 0.01246688049286604, + 0.009524578228592873, + 0.00866182241588831, + 0.014772133901715279, + 0.008324460126459599, + 0.007813932374119759, + 0.009283025749027729, + 0.008573328144848347, + 0.00615277886390686, + 0.00658451160416007, + 0.02496659941971302, + 0.013262991793453693, + 0.01645446941256523, + 0.0152580002322793, + 0.013017878867685795, + 0.012769386172294617, + 0.013993429020047188, + 0.013654501177370548, + 0.008873450569808483, + 0.016921408474445343, + 0.008213057182729244, + 0.01026070211082697, + 0.01613776385784149, + 0.013896845281124115, + 0.009240670129656792, + 0.0074323867447674274, + 0.009630807675421238, + 0.029708752408623695, + 0.007806372828781605, + 0.006054625380784273, + 0.008081166073679924, + 0.015240395441651344, + 0.0034268919844180346, + 0.006749601103365421, + 0.005157615523785353, + 0.01328294351696968, + 0.006714776158332825, + 0.009199140593409538, + 0.005543573759496212, + 0.00922352448105812, + 0.005798057187348604, + 0.004750724416226149, + 0.007839146070182323, + 0.009995509870350361, + 0.008882444351911545, + 0.00820375420153141, + 0.005955786444246769, + 0.008520347066223621, + 0.006432898808270693, + 0.007286977954208851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.07896454632282257 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336986780166626 + ], + "opposite_logprob_tokens": [ + -3.408698558807373 + ], + "logprob_reward": -0.0336986780166626, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336986780166626, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1271725744009018, + 0.07660993933677673, + 0.07970598340034485, + 0.1965920627117157, + 0.07895895838737488, + 0.07471968978643417, + 0.10775324702262878, + 0.12621314823627472, + 0.05582344904541969, + 0.07645092904567719 + ], + "credit_map": [ + 0.017695065587759018, + 0.016046013683080673, + 0.016328204423189163, + 0.015325481072068214, + 0.013554830104112625, + 0.024662107229232788, + 0.01056442130357027, + 0.012996450066566467, + 0.009019427932798862, + 0.007728981785476208, + 0.016811713576316833, + 0.010336966253817081, + 0.006989523768424988, + 0.008279567584395409, + 0.005440657492727041, + 0.012003096751868725, + 0.0047926832921803, + 0.008363615721464157, + 0.009974361397325993, + 0.011167185381054878, + 0.008629859425127506, + 0.004516756162047386, + 0.01622917503118515, + 0.016032345592975616, + 0.01725332997739315, + 0.04016474634408951, + 0.021572502329945564, + 0.02211722917854786, + 0.035978611558675766, + 0.01857461780309677, + 0.024886514991521835, + 0.01604451984167099, + 0.00930874515324831, + 0.010911037214100361, + 0.009016858413815498, + 0.007960581220686436, + 0.007983355782926083, + 0.01128409057855606, + 0.007236463017761707, + 0.015257826074957848, + 0.013276497833430767, + 0.008082079701125622, + 0.009737754240632057, + 0.011441386304795742, + 0.010997879318892956, + 0.004820703063160181, + 0.008274984546005726, + 0.008088408038020134, + 0.009897422045469284, + 0.02202681079506874, + 0.010200342163443565, + 0.01396134216338396, + 0.020523644983768463, + 0.013436942361295223, + 0.009785790927708149, + 0.0079209478572011, + 0.012875320389866829, + 0.045454446226358414, + 0.011829700320959091, + 0.010457174852490425, + 0.014573164284229279, + 0.012114313431084156, + 0.005492487456649542, + 0.013416540808975697, + 0.005773615092039108, + 0.01176249235868454, + 0.006729788146913052, + 0.007179801352322102, + 0.005194629076868296, + 0.009111398831009865, + 0.0052275387570261955, + 0.004844182170927525, + 0.01133161224424839, + 0.014467980712652206, + 0.008843989111483097, + 0.007360382936894894, + 0.007588332053273916, + 0.013424107804894447, + 0.005747820250689983, + 0.00768670067191124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.0336986780166626 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25205764174461365 + ], + "opposite_logprob_tokens": [ + -1.502057671546936 + ], + "logprob_reward": -0.25205764174461365, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25205764174461365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062393996864557266, + 0.07066570222377777, + 0.05640903487801552, + 0.10565538704395294, + 0.08940044045448303, + 0.0891762524843216, + 0.13344629108905792, + 0.1382545828819275, + 0.16802093386650085, + 0.08657742291688919 + ], + "credit_map": [ + 0.007819024845957756, + 0.007450076751410961, + 0.007522525265812874, + 0.008619367145001888, + 0.00679803965613246, + 0.012300625443458557, + 0.005777168553322554, + 0.006107167806476355, + 0.011074005626142025, + 0.0073842923156917095, + 0.015433224849402905, + 0.006260322872549295, + 0.006546567659825087, + 0.00972861796617508, + 0.004946551751345396, + 0.009292121045291424, + 0.0038844835944473743, + 0.0034387826453894377, + 0.011486702598631382, + 0.005324567202478647, + 0.003760731080546975, + 0.004733132664114237, + 0.012783171609044075, + 0.010997465811669827, + 0.008790522813796997, + 0.022557003423571587, + 0.013075326569378376, + 0.011612929403781891, + 0.01925252005457878, + 0.008551213890314102, + 0.012737508863210678, + 0.009078373201191425, + 0.010050174780189991, + 0.032936159521341324, + 0.007818706333637238, + 0.006053310353308916, + 0.00943707674741745, + 0.011326486244797707, + 0.0045217666774988174, + 0.0072567579336464405, + 0.013907588087022305, + 0.00952411163598299, + 0.012851549312472343, + 0.013737894594669342, + 0.010527304373681545, + 0.011390523985028267, + 0.00972859188914299, + 0.007508688140660524, + 0.015157440677285194, + 0.022568223997950554, + 0.015202848240733147, + 0.015447653830051422, + 0.023105574771761894, + 0.019110381603240967, + 0.010936040431261063, + 0.011918123811483383, + 0.013847899623215199, + 0.05250753089785576, + 0.010473071597516537, + 0.010849134996533394, + 0.015861161053180695, + 0.014426160603761673, + 0.0061848219484090805, + 0.01410481333732605, + 0.018769636750221252, + 0.05144299194216728, + 0.010447884909808636, + 0.01372772827744484, + 0.010685119777917862, + 0.04374616593122482, + 0.00714905746281147, + 0.012052339501678944, + 0.013054716400802135, + 0.013111741282045841, + 0.010545449331402779, + 0.01101645641028881, + 0.010923663154244423, + 0.00934002548456192, + 0.009508909657597542, + 0.00907646119594574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.25205764174461365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020580438897013664 + ], + "opposite_logprob_tokens": [ + -3.895580530166626 + ], + "logprob_reward": -0.020580438897013664, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020580438897013664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07461937516927719, + 0.052085548639297485, + 0.09186145663261414, + 0.0802416056394577, + 0.16456885635852814, + 0.15811769664287567, + 0.1994946002960205, + 0.053099311888217926, + 0.11241338402032852, + 0.01349816657602787 + ], + "credit_map": [ + 0.01095238234847784, + 0.008123496547341347, + 0.014086521230638027, + 0.006721293553709984, + 0.007792606949806213, + 0.013308539055287838, + 0.005072619765996933, + 0.008561917580664158, + 0.010616835206747055, + 0.0044991085305809975, + 0.0065736351534724236, + 0.004256183747202158, + 0.0038282230962067842, + 0.011432692408561707, + 0.003675900399684906, + 0.007202968001365662, + 0.00680800573900342, + 0.004724700469523668, + 0.019335633143782616, + 0.007381592411547899, + 0.005862374324351549, + 0.009568799287080765, + 0.019491517916321754, + 0.018688829615712166, + 0.006861411966383457, + 0.013675837777554989, + 0.008987686596810818, + 0.009537308476865292, + 0.013688236474990845, + 0.007408510893583298, + 0.012506411410868168, + 0.0075762043707072735, + 0.017544660717248917, + 0.04381890594959259, + 0.015539043582975864, + 0.01383910421282053, + 0.01722579635679722, + 0.018668746575713158, + 0.010988258756697178, + 0.026944322511553764, + 0.02449602633714676, + 0.012669511139392853, + 0.032609064131975174, + 0.019193122163414955, + 0.01253773458302021, + 0.02121649868786335, + 0.010829528793692589, + 0.02456621080636978, + 0.022835593670606613, + 0.04291843995451927, + 0.020694810897111893, + 0.021600667387247086, + 0.024089710786938667, + 0.03100520372390747, + 0.015213037841022015, + 0.021137135103344917, + 0.007160530425608158, + 0.005607899744063616, + 0.008802739903330803, + 0.004597095772624016, + 0.004041089676320553, + 0.009542478248476982, + 0.0063850367441773415, + 0.006962439510971308, + 0.012016753666102886, + 0.028398700058460236, + 0.007875760085880756, + 0.008460515178740025, + 0.00832996517419815, + 0.030451832339167595, + 0.009226549416780472, + 0.007653300650417805, + 0.0018177902093157172, + 0.0027625439688563347, + 0.0017380559584125876, + 0.0012999882455915213, + 0.0017370437271893024, + 0.0014384695095941424, + 0.0012193385045975447, + 0.0014849368017166853 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 5 + ], + "selected_reward": -0.020580438897013664 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene reminiscent of a child's drawing, replete with childish brushstrokes and a flattened perspective, the camera begins with a dynamic, upward-looking angle, following a plump piglet, painted all over in vibrant pink, with simple black polka-dot eyes and a spiral tail. From the top of a vast haystack constructed from rough, thick yellow crayon lines, it glides rapidly downward, a trail of white, dynamic lines representing its speed in its wake. As it reaches the edge of the haystack, it leaps with all its might, performing a fluid, dramatic mid-air flip, its four tiny hooves stomping joyfully in the air. At the apex of its flip, the camera begins a rapid-fire cycle of focus shifts and jump cuts: one moment a close-up of the piglet in mid-air, the next abruptly cutting to the lower reaches of the haystack, where an audience of animals, also composed of simple geometric shapes—a wide-eyed cow, a clucking hen, and a fluffy sheep—raise their heads and open their mouths, cheering in the purest of ways. The rapid-fire editing is accompanied by a violent flickering of the screen, as if from a bright white flashbulb, each flicker fading the saturated colors of the frame into white. The entire scene is illuminated by strings of giant, brightly colored spheres suspended against a deep blue night sky, casting a warm, dappled glow and creating a party atmosphere that is both chaotic and joyful, brimming with childlike innocence and energy.", + "seed": 31623, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31623_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.061551318317651746, + "vlm_logprob": -0.061551318317651746, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.061551318317651746, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004168274812400341 + ], + "opposite_logprob_tokens": [ + -5.879168510437012 + ], + "logprob_reward": -0.004168274812400341, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004168274812400341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0896751806139946, + 0.034297745674848557, + 0.04592037573456764, + 0.03230646252632141, + 0.11463188380002975, + 0.09625344723463058, + 0.11311514675617218, + 0.059048574417829514, + 0.22215119004249573, + 0.19260001182556152 + ], + "credit_map": [ + 0.006072186399251223, + 0.011191231198608875, + 0.022427184507250786, + 0.007524783723056316, + 0.008754391223192215, + 0.010296676307916641, + 0.008120069280266762, + 0.015288658440113068, + 0.0025632292963564396, + 0.0030990734230726957, + 0.003991183824837208, + 0.004346521105617285, + 0.005327859427779913, + 0.005694732069969177, + 0.0040189046412706375, + 0.00525624118745327, + 0.0029759316239506006, + 0.004074661061167717, + 0.0070185912773013115, + 0.005210984498262405, + 0.00974211748689413, + 0.00542259868234396, + 0.005411853548139334, + 0.006063643377274275, + 0.0032329789828509092, + 0.006016191560775042, + 0.0030543645843863487, + 0.005520207807421684, + 0.0026991460472345352, + 0.004317029379308224, + 0.004030449315905571, + 0.0034360960125923157, + 0.02126799151301384, + 0.0188653152436018, + 0.009327853098511696, + 0.012544067576527596, + 0.014459644444286823, + 0.013853616081178188, + 0.012467044405639172, + 0.01184635329991579, + 0.017370328307151794, + 0.009377330541610718, + 0.005961187649518251, + 0.008305259048938751, + 0.008322526700794697, + 0.013332616537809372, + 0.021855458617210388, + 0.011728744953870773, + 0.011648288927972317, + 0.01807953417301178, + 0.00457032909616828, + 0.014763819985091686, + 0.015961291268467903, + 0.020936520770192146, + 0.020164690911769867, + 0.006990675814449787, + 0.008064123801887035, + 0.007721928879618645, + 0.0038582419510930777, + 0.008454273454844952, + 0.010367728769779205, + 0.007187454961240292, + 0.008154945448040962, + 0.005239875987172127, + 0.04455403983592987, + 0.029618369415402412, + 0.021144671365618706, + 0.015558702871203423, + 0.03986015543341637, + 0.03558998927474022, + 0.020990317687392235, + 0.014834929257631302, + 0.014071324840188026, + 0.016937483102083206, + 0.0054845125414431095, + 0.005724161863327026, + 0.07550564408302307, + 0.06207525357604027, + 0.006354639306664467, + 0.0064469873905181885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 4 + ], + "selected_reward": -0.004168274812400341 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278413355350494 + ], + "opposite_logprob_tokens": [ + -2.0177841186523438 + ], + "logprob_reward": -0.14278413355350494, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278413355350494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06635045260190964, + 0.09585387259721756, + 0.08202099055051804, + 0.07273125648498535, + 0.05972396209836006, + 0.0333375483751297, + 0.13506311178207397, + 0.11811979860067368, + 0.2712837755680084, + 0.06551526486873627 + ], + "credit_map": [ + 0.003999256528913975, + 0.013208866119384766, + 0.008857550099492073, + 0.008847031742334366, + 0.008920140564441681, + 0.006748378276824951, + 0.005671046674251556, + 0.010098176077008247, + 0.004900842439383268, + 0.01125253178179264, + 0.014615006744861603, + 0.016944637522101402, + 0.019834483042359352, + 0.012246036902070045, + 0.009435949847102165, + 0.006624372210353613, + 0.0052791680209338665, + 0.007667952217161655, + 0.012736346572637558, + 0.007938168942928314, + 0.013319858349859715, + 0.010068329982459545, + 0.009914182126522064, + 0.015096972696483135, + 0.006176568567752838, + 0.009728645905852318, + 0.016282828524708748, + 0.007731200195848942, + 0.007057224866002798, + 0.01121189072728157, + 0.005404982715845108, + 0.009137902408838272, + 0.009599421173334122, + 0.008342738263309002, + 0.005824380088597536, + 0.00901559367775917, + 0.005857956595718861, + 0.007511461619287729, + 0.008382146246731281, + 0.005190260708332062, + 0.006350122857838869, + 0.00460173049941659, + 0.002133383182808757, + 0.003115094965323806, + 0.0037150399293750525, + 0.004588172771036625, + 0.005529683083295822, + 0.0033043164294213057, + 0.026969609782099724, + 0.01749960146844387, + 0.004639102146029472, + 0.00816004816442728, + 0.016414372250437737, + 0.04819970205426216, + 0.0072259544394910336, + 0.005954698193818331, + 0.015085524879395962, + 0.01721620187163353, + 0.010009943507611752, + 0.015183418989181519, + 0.018197210505604744, + 0.013421972282230854, + 0.016545990481972694, + 0.012459526769816875, + 0.05227677896618843, + 0.03794305399060249, + 0.025516847148537636, + 0.02469543181359768, + 0.04210960492491722, + 0.04396684840321541, + 0.025137169286608696, + 0.019638042896986008, + 0.010165542364120483, + 0.009770669974386692, + 0.0047435821034014225, + 0.007724029012024403, + 0.009103238582611084, + 0.012481212615966797, + 0.006104537285864353, + 0.005422441754490137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.14278413355350494 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369913622736931 + ], + "opposite_logprob_tokens": [ + -3.4086990356445312 + ], + "logprob_reward": -0.03369913622736931, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369913622736931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07184195518493652, + 0.04496856406331062, + 0.07675804942846298, + 0.024472080171108246, + 0.09737414866685867, + 0.12468589842319489, + 0.06173601374030113, + 0.07467041164636612, + 0.23350805044174194, + 0.18998482823371887 + ], + "credit_map": [ + 0.007164896000176668, + 0.009027160704135895, + 0.007642889861017466, + 0.006292429752647877, + 0.006902805529534817, + 0.009765715338289738, + 0.009579765610396862, + 0.015466291457414627, + 0.0029328614473342896, + 0.004687280859798193, + 0.004760914482176304, + 0.005158969201147556, + 0.008245954290032387, + 0.007148635108023882, + 0.005561395548284054, + 0.006472554057836533, + 0.0044719744473695755, + 0.006245926953852177, + 0.012821773998439312, + 0.010367645882070065, + 0.013613875955343246, + 0.008846540935337543, + 0.00935764517635107, + 0.011032662354409695, + 0.0018604876240715384, + 0.0028843055479228497, + 0.0030020622070878744, + 0.003879738273099065, + 0.0035427312832325697, + 0.0029255806002765894, + 0.0029592406935989857, + 0.0034179335925728083, + 0.016228633001446724, + 0.01767941750586033, + 0.007528592366725206, + 0.013269015587866306, + 0.010568964295089245, + 0.012547187507152557, + 0.012155829928815365, + 0.0073965187184512615, + 0.021625950932502747, + 0.01382310502231121, + 0.008317217230796814, + 0.010995186865329742, + 0.012085939757525921, + 0.016380643472075462, + 0.026525281369686127, + 0.014932570047676563, + 0.008675857447087765, + 0.012722309678792953, + 0.0035843413788825274, + 0.005034280940890312, + 0.007980765774846077, + 0.008473057299852371, + 0.007669561076909304, + 0.007595841772854328, + 0.012112009339034557, + 0.010756018571555614, + 0.006207657977938652, + 0.008195417001843452, + 0.011905805207788944, + 0.014450986869633198, + 0.005258353427052498, + 0.005784160923212767, + 0.04780380800366402, + 0.031873978674411774, + 0.02096962369978428, + 0.017604881897568703, + 0.039041146636009216, + 0.03890902176499367, + 0.02123250998556614, + 0.0160730741918087, + 0.013053108006715775, + 0.015560047701001167, + 0.006593520753085613, + 0.010158066637814045, + 0.06299655139446259, + 0.06546924263238907, + 0.007979916408658028, + 0.008174360729753971 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.03369913622736931 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.112911157310009 + ], + "opposite_logprob_tokens": [ + -2.2379112243652344 + ], + "logprob_reward": -0.112911157310009, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.112911157310009, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04561373591423035, + 0.035752344876527786, + 0.020944660529494286, + 0.014845219440758228, + 0.10345923900604248, + 0.06498494744300842, + 0.11137546598911285, + 0.06650081276893616, + 0.4308408498764038, + 0.10568270087242126 + ], + "credit_map": [ + 0.005511225201189518, + 0.006847847253084183, + 0.004531444516032934, + 0.005509186070412397, + 0.005322162061929703, + 0.0033471377100795507, + 0.0050697349943220615, + 0.009474998340010643, + 0.0020680278539657593, + 0.004527570679783821, + 0.00412598205730319, + 0.005339051131159067, + 0.006751617416739464, + 0.00516800582408905, + 0.003780325409024954, + 0.003991765435785055, + 0.0010227721650153399, + 0.0018660661298781633, + 0.0029405432287603617, + 0.0032925454434007406, + 0.004575757775455713, + 0.00254066102206707, + 0.002872761804610491, + 0.0018335526110604405, + 0.001282371231354773, + 0.0017029711743816733, + 0.0021685839165002108, + 0.0018172094132751226, + 0.0020868233405053616, + 0.002278425032272935, + 0.0016672435449436307, + 0.0018415922531858087, + 0.019259246066212654, + 0.016735538840293884, + 0.008446448482573032, + 0.010779237374663353, + 0.013700203038752079, + 0.01357591524720192, + 0.01034046895802021, + 0.010622179135680199, + 0.012147689238190651, + 0.007778100203722715, + 0.0037150559946894646, + 0.00635373592376709, + 0.005887391045689583, + 0.008684732019901276, + 0.012158384546637535, + 0.008259856142103672, + 0.01803257316350937, + 0.014775280840694904, + 0.006069863215088844, + 0.007673923391848803, + 0.019190389662981033, + 0.02960227243602276, + 0.006922220811247826, + 0.009108946658670902, + 0.008319714106619358, + 0.01198015920817852, + 0.004971397574990988, + 0.011076600290834904, + 0.007752721197903156, + 0.007349961902946234, + 0.009545582346618176, + 0.005504675209522247, + 0.08323445916175842, + 0.060150016099214554, + 0.037971995770931244, + 0.03564484417438507, + 0.07058977335691452, + 0.0705292820930481, + 0.042419981211423874, + 0.030300533398985863, + 0.014776792377233505, + 0.014454158022999763, + 0.006926453672349453, + 0.013877230696380138, + 0.010463357903063297, + 0.03052450530230999, + 0.007505093235522509, + 0.007155110593885183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.112911157310009 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193889684975147 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014193889684975147, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193889684975147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08351580798625946, + 0.038171201944351196, + 0.03254484757781029, + 0.017615819349884987, + 0.15997709333896637, + 0.029353810474276543, + 0.015830930322408676, + 0.03237174078822136, + 0.5022850632667542, + 0.08833370357751846 + ], + "credit_map": [ + 0.008414700627326965, + 0.006393324118107557, + 0.023524459451436996, + 0.007746128831058741, + 0.004738065414130688, + 0.004532571416348219, + 0.006662917323410511, + 0.021503645926713943, + 0.0019147973507642746, + 0.0042203557677567005, + 0.0047491793520748615, + 0.005050064064562321, + 0.004591848235577345, + 0.005583315622061491, + 0.005780722945928574, + 0.006280918139964342, + 0.0018119853921234608, + 0.002684904495254159, + 0.004744797479361296, + 0.004026111215353012, + 0.005928255617618561, + 0.0035925565753132105, + 0.004217514768242836, + 0.005538723431527615, + 0.0019441924523562193, + 0.001954051200300455, + 0.0023640047293156385, + 0.0030098464339971542, + 0.001755500677973032, + 0.0023042955435812473, + 0.002097301883623004, + 0.0021866278257220984, + 0.027189668267965317, + 0.02707306109368801, + 0.012058417312800884, + 0.018878750503063202, + 0.021354693919420242, + 0.02109183371067047, + 0.018232787027955055, + 0.01409787219017744, + 0.005310436710715294, + 0.003214211668819189, + 0.0014524284051731229, + 0.002310252981260419, + 0.0027655025478452444, + 0.00426096748560667, + 0.0067247916013002396, + 0.003315218025818467, + 0.002680857665836811, + 0.00294622709043324, + 0.0008843302493914962, + 0.0013113118475303054, + 0.002069842303171754, + 0.002754871267825365, + 0.001765656634233892, + 0.0014178331475704908, + 0.00413139071315527, + 0.006472473964095116, + 0.0020254277624189854, + 0.004477395210415125, + 0.004929812625050545, + 0.004320248495787382, + 0.0031762158032506704, + 0.0028387759812176228, + 0.09235220402479172, + 0.07274588197469711, + 0.049585260450839996, + 0.04155811294913292, + 0.07772912830114365, + 0.0796443447470665, + 0.05234450101852417, + 0.036325644701719284, + 0.00649512792006135, + 0.00810821820050478, + 0.003261271398514509, + 0.004848533309996128, + 0.03123885951936245, + 0.02642112970352173, + 0.003848303109407425, + 0.004112257156521082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.014193889684975147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene reminiscent of a child's drawing, replete with childish brushstrokes and a flattened perspective, the camera begins with a dynamic, upward-looking angle, following a plump piglet, painted all over in vibrant pink, with simple black polka-dot eyes and a spiral tail. From the top of a vast haystack constructed from rough, thick yellow crayon lines, it glides rapidly downward, a trail of white, dynamic lines representing its speed in its wake. As it reaches the edge of the haystack, it leaps with all its might, performing a fluid, dramatic mid-air flip, its four tiny hooves stomping joyfully in the air. At the apex of its flip, the camera begins a rapid-fire cycle of focus shifts and jump cuts: one moment a close-up of the piglet in mid-air, the next abruptly cutting to the lower reaches of the haystack, where an audience of animals, also composed of simple geometric shapes—a wide-eyed cow, a clucking hen, and a fluffy sheep—raise their heads and open their mouths, cheering in the purest of ways. The rapid-fire editing is accompanied by a violent flickering of the screen, as if from a bright white flashbulb, each flicker fading the saturated colors of the frame into white. The entire scene is illuminated by strings of giant, brightly colored spheres suspended against a deep blue night sky, casting a warm, dappled glow and creating a party atmosphere that is both chaotic and joyful, brimming with childlike innocence and energy.", + "seed": 31624, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31624_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03597155027091503, + "vlm_logprob": -0.03597155027091503, + "vlm_yes_no_margin": 3.849999952316284, + "vlm_token_credit": -0.03597155027091503, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004910314455628395 + ], + "opposite_logprob_tokens": [ + -5.504910469055176 + ], + "logprob_reward": -0.004910314455628395, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004910314455628395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08146434277296066, + 0.10040727257728577, + 0.11589246988296509, + 0.13356520235538483, + 0.11306390911340714, + 0.1116257905960083, + 0.1492381989955902, + 0.03767775371670723, + 0.0826273262500763, + 0.07443775236606598 + ], + "credit_map": [ + 0.006591059733182192, + 0.014375695027410984, + 0.005666316021233797, + 0.01125187799334526, + 0.011174147948622704, + 0.009736135601997375, + 0.013609228655695915, + 0.009059876203536987, + 0.010068204253911972, + 0.01024380512535572, + 0.011617359705269337, + 0.010902106761932373, + 0.009603935293853283, + 0.013282095082104206, + 0.011887817643582821, + 0.022801946848630905, + 0.010909334756433964, + 0.019958684220910072, + 0.014372054487466812, + 0.012805512174963951, + 0.015594257973134518, + 0.01602182351052761, + 0.0143785635009408, + 0.011852235533297062, + 0.019091244786977768, + 0.008824747987091541, + 0.023740237578749657, + 0.009244979359209538, + 0.009601254016160965, + 0.03371162340044975, + 0.007869189605116844, + 0.021481921896338463, + 0.018516743555665016, + 0.016563335433602333, + 0.011792809702455997, + 0.014662196859717369, + 0.015374132432043552, + 0.01567440666258335, + 0.009606922045350075, + 0.010873346589505672, + 0.0150078684091568, + 0.024044817313551903, + 0.008281112648546696, + 0.009742839261889458, + 0.016551459208130836, + 0.012000098824501038, + 0.013423353433609009, + 0.012574232183396816, + 0.014814909547567368, + 0.01821104623377323, + 0.011092673987150192, + 0.030185235664248466, + 0.019148604944348335, + 0.020406974479556084, + 0.019427720457315445, + 0.015951024368405342, + 0.004375799559056759, + 0.004401534795761108, + 0.003492177464067936, + 0.005783651024103165, + 0.005262086633592844, + 0.004371450282633305, + 0.0047073024325072765, + 0.005283745937049389, + 0.009137901477515697, + 0.01715034991502762, + 0.005083299241960049, + 0.011234156787395477, + 0.011918840929865837, + 0.012846582569181919, + 0.009233721531927586, + 0.006022467743605375, + 0.010786289349198341, + 0.011183785274624825, + 0.00586533872410655, + 0.012602382339537144, + 0.007634828798472881, + 0.0064020403660833836, + 0.01162287499755621, + 0.008340205065906048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.004910314455628395 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290295422077179 + ], + "opposite_logprob_tokens": [ + -2.237902879714966 + ], + "logprob_reward": -0.11290295422077179, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290295422077179, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08428622037172318, + 0.13633845746517181, + 0.07753153890371323, + 0.18208351731300354, + 0.12284892797470093, + 0.08497139811515808, + 0.162864089012146, + 0.03744048997759819, + 0.08691752701997757, + 0.02471783012151718 + ], + "credit_map": [ + 0.007183994632214308, + 0.009830160066485405, + 0.007545458618551493, + 0.011505118571221828, + 0.01338470820337534, + 0.010134668089449406, + 0.01463439129292965, + 0.010067716240882874, + 0.012911776080727577, + 0.016426363959908485, + 0.014005951583385468, + 0.013595838099718094, + 0.015907904133200645, + 0.0182429701089859, + 0.012174573726952076, + 0.033073075115680695, + 0.006212011445313692, + 0.014836587011814117, + 0.008970149792730808, + 0.010157044045627117, + 0.01140793040394783, + 0.008753875270485878, + 0.009156089276075363, + 0.008037851192057133, + 0.03512435033917427, + 0.009787455201148987, + 0.02588174305856228, + 0.012086649425327778, + 0.013345650397241116, + 0.028170859441161156, + 0.01060605701059103, + 0.04708074778318405, + 0.017800869420170784, + 0.01197595801204443, + 0.02708195336163044, + 0.013567985035479069, + 0.007882020436227322, + 0.0076112146489322186, + 0.017955593764781952, + 0.01897333562374115, + 0.013217604719102383, + 0.009329753927886486, + 0.0043556587770581245, + 0.008500060066580772, + 0.007586158346384764, + 0.021949172019958496, + 0.009251374751329422, + 0.010781614109873772, + 0.013535814359784126, + 0.01862815022468567, + 0.011203157715499401, + 0.039158161729574203, + 0.02542726881802082, + 0.015062463469803333, + 0.024790311232209206, + 0.015058755874633789, + 0.0050984895788133144, + 0.00495640654116869, + 0.0030851562041789293, + 0.006770555395632982, + 0.004197603557258844, + 0.004813354462385178, + 0.004552160389721394, + 0.00396676454693079, + 0.016422705724835396, + 0.013267353177070618, + 0.004677248187363148, + 0.013999984599649906, + 0.00827076192945242, + 0.015128307044506073, + 0.009088324382901192, + 0.006062848027795553, + 0.0031273607164621353, + 0.0029751360416412354, + 0.0022846299689263105, + 0.0038382706698030233, + 0.003766265930607915, + 0.0025943347718566656, + 0.002994575072079897, + 0.0031372560188174248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 1 + ], + "selected_reward": -0.11290295422077179 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581139251589775 + ], + "opposite_logprob_tokens": [ + -3.8955812454223633 + ], + "logprob_reward": -0.020581139251589775, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581139251589775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08196496963500977, + 0.16669288277626038, + 0.06471003592014313, + 0.06392497569322586, + 0.0971473678946495, + 0.05324279144406319, + 0.21545392274856567, + 0.025427261367440224, + 0.11876997351646423, + 0.1126658096909523 + ], + "credit_map": [ + 0.007608609274029732, + 0.011695942841470242, + 0.007459086365997791, + 0.013219023123383522, + 0.011195411905646324, + 0.008572738617658615, + 0.013871708884835243, + 0.008342447690665722, + 0.013789411634206772, + 0.017825279384851456, + 0.02353198267519474, + 0.016892777755856514, + 0.015836775302886963, + 0.021189803257584572, + 0.016363052651286125, + 0.041263800114393234, + 0.008129633031785488, + 0.008262056857347488, + 0.008333699777722359, + 0.0069674598053097725, + 0.008685487322509289, + 0.007623712066560984, + 0.006727715488523245, + 0.009980271570384502, + 0.0087441923096776, + 0.0046044886112213135, + 0.009063825942575932, + 0.004984172061085701, + 0.005595140624791384, + 0.008528182283043861, + 0.005571005400270224, + 0.016833970323204994, + 0.012455619871616364, + 0.014141056686639786, + 0.017213204875588417, + 0.010964818298816681, + 0.008406508713960648, + 0.00740438885986805, + 0.0159196387976408, + 0.010642127133905888, + 0.007093899883329868, + 0.00734602939337492, + 0.0034464604686945677, + 0.006020521279424429, + 0.004597393795847893, + 0.008012673817574978, + 0.009821350686252117, + 0.006904462818056345, + 0.021664170548319817, + 0.02664029411971569, + 0.016962196677923203, + 0.046022769063711166, + 0.028206970542669296, + 0.02554374933242798, + 0.029878022149205208, + 0.020535742864012718, + 0.00396362878382206, + 0.0032642141450196505, + 0.0027156295254826546, + 0.0030872358474880457, + 0.0035858312621712685, + 0.003313377033919096, + 0.0025998970959335566, + 0.002897450001910329, + 0.025252923369407654, + 0.01467463281005621, + 0.0056367963552474976, + 0.016979306936264038, + 0.010870746336877346, + 0.024647923186421394, + 0.013112680986523628, + 0.007594952825456858, + 0.014475786127150059, + 0.012067551724612713, + 0.010386756621301174, + 0.023415379226207733, + 0.009385570883750916, + 0.009158359840512276, + 0.02089194394648075, + 0.012884458526968956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.020581139251589775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018179578706622124 + ], + "opposite_logprob_tokens": [ + -4.018179416656494 + ], + "logprob_reward": -0.018179578706622124, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018179578706622124, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1340123862028122, + 0.033181384205818176, + 0.030149158090353012, + 0.02469797618687153, + 0.05055768042802811, + 0.07121511548757553, + 0.3394889533519745, + 0.027331899851560593, + 0.2719363272190094, + 0.017429161816835403 + ], + "credit_map": [ + 0.01273254957050085, + 0.02564561367034912, + 0.012169030494987965, + 0.014208895154297352, + 0.028084326535463333, + 0.014326514676213264, + 0.014836679212749004, + 0.012008768506348133, + 0.003391075413674116, + 0.0026878390926867723, + 0.004034906160086393, + 0.004677391145378351, + 0.0028549986891448498, + 0.006123364437371492, + 0.004832092206925154, + 0.0045797149650752544, + 0.002947453409433365, + 0.0031633994076400995, + 0.004126416984945536, + 0.003474991302937269, + 0.002960789017379284, + 0.003863096470013261, + 0.0036047417670488358, + 0.006008267402648926, + 0.004210468847304583, + 0.0017693424597382545, + 0.0034532963763922453, + 0.002688873326405883, + 0.002029730938374996, + 0.004636087454855442, + 0.00213346048258245, + 0.0037767146714031696, + 0.007766711991280317, + 0.007149478420615196, + 0.004103210289031267, + 0.0059371995739638805, + 0.006858071777969599, + 0.007907659746706486, + 0.00629678787663579, + 0.004538557026535273, + 0.012534156441688538, + 0.011166335083544254, + 0.0038561238907277584, + 0.005676941014826298, + 0.009949405677616596, + 0.01726708561182022, + 0.005215238779783249, + 0.005549823399633169, + 0.03838379308581352, + 0.04500753432512283, + 0.028696617111563683, + 0.06198329105973244, + 0.041641075164079666, + 0.04282670468091965, + 0.04467976465821266, + 0.03627016395330429, + 0.0029496862553060055, + 0.003084413008764386, + 0.002720080316066742, + 0.004412610083818436, + 0.0031954110600054264, + 0.0029606337193399668, + 0.004640565719455481, + 0.003368501318618655, + 0.04716669023036957, + 0.04236724600195885, + 0.015347575768828392, + 0.03162975236773491, + 0.028598930686712265, + 0.05812980607151985, + 0.025019511580467224, + 0.023676816374063492, + 0.0023107994347810745, + 0.0023378555197268724, + 0.0016013995045796037, + 0.003410204080864787, + 0.0020499497186392546, + 0.001456417259760201, + 0.0025307289324700832, + 0.0017318056197836995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 0 + ], + "selected_reward": -0.018179578706622124 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023283764719963074 + ], + "opposite_logprob_tokens": [ + -3.7732837200164795 + ], + "logprob_reward": -0.023283764719963074, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023283764719963074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10653480887413025, + 0.16755129396915436, + 0.04657614231109619, + 0.024020928889513016, + 0.02798164077103138, + 0.022450458258390427, + 0.2144802212715149, + 0.08607597649097443, + 0.1528015434741974, + 0.15152698755264282 + ], + "credit_map": [ + 0.013336951844394207, + 0.013030857779085636, + 0.01030142605304718, + 0.014391009695827961, + 0.01381280180066824, + 0.018982229754328728, + 0.012090219184756279, + 0.01058930717408657, + 0.014833438210189342, + 0.018907830119132996, + 0.017627928406000137, + 0.01741248182952404, + 0.0159759484231472, + 0.02631567232310772, + 0.01178772747516632, + 0.04469026252627373, + 0.005836110096424818, + 0.005063147749751806, + 0.007548303343355656, + 0.004870918579399586, + 0.004607821349054575, + 0.005773217882961035, + 0.004462512210011482, + 0.008414112031459808, + 0.004528313409537077, + 0.002149656182155013, + 0.003545608138665557, + 0.0023931232281029224, + 0.0020007938146591187, + 0.003887119935825467, + 0.0026385788805782795, + 0.0028777357656508684, + 0.004069654271006584, + 0.004272417165338993, + 0.004170156549662352, + 0.003099683439359069, + 0.0035125187132507563, + 0.004115753807127476, + 0.00226481887511909, + 0.002476637251675129, + 0.003504588035866618, + 0.0031920813489705324, + 0.0016038321191444993, + 0.002248094417154789, + 0.0032899321522563696, + 0.003009428968653083, + 0.002758399583399296, + 0.0028440996538847685, + 0.015423827804625034, + 0.027190672233700752, + 0.016640299931168556, + 0.049506548792123795, + 0.033216968178749084, + 0.01785723678767681, + 0.03366798162460327, + 0.020976683124899864, + 0.012729808688163757, + 0.013471090234816074, + 0.007326800376176834, + 0.012998196296393871, + 0.012456764467060566, + 0.010493508540093899, + 0.01012350432574749, + 0.0064762975089251995, + 0.03138284757733345, + 0.01664557494223118, + 0.010922740213572979, + 0.017065977677702904, + 0.01377884205430746, + 0.032870177179574966, + 0.015157191082835197, + 0.014978191815316677, + 0.02414577640593052, + 0.01234884187579155, + 0.013393131084740162, + 0.023782435804605484, + 0.0301088597625494, + 0.012399369850754738, + 0.024958331137895584, + 0.010390241630375385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 8 + ], + "selected_reward": -0.023283764719963074 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a scene reminiscent of a child's drawing, replete with childish brushstrokes and a flattened perspective, the camera begins with a dynamic, upward-looking angle, following a plump piglet, painted all over in vibrant pink, with simple black polka-dot eyes and a spiral tail. From the top of a vast haystack constructed from rough, thick yellow crayon lines, it glides rapidly downward, a trail of white, dynamic lines representing its speed in its wake. As it reaches the edge of the haystack, it leaps with all its might, performing a fluid, dramatic mid-air flip, its four tiny hooves stomping joyfully in the air. At the apex of its flip, the camera begins a rapid-fire cycle of focus shifts and jump cuts: one moment a close-up of the piglet in mid-air, the next abruptly cutting to the lower reaches of the haystack, where an audience of animals, also composed of simple geometric shapes—a wide-eyed cow, a clucking hen, and a fluffy sheep—raise their heads and open their mouths, cheering in the purest of ways. The rapid-fire editing is accompanied by a violent flickering of the screen, as if from a bright white flashbulb, each flicker fading the saturated colors of the frame into white. The entire scene is illuminated by strings of giant, brightly colored spheres suspended against a deep blue night sky, casting a warm, dappled glow and creating a party atmosphere that is both chaotic and joyful, brimming with childlike innocence and energy.", + "seed": 31625, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_13/index_7625_seed_31625_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19399956688284875, + "vlm_logprob": -0.19399956688284875, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.19399956688284875, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003721337765455246 + ], + "opposite_logprob_tokens": [ + -6.128721237182617 + ], + "logprob_reward": -0.003721337765455246, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.003721337765455246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11147794872522354, + 0.07068079710006714, + 0.06833065301179886, + 0.13804438710212708, + 0.07349283993244171, + 0.05763937160372734, + 0.1755688339471817, + 0.05475990101695061, + 0.1699218451976776, + 0.08008342981338501 + ], + "credit_map": [ + 0.013636352494359016, + 0.01669095829129219, + 0.01686018332839012, + 0.014297443442046642, + 0.008841694332659245, + 0.011071099899709225, + 0.015386179089546204, + 0.0146940341219306, + 0.011474556289613247, + 0.008549072779715061, + 0.010638708248734474, + 0.007311016321182251, + 0.0068642389960587025, + 0.007831674069166183, + 0.008386997506022453, + 0.009624526835978031, + 0.006938307546079159, + 0.008404847234487534, + 0.005637791007757187, + 0.009044849313795567, + 0.009028295055031776, + 0.013039394281804562, + 0.008594496175646782, + 0.0076426733285188675, + 0.01322664599865675, + 0.013487833552062511, + 0.01709182746708393, + 0.015055841766297817, + 0.021413220092654228, + 0.02601444348692894, + 0.01876792684197426, + 0.012986634857952595, + 0.00874996930360794, + 0.009678979404270649, + 0.007451939396560192, + 0.008631780743598938, + 0.013428445905447006, + 0.0078938789665699, + 0.010463236831128597, + 0.007194602396339178, + 0.007797760423272848, + 0.006094531621783972, + 0.006131570320576429, + 0.006213714834302664, + 0.011165693402290344, + 0.006275438703596592, + 0.007533600088208914, + 0.006427059415727854, + 0.011720939539372921, + 0.028123514726758003, + 0.01856246031820774, + 0.04643212631344795, + 0.024983476847410202, + 0.011673479340970516, + 0.015340646728873253, + 0.018732186406850815, + 0.005645351950079203, + 0.008118504658341408, + 0.006993401795625687, + 0.006813128478825092, + 0.008698779158294201, + 0.00756027502939105, + 0.005733725614845753, + 0.0051967306062579155, + 0.014746989123523235, + 0.023984389379620552, + 0.0163490679115057, + 0.009828911162912846, + 0.028292018920183182, + 0.037122294306755066, + 0.017187384888529778, + 0.022410782054066658, + 0.008867652155458927, + 0.01133083738386631, + 0.00757020665332675, + 0.007871071808040142, + 0.011191544122993946, + 0.015815401449799538, + 0.009205173701047897, + 0.008231532759964466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 3 + ], + "selected_reward": -0.003721337765455246 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31340622901916504 + ], + "opposite_logprob_tokens": [ + -1.313406229019165 + ], + "logprob_reward": -0.31340622901916504, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31340622901916504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13197796046733856, + 0.06871327012777328, + 0.11313293874263763, + 0.16194775700569153, + 0.03881588578224182, + 0.07800377905368805, + 0.1442514955997467, + 0.05693187937140465, + 0.14332129061222076, + 0.06290376931428909 + ], + "credit_map": [ + 0.01198034267872572, + 0.011030004359781742, + 0.029585909098386765, + 0.019287245348095894, + 0.007824794389307499, + 0.016039254143834114, + 0.019879236817359924, + 0.016351154074072838, + 0.008933361619710922, + 0.006981575395911932, + 0.009658250957727432, + 0.009526662528514862, + 0.004586848895996809, + 0.008827186189591885, + 0.008536264300346375, + 0.01166311465203762, + 0.010984488762915134, + 0.013435262255370617, + 0.015106293372809887, + 0.015683382749557495, + 0.01055192667990923, + 0.017405100166797638, + 0.016769953072071075, + 0.013196524232625961, + 0.016094354912638664, + 0.011345786973834038, + 0.024933984503149986, + 0.014844154939055443, + 0.018542461097240448, + 0.02464349754154682, + 0.03174983337521553, + 0.019793663173913956, + 0.004404180683195591, + 0.004840549547225237, + 0.005099138244986534, + 0.004593485966324806, + 0.006941433995962143, + 0.0045888107270002365, + 0.004349980503320694, + 0.003998305182904005, + 0.0071333167143166065, + 0.00866422988474369, + 0.00804240070283413, + 0.008738420903682709, + 0.02072952501475811, + 0.008380405604839325, + 0.008949537761509418, + 0.007365941070020199, + 0.0092883026227355, + 0.021166786551475525, + 0.017221231013536453, + 0.03670283779501915, + 0.020768169313669205, + 0.011315654031932354, + 0.0117986174300313, + 0.015989892184734344, + 0.006790099199861288, + 0.010068279691040516, + 0.006476470269262791, + 0.007935212925076485, + 0.007530563045293093, + 0.00641606142744422, + 0.005819737911224365, + 0.005895447451621294, + 0.012987262569367886, + 0.024679064750671387, + 0.010611395351588726, + 0.011239644140005112, + 0.02241084910929203, + 0.03204401582479477, + 0.01707274094223976, + 0.012276312336325645, + 0.006996346637606621, + 0.006887740921229124, + 0.007230483461171389, + 0.008698167279362679, + 0.008089708164334297, + 0.009865458123385906, + 0.007698390632867813, + 0.007437472231686115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.31340622901916504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043071962893009186 + ], + "opposite_logprob_tokens": [ + -3.168071985244751 + ], + "logprob_reward": -0.043071962893009186, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043071962893009186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1098565012216568, + 0.06102163717150688, + 0.06341886520385742, + 0.20760305225849152, + 0.04859597980976105, + 0.04064454883337021, + 0.17390230298042297, + 0.03836488723754883, + 0.18967071175575256, + 0.06692145019769669 + ], + "credit_map": [ + 0.008376708254218102, + 0.007432559505105019, + 0.017198162153363228, + 0.018012477084994316, + 0.009162466041743755, + 0.015366637147963047, + 0.021797534078359604, + 0.01250995509326458, + 0.008939698338508606, + 0.005391641985625029, + 0.008128375746309757, + 0.009617211297154427, + 0.004484092351049185, + 0.008074714802205563, + 0.007957239635288715, + 0.008428662084043026, + 0.007210886105895042, + 0.005977933760732412, + 0.007822394371032715, + 0.008553820662200451, + 0.00680876336991787, + 0.010228738188743591, + 0.009523528628051281, + 0.00729279825463891, + 0.021496761590242386, + 0.01506879087537527, + 0.039904214441776276, + 0.023425359278917313, + 0.026095720008015633, + 0.030298149213194847, + 0.028523536399006844, + 0.02279052697122097, + 0.005101691000163555, + 0.007853951305150986, + 0.005510789807885885, + 0.004437727387994528, + 0.012755037285387516, + 0.004709064960479736, + 0.003981600981205702, + 0.004246115684509277, + 0.0033056973479688168, + 0.00429165456444025, + 0.00507391057908535, + 0.004528575111180544, + 0.01028160098940134, + 0.004446650855243206, + 0.004647963214665651, + 0.004068493843078613, + 0.011003918945789337, + 0.02971571497619152, + 0.01762266829609871, + 0.04843779280781746, + 0.024206314235925674, + 0.009564680978655815, + 0.015115158632397652, + 0.018236059695482254, + 0.0033956272527575493, + 0.006944631692022085, + 0.005740771070122719, + 0.003338472917675972, + 0.005210858304053545, + 0.0040105190128088, + 0.004959550686180592, + 0.004764454439282417, + 0.01580478437244892, + 0.030291404575109482, + 0.015468859113752842, + 0.008568951860070229, + 0.030531728640198708, + 0.04601098969578743, + 0.01865982823073864, + 0.024334164336323738, + 0.006740436423569918, + 0.009925514459609985, + 0.0049258628860116005, + 0.009679067879915237, + 0.009977695532143116, + 0.008994861505925655, + 0.010809037834405899, + 0.005868977401405573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.043071962893009186 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5760959386825562 + ], + "opposite_logprob_tokens": [ + -0.8260959386825562 + ], + "logprob_reward": -0.5760959386825562, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5760959386825562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09881920367479324, + 0.09183745831251144, + 0.04112974926829338, + 0.06841125339269638, + 0.10856518894433975, + 0.06816408783197403, + 0.11142132431268692, + 0.06535521149635315, + 0.27030980587005615, + 0.07598677277565002 + ], + "credit_map": [ + 0.010133747011423111, + 0.009902462363243103, + 0.01598438434302807, + 0.014058302156627178, + 0.007017495576292276, + 0.013246501795947552, + 0.015270482748746872, + 0.013205827213823795, + 0.012896491214632988, + 0.012378430925309658, + 0.01863216422498226, + 0.011945684440433979, + 0.008146165870130062, + 0.007518874481320381, + 0.010501108132302761, + 0.009818537160754204, + 0.00517166405916214, + 0.0054169283248484135, + 0.004159461706876755, + 0.006611799355596304, + 0.005121038760989904, + 0.005621044430881739, + 0.004769157152622938, + 0.004258657339960337, + 0.005783486645668745, + 0.005636925809085369, + 0.007416756357997656, + 0.006153651978820562, + 0.014419272541999817, + 0.012314503081142902, + 0.00915110856294632, + 0.007535547949373722, + 0.010797840543091297, + 0.017145724967122078, + 0.009448500350117683, + 0.012418272905051708, + 0.026838673278689384, + 0.010919440537691116, + 0.011268431320786476, + 0.009728305973112583, + 0.005499996244907379, + 0.008054198697209358, + 0.00706026516854763, + 0.0085475854575634, + 0.018462510779500008, + 0.007316102739423513, + 0.006594672799110413, + 0.00662876246497035, + 0.008363274857401848, + 0.01589479297399521, + 0.013919834047555923, + 0.0253697969019413, + 0.014073428697884083, + 0.010944729670882225, + 0.010495418682694435, + 0.012360045686364174, + 0.006729653105139732, + 0.00907046440988779, + 0.01056140661239624, + 0.006422821432352066, + 0.010210266336798668, + 0.008031623438000679, + 0.007423041388392448, + 0.006905936170369387, + 0.022888515144586563, + 0.04316682368516922, + 0.020016750320792198, + 0.01424605306237936, + 0.04245983436703682, + 0.06545065343379974, + 0.026868339627981186, + 0.035212837159633636, + 0.007572988048195839, + 0.007858558557927608, + 0.006460351869463921, + 0.008969530463218689, + 0.011791684664785862, + 0.011612871661782265, + 0.0132487453520298, + 0.008472041226923466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 4 + ], + "selected_reward": -0.5760959386825562 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a plump piglet painted in vibrant pink with black polka-dot eyes and a spiral tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the piglet glide rapidly downward from a vast haystack made of rough, thick yellow crayon lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the piglet perform a fluid mid-air flip with its four tiny hooves stomping joyfully?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene accompanied by a rapid-fire cycle of focus shifts, jump cuts, and violent screen flickering that fades colors to white?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are there giant brightly colored spheres suspended against a deep blue night sky casting a warm, dappled glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033702366054058075 + ], + "opposite_logprob_tokens": [ + -3.4087023735046387 + ], + "logprob_reward": -0.033702366054058075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033702366054058075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5114570260047913, + 0.07764028757810593, + 0.024184580892324448, + 0.07988101243972778, + 0.013185138814151287, + 0.0492180660367012, + 0.062054235488176346, + 0.04638851433992386, + 0.08852118253707886, + 0.04747001454234123 + ], + "credit_map": [ + 0.04554180055856705, + 0.05014047399163246, + 0.09140612930059433, + 0.06772342324256897, + 0.04525812715291977, + 0.07611288130283356, + 0.07842333614826202, + 0.05685079097747803, + 0.009631811641156673, + 0.008619165048003197, + 0.01580541580915451, + 0.00865188054740429, + 0.00805725809186697, + 0.00583283044397831, + 0.010268484242260456, + 0.010773439891636372, + 0.0026107397861778736, + 0.0030658026225864887, + 0.0035327081568539143, + 0.003141394816339016, + 0.002579659456387162, + 0.0038903378881514072, + 0.0025370505172759295, + 0.0028268853202462196, + 0.011484733782708645, + 0.00663407426327467, + 0.012698356993496418, + 0.009126354940235615, + 0.006992594804614782, + 0.01291567087173462, + 0.01115713082253933, + 0.008872093632817268, + 0.0014346130192279816, + 0.001679272623732686, + 0.0015071145026013255, + 0.0016517387703061104, + 0.0020822640508413315, + 0.0020461352542042732, + 0.001440140651538968, + 0.0013438590103760362, + 0.004769701045006514, + 0.005733908154070377, + 0.005842023063451052, + 0.006698362063616514, + 0.01220187172293663, + 0.004276348743587732, + 0.00479071494191885, + 0.004905133508145809, + 0.0048366389237344265, + 0.009011518210172653, + 0.0069535099901258945, + 0.014221834018826485, + 0.0091465525329113, + 0.006758969277143478, + 0.004762918688356876, + 0.0063622877933084965, + 0.004198713693767786, + 0.0067531210370361805, + 0.006843936163932085, + 0.008643974550068378, + 0.005087384022772312, + 0.005948592443019152, + 0.004942696541547775, + 0.003970093093812466, + 0.011928252875804901, + 0.014592253603041172, + 0.004929607268422842, + 0.007855284959077835, + 0.01311216689646244, + 0.01959001086652279, + 0.006821097806096077, + 0.009692502208054066, + 0.0052788653410971165, + 0.004814645275473595, + 0.003734590718522668, + 0.005418703425675631, + 0.007266558241099119, + 0.007977675646543503, + 0.008384610526263714, + 0.004594366531819105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.033702366054058075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_tensor([3070, 3070])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_tensor([3070, 3070])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..30bedefd88139446815ae2db03332a232e5aabae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_tensor([3070, 3070])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly descending drone shot captures a vast, tranquil lake, a deep indigo blue, like a flawless gem, bathed in the twilight's early pink-purple light. An elegant wooden flower boat, composed of dark, polished wood, glides steadily and deliberately from one side of the frame toward the center. As the camera descends and moves forward, details of the boat become increasingly clear: its hull and sails are adorned with an intricate garland of blooming pink peonies, white lilies, and lavender hyacinths, their petals dripping with glistening dewdrops. Rows of delicate lotus-shaped paper lanterns hang from the boat's sides, emitting a warm, soft orange glow that interplays with translucent gold and pink ribbons that flutter gently in the breeze. The boat's stern carves a V-shaped trail across the still water, transforming into soft, spreading ripples, shattering the lake's tranquility. The soft light of dusk and the warm light of lanterns interweave, casting dancing spots of light on the hull and giving the petals a dreamy golden edge. The whole picture is shrouded in a tranquil, romantic atmosphere full of fairy-tale magic.", + "seed": 28414, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_3070_seed_28414_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1565457542426884, + "vlm_logprob": -0.1565457542426884, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.1565457542426884, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5483748316764832 + ], + "opposite_logprob_tokens": [ + -0.9233748316764832 + ], + "logprob_reward": -0.5483748316764832, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5483748316764832, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20000828802585602, + 0.1370091289281845, + 0.03925667330622673, + 0.1430923044681549, + 0.11296158283948898, + 0.10390393435955048, + 0.11000508069992065, + 0.07325422018766403, + 0.06140168011188507, + 0.01910712569952011 + ], + "credit_map": [ + 0.022879203781485558, + 0.028205085545778275, + 0.020419493317604065, + 0.018736816942691803, + 0.03510606288909912, + 0.024861956015229225, + 0.022380400449037552, + 0.027419254183769226, + 0.03805461525917053, + 0.025169558823108673, + 0.018465841189026833, + 0.00781434215605259, + 0.008930373005568981, + 0.010146535001695156, + 0.007264651823788881, + 0.021163200959563255, + 0.00821468885987997, + 0.005053759086877108, + 0.0035069778095930815, + 0.0037665963172912598, + 0.004679248668253422, + 0.004199829418212175, + 0.00416623055934906, + 0.005669345147907734, + 0.01624635048210621, + 0.025428105145692825, + 0.011339560151100159, + 0.013687376864254475, + 0.023490792140364647, + 0.018225066363811493, + 0.013421363197267056, + 0.02125367522239685, + 0.016897911205887794, + 0.013599543832242489, + 0.016557609662413597, + 0.010044191963970661, + 0.016241690143942833, + 0.011552844196557999, + 0.011721979826688766, + 0.016345813870429993, + 0.01586119644343853, + 0.010269956663250923, + 0.014114121906459332, + 0.014566866680979729, + 0.017837615683674812, + 0.009762843139469624, + 0.00987865962088108, + 0.011612673290073872, + 0.012554937042295933, + 0.011903547681868076, + 0.016773955896496773, + 0.012283477932214737, + 0.015726668760180473, + 0.01160235982388258, + 0.011737198568880558, + 0.017422929406166077, + 0.016369326040148735, + 0.010634512640535831, + 0.004873861093074083, + 0.005351051222532988, + 0.0072578443214297295, + 0.0059878709726035595, + 0.005653033964335918, + 0.017126722261309624, + 0.01018582470715046, + 0.007256508804857731, + 0.009519253857433796, + 0.00756834028288722, + 0.004514038097113371, + 0.0043115499429404736, + 0.007888774387538433, + 0.01015738770365715, + 0.0030437856912612915, + 0.002287579933181405, + 0.00197820202447474, + 0.0021307426504790783, + 0.0018143127672374249, + 0.0023130851332098246, + 0.002095315605401993, + 0.0034441016614437103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.5483748316764832 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011100560426712036 + ], + "opposite_logprob_tokens": [ + -4.511100769042969 + ], + "logprob_reward": -0.011100560426712036, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011100560426712036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16127945482730865, + 0.29961398243904114, + 0.1170811802148819, + 0.03954460471868515, + 0.050607092678546906, + 0.038173750042915344, + 0.059532780200242996, + 0.07779143005609512, + 0.06665869802236557, + 0.08971698582172394 + ], + "credit_map": [ + 0.028943322598934174, + 0.024362487718462944, + 0.023075252771377563, + 0.01145164668560028, + 0.016976578161120415, + 0.02337021939456463, + 0.011544927954673767, + 0.021555017679929733, + 0.09649669378995895, + 0.06261317431926727, + 0.04116976261138916, + 0.015955034643411636, + 0.019096210598945618, + 0.01864035055041313, + 0.0140182850882411, + 0.03162446618080139, + 0.01836147904396057, + 0.016861381009221077, + 0.01217920146882534, + 0.010693076997995377, + 0.012404052540659904, + 0.011342816054821014, + 0.011503450572490692, + 0.02373572625219822, + 0.005364081356674433, + 0.004553772043436766, + 0.004034919198602438, + 0.005871175322681665, + 0.004679796285927296, + 0.005475545302033424, + 0.0047944518737494946, + 0.0047708614729344845, + 0.008989298716187477, + 0.007251963019371033, + 0.006731439381837845, + 0.0059981513768434525, + 0.00673914747312665, + 0.00499523151665926, + 0.004952747840434313, + 0.004949111025780439, + 0.0062564038671553135, + 0.005124156828969717, + 0.004210969898849726, + 0.004199748393148184, + 0.004712903406471014, + 0.004105071071535349, + 0.004691486246883869, + 0.0048730093985795975, + 0.00731433043256402, + 0.006788676604628563, + 0.008544299751520157, + 0.005681485868990421, + 0.010192200541496277, + 0.006223768461495638, + 0.005984873045235872, + 0.008803144097328186, + 0.013573205098509789, + 0.01546931266784668, + 0.007113035302609205, + 0.006259869784116745, + 0.010575184598565102, + 0.006026252638548613, + 0.005644192919135094, + 0.013130376115441322, + 0.009108912199735641, + 0.009769558906555176, + 0.006101385224610567, + 0.008193221874535084, + 0.010569796897470951, + 0.007799612823873758, + 0.0075719114392995834, + 0.0075443037785589695, + 0.0142213748767972, + 0.018806176260113716, + 0.007397088222205639, + 0.006094155367463827, + 0.010204692371189594, + 0.007773620542138815, + 0.00830922182649374, + 0.01691066287457943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.011100560426712036 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038089726120233536 + ], + "opposite_logprob_tokens": [ + -3.2880897521972656 + ], + "logprob_reward": -0.038089726120233536, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038089726120233536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2096904069185257, + 0.0647457018494606, + 0.05792679637670517, + 0.15839043259620667, + 0.08332176506519318, + 0.07829020172357559, + 0.13673551380634308, + 0.11190446466207504, + 0.04096188768744469, + 0.058032773435115814 + ], + "credit_map": [ + 0.027827469632029533, + 0.03577711805701256, + 0.027105286717414856, + 0.01671191304922104, + 0.026558632031083107, + 0.024012096226215363, + 0.021024882793426514, + 0.03067300282418728, + 0.017413433641195297, + 0.012405209243297577, + 0.005184774752706289, + 0.005516592413187027, + 0.005481258500367403, + 0.005261757876724005, + 0.007397437002509832, + 0.006085241213440895, + 0.01000925526022911, + 0.007340462878346443, + 0.0062905605882406235, + 0.004940549377351999, + 0.005902674049139023, + 0.0069036115892231464, + 0.007039556745439768, + 0.009500128217041492, + 0.016840113326907158, + 0.02834688313305378, + 0.013723256066441536, + 0.017360869795084, + 0.02588696777820587, + 0.01973029598593712, + 0.01512552797794342, + 0.02137651853263378, + 0.012154613621532917, + 0.013064336031675339, + 0.011962034739553928, + 0.007740434259176254, + 0.01001693494617939, + 0.008481299504637718, + 0.009959230199456215, + 0.009942884556949139, + 0.013215367682278156, + 0.009657998569309711, + 0.00830153003334999, + 0.008595820516347885, + 0.007953369989991188, + 0.008255076594650745, + 0.010466176085174084, + 0.011844863183796406, + 0.016195029020309448, + 0.014737563207745552, + 0.021697470918297768, + 0.015011093579232693, + 0.019821301102638245, + 0.013958057388663292, + 0.013029750436544418, + 0.022285252809524536, + 0.026689672842621803, + 0.014298463240265846, + 0.00881658773869276, + 0.008539867587387562, + 0.010196089744567871, + 0.008082429878413677, + 0.008221962489187717, + 0.027059394866228104, + 0.005504947621375322, + 0.004103758372366428, + 0.006638575345277786, + 0.00470610661432147, + 0.00459852023050189, + 0.004111625719815493, + 0.005177454091608524, + 0.006120902020484209, + 0.010151985101401806, + 0.009769044816493988, + 0.006525477394461632, + 0.005260446574538946, + 0.006257586646825075, + 0.006275682710111141, + 0.0067983875051140785, + 0.0069941626861691475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.038089726120233536 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799161434173584 + ], + "opposite_logprob_tokens": [ + -1.8049161434173584 + ], + "logprob_reward": -0.1799161434173584, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799161434173584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18129508197307587, + 0.10844236612319946, + 0.04134167730808258, + 0.06423214823007584, + 0.10858537256717682, + 0.05911707133054733, + 0.035626135766506195, + 0.13671831786632538, + 0.19922499358654022, + 0.06541687250137329 + ], + "credit_map": [ + 0.01896236278116703, + 0.028802499175071716, + 0.01915566250681877, + 0.014940918423235416, + 0.02313247136771679, + 0.03461678326129913, + 0.016863668337464333, + 0.024820715188980103, + 0.02255205251276493, + 0.01780281402170658, + 0.013830050826072693, + 0.0071966396644711494, + 0.009745652787387371, + 0.009584547951817513, + 0.007573990151286125, + 0.02015662007033825, + 0.007105878554284573, + 0.00794784352183342, + 0.0040588644333183765, + 0.003715335391461849, + 0.005571889225393534, + 0.004632105585187674, + 0.003926270641386509, + 0.004383492283523083, + 0.010724685154855251, + 0.00989770982414484, + 0.005185937508940697, + 0.006856649648398161, + 0.00933026522397995, + 0.0073308562859892845, + 0.0069647179916501045, + 0.007941328920423985, + 0.021815374493598938, + 0.01564626581966877, + 0.012048210017383099, + 0.010593310929834843, + 0.00974039826542139, + 0.011978176422417164, + 0.010681533254683018, + 0.016082102432847023, + 0.009348170831799507, + 0.00918135792016983, + 0.0045623816549777985, + 0.006770680192857981, + 0.008077374659478664, + 0.006918954662978649, + 0.008661935105919838, + 0.0055962176993489265, + 0.004288224037736654, + 0.004252997692674398, + 0.006003026384860277, + 0.003653389634564519, + 0.0049215881153941154, + 0.00455569988116622, + 0.0033653073478490114, + 0.004585901275277138, + 0.030339036136865616, + 0.01991799846291542, + 0.010887607000768185, + 0.008700081147253513, + 0.015703314915299416, + 0.010606639087200165, + 0.00918551441282034, + 0.0313781313598156, + 0.03463741019368172, + 0.02699536643922329, + 0.017037099227309227, + 0.023756591603159904, + 0.026458511129021645, + 0.023357482627034187, + 0.020211096853017807, + 0.026771435514092445, + 0.008167891763150692, + 0.01198447123169899, + 0.00887230969965458, + 0.007394240237772465, + 0.008772043511271477, + 0.00566795002669096, + 0.007602235767990351, + 0.006955727934837341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.1799161434173584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052475095726549625 + ], + "opposite_logprob_tokens": [ + -5.255247592926025 + ], + "logprob_reward": -0.0052475095726549625, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052475095726549625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1553863137960434, + 0.04632875695824623, + 0.04591265693306923, + 0.06823023408651352, + 0.12786966562271118, + 0.07260341197252274, + 0.0333968922495842, + 0.08603708446025848, + 0.21047450602054596, + 0.15376043319702148 + ], + "credit_map": [ + 0.018191074952483177, + 0.025785235688090324, + 0.017578037455677986, + 0.013587236404418945, + 0.021588928997516632, + 0.025622393935918808, + 0.012489323504269123, + 0.020544085651636124, + 0.006622547749429941, + 0.00733571732416749, + 0.006443163380026817, + 0.0036392156034708023, + 0.006251094862818718, + 0.004366416018456221, + 0.004941902123391628, + 0.006728696171194315, + 0.0074297706596553326, + 0.008888214826583862, + 0.004363842774182558, + 0.004201891366392374, + 0.0057791005820035934, + 0.004675633739680052, + 0.005051227752119303, + 0.005522978026419878, + 0.009629456326365471, + 0.011104806326329708, + 0.012637177482247353, + 0.009091392159461975, + 0.007425263058394194, + 0.006715507246553898, + 0.005955852102488279, + 0.00567077798768878, + 0.022483063861727715, + 0.01844150386750698, + 0.01434446033090353, + 0.012572004459798336, + 0.017500603571534157, + 0.013415081426501274, + 0.012748447246849537, + 0.01636449247598648, + 0.010101895779371262, + 0.007563162595033646, + 0.010854854248464108, + 0.010863825678825378, + 0.012148663401603699, + 0.005986945237964392, + 0.006428791210055351, + 0.00865528080612421, + 0.005125845782458782, + 0.004434723407030106, + 0.004683505743741989, + 0.0043021696619689465, + 0.004037944599986076, + 0.0036918180994689465, + 0.002991940127685666, + 0.004128946457058191, + 0.01327566895633936, + 0.011891077272593975, + 0.00796758197247982, + 0.011009778827428818, + 0.0091857323423028, + 0.009147895500063896, + 0.0087756821885705, + 0.014783668331801891, + 0.039144255220890045, + 0.026854055002331734, + 0.014940452761948109, + 0.025124212726950645, + 0.027742961421608925, + 0.026101870462298393, + 0.022879011929035187, + 0.027687693014740944, + 0.022027624770998955, + 0.01861370913684368, + 0.017601830884814262, + 0.01988311856985092, + 0.01899072900414467, + 0.016253555193543434, + 0.019181735813617706, + 0.021208129823207855 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0052475095726549625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending drone shot captures a vast, tranquil lake, a deep indigo blue, like a flawless gem, bathed in the twilight's early pink-purple light. An elegant wooden flower boat, composed of dark, polished wood, glides steadily and deliberately from one side of the frame toward the center. As the camera descends and moves forward, details of the boat become increasingly clear: its hull and sails are adorned with an intricate garland of blooming pink peonies, white lilies, and lavender hyacinths, their petals dripping with glistening dewdrops. Rows of delicate lotus-shaped paper lanterns hang from the boat's sides, emitting a warm, soft orange glow that interplays with translucent gold and pink ribbons that flutter gently in the breeze. The boat's stern carves a V-shaped trail across the still water, transforming into soft, spreading ripples, shattering the lake's tranquility. The soft light of dusk and the warm light of lanterns interweave, casting dancing spots of light on the hull and giving the petals a dreamy golden edge. The whole picture is shrouded in a tranquil, romantic atmosphere full of fairy-tale magic.", + "seed": 28415, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_3070_seed_28415_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11630709320306779, + "vlm_logprob": -0.11630709320306779, + "vlm_yes_no_margin": 2.7250000119209288, + "vlm_token_credit": -0.11630709320306779, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3953356146812439 + ], + "opposite_logprob_tokens": [ + -1.1453356742858887 + ], + "logprob_reward": -0.3953356146812439, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.3953356146812439, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15545307099819183, + 0.044161271303892136, + 0.021422551944851875, + 0.046093832701444626, + 0.02475450560450554, + 0.08957517147064209, + 0.505336582660675, + 0.08174262195825577, + 0.027453236281871796, + 0.004007222130894661 + ], + "credit_map": [ + 0.014028077013790607, + 0.0304560624063015, + 0.014916555024683475, + 0.005733004305511713, + 0.06464867293834686, + 0.0129185626283288, + 0.007185335271060467, + 0.005566789768636227, + 0.014254234731197357, + 0.004544026218354702, + 0.01020293403416872, + 0.0023529084865003824, + 0.004976077005267143, + 0.002893976401537657, + 0.0020234743133187294, + 0.0029136454686522484, + 0.0026459358632564545, + 0.0019635746721178293, + 0.0037884979974478483, + 0.0018147187074646354, + 0.0023675872944295406, + 0.005793198477476835, + 0.0013976962072774768, + 0.0016513412119820714, + 0.004769049119204283, + 0.005620752461254597, + 0.004114334937185049, + 0.007014465052634478, + 0.006340008694678545, + 0.008814577013254166, + 0.005440867505967617, + 0.003979780245572329, + 0.0032695613335818052, + 0.0025718738324940205, + 0.002520886482670903, + 0.00480961799621582, + 0.003135652979835868, + 0.002810790902003646, + 0.003038633381947875, + 0.002597488695755601, + 0.012519610114395618, + 0.008739720098674297, + 0.0068201967515051365, + 0.012221325188875198, + 0.008251406252384186, + 0.017981404438614845, + 0.014780369587242603, + 0.008261144161224365, + 0.13445448875427246, + 0.030918052420020103, + 0.029726240783929825, + 0.05371616780757904, + 0.032344818115234375, + 0.13552749156951904, + 0.05924437195062637, + 0.029404934495687485, + 0.014658365398645401, + 0.005845073144882917, + 0.00965891219675541, + 0.01877540349960327, + 0.003993663005530834, + 0.010744976811110973, + 0.010560977272689342, + 0.007505247835069895, + 0.004092104267328978, + 0.0024592664558440447, + 0.005440681707113981, + 0.0031432551331818104, + 0.002713510999456048, + 0.0044293152168393135, + 0.00236097676679492, + 0.0028141268994659185, + 0.000607054156716913, + 0.00046544874203391373, + 0.0006026109331287444, + 0.0005463625420816243, + 0.00047987684956751764, + 0.00035945334821008146, + 0.0004935821052640676, + 0.0004528338322415948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.3953356146812439 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01420564204454422 + ], + "opposite_logprob_tokens": [ + -4.264205455780029 + ], + "logprob_reward": -0.01420564204454422, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01420564204454422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30022066831588745, + 0.042309798300266266, + 0.06647598743438721, + 0.04759110510349274, + 0.03239532187581062, + 0.13082805275917053, + 0.2633496820926666, + 0.08106856048107147, + 0.026875581592321396, + 0.0088852159678936 + ], + "credit_map": [ + 0.02498650550842285, + 0.058252133429050446, + 0.03180670738220215, + 0.009948991239070892, + 0.12429410219192505, + 0.027398735284805298, + 0.013291816227138042, + 0.010241675190627575, + 0.012758045457303524, + 0.0041595459915697575, + 0.00917840376496315, + 0.0028038816526532173, + 0.004523056093603373, + 0.0033874674700200558, + 0.002467567566782236, + 0.003031832166016102, + 0.00838789064437151, + 0.004778710659593344, + 0.012736286036670208, + 0.008424692787230015, + 0.00654606381431222, + 0.015008240006864071, + 0.0053328718058764935, + 0.005261240992695093, + 0.006119313649833202, + 0.005166187882423401, + 0.005967047531157732, + 0.005955789238214493, + 0.005307184532284737, + 0.011114314198493958, + 0.0038482837844640017, + 0.004112987779080868, + 0.003734622150659561, + 0.003381771268323064, + 0.0032856271136552095, + 0.006414737086743116, + 0.004046314395964146, + 0.0032982262782752514, + 0.004519481677561998, + 0.0037145446985960007, + 0.020977802574634552, + 0.011809531599283218, + 0.010833818465471268, + 0.017246387898921967, + 0.009431866928935051, + 0.02788170985877514, + 0.01977497525513172, + 0.012871971353888512, + 0.08542394638061523, + 0.014295379631221294, + 0.01203664019703865, + 0.01536340732127428, + 0.016209306195378304, + 0.08997348695993423, + 0.016290873289108276, + 0.013756667263805866, + 0.012384770438075066, + 0.0074643720872700214, + 0.009924519807100296, + 0.015725430101156235, + 0.006721665617078543, + 0.008481447584927082, + 0.011238656006753445, + 0.00912771187722683, + 0.004777912981808186, + 0.002603895030915737, + 0.00550931366160512, + 0.0037950007244944572, + 0.0021744309924542904, + 0.0027499005664139986, + 0.002550277393311262, + 0.002714852336794138, + 0.0014362391084432602, + 0.0013946404214948416, + 0.0011360624339431524, + 0.0010918674524873495, + 0.001213697250932455, + 0.0008152658701874316, + 0.0009323233389295638, + 0.0008651208481751382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.01420564204454422 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054932743310928345 + ], + "opposite_logprob_tokens": [ + -2.9299328327178955 + ], + "logprob_reward": -0.054932743310928345, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054932743310928345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15188157558441162, + 0.018809909000992775, + 0.03047989308834076, + 0.037215568125247955, + 0.026433046907186508, + 0.040092311799526215, + 0.4962064027786255, + 0.14522412419319153, + 0.0335792638361454, + 0.020077912136912346 + ], + "credit_map": [ + 0.010426883585751057, + 0.025063885375857353, + 0.01969454623758793, + 0.010941402986645699, + 0.052668001502752304, + 0.01416498702019453, + 0.008706741966307163, + 0.01021514367312193, + 0.0026468730065971613, + 0.00242369482293725, + 0.0031412832904607058, + 0.0026119700632989407, + 0.002194356871768832, + 0.0019553538877516985, + 0.0023872912861406803, + 0.001449086586944759, + 0.0032398414332419634, + 0.002504325471818447, + 0.005230089649558067, + 0.003248482011258602, + 0.002999659162014723, + 0.008650995790958405, + 0.0018988046795129776, + 0.002707693725824356, + 0.005075486376881599, + 0.0029963974375277758, + 0.0043654548935592175, + 0.003925052471458912, + 0.003935111686587334, + 0.01084859948605299, + 0.0024316662456840277, + 0.0036378030199557543, + 0.003525317879393697, + 0.003429321339353919, + 0.002007425297051668, + 0.004351033363491297, + 0.0031581451185047626, + 0.004335594829171896, + 0.003189551644027233, + 0.002436659764498472, + 0.004980167374014854, + 0.0041291736997663975, + 0.0029349764809012413, + 0.005151910241693258, + 0.005652749445289373, + 0.006553206127136946, + 0.0070490045472979546, + 0.0036411236505955458, + 0.1569875180721283, + 0.028631487861275673, + 0.02302050031721592, + 0.03247348219156265, + 0.032452769577503204, + 0.15874087810516357, + 0.0398048534989357, + 0.02409488894045353, + 0.027045389637351036, + 0.01344200037419796, + 0.020872294902801514, + 0.02025582268834114, + 0.01064615324139595, + 0.019512705504894257, + 0.016161402687430382, + 0.017288358882069588, + 0.004793452564626932, + 0.004983320366591215, + 0.0036670370027422905, + 0.0027670571580529213, + 0.004530787467956543, + 0.006959679536521435, + 0.002695923438295722, + 0.0031820049043744802, + 0.004574124235659838, + 0.0026447835844010115, + 0.0025759015697985888, + 0.0026103733107447624, + 0.00203153514303267, + 0.0018163269851356745, + 0.0016051864949986339, + 0.0022196786012500525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.054932743310928345 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898183912038803 + ], + "opposite_logprob_tokens": [ + -2.578981876373291 + ], + "logprob_reward": -0.07898183912038803, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898183912038803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2806571125984192, + 0.061800383031368256, + 0.038023628294467926, + 0.044710855931043625, + 0.027632225304841995, + 0.03846548870205879, + 0.29700416326522827, + 0.07706286013126373, + 0.09344979375600815, + 0.04119346663355827 + ], + "credit_map": [ + 0.023866452276706696, + 0.05462199077010155, + 0.029005125164985657, + 0.009836884215474129, + 0.11684667319059372, + 0.023896127939224243, + 0.013206157833337784, + 0.009377729147672653, + 0.01893293485045433, + 0.005641522817313671, + 0.016299506649374962, + 0.003139605512842536, + 0.006915524136275053, + 0.004012784920632839, + 0.0021627508103847504, + 0.00469575310125947, + 0.004082296043634415, + 0.003582495730370283, + 0.006199134513735771, + 0.004033152479678392, + 0.004583434667438269, + 0.010674293152987957, + 0.0026986354496330023, + 0.0021701909136027098, + 0.005083210300654173, + 0.005789758171886206, + 0.004495191387832165, + 0.0050546214915812016, + 0.006068766117095947, + 0.010107826441526413, + 0.003904615994542837, + 0.004206871148198843, + 0.0043830894865095615, + 0.0027455734089016914, + 0.002832758938893676, + 0.005165891256183386, + 0.0036822431720793247, + 0.0030869352631270885, + 0.0029741320759058, + 0.0027616058941930532, + 0.005521479528397322, + 0.003416848136112094, + 0.003876039292663336, + 0.005727460142225027, + 0.004376072436571121, + 0.007068513426929712, + 0.004255800507962704, + 0.00422328244894743, + 0.0982825756072998, + 0.01554288249462843, + 0.012687119655311108, + 0.016518743708729744, + 0.0188185703009367, + 0.10555918514728546, + 0.013834618963301182, + 0.015760501846671104, + 0.008993790484964848, + 0.0119051244109869, + 0.008886989206075668, + 0.010052274912595749, + 0.01168198324739933, + 0.006613192614167929, + 0.007977847941219807, + 0.010951663367450237, + 0.014022941701114178, + 0.009380600415170193, + 0.017528904601931572, + 0.009387251920998096, + 0.01048694271594286, + 0.016102945432066917, + 0.007389521691948175, + 0.009150695987045765, + 0.009806729853153229, + 0.006327817216515541, + 0.005173236131668091, + 0.0038349665701389313, + 0.0053837182931602, + 0.00332356127910316, + 0.003571480279788375, + 0.0037719635292887688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.07898183912038803 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038079626858234406 + ], + "opposite_logprob_tokens": [ + -3.2880797386169434 + ], + "logprob_reward": -0.038079626858234406, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038079626858234406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07890564203262329, + 0.029236650094389915, + 0.07213623821735382, + 0.03400560840964317, + 0.04896767809987068, + 0.07947810739278793, + 0.2364027053117752, + 0.28589072823524475, + 0.10130808502435684, + 0.033668600022792816 + ], + "credit_map": [ + 0.005018565338104963, + 0.011336005292832851, + 0.0117079708725214, + 0.007907500490546227, + 0.021020878106355667, + 0.010473831556737423, + 0.005392969120293856, + 0.006047917529940605, + 0.0039779432117938995, + 0.002889044117182493, + 0.003949886187911034, + 0.004744215868413448, + 0.0023747652303427458, + 0.004658110905438662, + 0.0035643018782138824, + 0.0030783831607550383, + 0.010366242378950119, + 0.007848023436963558, + 0.008647809736430645, + 0.008835485205054283, + 0.008517012000083923, + 0.01701824925839901, + 0.006179014220833778, + 0.00472440617159009, + 0.004431671928614378, + 0.004179442301392555, + 0.003492942778393626, + 0.0038760609459131956, + 0.004366907756775618, + 0.008155296556651592, + 0.0027455445379018784, + 0.002757744397968054, + 0.005245652049779892, + 0.0057508558966219425, + 0.004289548844099045, + 0.009298278018832207, + 0.00722681125625968, + 0.006471049040555954, + 0.00546870194375515, + 0.005216783378273249, + 0.011758361011743546, + 0.008784787729382515, + 0.005808308720588684, + 0.009404250420629978, + 0.010815351270139217, + 0.014965012669563293, + 0.007993421517312527, + 0.009948614053428173, + 0.07078683376312256, + 0.017100604251027107, + 0.012386036105453968, + 0.014179276302456856, + 0.022393299266695976, + 0.07419095188379288, + 0.013208851218223572, + 0.01215684786438942, + 0.06255409866571426, + 0.030860794708132744, + 0.023851368576288223, + 0.05509406328201294, + 0.020526256412267685, + 0.038409944623708725, + 0.028835346922278404, + 0.02575884759426117, + 0.01531349215656519, + 0.010187952779233456, + 0.016072453930974007, + 0.009699766524136066, + 0.012957613915205002, + 0.02167097106575966, + 0.006516312714666128, + 0.008889526128768921, + 0.005383872892707586, + 0.004982641898095608, + 0.004159117583185434, + 0.0037869063671678305, + 0.0050522880628705025, + 0.003224916523322463, + 0.0032354036811739206, + 0.003843453014269471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.038079626858234406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending drone shot captures a vast, tranquil lake, a deep indigo blue, like a flawless gem, bathed in the twilight's early pink-purple light. An elegant wooden flower boat, composed of dark, polished wood, glides steadily and deliberately from one side of the frame toward the center. As the camera descends and moves forward, details of the boat become increasingly clear: its hull and sails are adorned with an intricate garland of blooming pink peonies, white lilies, and lavender hyacinths, their petals dripping with glistening dewdrops. Rows of delicate lotus-shaped paper lanterns hang from the boat's sides, emitting a warm, soft orange glow that interplays with translucent gold and pink ribbons that flutter gently in the breeze. The boat's stern carves a V-shaped trail across the still water, transforming into soft, spreading ripples, shattering the lake's tranquility. The soft light of dusk and the warm light of lanterns interweave, casting dancing spots of light on the hull and giving the petals a dreamy golden edge. The whole picture is shrouded in a tranquil, romantic atmosphere full of fairy-tale magic.", + "seed": 28416, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_3070_seed_28416_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1350505013950169, + "vlm_logprob": -0.1350505013950169, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.1350505013950169, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5888986587524414 + ], + "opposite_logprob_tokens": [ + -0.8388986587524414 + ], + "logprob_reward": -0.5888986587524414, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5888986587524414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09056717157363892, + 0.05663413926959038, + 0.041263397783041, + 0.2620982527732849, + 0.08166787028312683, + 0.032354895025491714, + 0.2233472764492035, + 0.06935588270425797, + 0.11746779084205627, + 0.025243351235985756 + ], + "credit_map": [ + 0.014863190241158009, + 0.013889696449041367, + 0.011714191175997257, + 0.006306651514023542, + 0.011242331005632877, + 0.014976240694522858, + 0.008662755601108074, + 0.008912119083106518, + 0.005920749623328447, + 0.008411630056798458, + 0.008683140389621258, + 0.003979293163865805, + 0.013231965713202953, + 0.006791966967284679, + 0.005505930166691542, + 0.004109461326152086, + 0.005031939595937729, + 0.007399883586913347, + 0.00590919703245163, + 0.003958558663725853, + 0.007365050725638866, + 0.005370014812797308, + 0.003126362105831504, + 0.0031023898627609015, + 0.11220768839120865, + 0.099093496799469, + 0.011662120930850506, + 0.006752049550414085, + 0.01170506700873375, + 0.011595355346798897, + 0.005424593575298786, + 0.0036578839644789696, + 0.01198040321469307, + 0.008892864920198917, + 0.010522641241550446, + 0.00572597561404109, + 0.016743171960115433, + 0.007872809655964375, + 0.012256111949682236, + 0.007673890329897404, + 0.0042668720707297325, + 0.0035856030881404877, + 0.0028916369192302227, + 0.002980927238240838, + 0.009201286360621452, + 0.002934094285592437, + 0.0037008619401603937, + 0.0027936112601310015, + 0.02179747074842453, + 0.03225411847233772, + 0.02043025754392147, + 0.03421647474169731, + 0.04314016178250313, + 0.018883205950260162, + 0.034664180129766464, + 0.017961397767066956, + 0.006581481080502272, + 0.014128120616078377, + 0.0048512984067201614, + 0.008399922400712967, + 0.00945264846086502, + 0.01005085464566946, + 0.011140238493680954, + 0.004751319065690041, + 0.008629833348095417, + 0.016926202923059464, + 0.007815388031303883, + 0.0259409137070179, + 0.018605319783091545, + 0.008913966827094555, + 0.024321673437952995, + 0.006314494181424379, + 0.0031401016749441624, + 0.0042723375372588634, + 0.00405143154785037, + 0.002705513732507825, + 0.002981538651511073, + 0.002906464273110032, + 0.003245556727051735, + 0.0019404069753363729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.5888986587524414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046500167809426785 + ], + "opposite_logprob_tokens": [ + -5.379650115966797 + ], + "logprob_reward": -0.0046500167809426785, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046500167809426785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14400072395801544, + 0.10229913890361786, + 0.07458467036485672, + 0.20817561447620392, + 0.11248492449522018, + 0.06145612522959709, + 0.030261214822530746, + 0.05746397003531456, + 0.11748068779706955, + 0.09179291129112244 + ], + "credit_map": [ + 0.01590835303068161, + 0.012596904300153255, + 0.02636750042438507, + 0.009389875456690788, + 0.02326948381960392, + 0.028699297457933426, + 0.01453461404889822, + 0.013234695419669151, + 0.011791902594268322, + 0.012369553558528423, + 0.01211492158472538, + 0.008290155790746212, + 0.021425707265734673, + 0.02175474725663662, + 0.00782860442996025, + 0.006723539903759956, + 0.0161798857152462, + 0.010358658619225025, + 0.020481573417782784, + 0.004015786107629538, + 0.009077071212232113, + 0.005634865257889032, + 0.004408332984894514, + 0.0044285026378929615, + 0.05441619083285332, + 0.07509391009807587, + 0.027700653299689293, + 0.009531104005873203, + 0.016230538487434387, + 0.010834859684109688, + 0.007065588142722845, + 0.0073027582839131355, + 0.01395538542419672, + 0.01186183001846075, + 0.021614542230963707, + 0.007327331695705652, + 0.029234519228339195, + 0.010073145851492882, + 0.009042314253747463, + 0.009375862777233124, + 0.007305430248379707, + 0.006756712682545185, + 0.00981734599918127, + 0.005201095249503851, + 0.014485223218798637, + 0.005940590985119343, + 0.005747608374804258, + 0.006202115211635828, + 0.0036613629199564457, + 0.005534432828426361, + 0.003944015130400658, + 0.0038002184592187405, + 0.004938474390655756, + 0.003014293033629656, + 0.003353440435603261, + 0.0020149764604866505, + 0.005619417876005173, + 0.011920589953660965, + 0.0040921540930867195, + 0.00625685416162014, + 0.00751119339838624, + 0.008167104795575142, + 0.010032694786787033, + 0.0038639619015157223, + 0.011561241000890732, + 0.016936182975769043, + 0.00938172172755003, + 0.035300303250551224, + 0.010077587328851223, + 0.0059594884514808655, + 0.02383589744567871, + 0.004428282845765352, + 0.009352987632155418, + 0.010217411443591118, + 0.009154163300991058, + 0.010580934584140778, + 0.011127294041216373, + 0.013673478737473488, + 0.015466518700122833, + 0.012220116332173347 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.0046500167809426785 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863445460796356 + ], + "opposite_logprob_tokens": [ + -3.0486345291137695 + ], + "logprob_reward": -0.04863445460796356, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863445460796356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16961254179477692, + 0.05144275724887848, + 0.11559682339429855, + 0.13182319700717926, + 0.10389944911003113, + 0.03768942505121231, + 0.04860679432749748, + 0.09400174766778946, + 0.14158335328102112, + 0.10574392229318619 + ], + "credit_map": [ + 0.02540627308189869, + 0.013508198782801628, + 0.024076996371150017, + 0.015049569308757782, + 0.023516075685620308, + 0.026432042941451073, + 0.024664221331477165, + 0.016959160566329956, + 0.00686269998550415, + 0.006782702635973692, + 0.006512910593301058, + 0.0050804647617042065, + 0.011831439100205898, + 0.004667229484766722, + 0.005449382122606039, + 0.004255928564816713, + 0.016540557146072388, + 0.019978098571300507, + 0.019065624102950096, + 0.0089713828638196, + 0.024711793288588524, + 0.010324550792574883, + 0.01038761530071497, + 0.005617199465632439, + 0.042413901537656784, + 0.04035330191254616, + 0.007785385474562645, + 0.0047501386143267155, + 0.01862287148833275, + 0.006758430972695351, + 0.006853014696389437, + 0.004286146257072687, + 0.01477628480643034, + 0.010284285992383957, + 0.015201523900032043, + 0.008000150322914124, + 0.027280695736408234, + 0.01009168941527605, + 0.0102786710485816, + 0.007986149750649929, + 0.004689557012170553, + 0.003917968366295099, + 0.003781770123168826, + 0.002349100774154067, + 0.010830803774297237, + 0.004344495013356209, + 0.004442759323865175, + 0.0033329715952277184, + 0.005294021684676409, + 0.006860308349132538, + 0.005054113920778036, + 0.006462018005549908, + 0.00850845966488123, + 0.0043100290931761265, + 0.008933479897677898, + 0.0031843630131334066, + 0.008375575765967369, + 0.02181709185242653, + 0.004560209810733795, + 0.010702334344387054, + 0.01447165384888649, + 0.013516116887331009, + 0.014929273165762424, + 0.0056294891983270645, + 0.010899486020207405, + 0.018409207463264465, + 0.009190446697175503, + 0.033248938620090485, + 0.021362164989113808, + 0.011211718432605267, + 0.02852591685950756, + 0.008735466748476028, + 0.014830517582595348, + 0.014436684548854828, + 0.011957644484937191, + 0.011069662868976593, + 0.013972013257443905, + 0.012259445153176785, + 0.016310619190335274, + 0.010907336138188839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.04863445460796356 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0886656790971756, + 0.08296970278024673, + 0.026958893984556198, + 0.15203215181827545, + 0.18936949968338013, + 0.01692400500178337, + 0.026384945958852768, + 0.22967523336410522, + 0.12302042543888092, + 0.06399951875209808 + ], + "credit_map": [ + 0.010305440984666348, + 0.009589103981852531, + 0.014097214676439762, + 0.005319597199559212, + 0.012229438871145248, + 0.019802996888756752, + 0.009573194198310375, + 0.007748683448880911, + 0.009388030506670475, + 0.010318535380065441, + 0.009626802057027817, + 0.007308775093406439, + 0.02072061039507389, + 0.012342085130512714, + 0.007616671733558178, + 0.00564819248393178, + 0.003710420336574316, + 0.004613989032804966, + 0.00415864447131753, + 0.002340457635000348, + 0.005201317369937897, + 0.002671235939487815, + 0.0024906876496970654, + 0.0017721413169056177, + 0.04492179676890373, + 0.04003923013806343, + 0.008111909963190556, + 0.004561503883451223, + 0.03187479451298714, + 0.009812891483306885, + 0.007697202265262604, + 0.005012805573642254, + 0.0245194248855114, + 0.021258126944303513, + 0.019363809376955032, + 0.010680324397981167, + 0.06294436752796173, + 0.02013321779668331, + 0.015324189327657223, + 0.01514603104442358, + 0.0021816084627062082, + 0.0019792295061051846, + 0.001822686637751758, + 0.0013134934706613421, + 0.00404795166105032, + 0.0016620536334812641, + 0.0023954887874424458, + 0.001521491096355021, + 0.0029562576673924923, + 0.003406818024814129, + 0.003353671869263053, + 0.004367184359580278, + 0.003318668343126774, + 0.002269256627187133, + 0.0038872009608894587, + 0.00282588810659945, + 0.02004578337073326, + 0.05775581672787666, + 0.014395169913768768, + 0.029219212010502815, + 0.031017936766147614, + 0.031820911914110184, + 0.031280435621738434, + 0.014139946550130844, + 0.011027373373508453, + 0.018653646111488342, + 0.009683762677013874, + 0.03349161148071289, + 0.013665908016264439, + 0.007425157818943262, + 0.023993706330657005, + 0.005079254973679781, + 0.005670439917594194, + 0.008894500322639942, + 0.008224556222558022, + 0.007566813845187426, + 0.007909717038273811, + 0.007886658422648907, + 0.009006865322589874, + 0.008839963003993034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 3 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009773616679012775 + ], + "opposite_logprob_tokens": [ + -4.6347737312316895 + ], + "logprob_reward": -0.009773616679012775, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009773616679012775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051661327481269836, + 0.05170345678925514, + 0.056962378323078156, + 0.22659505903720856, + 0.052561018615961075, + 0.023145321756601334, + 0.10677070915699005, + 0.07062318176031113, + 0.11027327179908752, + 0.2497042417526245 + ], + "credit_map": [ + 0.0074827540665864944, + 0.005498370621353388, + 0.005737797357141972, + 0.004378130193799734, + 0.009082765318453312, + 0.008416461758315563, + 0.005552269052714109, + 0.005512776784598827, + 0.005218039732426405, + 0.005981395486742258, + 0.007827534340322018, + 0.0034653348848223686, + 0.014853714033961296, + 0.00551200658082962, + 0.0051894765347242355, + 0.0036559519357979298, + 0.008935246616601944, + 0.01182320062071085, + 0.007330767810344696, + 0.0037925883661955595, + 0.008593764156103134, + 0.007809960749000311, + 0.004068168345838785, + 0.00460868701338768, + 0.0886002853512764, + 0.07858414202928543, + 0.00966944545507431, + 0.006339493207633495, + 0.021944697946310043, + 0.009293973445892334, + 0.006423793267458677, + 0.0057392423041164875, + 0.005885284394025803, + 0.005640726070851088, + 0.00708152586594224, + 0.005360954906791449, + 0.012066018767654896, + 0.00486197741702199, + 0.005710184574127197, + 0.00595434382557869, + 0.00259921932592988, + 0.0023944273125380278, + 0.0031400215812027454, + 0.0015173607971519232, + 0.006628329865634441, + 0.00253142393194139, + 0.002412061206996441, + 0.0019224778516218066, + 0.010961058549582958, + 0.013752993196249008, + 0.010194112546741962, + 0.01718279719352722, + 0.019421223551034927, + 0.008133468218147755, + 0.019291788339614868, + 0.007833275012671947, + 0.0064324988052248955, + 0.017472339794039726, + 0.004964310675859451, + 0.012882553972303867, + 0.008599985390901566, + 0.006401271093636751, + 0.010325654409825802, + 0.003544572275131941, + 0.008887371979653835, + 0.011959354393184185, + 0.010162693448364735, + 0.015325453132390976, + 0.01878724992275238, + 0.014248817227780819, + 0.023007189854979515, + 0.007895143702626228, + 0.028980987146496773, + 0.02801980823278427, + 0.027406418696045876, + 0.029916847124695778, + 0.033101800829172134, + 0.026162046939134598, + 0.05212432146072388, + 0.023992009460926056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 8 + ], + "selected_reward": -0.009773616679012775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly descending drone shot captures a vast, tranquil lake, a deep indigo blue, like a flawless gem, bathed in the twilight's early pink-purple light. An elegant wooden flower boat, composed of dark, polished wood, glides steadily and deliberately from one side of the frame toward the center. As the camera descends and moves forward, details of the boat become increasingly clear: its hull and sails are adorned with an intricate garland of blooming pink peonies, white lilies, and lavender hyacinths, their petals dripping with glistening dewdrops. Rows of delicate lotus-shaped paper lanterns hang from the boat's sides, emitting a warm, soft orange glow that interplays with translucent gold and pink ribbons that flutter gently in the breeze. The boat's stern carves a V-shaped trail across the still water, transforming into soft, spreading ripples, shattering the lake's tranquility. The soft light of dusk and the warm light of lanterns interweave, casting dancing spots of light on the hull and giving the petals a dreamy golden edge. The whole picture is shrouded in a tranquil, romantic atmosphere full of fairy-tale magic.", + "seed": 28417, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_14/index_3070_seed_28417_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09690394401550292, + "vlm_logprob": -0.09690394401550292, + "vlm_yes_no_margin": 3.074999988079071, + "vlm_token_credit": -0.09690394401550292, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3646610379219055 + ], + "opposite_logprob_tokens": [ + -1.2396609783172607 + ], + "logprob_reward": -0.3646610379219055, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.3646610379219055, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5613476634025574, + 0.02063051611185074, + 0.031231321394443512, + 0.04090406745672226, + 0.13064751029014587, + 0.0382559671998024, + 0.04520457237958908, + 0.08213219046592712, + 0.031741172075271606, + 0.01790502481162548 + ], + "credit_map": [ + 0.022413836792111397, + 0.24129073321819305, + 0.011463670991361141, + 0.011173411272466183, + 0.07959772646427155, + 0.16445067524909973, + 0.012656337581574917, + 0.018301257863640785, + 0.00305460998788476, + 0.0016022615600377321, + 0.001209246227517724, + 0.0016081170178949833, + 0.002982483245432377, + 0.006370742339640856, + 0.002093876712024212, + 0.0017091799527406693, + 0.003956572152674198, + 0.0028263754211366177, + 0.004585864022374153, + 0.0036373359616845846, + 0.0037924335338175297, + 0.00488679576665163, + 0.004605764057487249, + 0.0029401807114481926, + 0.006378520745784044, + 0.008399189449846745, + 0.003595285350456834, + 0.004970659967511892, + 0.006954842712730169, + 0.005309619475156069, + 0.0033042102586477995, + 0.0019917364697903395, + 0.040848299860954285, + 0.011082017794251442, + 0.008550974540412426, + 0.004568262491375208, + 0.005273012910038233, + 0.011318933218717575, + 0.011475227773189545, + 0.03753077611327171, + 0.004666744265705347, + 0.004464494064450264, + 0.006390092894434929, + 0.006884835194796324, + 0.004497827496379614, + 0.003616000758484006, + 0.004723389632999897, + 0.0030125805642455816, + 0.0055759986862540245, + 0.004567313008010387, + 0.007300089113414288, + 0.004881287459284067, + 0.005014933180063963, + 0.006717325188219547, + 0.0054824152030050755, + 0.005665209610015154, + 0.008461929857730865, + 0.016348587349057198, + 0.00908200815320015, + 0.007365713827311993, + 0.015232661738991737, + 0.013608316890895367, + 0.005788328591734171, + 0.00624464824795723, + 0.0062073650769889355, + 0.007059175055474043, + 0.003644424257799983, + 0.002804611576721072, + 0.0028259053360670805, + 0.0026449686847627163, + 0.0028827579226344824, + 0.003671964630484581, + 0.0016027992824092507, + 0.001138264429755509, + 0.004420040175318718, + 0.0032772505655884743, + 0.0013119076611474156, + 0.002436242066323757, + 0.002329309703782201, + 0.0013892104616388679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.3646610379219055 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012553717941045761 + ], + "opposite_logprob_tokens": [ + -4.387553691864014 + ], + "logprob_reward": -0.012553717941045761, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012553717941045761, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8479714393615723, + 0.061191774904727936, + 0.02750837244093418, + 0.004394846502691507, + 0.011795008555054665, + 0.0025673750787973404, + 0.014264875091612339, + 0.014307624660432339, + 0.008553003892302513, + 0.007445626892149448 + ], + "credit_map": [ + 0.03271094337105751, + 0.38731393218040466, + 0.00883195549249649, + 0.012098437175154686, + 0.14161622524261475, + 0.24035245180130005, + 0.011077815666794777, + 0.013969724997878075, + 0.01345084048807621, + 0.0038493566680699587, + 0.003966623917222023, + 0.004070747643709183, + 0.006667042151093483, + 0.017611656337976456, + 0.005630635656416416, + 0.0059448713436722755, + 0.0035953011829406023, + 0.002434012945741415, + 0.0025318660773336887, + 0.002746143378317356, + 0.0037180008366703987, + 0.005096908193081617, + 0.0033798564691096544, + 0.004006286151707172, + 0.0007141617825254798, + 0.0006231799488887191, + 0.0004722709418274462, + 0.0004567459982354194, + 0.0006294136983342469, + 0.0006566148949787021, + 0.0004223268188070506, + 0.00042013218626379967, + 0.0035637374967336655, + 0.0010164157720282674, + 0.0007665497832931578, + 0.00041338306618854403, + 0.0006216527544893324, + 0.001140050357207656, + 0.0010511549189686775, + 0.0032220648135989904, + 0.00035621554707176983, + 0.0002363455278100446, + 0.00024807584122754633, + 0.0003404677554499358, + 0.00028613212634809315, + 0.0003049817169085145, + 0.0002999529824592173, + 0.000495203654281795, + 0.0015794342616572976, + 0.0014113234356045723, + 0.0020189466886222363, + 0.0015878233825787902, + 0.0023539995308965445, + 0.0023828321136534214, + 0.0015337711665779352, + 0.0013967440463602543, + 0.0016760664293542504, + 0.0025387764908373356, + 0.0022832469549030066, + 0.0016717567341402173, + 0.001758093829266727, + 0.001926672412082553, + 0.0012481559533625841, + 0.0012048549251630902, + 0.0008702711202204227, + 0.0013336543925106525, + 0.0019603807013481855, + 0.0008033965132199228, + 0.0006739053642377257, + 0.0006389425252564251, + 0.0006564768846146762, + 0.0016159757506102324, + 0.00046645902330055833, + 0.00036981748417019844, + 0.0022773484233766794, + 0.002649732166901231, + 0.0004247649048920721, + 0.00041878074989654124, + 0.0005208931397646666, + 0.0003178308252245188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012553717941045761 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0486358143389225 + ], + "opposite_logprob_tokens": [ + -3.048635721206665 + ], + "logprob_reward": -0.0486358143389225, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0486358143389225, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8600149154663086, + 0.026496486738324165, + 0.019367117434740067, + 0.010238184593617916, + 0.037153709679841995, + 0.017131183296442032, + 0.008398487232625484, + 0.005897194147109985, + 0.012088282033801079, + 0.003214371856302023 + ], + "credit_map": [ + 0.03252803534269333, + 0.393616646528244, + 0.008403217419981956, + 0.012185684405267239, + 0.1334306299686432, + 0.25593262910842896, + 0.00983391422778368, + 0.014084188267588615, + 0.0046721878461539745, + 0.0015456982655450702, + 0.0012071550590917468, + 0.0018039497081190348, + 0.003312316257506609, + 0.008840627036988735, + 0.00279025174677372, + 0.002324299653992057, + 0.002708430401980877, + 0.002277093706652522, + 0.001962283393368125, + 0.0017774251755326986, + 0.0027150241658091545, + 0.003363149706274271, + 0.0028563037049025297, + 0.0017074078787118196, + 0.0014918388333171606, + 0.001796870375983417, + 0.0009010802605189383, + 0.001116182771511376, + 0.0016411874676123261, + 0.001740753068588674, + 0.0008693718700669706, + 0.0006808999460190535, + 0.01118176244199276, + 0.003198907943442464, + 0.002624249318614602, + 0.0018046338809654117, + 0.0018307336140424013, + 0.003195452969521284, + 0.0033675385639071465, + 0.009950431063771248, + 0.0022360614966601133, + 0.0023754455614835024, + 0.003222991479560733, + 0.0023044408299028873, + 0.002261863788589835, + 0.0016575410263612866, + 0.001757620251737535, + 0.0013152188621461391, + 0.0009918509749695659, + 0.0010931871365755796, + 0.0009908891515806317, + 0.0008963618893176317, + 0.0014741661725565791, + 0.001387180876918137, + 0.0007732590311206877, + 0.0007915918249636889, + 0.000735467008780688, + 0.0009721456444822252, + 0.0007877141470089555, + 0.0007452642894349992, + 0.0007849663961678743, + 0.0007066153921186924, + 0.000600648985709995, + 0.0005643725744448602, + 0.0018900485010817647, + 0.0012616817839443684, + 0.0018353896448388696, + 0.0011308389948680997, + 0.0010427606757730246, + 0.001221740385517478, + 0.0013440668117254972, + 0.002361754886806011, + 0.00026594195514917374, + 0.0002423008845653385, + 0.0005412918981164694, + 0.0007384067284874618, + 0.000248235824983567, + 0.00036622618790715933, + 0.0005529560730792582, + 0.000259012304013595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.0486358143389225 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810085728764534 + ], + "opposite_logprob_tokens": [ + -3.2881009578704834 + ], + "logprob_reward": -0.03810085728764534, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810085728764534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6641125679016113, + 0.04456767812371254, + 0.018458055332303047, + 0.010848748497664928, + 0.11302756518125534, + 0.01865549013018608, + 0.06676878780126572, + 0.023943541571497917, + 0.018901757895946503, + 0.020715847611427307 + ], + "credit_map": [ + 0.02331548184156418, + 0.301779180765152, + 0.00980041641741991, + 0.009617597796022892, + 0.1591116487979889, + 0.14151835441589355, + 0.00823880359530449, + 0.010731053538620472, + 0.010254474356770515, + 0.003092994447797537, + 0.0029366114176809788, + 0.0031989843118935823, + 0.004738697782158852, + 0.012198125943541527, + 0.003922725096344948, + 0.004225066397339106, + 0.002908386755734682, + 0.0017932524206116796, + 0.00127842603251338, + 0.0019892784766852856, + 0.00301956245675683, + 0.0029067075811326504, + 0.002694203983992338, + 0.0018682368099689484, + 0.001762943109497428, + 0.0015378044918179512, + 0.001229457906447351, + 0.0011386446421965957, + 0.0015632042195647955, + 0.001771870767697692, + 0.001136038568802178, + 0.0007087851408869028, + 0.03458890691399574, + 0.010211636312305927, + 0.007192355114966631, + 0.004213981796056032, + 0.00558629771694541, + 0.009245575405657291, + 0.010030503384768963, + 0.03195830434560776, + 0.002620363375172019, + 0.0018435384845361114, + 0.00304919620975852, + 0.0022662216797471046, + 0.0020175850950181484, + 0.001384939649142325, + 0.0019904552027583122, + 0.003483190666884184, + 0.007713571656495333, + 0.008219779469072819, + 0.007542754523456097, + 0.007400953210890293, + 0.011951831169426441, + 0.010505057871341705, + 0.0070257047191262245, + 0.006409136112779379, + 0.0026271373499184847, + 0.003317734692245722, + 0.002910969778895378, + 0.0034311015624552965, + 0.0044102296233177185, + 0.0026600572746247053, + 0.002370474161580205, + 0.0022158371284604073, + 0.004146869760006666, + 0.0036898497492074966, + 0.0029871801380068064, + 0.0017913111951202154, + 0.0019787480123341084, + 0.0017487411387264729, + 0.0013944340171292424, + 0.001164624118246138, + 0.0013390477979555726, + 0.001248038373887539, + 0.004348231945186853, + 0.007145120296627283, + 0.0013590733287855983, + 0.0012428174959495664, + 0.0028270387556403875, + 0.0012064805487170815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.03810085728764534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a descending drone view?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the lake deep indigo blue and bathed in early twilight pink-purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an elegant wooden flower boat gliding from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the boat's hull and sails adorned with pink peonies, white lilies, and lavender hyacinths dripping with dewdrops?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do rows of lotus-shaped paper lanterns emit a warm orange glow and hang from the boat's sides?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056829258799553 + ], + "opposite_logprob_tokens": [ + -3.895568370819092 + ], + "logprob_reward": -0.02056829258799553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056829258799553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5081261992454529, + 0.02981998398900032, + 0.030795494094491005, + 0.04015505686402321, + 0.2639210522174835, + 0.020514655858278275, + 0.011614999733865261, + 0.02774857170879841, + 0.02765454165637493, + 0.03964948281645775 + ], + "credit_map": [ + 0.01804587058722973, + 0.23385798931121826, + 0.004761704243719578, + 0.0064900824800133705, + 0.11673963069915771, + 0.11565983295440674, + 0.0052236285991966724, + 0.007347458973526955, + 0.005847462918609381, + 0.002069963840767741, + 0.0013588935835286975, + 0.0015571002149954438, + 0.003859610529616475, + 0.010169479064643383, + 0.00230780360288918, + 0.0026496665086597204, + 0.004582508932799101, + 0.0038644131273031235, + 0.003701812354847789, + 0.003455697325989604, + 0.003750067437067628, + 0.004959410522133112, + 0.0030080971773713827, + 0.0034734862856566906, + 0.005852376576513052, + 0.007388186175376177, + 0.003918769769370556, + 0.0042317332699894905, + 0.007546989247202873, + 0.00660725636407733, + 0.002614135853946209, + 0.001995611237362027, + 0.07980367541313171, + 0.021960802376270294, + 0.020290745422244072, + 0.012460864149034023, + 0.011506635695695877, + 0.023091118782758713, + 0.023198315873742104, + 0.07160888612270355, + 0.0028723464347422123, + 0.0021954288240522146, + 0.0038722294848412275, + 0.002950719790533185, + 0.0018231532303616405, + 0.0017424506368115544, + 0.0019328530179336667, + 0.0031254757195711136, + 0.0011520544067025185, + 0.001624109223484993, + 0.0021935822442173958, + 0.0011385990073904395, + 0.0017994862282648683, + 0.00157754251267761, + 0.0010812769178301096, + 0.0010483491932973266, + 0.0030282377265393734, + 0.003492710879072547, + 0.004331307485699654, + 0.0034142101649194956, + 0.004078367725014687, + 0.004149499349296093, + 0.002723053330555558, + 0.002531185280531645, + 0.005443172995001078, + 0.004138307645916939, + 0.0037536765448749065, + 0.0026946121361106634, + 0.0035095238126814365, + 0.0024756318889558315, + 0.002101748948916793, + 0.003537866985425353, + 0.004496571142226458, + 0.0035988581366837025, + 0.00885610468685627, + 0.006203454453498125, + 0.004387883935123682, + 0.003298445837572217, + 0.005465399939566851, + 0.003342763753607869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.02056829258799553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_tensor([4081, 4081])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_tensor([4081, 4081])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ddcc924ea0d23513a253298f2d029be294495fb4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_tensor([4081, 4081])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "This static, eye-level medium shot captures a woman with a chic, short, slightly tousled bob, sitting at an outdoor cafe. She wears a short-sleeved shirt made of lightweight fabric with a delicate, pastel floral pattern against a cream background. She sits comfortably at a small, round cast-iron table on the sunny restaurant terrace, leaning slightly forward and holding a modern smartphone to her right ear. Amidst a lively conversation, a warm, relaxed smile graces her lips, her eyes crinkling slightly as she speaks, her head tilted with interest. Her left hand rests on the tabletop, her fingers leisurely stroking the rim of a glass of espresso iced tea. In the background, the restaurant's blurred facade, with its large windows and dark green awning, is visible, while the soft, out-of-focus shapes of other patrons create a gentle sense of ambient activity. The entire scene is illuminated by the warm midday sun, which filters through a large parasol, casting a soft, diffuse glow that eliminates harsh shadows and creates an overall relaxed, down-to-earth, and joyful atmosphere of a pleasant afternoon.", + "seed": 18320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_4081_seed_18320_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01775651201605797, + "vlm_logprob": -0.01775651201605797, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.01775651201605797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015761440619826317 + ], + "opposite_logprob_tokens": [ + -5.140761375427246 + ], + "logprob_reward": -0.015761440619826317, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.015761440619826317, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7413746118545532, + 0.08636138588190079, + 0.023488769307732582, + 0.029980560764670372, + 0.01464762818068266, + 0.014299566857516766, + 0.018954286351799965, + 0.01650881953537464, + 0.028218762949109077, + 0.026165630668401718 + ], + "credit_map": [ + 0.0575927197933197, + 0.06363871693611145, + 0.24088548123836517, + 0.044080961495637894, + 0.24448807537555695, + 0.0404314287006855, + 0.023264693096280098, + 0.0269925557076931, + 0.008801800198853016, + 0.008685954846441746, + 0.02304447442293167, + 0.004913691896945238, + 0.025299793109297752, + 0.006188974250108004, + 0.003647104138508439, + 0.005779589526355267, + 0.004211886320263147, + 0.0021438784897327423, + 0.004638066980987787, + 0.002483072690665722, + 0.0024977000430226326, + 0.0033232993446290493, + 0.0016367787029594183, + 0.0025540885981172323, + 0.00856172014027834, + 0.003113703103736043, + 0.004588671028614044, + 0.0022822527680546045, + 0.0035642795264720917, + 0.003180096857249737, + 0.00195417832583189, + 0.0027356594800949097, + 0.002847143914550543, + 0.0018793533090502024, + 0.0022658228408545256, + 0.001175743993371725, + 0.001966047566384077, + 0.0012518750736489892, + 0.0011919249081984162, + 0.0020697161089628935, + 0.0019161808304488659, + 0.0012760377721861005, + 0.0036634576972573996, + 0.0013673835201188922, + 0.0017886013956740499, + 0.001615240122191608, + 0.0012210301356390119, + 0.0014516360824927688, + 0.002954853931441903, + 0.0015512699028477073, + 0.003578047500923276, + 0.002132752211764455, + 0.0019729495979845524, + 0.002385711995884776, + 0.001848201616667211, + 0.002530500292778015, + 0.0021129008382558823, + 0.001637998386286199, + 0.0033888877369463444, + 0.0022182094398885965, + 0.0021295526530593634, + 0.0019022616324946284, + 0.0014727303059771657, + 0.0016462794737890363, + 0.004523810930550098, + 0.0023179911077022552, + 0.004294249694794416, + 0.0020854915492236614, + 0.0054237316362559795, + 0.002856447361409664, + 0.0020866321865469217, + 0.004630411509424448, + 0.0034972880966961384, + 0.0020860927179455757, + 0.0053697009570896626, + 0.0023473433684557676, + 0.005265918094664812, + 0.002763034775853157, + 0.0026162920985370874, + 0.002219961490482092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.015761440619826317 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555483728647232 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.012555483728647232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555483728647232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21220001578330994, + 0.18677039444446564, + 0.09546233713626862, + 0.10075315088033676, + 0.07043012976646423, + 0.07027602940797806, + 0.058560457080602646, + 0.09445463865995407, + 0.026883531361818314, + 0.08420931547880173 + ], + "credit_map": [ + 0.026330430060625076, + 0.019665898755192757, + 0.04491618648171425, + 0.028570029884576797, + 0.03594861179590225, + 0.020092612132430077, + 0.017608778551220894, + 0.019067471846938133, + 0.024002330377697945, + 0.021439773961901665, + 0.038248080760240555, + 0.0220999363809824, + 0.04099641740322113, + 0.01321779191493988, + 0.015050527639687061, + 0.011715533211827278, + 0.015685053542256355, + 0.009569794870913029, + 0.018902983516454697, + 0.013926160521805286, + 0.014080807566642761, + 0.009890985675156116, + 0.006886369548738003, + 0.006520182825624943, + 0.023710912093520164, + 0.010737212374806404, + 0.018841970711946487, + 0.009377769194543362, + 0.009868254885077477, + 0.011303898878395557, + 0.0063659329898655415, + 0.010547192767262459, + 0.012720992788672447, + 0.006991197355091572, + 0.013844478875398636, + 0.008992742747068405, + 0.011811649426817894, + 0.00559036573395133, + 0.0046038320288062096, + 0.0058748722076416016, + 0.010636790655553341, + 0.005181582178920507, + 0.018410826101899147, + 0.006343464832752943, + 0.016254901885986328, + 0.005139347165822983, + 0.004333545919507742, + 0.003975569736212492, + 0.008469797670841217, + 0.005140299908816814, + 0.00948118232190609, + 0.005824859254062176, + 0.005642622709274292, + 0.008979640901088715, + 0.005070994608104229, + 0.00995105691254139, + 0.011766519397497177, + 0.008158748038113117, + 0.017977027222514153, + 0.008519170805811882, + 0.0236449483782053, + 0.00821894221007824, + 0.007004871964454651, + 0.009164409711956978, + 0.00605061836540699, + 0.0028243025299161673, + 0.004470016807317734, + 0.0032442042138427496, + 0.003361559472978115, + 0.002430281601846218, + 0.002318499842658639, + 0.002184047596529126, + 0.006180585362017155, + 0.01007843017578125, + 0.015389603562653065, + 0.01854017749428749, + 0.014277763664722443, + 0.006183757446706295, + 0.008183878846466541, + 0.00537512032315135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.012555483728647232 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781880304217339 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781880304217339, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781880304217339, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33494308590888977, + 0.257377952337265, + 0.09188414365053177, + 0.07837075740098953, + 0.0449720062315464, + 0.07669481635093689, + 0.03322029858827591, + 0.032988060265779495, + 0.02181745506823063, + 0.02773147262632847 + ], + "credit_map": [ + 0.02654927596449852, + 0.027696970850229263, + 0.09918445348739624, + 0.026329465210437775, + 0.09312251955270767, + 0.027424650266766548, + 0.015927128493785858, + 0.0187086071819067, + 0.01910421997308731, + 0.020121101289987564, + 0.0822562649846077, + 0.01940096728503704, + 0.07603422552347183, + 0.016558699309825897, + 0.010616336017847061, + 0.01328614354133606, + 0.009101781994104385, + 0.007118747103959322, + 0.024809829890727997, + 0.01080633606761694, + 0.018400222063064575, + 0.00802650023251772, + 0.0071099004708230495, + 0.0065108188427984715, + 0.014208409935235977, + 0.0073147304356098175, + 0.017464835196733475, + 0.008044415153563023, + 0.010464424267411232, + 0.006816677283495665, + 0.005684150382876396, + 0.008373119868338108, + 0.006996389478445053, + 0.004947088658809662, + 0.009408939629793167, + 0.005450527183711529, + 0.006816687062382698, + 0.004160153679549694, + 0.0036225360818207264, + 0.0035696865525096655, + 0.0063635543920099735, + 0.006739764008671045, + 0.02305949665606022, + 0.007247025612741709, + 0.02128080651164055, + 0.004582890309393406, + 0.004008660092949867, + 0.0034126141108572483, + 0.0054853069595992565, + 0.0031429112423211336, + 0.005588338244706392, + 0.0044013760052621365, + 0.0034063151106238365, + 0.0038098131772130728, + 0.003314586589112878, + 0.004071652889251709, + 0.003501954022794962, + 0.004319478757679462, + 0.007091898936778307, + 0.004254162777215242, + 0.004570085555315018, + 0.003113572718575597, + 0.003834580769762397, + 0.0023023257963359356, + 0.0032802773639559746, + 0.0028543169610202312, + 0.0032289111986756325, + 0.00226525217294693, + 0.0027361151296645403, + 0.003012285102158785, + 0.0024850021582096815, + 0.0019552945159375668, + 0.0036652155686169863, + 0.0035251029767096043, + 0.0035729906521737576, + 0.002204687800258398, + 0.0042069354094564915, + 0.004381892271339893, + 0.002973122987896204, + 0.0032015247270464897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.009781880304217339 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305894672870636 + ], + "opposite_logprob_tokens": [ + -3.1680588722229004 + ], + "logprob_reward": -0.04305894672870636, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305894672870636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6718087792396545, + 0.11002275347709656, + 0.02731192857027054, + 0.06989990919828415, + 0.019414136186242104, + 0.02371184341609478, + 0.024752600118517876, + 0.0276965219527483, + 0.01047111302614212, + 0.014910375699400902 + ], + "credit_map": [ + 0.05148644000291824, + 0.06053544953465462, + 0.21438461542129517, + 0.03974977508187294, + 0.21686702966690063, + 0.03815171495079994, + 0.023834671825170517, + 0.026799092069268227, + 0.010282974690198898, + 0.011912999674677849, + 0.02837064117193222, + 0.007168745156377554, + 0.03279824182391167, + 0.0077224490232765675, + 0.004607828333973885, + 0.0071588787250220776, + 0.004724493715912104, + 0.0027339619118720293, + 0.004619115497916937, + 0.0030761128291487694, + 0.0026872099842876196, + 0.004178842529654503, + 0.002415329683572054, + 0.002876861486583948, + 0.010414649732410908, + 0.005344030447304249, + 0.014194079674780369, + 0.003701224457472563, + 0.02008356712758541, + 0.0061716921627521515, + 0.004334410652518272, + 0.005656256340444088, + 0.0023123188875615597, + 0.00203114515170455, + 0.003728340845555067, + 0.003781309351325035, + 0.002982861129567027, + 0.0012025343021377921, + 0.001693764585070312, + 0.001681861816905439, + 0.004082684405148029, + 0.0016878199530765414, + 0.005561180412769318, + 0.002507543656975031, + 0.005253284703940153, + 0.00161191716324538, + 0.00156252714805305, + 0.0014448873698711395, + 0.0031778658740222454, + 0.0023427936248481274, + 0.0036778568755835295, + 0.002069719135761261, + 0.0026969348546117544, + 0.003917911555618048, + 0.002195041859522462, + 0.004674477037042379, + 0.003124796086922288, + 0.002494339132681489, + 0.005078651010990143, + 0.002145949052646756, + 0.0066503188572824, + 0.002510479651391506, + 0.002232096390798688, + 0.0034598889760673046, + 0.001440082909539342, + 0.0011249551316723228, + 0.0016862984048202634, + 0.0011089671170338988, + 0.0012855478562414646, + 0.0013727101031690836, + 0.0013707599136978388, + 0.0010817924048751593, + 0.002732819877564907, + 0.001562121557071805, + 0.002263738075271249, + 0.00146809674333781, + 0.0016413085395470262, + 0.0022213251795619726, + 0.001477598212659359, + 0.0015433670487254858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.04305894672870636 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624808698892593 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624808698892593, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624808698892593, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7791239023208618, + 0.07110124826431274, + 0.02198263816535473, + 0.013339563272893429, + 0.01624160259962082, + 0.019615434110164642, + 0.01633352041244507, + 0.03414899483323097, + 0.015397497452795506, + 0.012715693563222885 + ], + "credit_map": [ + 0.05815623328089714, + 0.06653165817260742, + 0.2573559880256653, + 0.04634612798690796, + 0.25468316674232483, + 0.04319823905825615, + 0.025065667927265167, + 0.027786724269390106, + 0.005764496512711048, + 0.006268197670578957, + 0.022182388231158257, + 0.004669009242206812, + 0.02238890901207924, + 0.003955580294132233, + 0.002508071018382907, + 0.0033645902294665575, + 0.0025268341414630413, + 0.0019384358311071992, + 0.005204236600548029, + 0.002533707534894347, + 0.0038191049825400114, + 0.0018848944455385208, + 0.0017562618013471365, + 0.002319159684702754, + 0.0024820303078740835, + 0.001512020593509078, + 0.0017205722397193313, + 0.001485019107349217, + 0.0018068138742819428, + 0.0014796658651903272, + 0.0012274399632588029, + 0.0016259995754808187, + 0.003042778465896845, + 0.0016608687583357096, + 0.00298490421846509, + 0.002297468949109316, + 0.0014805634273216128, + 0.001436106744222343, + 0.0015500356676056981, + 0.0017888726433739066, + 0.0020280356984585524, + 0.0016293451189994812, + 0.005188962444663048, + 0.0015781607944518328, + 0.005196271929889917, + 0.0013893930008634925, + 0.0011524047004058957, + 0.001452856813557446, + 0.002547441516071558, + 0.00141718122176826, + 0.002562896115705371, + 0.0015274178003892303, + 0.0017953346250578761, + 0.002221250906586647, + 0.0013811676762998104, + 0.0028808279894292355, + 0.003487240057438612, + 0.003359361318871379, + 0.005697993095964193, + 0.002406833926215768, + 0.007450570818036795, + 0.0032927519641816616, + 0.0027693435549736023, + 0.005684895906597376, + 0.0017293983837589622, + 0.0015217815525829792, + 0.0027011472266167402, + 0.0016707050381228328, + 0.002655911725014448, + 0.0012560023460537195, + 0.0013215224025771022, + 0.0025410265661776066, + 0.0011522592976689339, + 0.0012645409442484379, + 0.002533978084102273, + 0.0010842077899724245, + 0.0026816874742507935, + 0.0010757965501397848, + 0.0012043199967592955, + 0.0017189019126817584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.007624808698892593 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium shot captures a woman with a chic, short, slightly tousled bob, sitting at an outdoor cafe. She wears a short-sleeved shirt made of lightweight fabric with a delicate, pastel floral pattern against a cream background. She sits comfortably at a small, round cast-iron table on the sunny restaurant terrace, leaning slightly forward and holding a modern smartphone to her right ear. Amidst a lively conversation, a warm, relaxed smile graces her lips, her eyes crinkling slightly as she speaks, her head tilted with interest. Her left hand rests on the tabletop, her fingers leisurely stroking the rim of a glass of espresso iced tea. In the background, the restaurant's blurred facade, with its large windows and dark green awning, is visible, while the soft, out-of-focus shapes of other patrons create a gentle sense of ambient activity. The entire scene is illuminated by the warm midday sun, which filters through a large parasol, casting a soft, diffuse glow that eliminates harsh shadows and creates an overall relaxed, down-to-earth, and joyful atmosphere of a pleasant afternoon.", + "seed": 18321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_4081_seed_18321_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.027269671391695737, + "vlm_logprob": -0.027269671391695737, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.027269671391695737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03240705654025078 + ], + "opposite_logprob_tokens": [ + -4.032407283782959 + ], + "logprob_reward": -0.03240705654025078, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.03240705654025078, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5522868037223816, + 0.03407535329461098, + 0.037502024322748184, + 0.08923428505659103, + 0.07958924770355225, + 0.05780171602964401, + 0.03714568912982941, + 0.06980664283037186, + 0.020734228193759918, + 0.021824048832058907 + ], + "credit_map": [ + 0.049451008439064026, + 0.05398201197385788, + 0.03215065598487854, + 0.135890394449234, + 0.03911203518509865, + 0.13154105842113495, + 0.07560106366872787, + 0.03455854579806328, + 0.0026170597411692142, + 0.004546963609755039, + 0.002970999339595437, + 0.005831991787999868, + 0.006012949626892805, + 0.004901384934782982, + 0.004543785471469164, + 0.002650216454640031, + 0.0030608316883444786, + 0.005253307055681944, + 0.0031700190156698227, + 0.005318745505064726, + 0.004305075388401747, + 0.0032075641211122274, + 0.010459136217832565, + 0.0027273432351648808, + 0.009518824517726898, + 0.006944108288735151, + 0.004674871917814016, + 0.02519497647881508, + 0.00552118755877018, + 0.02183402329683304, + 0.0093442527577281, + 0.006202029995620251, + 0.008186304941773415, + 0.007895155809819698, + 0.004718356765806675, + 0.019798679277300835, + 0.007110911887139082, + 0.017299091443419456, + 0.010152793489396572, + 0.0044279443100094795, + 0.005645092111080885, + 0.00715608149766922, + 0.004082648549228907, + 0.012758207507431507, + 0.007039369083940983, + 0.00799579732120037, + 0.009300779551267624, + 0.003823736449703574, + 0.004437235649675131, + 0.0035654606763273478, + 0.00324668874964118, + 0.007879096083343029, + 0.003641513641923666, + 0.0046241022646427155, + 0.006134465802460909, + 0.003617125330492854, + 0.006653815507888794, + 0.01001301221549511, + 0.006645684130489826, + 0.011985601857304573, + 0.0082932747900486, + 0.00791978370398283, + 0.012467882595956326, + 0.005827585235238075, + 0.0021447576582431793, + 0.0033861228730529547, + 0.0018064171308651567, + 0.004367463756352663, + 0.0015035656979307532, + 0.0031317221000790596, + 0.0022612682078033686, + 0.002132908208295703, + 0.0020299453753978014, + 0.004661476705223322, + 0.001848401501774788, + 0.002931251423433423, + 0.0020044823177158833, + 0.003001188160851598, + 0.003150259843096137, + 0.002197044203057885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.03240705654025078 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01609065569937229 + ], + "opposite_logprob_tokens": [ + -4.1410908699035645 + ], + "logprob_reward": -0.01609065569937229, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01609065569937229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42462483048439026, + 0.10222011804580688, + 0.04760236665606499, + 0.105254165828228, + 0.1391066610813141, + 0.0431375578045845, + 0.018573926761746407, + 0.04770534113049507, + 0.018665902316570282, + 0.05310908332467079 + ], + "credit_map": [ + 0.039525460451841354, + 0.040845837444067, + 0.024733131751418114, + 0.1165693923830986, + 0.0230672936886549, + 0.11563879251480103, + 0.03284269571304321, + 0.031402263790369034, + 0.008042367175221443, + 0.01193773653358221, + 0.011517654173076153, + 0.017904063686728477, + 0.011129003018140793, + 0.013868940062820911, + 0.021158140152692795, + 0.006662201136350632, + 0.003612791420891881, + 0.00541331572458148, + 0.0038060809019953012, + 0.008177262730896473, + 0.0077255466021597385, + 0.006349792238324881, + 0.009282208047807217, + 0.0032353687565773726, + 0.011524777859449387, + 0.008909530006349087, + 0.005662959534674883, + 0.03007327765226364, + 0.006623142398893833, + 0.026045748963952065, + 0.008972530253231525, + 0.007442205213010311, + 0.014032539911568165, + 0.017550991848111153, + 0.008164777420461178, + 0.03380453959107399, + 0.014067953452467918, + 0.0257112979888916, + 0.017309270799160004, + 0.008465291932225227, + 0.0034216647036373615, + 0.006333678960800171, + 0.00465205404907465, + 0.0056555308401584625, + 0.004204345401376486, + 0.005856935400515795, + 0.007916177622973919, + 0.0050971717573702335, + 0.001923207426443696, + 0.0025584541726857424, + 0.002383903134614229, + 0.0023922512773424387, + 0.0019590859301388264, + 0.0019022268243134022, + 0.002967802109196782, + 0.002486995654180646, + 0.004846736788749695, + 0.006885917857289314, + 0.004024308640509844, + 0.008187566883862019, + 0.005209635011851788, + 0.005088457837700844, + 0.009481336921453476, + 0.00398138165473938, + 0.0016122658271342516, + 0.0027175622526556253, + 0.0024089592043310404, + 0.003228019457310438, + 0.002481058705598116, + 0.0019462734926491976, + 0.002527835313230753, + 0.001743927365168929, + 0.004684313666075468, + 0.008338015526533127, + 0.0038658338598906994, + 0.008207185193896294, + 0.006403753999620676, + 0.005071016028523445, + 0.012444732710719109, + 0.004094228148460388 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.01609065569937229 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018184026703238487 + ], + "opposite_logprob_tokens": [ + -4.018184185028076 + ], + "logprob_reward": -0.018184026703238487, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018184026703238487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2458336353302002, + 0.1212783083319664, + 0.04510565847158432, + 0.044000763446092606, + 0.3733971118927002, + 0.04421082139015198, + 0.07372099161148071, + 0.01852383092045784, + 0.01100111473351717, + 0.022927729412913322 + ], + "credit_map": [ + 0.019446343183517456, + 0.025954965502023697, + 0.014634703285992146, + 0.059662144631147385, + 0.01914079114794731, + 0.07108350843191147, + 0.018715007230639458, + 0.017196204513311386, + 0.010571942664682865, + 0.017650116235017776, + 0.006252976134419441, + 0.023195341229438782, + 0.016466699540615082, + 0.015608556568622589, + 0.024070924147963524, + 0.007461758330464363, + 0.0034875329583883286, + 0.005872517824172974, + 0.0028804978355765343, + 0.00705735944211483, + 0.007290919777005911, + 0.005330656189471483, + 0.00865229219198227, + 0.004533884581178427, + 0.004393254406750202, + 0.00347883440554142, + 0.00282662152312696, + 0.011589737609028816, + 0.0030097858980298042, + 0.010153306648135185, + 0.0043875607661902905, + 0.004161670804023743, + 0.03720001131296158, + 0.03565042093396187, + 0.02185234986245632, + 0.09717914462089539, + 0.03851313889026642, + 0.07913711667060852, + 0.0438113659620285, + 0.020053623244166374, + 0.003954646643251181, + 0.00491405138745904, + 0.0048012761399149895, + 0.006874318700283766, + 0.005244510248303413, + 0.005377291236072779, + 0.00904056429862976, + 0.004004167392849922, + 0.006152585614472628, + 0.010258185677230358, + 0.004348319489508867, + 0.012889878824353218, + 0.010603494942188263, + 0.008284392766654491, + 0.016513878479599953, + 0.004670262802392244, + 0.001646319986321032, + 0.005012921988964081, + 0.0015887884655967355, + 0.0022357706911861897, + 0.0021243717055767775, + 0.0019598575308918953, + 0.0022264777217060328, + 0.0017293259734287858, + 0.0011061137774959207, + 0.0019815678242594004, + 0.0008545711170881987, + 0.0017937642987817526, + 0.0008394212345592678, + 0.0012544890632852912, + 0.001706184004433453, + 0.0014650053344666958, + 0.002084206324070692, + 0.006048927549272776, + 0.0019488385878503323, + 0.0029480503872036934, + 0.0021162203047424555, + 0.0019042944768443704, + 0.0028029975946992636, + 0.003074196632951498 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.018184026703238487 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204346567392349 + ], + "opposite_logprob_tokens": [ + -2.8120434284210205 + ], + "logprob_reward": -0.06204346567392349, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204346567392349, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34371694922447205, + 0.13160675764083862, + 0.04024383798241615, + 0.09547244757413864, + 0.15327893197536469, + 0.042353156954050064, + 0.08383011817932129, + 0.051349081099033356, + 0.024727992713451385, + 0.033420730382204056 + ], + "credit_map": [ + 0.03244727477431297, + 0.035556595772504807, + 0.021510884165763855, + 0.0767173022031784, + 0.04341446980834007, + 0.06847584992647171, + 0.036343738436698914, + 0.02925082854926586, + 0.009559654630720615, + 0.01903153769671917, + 0.007043465040624142, + 0.020855993032455444, + 0.020678026601672173, + 0.015081099234521389, + 0.030437199398875237, + 0.008919794112443924, + 0.003333898726850748, + 0.005308729130774736, + 0.004105828702449799, + 0.004965238273143768, + 0.0038901008665561676, + 0.005254013929516077, + 0.006630510091781616, + 0.006755518727004528, + 0.010272878222167492, + 0.00912559125572443, + 0.005914045497775078, + 0.02328677847981453, + 0.010783870704472065, + 0.0170577522367239, + 0.01035084668546915, + 0.008680684491991997, + 0.015893565490841866, + 0.013821421191096306, + 0.009282820858061314, + 0.041371677070856094, + 0.013327440246939659, + 0.03500073030591011, + 0.0161251462996006, + 0.008456124924123287, + 0.004998095333576202, + 0.003735622623935342, + 0.005464568268507719, + 0.006388248410075903, + 0.003722492139786482, + 0.005530917085707188, + 0.007146693300455809, + 0.005366520490497351, + 0.0069381981156766415, + 0.010354644618928432, + 0.004747355356812477, + 0.014463918283581734, + 0.013926750980317593, + 0.009981838054955006, + 0.018083348870277405, + 0.005334064830094576, + 0.0045418995432555676, + 0.009098435752093792, + 0.0047818017192184925, + 0.009358367882668972, + 0.0070201195776462555, + 0.006397258955985308, + 0.006263596471399069, + 0.0038876018952578306, + 0.0023887907154858112, + 0.004045203328132629, + 0.0034071411937475204, + 0.0031293476931750774, + 0.0026102338451892138, + 0.002381189027801156, + 0.003790296148508787, + 0.002975791459903121, + 0.0029085257556289434, + 0.005182703956961632, + 0.002848114352673292, + 0.007346325553953648, + 0.0027059102430939674, + 0.004754831548780203, + 0.004536963999271393, + 0.0031373542733490467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.06204346567392349 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762315234169364 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.00762315234169364, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762315234169364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34061214327812195, + 0.17665629088878632, + 0.054326049983501434, + 0.05510008707642555, + 0.10570168495178223, + 0.057152073830366135, + 0.03595126420259476, + 0.09508874267339706, + 0.029670026153326035, + 0.0497417114675045 + ], + "credit_map": [ + 0.03254149854183197, + 0.05447165668010712, + 0.0425083301961422, + 0.06950333714485168, + 0.02081584744155407, + 0.05508628860116005, + 0.031012732535600662, + 0.034672390669584274, + 0.017726672813296318, + 0.026024049147963524, + 0.022154148668050766, + 0.02514505572617054, + 0.018486134707927704, + 0.01999366283416748, + 0.03382471576333046, + 0.013301840983331203, + 0.005565169267356396, + 0.007327580824494362, + 0.003688458353281021, + 0.008353469893336296, + 0.005315047688782215, + 0.005519727244973183, + 0.006844919640570879, + 0.011711673811078072, + 0.0057685053907334805, + 0.006355340592563152, + 0.004670928698033094, + 0.013006981462240219, + 0.005830026231706142, + 0.009824758395552635, + 0.005963551811873913, + 0.00367999030277133, + 0.009889288805425167, + 0.011778601445257664, + 0.007720350753515959, + 0.024809788912534714, + 0.010547824203968048, + 0.02203192189335823, + 0.013141127303242683, + 0.005782771855592728, + 0.005273403599858284, + 0.005938749294728041, + 0.007195433601737022, + 0.008165308274328709, + 0.005029246211051941, + 0.006937885656952858, + 0.008320467546582222, + 0.010291577316820621, + 0.003394368104636669, + 0.005576795898377895, + 0.005758190061897039, + 0.004959475714713335, + 0.002567516639828682, + 0.0037013902328908443, + 0.0044207144528627396, + 0.0055728089064359665, + 0.008900744840502739, + 0.012430782429873943, + 0.00909842737019062, + 0.015334016643464565, + 0.010650035925209522, + 0.0076805162243545055, + 0.0235423780977726, + 0.007451837416738272, + 0.0021396575029939413, + 0.0048239510506391525, + 0.0029109343886375427, + 0.004188716411590576, + 0.0024876368697732687, + 0.0030109568033367395, + 0.006031477823853493, + 0.004076692275702953, + 0.004672624170780182, + 0.00955889094620943, + 0.005179479252547026, + 0.0065945107489824295, + 0.005727965384721756, + 0.005478037055581808, + 0.00802753958851099, + 0.004502659663558006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00762315234169364 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium shot captures a woman with a chic, short, slightly tousled bob, sitting at an outdoor cafe. She wears a short-sleeved shirt made of lightweight fabric with a delicate, pastel floral pattern against a cream background. She sits comfortably at a small, round cast-iron table on the sunny restaurant terrace, leaning slightly forward and holding a modern smartphone to her right ear. Amidst a lively conversation, a warm, relaxed smile graces her lips, her eyes crinkling slightly as she speaks, her head tilted with interest. Her left hand rests on the tabletop, her fingers leisurely stroking the rim of a glass of espresso iced tea. In the background, the restaurant's blurred facade, with its large windows and dark green awning, is visible, while the soft, out-of-focus shapes of other patrons create a gentle sense of ambient activity. The entire scene is illuminated by the warm midday sun, which filters through a large parasol, casting a soft, diffuse glow that eliminates harsh shadows and creates an overall relaxed, down-to-earth, and joyful atmosphere of a pleasant afternoon.", + "seed": 18322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_4081_seed_18322_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.027297281380742788, + "vlm_logprob": -0.027297281380742788, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.027297281380742788, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018434155732393265 + ], + "opposite_logprob_tokens": [ + -4.768434047698975 + ], + "logprob_reward": -0.018434155732393265, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018434155732393265, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2523181438446045, + 0.08310531824827194, + 0.07658332586288452, + 0.06911701709032059, + 0.08581933379173279, + 0.07834906131029129, + 0.11423710733652115, + 0.056764155626297, + 0.0791323110461235, + 0.10457421094179153 + ], + "credit_map": [ + 0.030445696786046028, + 0.02825877256691456, + 0.02677319385111332, + 0.047706443816423416, + 0.02319221943616867, + 0.028532724827528, + 0.03608061373233795, + 0.03132849186658859, + 0.007146755233407021, + 0.007600194308906794, + 0.008924484252929688, + 0.018850702792406082, + 0.010172859765589237, + 0.0076699997298419476, + 0.015376332215964794, + 0.0073639946058392525, + 0.005974342580884695, + 0.008531233295798302, + 0.008955206722021103, + 0.01612279936671257, + 0.010359163396060467, + 0.007472275290638208, + 0.01261715404689312, + 0.006551145110279322, + 0.005751768127083778, + 0.013323311693966389, + 0.0054524983279407024, + 0.011396873742341995, + 0.006146632134914398, + 0.0058972216211259365, + 0.010602126829326153, + 0.010546584613621235, + 0.008659246377646923, + 0.015314965508878231, + 0.0069351401180028915, + 0.015823103487491608, + 0.009172871708869934, + 0.007605002727359533, + 0.015150387771427631, + 0.007158623076975346, + 0.007330415304750204, + 0.006477802060544491, + 0.00720088928937912, + 0.01682323031127453, + 0.008525043725967407, + 0.008338779211044312, + 0.01566183753311634, + 0.007991059683263302, + 0.010745054110884666, + 0.017682190984487534, + 0.009862610138952732, + 0.02032112516462803, + 0.01237097941339016, + 0.008491797372698784, + 0.01820196583867073, + 0.016561385244131088, + 0.009199153631925583, + 0.00623740628361702, + 0.0061667924746870995, + 0.007807022426277399, + 0.006216943264007568, + 0.0077039506286382675, + 0.00648637767881155, + 0.006946512497961521, + 0.00811434630304575, + 0.0066625322215259075, + 0.007678130641579628, + 0.013878806494176388, + 0.00886812899261713, + 0.009650591760873795, + 0.01584882289171219, + 0.008430957794189453, + 0.013942628167569637, + 0.01007797010242939, + 0.009242609143257141, + 0.018351959064602852, + 0.010122057981789112, + 0.014622456394135952, + 0.01758182793855667, + 0.010632703080773354 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.018434155732393265 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012561723589897156 + ], + "opposite_logprob_tokens": [ + -4.387561798095703 + ], + "logprob_reward": -0.012561723589897156, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012561723589897156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2416338175535202, + 0.24447636306285858, + 0.05592373013496399, + 0.06588082760572433, + 0.07201037555932999, + 0.05603061243891716, + 0.08330157399177551, + 0.04146549105644226, + 0.08235006034374237, + 0.05692712217569351 + ], + "credit_map": [ + 0.022346412762999535, + 0.0331287756562233, + 0.024128280580043793, + 0.04864637553691864, + 0.02672831527888775, + 0.023916244506835938, + 0.042901452630758286, + 0.019837960600852966, + 0.01949046552181244, + 0.036514248698949814, + 0.016934972256422043, + 0.05390250310301781, + 0.029278112575411797, + 0.018145164474844933, + 0.04311775043606758, + 0.027093160897493362, + 0.005179155617952347, + 0.006197376176714897, + 0.005981753580272198, + 0.0113271065056324, + 0.0051147593185305595, + 0.007969425059854984, + 0.008588794618844986, + 0.005565362516790628, + 0.005921506322920322, + 0.013334381394088268, + 0.005911691579967737, + 0.011398056522011757, + 0.006779712624847889, + 0.00595160573720932, + 0.010169782675802708, + 0.006414097733795643, + 0.010116932913661003, + 0.014798000454902649, + 0.005747104994952679, + 0.010437398217618465, + 0.008983135223388672, + 0.006176625844091177, + 0.010248755104839802, + 0.0055024279281497, + 0.007058567833155394, + 0.00769174424931407, + 0.004910371266305447, + 0.010536856018006802, + 0.006364408880472183, + 0.004588254261761904, + 0.009989227168262005, + 0.004891180898994207, + 0.008681487292051315, + 0.01000016089528799, + 0.00738192442804575, + 0.017691299319267273, + 0.009489466436207294, + 0.005642903037369251, + 0.015542341396212578, + 0.008871993981301785, + 0.0054996726103127, + 0.0066454969346523285, + 0.004227471072226763, + 0.0076574585400521755, + 0.0035487220156937838, + 0.004514849744737148, + 0.006042595952749252, + 0.0033292288426309824, + 0.014482381753623486, + 0.006535849533975124, + 0.0057608564384281635, + 0.016699064522981644, + 0.00769162317737937, + 0.005768304690718651, + 0.017188943922519684, + 0.008223037235438824, + 0.007042342331260443, + 0.0051510995253920555, + 0.004530891310423613, + 0.011342437006533146, + 0.006391165778040886, + 0.007188013289123774, + 0.009920311160385609, + 0.005360864568501711 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.012561723589897156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020577168092131615 + ], + "opposite_logprob_tokens": [ + -3.8955771923065186 + ], + "logprob_reward": -0.020577168092131615, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020577168092131615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23665012419223785, + 0.20361587405204773, + 0.07334724068641663, + 0.08105917274951935, + 0.08416541665792465, + 0.06698844581842422, + 0.08628330379724503, + 0.05804818496108055, + 0.04586943984031677, + 0.06397281587123871 + ], + "credit_map": [ + 0.02330147475004196, + 0.02321280725300312, + 0.026418130844831467, + 0.05134010314941406, + 0.026881247758865356, + 0.02031979151070118, + 0.04567582160234451, + 0.01950075291097164, + 0.01564251445233822, + 0.01757718063890934, + 0.016124293208122253, + 0.04008990153670311, + 0.02524411492049694, + 0.024401383474469185, + 0.03688286617398262, + 0.027653614059090614, + 0.009374995715916157, + 0.00564089510589838, + 0.00898862536996603, + 0.0159661415964365, + 0.007142063230276108, + 0.007143394090235233, + 0.012901853770017624, + 0.006189273204654455, + 0.008641918189823627, + 0.010760558769106865, + 0.007153250742703676, + 0.01741613633930683, + 0.007943624630570412, + 0.005984737537801266, + 0.01479456014931202, + 0.008364387787878513, + 0.01353461667895317, + 0.008205574005842209, + 0.006866760086268187, + 0.016981936991214752, + 0.008376454934477806, + 0.008004482835531235, + 0.01654219441115856, + 0.005653398111462593, + 0.008582922630012035, + 0.005648021120578051, + 0.006259411107748747, + 0.014414762146770954, + 0.007308217231184244, + 0.0062897587195038795, + 0.013849387876689434, + 0.004635962657630444, + 0.006905803922563791, + 0.006555738393217325, + 0.007290619425475597, + 0.019711824133992195, + 0.011487502604722977, + 0.007843498140573502, + 0.017376400530338287, + 0.009111913852393627, + 0.009117878042161465, + 0.005168180447071791, + 0.0072702462784945965, + 0.008285625837743282, + 0.005112869199365377, + 0.006597219500690699, + 0.007671143393963575, + 0.008825019001960754, + 0.006055040750652552, + 0.005107098259031773, + 0.004099207930266857, + 0.008334606885910034, + 0.0066693187691271305, + 0.0031413829419761896, + 0.007330136373639107, + 0.00513265049085021, + 0.008563061244785786, + 0.005122319795191288, + 0.0064362576231360435, + 0.01072663813829422, + 0.006300817709416151, + 0.007310124114155769, + 0.008690016344189644, + 0.010823584161698818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020577168092131615 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07896928489208221 + ], + "opposite_logprob_tokens": [ + -2.5789692401885986 + ], + "logprob_reward": -0.07896928489208221, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07896928489208221, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2659136950969696, + 0.17477542161941528, + 0.06872087717056274, + 0.07170984148979187, + 0.0919102355837822, + 0.07098617404699326, + 0.06203016638755798, + 0.0798305869102478, + 0.060136258602142334, + 0.05398673936724663 + ], + "credit_map": [ + 0.029611987993121147, + 0.02981872297823429, + 0.021894235163927078, + 0.05174022912979126, + 0.02979985438287258, + 0.02482413314282894, + 0.05486399680376053, + 0.023360557854175568, + 0.01379119511693716, + 0.01521549466997385, + 0.014814871363341808, + 0.03706652298569679, + 0.021111080422997475, + 0.015084149315953255, + 0.03340400382876396, + 0.02428809553384781, + 0.009840739890933037, + 0.00589993130415678, + 0.00573391979560256, + 0.012859080918133259, + 0.006977382116019726, + 0.0072878687642514706, + 0.014426291920244694, + 0.005695660598576069, + 0.006087945308536291, + 0.01005591545253992, + 0.0070304893888533115, + 0.014728499576449394, + 0.0073075927793979645, + 0.0044794450514018536, + 0.012331766076385975, + 0.009688184596598148, + 0.014286798425018787, + 0.012962430715560913, + 0.006805954035371542, + 0.013957850635051727, + 0.013827440328896046, + 0.009786139242351055, + 0.013002882711589336, + 0.00728073762729764, + 0.009411045350134373, + 0.007279470097273588, + 0.005473919212818146, + 0.015224271453917027, + 0.008053272031247616, + 0.005595155525952578, + 0.015017827972769737, + 0.004931211471557617, + 0.00864383950829506, + 0.00579319940879941, + 0.00628952169790864, + 0.012430545873939991, + 0.007388753816485405, + 0.005172723904252052, + 0.011081738397479057, + 0.00522984704002738, + 0.008460004813969135, + 0.00592298898845911, + 0.009279928170144558, + 0.011934375390410423, + 0.011151746846735477, + 0.009942639619112015, + 0.01123945601284504, + 0.01189944613724947, + 0.007770319934934378, + 0.007376112975180149, + 0.004731772001832724, + 0.009125839918851852, + 0.006136302836239338, + 0.005843077786266804, + 0.011562737636268139, + 0.007590092718601227, + 0.006453525274991989, + 0.00442466838285327, + 0.004101330414414406, + 0.009686347097158432, + 0.008115652948617935, + 0.0074847727082669735, + 0.007455006241798401, + 0.006265436764806509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.07896928489208221 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005944074597209692 + ], + "opposite_logprob_tokens": [ + -5.13094425201416 + ], + "logprob_reward": -0.005944074597209692, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005944074597209692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25059840083122253, + 0.1285584419965744, + 0.07707837224006653, + 0.042392659932374954, + 0.08575519919395447, + 0.06264318525791168, + 0.11675316840410233, + 0.07394154369831085, + 0.09220030158758163, + 0.07007869333028793 + ], + "credit_map": [ + 0.028114991262555122, + 0.028558412566781044, + 0.018755316734313965, + 0.03765318542718887, + 0.03853471949696541, + 0.029251081869006157, + 0.03946473449468613, + 0.030265983194112778, + 0.009552679024636745, + 0.012988272123038769, + 0.010587402619421482, + 0.02513345144689083, + 0.016532378271222115, + 0.015324447304010391, + 0.02378367818892002, + 0.01465613953769207, + 0.007027348969131708, + 0.0063710566610097885, + 0.007446296513080597, + 0.015242224559187889, + 0.008816052228212357, + 0.009012851864099503, + 0.01659265346825123, + 0.006569886580109596, + 0.004219702444970608, + 0.006692174356430769, + 0.0038333546835929155, + 0.005559765733778477, + 0.007025483995676041, + 0.004112087190151215, + 0.006249398458749056, + 0.004700693301856518, + 0.00697776535525918, + 0.007224394008517265, + 0.008432860486209393, + 0.010871236212551594, + 0.011111248284578323, + 0.016895519569516182, + 0.013776897452771664, + 0.01046527735888958, + 0.007385582663118839, + 0.009747854433953762, + 0.005265568848699331, + 0.008436423726379871, + 0.00915730744600296, + 0.008946651592850685, + 0.008638257160782814, + 0.005065540317445993, + 0.008998407982289791, + 0.01087444182485342, + 0.007885594852268696, + 0.025017155334353447, + 0.014085359871387482, + 0.012130514718592167, + 0.024470701813697815, + 0.013290993869304657, + 0.00971005018800497, + 0.005448351614177227, + 0.008199336938560009, + 0.01250192429870367, + 0.008670167066156864, + 0.0058038183487951756, + 0.013407968915998936, + 0.010199923068284988, + 0.014388052746653557, + 0.007580157369375229, + 0.00861433707177639, + 0.01715630665421486, + 0.010219362564384937, + 0.005659088026732206, + 0.0198861975222826, + 0.008696798235177994, + 0.008296966552734375, + 0.009488292038440704, + 0.0052970764227211475, + 0.007950996980071068, + 0.009211663156747818, + 0.013402441516518593, + 0.008790349587798119, + 0.007640906609594822 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.005944074597209692 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "This static, eye-level medium shot captures a woman with a chic, short, slightly tousled bob, sitting at an outdoor cafe. She wears a short-sleeved shirt made of lightweight fabric with a delicate, pastel floral pattern against a cream background. She sits comfortably at a small, round cast-iron table on the sunny restaurant terrace, leaning slightly forward and holding a modern smartphone to her right ear. Amidst a lively conversation, a warm, relaxed smile graces her lips, her eyes crinkling slightly as she speaks, her head tilted with interest. Her left hand rests on the tabletop, her fingers leisurely stroking the rim of a glass of espresso iced tea. In the background, the restaurant's blurred facade, with its large windows and dark green awning, is visible, while the soft, out-of-focus shapes of other patrons create a gentle sense of ambient activity. The entire scene is illuminated by the warm midday sun, which filters through a large parasol, casting a soft, diffuse glow that eliminates harsh shadows and creates an overall relaxed, down-to-earth, and joyful atmosphere of a pleasant afternoon.", + "seed": 18323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_15/index_4081_seed_18323_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016005346179008485, + "vlm_logprob": -0.016005346179008485, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.016005346179008485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010852009989321232 + ], + "opposite_logprob_tokens": [ + -5.510851860046387 + ], + "logprob_reward": -0.010852009989321232, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.010852009989321232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23690563440322876, + 0.11697316914796829, + 0.07054276764392853, + 0.0507955439388752, + 0.09606169164180756, + 0.03152654692530632, + 0.2043479084968567, + 0.059986412525177, + 0.04723644256591797, + 0.08562386780977249 + ], + "credit_map": [ + 0.023911157622933388, + 0.02281332015991211, + 0.032975997775793076, + 0.042645253241062164, + 0.02998012863099575, + 0.020956462249159813, + 0.04038422182202339, + 0.02323911339044571, + 0.020468618720769882, + 0.010085081681609154, + 0.01259709894657135, + 0.016650833189487457, + 0.014879679307341576, + 0.008379856124520302, + 0.022568490356206894, + 0.011343525722622871, + 0.011690450832247734, + 0.006203139666467905, + 0.007532033137977123, + 0.009441457688808441, + 0.009348016232252121, + 0.004637317731976509, + 0.01433280948549509, + 0.0073575484566390514, + 0.007852558977901936, + 0.0054438766092062, + 0.0061534312553703785, + 0.007141941227018833, + 0.005513947457075119, + 0.004783533979207277, + 0.007430289406329393, + 0.006475971080362797, + 0.01758207194507122, + 0.010454261675477028, + 0.012147034518420696, + 0.011759448796510696, + 0.01165530364960432, + 0.01082016620784998, + 0.010716546326875687, + 0.010926864109933376, + 0.006083060521632433, + 0.003722146386280656, + 0.0041297548450529575, + 0.003835418028756976, + 0.0031000380404293537, + 0.0028445429634302855, + 0.0036733003798872232, + 0.0041382890194654465, + 0.028873909264802933, + 0.017443453893065453, + 0.022211864590644836, + 0.03110518492758274, + 0.029767315834760666, + 0.01659310795366764, + 0.03505021706223488, + 0.023302877321839333, + 0.007171437609940767, + 0.0048148552887141705, + 0.010384149849414825, + 0.010071272030472755, + 0.007946405559778214, + 0.004171175882220268, + 0.009655655361711979, + 0.005771465599536896, + 0.00613238662481308, + 0.0054712798446416855, + 0.006647171918302774, + 0.0077860914170742035, + 0.00572566781193018, + 0.004745544400066137, + 0.00672259833663702, + 0.0040057054720819, + 0.015595332719385624, + 0.01787245273590088, + 0.010357091203331947, + 0.008517754264175892, + 0.009471604600548744, + 0.006473081186413765, + 0.007975563406944275, + 0.009360995143651962 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.010852009989321232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009796281345188618 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.009796281345188618, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009796281345188618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1717970222234726, + 0.19342823326587677, + 0.13601495325565338, + 0.04732395336031914, + 0.07649915665388107, + 0.1062440499663353, + 0.0933469608426094, + 0.037093568593263626, + 0.07074150443077087, + 0.06751050800085068 + ], + "credit_map": [ + 0.016941674053668976, + 0.02706019952893257, + 0.016475582495331764, + 0.024351630359888077, + 0.021742530167102814, + 0.018114598467946053, + 0.0319727323949337, + 0.015138085931539536, + 0.024437133222818375, + 0.017189841717481613, + 0.020270008593797684, + 0.030934445559978485, + 0.026378674432635307, + 0.01663261279463768, + 0.0398174524307251, + 0.017768090590834618, + 0.017252041026949883, + 0.012789220549166203, + 0.01532626710832119, + 0.021022142842411995, + 0.019686762243509293, + 0.010495485737919807, + 0.02624385431408882, + 0.013199186883866787, + 0.00885575171560049, + 0.004124453291296959, + 0.006469598039984703, + 0.005905558820813894, + 0.006572828162461519, + 0.003701483830809593, + 0.007436580024659634, + 0.004257706925272942, + 0.01358898263424635, + 0.006730563938617706, + 0.010800735093653202, + 0.011003377847373486, + 0.007672786712646484, + 0.0075164963491261005, + 0.010395556688308716, + 0.008790665306150913, + 0.015390424989163876, + 0.008914105594158173, + 0.012000062502920628, + 0.017539430409669876, + 0.014107916504144669, + 0.008943568915128708, + 0.018387777730822563, + 0.010960773564875126, + 0.012174882926046848, + 0.007546587847173214, + 0.010570673272013664, + 0.015285749919712543, + 0.013698785565793514, + 0.007835903204977512, + 0.016502168029546738, + 0.00973221380263567, + 0.004475641064345837, + 0.003631416242569685, + 0.004626024980098009, + 0.007122876588255167, + 0.0043458095751702785, + 0.0029232839588075876, + 0.006116434466093779, + 0.0038520849775522947, + 0.007166330236941576, + 0.005597573705017567, + 0.007933699525892735, + 0.013615565374493599, + 0.010634242556989193, + 0.004987823776900768, + 0.0139140784740448, + 0.00689220055937767, + 0.007729441858828068, + 0.008214902132749557, + 0.007809516973793507, + 0.010442929342389107, + 0.00942709855735302, + 0.006547403987497091, + 0.009843301959335804, + 0.007495927158743143 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009796281345188618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071676388382912 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071676388382912, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071676388382912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18617568910121918, + 0.10368885099887848, + 0.07138130813837051, + 0.05796237662434578, + 0.0612766332924366, + 0.07140839844942093, + 0.04824255779385567, + 0.07496777921915054, + 0.15507297217845917, + 0.16982348263263702 + ], + "credit_map": [ + 0.017067324370145798, + 0.01953849568963051, + 0.027425508946180344, + 0.034457977861166, + 0.018410421907901764, + 0.02053319849073887, + 0.031956747174263, + 0.016785990446805954, + 0.019347431138157845, + 0.018066611140966415, + 0.012928768992424011, + 0.014532020315527916, + 0.0096436170861125, + 0.007371365092694759, + 0.013956411741673946, + 0.007842611521482468, + 0.01037728600203991, + 0.008462312631309032, + 0.00931856781244278, + 0.009129159152507782, + 0.008320990018546581, + 0.007561032194644213, + 0.009313240647315979, + 0.008898712694644928, + 0.010530282743275166, + 0.006018723361194134, + 0.00749374320730567, + 0.008488924242556095, + 0.006290750578045845, + 0.004271451383829117, + 0.009389149025082588, + 0.005479349289089441, + 0.013674411922693253, + 0.0062851752154529095, + 0.01139526255428791, + 0.006499833893030882, + 0.004981979727745056, + 0.00483184726908803, + 0.0067265769466757774, + 0.006881533190608025, + 0.009873094037175179, + 0.007031374145299196, + 0.008076812140643597, + 0.009172117337584496, + 0.010075587779283524, + 0.006966735702008009, + 0.010447975248098373, + 0.009764689952135086, + 0.006566736847162247, + 0.005584587808698416, + 0.0077583203092217445, + 0.007150503806769848, + 0.00664143031463027, + 0.003970086574554443, + 0.005786922760307789, + 0.004783960059285164, + 0.012146980501711369, + 0.009681827388703823, + 0.009790966287255287, + 0.009012262336909771, + 0.008869625627994537, + 0.007227134890854359, + 0.008498440496623516, + 0.009740526787936687, + 0.025838010013103485, + 0.026982130482792854, + 0.01940266415476799, + 0.021329114213585854, + 0.01706315204501152, + 0.013077199459075928, + 0.019390610978007317, + 0.011990086175501347, + 0.022849731147289276, + 0.020902419462800026, + 0.02090112864971161, + 0.019585026428103447, + 0.022091040387749672, + 0.019149737432599068, + 0.020458199083805084, + 0.023886194452643394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.011071676388382912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306020215153694 + ], + "opposite_logprob_tokens": [ + -3.168060302734375 + ], + "logprob_reward": -0.04306020215153694, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306020215153694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13464708626270294, + 0.09404191374778748, + 0.047270432114601135, + 0.09254005551338196, + 0.033740248531103134, + 0.09175065904855728, + 0.33236968517303467, + 0.062020204961299896, + 0.05490445718169212, + 0.05671526491641998 + ], + "credit_map": [ + 0.019837098196148872, + 0.01686086878180504, + 0.023404033854603767, + 0.022163351997733116, + 0.012407710775732994, + 0.009524557739496231, + 0.0214230976998806, + 0.009026370011270046, + 0.012889881618320942, + 0.011350562795996666, + 0.012216507457196712, + 0.01607198640704155, + 0.010257037356495857, + 0.008121022023260593, + 0.0154344467446208, + 0.007700474467128515, + 0.007295635528862476, + 0.0036169409286230803, + 0.0067315176129341125, + 0.006734542548656464, + 0.005789398681372404, + 0.0030897988472133875, + 0.009678575210273266, + 0.004334022291004658, + 0.016028571873903275, + 0.008250459097325802, + 0.014794946648180485, + 0.012961974367499352, + 0.011204307898879051, + 0.007093524094671011, + 0.014028087258338928, + 0.00817818008363247, + 0.0058411951176822186, + 0.003839839017018676, + 0.0060250950045883656, + 0.004422244615852833, + 0.0034899432212114334, + 0.002989324275404215, + 0.004215099848806858, + 0.002917506964877248, + 0.012704619206488132, + 0.008192182518541813, + 0.010222655721008778, + 0.01401779055595398, + 0.012411219999194145, + 0.007492764852941036, + 0.01629653573036194, + 0.010412892326712608, + 0.04243769496679306, + 0.027110805734992027, + 0.03591595217585564, + 0.05340692028403282, + 0.05049682781100273, + 0.026396185159683228, + 0.05993557721376419, + 0.036669738590717316, + 0.010596361942589283, + 0.009920919314026833, + 0.009799701161682606, + 0.008136582560837269, + 0.0068745482712984085, + 0.005254685878753662, + 0.007587167900055647, + 0.0038502372335642576, + 0.006054603960365057, + 0.005025678314268589, + 0.007446198258548975, + 0.008050474338233471, + 0.007042943499982357, + 0.004457203671336174, + 0.00921971257776022, + 0.007607641164213419, + 0.009216980077326298, + 0.010775836184620857, + 0.00785816926509142, + 0.006387823726981878, + 0.005926099605858326, + 0.006300817243754864, + 0.004916962236166, + 0.005332577507942915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.04306020215153694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as a static, eye-level medium shot of a woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the woman have a short, slightly tousled bob hairstyle and wear a short-sleeved shirt with a pastel floral pattern on a cream background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman sitting at a small, round cast-iron table on a sunny restaurant terrace, holding a smartphone to her right ear?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman smiling with crinkled eyes and leaning forward while her left hand strokes the rim of a glass of iced tea?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background visible as a blurred restaurant facade with large windows and a dark green awning, illuminated by warm midday sun filtering through a parasol?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005246561020612717 + ], + "opposite_logprob_tokens": [ + -5.255246639251709 + ], + "logprob_reward": -0.005246561020612717, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005246561020612717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20478662848472595, + 0.15829510986804962, + 0.04074130207300186, + 0.1353330910205841, + 0.04115832969546318, + 0.045699771493673325, + 0.09892817586660385, + 0.037461649626493454, + 0.0446467287838459, + 0.19294922053813934 + ], + "credit_map": [ + 0.016861502081155777, + 0.026179345324635506, + 0.01887962967157364, + 0.029252316802740097, + 0.030522478744387627, + 0.02417048066854477, + 0.04038763418793678, + 0.018533241003751755, + 0.021035216748714447, + 0.013099007308483124, + 0.01711800880730152, + 0.02481601946055889, + 0.022067703306674957, + 0.012607758864760399, + 0.031737543642520905, + 0.01581384427845478, + 0.006204464007169008, + 0.004045641981065273, + 0.004644459113478661, + 0.005936159752309322, + 0.005512402858585119, + 0.0030216409359127283, + 0.007085726130753756, + 0.004290806595236063, + 0.01898796670138836, + 0.0100829703733325, + 0.015840405598282814, + 0.02046985737979412, + 0.01845329813659191, + 0.0123477503657341, + 0.024748655036091805, + 0.014402193948626518, + 0.006250227801501751, + 0.0072274766862392426, + 0.005603357218205929, + 0.005271772854030132, + 0.004118950106203556, + 0.0048690298572182655, + 0.004090764559805393, + 0.003726748749613762, + 0.0074120499193668365, + 0.004815141204744577, + 0.004976163152605295, + 0.0066551947966217995, + 0.005688781384378672, + 0.005015254486352205, + 0.0059354049153625965, + 0.005201784428209066, + 0.012065403163433075, + 0.007912357337772846, + 0.010633028112351894, + 0.016011808067560196, + 0.014802739955484867, + 0.008143960498273373, + 0.018246209248900414, + 0.011112663894891739, + 0.00540133286267519, + 0.004929378163069487, + 0.005271051079034805, + 0.005176809150725603, + 0.004139058757573366, + 0.0035012138541787863, + 0.004743189550936222, + 0.004299614578485489, + 0.008572150953114033, + 0.009885952807962894, + 0.005996998865157366, + 0.004515074659138918, + 0.004327328875660896, + 0.004247064236551523, + 0.003979064058512449, + 0.00312309549190104, + 0.028155414387583733, + 0.03141479939222336, + 0.023211222141981125, + 0.024836726486682892, + 0.024557186290621758, + 0.01837068423628807, + 0.025704730302095413, + 0.016698462888598442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.005246561020612717 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_tensor([2658, 2658])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_tensor([2658, 2658])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c157240fc974fe0847d263014c7d2a9925d6000b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_tensor([2658, 2658])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "cinematic shot begins in mid-ground and slowly pans rightward, synchronizing with a small, carved boat. At dusk, Yang Guifei, dressed in a Tang Dynasty-style, chest-length ruqu (a traditional Chinese dress) in a criss-cross pattern of crimson and bright yellow, reclines on the brocade cushions of the boat's cabin. Her towering bun, adorned with delicate gold hairpins and pearl ornaments, trembles with the boat's gentle sway. The soft golden light of the setting sun filters through the mist, illuminating her elegant silk attire, the phoenix and peony designs embroidered in gold thread on the sleeves. Her face is like a silver disk, her eyes slightly tipsy, a languid smile lingering, her posture relaxed and relaxed. The boat, constructed of rich, dark wood, its sides and supports intricately carved with scrolling patterns, silently cuts across the mirror-like surface of the lake, leaving behind slowly spreading ripples that reflect the vibrant sunset glow on the horizon and the distant, misty silhouettes of green mountains. One of her delicate hands rested casually on the side of the boat, and her slender fingertips occasionally casually stroked the cool lake water. The whole picture was filled with an extremely tranquil, luxurious and poetic atmosphere in warm golden tones.", + "seed": 4194, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4194_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13445593025535346, + "vlm_logprob": -0.13445593025535346, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.13445593025535346, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07437409460544586 + ], + "opposite_logprob_tokens": [ + -2.824374198913574 + ], + "logprob_reward": -0.07437409460544586, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07437409460544586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09450800716876984, + 0.02983243390917778, + 0.17011688649654388, + 0.013834738172590733, + 0.047954343259334564, + 0.2388976663351059, + 0.030740762129426003, + 0.3158811628818512, + 0.02379799447953701, + 0.03443599492311478 + ], + "credit_map": [ + 0.0033553829416632652, + 0.0047214943915605545, + 0.004841918591409922, + 0.004105998668819666, + 0.031690724194049835, + 0.03431498259305954, + 0.006183403544127941, + 0.005294101312756538, + 0.0015472983941435814, + 0.0019445096841081977, + 0.0025606886483728886, + 0.003051658859476447, + 0.00612608389928937, + 0.006056167650967836, + 0.004896759055554867, + 0.0036492699291557074, + 0.004618669860064983, + 0.004516027867794037, + 0.012155093252658844, + 0.006486358121037483, + 0.05914962291717529, + 0.06970252841711044, + 0.006643548607826233, + 0.006845047697424889, + 0.0011148098856210709, + 0.0015739075606688857, + 0.0014808938140049577, + 0.0013049470726400614, + 0.0026991867925971746, + 0.0024533008690923452, + 0.0019792725797742605, + 0.0012284195981919765, + 0.005750138778239489, + 0.0042086876928806305, + 0.006190024316310883, + 0.00463319756090641, + 0.006415154319256544, + 0.005352043081074953, + 0.007294571492820978, + 0.008110523223876953, + 0.0077230255119502544, + 0.007513490039855242, + 0.026493947952985764, + 0.006790222134441137, + 0.07563386112451553, + 0.08957042545080185, + 0.013816829770803452, + 0.01135585829615593, + 0.0023119912948459387, + 0.0025467639788985252, + 0.006313272286206484, + 0.002355898730456829, + 0.005367200821638107, + 0.00429928582161665, + 0.004256822168827057, + 0.00328952819108963, + 0.008845493197441101, + 0.01807100884616375, + 0.021327629685401917, + 0.004822239745408297, + 0.12741687893867493, + 0.11213374882936478, + 0.013096041046082973, + 0.01016809232532978, + 0.0020983165595680475, + 0.0024244373198598623, + 0.0019420960452407598, + 0.002333440352231264, + 0.004038549028337002, + 0.005306824576109648, + 0.0033593899570405483, + 0.0022949385456740856, + 0.0016991870943456888, + 0.004022250883281231, + 0.0028459466993808746, + 0.0018949746154248714, + 0.010127258487045765, + 0.004319704603403807, + 0.0023434360045939684, + 0.007183237932622433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 2 + ], + "selected_reward": -0.07437409460544586 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0111062191426754 + ], + "opposite_logprob_tokens": [ + -4.511106014251709 + ], + "logprob_reward": -0.0111062191426754, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.0111062191426754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06579221785068512, + 0.01250686589628458, + 0.6536599397659302, + 0.013522376306355, + 0.01150447502732277, + 0.058676544576883316, + 0.08273886889219284, + 0.07837192714214325, + 0.011605380102992058, + 0.011621402576565742 + ], + "credit_map": [ + 0.003078262321650982, + 0.0038361113984137774, + 0.003738740924745798, + 0.0042757694609463215, + 0.02109876088798046, + 0.022177351638674736, + 0.003955822438001633, + 0.0036313950549811125, + 0.0013198042288422585, + 0.0010593963088467717, + 0.001338787260465324, + 0.001230355934239924, + 0.001969788921996951, + 0.0025526906829327345, + 0.0016595192719250917, + 0.0013765220064669847, + 0.013958628289401531, + 0.011392812244594097, + 0.04240225628018379, + 0.011932005174458027, + 0.2536744475364685, + 0.28238704800605774, + 0.017577001824975014, + 0.02033567987382412, + 0.0008638857980258763, + 0.0008179475553333759, + 0.0017197192646563053, + 0.0013844860950484872, + 0.002607114613056183, + 0.002720138756558299, + 0.0018996578874066472, + 0.001509424764662981, + 0.001142179942689836, + 0.0009438712731935084, + 0.0015470721991732717, + 0.0009665713296271861, + 0.0016538390191271901, + 0.0017047624569386244, + 0.0019308744231238961, + 0.001615302637219429, + 0.0027150046080350876, + 0.004028031602501869, + 0.005583749618381262, + 0.004797938745468855, + 0.016972489655017853, + 0.016484925523400307, + 0.005109630525112152, + 0.002984773600474, + 0.005162402056157589, + 0.004504181910306215, + 0.023160163313150406, + 0.004342623520642519, + 0.02685023657977581, + 0.009436942636966705, + 0.005863650236278772, + 0.003418669803068042, + 0.004702266305685043, + 0.007706090807914734, + 0.005428938195109367, + 0.008080858737230301, + 0.0228903628885746, + 0.01989835686981678, + 0.006376296281814575, + 0.003288750536739826, + 0.0011341256322339177, + 0.0015998025191947818, + 0.0010195214999839664, + 0.001604354358278215, + 0.0017117655370384455, + 0.0019571713637560606, + 0.001465021283365786, + 0.0011136168614029884, + 0.0006301985704340041, + 0.0014921689871698618, + 0.0009619408519938588, + 0.0007025193190202117, + 0.00341025204397738, + 0.001044542994350195, + 0.0009772421326488256, + 0.002402537502348423 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.0111062191426754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012555836699903011 + ], + "opposite_logprob_tokens": [ + -4.3875555992126465 + ], + "logprob_reward": -0.012555836699903011, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012555836699903011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08065572381019592, + 0.08982695639133453, + 0.24488237500190735, + 0.022150157019495964, + 0.019419776275753975, + 0.13046997785568237, + 0.2476024329662323, + 0.06541629880666733, + 0.08448363840579987, + 0.015092677436769009 + ], + "credit_map": [ + 0.004236611071974039, + 0.005417628679424524, + 0.0054937852546572685, + 0.00531733687967062, + 0.02339424379169941, + 0.024786662310361862, + 0.005255402531474829, + 0.006754057947546244, + 0.0034199857618659735, + 0.006411213427782059, + 0.005675800610333681, + 0.006981103681027889, + 0.025753185153007507, + 0.026894710958003998, + 0.008295944891870022, + 0.006395014934241772, + 0.0076555595733225346, + 0.013839244842529297, + 0.01676866039633751, + 0.01459892001003027, + 0.07801688462495804, + 0.08329049497842789, + 0.014737063087522984, + 0.01597554236650467, + 0.0020674443803727627, + 0.0018846296006813645, + 0.003020711475983262, + 0.002494652522727847, + 0.003022675635293126, + 0.0026822967920452356, + 0.0041296775452792645, + 0.0028480682522058487, + 0.002095166826620698, + 0.0024421659763902426, + 0.00334469904191792, + 0.0016249456675723195, + 0.002122321864590049, + 0.002483248943462968, + 0.0022907431703060865, + 0.0030164839699864388, + 0.005246239714324474, + 0.006627913564443588, + 0.014589752070605755, + 0.007383184041827917, + 0.03532858192920685, + 0.04348696395754814, + 0.011054398491978645, + 0.006752942223101854, + 0.015937017276883125, + 0.011585695669054985, + 0.07445239275693893, + 0.01225037407130003, + 0.07713969051837921, + 0.026146838441491127, + 0.019160594791173935, + 0.010929838754236698, + 0.0026996261440217495, + 0.003866145620122552, + 0.005130084231495857, + 0.0035914473701268435, + 0.02043771557509899, + 0.021320823580026627, + 0.004274932201951742, + 0.004095523152500391, + 0.008439254015684128, + 0.009619624353945255, + 0.006853647064417601, + 0.00832054391503334, + 0.012373915873467922, + 0.018723007291555405, + 0.011627272702753544, + 0.008526367135345936, + 0.0013126059202477336, + 0.0018408215837553144, + 0.0013164776610210538, + 0.0013687650207430124, + 0.0029298365116119385, + 0.002177590038627386, + 0.0021910914219915867, + 0.0019554884638637304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 5 + ], + "selected_reward": -0.012555836699903011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22563202679157257 + ], + "opposite_logprob_tokens": [ + -1.6006320714950562 + ], + "logprob_reward": -0.22563202679157257, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22563202679157257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15762484073638916, + 0.02078886143863201, + 0.32463693618774414, + 0.019861875101923943, + 0.025785785168409348, + 0.3036442697048187, + 0.06408470124006271, + 0.03283532336354256, + 0.030072204768657684, + 0.020665278658270836 + ], + "credit_map": [ + 0.007624430116266012, + 0.006013714242726564, + 0.007992909289896488, + 0.008025143295526505, + 0.05223994329571724, + 0.05511895567178726, + 0.01100306399166584, + 0.009606679901480675, + 0.001348743331618607, + 0.002010657452046871, + 0.001865012221969664, + 0.001883667311631143, + 0.0038074934855103493, + 0.004306364338845015, + 0.0032645268365740776, + 0.0023023970425128937, + 0.007003158330917358, + 0.007831818424165249, + 0.02098231017589569, + 0.009792406111955643, + 0.12055511027574539, + 0.1361064910888672, + 0.010307894088327885, + 0.01205772627145052, + 0.0012136324075981975, + 0.001996314385905862, + 0.002549010794609785, + 0.0016655700746923685, + 0.00329801207408309, + 0.00373951462097466, + 0.0028890003450214863, + 0.0025108216796070337, + 0.0023749968968331814, + 0.0021554166451096535, + 0.0034387600608170033, + 0.002135111717507243, + 0.0039560915902256966, + 0.004010949283838272, + 0.004529462195932865, + 0.003184997709468007, + 0.008968031033873558, + 0.011494984850287437, + 0.02917267009615898, + 0.008448026143014431, + 0.10517926514148712, + 0.11382728070020676, + 0.014793982729315758, + 0.011760017834603786, + 0.0054410588927567005, + 0.004947334062308073, + 0.013136123307049274, + 0.006937613245099783, + 0.008847370743751526, + 0.006866079289466143, + 0.012400885112583637, + 0.005508235190063715, + 0.001136792590841651, + 0.00190096499864012, + 0.0021109306253492832, + 0.0010225262958556414, + 0.012545444071292877, + 0.01086826529353857, + 0.0016766224289312959, + 0.001573776127770543, + 0.0025884888600558043, + 0.003981631249189377, + 0.00285671791061759, + 0.006249244790524244, + 0.004661716055124998, + 0.002898318925872445, + 0.004377597477287054, + 0.0024584904313087463, + 0.002254937309771776, + 0.0028683145064860582, + 0.0022787090856581926, + 0.0029444689862430096, + 0.003780952887609601, + 0.0015377679374068975, + 0.002168081933632493, + 0.0028320462442934513 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.22563202679157257 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486114740371704 + ], + "opposite_logprob_tokens": [ + -1.2236114740371704 + ], + "logprob_reward": -0.3486114740371704, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486114740371704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21098192036151886, + 0.04117780923843384, + 0.3479013741016388, + 0.012482944875955582, + 0.01930372230708599, + 0.07961488515138626, + 0.0267702117562294, + 0.10090848803520203, + 0.04408534616231918, + 0.11677324026823044 + ], + "credit_map": [ + 0.005895739421248436, + 0.002937792567536235, + 0.008908120915293694, + 0.004495103843510151, + 0.08674630522727966, + 0.09174269437789917, + 0.004680457059293985, + 0.0055757262744009495, + 0.0029434547759592533, + 0.0027349151205271482, + 0.003273364622145891, + 0.003981278743594885, + 0.008635633625090122, + 0.009227078408002853, + 0.006071609910577536, + 0.004310480784624815, + 0.007397559005767107, + 0.005370752420276403, + 0.023018429055809975, + 0.005185221321880817, + 0.13633018732070923, + 0.15310923755168915, + 0.007757487706840038, + 0.009732541628181934, + 0.001206323504447937, + 0.0012028360506519675, + 0.0020857995841652155, + 0.0019514550222083926, + 0.001511343172751367, + 0.001503112493082881, + 0.0018862823490053415, + 0.0011357935145497322, + 0.0019185392884537578, + 0.001238649827428162, + 0.0025328933261334896, + 0.0013386189239099622, + 0.004017556086182594, + 0.0032736589200794697, + 0.002677791751921177, + 0.002306015230715275, + 0.0027616024017333984, + 0.003099725116044283, + 0.008533534593880177, + 0.0036142091266810894, + 0.02498209848999977, + 0.029434693977236748, + 0.0035317447036504745, + 0.003657276276499033, + 0.001746528665535152, + 0.001431245356798172, + 0.007329557556658983, + 0.0017016270430758595, + 0.006831763777881861, + 0.0028177835047245026, + 0.0031940052285790443, + 0.001717702834866941, + 0.0028926851227879524, + 0.005676642060279846, + 0.007362703327089548, + 0.0020120027475059032, + 0.039224959909915924, + 0.03505709394812584, + 0.004541612695902586, + 0.004140789620578289, + 0.003748286282643676, + 0.005724871065467596, + 0.005024675279855728, + 0.007914140820503235, + 0.0061882114969193935, + 0.004845648538321257, + 0.007205945905297995, + 0.0034335709642618895, + 0.005186513997614384, + 0.014849219471216202, + 0.010415475815534592, + 0.005094143562018871, + 0.03736347332596779, + 0.010197339579463005, + 0.006262775044888258, + 0.027404308319091797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.3486114740371704 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic shot begins in mid-ground and slowly pans rightward, synchronizing with a small, carved boat. At dusk, Yang Guifei, dressed in a Tang Dynasty-style, chest-length ruqu (a traditional Chinese dress) in a criss-cross pattern of crimson and bright yellow, reclines on the brocade cushions of the boat's cabin. Her towering bun, adorned with delicate gold hairpins and pearl ornaments, trembles with the boat's gentle sway. The soft golden light of the setting sun filters through the mist, illuminating her elegant silk attire, the phoenix and peony designs embroidered in gold thread on the sleeves. Her face is like a silver disk, her eyes slightly tipsy, a languid smile lingering, her posture relaxed and relaxed. The boat, constructed of rich, dark wood, its sides and supports intricately carved with scrolling patterns, silently cuts across the mirror-like surface of the lake, leaving behind slowly spreading ripples that reflect the vibrant sunset glow on the horizon and the distant, misty silhouettes of green mountains. One of her delicate hands rested casually on the side of the boat, and her slender fingertips occasionally casually stroked the cool lake water. The whole picture was filled with an extremely tranquil, luxurious and poetic atmosphere in warm golden tones.", + "seed": 4195, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4195_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.18290479592978953, + "vlm_logprob": -0.18290479592978953, + "vlm_yes_no_margin": 2.525, + "vlm_token_credit": -0.18290479592978953, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09985937923192978 + ], + "opposite_logprob_tokens": [ + -2.4748594760894775 + ], + "logprob_reward": -0.09985937923192978, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09985937923192978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2222825437784195, + 0.06019739806652069, + 0.07437901198863983, + 0.028255373239517212, + 0.09655672311782837, + 0.10438302159309387, + 0.07312016189098358, + 0.08181504160165787, + 0.14033231139183044, + 0.11867836862802505 + ], + "credit_map": [ + 0.02577142044901848, + 0.030632562935352325, + 0.02362598106265068, + 0.026441402733325958, + 0.020717589184641838, + 0.028829269111156464, + 0.030928177759051323, + 0.035336144268512726, + 0.008371703326702118, + 0.008129853755235672, + 0.006854167673736811, + 0.006504645571112633, + 0.005798921454697847, + 0.007340892683714628, + 0.008730935864150524, + 0.00846627727150917, + 0.00918449554592371, + 0.008604675531387329, + 0.009096029214560986, + 0.006487330887466669, + 0.008375944569706917, + 0.010262061841785908, + 0.010928465984761715, + 0.011440012603998184, + 0.0037905978970229626, + 0.0037078873720020056, + 0.002852475270628929, + 0.003693273290991783, + 0.003196363104507327, + 0.003876641858369112, + 0.004511750768870115, + 0.0026263834442943335, + 0.01767304539680481, + 0.005102229304611683, + 0.00675019109621644, + 0.005832039751112461, + 0.025139492005109787, + 0.008955775760114193, + 0.006592835299670696, + 0.020511113107204437, + 0.02126455307006836, + 0.005157621577382088, + 0.009806983172893524, + 0.005772864446043968, + 0.02917363867163658, + 0.012107123620808125, + 0.009569059126079082, + 0.011531179770827293, + 0.010031698271632195, + 0.005089067853987217, + 0.007607811130583286, + 0.005102332215756178, + 0.011763986200094223, + 0.01377258449792862, + 0.0087971743196249, + 0.010955505073070526, + 0.019487254321575165, + 0.0068537043407559395, + 0.00710010901093483, + 0.004483663011342287, + 0.007700254675000906, + 0.018239352852106094, + 0.007970882579684258, + 0.009979823604226112, + 0.04217389598488808, + 0.007600924465805292, + 0.008766332641243935, + 0.006768259219825268, + 0.013690770603716373, + 0.04228837043046951, + 0.008290023542940617, + 0.010753722861409187, + 0.011290854774415493, + 0.0066219777800142765, + 0.014496976509690285, + 0.010315747931599617, + 0.0221537034958601, + 0.02086377516388893, + 0.01086015347391367, + 0.022075191140174866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.09985937923192978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822066679596901 + ], + "opposite_logprob_tokens": [ + -4.018220901489258 + ], + "logprob_reward": -0.01822066679596901, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822066679596901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2850538492202759, + 0.04773508757352829, + 0.03425874933600426, + 0.06856067478656769, + 0.0520855113863945, + 0.025438208132982254, + 0.09539344906806946, + 0.11549755930900574, + 0.17618905007839203, + 0.0997878909111023 + ], + "credit_map": [ + 0.036215219646692276, + 0.04638371244072914, + 0.024571768939495087, + 0.04278577119112015, + 0.02901916764676571, + 0.030985210090875626, + 0.04319123178720474, + 0.03190174326300621, + 0.005839002318680286, + 0.007707138080149889, + 0.004561488050967455, + 0.006396700162440538, + 0.005059417337179184, + 0.005484218709170818, + 0.005542846396565437, + 0.00714427325874567, + 0.004226241260766983, + 0.004240898881107569, + 0.004254575353115797, + 0.0036055338568985462, + 0.003001399803906679, + 0.004715899005532265, + 0.004180322866886854, + 0.006033872254192829, + 0.007244835142046213, + 0.0102041345089674, + 0.0063496665097773075, + 0.013223346322774887, + 0.007666360586881638, + 0.005323964636772871, + 0.01092366874217987, + 0.007624693214893341, + 0.0063598426058888435, + 0.008687430992722511, + 0.004619322717189789, + 0.00971611775457859, + 0.008199806325137615, + 0.005049373488873243, + 0.004339896142482758, + 0.005113719496876001, + 0.002306480659171939, + 0.001993495738133788, + 0.003799529280513525, + 0.0019899301696568727, + 0.00290096178650856, + 0.0037927397061139345, + 0.003642788389697671, + 0.005012279376387596, + 0.029238170012831688, + 0.005134738516062498, + 0.007066431455314159, + 0.004183609038591385, + 0.011437484063208103, + 0.025559639558196068, + 0.006899523548781872, + 0.005873843561857939, + 0.03996969386935234, + 0.007168065290898085, + 0.006898818537592888, + 0.006057484541088343, + 0.009962789714336395, + 0.03541289642453194, + 0.004690171685069799, + 0.0053376443684101105, + 0.05864619463682175, + 0.008925563655793667, + 0.01047063060104847, + 0.007618324365466833, + 0.017815159633755684, + 0.05633421987295151, + 0.006041719112545252, + 0.010337237268686295, + 0.024461403489112854, + 0.005513192154467106, + 0.007622282952070236, + 0.006081503350287676, + 0.013340475037693977, + 0.02947162464261055, + 0.005126539152115583, + 0.008170864544808865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.01822066679596901 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008650803938508034 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.008650803938508034, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008650803938508034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292460769414902, + 0.02787654474377632, + 0.02588372305035591, + 0.030085792765021324, + 0.0639561116695404, + 0.0995640680193901, + 0.25027909874916077, + 0.02508993074297905, + 0.30809661746025085, + 0.0662434846162796 + ], + "credit_map": [ + 0.011759446933865547, + 0.014631280675530434, + 0.010099105536937714, + 0.014316773042082787, + 0.012083785608410835, + 0.01240141037851572, + 0.011852838099002838, + 0.015779968351125717, + 0.0033909182529896498, + 0.004675013013184071, + 0.0021930241491645575, + 0.0034690184984356165, + 0.002123230369761586, + 0.0036225593648850918, + 0.005214366130530834, + 0.003188414266332984, + 0.0034395698457956314, + 0.003172712866216898, + 0.0021735907066613436, + 0.0025574080646038055, + 0.0028930343687534332, + 0.0033010905608534813, + 0.003707312047481537, + 0.004639002960175276, + 0.00396826583892107, + 0.003909884486347437, + 0.0024436519015580416, + 0.004209986422210932, + 0.004326304886490107, + 0.0033360051456838846, + 0.0049508800730109215, + 0.0029408177360892296, + 0.010306084528565407, + 0.004337972030043602, + 0.003936687018722296, + 0.007452961057424545, + 0.01574345864355564, + 0.005401567555963993, + 0.005640699528157711, + 0.011136678047478199, + 0.02179241180419922, + 0.004699293058365583, + 0.006319417152553797, + 0.0025846364442259073, + 0.03146187216043472, + 0.00911230780184269, + 0.00719454325735569, + 0.016399584710597992, + 0.08670860528945923, + 0.012519268319010735, + 0.01728169620037079, + 0.007603146601468325, + 0.029681196436285973, + 0.07815717160701752, + 0.008707736618816853, + 0.00962029118090868, + 0.0030210756231099367, + 0.003251552814617753, + 0.003408326767385006, + 0.005037418566644192, + 0.0021499809809029102, + 0.002554092789068818, + 0.0027330939192324877, + 0.0029343890491873026, + 0.08766235411167145, + 0.014858683571219444, + 0.022600896656513214, + 0.015209474600851536, + 0.03216175734996796, + 0.08971159905195236, + 0.015231795608997345, + 0.030660049989819527, + 0.009713406674563885, + 0.003991177771240473, + 0.006200982257723808, + 0.005991026759147644, + 0.01106279157102108, + 0.014898757450282574, + 0.005119567736983299, + 0.009265774860978127 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.008650803938508034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135206997394562 + ], + "opposite_logprob_tokens": [ + -1.3135206699371338 + ], + "logprob_reward": -0.3135206997394562, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135206997394562, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1236281543970108, + 0.04492601379752159, + 0.048963431268930435, + 0.09035852551460266, + 0.056458499282598495, + 0.07715395838022232, + 0.12533575296401978, + 0.05158580094575882, + 0.2821275591850281, + 0.09946230798959732 + ], + "credit_map": [ + 0.016195038333535194, + 0.019630473107099533, + 0.010586204938590527, + 0.017520541325211525, + 0.014616932719945908, + 0.015324744395911694, + 0.016883838921785355, + 0.012870371341705322, + 0.006083266343921423, + 0.0056127323769032955, + 0.004943964071571827, + 0.006802708841860294, + 0.005383441224694252, + 0.005933755077421665, + 0.006603464018553495, + 0.0035626874305307865, + 0.005899242125451565, + 0.005837529432028532, + 0.004769500810652971, + 0.0050858790054917336, + 0.007486394140869379, + 0.006014315877109766, + 0.0070770722813904285, + 0.006793500389903784, + 0.011803894303739071, + 0.009720875881612301, + 0.008182878606021404, + 0.011436988599598408, + 0.011598717421293259, + 0.011076000519096851, + 0.013015826232731342, + 0.013523340225219727, + 0.00854577962309122, + 0.006165673490613699, + 0.00439680740237236, + 0.005226625129580498, + 0.013058040291070938, + 0.0064635369926691055, + 0.004390203393995762, + 0.008211829699575901, + 0.014762688428163528, + 0.004837087821215391, + 0.005329903680831194, + 0.004225458484143019, + 0.0236153956502676, + 0.006260847207158804, + 0.005642368923872709, + 0.012480207718908787, + 0.03294573724269867, + 0.008962040767073631, + 0.009933041408658028, + 0.009655074216425419, + 0.01835191249847412, + 0.022267429158091545, + 0.013675214722752571, + 0.009545308537781239, + 0.009931854903697968, + 0.005827062763273716, + 0.00510493665933609, + 0.004917820915579796, + 0.005065826699137688, + 0.010502701625227928, + 0.005660735536366701, + 0.004574861377477646, + 0.03213457018136978, + 0.023553689941763878, + 0.04137018322944641, + 0.02425185777246952, + 0.03409121185541153, + 0.037546366453170776, + 0.03090505488216877, + 0.05827462300658226, + 0.01721535250544548, + 0.009113199077546597, + 0.00957527942955494, + 0.007865038700401783, + 0.01290860865265131, + 0.018423480913043022, + 0.008533007465302944, + 0.015828343108296394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.3135206997394562 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742724299430847 + ], + "opposite_logprob_tokens": [ + -0.9742724299430847 + ], + "logprob_reward": -0.4742724299430847, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742724299430847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09510037302970886, + 0.030150938779115677, + 0.022509774193167686, + 0.02159825526177883, + 0.09015606343746185, + 0.06815356761217117, + 0.08144012093544006, + 0.042499933391809464, + 0.47247371077537537, + 0.07591725140810013 + ], + "credit_map": [ + 0.012907779775559902, + 0.012926892377436161, + 0.008525559678673744, + 0.012855925597250462, + 0.01029521320015192, + 0.013899511657655239, + 0.012875406071543694, + 0.010814090259373188, + 0.0042926715686917305, + 0.004750996362417936, + 0.003379746340215206, + 0.004476725123822689, + 0.003567988518625498, + 0.0040781269781291485, + 0.00367916002869606, + 0.0019255246734246612, + 0.003940058872103691, + 0.003835192648693919, + 0.002121328841894865, + 0.0019404144259169698, + 0.0028714777436107397, + 0.0033037990797311068, + 0.0022117893677204847, + 0.0022857135627418756, + 0.0029159800615161657, + 0.003096839413046837, + 0.0023747102823108435, + 0.003019433468580246, + 0.0030728422570973635, + 0.002221294678747654, + 0.0021792789921164513, + 0.002717877272516489, + 0.014790976420044899, + 0.0044878264889121056, + 0.006555426865816116, + 0.005957549437880516, + 0.01960562914609909, + 0.012252791784703732, + 0.007551874965429306, + 0.018953990191221237, + 0.01777326874434948, + 0.004172063898295164, + 0.005722000729292631, + 0.0031367798801511526, + 0.021144606173038483, + 0.005107236560434103, + 0.0065669347532093525, + 0.004530681297183037, + 0.0261620432138443, + 0.00495994184166193, + 0.005898633506149054, + 0.002836014609783888, + 0.014695689082145691, + 0.015147379599511623, + 0.005880073644220829, + 0.005860346835106611, + 0.010265171527862549, + 0.002778546418994665, + 0.002971586072817445, + 0.002767757745459676, + 0.005422639660537243, + 0.008170058019459248, + 0.005750035401433706, + 0.004374135751277208, + 0.14112591743469238, + 0.0223423819988966, + 0.030970679596066475, + 0.023109272122383118, + 0.05187788978219032, + 0.14183323085308075, + 0.02229928970336914, + 0.03891504555940628, + 0.011044381186366081, + 0.007848795503377914, + 0.006208834704011679, + 0.009453678503632545, + 0.01045587845146656, + 0.010713119991123676, + 0.009301545098423958, + 0.010891018435359001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.4742724299430847 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic shot begins in mid-ground and slowly pans rightward, synchronizing with a small, carved boat. At dusk, Yang Guifei, dressed in a Tang Dynasty-style, chest-length ruqu (a traditional Chinese dress) in a criss-cross pattern of crimson and bright yellow, reclines on the brocade cushions of the boat's cabin. Her towering bun, adorned with delicate gold hairpins and pearl ornaments, trembles with the boat's gentle sway. The soft golden light of the setting sun filters through the mist, illuminating her elegant silk attire, the phoenix and peony designs embroidered in gold thread on the sleeves. Her face is like a silver disk, her eyes slightly tipsy, a languid smile lingering, her posture relaxed and relaxed. The boat, constructed of rich, dark wood, its sides and supports intricately carved with scrolling patterns, silently cuts across the mirror-like surface of the lake, leaving behind slowly spreading ripples that reflect the vibrant sunset glow on the horizon and the distant, misty silhouettes of green mountains. One of her delicate hands rested casually on the side of the boat, and her slender fingertips occasionally casually stroked the cool lake water. The whole picture was filled with an extremely tranquil, luxurious and poetic atmosphere in warm golden tones.", + "seed": 4196, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4196_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.16227782871574165, + "vlm_logprob": -0.16227782871574165, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.16227782871574165, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05327378958463669 + ], + "opposite_logprob_tokens": [ + -3.0532736778259277 + ], + "logprob_reward": -0.05327378958463669, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05327378958463669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06926532834768295, + 0.03067937307059765, + 0.0228183064609766, + 0.04506635293364525, + 0.07561992853879929, + 0.1252361387014389, + 0.3359002470970154, + 0.08298904448747635, + 0.19513903558254242, + 0.01728629507124424 + ], + "credit_map": [ + 0.015501197427511215, + 0.015783702954649925, + 0.006037624552845955, + 0.005123271606862545, + 0.005209040828049183, + 0.01146750245243311, + 0.004423293750733137, + 0.005719688721001148, + 0.005832647439092398, + 0.007172940764576197, + 0.00265920115634799, + 0.0018924535252153873, + 0.002548309275880456, + 0.00623998511582613, + 0.0019288139883428812, + 0.0024050194770097733, + 0.0032508661970496178, + 0.004703976679593325, + 0.0029055706690996885, + 0.0016283582663163543, + 0.0019349332433193922, + 0.0042919558472931385, + 0.0018045315518975258, + 0.00229811342433095, + 0.011469326913356781, + 0.004834593273699284, + 0.002775801345705986, + 0.0018054574029520154, + 0.002431495115160942, + 0.014712275005877018, + 0.0024568086955696344, + 0.004580596461892128, + 0.01765771582722664, + 0.005479286890476942, + 0.006306980736553669, + 0.0038517089560627937, + 0.003913235384970903, + 0.026964975520968437, + 0.0031319649424403906, + 0.008314055390655994, + 0.03843209892511368, + 0.0162790659815073, + 0.003570278873667121, + 0.0032248173374682665, + 0.005187121219933033, + 0.04395831748843193, + 0.0036353489849716425, + 0.010949081741273403, + 0.08493934571743011, + 0.13276362419128418, + 0.005760956089943647, + 0.005291170906275511, + 0.010309266857802868, + 0.07752290368080139, + 0.005810466129332781, + 0.013502496294677258, + 0.015378287993371487, + 0.009004288353025913, + 0.003915397450327873, + 0.0032394130248576403, + 0.0037999351043254137, + 0.036118846386671066, + 0.0032675089314579964, + 0.008265356533229351, + 0.06910602003335953, + 0.07111641019582748, + 0.009199854917824268, + 0.0055252425372600555, + 0.009178044274449348, + 0.014251385815441608, + 0.006138592027127743, + 0.01062347088009119, + 0.0024975407868623734, + 0.0021050579380244017, + 0.001981372246518731, + 0.001846767496317625, + 0.002376724500209093, + 0.0027660669293254614, + 0.0017195140244439244, + 0.0019932494033128023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.05327378958463669 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012562076561152935 + ], + "opposite_logprob_tokens": [ + -4.387562274932861 + ], + "logprob_reward": -0.012562076561152935, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012562076561152935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10183493047952652, + 0.06357279419898987, + 0.019312776625156403, + 0.038860633969306946, + 0.01918048970401287, + 0.016533849760890007, + 0.0792946144938469, + 0.18890783190727234, + 0.33801978826522827, + 0.13448230922222137 + ], + "credit_map": [ + 0.014488371089100838, + 0.02563488855957985, + 0.01150499191135168, + 0.010542499832808971, + 0.008990622125566006, + 0.011810173280537128, + 0.009721665643155575, + 0.009141717106103897, + 0.006108776666224003, + 0.011801397427916527, + 0.009082269854843616, + 0.007094486616551876, + 0.007405132986605167, + 0.008028516545891762, + 0.008296290412545204, + 0.005755928345024586, + 0.0020336806774139404, + 0.004490591585636139, + 0.0026059632655233145, + 0.0017950658220797777, + 0.0023202928714454174, + 0.002192229963839054, + 0.002094859257340431, + 0.001780093414708972, + 0.010417578741908073, + 0.004062623251229525, + 0.0018174370052292943, + 0.0017890582093968987, + 0.0023619732819497585, + 0.012678313069045544, + 0.00212755030952394, + 0.0036060998681932688, + 0.0026448701974004507, + 0.002199834678322077, + 0.001766767236404121, + 0.0022073877044022083, + 0.002891533775255084, + 0.0033880388364195824, + 0.0023806081153452396, + 0.0017014489276334643, + 0.0044263978488743305, + 0.00216210656799376, + 0.0009061151649802923, + 0.001150525757111609, + 0.0011603492312133312, + 0.004408762324601412, + 0.0010871614795178175, + 0.0012324313865974545, + 0.022807689383625984, + 0.005453299731016159, + 0.0026600852143019438, + 0.001956729916855693, + 0.00322144222445786, + 0.03281412646174431, + 0.0026835643220692873, + 0.007697673514485359, + 0.026570450514554977, + 0.06595653295516968, + 0.0040362062864005566, + 0.004313947167247534, + 0.006517370231449604, + 0.06470203399658203, + 0.004879693500697613, + 0.0119315842166543, + 0.1274360716342926, + 0.14385323226451874, + 0.008664977736771107, + 0.0066092186607420444, + 0.01144258864223957, + 0.024762796238064766, + 0.007213369011878967, + 0.008037539198994637, + 0.026783866807818413, + 0.0287080816924572, + 0.00850900262594223, + 0.014024320058524609, + 0.019545242190361023, + 0.011400246061384678, + 0.01691865548491478, + 0.008592890575528145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.012562076561152935 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00979604572057724 + ], + "opposite_logprob_tokens": [ + -4.634796142578125 + ], + "logprob_reward": -0.00979604572057724, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00979604572057724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11299379169940948, + 0.07124453037977219, + 0.056815557181835175, + 0.07025869190692902, + 0.025931840762495995, + 0.09394911676645279, + 0.1389557123184204, + 0.08010195940732956, + 0.31253254413604736, + 0.03721620887517929 + ], + "credit_map": [ + 0.014670665375888348, + 0.027023373171687126, + 0.01059311255812645, + 0.011663640849292278, + 0.013373148627579212, + 0.014278911054134369, + 0.011868037283420563, + 0.00952290277928114, + 0.008764290250837803, + 0.015878643840551376, + 0.008225133642554283, + 0.006371926050633192, + 0.00818998645991087, + 0.010650023818016052, + 0.007067097816616297, + 0.0060974303632974625, + 0.009542387910187244, + 0.006953825708478689, + 0.004252953454852104, + 0.005157201085239649, + 0.0063512674532830715, + 0.013447298668324947, + 0.005851248279213905, + 0.005259376484900713, + 0.011513900943100452, + 0.00492042675614357, + 0.006743260193616152, + 0.00670494232326746, + 0.006655213423073292, + 0.017848938703536987, + 0.009224540553987026, + 0.006647469475865364, + 0.004672814626246691, + 0.0024515590630471706, + 0.003176815574988723, + 0.0024527255445718765, + 0.002816147869452834, + 0.004997061565518379, + 0.0025844722986221313, + 0.00278024235740304, + 0.0282839797437191, + 0.013081250712275505, + 0.003044453449547291, + 0.002956469776108861, + 0.005383171606808901, + 0.030416037887334824, + 0.0033162261825054884, + 0.007467527408152819, + 0.04183126986026764, + 0.05566553398966789, + 0.003174852579832077, + 0.003528643399477005, + 0.005927966441959143, + 0.020642882212996483, + 0.0040961578488349915, + 0.004088403191417456, + 0.004739383701235056, + 0.021974947303533554, + 0.00271407887339592, + 0.0031499527394771576, + 0.003954634070396423, + 0.03346383944153786, + 0.003536527045071125, + 0.006568591576069593, + 0.11552248150110245, + 0.1303037852048874, + 0.007648597937077284, + 0.007193796336650848, + 0.010318809188902378, + 0.02388760633766651, + 0.008525771088898182, + 0.009131700731813908, + 0.007477347273379564, + 0.009056517854332924, + 0.0029380880296230316, + 0.0030679581686854362, + 0.0033522541634738445, + 0.00439426489174366, + 0.004053089767694473, + 0.0028766889590770006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.00979604572057724 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34871748089790344 + ], + "opposite_logprob_tokens": [ + -1.223717451095581 + ], + "logprob_reward": -0.34871748089790344, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34871748089790344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10459891706705093, + 0.025354094803333282, + 0.05252388119697571, + 0.04164723679423332, + 0.025994908064603806, + 0.08629120141267776, + 0.1183551773428917, + 0.042366933077573776, + 0.4842053949832916, + 0.018662266433238983 + ], + "credit_map": [ + 0.013021748512983322, + 0.023559555411338806, + 0.014241239987313747, + 0.009190009906888008, + 0.010723967105150223, + 0.014422258362174034, + 0.008818225003778934, + 0.010621911846101284, + 0.003935099113732576, + 0.003979585133492947, + 0.0023172926157712936, + 0.001773196505382657, + 0.0023421989753842354, + 0.006055090110749006, + 0.0022942994255572557, + 0.0026573343202471733, + 0.01245846413075924, + 0.004433184862136841, + 0.0031876619905233383, + 0.001967519987374544, + 0.0036571682430803776, + 0.01779005490243435, + 0.00290746265091002, + 0.006122368387877941, + 0.010028033517301083, + 0.003371103201061487, + 0.0025962193030864, + 0.002416312927380204, + 0.002845912706106901, + 0.013351035304367542, + 0.0030015516094863415, + 0.004037071485072374, + 0.0023022156674414873, + 0.0034615257754921913, + 0.0030254710000008345, + 0.0035759578458964825, + 0.003917890600860119, + 0.00333630433306098, + 0.003375314874574542, + 0.003000229597091675, + 0.02541612833738327, + 0.010016394779086113, + 0.003294636495411396, + 0.0032364593353122473, + 0.004556338768452406, + 0.028479719534516335, + 0.0038624522276222706, + 0.007429076824337244, + 0.02278757281601429, + 0.03964737430214882, + 0.003971027210354805, + 0.005452071316540241, + 0.004663675092160702, + 0.030567947775125504, + 0.004428701940923929, + 0.006836810614913702, + 0.00401895260438323, + 0.007710082456469536, + 0.002547660144045949, + 0.002558407373726368, + 0.0033320023212581873, + 0.01605389080941677, + 0.002772756852209568, + 0.003373181913048029, + 0.17254389822483063, + 0.20614206790924072, + 0.008960647508502007, + 0.00912809744477272, + 0.016553884372115135, + 0.050266437232494354, + 0.009423438459634781, + 0.011186925694346428, + 0.0036579612642526627, + 0.0035355789586901665, + 0.001978182001039386, + 0.0012283200630918145, + 0.0026423281524330378, + 0.0027773967012763023, + 0.0013696816749870777, + 0.0014728194801136851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.34871748089790344 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38703975081443787 + ], + "opposite_logprob_tokens": [ + -1.1370397806167603 + ], + "logprob_reward": -0.38703975081443787, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38703975081443787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06611838936805725, + 0.02369575761258602, + 0.02700178138911724, + 0.062294334173202515, + 0.035369906574487686, + 0.0692153349518776, + 0.4524098336696625, + 0.09843358397483826, + 0.1517120748758316, + 0.013748976401984692 + ], + "credit_map": [ + 0.007458477281033993, + 0.016781968995928764, + 0.008126791566610336, + 0.005342573393136263, + 0.00799386017024517, + 0.007805630564689636, + 0.006239661015570164, + 0.006369429640471935, + 0.0025492513086646795, + 0.005124735180288553, + 0.003225842723622918, + 0.0023226039484143257, + 0.0026771314442157745, + 0.0031040124595165253, + 0.002640040358528495, + 0.002052145078778267, + 0.004917329177260399, + 0.0040804860182106495, + 0.0026747945230454206, + 0.0022536886390298605, + 0.0022694149520248175, + 0.005593028850853443, + 0.002694157650694251, + 0.0025188857689499855, + 0.016503699123859406, + 0.006204552948474884, + 0.0029382200445979834, + 0.00206854403950274, + 0.0033017946407198906, + 0.021470101550221443, + 0.0033988170325756073, + 0.006408605724573135, + 0.007019215263426304, + 0.0026181223802268505, + 0.0023892074823379517, + 0.002624846762046218, + 0.002913421019911766, + 0.011221031658351421, + 0.0025262897834181786, + 0.004057773854583502, + 0.020143043249845505, + 0.008268828503787518, + 0.0028760016430169344, + 0.002646156819537282, + 0.0036945377942174673, + 0.023119483143091202, + 0.0029180587735027075, + 0.005549231544137001, + 0.06502078473567963, + 0.15376661717891693, + 0.009792949073016644, + 0.009442905895411968, + 0.015959298238158226, + 0.15540774166584015, + 0.011725397780537605, + 0.031294167041778564, + 0.0068511078134179115, + 0.029082681983709335, + 0.0028140516951680183, + 0.002373205730691552, + 0.004052007105201483, + 0.0416521281003952, + 0.003735966281965375, + 0.007872454822063446, + 0.03781435638666153, + 0.05032065883278847, + 0.008500869385898113, + 0.006146726664155722, + 0.009370519779622555, + 0.023224949836730957, + 0.008410164155066013, + 0.007923838682472706, + 0.002339314203709364, + 0.0022124534007161856, + 0.0012747204164043069, + 0.0012992934789508581, + 0.0021693783346563578, + 0.0017513082129880786, + 0.0012248816201463342, + 0.001477628480643034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.38703975081443787 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "cinematic shot begins in mid-ground and slowly pans rightward, synchronizing with a small, carved boat. At dusk, Yang Guifei, dressed in a Tang Dynasty-style, chest-length ruqu (a traditional Chinese dress) in a criss-cross pattern of crimson and bright yellow, reclines on the brocade cushions of the boat's cabin. Her towering bun, adorned with delicate gold hairpins and pearl ornaments, trembles with the boat's gentle sway. The soft golden light of the setting sun filters through the mist, illuminating her elegant silk attire, the phoenix and peony designs embroidered in gold thread on the sleeves. Her face is like a silver disk, her eyes slightly tipsy, a languid smile lingering, her posture relaxed and relaxed. The boat, constructed of rich, dark wood, its sides and supports intricately carved with scrolling patterns, silently cuts across the mirror-like surface of the lake, leaving behind slowly spreading ripples that reflect the vibrant sunset glow on the horizon and the distant, misty silhouettes of green mountains. One of her delicate hands rested casually on the side of the boat, and her slender fingertips occasionally casually stroked the cool lake water. The whole picture was filled with an extremely tranquil, luxurious and poetic atmosphere in warm golden tones.", + "seed": 4197, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_2/index_2658_seed_4197_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18805084209889172, + "vlm_logprob": -0.18805084209889172, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.18805084209889172, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2873218059539795 + ], + "opposite_logprob_tokens": [ + -1.4123218059539795 + ], + "logprob_reward": -0.2873218059539795, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2873218059539795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43363305926322937, + 0.06075205281376839, + 0.02020702138543129, + 0.09231511503458023, + 0.08354593813419342, + 0.12755434215068817, + 0.028031421825289726, + 0.05329931899905205, + 0.030221804976463318, + 0.07043995708227158 + ], + "credit_map": [ + 0.03651005029678345, + 0.02083957940340042, + 0.09357491135597229, + 0.021358920261263847, + 0.04167848080396652, + 0.17309288680553436, + 0.020866936072707176, + 0.02571127749979496, + 0.009768646210432053, + 0.007893159985542297, + 0.01112819742411375, + 0.0046716961078345776, + 0.006813062820583582, + 0.006800096947699785, + 0.007097058463841677, + 0.0065801325254142284, + 0.0025373012758791447, + 0.0022784871980547905, + 0.002348195994272828, + 0.0032903666142374277, + 0.0018804804421961308, + 0.002384586026892066, + 0.002110427478328347, + 0.003377176122739911, + 0.009789847768843174, + 0.011195065453648567, + 0.006849509198218584, + 0.013131346553564072, + 0.009563473053276539, + 0.011687533929944038, + 0.016561970114707947, + 0.013536369428038597, + 0.008192387409508228, + 0.009780284017324448, + 0.008123858831822872, + 0.009044075384736061, + 0.007782901171594858, + 0.011777509935200214, + 0.016688164323568344, + 0.012156762182712555, + 0.014603549614548683, + 0.013530856929719448, + 0.016899440437555313, + 0.014521459117531776, + 0.011097386479377747, + 0.014381852000951767, + 0.021217769011855125, + 0.02130202017724514, + 0.003037847578525543, + 0.00496658543124795, + 0.0022949515841901302, + 0.003197847632691264, + 0.002653513103723526, + 0.0037996266037225723, + 0.0037660414818674326, + 0.004315007012337446, + 0.006574670784175396, + 0.005416890140622854, + 0.007410743273794651, + 0.006076544988900423, + 0.004872974008321762, + 0.006120137870311737, + 0.006777840666472912, + 0.010049518197774887, + 0.004661201033741236, + 0.0032728060614317656, + 0.002172956708818674, + 0.002951915143057704, + 0.0026942540425807238, + 0.003122159978374839, + 0.004054551478475332, + 0.007291959132999182, + 0.009582811035215855, + 0.010372420772910118, + 0.007534608244895935, + 0.00978690106421709, + 0.007236344274133444, + 0.007688345853239298, + 0.00810064934194088, + 0.010137876495718956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.2873218059539795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012572435662150383 + ], + "opposite_logprob_tokens": [ + -4.387572288513184 + ], + "logprob_reward": -0.012572435662150383, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012572435662150383, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20790508389472961, + 0.13269756734371185, + 0.12827375531196594, + 0.0428912378847599, + 0.07485666126012802, + 0.06940889358520508, + 0.048747073858976364, + 0.11675121635198593, + 0.07579023391008377, + 0.10267826914787292 + ], + "credit_map": [ + 0.02031715027987957, + 0.019595127552747726, + 0.035781893879175186, + 0.014364260248839855, + 0.018488407135009766, + 0.06683241575956345, + 0.02219044789671898, + 0.010335378348827362, + 0.012194912880659103, + 0.011334340088069439, + 0.010144596919417381, + 0.014816399663686752, + 0.012074283324182034, + 0.01788710057735443, + 0.029299819841980934, + 0.02494610846042633, + 0.015370907261967659, + 0.013404222205281258, + 0.011627957224845886, + 0.019280849024653435, + 0.011806683614850044, + 0.014974265359342098, + 0.020466908812522888, + 0.021341970190405846, + 0.004723052028566599, + 0.0039016264490783215, + 0.0056260721758008, + 0.005379789508879185, + 0.004246399737894535, + 0.005162384826689959, + 0.006565847899764776, + 0.007286062464118004, + 0.008167203515768051, + 0.006688601803034544, + 0.009510532021522522, + 0.009292013011872768, + 0.00831008143723011, + 0.007800659164786339, + 0.01465618796646595, + 0.010431388393044472, + 0.007838871330022812, + 0.010541234165430069, + 0.010031338781118393, + 0.006920481100678444, + 0.0072051603347063065, + 0.009562713094055653, + 0.006534048821777105, + 0.01077504176646471, + 0.00707963714376092, + 0.006584655959159136, + 0.0052189514972269535, + 0.005493004806339741, + 0.004932407755404711, + 0.006027792580425739, + 0.00694267125800252, + 0.006467955186963081, + 0.017118504270911217, + 0.013452248647809029, + 0.011355987749993801, + 0.013906491920351982, + 0.009201000444591045, + 0.01139950379729271, + 0.014690515585243702, + 0.025626961141824722, + 0.010006298311054707, + 0.007959703914821148, + 0.006594075821340084, + 0.007459611631929874, + 0.005846643354743719, + 0.008526233024895191, + 0.011137476190924644, + 0.01826019026339054, + 0.012645532377064228, + 0.010211294516921043, + 0.00971309281885624, + 0.017330849543213844, + 0.009602930396795273, + 0.01068983506411314, + 0.014461487531661987, + 0.018023252487182617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.012572435662150383 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009792149998247623 + ], + "opposite_logprob_tokens": [ + -4.634792327880859 + ], + "logprob_reward": -0.009792149998247623, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009792149998247623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1480948030948639, + 0.05289461836218834, + 0.052478138357400894, + 0.11283872276544571, + 0.1264226883649826, + 0.19423431158065796, + 0.08857408165931702, + 0.04582814499735832, + 0.0539214089512825, + 0.12471308559179306 + ], + "credit_map": [ + 0.01337496004998684, + 0.012408463284373283, + 0.021724626421928406, + 0.011229302734136581, + 0.014871148392558098, + 0.036800164729356766, + 0.023566605523228645, + 0.014119538478553295, + 0.008106371387839317, + 0.007109007798135281, + 0.004158862866461277, + 0.0067537883296608925, + 0.008187641389667988, + 0.006830128375440836, + 0.0058762021362781525, + 0.005872619338333607, + 0.004826755728572607, + 0.00574695598334074, + 0.004878397099673748, + 0.008405654691159725, + 0.005312500521540642, + 0.007196109276264906, + 0.010201440192759037, + 0.005910327658057213, + 0.012165015563368797, + 0.013847765512764454, + 0.008328622207045555, + 0.016852719709277153, + 0.012655946426093578, + 0.013145942240953445, + 0.020938776433467865, + 0.014903941191732883, + 0.017693858593702316, + 0.014184873551130295, + 0.019566474482417107, + 0.011455824598670006, + 0.010846086777746677, + 0.016977207735180855, + 0.021583763882517815, + 0.014114607125520706, + 0.02436860091984272, + 0.01643148437142372, + 0.023094171658158302, + 0.024273676797747612, + 0.015728365629911423, + 0.020471977069973946, + 0.03915552794933319, + 0.030710525810718536, + 0.010536227375268936, + 0.01361538004130125, + 0.006095541641116142, + 0.010242287069559097, + 0.009313574060797691, + 0.013669580221176147, + 0.012952563352882862, + 0.01214892789721489, + 0.005458307918161154, + 0.005175380501896143, + 0.004892135504633188, + 0.005232177209109068, + 0.004554301034659147, + 0.005568358115851879, + 0.005996656138449907, + 0.008950832299888134, + 0.008513981476426125, + 0.005548315588384867, + 0.004629815462976694, + 0.0055080014280974865, + 0.004717703443020582, + 0.006256815046072006, + 0.006871664896607399, + 0.011875116266310215, + 0.015606618486344814, + 0.012564070522785187, + 0.01574583910405636, + 0.018696850165724754, + 0.010579937137663364, + 0.014019818976521492, + 0.017224833369255066, + 0.020275115966796875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.009792149998247623 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20166242122650146 + ], + "opposite_logprob_tokens": [ + -1.7016624212265015 + ], + "logprob_reward": -0.20166242122650146, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20166242122650146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18633967638015747, + 0.12306573241949081, + 0.0599198155105114, + 0.0630192831158638, + 0.10844669491052628, + 0.1128695160150528, + 0.06871985644102097, + 0.02898683212697506, + 0.06407181173563004, + 0.18456073105335236 + ], + "credit_map": [ + 0.021211611106991768, + 0.019359413534402847, + 0.01978171616792679, + 0.020112406462430954, + 0.022229501977562904, + 0.048545438796281815, + 0.014415299519896507, + 0.02068428136408329, + 0.019032256677746773, + 0.010262302123010159, + 0.02442782372236252, + 0.01413850486278534, + 0.010979857295751572, + 0.016291141510009766, + 0.010867211036384106, + 0.017066631466150284, + 0.006758257746696472, + 0.006649139802902937, + 0.007343529257923365, + 0.006751952227205038, + 0.00585390068590641, + 0.008220025338232517, + 0.01013057678937912, + 0.008212433196604252, + 0.007859750650823116, + 0.005938997026532888, + 0.007937375456094742, + 0.007326232735067606, + 0.0059792096726596355, + 0.007944461889564991, + 0.009003926999866962, + 0.01102933194488287, + 0.009885464794933796, + 0.010979505255818367, + 0.008164258673787117, + 0.017445921897888184, + 0.009071378968656063, + 0.015254591591656208, + 0.02483082190155983, + 0.01281475555151701, + 0.016377042979002, + 0.008601042442023754, + 0.01435395609587431, + 0.011020141653716564, + 0.008641691878437996, + 0.01562297623604536, + 0.016461145132780075, + 0.021791521459817886, + 0.008756346069276333, + 0.00948362611234188, + 0.005032052285969257, + 0.008557300083339214, + 0.005715386476367712, + 0.007270012982189655, + 0.010724452324211597, + 0.013180689886212349, + 0.004259226378053427, + 0.0036856650840491056, + 0.0041810437105596066, + 0.0036381979007273912, + 0.0028766740579158068, + 0.0036185518838465214, + 0.003395537845790386, + 0.0033319343347102404, + 0.008012935519218445, + 0.011356228031218052, + 0.005541714373975992, + 0.007013046648353338, + 0.009726758114993572, + 0.007125880103558302, + 0.0074026682414114475, + 0.007892581634223461, + 0.028242306783795357, + 0.02646377868950367, + 0.015257460065186024, + 0.028868671506643295, + 0.023318035528063774, + 0.019201861694455147, + 0.01954714022576809, + 0.023661470040678978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.20166242122650146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a small, carved boat constructed of rich dark wood with scrolling patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a woman wearing a Tang Dynasty-style chest-length ruqu dress in a criss-cross pattern of crimson and bright yellow visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman reclining on brocade cushions inside the boat's cabin with her face pale and a languid smile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the woman's hair bun adorned with gold hairpins and pearl ornaments tremble as the boat sways?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in the mid-ground and slowly pan rightward while the boat cuts across the lake?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289053976535797 + ], + "opposite_logprob_tokens": [ + -1.0539053678512573 + ], + "logprob_reward": -0.4289053976535797, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289053976535797, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14285056293010712, + 0.10355378687381744, + 0.17460548877716064, + 0.050362106412649155, + 0.0470765084028244, + 0.14516989886760712, + 0.0928194522857666, + 0.0499347485601902, + 0.07704898715019226, + 0.11657854169607162 + ], + "credit_map": [ + 0.01150117814540863, + 0.01563066616654396, + 0.022882195189595222, + 0.012326219119131565, + 0.014939901418983936, + 0.03303951397538185, + 0.01694696769118309, + 0.01558389887213707, + 0.01426265761256218, + 0.009828620590269566, + 0.01460359338670969, + 0.011989763006567955, + 0.012754368595778942, + 0.013785956427454948, + 0.014163720421493053, + 0.012165102176368237, + 0.01779489405453205, + 0.01647155173122883, + 0.012130596674978733, + 0.025869933888316154, + 0.017781207337975502, + 0.01834387518465519, + 0.033316973596811295, + 0.03289642930030823, + 0.0059740254655480385, + 0.006613363046199083, + 0.004537875298410654, + 0.006697993725538254, + 0.006270510610193014, + 0.006226660683751106, + 0.008785254321992397, + 0.00525641767308116, + 0.0037921597249805927, + 0.005023995880037546, + 0.003124627750366926, + 0.007452962454408407, + 0.0042460220865905285, + 0.0073414696380496025, + 0.00993393361568451, + 0.006161332596093416, + 0.013910330832004547, + 0.015980718657374382, + 0.013647940009832382, + 0.019931389018893242, + 0.014305991120636463, + 0.015017152763903141, + 0.02974856272339821, + 0.022627798840403557, + 0.010149175301194191, + 0.012261597439646721, + 0.007143986877053976, + 0.010948069393634796, + 0.008348898030817509, + 0.010583112947642803, + 0.012582490220665932, + 0.020802119746804237, + 0.006483993027359247, + 0.007008811458945274, + 0.005561819300055504, + 0.005624738987535238, + 0.005152470897883177, + 0.005984850227832794, + 0.005031854845583439, + 0.009086204692721367, + 0.008857419714331627, + 0.010278698988258839, + 0.00968221202492714, + 0.010143724270164967, + 0.006484555546194315, + 0.008170155808329582, + 0.0096797039732337, + 0.013752512633800507, + 0.019689315930008888, + 0.01396049465984106, + 0.01135150995105505, + 0.01800181157886982, + 0.013845240697264671, + 0.012031838297843933, + 0.011117205023765564, + 0.01658111996948719 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.4289053976535797 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_tensor([3137, 3137])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_tensor([3137, 3137])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..bd918f9dd1299e8e95091c5889049d4fc0a095d0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_tensor([3137, 3137])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning high-angle wide shot captures the scene, the camera panning slowly and steadily from left to right, revealing a seemingly boundless expanse of a ripe wheat field bathed in the brilliant midday sun. Dense golden stalks, laden with ripe grains, create a richly textured, undulating sea of amber and gold that stretches to the distant, hazy horizon. Powerful, unseen gusts of wind sweep across the land, choreographing the field into a mesmerizing dance; the stalks bend and rise in flowing, rippling waves, forming intricate, ever-shifting patterns of swirls and ephemeral spirals that chase each other across the surface. Above, the perfectly clear, azure sky contrasts starkly and beautifully with the warm tones below, its vastness magnifying the scale of the scene. The intense sunlight casts minimal shadows, causing the entire field to shimmer with a dazzling, breathtaking light, creating an atmosphere of profound peace, quiet majesty, and the raw, untamed beauty of nature in motion.", + "seed": 23250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23250_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11462459443137049, + "vlm_logprob": -0.11462459443137049, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.11462459443137049, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029951972886919975 + ], + "opposite_logprob_tokens": [ + -5.654952049255371 + ], + "logprob_reward": -0.029951972886919975, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.029951972886919975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018464108929038048, + 0.48833897709846497, + 0.009505907073616982, + 0.008257897570729256, + 0.013959360308945179, + 0.0075891693122684956, + 0.003743619192391634, + 0.005164139438420534, + 0.0035858137998729944, + 0.4413909614086151 + ], + "credit_map": [ + 0.0014854397159069777, + 0.003122104099020362, + 0.0022993178572505713, + 0.003348882542923093, + 0.001563251600600779, + 0.0013771296944469213, + 0.003097805893048644, + 0.0021701790392398834, + 0.010005765594542027, + 0.009871572256088257, + 0.19570867717266083, + 0.21814273297786713, + 0.005979826208204031, + 0.005063476040959358, + 0.024749720469117165, + 0.018817203119397163, + 0.0005707742529921234, + 0.00084004329983145, + 0.0016914636362344027, + 0.0014204623876139522, + 0.000906301022041589, + 0.0008056277292780578, + 0.0013040906051173806, + 0.001967144198715687, + 0.0005937695968896151, + 0.0007759823347441852, + 0.001234881579875946, + 0.0015934709226712584, + 0.0005630374071188271, + 0.0006534905405715108, + 0.0015149745158851147, + 0.0013282903237268329, + 0.0006410136120393872, + 0.0007963264943100512, + 0.002054446842521429, + 0.0028679962269961834, + 0.0008309025433845818, + 0.0009478596039116383, + 0.0026532646734267473, + 0.0031675505451858044, + 0.0004464329103939235, + 0.0006695148185826838, + 0.0009293573093600571, + 0.001237391377799213, + 0.000646744854748249, + 0.0004682289727497846, + 0.0016558088827878237, + 0.001535690505988896, + 0.0002455026551615447, + 0.0003620521747507155, + 0.000605150533374399, + 0.0006446781917475164, + 0.000325542816426605, + 0.0003105275973211974, + 0.0006677800556644797, + 0.0005823855171911418, + 0.0003698361979331821, + 0.0007912751752883196, + 0.0006750710890628397, + 0.0009229762363247573, + 0.0004746531485579908, + 0.000541279383469373, + 0.0006756843067705631, + 0.0007133634644560516, + 0.00035811495035886765, + 0.0003139469481538981, + 0.0006086569628678262, + 0.0005420896923169494, + 0.00040098634781315923, + 0.0003549981629475951, + 0.000579141138587147, + 0.0004278795386198908, + 0.012535460293293, + 0.004907576832920313, + 0.13875894248485565, + 0.033813074231147766, + 0.010930968448519707, + 0.004904133267700672, + 0.19337329268455505, + 0.04216751083731651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.029951972886919975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743231534957886 + ], + "opposite_logprob_tokens": [ + -0.9743231534957886 + ], + "logprob_reward": -0.4743231534957886, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743231534957886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010329881682991982, + 0.9168961048126221, + 0.004953271243721247, + 0.0021413699723780155, + 0.001957372762262821, + 0.001806184183806181, + 0.004727023653686047, + 0.002461919328197837, + 0.0009606091189198196, + 0.05376627296209335 + ], + "credit_map": [ + 0.0008873913320712745, + 0.001477280748076737, + 0.001219758647494018, + 0.0017225455958396196, + 0.0009539119782857597, + 0.0009246445260941982, + 0.0015300321392714977, + 0.001614316483028233, + 0.020718572661280632, + 0.017929181456565857, + 0.3760439157485962, + 0.4118936359882355, + 0.005013891030102968, + 0.006577847991138697, + 0.04449020326137543, + 0.03422881290316582, + 0.00029139546677470207, + 0.0007699894485995173, + 0.000759345362894237, + 0.0008081996347755194, + 0.0006163805956020951, + 0.00038179714465513825, + 0.000540627574082464, + 0.0007855353760533035, + 0.00012229851563461125, + 0.00017701163596939296, + 0.00036399936652742326, + 0.00043181469663977623, + 9.397002577316016e-05, + 0.0001332191313849762, + 0.00043726532021537423, + 0.0003817910619545728, + 0.0001065623655449599, + 0.000286890339339152, + 0.00024664850207045674, + 0.000382136000553146, + 0.0001859267067629844, + 0.00012285313277971, + 0.0003406963078305125, + 0.00028565910179167986, + 0.000131897977553308, + 0.00020939807291142642, + 0.00019503881048876792, + 0.00027715833857655525, + 0.00014715718862134963, + 0.00011916210496565327, + 0.00032830206328071654, + 0.000398069474613294, + 0.0004539928922895342, + 0.0005528844194486737, + 0.0005322526558302343, + 0.0006310614990070462, + 0.0007157536456361413, + 0.0005127619951963425, + 0.0005975667736493051, + 0.0007307492196559906, + 0.00024054752429947257, + 0.0004112514143344015, + 0.0002666270302142948, + 0.00031421275343745947, + 0.0003334249777253717, + 0.00027314218459650874, + 0.0002793092280626297, + 0.0003434040118008852, + 0.00011323665967211127, + 0.00011702637129928917, + 0.00014947705494705588, + 0.00010524624667596072, + 0.00012157680612290278, + 9.934437548508868e-05, + 0.00010607546573737636, + 0.00014862608804833144, + 0.0016541429795324802, + 0.0007273242226801813, + 0.021181022748351097, + 0.0040499367751181126, + 0.0012221382930874825, + 0.0010361761087551713, + 0.020182494074106216, + 0.003713030368089676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.4743231534957886 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013428251259028912 + ], + "opposite_logprob_tokens": [ + -6.6263427734375 + ], + "logprob_reward": -0.0013428251259028912, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013428251259028912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.021425217390060425, + 0.9113228917121887, + 0.008865204639732838, + 0.005436208099126816, + 0.007734017912298441, + 0.008359971456229687, + 0.00341980904340744, + 0.005114531610161066, + 0.0014009912265464664, + 0.026921164244413376 + ], + "credit_map": [ + 0.0014445786364376545, + 0.002937040524557233, + 0.0033698356710374355, + 0.004457646049559116, + 0.0015495455591008067, + 0.0014541224809363484, + 0.003349052043631673, + 0.0028633929323405027, + 0.024287093430757523, + 0.019431065768003464, + 0.3737799823284149, + 0.401950865983963, + 0.007035494782030582, + 0.009047683328390121, + 0.042418718338012695, + 0.0333719439804554, + 0.0006921332678757608, + 0.0007546775741502643, + 0.001216079224832356, + 0.0015363275306299329, + 0.000909836555365473, + 0.0007926571415737271, + 0.001257646712474525, + 0.0017058458179235458, + 0.0004265944880899042, + 0.0005628826911561191, + 0.0006620525382459164, + 0.0009181330096907914, + 0.0005624284967780113, + 0.0004948845016770065, + 0.0009058341383934021, + 0.0009033972746692598, + 0.0004716940165963024, + 0.0007813394768163562, + 0.0010203375713899732, + 0.0013748533092439175, + 0.0005046014557592571, + 0.0006367415189743042, + 0.0013816567370668054, + 0.001562792924232781, + 0.0007027231622487307, + 0.000702354998793453, + 0.0011762870708480477, + 0.0013187346048653126, + 0.0007045553647913039, + 0.0005609257495962083, + 0.001469313632696867, + 0.0017250749515369534, + 0.0002758184855338186, + 0.0004029528354294598, + 0.0004437190655153245, + 0.0005875994102098048, + 0.0004641036211978644, + 0.0003989562683273107, + 0.0004928488051518798, + 0.0003538098535500467, + 0.00028909704997204244, + 0.0007042918587103486, + 0.0006005627219565213, + 0.0012065359624102712, + 0.0003896098642144352, + 0.00037738712853752077, + 0.0007159304223023355, + 0.0008311157580465078, + 9.0972134785261e-05, + 9.42209007916972e-05, + 0.0002233951963717118, + 0.0003817490360233933, + 0.00011620752047747374, + 0.00012818585673812777, + 0.00022467627422884107, + 0.00014158406702335924, + 0.0011590853100642562, + 0.0006359366234391928, + 0.004492128733545542, + 0.0033868150785565376, + 0.0008781235083006322, + 0.0006313813501037657, + 0.011414740234613419, + 0.0043229470029473305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0013428251259028912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496366322040558 + ], + "opposite_logprob_tokens": [ + -2.9299635887145996 + ], + "logprob_reward": -0.05496366322040558, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496366322040558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02581131085753441, + 0.8854336738586426, + 0.011419514194130898, + 0.015060901641845703, + 0.017765028402209282, + 0.009595238603651524, + 0.004874960519373417, + 0.010817185044288635, + 0.0016721428837627172, + 0.017550069838762283 + ], + "credit_map": [ + 0.002139634219929576, + 0.0018567239167168736, + 0.0040346370078623295, + 0.0059086172841489315, + 0.0013132264139130712, + 0.0018206191016361117, + 0.004721880424767733, + 0.004015973303467035, + 0.02316848561167717, + 0.020676221698522568, + 0.3614449203014374, + 0.3881230056285858, + 0.007753592915832996, + 0.009993362240493298, + 0.0412323921918869, + 0.03304169699549675, + 0.0008670406532473862, + 0.001582141499966383, + 0.0022502848878502846, + 0.0016277293907478452, + 0.0010030338307842612, + 0.000860069238115102, + 0.0013767087366431952, + 0.0018525064224377275, + 0.0009445689502172172, + 0.0013082802761346102, + 0.0018972541438415647, + 0.003047650447115302, + 0.0013328467030078173, + 0.0009404342854395509, + 0.002903923625126481, + 0.00268594385124743, + 0.0011250792304053903, + 0.0012588929384946823, + 0.0024339037481695414, + 0.0030590142123401165, + 0.0009955373825505376, + 0.0014091491466388106, + 0.0033651848789304495, + 0.00411826791241765, + 0.0005065920995548368, + 0.0007786951609887183, + 0.0014787232503294945, + 0.0017255457350984216, + 0.0009569153189659119, + 0.0006641540676355362, + 0.0019272429635748267, + 0.0015573700657114387, + 0.0004906654357910156, + 0.0008123407606035471, + 0.0007919553318060935, + 0.0006200448842719197, + 0.0003713818150572479, + 0.0004306513292249292, + 0.0006593967555090785, + 0.0006985244108363986, + 0.0005798098281957209, + 0.0012323206756263971, + 0.0015894039534032345, + 0.0028322122525423765, + 0.000599827035330236, + 0.0006766016595065594, + 0.0013785766204819083, + 0.001928433426655829, + 0.00010535680485190824, + 0.00014319937326945364, + 0.0002655661082826555, + 0.000363975646905601, + 0.00016202114056795835, + 0.00014968404138926417, + 0.00028173098689876497, + 0.0002006088470807299, + 0.0012473955284804106, + 0.001182124949991703, + 0.00487171858549118, + 0.0025252762716263533, + 0.001621139352209866, + 0.000969411397818476, + 0.0016444388311356306, + 0.0034885662607848644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.05496366322040558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541357427835464 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541357427835464, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541357427835464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030180055648088455, + 0.8339049816131592, + 0.008056586608290672, + 0.008051758632063866, + 0.005790713243186474, + 0.00993769895285368, + 0.017694979906082153, + 0.006284431088715792, + 0.0037282833363860846, + 0.07637055963277817 + ], + "credit_map": [ + 0.002501639537513256, + 0.004702170379459858, + 0.00450592627748847, + 0.005174473859369755, + 0.0024676911998540163, + 0.002067595487460494, + 0.004791355691850185, + 0.0039691985584795475, + 0.01939946971833706, + 0.016154779121279716, + 0.34309932589530945, + 0.3740389347076416, + 0.0048101358115673065, + 0.00585637241601944, + 0.03985367342829704, + 0.0306922048330307, + 0.0007049132254905999, + 0.001231831032782793, + 0.001015171525068581, + 0.0012175634037703276, + 0.001061613904312253, + 0.0006755976937711239, + 0.00107527244836092, + 0.0010746222687885165, + 0.0006808722391724586, + 0.0005651985411532223, + 0.0011669142404571176, + 0.0016354562249034643, + 0.00043868739157915115, + 0.0005405626143328846, + 0.0016884795622900128, + 0.0013355867704376578, + 0.00035886812838725746, + 0.0008408436551690102, + 0.0008135840762406588, + 0.0008844271069392562, + 0.0006635074387304485, + 0.000548436539247632, + 0.000900333165191114, + 0.0007807120564393699, + 0.0007085291435942054, + 0.0018629062687978148, + 0.0009713685140013695, + 0.001376757281832397, + 0.0011039222590625286, + 0.0010063537629321218, + 0.0015271190786734223, + 0.0013807413633912802, + 0.001418428961187601, + 0.0019804947078227997, + 0.002411782741546631, + 0.0030495268292725086, + 0.001815056661143899, + 0.0016125041292980313, + 0.002445886842906475, + 0.002961296821013093, + 0.0003177934268023819, + 0.0005958836991339922, + 0.0008998197736218572, + 0.0018358256202191114, + 0.0004221764102112502, + 0.00039787025889381766, + 0.0008578670676797628, + 0.0009571944246999919, + 0.0002947159227915108, + 0.0002930984483100474, + 0.000713946355972439, + 0.0005056884256191552, + 0.00022254181385505944, + 0.00035287707578390837, + 0.0005439742817543447, + 0.000801440270151943, + 0.0023008219432085752, + 0.0008620540029369295, + 0.026809226721525192, + 0.00516381673514843, + 0.0021349939052015543, + 0.0008732487913221121, + 0.03196260333061218, + 0.00626378133893013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.012541357427835464 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot captures the scene, the camera panning slowly and steadily from left to right, revealing a seemingly boundless expanse of a ripe wheat field bathed in the brilliant midday sun. Dense golden stalks, laden with ripe grains, create a richly textured, undulating sea of amber and gold that stretches to the distant, hazy horizon. Powerful, unseen gusts of wind sweep across the land, choreographing the field into a mesmerizing dance; the stalks bend and rise in flowing, rippling waves, forming intricate, ever-shifting patterns of swirls and ephemeral spirals that chase each other across the surface. Above, the perfectly clear, azure sky contrasts starkly and beautifully with the warm tones below, its vastness magnifying the scale of the scene. The intense sunlight casts minimal shadows, causing the entire field to shimmer with a dazzling, breathtaking light, creating an atmosphere of profound peace, quiet majesty, and the raw, untamed beauty of nature in motion.", + "seed": 23251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23251_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.15278832367621362, + "vlm_logprob": -0.15278832367621362, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.15278832367621362, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02132335677742958 + ], + "opposite_logprob_tokens": [ + -5.271323204040527 + ], + "logprob_reward": -0.02132335677742958, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.02132335677742958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3600165545940399, + 0.13986240327358246, + 0.05526910349726677, + 0.012543887831270695, + 0.09381594508886337, + 0.009865224361419678, + 0.1935911625623703, + 0.09737075865268707, + 0.01461863424628973, + 0.02304629608988762 + ], + "credit_map": [ + 0.0055845859460532665, + 0.008964650332927704, + 0.020708225667476654, + 0.07928634434938431, + 0.01188093051314354, + 0.010381080210208893, + 0.1585151106119156, + 0.06469565629959106, + 0.011374406516551971, + 0.007539763115346432, + 0.015994731336832047, + 0.04352481663227081, + 0.009324144572019577, + 0.015608194284141064, + 0.019265219569206238, + 0.01723112352192402, + 0.0035582042764872313, + 0.011175299994647503, + 0.0034332273062318563, + 0.006242629140615463, + 0.00521486671641469, + 0.008056776598095894, + 0.012040385976433754, + 0.005547711160033941, + 0.0007127067656256258, + 0.0007763077737763524, + 0.0018569182138890028, + 0.001974459271878004, + 0.0011776478495448828, + 0.0009485603077337146, + 0.0033033215440809727, + 0.0017939668614417315, + 0.0029642777517437935, + 0.009633626788854599, + 0.01809951476752758, + 0.03365208953619003, + 0.005390490870922804, + 0.007703144568949938, + 0.007215925958007574, + 0.009156866930425167, + 0.0006330932956188917, + 0.0012728706933557987, + 0.001340813236311078, + 0.0018551223911345005, + 0.0012158865574747324, + 0.0006356138619594276, + 0.0014327972894534469, + 0.0014790268614888191, + 0.011527404189109802, + 0.012497831135988235, + 0.0397113598883152, + 0.028217632323503494, + 0.021261515095829964, + 0.033640243113040924, + 0.023468628525733948, + 0.02326655574142933, + 0.004511354956775904, + 0.01195798721164465, + 0.01079940889030695, + 0.020025914534926414, + 0.006619764491915703, + 0.0059585548005998135, + 0.02793746441602707, + 0.009560313075780869, + 0.000923330313526094, + 0.0013965492835268378, + 0.0012495318660512567, + 0.0026960892137140036, + 0.001550829503685236, + 0.0030507470946758986, + 0.001893688808195293, + 0.001857868512161076, + 0.001186038600280881, + 0.0012551494874060154, + 0.007325904909521341, + 0.0034072287380695343, + 0.001384940231218934, + 0.0013570613227784634, + 0.0030795081984251738, + 0.004050463438034058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02132335677742958 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933870315551758 + ], + "opposite_logprob_tokens": [ + -0.6933870315551758 + ], + "logprob_reward": -0.6933870315551758, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933870315551758, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.4565902352333069, + 0.029720859602093697, + 0.05197630822658539, + 0.03179299458861351, + 0.2610184848308563, + 0.01896471343934536, + 0.016982030123472214, + 0.045556195080280304, + 0.028343407437205315, + 0.05905478075146675 + ], + "credit_map": [ + 0.008546153083443642, + 0.007438342086970806, + 0.028226401656866074, + 0.0820574164390564, + 0.018960103392601013, + 0.02066878229379654, + 0.19740216434001923, + 0.09329091012477875, + 0.0020683829206973314, + 0.002610589610412717, + 0.003699576249346137, + 0.009413927793502808, + 0.001974404091015458, + 0.0016592151951044798, + 0.004511435516178608, + 0.003783327993005514, + 0.0031720255501568317, + 0.004166169557720423, + 0.003936907276511192, + 0.011055241338908672, + 0.0034298186656087637, + 0.005988347344100475, + 0.015091515146195889, + 0.005136283114552498, + 0.002556929597631097, + 0.004427161067724228, + 0.0036887754686176777, + 0.005871400237083435, + 0.004097857978194952, + 0.0028264280408620834, + 0.004317147191613913, + 0.004007298033684492, + 0.016767777502536774, + 0.06391464918851852, + 0.02324630320072174, + 0.03486381843686104, + 0.03201284632086754, + 0.05368593707680702, + 0.014624863862991333, + 0.021902281790971756, + 0.0012541969772428274, + 0.0012922115856781602, + 0.002877696417272091, + 0.002881882479414344, + 0.0015609115362167358, + 0.0018100838642567396, + 0.004347026813775301, + 0.002940703881904483, + 0.0010300432331860065, + 0.0017890293383970857, + 0.00250032264739275, + 0.002244481584057212, + 0.0012456448748707771, + 0.002628103131428361, + 0.002801756141707301, + 0.002742649521678686, + 0.002981183584779501, + 0.006301491521298885, + 0.009703547693789005, + 0.004400129429996014, + 0.003898587077856064, + 0.0039646509103477, + 0.005463209468871355, + 0.008843393996357918, + 0.0019344753818586469, + 0.006344032473862171, + 0.002776404842734337, + 0.004427456762641668, + 0.0027981942985206842, + 0.004359957296401262, + 0.0029016579501330853, + 0.0028012297116219997, + 0.0036011359188705683, + 0.003759233048185706, + 0.00759067153558135, + 0.013988207094371319, + 0.005386606324464083, + 0.008434049785137177, + 0.00957715604454279, + 0.006717722862958908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.6933870315551758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013433012645691633 + ], + "opposite_logprob_tokens": [ + -6.626343250274658 + ], + "logprob_reward": -0.0013433012645691633, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013433012645691633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4573774039745331, + 0.04132392257452011, + 0.014250722713768482, + 0.036888107657432556, + 0.18969634175300598, + 0.014310787431895733, + 0.1130685806274414, + 0.10217594355344772, + 0.006262925919145346, + 0.024645252153277397 + ], + "credit_map": [ + 0.004736958537250757, + 0.017758546397089958, + 0.022152403369545937, + 0.1099088117480278, + 0.02350577525794506, + 0.010452481918036938, + 0.1961139589548111, + 0.07274844497442245, + 0.0027797608636319637, + 0.007103332784026861, + 0.003621969372034073, + 0.004770231433212757, + 0.004377728793770075, + 0.007979540154337883, + 0.006101549137383699, + 0.004589809570461512, + 0.0011484835995361209, + 0.0017575289821252227, + 0.0012832520296797156, + 0.001334276981651783, + 0.0022835079580545425, + 0.0031996106263250113, + 0.0012498111464083195, + 0.001994251273572445, + 0.003115080064162612, + 0.008336269296705723, + 0.003828952554613352, + 0.005964027717709541, + 0.004803648218512535, + 0.003254130482673645, + 0.003986350726336241, + 0.0035996458027511835, + 0.011360209435224533, + 0.04192633926868439, + 0.02155127376317978, + 0.03374601900577545, + 0.019006922841072083, + 0.035845715552568436, + 0.009471984580159187, + 0.01678786426782608, + 0.0009420915157534182, + 0.0018597424495965242, + 0.0016744681634008884, + 0.002289124531671405, + 0.0014666655333712697, + 0.0011493114288896322, + 0.0023918026126921177, + 0.0025375792756676674, + 0.006809747777879238, + 0.007578428369015455, + 0.022931566461920738, + 0.016045883297920227, + 0.01208554394543171, + 0.02057153917849064, + 0.01364833489060402, + 0.013397536240518093, + 0.005180379841476679, + 0.015569420531392097, + 0.011103636585175991, + 0.018676474690437317, + 0.008303276263177395, + 0.008861605077981949, + 0.02483663521707058, + 0.009644518606364727, + 0.0003838418342638761, + 0.0004571913741528988, + 0.0006425482570193708, + 0.0010572324972599745, + 0.0010276754619553685, + 0.0010378800798207521, + 0.0006600371561944485, + 0.0009965194622054696, + 0.0017117528477683663, + 0.002483397489413619, + 0.0048753321170806885, + 0.004752198699861765, + 0.0020833213347941637, + 0.0015976954018697143, + 0.004432570189237595, + 0.002708981977775693 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.0013433012645691633 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810510411858559 + ], + "opposite_logprob_tokens": [ + -3.288105010986328 + ], + "logprob_reward": -0.03810510411858559, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810510411858559, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04371357336640358, + 0.14001162350177765, + 0.039673469960689545, + 0.027736736461520195, + 0.1499011218547821, + 0.01390840858221054, + 0.16196364164352417, + 0.3481545150279999, + 0.020849715918302536, + 0.054087214171886444 + ], + "credit_map": [ + 0.0029039804358035326, + 0.008422261103987694, + 0.003649518359452486, + 0.0083316033706069, + 0.005133708473294973, + 0.0057794442400336266, + 0.0027431242633610964, + 0.006749933119863272, + 0.011347334831953049, + 0.008528806269168854, + 0.015256866812705994, + 0.039636414498090744, + 0.010254797525703907, + 0.019212964922189713, + 0.01908021606504917, + 0.016694216057658195, + 0.0025794252287596464, + 0.0033083846792578697, + 0.0020941379480063915, + 0.008515426889061928, + 0.0024833683855831623, + 0.004124228376895189, + 0.012169897556304932, + 0.004398598801344633, + 0.002162395277991891, + 0.007550748996436596, + 0.0020104350987821817, + 0.0026056861970573664, + 0.0030468071345239878, + 0.004938937723636627, + 0.002852360252290964, + 0.0025693662464618683, + 0.010587931610643864, + 0.03731584921479225, + 0.012719398364424706, + 0.01606125198304653, + 0.016407903283834457, + 0.03465980291366577, + 0.008950281888246536, + 0.01319870539009571, + 0.0010773494141176343, + 0.001166953588835895, + 0.0019109116401523352, + 0.0019134941976517439, + 0.0009451357764191926, + 0.0019650249741971493, + 0.002132146619260311, + 0.002797391964122653, + 0.00987024325877428, + 0.010589788667857647, + 0.03306092694401741, + 0.022767910733819008, + 0.017799334600567818, + 0.029828518629074097, + 0.018745746463537216, + 0.019301172345876694, + 0.016480980440974236, + 0.043653350323438644, + 0.039047155529260635, + 0.06964073330163956, + 0.02409796044230461, + 0.023178430274128914, + 0.09662309288978577, + 0.03543280437588692, + 0.0012062093010172248, + 0.001871942775323987, + 0.0018791130278259516, + 0.003491149516776204, + 0.002767899539321661, + 0.004493608605116606, + 0.002606963040307164, + 0.0025328288320451975, + 0.0038857997860759497, + 0.006194916553795338, + 0.008792394772171974, + 0.00850700493901968, + 0.006035607308149338, + 0.004144898150116205, + 0.010733046568930149, + 0.005793548654764891 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.03810510411858559 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782824665307999 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782824665307999, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782824665307999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.48761996626853943, + 0.019170580431818962, + 0.02785787731409073, + 0.011530199088156223, + 0.07318849861621857, + 0.019441545009613037, + 0.21990108489990234, + 0.10994875431060791, + 0.01516774296760559, + 0.0161737110465765 + ], + "credit_map": [ + 0.005486168432980776, + 0.0071744415909051895, + 0.027462676167488098, + 0.10947325825691223, + 0.014645248651504517, + 0.008409225381910801, + 0.22215694189071655, + 0.09281203150749207, + 0.0014070018660277128, + 0.0011825637193396688, + 0.0026078629307448864, + 0.004922747146338224, + 0.0012588195968419313, + 0.0022355683613568544, + 0.002605810994282365, + 0.00295020523481071, + 0.0021419059485197067, + 0.007247517351061106, + 0.002001603599637747, + 0.0026803528890013695, + 0.004065594170242548, + 0.003440715605393052, + 0.0038368108216673136, + 0.0024433783255517483, + 0.0009787635644897819, + 0.0007050495478324592, + 0.0019117508782073855, + 0.001815931755118072, + 0.0011109113693237305, + 0.0007553458563052118, + 0.0021307216957211494, + 0.0021217248868197203, + 0.005063874647021294, + 0.02127588912844658, + 0.004103097598999739, + 0.00688087148591876, + 0.01267623994499445, + 0.013230541720986366, + 0.00456980848684907, + 0.005388182122260332, + 0.00104780332185328, + 0.003128254320472479, + 0.003008913015946746, + 0.002069275826215744, + 0.0015204489463940263, + 0.00288293045014143, + 0.0030575143173336983, + 0.0027264065574854612, + 0.01281263493001461, + 0.015393680892884731, + 0.044095247983932495, + 0.03183884546160698, + 0.023123018443584442, + 0.0401211753487587, + 0.0271914079785347, + 0.025325080379843712, + 0.005360935814678669, + 0.012538556940853596, + 0.014878307469189167, + 0.019973378628492355, + 0.006812660023570061, + 0.0087615717202425, + 0.02743707224726677, + 0.014186279848217964, + 0.0009335720096714795, + 0.0025550806894898415, + 0.0010989694856107235, + 0.0021103438921272755, + 0.0018476136028766632, + 0.003867087885737419, + 0.0013430495746433735, + 0.0014120247215032578, + 0.0008794562891125679, + 0.0010742134181782603, + 0.0027732420712709427, + 0.003737466875463724, + 0.0013935680035501719, + 0.0018318661022931337, + 0.002425400773063302, + 0.002058496465906501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.009782824665307999 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot captures the scene, the camera panning slowly and steadily from left to right, revealing a seemingly boundless expanse of a ripe wheat field bathed in the brilliant midday sun. Dense golden stalks, laden with ripe grains, create a richly textured, undulating sea of amber and gold that stretches to the distant, hazy horizon. Powerful, unseen gusts of wind sweep across the land, choreographing the field into a mesmerizing dance; the stalks bend and rise in flowing, rippling waves, forming intricate, ever-shifting patterns of swirls and ephemeral spirals that chase each other across the surface. Above, the perfectly clear, azure sky contrasts starkly and beautifully with the warm tones below, its vastness magnifying the scale of the scene. The intense sunlight casts minimal shadows, causing the entire field to shimmer with a dazzling, breathtaking light, creating an atmosphere of profound peace, quiet majesty, and the raw, untamed beauty of nature in motion.", + "seed": 23252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23252_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13902453447226434, + "vlm_logprob": -0.13902453447226434, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.13902453447226434, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02377512864768505 + ], + "opposite_logprob_tokens": [ + -5.773775100708008 + ], + "logprob_reward": -0.02377512864768505, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.02377512864768505, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0489526204764843, + 0.06847278773784637, + 0.15632809698581696, + 0.032691437751054764, + 0.031895965337753296, + 0.45504504442214966, + 0.04717803746461868, + 0.018763624131679535, + 0.046374108642339706, + 0.09429825842380524 + ], + "credit_map": [ + 0.003992892801761627, + 0.003924087155610323, + 0.007523803506046534, + 0.00949687696993351, + 0.004777326248586178, + 0.006209175568073988, + 0.005771571770310402, + 0.0072568804025650024, + 0.003530464367941022, + 0.004697410389780998, + 0.00971702579408884, + 0.007493933662772179, + 0.013545729219913483, + 0.017294786870479584, + 0.007470273412764072, + 0.004723160061985254, + 0.008392980322241783, + 0.015878157690167427, + 0.013983409851789474, + 0.026315150782465935, + 0.018533922731876373, + 0.021986691281199455, + 0.030102243646979332, + 0.021135540679097176, + 0.0027127342764288187, + 0.005401707254350185, + 0.0026133579667657614, + 0.0029823998920619488, + 0.004016413353383541, + 0.008950904943048954, + 0.0030328729189932346, + 0.0029810431879013777, + 0.0023995181545615196, + 0.0022975404281169176, + 0.005012757144868374, + 0.006147750187665224, + 0.0035664017777889967, + 0.006879114080220461, + 0.0025949114933609962, + 0.002997972536832094, + 0.015765385702252388, + 0.0223097987473011, + 0.17191511392593384, + 0.07604414224624634, + 0.04422459006309509, + 0.013375204987823963, + 0.08773719519376755, + 0.02367360331118107, + 0.003104737028479576, + 0.00510876439511776, + 0.005790955387055874, + 0.006531653925776482, + 0.008655132725834846, + 0.007998805493116379, + 0.004329069517552853, + 0.005658914800733328, + 0.001236748998053372, + 0.0018340091919526458, + 0.002030683448538184, + 0.004808907862752676, + 0.0013584606349468231, + 0.001601974363438785, + 0.0021891379728913307, + 0.003703700378537178, + 0.0041281902231276035, + 0.005707193166017532, + 0.003666802542284131, + 0.011743830516934395, + 0.0039633591659367085, + 0.004472780041396618, + 0.004665131215006113, + 0.008026818744838238, + 0.0031083913054317236, + 0.0036835740320384502, + 0.00471910834312439, + 0.023163815960288048, + 0.0060053193010389805, + 0.005513898096978664, + 0.034998420625925064, + 0.0131057258695364 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.02377512864768505 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762161016464233 + ], + "opposite_logprob_tokens": [ + -0.8262161016464233 + ], + "logprob_reward": -0.5762161016464233, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762161016464233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08519696444272995, + 0.05319228768348694, + 0.16849523782730103, + 0.015463371761143208, + 0.028799813240766525, + 0.30348339676856995, + 0.07959362119436264, + 0.11929453164339066, + 0.018725940957665443, + 0.127754807472229 + ], + "credit_map": [ + 0.006719588302075863, + 0.008016821928322315, + 0.019341539591550827, + 0.016456177458167076, + 0.009192083962261677, + 0.007215732242912054, + 0.007278298027813435, + 0.010976724326610565, + 0.0037150983698666096, + 0.006793678738176823, + 0.007326932158321142, + 0.00981870200484991, + 0.004797628149390221, + 0.009205816313624382, + 0.005281425081193447, + 0.00625300407409668, + 0.009777597151696682, + 0.01710006222128868, + 0.015785615891218185, + 0.029667669907212257, + 0.021331708878278732, + 0.021999746561050415, + 0.029679372906684875, + 0.02315347082912922, + 0.0012600974878296256, + 0.0009404923766851425, + 0.0020370164420455694, + 0.0030703528318554163, + 0.0012362611014395952, + 0.001503409817814827, + 0.002178980503231287, + 0.0032367606181651354, + 0.0020571521017700434, + 0.002938172547146678, + 0.0022253263741731644, + 0.0028045549988746643, + 0.004587599076330662, + 0.008498790673911572, + 0.0028358169365674257, + 0.0028524030931293964, + 0.01100053172558546, + 0.014927705749869347, + 0.11397457867860794, + 0.04970372095704079, + 0.030372168868780136, + 0.00909485761076212, + 0.058826759457588196, + 0.0155830392614007, + 0.005207004491239786, + 0.008126331493258476, + 0.008337740786373615, + 0.013859051279723644, + 0.012709801085293293, + 0.01107152458280325, + 0.00915277749300003, + 0.011129393242299557, + 0.0034048170782625675, + 0.003589311148971319, + 0.008046680130064487, + 0.05107264220714569, + 0.0027389100287109613, + 0.003397692460566759, + 0.03854282200336456, + 0.008501658216118813, + 0.002021583030000329, + 0.0017159049166366458, + 0.0020684131886810064, + 0.003208603709936142, + 0.0023107111919671297, + 0.0022955969907343388, + 0.003027094528079033, + 0.002078034682199359, + 0.00790442805737257, + 0.005704890470951796, + 0.011339349672198296, + 0.040165167301893234, + 0.012258249334990978, + 0.010365660302340984, + 0.027953265234827995, + 0.012063790112733841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.5762161016464233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019452712731435895 + ], + "opposite_logprob_tokens": [ + -6.251945495605469 + ], + "logprob_reward": -0.0019452712731435895, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019452712731435895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06714306771755219, + 0.1748853176832199, + 0.04110710322856903, + 0.02719913050532341, + 0.19386668503284454, + 0.23627324402332306, + 0.08901087194681168, + 0.03791196644306183, + 0.04726812243461609, + 0.08533451706171036 + ], + "credit_map": [ + 0.004775953944772482, + 0.006110989488661289, + 0.007346197497099638, + 0.011003546416759491, + 0.013177662156522274, + 0.008908561430871487, + 0.009322023950517178, + 0.006498132366687059, + 0.007928725332021713, + 0.011710230261087418, + 0.048888884484767914, + 0.019660944119095802, + 0.015414943918585777, + 0.02933633327484131, + 0.011914221569895744, + 0.030031032860279083, + 0.0022992442827671766, + 0.004652995616197586, + 0.004811710212379694, + 0.00715270172804594, + 0.004927224945276976, + 0.007619921118021011, + 0.004441364202648401, + 0.005201941356062889, + 0.0018320459639653563, + 0.0014641651650890708, + 0.003260448109358549, + 0.005581360310316086, + 0.0025418135337531567, + 0.004261274356395006, + 0.004468302242457867, + 0.003789722453802824, + 0.01319407019764185, + 0.019132548943161964, + 0.011746413074433804, + 0.01990646682679653, + 0.03165018558502197, + 0.06046421825885773, + 0.01655670814216137, + 0.021216068416833878, + 0.0068833124823868275, + 0.0107413399964571, + 0.09042912721633911, + 0.03792044520378113, + 0.01692228950560093, + 0.010492613539099693, + 0.05002802610397339, + 0.012856075540184975, + 0.007010295055806637, + 0.007043038494884968, + 0.008809121325612068, + 0.015868840739130974, + 0.011123906821012497, + 0.01318410225212574, + 0.012132218107581139, + 0.013839349150657654, + 0.0014403359964489937, + 0.002052515745162964, + 0.003983776550740004, + 0.014082462526857853, + 0.0023282459005713463, + 0.002272556070238352, + 0.006703271996229887, + 0.005048803053796291, + 0.0037896609865128994, + 0.004776520188897848, + 0.004356400575488806, + 0.010812738910317421, + 0.006325215566903353, + 0.006375281140208244, + 0.005678013898432255, + 0.005154290236532688, + 0.003619391703978181, + 0.0029596115928143263, + 0.004051729571074247, + 0.016220152378082275, + 0.004068466834723949, + 0.002701858524233103, + 0.03188423067331314, + 0.01982908323407173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.0019452712731435895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899098843336105 + ], + "opposite_logprob_tokens": [ + -2.578990936279297 + ], + "logprob_reward": -0.07899098843336105, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899098843336105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046051450073719025, + 0.07622909545898438, + 0.047775670886039734, + 0.044738732278347015, + 0.0336829349398613, + 0.5774112343788147, + 0.05359414592385292, + 0.01063761580735445, + 0.03364479914307594, + 0.07623431831598282 + ], + "credit_map": [ + 0.004242215305566788, + 0.004466509912163019, + 0.010064252652227879, + 0.010081754066050053, + 0.004794202279299498, + 0.004104327876120806, + 0.0045996978878974915, + 0.0036984914913773537, + 0.002863919595256448, + 0.0031788747292011976, + 0.022207722067832947, + 0.009965585544705391, + 0.008991614915430546, + 0.010016613639891148, + 0.005534091964364052, + 0.013470674864947796, + 0.0015303269028663635, + 0.0034688287414610386, + 0.0029792480636388063, + 0.009979485534131527, + 0.003543923841789365, + 0.003432023571804166, + 0.01629451848566532, + 0.006547317840158939, + 0.001327318837866187, + 0.00169886345975101, + 0.004114045761525631, + 0.016953347250819206, + 0.0009754924685694277, + 0.0018055055988952518, + 0.0034630163572728634, + 0.014401141554117203, + 0.0025630330201238394, + 0.0019659914541989565, + 0.004064559005200863, + 0.004698988515883684, + 0.005494644399732351, + 0.008054274134337902, + 0.0034496677108108997, + 0.003391775069758296, + 0.018572984263300896, + 0.026201637461781502, + 0.2229810655117035, + 0.0953235775232315, + 0.04689677804708481, + 0.02095809206366539, + 0.11679553985595703, + 0.029681559652090073, + 0.0022569471038877964, + 0.0045159426517784595, + 0.007283182814717293, + 0.013695626519620419, + 0.006141324527561665, + 0.0038323046173900366, + 0.007231018040329218, + 0.008637798018753529, + 0.0009563350467942655, + 0.0012020118301734328, + 0.0011049521854147315, + 0.0014909127494320273, + 0.0014065307332202792, + 0.0018607935635372996, + 0.0012331309262663126, + 0.0013829497620463371, + 0.002866478869691491, + 0.0029664174653589725, + 0.00359362899325788, + 0.005625855177640915, + 0.0042545306496322155, + 0.004632845055311918, + 0.005351954605430365, + 0.004353087395429611, + 0.0030849985778331757, + 0.004159201867878437, + 0.003834074828773737, + 0.00830439105629921, + 0.005903176032006741, + 0.003664074931293726, + 0.025447474792599678, + 0.02183692343533039 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 1 + ], + "selected_reward": -0.07899098843336105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014195182360708714 + ], + "opposite_logprob_tokens": [ + -4.264194965362549 + ], + "logprob_reward": -0.014195182360708714, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014195182360708714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13859538733959198, + 0.15044330060482025, + 0.19871214032173157, + 0.03916969150304794, + 0.049890488386154175, + 0.22812128067016602, + 0.03680623322725296, + 0.0281204991042614, + 0.037457771599292755, + 0.09268316626548767 + ], + "credit_map": [ + 0.009890798479318619, + 0.013566342182457447, + 0.013130252249538898, + 0.023181259632110596, + 0.014789914712309837, + 0.02943175658583641, + 0.02023709937930107, + 0.014367979019880295, + 0.006156862713396549, + 0.007535416632890701, + 0.038204070180654526, + 0.021739233285188675, + 0.02459539845585823, + 0.02774948813021183, + 0.009614625945687294, + 0.014848211780190468, + 0.012199685908854008, + 0.020415188744664192, + 0.018868574872612953, + 0.037774987518787384, + 0.028252115473151207, + 0.021048182621598244, + 0.030640576034784317, + 0.02951284497976303, + 0.0028693582862615585, + 0.002359293634071946, + 0.0032828529365360737, + 0.007242091465741396, + 0.005282561760395765, + 0.007072925567626953, + 0.003188234753906727, + 0.007872371934354305, + 0.003851523157209158, + 0.003222046187147498, + 0.007558919955044985, + 0.008056453429162502, + 0.006931806914508343, + 0.008375085890293121, + 0.005630843341350555, + 0.0062638139352202415, + 0.007695929147303104, + 0.009420151822268963, + 0.089312843978405, + 0.036910176277160645, + 0.01757679507136345, + 0.007156353443861008, + 0.048684705048799515, + 0.011364336125552654, + 0.0025905428919941187, + 0.0034307059831917286, + 0.004295077174901962, + 0.00737189082428813, + 0.0040010008960962296, + 0.005112731829285622, + 0.006023145280778408, + 0.003981136716902256, + 0.0018397979438304901, + 0.0022146599367260933, + 0.003212538780644536, + 0.0031611714512109756, + 0.0023282307665795088, + 0.0037467707879841328, + 0.0073701199144124985, + 0.004247208591550589, + 0.003906560596078634, + 0.004472251981496811, + 0.004369667265564203, + 0.004745046608150005, + 0.005507034715265036, + 0.004835246130824089, + 0.005196973215788603, + 0.004424990154802799, + 0.0026699707377701998, + 0.004102735780179501, + 0.00405638525262475, + 0.013511803932487965, + 0.006428793538361788, + 0.004114225506782532, + 0.03300100192427635, + 0.024798249825835228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.014195182360708714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning high-angle wide shot captures the scene, the camera panning slowly and steadily from left to right, revealing a seemingly boundless expanse of a ripe wheat field bathed in the brilliant midday sun. Dense golden stalks, laden with ripe grains, create a richly textured, undulating sea of amber and gold that stretches to the distant, hazy horizon. Powerful, unseen gusts of wind sweep across the land, choreographing the field into a mesmerizing dance; the stalks bend and rise in flowing, rippling waves, forming intricate, ever-shifting patterns of swirls and ephemeral spirals that chase each other across the surface. Above, the perfectly clear, azure sky contrasts starkly and beautifully with the warm tones below, its vastness magnifying the scale of the scene. The intense sunlight casts minimal shadows, causing the entire field to shimmer with a dazzling, breathtaking light, creating an atmosphere of profound peace, quiet majesty, and the raw, untamed beauty of nature in motion.", + "seed": 23253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_3/index_3137_seed_23253_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13171763974241912, + "vlm_logprob": -0.13171763974241912, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.13171763974241912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.035616934299468994 + ], + "opposite_logprob_tokens": [ + -6.285616874694824 + ], + "logprob_reward": -0.035616934299468994, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.035616934299468994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05764853581786156, + 0.2188333421945572, + 0.06409623473882675, + 0.04102610796689987, + 0.06641873717308044, + 0.03978988528251648, + 0.08415563404560089, + 0.04980522766709328, + 0.3323320150375366, + 0.04589430242776871 + ], + "credit_map": [ + 0.003582786535844207, + 0.004602181725203991, + 0.012949145399034023, + 0.007230969611555338, + 0.004575522616505623, + 0.004456569906324148, + 0.013066520914435387, + 0.007184841204434633, + 0.008494460955262184, + 0.011374382302165031, + 0.06951253861188889, + 0.08067222684621811, + 0.007959467358887196, + 0.012877646833658218, + 0.017801079899072647, + 0.010141532868146896, + 0.0031616948544979095, + 0.007059536874294281, + 0.006658671889454126, + 0.016854174435138702, + 0.013114210218191147, + 0.005156076978892088, + 0.007067066617310047, + 0.005024804268032312, + 0.003073075320571661, + 0.004209358710795641, + 0.004482893738895655, + 0.005676207132637501, + 0.006631551776081324, + 0.0065877740271389484, + 0.006240592338144779, + 0.004124654456973076, + 0.007411234546452761, + 0.007072790525853634, + 0.005714159458875656, + 0.00685441680252552, + 0.009016203694045544, + 0.011358592659235, + 0.007644610945135355, + 0.011346731334924698, + 0.002864426001906395, + 0.006777682341635227, + 0.00484171649441123, + 0.005503233056515455, + 0.0037832462694495916, + 0.005241534672677517, + 0.005437603686004877, + 0.00534044299274683, + 0.002925050677731633, + 0.007123854011297226, + 0.010233321227133274, + 0.005458596162497997, + 0.012127561494708061, + 0.006260210182517767, + 0.025805316865444183, + 0.014221725054085255, + 0.004462436772882938, + 0.004899355582892895, + 0.0043326374143362045, + 0.004721537232398987, + 0.007134057115763426, + 0.008360529318451881, + 0.010427296161651611, + 0.0054673790000379086, + 0.018702123314142227, + 0.03126432001590729, + 0.10314379632472992, + 0.08030644804239273, + 0.02859167568385601, + 0.03182980418205261, + 0.020451681688427925, + 0.018042167648673058, + 0.002306234324350953, + 0.001466818735934794, + 0.015973050147294998, + 0.006180491764098406, + 0.0016646169824525714, + 0.001736102276481688, + 0.003936283756047487, + 0.012630707584321499 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.035616934299468994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5761746764183044 + ], + "opposite_logprob_tokens": [ + -0.8261746764183044 + ], + "logprob_reward": -0.5761746764183044, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5761746764183044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09874778985977173, + 0.3050283193588257, + 0.20434266328811646, + 0.11309237033128738, + 0.04638870060443878, + 0.016973942518234253, + 0.0836111307144165, + 0.03597436845302582, + 0.05486484244465828, + 0.04097587615251541 + ], + "credit_map": [ + 0.004164400510489941, + 0.006507899612188339, + 0.02329055219888687, + 0.015265511348843575, + 0.007010384928435087, + 0.007948117330670357, + 0.023360123857855797, + 0.011200794018805027, + 0.011830401606857777, + 0.016324099153280258, + 0.09723323583602905, + 0.10998532176017761, + 0.017724791541695595, + 0.016368478536605835, + 0.022894175723195076, + 0.012667816132307053, + 0.007729780860245228, + 0.01625736989080906, + 0.020014474168419838, + 0.0713033527135849, + 0.03397784382104874, + 0.008722487837076187, + 0.032291118055582047, + 0.01404623594135046, + 0.008512797765433788, + 0.014916735701262951, + 0.00856065098196268, + 0.013289800845086575, + 0.020779576152563095, + 0.02301836758852005, + 0.015295826829969883, + 0.00871861819177866, + 0.004813438281416893, + 0.005939925089478493, + 0.0034484700299799442, + 0.003826908068731427, + 0.0076068369671702385, + 0.00903954915702343, + 0.004608308896422386, + 0.007105267606675625, + 0.0013843333581462502, + 0.0023434252943843603, + 0.00259615876711905, + 0.0023740106262266636, + 0.001760147511959076, + 0.002603093395009637, + 0.0020646797493100166, + 0.0018480928847566247, + 0.004944601096212864, + 0.010520415380597115, + 0.007560902740806341, + 0.004332316108047962, + 0.021332532167434692, + 0.008819704875349998, + 0.018220080062747, + 0.007880576886236668, + 0.003441820153966546, + 0.005007951520383358, + 0.003674487816169858, + 0.0037643280811607838, + 0.005157520994544029, + 0.004409335553646088, + 0.00670546805486083, + 0.0038134523201733828, + 0.005434453021734953, + 0.004294913727790117, + 0.007344560232013464, + 0.012629113160073757, + 0.012349310331046581, + 0.005115273874253035, + 0.0039479793049395084, + 0.0037492401897907257, + 0.002269592834636569, + 0.0023925371933728456, + 0.007652207277715206, + 0.004556980915367603, + 0.007182515226304531, + 0.005093745421618223, + 0.005490080453455448, + 0.006338216830044985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.5761746764183044 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005647733341902494 + ], + "opposite_logprob_tokens": [ + -7.5005645751953125 + ], + "logprob_reward": -0.0005647733341902494, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005647733341902494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12863369286060333, + 0.14917032420635223, + 0.20976896584033966, + 0.02394322119653225, + 0.05967757850885391, + 0.07070684432983398, + 0.1367815136909485, + 0.06198442727327347, + 0.08221304416656494, + 0.07712042331695557 + ], + "credit_map": [ + 0.006751348730176687, + 0.0072909193113446236, + 0.03220417723059654, + 0.017190616577863693, + 0.0109306201338768, + 0.01550639420747757, + 0.025654638186097145, + 0.013104974292218685, + 0.00770964240655303, + 0.007198285311460495, + 0.04114528000354767, + 0.055638331919908524, + 0.005456170532852411, + 0.007072110660374165, + 0.015699269250035286, + 0.009251223877072334, + 0.009864741005003452, + 0.010847899131476879, + 0.02415630593895912, + 0.08589247614145279, + 0.007483329623937607, + 0.00976432953029871, + 0.04515621438622475, + 0.01660364679992199, + 0.001512745046056807, + 0.004377072211354971, + 0.0026739987079054117, + 0.003921045921742916, + 0.0027124749030917883, + 0.003724197158589959, + 0.0029912274330854416, + 0.00203045760281384, + 0.006247757002711296, + 0.0066489302553236485, + 0.005312642548233271, + 0.0064626126550138, + 0.00856701098382473, + 0.009507904760539532, + 0.007339876610785723, + 0.00959083903580904, + 0.005731768440455198, + 0.010722417384386063, + 0.008610481396317482, + 0.01048216875642538, + 0.00640918780118227, + 0.009348971769213676, + 0.009643695317208767, + 0.00975815113633871, + 0.011140805669128895, + 0.023329833522439003, + 0.01189928688108921, + 0.010077692568302155, + 0.024022569879889488, + 0.017197955399751663, + 0.022290335968136787, + 0.01682303287088871, + 0.005408359691500664, + 0.00451311282813549, + 0.006441293749958277, + 0.0067903101444244385, + 0.008916571736335754, + 0.014079934917390347, + 0.008587148040533066, + 0.007247693371027708, + 0.00791719276458025, + 0.004525186028331518, + 0.018639322370290756, + 0.012890116311609745, + 0.020260322839021683, + 0.008618418127298355, + 0.00545039027929306, + 0.003912091720849276, + 0.004716932307928801, + 0.008981450460851192, + 0.016091367229819298, + 0.00992522295564413, + 0.01849108375608921, + 0.005946021992713213, + 0.006904940586537123, + 0.006063398439437151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 6 + ], + "selected_reward": -0.0005647733341902494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369763866066933 + ], + "opposite_logprob_tokens": [ + -3.4086976051330566 + ], + "logprob_reward": -0.03369763866066933, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369763866066933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039447080343961716, + 0.183014914393425, + 0.10536430031061172, + 0.169389009475708, + 0.0339677631855011, + 0.1163708046078682, + 0.13739554584026337, + 0.03512249514460564, + 0.15845297276973724, + 0.021475085988640785 + ], + "credit_map": [ + 0.0031082178466022015, + 0.0041880011558532715, + 0.0055529046803712845, + 0.008580319583415985, + 0.0026607103645801544, + 0.003933710511773825, + 0.005586967337876558, + 0.005836252123117447, + 0.008288362994790077, + 0.010269940830767155, + 0.04367891326546669, + 0.06486819684505463, + 0.013950105756521225, + 0.010913640260696411, + 0.02009744942188263, + 0.010948310606181622, + 0.012408795766532421, + 0.015181204304099083, + 0.007240078877657652, + 0.010082398541271687, + 0.03759988024830818, + 0.011367076076567173, + 0.007205295376479626, + 0.004279579035937786, + 0.0131495650857687, + 0.020083408802747726, + 0.015406275168061256, + 0.02195332571864128, + 0.028521478176116943, + 0.030737651512026787, + 0.024538230150938034, + 0.014999069273471832, + 0.0025858068838715553, + 0.005650780629366636, + 0.003006031271070242, + 0.004456012975424528, + 0.005664200056344271, + 0.004141770303249359, + 0.003831661306321621, + 0.004631501156836748, + 0.008657954633235931, + 0.009021748788654804, + 0.019031517207622528, + 0.018184421584010124, + 0.009870888665318489, + 0.01701858639717102, + 0.014336858876049519, + 0.020248832181096077, + 0.007393877021968365, + 0.016759542748332024, + 0.014543157070875168, + 0.010091706179082394, + 0.02468663454055786, + 0.014352252706885338, + 0.032557450234889984, + 0.017010927200317383, + 0.0034686739090830088, + 0.005632651504129171, + 0.004062558524310589, + 0.004436594899743795, + 0.00495557114481926, + 0.002613351447507739, + 0.005764078814536333, + 0.0041890135034918785, + 0.008679098449647427, + 0.0061380560509860516, + 0.04780261963605881, + 0.04289643466472626, + 0.02477966621518135, + 0.010612504556775093, + 0.010386046953499317, + 0.007158545311540365, + 0.0012174465227872133, + 0.0019908163230866194, + 0.004305096808820963, + 0.0030576251447200775, + 0.0020613777451217175, + 0.0015020507853478193, + 0.0033942381851375103, + 0.003946436103433371 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.03369763866066933 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene shot from a high-angle wide perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan slowly and steadily from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the environment a ripe wheat field with dense golden stalks under a clear azure sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the wheat stalks bend and rise in flowing, rippling waves due to wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting bright midday sunlight with minimal shadows causing the field to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012534175999462605 + ], + "opposite_logprob_tokens": [ + -4.387534141540527 + ], + "logprob_reward": -0.012534175999462605, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012534175999462605, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0654105618596077, + 0.09156902879476547, + 0.050791703164577484, + 0.04241224005818367, + 0.04132014885544777, + 0.04447005316615105, + 0.49958136677742004, + 0.010949634946882725, + 0.12916365265846252, + 0.024331597611308098 + ], + "credit_map": [ + 0.003845209488645196, + 0.004938831552863121, + 0.015233034268021584, + 0.008423927240073681, + 0.008295142091810703, + 0.006456382572650909, + 0.011963007040321827, + 0.006255023181438446, + 0.0033986433409154415, + 0.003955107182264328, + 0.029878297820687294, + 0.03522317111492157, + 0.0036107180640101433, + 0.0039108870550990105, + 0.007325439248234034, + 0.0042667631059885025, + 0.0042920587584376335, + 0.007571863941848278, + 0.0034552263095974922, + 0.004341269843280315, + 0.016443300992250443, + 0.007307957857847214, + 0.003313749562948942, + 0.004066276364028454, + 0.0032952961046248674, + 0.004597201012074947, + 0.004598639439791441, + 0.005749541334807873, + 0.006341102998703718, + 0.00760903162881732, + 0.006213171873241663, + 0.004008257295936346, + 0.003878029529005289, + 0.005675936117768288, + 0.0028465569484978914, + 0.0022770357318222523, + 0.009602073580026627, + 0.005609709769487381, + 0.0045659285970032215, + 0.006864878814667463, + 0.003945495001971722, + 0.008480516262352467, + 0.0043571023270487785, + 0.0036059359554201365, + 0.004794429987668991, + 0.009367079474031925, + 0.0034999018535017967, + 0.006419590674340725, + 0.024082379415631294, + 0.05252445489168167, + 0.0640900731086731, + 0.0325494185090065, + 0.03594769909977913, + 0.05290823429822922, + 0.15400612354278564, + 0.08347298949956894, + 0.0008295110892504454, + 0.0013907605316489935, + 0.0008930549374781549, + 0.0009907400235533714, + 0.0014514881186187267, + 0.0026715227868407965, + 0.001097347354516387, + 0.0016252101631835103, + 0.011359576135873795, + 0.007325379643589258, + 0.032376404851675034, + 0.016934216022491455, + 0.02822328731417656, + 0.016813861206173897, + 0.00819326937198639, + 0.007937662303447723, + 0.0011717895977199078, + 0.0012142907362431288, + 0.008111032657325268, + 0.0028997818008065224, + 0.0012520246673375368, + 0.0014583147130906582, + 0.0021907237824052572, + 0.006033638957887888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.012534175999462605 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_tensor([5960, 5960])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_tensor([5960, 5960])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2e7baa08bc7c117a148275e1074fed601d281246 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_tensor([5960, 5960])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot depicts a tense, silent confrontation in the hallowed quiet of a wood-paneled academic library. At a large, polished mahogany table, a young man with shaggy black hair leans forward, an impassioned look of frustration written across his face. Wearing a comfortable, dark green cardigan, they repeatedly gesture with their index finger to a tablet on the table. The tablet's screen emits a bright, cool digital glow, casting a stark blue light on their faces and hands, illuminating the complex data charts visible on the display. An older man seated opposite, dressed in a formal dark tweed blazer and a crisp white collared shirt, remains erect, his posture a wall of defiance. His silver-gray hair is impossibly neatly trimmed, his face a mask of resolute misunderstanding. He shakes his head slowly and stubbornly, his lips pressed into a thin line, his gaze deliberately averted, refusing even a glance at the glowing screen. The warm golden light from a classic green-shaded bank lamp on the table illuminates the scene, creating deep shadows that contrast with the cool glow of the tablet, visually emphasizing the generational and ideological clash. The background is filled with tall, dimly lit bookshelves filled with old leather-bound books, creating a heavy academic atmosphere, imbued with the weight of tradition and unspoken disagreements.", + "seed": 20021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_5960_seed_20021_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.498429636657238, + "vlm_logprob": -0.498429636657238, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.498429636657238, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.351696491241455 + ], + "opposite_logprob_tokens": [ + -0.10169650614261627 + ], + "logprob_reward": -2.351696491241455, + "yes_no_margin_reward": -2.25, + "token_credit_reward": -2.351696491241455, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2743343412876129, + 0.09165912866592407, + 0.12827548384666443, + 0.11598151177167892, + 0.0734831839799881, + 0.051826056092977524, + 0.07338600605726242, + 0.1145307794213295, + 0.023893989622592926, + 0.0526295006275177 + ], + "credit_map": [ + 0.026105929166078568, + 0.051262155175209045, + 0.030651897192001343, + 0.041377343237400055, + 0.02379956655204296, + 0.02615797519683838, + 0.030925016850233078, + 0.044054459780454636, + 0.008003121241927147, + 0.018029766157269478, + 0.008440635167062283, + 0.014227847568690777, + 0.009912683628499508, + 0.010483683086931705, + 0.00953796785324812, + 0.013023419305682182, + 0.009118946269154549, + 0.02681652083992958, + 0.010646757669746876, + 0.01773814670741558, + 0.017076199874281883, + 0.0131217110902071, + 0.014954691752791405, + 0.018802499398589134, + 0.00824940949678421, + 0.027081698179244995, + 0.009755999781191349, + 0.016913307830691338, + 0.017322493717074394, + 0.010768592357635498, + 0.01055473368614912, + 0.01533527486026287, + 0.007429727353155613, + 0.015295508317649364, + 0.006646098103374243, + 0.008532005362212658, + 0.011007496155798435, + 0.008921311236917973, + 0.005994086619466543, + 0.009656952694058418, + 0.004351833835244179, + 0.01079900935292244, + 0.004746900871396065, + 0.007372593507170677, + 0.00718632061034441, + 0.005098426714539528, + 0.005236840806901455, + 0.007034130394458771, + 0.006680588703602552, + 0.01203128881752491, + 0.006982262246310711, + 0.012191534973680973, + 0.006471249274909496, + 0.01025859359651804, + 0.008968835696578026, + 0.009801648557186127, + 0.009055441245436668, + 0.02233768254518509, + 0.012683464214205742, + 0.014962540939450264, + 0.017482351511716843, + 0.008991295471787453, + 0.013645359314978123, + 0.015372644178569317, + 0.003094549523666501, + 0.004007526207715273, + 0.0027315733022987843, + 0.0038671870715916157, + 0.002051878022029996, + 0.0020142202265560627, + 0.002887795679271221, + 0.003239257261157036, + 0.004715370945632458, + 0.011242943815886974, + 0.004980785306543112, + 0.006593497470021248, + 0.006956327706575394, + 0.0044188243336975574, + 0.006323674693703651, + 0.007398080080747604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -2.351696491241455 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04322199150919914 + ], + "opposite_logprob_tokens": [ + -3.168221950531006 + ], + "logprob_reward": -0.04322199150919914, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04322199150919914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3088172674179077, + 0.12267785519361496, + 0.11223538964986801, + 0.10134068131446838, + 0.08735065907239914, + 0.047729745507240295, + 0.07761187851428986, + 0.06505686789751053, + 0.02792643941938877, + 0.049253202974796295 + ], + "credit_map": [ + 0.025939103215932846, + 0.059447817504405975, + 0.04181033745408058, + 0.04140011966228485, + 0.030300576239824295, + 0.03358238935470581, + 0.02887258119881153, + 0.04746435955166817, + 0.011516505852341652, + 0.022280246019363403, + 0.013633335940539837, + 0.016782335937023163, + 0.01272317860275507, + 0.0105303218588233, + 0.016867462545633316, + 0.018344469368457794, + 0.00841815397143364, + 0.02271737903356552, + 0.010219600982964039, + 0.01785176992416382, + 0.014696209691464901, + 0.011346129700541496, + 0.010651824995875359, + 0.016334332525730133, + 0.00816604308784008, + 0.022665418684482574, + 0.008468622341752052, + 0.015596278011798859, + 0.014458628371357918, + 0.008658727630972862, + 0.009421944618225098, + 0.013905012980103493, + 0.008766759186983109, + 0.011770355515182018, + 0.010541300289332867, + 0.013190376572310925, + 0.007019159384071827, + 0.012553025968372822, + 0.010792560875415802, + 0.012717117555439472, + 0.005811740178614855, + 0.007990102283656597, + 0.005801666993647814, + 0.005632456392049789, + 0.004663190338760614, + 0.007047317456454039, + 0.004037391394376755, + 0.006745876744389534, + 0.006850927136838436, + 0.01415371336042881, + 0.006646746303886175, + 0.01395720150321722, + 0.008304426446557045, + 0.007480313535779715, + 0.009274729527533054, + 0.010943812318146229, + 0.005114310421049595, + 0.011441288515925407, + 0.007145389448851347, + 0.009036564268171787, + 0.0092525789514184, + 0.005349413026124239, + 0.008158504031598568, + 0.00955882016569376, + 0.003199617378413677, + 0.004009163472801447, + 0.003121504094451666, + 0.004633575212210417, + 0.0028909631073474884, + 0.002209288068115711, + 0.0038184854201972485, + 0.004043844528496265, + 0.006343536544591188, + 0.007128825876861811, + 0.005617763381451368, + 0.006195091642439365, + 0.005743870977312326, + 0.005447381641715765, + 0.0049830153584480286, + 0.00779371615499258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04322199150919914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371942415833473 + ], + "opposite_logprob_tokens": [ + -3.408719539642334 + ], + "logprob_reward": -0.03371942415833473, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371942415833473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2631221115589142, + 0.10386661440134048, + 0.11720944941043854, + 0.12843047082424164, + 0.10530778765678406, + 0.11197703331708908, + 0.05645988881587982, + 0.046021588146686554, + 0.028576606884598732, + 0.039028462022542953 + ], + "credit_map": [ + 0.027625635266304016, + 0.052036184817552567, + 0.03441125899553299, + 0.0307058934122324, + 0.021644555032253265, + 0.03287820145487785, + 0.024210281670093536, + 0.039610110223293304, + 0.011021839454770088, + 0.019015617668628693, + 0.011157291010022163, + 0.010137739591300488, + 0.012774752452969551, + 0.012521930038928986, + 0.011519214138388634, + 0.015718234702944756, + 0.008392736315727234, + 0.02362835593521595, + 0.01327959168702364, + 0.019045980647206306, + 0.012375729158520699, + 0.01158158853650093, + 0.012904477305710316, + 0.016000987961888313, + 0.010040432214736938, + 0.030920976772904396, + 0.011606075800955296, + 0.017321361228823662, + 0.01922793872654438, + 0.00983807910233736, + 0.011105470359325409, + 0.01837012730538845, + 0.010580713860690594, + 0.014391273260116577, + 0.01196028757840395, + 0.0168771930038929, + 0.009222263470292091, + 0.017167940735816956, + 0.01104908436536789, + 0.014059028588235378, + 0.011007935740053654, + 0.024762025102972984, + 0.012205233797430992, + 0.009638523682951927, + 0.01907598227262497, + 0.012440602295100689, + 0.008047599345445633, + 0.014799129217863083, + 0.004898535553365946, + 0.008369163610041142, + 0.005621110089123249, + 0.010768137872219086, + 0.006381351500749588, + 0.004080778453499079, + 0.008558926172554493, + 0.007781880907714367, + 0.0033493333030492067, + 0.007352169137448072, + 0.00459591718390584, + 0.0076649305410683155, + 0.0053365216590464115, + 0.004857887513935566, + 0.006124264560639858, + 0.0067405616864562035, + 0.0026482478715479374, + 0.005053110886365175, + 0.0033730727154761553, + 0.0037045057397335768, + 0.00447260495275259, + 0.0023897786159068346, + 0.003024424659088254, + 0.003910860978066921, + 0.0036620325408875942, + 0.007178041152656078, + 0.0048413630574941635, + 0.00496636051684618, + 0.005140967667102814, + 0.0044614155776798725, + 0.0033887077588588, + 0.0053895749151706696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.03371942415833473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370640054345131 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370640054345131, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370640054345131, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27288466691970825, + 0.1718219518661499, + 0.06602799147367477, + 0.08547482639551163, + 0.09155277907848358, + 0.10435891896486282, + 0.05708729103207588, + 0.07964253425598145, + 0.026381637901067734, + 0.04476749524474144 + ], + "credit_map": [ + 0.02814737893640995, + 0.04414515569806099, + 0.03663590922951698, + 0.036056552082300186, + 0.027958905324339867, + 0.03606700524687767, + 0.02420320175588131, + 0.03967052325606346, + 0.01920478045940399, + 0.027965430170297623, + 0.01868811808526516, + 0.025836387649178505, + 0.019282596185803413, + 0.01831364631652832, + 0.01944204419851303, + 0.023088928312063217, + 0.007445130031555891, + 0.01165056973695755, + 0.0070267789997160435, + 0.010774140246212482, + 0.00625850073993206, + 0.007827741093933582, + 0.006514937616884708, + 0.008530191145837307, + 0.0067624361254274845, + 0.019969157874584198, + 0.007130092941224575, + 0.012933912687003613, + 0.011201826855540276, + 0.0075005581602454185, + 0.007840319536626339, + 0.012136511504650116, + 0.008780676871538162, + 0.017616458237171173, + 0.008948073722422123, + 0.010675456374883652, + 0.014661421999335289, + 0.008969112299382687, + 0.010283427312970161, + 0.01161813922226429, + 0.010154863819479942, + 0.024332907050848007, + 0.010964654386043549, + 0.009373892098665237, + 0.01834026910364628, + 0.010630715638399124, + 0.0072496808134019375, + 0.013311924412846565, + 0.005005076061934233, + 0.009215395897626877, + 0.005314407404512167, + 0.010046151466667652, + 0.006547254975885153, + 0.005992255639284849, + 0.007387334015220404, + 0.0075794123113155365, + 0.0060028498992323875, + 0.014004068449139595, + 0.009349202737212181, + 0.011878439225256443, + 0.009653755463659763, + 0.006557148415595293, + 0.01036896277219057, + 0.01182809192687273, + 0.002480649622157216, + 0.004276425112038851, + 0.003281503915786743, + 0.004418229218572378, + 0.003022263292223215, + 0.002931641647592187, + 0.0030208381358534098, + 0.0029500825330615044, + 0.003519969992339611, + 0.006633793003857136, + 0.004431491252034903, + 0.007218539714813232, + 0.006100763101130724, + 0.003371849190443754, + 0.006498159375041723, + 0.00699292216449976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.03370640054345131 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980387583374977 + ], + "opposite_logprob_tokens": [ + -3.529803991317749 + ], + "logprob_reward": -0.02980387583374977, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980387583374977, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25149187445640564, + 0.11242016404867172, + 0.15103408694267273, + 0.09541956335306168, + 0.09375475347042084, + 0.03846723213791847, + 0.03454453498125076, + 0.06541111320257187, + 0.02982916124165058, + 0.12762749195098877 + ], + "credit_map": [ + 0.025742914527654648, + 0.039739515632390976, + 0.027532748878002167, + 0.04064924642443657, + 0.025171488523483276, + 0.02618877962231636, + 0.031533122062683105, + 0.03493403270840645, + 0.010321596637368202, + 0.020640138536691666, + 0.009845346212387085, + 0.0192374549806118, + 0.012552541680634022, + 0.012374001555144787, + 0.011608565226197243, + 0.015840521082282066, + 0.011662014760077, + 0.0340183787047863, + 0.014343315735459328, + 0.018878500908613205, + 0.019644470885396004, + 0.01568804495036602, + 0.015617660246789455, + 0.021181700751185417, + 0.007702457252889872, + 0.023267202079296112, + 0.008372606709599495, + 0.01247519999742508, + 0.013579421676695347, + 0.007710733916610479, + 0.008607030846178532, + 0.013704911805689335, + 0.008518517948687077, + 0.020565131679177284, + 0.008157244883477688, + 0.010780813172459602, + 0.014640516601502895, + 0.009715099819004536, + 0.009132484905421734, + 0.012244943529367447, + 0.0032451527658849955, + 0.007698786444962025, + 0.004163887817412615, + 0.006392378825694323, + 0.004837029613554478, + 0.0027455249801278114, + 0.004159422591328621, + 0.005225049331784248, + 0.0032147192396223545, + 0.005668682046234608, + 0.0035984443966299295, + 0.004525858908891678, + 0.00495496392250061, + 0.004499330651015043, + 0.003606615588068962, + 0.004475919995456934, + 0.005749243311583996, + 0.012754015624523163, + 0.007503969594836235, + 0.007975053042173386, + 0.009811468422412872, + 0.004904111381620169, + 0.007598160766065121, + 0.009115089662373066, + 0.0026861894875764847, + 0.005352679640054703, + 0.003223876003175974, + 0.0039634695276618, + 0.004633639007806778, + 0.002623197855427861, + 0.003555829171091318, + 0.003790280781686306, + 0.012351133860647678, + 0.020004473626613617, + 0.010947897098958492, + 0.020513631403446198, + 0.014484318904578686, + 0.012414136901497841, + 0.019241688773036003, + 0.017670216038823128 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 9 + ], + "selected_reward": -0.02980387583374977 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts a tense, silent confrontation in the hallowed quiet of a wood-paneled academic library. At a large, polished mahogany table, a young man with shaggy black hair leans forward, an impassioned look of frustration written across his face. Wearing a comfortable, dark green cardigan, they repeatedly gesture with their index finger to a tablet on the table. The tablet's screen emits a bright, cool digital glow, casting a stark blue light on their faces and hands, illuminating the complex data charts visible on the display. An older man seated opposite, dressed in a formal dark tweed blazer and a crisp white collared shirt, remains erect, his posture a wall of defiance. His silver-gray hair is impossibly neatly trimmed, his face a mask of resolute misunderstanding. He shakes his head slowly and stubbornly, his lips pressed into a thin line, his gaze deliberately averted, refusing even a glance at the glowing screen. The warm golden light from a classic green-shaded bank lamp on the table illuminates the scene, creating deep shadows that contrast with the cool glow of the tablet, visually emphasizing the generational and ideological clash. The background is filled with tall, dimly lit bookshelves filled with old leather-bound books, creating a heavy academic atmosphere, imbued with the weight of tradition and unspoken disagreements.", + "seed": 20022, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_5960_seed_20022_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22206272250041365, + "vlm_logprob": -0.22206272250041365, + "vlm_yes_no_margin": 3.450000011920929, + "vlm_token_credit": -0.22206272250041365, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0608091354370117 + ], + "opposite_logprob_tokens": [ + -0.4358091950416565 + ], + "logprob_reward": -1.0608091354370117, + "yes_no_margin_reward": -0.6249999403953552, + "token_credit_reward": -1.0608091354370117, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17042556405067444, + 0.05729945749044418, + 0.05737696588039398, + 0.5052191019058228, + 0.03354882821440697, + 0.03489755839109421, + 0.05638458579778671, + 0.02019430138170719, + 0.04416561499238014, + 0.020487993955612183 + ], + "credit_map": [ + 0.024328574538230896, + 0.03376154601573944, + 0.024302855134010315, + 0.013047657907009125, + 0.01164044439792633, + 0.01615539938211441, + 0.01652546599507332, + 0.03066362999379635, + 0.00887747947126627, + 0.009717660024762154, + 0.011031942442059517, + 0.003494231030344963, + 0.003889716463163495, + 0.004390731453895569, + 0.004667006898671389, + 0.011230695061385632, + 0.006384134758263826, + 0.006076608784496784, + 0.01419010665267706, + 0.0043876259587705135, + 0.0024981172755360603, + 0.003244029823690653, + 0.003830260830000043, + 0.01676608994603157, + 0.056080352514982224, + 0.011565813794732094, + 0.18808527290821075, + 0.01610080897808075, + 0.006851102691143751, + 0.017981944605708122, + 0.01154319103807211, + 0.19701063632965088, + 0.003681046422570944, + 0.003536627860739827, + 0.0067010801285505295, + 0.003239379031583667, + 0.0023951490875333548, + 0.0031163094099611044, + 0.00341385998763144, + 0.007465377449989319, + 0.004343423061072826, + 0.005401712842285633, + 0.005583616904914379, + 0.004294866230338812, + 0.003049037419259548, + 0.002945537446066737, + 0.003977284766733646, + 0.005302080884575844, + 0.006079988554120064, + 0.006673273630440235, + 0.013357006013393402, + 0.004611559212207794, + 0.003923892509192228, + 0.00383130949921906, + 0.004414406605064869, + 0.013493156060576439, + 0.0024147899821400642, + 0.003101966343820095, + 0.0032386393286287785, + 0.0019095526076853275, + 0.0019653087947517633, + 0.0022014754358679056, + 0.0020451396703720093, + 0.0033174322452396154, + 0.004495086148381233, + 0.005855122581124306, + 0.009024931117892265, + 0.005219209007918835, + 0.0032153381034731865, + 0.0030578135047107935, + 0.004559224937111139, + 0.008738894946873188, + 0.002053650561720133, + 0.002572579076513648, + 0.004470973275601864, + 0.0019166244892403483, + 0.0017311868723481894, + 0.001539207180030644, + 0.0016471127746626735, + 0.00455666147172451 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -1.0608091354370117 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011117654852569103 + ], + "opposite_logprob_tokens": [ + -4.511117458343506 + ], + "logprob_reward": -0.011117654852569103, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011117654852569103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16613921523094177, + 0.08383515477180481, + 0.0500301793217659, + 0.5291110277175903, + 0.028348082676529884, + 0.03209078684449196, + 0.04779709875583649, + 0.013702835887670517, + 0.02619946002960205, + 0.02274618111550808 + ], + "credit_map": [ + 0.020031485706567764, + 0.017699604853987694, + 0.03794250637292862, + 0.007565604057163, + 0.00913275871425867, + 0.015825020149350166, + 0.01277377549558878, + 0.045168448239564896, + 0.009276799857616425, + 0.005965515039861202, + 0.020726842805743217, + 0.0061102635227143764, + 0.004435574170202017, + 0.008305305615067482, + 0.004584831185638905, + 0.024430016055703163, + 0.005282674916088581, + 0.0027269944548606873, + 0.011964365839958191, + 0.004352902993559837, + 0.003873215289786458, + 0.005348727107048035, + 0.003797316923737526, + 0.012683974578976631, + 0.058602094650268555, + 0.012160726822912693, + 0.19707831740379333, + 0.017117047682404518, + 0.007177244406193495, + 0.018253175541758537, + 0.012365199625492096, + 0.2063572108745575, + 0.003115719649940729, + 0.0033895878586918116, + 0.004620578605681658, + 0.002937384182587266, + 0.0025396489072591066, + 0.0031359808053821325, + 0.003657599678263068, + 0.004951581358909607, + 0.0034552342258393764, + 0.0036293254233896732, + 0.007764006964862347, + 0.002425751183182001, + 0.002188154263421893, + 0.0024958420544862747, + 0.0026633245870471, + 0.007469143718481064, + 0.004923579283058643, + 0.004494726192206144, + 0.013709694147109985, + 0.0027855588123202324, + 0.0026243303436785936, + 0.002651760121807456, + 0.0026704955380409956, + 0.013936949893832207, + 0.0017591404030099511, + 0.002127822022885084, + 0.0023274098057299852, + 0.0011617555283010006, + 0.0012586073717102408, + 0.0017412715824320912, + 0.0012433481169864535, + 0.0020834794268012047, + 0.0026848556008189917, + 0.0037232364993542433, + 0.005024690646678209, + 0.0030573203694075346, + 0.0023715191055089235, + 0.001990512479096651, + 0.0026747207157313824, + 0.004672605078667402, + 0.002211453625932336, + 0.002376049757003784, + 0.005753844045102596, + 0.001500714453868568, + 0.001863787416368723, + 0.0017056689830496907, + 0.0017994822701439261, + 0.005535177420824766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.011117654852569103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006738320458680391 + ], + "opposite_logprob_tokens": [ + -5.006738185882568 + ], + "logprob_reward": -0.006738320458680391, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006738320458680391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12223245948553085, + 0.08699886500835419, + 0.05080125853419304, + 0.5992474555969238, + 0.017262408509850502, + 0.027726922184228897, + 0.04088358208537102, + 0.01585816591978073, + 0.026940546929836273, + 0.012048373930156231 + ], + "credit_map": [ + 0.016012076288461685, + 0.012225938960909843, + 0.02635384537279606, + 0.004866160452365875, + 0.010071596130728722, + 0.016294650733470917, + 0.010059718042612076, + 0.026348477229475975, + 0.009256631135940552, + 0.0042703780345618725, + 0.023172007873654366, + 0.004305416718125343, + 0.005543499253690243, + 0.010401112958788872, + 0.004291779361665249, + 0.025758039206266403, + 0.005411958787590265, + 0.002531904261559248, + 0.013003995642066002, + 0.003160363296046853, + 0.0037543370854109526, + 0.0061723156832158566, + 0.0033205761574208736, + 0.013445807620882988, + 0.0668342337012291, + 0.013077514246106148, + 0.2231827676296234, + 0.01923549920320511, + 0.00651892414316535, + 0.02042405866086483, + 0.013571392744779587, + 0.23640306293964386, + 0.002130538458004594, + 0.0024186281953006983, + 0.0025482235942035913, + 0.002213804516941309, + 0.001743484172038734, + 0.0019986641127616167, + 0.0019144726684316993, + 0.002294591860845685, + 0.0029548793099820614, + 0.003073536790907383, + 0.0068019297905266285, + 0.002187673468142748, + 0.001982208341360092, + 0.002264442155137658, + 0.002131109591573477, + 0.0063311439007520676, + 0.004257530439645052, + 0.003530289279296994, + 0.012146652676165104, + 0.0022458184976130724, + 0.0019551704172044992, + 0.002130204811692238, + 0.002118978649377823, + 0.012498936615884304, + 0.0019174751359969378, + 0.002168903825804591, + 0.003144357353448868, + 0.001519239624030888, + 0.0012337075313553214, + 0.0015742076793685555, + 0.001477423356845975, + 0.0028228515293449163, + 0.0027065251488238573, + 0.0030342149548232555, + 0.006597890518605709, + 0.0025843032635748386, + 0.0018948712386190891, + 0.0017198604764416814, + 0.0021209914702922106, + 0.006281890440732241, + 0.0015067203203216195, + 0.0018643451621755958, + 0.0016707985196262598, + 0.001422740751877427, + 0.0014334529405459762, + 0.0012465318432077765, + 0.001433918485417962, + 0.0014698660233989358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.006738320458680391 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020576000213623047 + ], + "opposite_logprob_tokens": [ + -3.895576000213623 + ], + "logprob_reward": -0.020576000213623047, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020576000213623047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1336490511894226, + 0.28605714440345764, + 0.18015970289707184, + 0.20525209605693817, + 0.07494731992483139, + 0.04029101878404617, + 0.024226976558566093, + 0.018897011876106262, + 0.01576140522956848, + 0.02075822837650776 + ], + "credit_map": [ + 0.012767232954502106, + 0.010632483288645744, + 0.03325239196419716, + 0.008088615722954273, + 0.006516972556710243, + 0.01129898615181446, + 0.009828152135014534, + 0.041264209896326065, + 0.029551487416028976, + 0.013039255514740944, + 0.08539670705795288, + 0.010989268310368061, + 0.01144692488014698, + 0.026243725791573524, + 0.014092463068664074, + 0.09529731422662735, + 0.01920982636511326, + 0.009273887611925602, + 0.04923298582434654, + 0.012822102755308151, + 0.009914393536746502, + 0.01538320817053318, + 0.011002479121088982, + 0.053320836275815964, + 0.0225499477237463, + 0.005378242116421461, + 0.07549300044775009, + 0.006662216503173113, + 0.0035540894605219364, + 0.008639423176646233, + 0.005070415325462818, + 0.07790476083755493, + 0.007850496098399162, + 0.005024882964789867, + 0.01687478832900524, + 0.00709161814302206, + 0.0053927400149405, + 0.01025147270411253, + 0.0075647952035069466, + 0.014896524138748646, + 0.0044353739358484745, + 0.0046277884393930435, + 0.009034475311636925, + 0.004117821808904409, + 0.002884349785745144, + 0.0033162999898195267, + 0.004119356628507376, + 0.007755549158900976, + 0.002642752369865775, + 0.0025862648617476225, + 0.00495232455432415, + 0.002223061863332987, + 0.0019550553988665342, + 0.0027724672108888626, + 0.002250334480777383, + 0.004844714421778917, + 0.0023210768122226, + 0.0018173456192016602, + 0.003245583036914468, + 0.0014198608696460724, + 0.0017749785911291838, + 0.0031781375873833895, + 0.0014061155961826444, + 0.003733913181349635, + 0.00186064257286489, + 0.0024958651047199965, + 0.0023693786934018135, + 0.0019538209307938814, + 0.0016291955253109336, + 0.001686488976702094, + 0.001714934827759862, + 0.0020510784815996885, + 0.0026246937923133373, + 0.0025577733758836985, + 0.002499696798622608, + 0.0023496083449572325, + 0.0022567021660506725, + 0.0032433061860501766, + 0.0025035329163074493, + 0.002722913632169366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.020576000213623047 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072501540184021 + ], + "opposite_logprob_tokens": [ + -4.511072635650635 + ], + "logprob_reward": -0.011072501540184021, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072501540184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2035883069038391, + 0.17877168953418732, + 0.16869467496871948, + 0.12083877623081207, + 0.09844077378511429, + 0.03878851234912872, + 0.05080471560359001, + 0.036833830177783966, + 0.06385824829339981, + 0.03938049077987671 + ], + "credit_map": [ + 0.030014727264642715, + 0.024268409237265587, + 0.034369293600320816, + 0.016944682225584984, + 0.015696652233600616, + 0.02058197185397148, + 0.020642699673771858, + 0.041069868952035904, + 0.021347684785723686, + 0.014890631660819054, + 0.04158603399991989, + 0.01386498287320137, + 0.011235406622290611, + 0.017796579748392105, + 0.010582567192614079, + 0.04746780917048454, + 0.019954130053520203, + 0.009742619469761848, + 0.05362356826663017, + 0.005646567791700363, + 0.004815909545868635, + 0.008489797823131084, + 0.0053350129164755344, + 0.061087071895599365, + 0.013036931864917278, + 0.00803403276950121, + 0.034575119614601135, + 0.004994583781808615, + 0.006340430583804846, + 0.00803982000797987, + 0.005360426381230354, + 0.040457434952259064, + 0.010601140558719635, + 0.006159849464893341, + 0.031240757554769516, + 0.0042718322947621346, + 0.004549096804112196, + 0.0069991047494113445, + 0.004339063074439764, + 0.03027992881834507, + 0.004986679647117853, + 0.005911995656788349, + 0.0054730684496462345, + 0.004536591935902834, + 0.00328752095811069, + 0.003905209479853511, + 0.005078721791505814, + 0.005608722567558289, + 0.005664251744747162, + 0.005028405226767063, + 0.012617472559213638, + 0.003761433996260166, + 0.0029993478674441576, + 0.0037919103633612394, + 0.0039253029972314835, + 0.0130165945738554, + 0.004121848382055759, + 0.003690019017085433, + 0.008444053120911121, + 0.0032156147062778473, + 0.0030516632832586765, + 0.002658059587702155, + 0.0034127964172512293, + 0.00823977217078209, + 0.006606625858694315, + 0.006993787828832865, + 0.015929650515317917, + 0.004651102237403393, + 0.003954481333494186, + 0.004289066884666681, + 0.004378504119813442, + 0.017055025324225426, + 0.004918512888252735, + 0.004481919575482607, + 0.008757052011787891, + 0.004955913871526718, + 0.002442257711663842, + 0.0032170831691473722, + 0.004531359765678644, + 0.006076393648982048 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011072501540184021 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts a tense, silent confrontation in the hallowed quiet of a wood-paneled academic library. At a large, polished mahogany table, a young man with shaggy black hair leans forward, an impassioned look of frustration written across his face. Wearing a comfortable, dark green cardigan, they repeatedly gesture with their index finger to a tablet on the table. The tablet's screen emits a bright, cool digital glow, casting a stark blue light on their faces and hands, illuminating the complex data charts visible on the display. An older man seated opposite, dressed in a formal dark tweed blazer and a crisp white collared shirt, remains erect, his posture a wall of defiance. His silver-gray hair is impossibly neatly trimmed, his face a mask of resolute misunderstanding. He shakes his head slowly and stubbornly, his lips pressed into a thin line, his gaze deliberately averted, refusing even a glance at the glowing screen. The warm golden light from a classic green-shaded bank lamp on the table illuminates the scene, creating deep shadows that contrast with the cool glow of the tablet, visually emphasizing the generational and ideological clash. The background is filled with tall, dimly lit bookshelves filled with old leather-bound books, creating a heavy academic atmosphere, imbued with the weight of tradition and unspoken disagreements.", + "seed": 20023, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_5960_seed_20023_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1432931823655963, + "vlm_logprob": -0.1432931823655963, + "vlm_yes_no_margin": 2.925, + "vlm_token_credit": -0.1432931823655963, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5812236070632935 + ], + "opposite_logprob_tokens": [ + -0.8312236070632935 + ], + "logprob_reward": -0.5812236070632935, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5812236070632935, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38282397389411926, + 0.08601918816566467, + 0.041366882622241974, + 0.046014562249183655, + 0.09684351086616516, + 0.027174757793545723, + 0.01583104580640793, + 0.013642944395542145, + 0.2646019458770752, + 0.025681160390377045 + ], + "credit_map": [ + 0.06506994366645813, + 0.026317274197936058, + 0.05558529123663902, + 0.05965600162744522, + 0.0370631068944931, + 0.03801315650343895, + 0.04810737818479538, + 0.05301183834671974, + 0.01608634740114212, + 0.006125365849584341, + 0.012238284572958946, + 0.009632287546992302, + 0.014231176115572453, + 0.010123147629201412, + 0.007563415914773941, + 0.010019171051681042, + 0.005753081291913986, + 0.002383441897109151, + 0.006273012608289719, + 0.0052381716668605804, + 0.009275387041270733, + 0.003666558302938938, + 0.0031769289635121822, + 0.005600304808467627, + 0.007030050735920668, + 0.004032447934150696, + 0.006450805813074112, + 0.004709081258624792, + 0.0074805449694395065, + 0.004359760787338018, + 0.004214279353618622, + 0.00773759838193655, + 0.02014801651239395, + 0.007997875101864338, + 0.013069961220026016, + 0.011665352620184422, + 0.018526585772633553, + 0.007696050684899092, + 0.007431259378790855, + 0.010308419354259968, + 0.004265202675014734, + 0.0016646870644763112, + 0.004429349210113287, + 0.0019264466827735305, + 0.004977465141564608, + 0.003261381993070245, + 0.001527524902485311, + 0.005122699309140444, + 0.002537439577281475, + 0.001016118098050356, + 0.0022205128334462643, + 0.0014358871849253774, + 0.0037704980932176113, + 0.001618708367459476, + 0.001221771934069693, + 0.0020101105328649282, + 0.002102619968354702, + 0.0007329013315029442, + 0.0024042646400630474, + 0.0007314807735383511, + 0.0033140110317617655, + 0.0012611083220690489, + 0.0009892518864944577, + 0.002107307780534029, + 0.018249357119202614, + 0.004891743417829275, + 0.10051155090332031, + 0.0066817887127399445, + 0.014866519719362259, + 0.011963184922933578, + 0.008038920350372791, + 0.09939888119697571, + 0.0037825724575668573, + 0.0018963709007948637, + 0.0040694307535886765, + 0.0017721395706757903, + 0.005926736164838076, + 0.0023545455187559128, + 0.002024616114795208, + 0.0038547494914382696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.5812236070632935 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03816339746117592 + ], + "opposite_logprob_tokens": [ + -3.288163423538208 + ], + "logprob_reward": -0.03816339746117592, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03816339746117592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2213132530450821, + 0.3048723042011261, + 0.036431510001420975, + 0.04743250459432602, + 0.03496794030070305, + 0.014112069271504879, + 0.0358387753367424, + 0.04927239939570427, + 0.13734565675258636, + 0.1184135377407074 + ], + "credit_map": [ + 0.03391697257757187, + 0.011283444240689278, + 0.028859345242381096, + 0.03265353664755821, + 0.02283572405576706, + 0.029924100264906883, + 0.032219599932432175, + 0.029620515182614326, + 0.03472701460123062, + 0.020382795482873917, + 0.040456436574459076, + 0.019517822191119194, + 0.10868845880031586, + 0.024272006005048752, + 0.024994071573019028, + 0.03183367848396301, + 0.004324419889599085, + 0.0015730143059045076, + 0.00473833829164505, + 0.0033193621784448624, + 0.007027040235698223, + 0.005243456922471523, + 0.0026284886989742517, + 0.007577387150377035, + 0.006810893304646015, + 0.003956260159611702, + 0.006214847322553396, + 0.0038942103274166584, + 0.011132837273180485, + 0.004687853157520294, + 0.003281397745013237, + 0.007454208564013243, + 0.006306502502411604, + 0.0016500592464581132, + 0.00477656489238143, + 0.0048343283124268055, + 0.004858517087996006, + 0.0040325382724404335, + 0.0032285128254443407, + 0.005280915182083845, + 0.0019082362996414304, + 0.0007891052518971264, + 0.001973578706383705, + 0.001152646611444652, + 0.0022434876300394535, + 0.002641685539856553, + 0.0007939471397548914, + 0.0026093819178640842, + 0.0045759910717606544, + 0.0016410283278673887, + 0.0049454523250460625, + 0.002338489517569542, + 0.011769144795835018, + 0.00274896202608943, + 0.0027212966233491898, + 0.00509840901941061, + 0.005378969945013523, + 0.002262323396280408, + 0.009046862833201885, + 0.0018183733336627483, + 0.015280669555068016, + 0.0043548839166760445, + 0.003259430406615138, + 0.007870885543525219, + 0.009345646016299725, + 0.0031968047842383385, + 0.04836897552013397, + 0.0042557804845273495, + 0.013296874240040779, + 0.007127006072551012, + 0.004114151932299137, + 0.047640420496463776, + 0.01495614368468523, + 0.006537414621561766, + 0.0223713181912899, + 0.007483202964067459, + 0.03203791752457619, + 0.0075042820535600185, + 0.01124006137251854, + 0.016283197328448296 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.03816339746117592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06207539886236191 + ], + "opposite_logprob_tokens": [ + -2.81207537651062 + ], + "logprob_reward": -0.06207539886236191, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06207539886236191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3725554943084717, + 0.10045809298753738, + 0.03516591340303421, + 0.06139230728149414, + 0.027325743809342384, + 0.010602021589875221, + 0.017432980239391327, + 0.02374030463397503, + 0.2594728171825409, + 0.09185433387756348 + ], + "credit_map": [ + 0.06368381530046463, + 0.02011267840862274, + 0.051151737570762634, + 0.04477408900856972, + 0.03001990169286728, + 0.0665954202413559, + 0.045994117856025696, + 0.050223711878061295, + 0.01191891822963953, + 0.006907575763761997, + 0.012660057283937931, + 0.006286881398409605, + 0.03366648033261299, + 0.009748979471623898, + 0.0077107809484004974, + 0.011558416299521923, + 0.004041606560349464, + 0.0023509713355451822, + 0.004491021856665611, + 0.004455979913473129, + 0.006658267695456743, + 0.004198448266834021, + 0.0032146091107279062, + 0.005755005870014429, + 0.008609508164227009, + 0.004644473548978567, + 0.009970646351575851, + 0.005212218966335058, + 0.01236035581678152, + 0.006458132062107325, + 0.004926670342683792, + 0.009210305288434029, + 0.006162246689200401, + 0.0015675947070121765, + 0.004226910416036844, + 0.0015549476956948638, + 0.0033030530903488398, + 0.0036388549488037825, + 0.0014348954427987337, + 0.00543724000453949, + 0.0014411743031814694, + 0.0008044003625400364, + 0.001737073645927012, + 0.001249494613148272, + 0.001045559998601675, + 0.0013854631688445807, + 0.001030883751809597, + 0.0019079719204455614, + 0.0030919217970222235, + 0.0008696477161720395, + 0.0023287315852940083, + 0.0012925732880830765, + 0.004076710902154446, + 0.0017832317389547825, + 0.0013689674669876695, + 0.0026211943477392197, + 0.002686926396563649, + 0.0009859660640358925, + 0.004062694031745195, + 0.00121011002920568, + 0.006046642549335957, + 0.0032847453840076923, + 0.0015232987934723496, + 0.003939921502023935, + 0.018498964607715607, + 0.005020814947783947, + 0.09669331461191177, + 0.006597056053578854, + 0.014266916550695896, + 0.013264978304505348, + 0.008358413353562355, + 0.09677238762378693, + 0.010752305388450623, + 0.00660218670964241, + 0.017771262675523758, + 0.007278480101376772, + 0.019806096330285072, + 0.009770495817065239, + 0.008020836859941483, + 0.011852666735649109 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.06207539886236191 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008644540794193745 + ], + "opposite_logprob_tokens": [ + -4.7586445808410645 + ], + "logprob_reward": -0.008644540794193745, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008644540794193745, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17147769033908844, + 0.09917300939559937, + 0.053368426859378815, + 0.0843416154384613, + 0.03703612461686134, + 0.02125413715839386, + 0.033274777233600616, + 0.014793487265706062, + 0.40648457407951355, + 0.07879623025655746 + ], + "credit_map": [ + 0.023748070001602173, + 0.010717418976128101, + 0.023604294285178185, + 0.030267175287008286, + 0.02034268155694008, + 0.01687294989824295, + 0.02382466569542885, + 0.02210042253136635, + 0.010570550337433815, + 0.006847536191344261, + 0.014019681140780449, + 0.008902373723685741, + 0.029603438451886177, + 0.008920030668377876, + 0.009027190506458282, + 0.011282198131084442, + 0.007835649885237217, + 0.0036418118979781866, + 0.0075057437643408775, + 0.007185920607298613, + 0.008057421073317528, + 0.006163241341710091, + 0.00407389784231782, + 0.008904736489057541, + 0.010898091830313206, + 0.005653770640492439, + 0.012584947980940342, + 0.009989812970161438, + 0.011541471816599369, + 0.0076260934583842754, + 0.007958314381539822, + 0.018089106306433678, + 0.005129452329128981, + 0.0026804974768310785, + 0.004781158175319433, + 0.004883543122559786, + 0.008859941735863686, + 0.0033886281307786703, + 0.0031543781515210867, + 0.004158518742769957, + 0.0027293043676763773, + 0.0014884751290082932, + 0.003126592142507434, + 0.003842240432277322, + 0.0029159625992178917, + 0.0020035968627780676, + 0.002092247363179922, + 0.0030557166319340467, + 0.006754867732524872, + 0.002519054338335991, + 0.00461347121745348, + 0.004865827038884163, + 0.004357593134045601, + 0.003001091070473194, + 0.003057704074308276, + 0.004105161875486374, + 0.0029574406798928976, + 0.0011249178787693381, + 0.002169156912714243, + 0.001492488430812955, + 0.0018930989317595959, + 0.0016397954896092415, + 0.0013733969535678625, + 0.002143190475180745, + 0.03007609024643898, + 0.010617814026772976, + 0.14541012048721313, + 0.016262054443359375, + 0.021929919719696045, + 0.021724145859479904, + 0.017914535477757454, + 0.14254984259605408, + 0.014860926195979118, + 0.004445015452802181, + 0.014317403547465801, + 0.00690100621432066, + 0.012525125406682491, + 0.00686210161074996, + 0.0052354298532009125, + 0.01364921685308218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.008644540794193745 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026358967646956444 + ], + "opposite_logprob_tokens": [ + -3.6513590812683105 + ], + "logprob_reward": -0.026358967646956444, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026358967646956444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14936645328998566, + 0.19738192856311798, + 0.11439736187458038, + 0.04153053089976311, + 0.09385589510202408, + 0.04059896245598793, + 0.03452500328421593, + 0.02660481445491314, + 0.2565859258174896, + 0.04515307769179344 + ], + "credit_map": [ + 0.023870723322033882, + 0.012948278337717056, + 0.019818510860204697, + 0.013196567073464394, + 0.017150085419416428, + 0.02301819622516632, + 0.01102391816675663, + 0.02834019437432289, + 0.04181760922074318, + 0.010998108424246311, + 0.02662261202931404, + 0.026371102780103683, + 0.025992538779973984, + 0.016937633976340294, + 0.019986942410469055, + 0.028655407950282097, + 0.014504487626254559, + 0.0085940882563591, + 0.015881327912211418, + 0.017498444765806198, + 0.02640872448682785, + 0.008232527412474155, + 0.010866669937968254, + 0.012411106377840042, + 0.004195481073111296, + 0.0026956370566040277, + 0.006546809338033199, + 0.003457897575572133, + 0.009576776064932346, + 0.004495625849813223, + 0.002804878866299987, + 0.00775742856785655, + 0.011790448799729347, + 0.0046689133159816265, + 0.012805520556867123, + 0.013420031405985355, + 0.024949811398983, + 0.007222665473818779, + 0.008501502685248852, + 0.010497009381651878, + 0.008007976226508617, + 0.002679298399016261, + 0.0061997659504413605, + 0.005056009162217379, + 0.0061432295478880405, + 0.003993328660726547, + 0.0033863696735352278, + 0.005132989026606083, + 0.0050571211613714695, + 0.0015308865113183856, + 0.00459505757316947, + 0.0024693619925528765, + 0.009341628290712833, + 0.003390131052583456, + 0.0028416935820132494, + 0.005299130454659462, + 0.004242622759193182, + 0.0015277067432180047, + 0.004396604839712381, + 0.0015094721456989646, + 0.007813324220478535, + 0.001740990555845201, + 0.0018685926916077733, + 0.0035055025946348906, + 0.023164067417383194, + 0.006400241982191801, + 0.08533377200365067, + 0.00645498838275671, + 0.031710926443338394, + 0.012446083128452301, + 0.009564129635691643, + 0.08151175081729889, + 0.007625165861099958, + 0.0033222974743694067, + 0.007114998530596495, + 0.002856695558875799, + 0.009493492543697357, + 0.0045243543572723866, + 0.0037004973273724318, + 0.0065155792981386185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.026358967646956444 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot depicts a tense, silent confrontation in the hallowed quiet of a wood-paneled academic library. At a large, polished mahogany table, a young man with shaggy black hair leans forward, an impassioned look of frustration written across his face. Wearing a comfortable, dark green cardigan, they repeatedly gesture with their index finger to a tablet on the table. The tablet's screen emits a bright, cool digital glow, casting a stark blue light on their faces and hands, illuminating the complex data charts visible on the display. An older man seated opposite, dressed in a formal dark tweed blazer and a crisp white collared shirt, remains erect, his posture a wall of defiance. His silver-gray hair is impossibly neatly trimmed, his face a mask of resolute misunderstanding. He shakes his head slowly and stubbornly, his lips pressed into a thin line, his gaze deliberately averted, refusing even a glance at the glowing screen. The warm golden light from a classic green-shaded bank lamp on the table illuminates the scene, creating deep shadows that contrast with the cool glow of the tablet, visually emphasizing the generational and ideological clash. The background is filled with tall, dimly lit bookshelves filled with old leather-bound books, creating a heavy academic atmosphere, imbued with the weight of tradition and unspoken disagreements.", + "seed": 20024, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_4/index_5960_seed_20024_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.18358870409429073, + "vlm_logprob": -0.18358870409429073, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.18358870409429073, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8335974812507629 + ], + "opposite_logprob_tokens": [ + -0.5835974812507629 + ], + "logprob_reward": -0.8335974812507629, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8335974812507629, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2898160219192505, + 0.15481247007846832, + 0.05871519818902016, + 0.060769062489271164, + 0.07136312127113342, + 0.04375139996409416, + 0.14381973445415497, + 0.13008348643779755, + 0.022442007437348366, + 0.02442748099565506 + ], + "credit_map": [ + 0.033847689628601074, + 0.03690159320831299, + 0.0351962074637413, + 0.028920551761984825, + 0.04671840742230415, + 0.04202847182750702, + 0.02853112667798996, + 0.037671979516744614, + 0.014263718388974667, + 0.022193966433405876, + 0.019016509875655174, + 0.016125407069921494, + 0.023358793929219246, + 0.022548353299498558, + 0.015762394294142723, + 0.02154332399368286, + 0.007445028051733971, + 0.007191648241132498, + 0.007877012714743614, + 0.0064031630754470825, + 0.009259690530598164, + 0.007584053091704845, + 0.006078708916902542, + 0.00687589542940259, + 0.005967950448393822, + 0.00733608053997159, + 0.010817046277225018, + 0.009210659191012383, + 0.005647697020322084, + 0.008717617020010948, + 0.0042117624543607235, + 0.008860252797603607, + 0.0071127102710306644, + 0.007997029460966587, + 0.007936524227261543, + 0.005412436090409756, + 0.01592751033604145, + 0.011505048722028732, + 0.006174264941364527, + 0.009297601878643036, + 0.004198738839477301, + 0.0047919610515236855, + 0.006449440494179726, + 0.003944651689380407, + 0.006767410319298506, + 0.007218230050057173, + 0.0032156086526811123, + 0.007165360730141401, + 0.016010606661438942, + 0.013566908426582813, + 0.025188138708472252, + 0.020033501088619232, + 0.013403883203864098, + 0.022680027410387993, + 0.016527440398931503, + 0.016409220173954964, + 0.01445339247584343, + 0.015065263956785202, + 0.019912809133529663, + 0.01862231083214283, + 0.012716489844024181, + 0.01933533325791359, + 0.015366421081125736, + 0.014611458405852318, + 0.0022960719652473927, + 0.003001757664605975, + 0.003279336029663682, + 0.0028609149158000946, + 0.0027922780718654394, + 0.003272301983088255, + 0.0019418112933635712, + 0.0029975338838994503, + 0.0025782703887671232, + 0.0029793642461299896, + 0.0036542171146720648, + 0.003117196960374713, + 0.002783829113468528, + 0.003018768737092614, + 0.0030669707339257, + 0.003228864399716258 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.8335974812507629 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004123044200241566 + ], + "opposite_logprob_tokens": [ + -5.504123210906982 + ], + "logprob_reward": -0.004123044200241566, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004123044200241566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3459610641002655, + 0.2004311978816986, + 0.07887376099824905, + 0.06800054013729095, + 0.09043758362531662, + 0.03489276394248009, + 0.07168589532375336, + 0.050967536866664886, + 0.027126476168632507, + 0.0316232293844223 + ], + "credit_map": [ + 0.03206929937005043, + 0.05276449769735336, + 0.05171836167573929, + 0.03517959266901016, + 0.04439595714211464, + 0.05057614669203758, + 0.029643915593624115, + 0.04961324855685234, + 0.014474384486675262, + 0.030840178951621056, + 0.036070872098207474, + 0.02536114491522312, + 0.022908350452780724, + 0.03045940212905407, + 0.01390350703150034, + 0.026413341984152794, + 0.005536095704883337, + 0.012392302043735981, + 0.011366262100636959, + 0.014005487784743309, + 0.01043372880667448, + 0.010566923767328262, + 0.005855001043528318, + 0.00871795229613781, + 0.00642351433634758, + 0.008780733682215214, + 0.013090514577925205, + 0.009236670099198818, + 0.0071323406882584095, + 0.010172855108976364, + 0.005008411128073931, + 0.00815549585968256, + 0.008028741925954819, + 0.01385174598544836, + 0.01602989435195923, + 0.012120313011109829, + 0.012120271101593971, + 0.012365549802780151, + 0.006453298963606358, + 0.009467759169638157, + 0.004010689910501242, + 0.0037691586185246706, + 0.005526116117835045, + 0.003162023378536105, + 0.006186941172927618, + 0.004809723701328039, + 0.0028267703019082546, + 0.004601338412612677, + 0.008917153812944889, + 0.007011887617409229, + 0.013406749814748764, + 0.009818094782531261, + 0.007026853039860725, + 0.01100099552422762, + 0.007131597492843866, + 0.007372553460299969, + 0.005149743519723415, + 0.0060743005014956, + 0.0074694110080599785, + 0.00737258605659008, + 0.005373388063162565, + 0.00783795490860939, + 0.0054209246300160885, + 0.006269220728427172, + 0.00284339627251029, + 0.003296774113550782, + 0.004822264425456524, + 0.0021717639174312353, + 0.004140172153711319, + 0.003792111063376069, + 0.0022008924279361963, + 0.003859098767861724, + 0.004775660112500191, + 0.003681098110973835, + 0.005391704384237528, + 0.003122293157503009, + 0.003999670501798391, + 0.004139948636293411, + 0.002606120426207781, + 0.003906731028109789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.004123044200241566 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011076628230512142 + ], + "opposite_logprob_tokens": [ + -4.5110764503479 + ], + "logprob_reward": -0.011076628230512142, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011076628230512142, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3943118453025818, + 0.15461893379688263, + 0.07367552816867828, + 0.06276243180036545, + 0.06800694763660431, + 0.0822521522641182, + 0.03920095041394234, + 0.04292098432779312, + 0.05015188455581665, + 0.03209838271141052 + ], + "credit_map": [ + 0.034639861434698105, + 0.052067823708057404, + 0.04675443470478058, + 0.03742416948080063, + 0.04736340790987015, + 0.06907451152801514, + 0.032685525715351105, + 0.07430211454629898, + 0.012808393687009811, + 0.02358252741396427, + 0.024295836687088013, + 0.015584250912070274, + 0.01877867989242077, + 0.02423086017370224, + 0.013385215774178505, + 0.021953167393803596, + 0.0059382678009569645, + 0.008311792276799679, + 0.009237634018063545, + 0.007265905849635601, + 0.010934284888207912, + 0.011974772438406944, + 0.006520054303109646, + 0.013492811471223831, + 0.006249049678444862, + 0.008215656504034996, + 0.008515727706253529, + 0.00819830410182476, + 0.008158986456692219, + 0.009509553201496601, + 0.00447241123765707, + 0.009442747570574284, + 0.007199289742857218, + 0.007731557358056307, + 0.011623491533100605, + 0.007784478832036257, + 0.011313878931105137, + 0.00933033600449562, + 0.0060006557032465935, + 0.007023262325674295, + 0.0076469723135232925, + 0.008925274014472961, + 0.009101424366235733, + 0.006199767347425222, + 0.017155004665255547, + 0.013802464120090008, + 0.006592345889657736, + 0.01282890047878027, + 0.0042676315642893314, + 0.005435606464743614, + 0.005694891791790724, + 0.00396762927994132, + 0.00422564335167408, + 0.006221708841621876, + 0.0035533669870346785, + 0.005834473296999931, + 0.0038295479025691748, + 0.00523710809648037, + 0.005907015409320593, + 0.005382838658988476, + 0.005985013209283352, + 0.006765000522136688, + 0.003274030750617385, + 0.006540426518768072, + 0.0042852312326431274, + 0.006354554556310177, + 0.00706505635753274, + 0.006143655627965927, + 0.008013221435248852, + 0.007307035848498344, + 0.0036745730321854353, + 0.007308553904294968, + 0.004387541674077511, + 0.003323192708194256, + 0.004747782368212938, + 0.0024607668165117502, + 0.004798174370080233, + 0.004339177627116442, + 0.003442594315856695, + 0.004599154461175203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.011076628230512142 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193536713719368 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193536713719368, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193536713719368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30199459195137024, + 0.19349342584609985, + 0.08238198608160019, + 0.050625115633010864, + 0.19573451578617096, + 0.033639807254076004, + 0.030267799273133278, + 0.03231188282370567, + 0.029604002833366394, + 0.04994695633649826 + ], + "credit_map": [ + 0.02385079301893711, + 0.04766096919775009, + 0.05115559324622154, + 0.03563033044338226, + 0.04919169470667839, + 0.03985597938299179, + 0.02288912981748581, + 0.0317600779235363, + 0.014119640924036503, + 0.025035111233592033, + 0.03335047513246536, + 0.02785775437951088, + 0.027351681143045425, + 0.02621716447174549, + 0.0163110438734293, + 0.023250551894307137, + 0.005969099700450897, + 0.012018399313092232, + 0.011634055525064468, + 0.011182366870343685, + 0.013698459602892399, + 0.01102752797305584, + 0.006385160144418478, + 0.010466919280588627, + 0.005200994201004505, + 0.0058033280074596405, + 0.010385345667600632, + 0.007804307620972395, + 0.004330041352659464, + 0.007553978357464075, + 0.004086201079189777, + 0.005460916552692652, + 0.019392328336834908, + 0.02338399924337864, + 0.027319950982928276, + 0.020146096125245094, + 0.03492587432265282, + 0.028899457305669785, + 0.017765628173947334, + 0.023901168256998062, + 0.0030375123023986816, + 0.003183528780937195, + 0.006332052405923605, + 0.003655968699604273, + 0.0040552616119384766, + 0.0050425054505467415, + 0.003676659893244505, + 0.004656318575143814, + 0.003905631136149168, + 0.004144123289734125, + 0.005229962058365345, + 0.003625900950282812, + 0.002978527918457985, + 0.004467396531254053, + 0.0025228995364159346, + 0.0033933569211512804, + 0.003456305479630828, + 0.003249942557886243, + 0.0049028522334992886, + 0.0033464222215116024, + 0.0048410347662866116, + 0.004501943476498127, + 0.002612794516608119, + 0.00540058733895421, + 0.00303352321498096, + 0.0034247515723109245, + 0.005362303927540779, + 0.0020754209253937006, + 0.0058358777314424515, + 0.0038622755091637373, + 0.002285690512508154, + 0.0037241592071950436, + 0.005355645902454853, + 0.004863399546593428, + 0.006812345236539841, + 0.005669181235134602, + 0.008972836658358574, + 0.006947106681764126, + 0.0046883258037269115, + 0.006638114340603352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.014193536713719368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a wood-paneled academic library with tall bookshelves filled with old leather-bound books in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a young man with shaggy black hair wearing a dark green cardigan visible, leaning forward at a large polished mahogany table?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is an older man with silver-gray hair wearing a dark tweed blazer and white collared shirt seated opposite, shaking his head slowly with his gaze averted?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a tablet on the table emitting a bright cool blue digital glow that illuminates the characters' faces and hands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a classic green-shaded bank lamp on the table emitting warm golden light that creates deep shadows contrasting with the tablet's glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495283007621765 + ], + "opposite_logprob_tokens": [ + -2.92995285987854 + ], + "logprob_reward": -0.05495283007621765, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495283007621765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3093383312225342, + 0.14253057539463043, + 0.09623308479785919, + 0.06377401947975159, + 0.09797631204128265, + 0.06223359331488609, + 0.05599679797887802, + 0.08852601051330566, + 0.04362437501549721, + 0.03976689279079437 + ], + "credit_map": [ + 0.027113797143101692, + 0.049057487398386, + 0.048218004405498505, + 0.03439081832766533, + 0.035188689827919006, + 0.0460323840379715, + 0.02460257150232792, + 0.04473457485437393, + 0.014073790982365608, + 0.01857885718345642, + 0.017917931079864502, + 0.020090190693736076, + 0.01563636213541031, + 0.02285894565284252, + 0.009944386780261993, + 0.023430110886693, + 0.01107469480484724, + 0.009892966598272324, + 0.014750178903341293, + 0.014148034155368805, + 0.015676192939281464, + 0.012752660550177097, + 0.006939197424799204, + 0.010999162681400776, + 0.005906001664698124, + 0.008022641763091087, + 0.009586367756128311, + 0.01426009926944971, + 0.007746860384941101, + 0.008085280656814575, + 0.004532745573669672, + 0.005634021945297718, + 0.010309085249900818, + 0.014509514905512333, + 0.01584060676395893, + 0.016814274713397026, + 0.008463853038847446, + 0.013199538923799992, + 0.008947591297328472, + 0.009891847148537636, + 0.006601674482226372, + 0.005876146722584963, + 0.011428557336330414, + 0.009564192034304142, + 0.007502098102122545, + 0.008401721715927124, + 0.004286247305572033, + 0.008572952821850777, + 0.006346649955958128, + 0.009086545556783676, + 0.009116651490330696, + 0.006213681306689978, + 0.005197130609303713, + 0.00817336980253458, + 0.0041120112873613834, + 0.007750761229544878, + 0.0089866416528821, + 0.009395795874297619, + 0.013482548296451569, + 0.01349557563662529, + 0.008404561318457127, + 0.01287044771015644, + 0.011625057086348534, + 0.010265388526022434, + 0.004816305357962847, + 0.004797699395567179, + 0.008208814077079296, + 0.004186545964330435, + 0.005263343453407288, + 0.006154471542686224, + 0.004175545647740364, + 0.006021653767675161, + 0.00430357875302434, + 0.004028982948511839, + 0.006627705879509449, + 0.004644451662898064, + 0.007479620631784201, + 0.004997397307306528, + 0.003256449941545725, + 0.004428705666214228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.05495283007621765 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_tensor([1416, 1416])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_tensor([1416, 1416])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..799ae0d2af3132183683d9d28f202ceed4933dd3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_tensor([1416, 1416])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fluid 360-degree tracking shot, beginning from a slightly upward angle, slowly orbits a mysterious figure. This figure is completely shrouded in a thick, matte black robe, its wide hood drawn down deeply, shrouding its features in impenetrable shadow. It stands before a massive door of dark, weathered oak, its panels riddled with deep grain and spider-web-like cracks. A massive, rusted iron knocker hangs from its center. As the camera pans, a bloodless, sickly pale hand slowly but firmly emerges from its wide sleeve. Its long, bony fingertips carefully curl into a fist, and its knuckles strike the heavy wooden door three times with distinct, dull thuds, a hollow sound that echoes through the silent stone corridor. The only illumination came from a flickering torch on the wall, its warm orange-yellow glow casting dancing shadows on the folds of the robe, sometimes illuminating the pale blue veins on the fingers, sometimes swallowing them up in darkness. The whole scene was filled with a suspenseful, solemn and unsettling sense of ritual.", + "seed": 28601, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_1416_seed_28601_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3755848132073879, + "vlm_logprob": -0.3755848132073879, + "vlm_yes_no_margin": 1.95, + "vlm_token_credit": -0.3755848132073879, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6124836206436157 + ], + "opposite_logprob_tokens": [ + -0.23748363554477692 + ], + "logprob_reward": -1.6124836206436157, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6124836206436157, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1967037320137024, + 0.11914242804050446, + 0.17575807869434357, + 0.07521228492259979, + 0.07196132838726044, + 0.18040074408054352, + 0.05679819732904434, + 0.0422736294567585, + 0.04142370447516441, + 0.04032592847943306 + ], + "credit_map": [ + 0.018997136503458023, + 0.03661187365651131, + 0.022580141201615334, + 0.031812701374292374, + 0.03614814206957817, + 0.014587071724236012, + 0.02661481499671936, + 0.009351829066872597, + 0.010557490400969982, + 0.017574379220604897, + 0.009229780174791813, + 0.027120083570480347, + 0.020803267136216164, + 0.007870078086853027, + 0.014623493887484074, + 0.01136385090649128, + 0.012396369129419327, + 0.01565048284828663, + 0.012125587090849876, + 0.05841286852955818, + 0.02377646043896675, + 0.018158001825213432, + 0.021902600303292274, + 0.013335700146853924, + 0.00866521243005991, + 0.00984533317387104, + 0.005218127276748419, + 0.010252165608108044, + 0.00817606970667839, + 0.01871420629322529, + 0.008050436154007912, + 0.006290731020271778, + 0.0062672048807144165, + 0.005234012845903635, + 0.011151130311191082, + 0.009572192095220089, + 0.004786157049238682, + 0.011958327144384384, + 0.010380271822214127, + 0.012612024322152138, + 0.01809042878448963, + 0.014252451248466969, + 0.03744451701641083, + 0.029440149664878845, + 0.01115573849529028, + 0.038152750581502914, + 0.015692448243498802, + 0.016172224655747414, + 0.004964751657098532, + 0.009699649177491665, + 0.006008591968566179, + 0.007869071327149868, + 0.004381732549518347, + 0.008684128522872925, + 0.0049589104019105434, + 0.010231362655758858, + 0.0035904471296817064, + 0.006577047053724527, + 0.0033217081800103188, + 0.006637485697865486, + 0.007126154378056526, + 0.005034353584051132, + 0.005880620330572128, + 0.004105810075998306, + 0.004729076754301786, + 0.004950689151883125, + 0.005282747559249401, + 0.0066692885011434555, + 0.0024772302713245153, + 0.006072836462408304, + 0.0060525136068463326, + 0.005189321935176849, + 0.004307399038225412, + 0.003050687490031123, + 0.003268551779910922, + 0.006903422065079212, + 0.006247134413570166, + 0.002983267419040203, + 0.007664792239665985, + 0.005900672636926174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -1.6124836206436157 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006505131721497 + ], + "opposite_logprob_tokens": [ + -2.6950650215148926 + ], + "logprob_reward": -0.07006505131721497, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006505131721497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31814736127853394, + 0.3517765998840332, + 0.10329698026180267, + 0.060978569090366364, + 0.04656495526432991, + 0.035003188997507095, + 0.01685906946659088, + 0.014186698012053967, + 0.026545386761426926, + 0.02664119377732277 + ], + "credit_map": [ + 0.02256263978779316, + 0.0636950358748436, + 0.04125344753265381, + 0.057572536170482635, + 0.04386426880955696, + 0.01646501198410988, + 0.059676673263311386, + 0.013057759031653404, + 0.026338651776313782, + 0.03699225187301636, + 0.03477644920349121, + 0.1443374902009964, + 0.04822191223502159, + 0.010424282401800156, + 0.03963768854737282, + 0.011047828942537308, + 0.006790323182940483, + 0.008435383439064026, + 0.005676065105944872, + 0.036407917737960815, + 0.014993608929216862, + 0.011405891738831997, + 0.01052626222372055, + 0.009061524644494057, + 0.006192326545715332, + 0.008217274211347103, + 0.005760563537478447, + 0.008848045952618122, + 0.009681876748800278, + 0.008787721395492554, + 0.008121808059513569, + 0.005368954502046108, + 0.004548158496618271, + 0.005244360771030188, + 0.007340569980442524, + 0.006451406516134739, + 0.0035429373383522034, + 0.008032403886318207, + 0.004514815751463175, + 0.006890304386615753, + 0.0042918771505355835, + 0.0038097361102700233, + 0.005324050318449736, + 0.005767405964434147, + 0.002148630563169718, + 0.007239507511258125, + 0.0026786194648593664, + 0.0037433598190546036, + 0.0014767601387575269, + 0.0023769917897880077, + 0.002259546425193548, + 0.002289228141307831, + 0.0016504775267094374, + 0.0027582617476582527, + 0.0020146628376096487, + 0.002033141441643238, + 0.0015174950240179896, + 0.0025051843840628862, + 0.0018361203838139772, + 0.0021198338363319635, + 0.0008306149393320084, + 0.0014944594586268067, + 0.0024256103206425905, + 0.0014573796652257442, + 0.003682403126731515, + 0.005696518812328577, + 0.002985746134072542, + 0.0040526981465518475, + 0.0014018516521900892, + 0.002233670325949788, + 0.0032482249662280083, + 0.0032442728988826275, + 0.0035973680205643177, + 0.0036188808735460043, + 0.002876193029806018, + 0.004722599871456623, + 0.002099408069625497, + 0.0020421503577381372, + 0.004001093562692404, + 0.0036835011560469866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.07006505131721497 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867317155003548 + ], + "opposite_logprob_tokens": [ + -3.048673152923584 + ], + "logprob_reward": -0.04867317155003548, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867317155003548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1604219526052475, + 0.21847209334373474, + 0.1593174934387207, + 0.06419115513563156, + 0.18221452832221985, + 0.06373973935842514, + 0.029118744656443596, + 0.029643917456269264, + 0.06031591072678566, + 0.032564427703619 + ], + "credit_map": [ + 0.012146648950874805, + 0.027887221425771713, + 0.022946897894144058, + 0.01882142573595047, + 0.033606041222810745, + 0.011614822782576084, + 0.024990126490592957, + 0.008408790454268456, + 0.01576520875096321, + 0.023283611983060837, + 0.028539549559354782, + 0.0882468968629837, + 0.026029128581285477, + 0.006386886816471815, + 0.024003736674785614, + 0.006217084359377623, + 0.010424133390188217, + 0.012325775809586048, + 0.009542394429445267, + 0.055192217230796814, + 0.023274006322026253, + 0.013402491807937622, + 0.02142143063247204, + 0.013735055923461914, + 0.005586558487266302, + 0.0040828026831150055, + 0.004591805394738913, + 0.006833332125097513, + 0.005587843246757984, + 0.014793301932513714, + 0.005977428052574396, + 0.016738086938858032, + 0.019670654088258743, + 0.016855567693710327, + 0.028273852542042732, + 0.021507324650883675, + 0.017869768664240837, + 0.030513400211930275, + 0.023902520537376404, + 0.02362145110964775, + 0.007639497518539429, + 0.005555495619773865, + 0.008585982024669647, + 0.009126337245106697, + 0.004322879482060671, + 0.012189115397632122, + 0.007077350746840239, + 0.00924308318644762, + 0.0023609730415046215, + 0.004832059144973755, + 0.0037004072219133377, + 0.004277742933481932, + 0.002460751449689269, + 0.004217909649014473, + 0.003969389945268631, + 0.0032995122019201517, + 0.0029441951774060726, + 0.0036941487342119217, + 0.003945864271372557, + 0.004843615461140871, + 0.002597477287054062, + 0.0028152931481599808, + 0.004252966493368149, + 0.004550359211862087, + 0.010462061502039433, + 0.008983510546386242, + 0.007150256074965, + 0.0082034757360816, + 0.0045617204159498215, + 0.008947583846747875, + 0.006407854612916708, + 0.005599452182650566, + 0.00406672665849328, + 0.0024881092831492424, + 0.005000007804483175, + 0.005943717435002327, + 0.0033739416394382715, + 0.002974247559905052, + 0.005008553620427847, + 0.003709125332534313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.04867317155003548 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129816472530365 + ], + "opposite_logprob_tokens": [ + -2.2379815578460693 + ], + "logprob_reward": -0.1129816472530365, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129816472530365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17308403551578522, + 0.3315431475639343, + 0.10863620042800903, + 0.08633363246917725, + 0.08737048506736755, + 0.023926446214318275, + 0.07674611359834671, + 0.02099454030394554, + 0.05412573367357254, + 0.037239670753479004 + ], + "credit_map": [ + 0.012402290478348732, + 0.030230022966861725, + 0.023860398679971695, + 0.030050113797187805, + 0.025477934628725052, + 0.00801143329590559, + 0.03657842427492142, + 0.0064733996987342834, + 0.019906722009181976, + 0.031819093972444534, + 0.0537429116666317, + 0.1350107342004776, + 0.040202490985393524, + 0.009231633506715298, + 0.03207405284047127, + 0.009555500000715256, + 0.008598756045103073, + 0.01650729589164257, + 0.010363820008933544, + 0.013693731278181076, + 0.015126517973840237, + 0.019697854295372963, + 0.012948242016136646, + 0.011699979193508625, + 0.004271104000508785, + 0.006128589157015085, + 0.0038701205048710108, + 0.00865645706653595, + 0.005249428562819958, + 0.029442844912409782, + 0.007223924156278372, + 0.021491166204214096, + 0.009259985759854317, + 0.003650798462331295, + 0.019633349031209946, + 0.010652802884578705, + 0.0046950895339250565, + 0.014998268336057663, + 0.008237837813794613, + 0.016242345795035362, + 0.002884336980059743, + 0.003782459534704685, + 0.002624214394018054, + 0.003427542746067047, + 0.002226074691861868, + 0.003560449229553342, + 0.00257410597987473, + 0.0028472605627030134, + 0.006849995814263821, + 0.010864592157304287, + 0.008707872591912746, + 0.012371586635708809, + 0.005473114084452391, + 0.009401018731296062, + 0.011167451739311218, + 0.011910470202565193, + 0.002756153466179967, + 0.003114235820248723, + 0.0035163876600563526, + 0.002896666992455721, + 0.0013272588839754462, + 0.002042591106146574, + 0.00256594130769372, + 0.002775302156805992, + 0.008502201177179813, + 0.00616036169230938, + 0.006571074482053518, + 0.008832103572785854, + 0.0030123305041342974, + 0.00874246098101139, + 0.007517517544329166, + 0.004787676502019167, + 0.005232964642345905, + 0.003073033643886447, + 0.006042178720235825, + 0.00596214272081852, + 0.0030320407822728157, + 0.004187039099633694, + 0.004303876310586929, + 0.005406394135206938 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.1129816472530365 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372057527303696 + ], + "opposite_logprob_tokens": [ + -3.4087204933166504 + ], + "logprob_reward": -0.03372057527303696, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372057527303696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1632004678249359, + 0.11089269816875458, + 0.08870440721511841, + 0.27217647433280945, + 0.04267188906669617, + 0.05994149670004845, + 0.03101401962339878, + 0.018053995445370674, + 0.1285066455602646, + 0.08483795076608658 + ], + "credit_map": [ + 0.01002382580190897, + 0.02943073958158493, + 0.020157361403107643, + 0.030104169622063637, + 0.02420593984425068, + 0.0076533895917236805, + 0.03460191190242767, + 0.0070231338031589985, + 0.007722171954810619, + 0.012806969694793224, + 0.011372635141015053, + 0.04354254528880119, + 0.014822680503129959, + 0.003928330726921558, + 0.012037348933517933, + 0.004660014994442463, + 0.005267464555799961, + 0.015002727508544922, + 0.007052741013467312, + 0.018920470029115677, + 0.008793074637651443, + 0.01273905299603939, + 0.010377601720392704, + 0.0105512710288167, + 0.02126489207148552, + 0.023964691907167435, + 0.01809544861316681, + 0.022059619426727295, + 0.021132100373506546, + 0.07595082372426987, + 0.030039910227060318, + 0.059668973088264465, + 0.004055666271597147, + 0.001450881827622652, + 0.007959617301821709, + 0.006178973242640495, + 0.003055555047467351, + 0.007207379210740328, + 0.004120200406759977, + 0.008643618784844875, + 0.00903504341840744, + 0.008098526857793331, + 0.008539821952581406, + 0.008230681531131268, + 0.004742579534649849, + 0.0061514596454799175, + 0.007798404898494482, + 0.007344977930188179, + 0.0031260126270353794, + 0.00527949957177043, + 0.003919668961316347, + 0.004711215849965811, + 0.00201099319383502, + 0.005062158685177565, + 0.0030914803501218557, + 0.003812991315498948, + 0.002364341402426362, + 0.0033778175711631775, + 0.0016400322783738375, + 0.0028003042098134756, + 0.0015908593777567148, + 0.002893376862630248, + 0.0017315148143097758, + 0.0016557492781430483, + 0.02166353166103363, + 0.024124713614583015, + 0.017229506745934486, + 0.017505856230854988, + 0.006220273207873106, + 0.012539711780846119, + 0.017206864431500435, + 0.012016195803880692, + 0.015310839749872684, + 0.0071635134518146515, + 0.012234115041792393, + 0.014827042818069458, + 0.005361335352063179, + 0.006687500048428774, + 0.013594566844403744, + 0.009659037925302982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.03372057527303696 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree tracking shot, beginning from a slightly upward angle, slowly orbits a mysterious figure. This figure is completely shrouded in a thick, matte black robe, its wide hood drawn down deeply, shrouding its features in impenetrable shadow. It stands before a massive door of dark, weathered oak, its panels riddled with deep grain and spider-web-like cracks. A massive, rusted iron knocker hangs from its center. As the camera pans, a bloodless, sickly pale hand slowly but firmly emerges from its wide sleeve. Its long, bony fingertips carefully curl into a fist, and its knuckles strike the heavy wooden door three times with distinct, dull thuds, a hollow sound that echoes through the silent stone corridor. The only illumination came from a flickering torch on the wall, its warm orange-yellow glow casting dancing shadows on the folds of the robe, sometimes illuminating the pale blue veins on the fingers, sometimes swallowing them up in darkness. The whole scene was filled with a suspenseful, solemn and unsettling sense of ritual.", + "seed": 28602, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_1416_seed_28602_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.6357398226857185, + "vlm_logprob": -0.6357398226857185, + "vlm_yes_no_margin": 1.675, + "vlm_token_credit": -0.6357398226857185, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.9372823238372803 + ], + "opposite_logprob_tokens": [ + -0.062282416969537735 + ], + "logprob_reward": -2.9372823238372803, + "yes_no_margin_reward": -2.875, + "token_credit_reward": -2.9372823238372803, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.42880889773368835, + 0.06669960170984268, + 0.15075576305389404, + 0.030671659857034683, + 0.04192965850234032, + 0.18879331648349762, + 0.031896695494651794, + 0.01659521833062172, + 0.03009016439318657, + 0.013759028166532516 + ], + "credit_map": [ + 0.012818609364330769, + 0.019015204161405563, + 0.012134947814047337, + 0.02314242348074913, + 0.031928952783346176, + 0.15561223030090332, + 0.020064692944288254, + 0.15409183502197266, + 0.003530028974637389, + 0.005663259886205196, + 0.003434604499489069, + 0.006410856731235981, + 0.009276328608393669, + 0.01247474830597639, + 0.01394884753972292, + 0.011960931122303009, + 0.005023142788559198, + 0.008108015172183514, + 0.006373124662786722, + 0.0070840646512806416, + 0.021319355815649033, + 0.05649833008646965, + 0.009428980760276318, + 0.03692075237631798, + 0.003766452195122838, + 0.003929860889911652, + 0.0037937837187200785, + 0.0028058900497853756, + 0.003268310334533453, + 0.004370604641735554, + 0.004937977530062199, + 0.003798779798671603, + 0.0021804494317620993, + 0.004018401261419058, + 0.003644510405138135, + 0.0040987334214150906, + 0.006887442898005247, + 0.005527558736503124, + 0.009802925400435925, + 0.005769636016339064, + 0.0060212756507098675, + 0.009003419429063797, + 0.012775648385286331, + 0.04000667482614517, + 0.07250779122114182, + 0.01869785040616989, + 0.016296759247779846, + 0.01348388846963644, + 0.004277385771274567, + 0.0024361847899854183, + 0.0057740723714232445, + 0.003274369053542614, + 0.004106709733605385, + 0.004462542477995157, + 0.003256289754062891, + 0.0043091438710689545, + 0.0018673664890229702, + 0.001173560507595539, + 0.0018352875486016273, + 0.0017276486614719033, + 0.0022843945771455765, + 0.0033657923340797424, + 0.0015974482521414757, + 0.0027437189128249884, + 0.0028390942607074976, + 0.0021056258119642735, + 0.0036286425311118364, + 0.005206592846661806, + 0.0029073304031044245, + 0.004163770470768213, + 0.004440573044121265, + 0.004798535257577896, + 0.0015632992144674063, + 0.0015507458010688424, + 0.0017496913205832243, + 0.0020449708681553602, + 0.0009722673567011952, + 0.002783783245831728, + 0.0011533494107425213, + 0.0019409202504903078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -2.9372823238372803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06215986609458923 + ], + "opposite_logprob_tokens": [ + -2.812159776687622 + ], + "logprob_reward": -0.06215986609458923, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06215986609458923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42964088916778564, + 0.0409737266600132, + 0.09145554900169373, + 0.02533888630568981, + 0.026311621069908142, + 0.313442200422287, + 0.032695475965738297, + 0.014567148871719837, + 0.014957286417484283, + 0.010617241263389587 + ], + "credit_map": [ + 0.01453499123454094, + 0.020145535469055176, + 0.018291106447577477, + 0.023225238546729088, + 0.032059382647275925, + 0.15202026069164276, + 0.02031714655458927, + 0.14904725551605225, + 0.0023350920528173447, + 0.004943087697029114, + 0.002482144394889474, + 0.006576217710971832, + 0.006061559543013573, + 0.005558037664741278, + 0.008368357084691525, + 0.0046492270193994045, + 0.003566327504813671, + 0.005817195400595665, + 0.005815850105136633, + 0.005082519259303808, + 0.012859825976192951, + 0.03059774823486805, + 0.0073314812034368515, + 0.02038460411131382, + 0.0017492898041382432, + 0.0029139493126422167, + 0.0023278933949768543, + 0.00186954066157341, + 0.0033519472926855087, + 0.005262463819235563, + 0.005065170582383871, + 0.00279863178730011, + 0.0013339791912585497, + 0.0024830361362546682, + 0.0023326948285102844, + 0.0025420214515179396, + 0.004241075366735458, + 0.0035271355882287025, + 0.006169096566736698, + 0.0036825835704803467, + 0.009194578044116497, + 0.013907724991440773, + 0.019573356956243515, + 0.06507734954357147, + 0.12307785451412201, + 0.03180646523833275, + 0.026381850242614746, + 0.02442304790019989, + 0.0035414332523941994, + 0.002417178126052022, + 0.004495465196669102, + 0.003215605625882745, + 0.0050549558363854885, + 0.005316710565239191, + 0.003750526113435626, + 0.004903601016849279, + 0.0014963536523282528, + 0.0015134476125240326, + 0.0017619759310036898, + 0.0017010067822411656, + 0.0019492005230858922, + 0.0023940096143633127, + 0.0016497769393026829, + 0.002101378981024027, + 0.0013081772485747933, + 0.000927406013943255, + 0.0018811251502484083, + 0.002557774307206273, + 0.0016257482348009944, + 0.0022320749703794718, + 0.0022925797384232283, + 0.002132400870323181, + 0.0009267089772038162, + 0.0011889546876773238, + 0.0013675348600372672, + 0.0014786013634875417, + 0.000903923821169883, + 0.0020713768899440765, + 0.0011247210204601288, + 0.001555419759824872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.06215986609458923 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03812966123223305 + ], + "opposite_logprob_tokens": [ + -3.2881295680999756 + ], + "logprob_reward": -0.03812966123223305, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03812966123223305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.584923505783081, + 0.05345021188259125, + 0.10656697303056717, + 0.08588303625583649, + 0.025104748085141182, + 0.06789795309305191, + 0.0326136089861393, + 0.01597740314900875, + 0.02016218565404415, + 0.007420361507683992 + ], + "credit_map": [ + 0.016358632594347, + 0.02751566655933857, + 0.01584072783589363, + 0.031717896461486816, + 0.04322969168424606, + 0.21139922738075256, + 0.02816755697131157, + 0.21069417893886566, + 0.0032188952900469303, + 0.0058807688765227795, + 0.003602981101721525, + 0.0047883098013699055, + 0.008439847268164158, + 0.0088350810110569, + 0.010222106240689754, + 0.008462229743599892, + 0.0041970741003751755, + 0.005878823343664408, + 0.0038596552331000566, + 0.004570877179503441, + 0.015596400015056133, + 0.0408448688685894, + 0.006212274543941021, + 0.025407014414668083, + 0.002844578353688121, + 0.009926032274961472, + 0.007631061132997274, + 0.006261094938963652, + 0.016852598637342453, + 0.011238577775657177, + 0.023285117000341415, + 0.007843993604183197, + 0.0013141517993062735, + 0.0023198523558676243, + 0.0019119849894195795, + 0.002569318749010563, + 0.0033950216602534056, + 0.003523671068251133, + 0.0065214005298912525, + 0.00354935135692358, + 0.0024127853102982044, + 0.0041274805553257465, + 0.0051495833322405815, + 0.013564099557697773, + 0.024830445647239685, + 0.007235068362206221, + 0.005483698099851608, + 0.00509480619803071, + 0.003887095721438527, + 0.0030800816603004932, + 0.004079727455973625, + 0.003666408360004425, + 0.003915281966328621, + 0.005386504344642162, + 0.003778828075155616, + 0.004819683730602264, + 0.0016374732367694378, + 0.0009907097555696964, + 0.00185963639523834, + 0.0017659984296187758, + 0.0026933352928608656, + 0.0031071703415364027, + 0.0013785980409011245, + 0.002544483868405223, + 0.0017160262214019895, + 0.001161753199994564, + 0.002148094354197383, + 0.0031941705383360386, + 0.0024816810619086027, + 0.0023851024452596903, + 0.003941794857382774, + 0.003133568447083235, + 0.0009295293712057173, + 0.0010589183075353503, + 0.0009047170751728117, + 0.0012440212303772569, + 0.0006832696381025016, + 0.000880742387380451, + 0.0006184902158565819, + 0.0011006742715835571 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.03812966123223305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903427630662918 + ], + "opposite_logprob_tokens": [ + -2.5790343284606934 + ], + "logprob_reward": -0.07903427630662918, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903427630662918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24569272994995117, + 0.047003887593746185, + 0.30100545287132263, + 0.06922987848520279, + 0.01908717304468155, + 0.2595924139022827, + 0.024403518065810204, + 0.01325749047100544, + 0.012718664482235909, + 0.008008827455341816 + ], + "credit_map": [ + 0.008473512716591358, + 0.01092741172760725, + 0.00724794389680028, + 0.013616063632071018, + 0.01797468401491642, + 0.08790132403373718, + 0.012422614730894566, + 0.08712917566299438, + 0.0028834696859121323, + 0.004547895863652229, + 0.002668827772140503, + 0.0043143173679709435, + 0.007740645669400692, + 0.007178747560828924, + 0.012074816972017288, + 0.005595169961452484, + 0.009780589491128922, + 0.01419598888605833, + 0.01130524929612875, + 0.01285562478005886, + 0.041880641132593155, + 0.12162061780691147, + 0.011124429292976856, + 0.07824230194091797, + 0.0029706633649766445, + 0.009276997298002243, + 0.005405786447227001, + 0.005282615777105093, + 0.012705533765256405, + 0.011434185318648815, + 0.014929113909602165, + 0.007224982604384422, + 0.0012592918938025832, + 0.001681092195212841, + 0.001461298787035048, + 0.0016066771931946278, + 0.0030600978061556816, + 0.0039027489256113768, + 0.00383742549456656, + 0.0022785400506109, + 0.007024080958217382, + 0.014243775978684425, + 0.013903331942856312, + 0.06383766233921051, + 0.09532234817743301, + 0.0239483080804348, + 0.019640393555164337, + 0.021672513335943222, + 0.0024471438955515623, + 0.0018417863175272942, + 0.003076347289606929, + 0.0027809967286884785, + 0.0031496176961809397, + 0.004930279217660427, + 0.002364580752328038, + 0.003812766168266535, + 0.0010954784229397774, + 0.0008635518606752157, + 0.001809942303225398, + 0.0015591016272082925, + 0.002435492118820548, + 0.0024442551657557487, + 0.0012647071853280067, + 0.0017849620198830962, + 0.0007695552194491029, + 0.0010597057407721877, + 0.0019349451176822186, + 0.0016450914554297924, + 0.0013359945733100176, + 0.002194043016061187, + 0.0020384970121085644, + 0.0017408322310075164, + 0.0008931225747801363, + 0.0012310650199651718, + 0.0009027039632201195, + 0.0011569707421585917, + 0.0006684823893010616, + 0.0012330448953434825, + 0.0007188210729509592, + 0.0012046170886605978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.07903427630662918 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06209298595786095 + ], + "opposite_logprob_tokens": [ + -2.8120930194854736 + ], + "logprob_reward": -0.06209298595786095, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06209298595786095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3125278651714325, + 0.05839013680815697, + 0.24848191440105438, + 0.029748447239398956, + 0.046580858528614044, + 0.19717323780059814, + 0.04193679243326187, + 0.025984548032283783, + 0.014062014408409595, + 0.02511419542133808 + ], + "credit_map": [ + 0.010768579319119453, + 0.01549581065773964, + 0.008794945664703846, + 0.016948213800787926, + 0.024093734100461006, + 0.11029617488384247, + 0.015111899934709072, + 0.11101851612329483, + 0.003377145854756236, + 0.005214158445596695, + 0.004267487209290266, + 0.0058384062722325325, + 0.01024191826581955, + 0.00794853176921606, + 0.01406489871442318, + 0.007437590043991804, + 0.009034902788698673, + 0.012046728283166885, + 0.00985628366470337, + 0.010349494405090809, + 0.035708215087652206, + 0.09959427267313004, + 0.010392810218036175, + 0.0614992156624794, + 0.002531725447624922, + 0.00398729695007205, + 0.003227760549634695, + 0.0024111757520586252, + 0.004024038091301918, + 0.003088585799559951, + 0.006781460717320442, + 0.0036964036989957094, + 0.0035656972322613, + 0.005058891139924526, + 0.003284597070887685, + 0.0046949684619903564, + 0.006732824258506298, + 0.007951127365231514, + 0.00966666080057621, + 0.005626089870929718, + 0.008486027829349041, + 0.00872081983834505, + 0.022370437160134315, + 0.021833723410964012, + 0.07280504703521729, + 0.02298254519701004, + 0.01724437065422535, + 0.022730253636837006, + 0.004352937918156385, + 0.003348334925249219, + 0.005763259716331959, + 0.004434274975210428, + 0.004879635758697987, + 0.007346486672759056, + 0.005070818588137627, + 0.006741043645888567, + 0.0014591950457543135, + 0.0008889290620572865, + 0.002753146691247821, + 0.002926431130617857, + 0.0031770148780196905, + 0.007269007153809071, + 0.002308631781488657, + 0.005202193278819323, + 0.0011214549886062741, + 0.0011575771495699883, + 0.0015747321303933859, + 0.0018248767592012882, + 0.0017894059419631958, + 0.0024374949280172586, + 0.0025155942421406507, + 0.001640878152102232, + 0.0018184861401095986, + 0.0018357984954491258, + 0.0026556134689599276, + 0.003906820435076952, + 0.0022046808153390884, + 0.00627228943631053, + 0.0020625093020498753, + 0.004357996396720409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.06209298595786095 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree tracking shot, beginning from a slightly upward angle, slowly orbits a mysterious figure. This figure is completely shrouded in a thick, matte black robe, its wide hood drawn down deeply, shrouding its features in impenetrable shadow. It stands before a massive door of dark, weathered oak, its panels riddled with deep grain and spider-web-like cracks. A massive, rusted iron knocker hangs from its center. As the camera pans, a bloodless, sickly pale hand slowly but firmly emerges from its wide sleeve. Its long, bony fingertips carefully curl into a fist, and its knuckles strike the heavy wooden door three times with distinct, dull thuds, a hollow sound that echoes through the silent stone corridor. The only illumination came from a flickering torch on the wall, its warm orange-yellow glow casting dancing shadows on the folds of the robe, sometimes illuminating the pale blue veins on the fingers, sometimes swallowing them up in darkness. The whole scene was filled with a suspenseful, solemn and unsettling sense of ritual.", + "seed": 28603, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_1416_seed_28603_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.6427547946572304, + "vlm_logprob": -0.6427547946572304, + "vlm_yes_no_margin": 1.625, + "vlm_token_credit": -0.6427547946572304, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.935758113861084 + ], + "opposite_logprob_tokens": [ + -0.060758013278245926 + ], + "logprob_reward": -2.935758113861084, + "yes_no_margin_reward": -2.875, + "token_credit_reward": -2.935758113861084, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15761874616146088, + 0.07445920258760452, + 0.06645173579454422, + 0.059398092329502106, + 0.10120708495378494, + 0.1052468940615654, + 0.09875664114952087, + 0.096110500395298, + 0.06768374145030975, + 0.1730673760175705 + ], + "credit_map": [ + 0.013650299049913883, + 0.02499251998960972, + 0.021058354526758194, + 0.019556373357772827, + 0.02287891134619713, + 0.0193814467638731, + 0.02441754750907421, + 0.011683298274874687, + 0.006972395349293947, + 0.01739659532904625, + 0.014679167419672012, + 0.011229374445974827, + 0.008578160777688026, + 0.00441795215010643, + 0.007972734048962593, + 0.003212820738554001, + 0.004727893974632025, + 0.012225001119077206, + 0.010345458053052425, + 0.01076219230890274, + 0.007235921919345856, + 0.0036276665050536394, + 0.01458686776459217, + 0.002940729958936572, + 0.005425812676548958, + 0.006411136593669653, + 0.019484998658299446, + 0.012042039074003696, + 0.004690598230808973, + 0.0029542590491473675, + 0.005052756983786821, + 0.0033364901319146156, + 0.006166600156575441, + 0.0148629667237401, + 0.011692840605974197, + 0.010908402502536774, + 0.018351074308156967, + 0.015153788961470127, + 0.010380533523857594, + 0.013690880499780178, + 0.007813580334186554, + 0.01342261303216219, + 0.010602601803839207, + 0.010532867163419724, + 0.018735982477664948, + 0.018893122673034668, + 0.014045357704162598, + 0.011200770735740662, + 0.01000477746129036, + 0.01077846810221672, + 0.018418623134493828, + 0.011640812270343304, + 0.008507049642503262, + 0.010510849766433239, + 0.005945470184087753, + 0.022950585931539536, + 0.004632230382412672, + 0.014186738058924675, + 0.008410696871578693, + 0.01213083602488041, + 0.019722966477274895, + 0.02193104848265648, + 0.007389385253190994, + 0.007706594653427601, + 0.003083375282585621, + 0.014578708447515965, + 0.004426333587616682, + 0.020397888496518135, + 0.005900299176573753, + 0.011051024310290813, + 0.003623128402978182, + 0.0046229856088757515, + 0.006081467028707266, + 0.05596449598670006, + 0.010172274895012379, + 0.06777570396661758, + 0.006717259529978037, + 0.01121363416314125, + 0.008991874754428864, + 0.006150674540549517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -2.935758113861084 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06215113028883934 + ], + "opposite_logprob_tokens": [ + -2.8121511936187744 + ], + "logprob_reward": -0.06215113028883934, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06215113028883934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13771100342273712, + 0.0598665252327919, + 0.06452302634716034, + 0.03809255734086037, + 0.13947875797748566, + 0.07231762260198593, + 0.07807835936546326, + 0.20443016290664673, + 0.09931826591491699, + 0.10618367791175842 + ], + "credit_map": [ + 0.010303822346031666, + 0.02677903138101101, + 0.016242200508713722, + 0.016593139618635178, + 0.021939925849437714, + 0.014204186387360096, + 0.022389095276594162, + 0.009259612299501896, + 0.008968230336904526, + 0.012976343743503094, + 0.00979990791529417, + 0.008954335935413837, + 0.007151296827942133, + 0.0031380823347717524, + 0.006682626903057098, + 0.002195707056671381, + 0.00670459633693099, + 0.009356111288070679, + 0.011965615674853325, + 0.011074732057750225, + 0.0075706494972109795, + 0.003420640714466572, + 0.010852990671992302, + 0.0035776926670223475, + 0.003294558497145772, + 0.0036169642116874456, + 0.012151009403169155, + 0.007213793229311705, + 0.003209315473213792, + 0.0019534605089575052, + 0.004379169084131718, + 0.002274292055517435, + 0.0091174291446805, + 0.014737947843968868, + 0.015518760308623314, + 0.013331474736332893, + 0.017918532714247704, + 0.016976049169898033, + 0.019299611449241638, + 0.0325789675116539, + 0.006296772044152021, + 0.009494359605014324, + 0.010170175693929195, + 0.008743862621486187, + 0.011620428413152695, + 0.013593457639217377, + 0.006167339161038399, + 0.006231239531189203, + 0.0059045120142400265, + 0.013307595625519753, + 0.014720016159117222, + 0.01211825106292963, + 0.010015731677412987, + 0.00787163246423006, + 0.005586325190961361, + 0.008554295636713505, + 0.013231567107141018, + 0.02950468100607395, + 0.024624299257993698, + 0.028090307489037514, + 0.018273118883371353, + 0.06635153293609619, + 0.008790391497313976, + 0.015564273111522198, + 0.008302678354084492, + 0.013508210889995098, + 0.00874459370970726, + 0.012378398329019547, + 0.015832480043172836, + 0.025561245158314705, + 0.005336335394531488, + 0.009654330089688301, + 0.003423066344112158, + 0.03307899460196495, + 0.005928662605583668, + 0.039960287511348724, + 0.004379838705062866, + 0.0066926185972988605, + 0.0064891367219388485, + 0.006231071427464485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.06215113028883934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0337345227599144 + ], + "opposite_logprob_tokens": [ + -3.4087345600128174 + ], + "logprob_reward": -0.0337345227599144, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0337345227599144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09148570150136948, + 0.02958584390580654, + 0.0612543448805809, + 0.058070216327905655, + 0.10499411821365356, + 0.08161173015832901, + 0.12511761486530304, + 0.1946054995059967, + 0.05603574216365814, + 0.1972392052412033 + ], + "credit_map": [ + 0.007410457823425531, + 0.012501470744609833, + 0.022329766303300858, + 0.014500617980957031, + 0.00842693168669939, + 0.009119560942053795, + 0.010237184353172779, + 0.006959713064134121, + 0.0028779248241335154, + 0.006033668760210276, + 0.004019733518362045, + 0.0046203890815377235, + 0.003716902108862996, + 0.002523330971598625, + 0.00381299015134573, + 0.001980907516553998, + 0.004512513522058725, + 0.009070985950529575, + 0.010526175610721111, + 0.011764644645154476, + 0.00733340485021472, + 0.0033848597668111324, + 0.011907188221812248, + 0.0027545674238353968, + 0.005024905316531658, + 0.005324319936335087, + 0.020358391106128693, + 0.01306383777409792, + 0.0038941747043281794, + 0.0037078550085425377, + 0.00410047359764576, + 0.0025962581858038902, + 0.006011460907757282, + 0.011419511400163174, + 0.013355395756661892, + 0.011827310547232628, + 0.02048669382929802, + 0.0159904845058918, + 0.011793269775807858, + 0.014109999872744083, + 0.007601804565638304, + 0.01134057529270649, + 0.009212900884449482, + 0.009586994536221027, + 0.009662482887506485, + 0.013776054605841637, + 0.01022748276591301, + 0.010203433223068714, + 0.010737561620771885, + 0.015791112557053566, + 0.014033547602593899, + 0.01222042366862297, + 0.013023685663938522, + 0.026901869103312492, + 0.010302419774234295, + 0.022106995806097984, + 0.01352227758616209, + 0.01913331262767315, + 0.015840424224734306, + 0.030481131747364998, + 0.018336134031414986, + 0.0725960060954094, + 0.010466200299561024, + 0.014230015687644482, + 0.0023798185866326094, + 0.01142963394522667, + 0.005327910650521517, + 0.014279008843004704, + 0.007653577718883753, + 0.006859226617962122, + 0.003762708744034171, + 0.004343859851360321, + 0.006582268048077822, + 0.06097415089607239, + 0.012426985427737236, + 0.07330512255430222, + 0.009493334218859673, + 0.010153925977647305, + 0.013710333965718746, + 0.010593092069029808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.0337345227599144 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12713077664375305 + ], + "opposite_logprob_tokens": [ + -2.1271307468414307 + ], + "logprob_reward": -0.12713077664375305, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12713077664375305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19633612036705017, + 0.07171407341957092, + 0.05047920346260071, + 0.05371496081352234, + 0.0946066826581955, + 0.05222742632031441, + 0.12185294181108475, + 0.1024269089102745, + 0.07020364701747894, + 0.18643805384635925 + ], + "credit_map": [ + 0.020152535289525986, + 0.019762860611081123, + 0.06514672935009003, + 0.03875400871038437, + 0.01363679114729166, + 0.010231731459498405, + 0.012433075346052647, + 0.016218405216932297, + 0.006406479515135288, + 0.013371996581554413, + 0.011846846900880337, + 0.011056497693061829, + 0.010047191753983498, + 0.0051238336600363255, + 0.010698294267058372, + 0.003162937005981803, + 0.005219961982220411, + 0.007597394287586212, + 0.009717363864183426, + 0.006447432562708855, + 0.005051891319453716, + 0.0027081191074103117, + 0.01044459082186222, + 0.003292449750006199, + 0.00424228748306632, + 0.005367259029299021, + 0.016593175008893013, + 0.010779134929180145, + 0.00366847962141037, + 0.0036218026652932167, + 0.00666796462610364, + 0.0027748553548008204, + 0.006688839755952358, + 0.011639289557933807, + 0.009068546816706657, + 0.0110037075355649, + 0.012729652225971222, + 0.008737794123589993, + 0.013784358277916908, + 0.020954495295882225, + 0.0041008200496435165, + 0.007691794540733099, + 0.005717211868613958, + 0.0052967313677072525, + 0.0061837309040129185, + 0.006686266511678696, + 0.008996918797492981, + 0.007553954143077135, + 0.01145932450890541, + 0.019282927736639977, + 0.027255164459347725, + 0.014876260422170162, + 0.011029209941625595, + 0.011786099523305893, + 0.007549441885203123, + 0.018614517524838448, + 0.006561501882970333, + 0.014519739896059036, + 0.016421034932136536, + 0.014405331574380398, + 0.018315285444259644, + 0.012141147628426552, + 0.010422494262456894, + 0.009640374220907688, + 0.0036893351934850216, + 0.01297434326261282, + 0.006140781566500664, + 0.016900764778256416, + 0.009821531362831593, + 0.010331804864108562, + 0.005025244317948818, + 0.005319834686815739, + 0.008460975252091885, + 0.05425987392663956, + 0.013138137757778168, + 0.07129642367362976, + 0.00998194795101881, + 0.010825001634657383, + 0.010409263893961906, + 0.008066444657742977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.12713077664375305 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499942973256111 + ], + "opposite_logprob_tokens": [ + -2.929999351501465 + ], + "logprob_reward": -0.05499942973256111, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499942973256111, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22995221614837646, + 0.046866524964571, + 0.030447250232100487, + 0.03233208879828453, + 0.2033180296421051, + 0.06607463955879211, + 0.1067114770412445, + 0.10811498016119003, + 0.09708891808986664, + 0.07909391075372696 + ], + "credit_map": [ + 0.01766541413962841, + 0.02946881577372551, + 0.0357414111495018, + 0.02841763198375702, + 0.03768954053521156, + 0.022396517917513847, + 0.041955988854169846, + 0.01661689579486847, + 0.0062434556894004345, + 0.008435710333287716, + 0.008445858024060726, + 0.006293118931353092, + 0.006477240938693285, + 0.0034940866753458977, + 0.005585703533142805, + 0.001891348627395928, + 0.0027835106011480093, + 0.004621824715286493, + 0.006002251990139484, + 0.004813032224774361, + 0.0023303453344851732, + 0.0018968124641105533, + 0.006305546034127474, + 0.0016939279157668352, + 0.0023212358355522156, + 0.004611639305949211, + 0.005566497333347797, + 0.005120778921991587, + 0.0028486158698797226, + 0.0029747732914984226, + 0.005821674130856991, + 0.0030668724793940783, + 0.011696449480950832, + 0.01816767454147339, + 0.02018948644399643, + 0.021801117807626724, + 0.0378919281065464, + 0.02758382447063923, + 0.03470363840460777, + 0.031283900141716, + 0.005506431218236685, + 0.010217303410172462, + 0.005938016343861818, + 0.007770772557705641, + 0.007500586099922657, + 0.008882604539394379, + 0.00977463461458683, + 0.010484288446605206, + 0.018013132736086845, + 0.010197505354881287, + 0.010885994881391525, + 0.012253515422344208, + 0.010518450289964676, + 0.011991508305072784, + 0.0059185996651649475, + 0.026932768523693085, + 0.006290160119533539, + 0.014009390026330948, + 0.01342159416526556, + 0.012549949809908867, + 0.019985197111964226, + 0.015195890329778194, + 0.013192491605877876, + 0.013470306992530823, + 0.007456032559275627, + 0.013412135653197765, + 0.008594351820647717, + 0.017266396433115005, + 0.014267982915043831, + 0.02232963591814041, + 0.005798943806439638, + 0.007963431067764759, + 0.0051893009804189205, + 0.015514474362134933, + 0.008142746984958649, + 0.02578640729188919, + 0.005243067629635334, + 0.007186781615018845, + 0.006621749140322208, + 0.005409378092736006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.05499942973256111 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fluid 360-degree tracking shot, beginning from a slightly upward angle, slowly orbits a mysterious figure. This figure is completely shrouded in a thick, matte black robe, its wide hood drawn down deeply, shrouding its features in impenetrable shadow. It stands before a massive door of dark, weathered oak, its panels riddled with deep grain and spider-web-like cracks. A massive, rusted iron knocker hangs from its center. As the camera pans, a bloodless, sickly pale hand slowly but firmly emerges from its wide sleeve. Its long, bony fingertips carefully curl into a fist, and its knuckles strike the heavy wooden door three times with distinct, dull thuds, a hollow sound that echoes through the silent stone corridor. The only illumination came from a flickering torch on the wall, its warm orange-yellow glow casting dancing shadows on the folds of the robe, sometimes illuminating the pale blue veins on the fingers, sometimes swallowing them up in darkness. The whole scene was filled with a suspenseful, solemn and unsettling sense of ritual.", + "seed": 28604, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_5/index_1416_seed_28604_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.6327403143048287, + "vlm_logprob": -0.6327403143048287, + "vlm_yes_no_margin": 1.725, + "vlm_token_credit": -0.6327403143048287, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.9356584548950195 + ], + "opposite_logprob_tokens": [ + -0.06065850332379341 + ], + "logprob_reward": -2.9356584548950195, + "yes_no_margin_reward": -2.875, + "token_credit_reward": -2.9356584548950195, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.19121834635734558, + 0.21250022947788239, + 0.07080353051424026, + 0.14415712654590607, + 0.1650327742099762, + 0.030646517872810364, + 0.05961429700255394, + 0.05195298418402672, + 0.021000122651457787, + 0.05307406932115555 + ], + "credit_map": [ + 0.021184131503105164, + 0.022038239985704422, + 0.019297922030091286, + 0.051450107246637344, + 0.01590689644217491, + 0.016385644674301147, + 0.018419373780488968, + 0.02653602510690689, + 0.02072472870349884, + 0.014527685008943081, + 0.01183830015361309, + 0.06250002235174179, + 0.06026570126414299, + 0.010714147239923477, + 0.019735107198357582, + 0.01219455897808075, + 0.006469538901001215, + 0.007166977506130934, + 0.007188194431364536, + 0.010385993868112564, + 0.007334763649851084, + 0.016707446426153183, + 0.0065703196451067924, + 0.008980297483503819, + 0.007918287999927998, + 0.006012278143316507, + 0.006841451860964298, + 0.010362413711845875, + 0.012577341869473457, + 0.0471617765724659, + 0.0069705406203866005, + 0.04631303995847702, + 0.00790037028491497, + 0.007538851350545883, + 0.007309141103178263, + 0.00866072066128254, + 0.011235605925321579, + 0.05891605094075203, + 0.01116947177797556, + 0.052302587777376175, + 0.001911027473397553, + 0.0032027987763285637, + 0.0025685846339911222, + 0.003217606572434306, + 0.0032666409388184547, + 0.005125979892909527, + 0.006163354497402906, + 0.005190527532249689, + 0.004396151285618544, + 0.007220865692943335, + 0.005901216063648462, + 0.004962040111422539, + 0.006075121928006411, + 0.0062275538221001625, + 0.013655493035912514, + 0.011175857856869698, + 0.004323736298829317, + 0.006952499505132437, + 0.005262945778667927, + 0.005575946532189846, + 0.008839927613735199, + 0.005826246924698353, + 0.00721439765766263, + 0.007957285270094872, + 0.0018551648827269673, + 0.0035037731286138296, + 0.001762400264851749, + 0.002804666059091687, + 0.002222511451691389, + 0.003456390928477049, + 0.0018810270121321082, + 0.0035141899716109037, + 0.003549823071807623, + 0.007823171094059944, + 0.00790319126099348, + 0.007503142114728689, + 0.006541449110955, + 0.007562711369246244, + 0.0046866219490766525, + 0.007503962609916925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -2.9356584548950195 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0487685464322567 + ], + "opposite_logprob_tokens": [ + -3.0487685203552246 + ], + "logprob_reward": -0.0487685464322567, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.0487685464322567, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10148204118013382, + 0.1623862385749817, + 0.06121877208352089, + 0.44360944628715515, + 0.05937874689698219, + 0.05010129511356354, + 0.02637111209332943, + 0.0418107770383358, + 0.014300641603767872, + 0.03934100270271301 + ], + "credit_map": [ + 0.01182932872325182, + 0.012378345243632793, + 0.010025539435446262, + 0.026334742084145546, + 0.007264187093824148, + 0.009520701132714748, + 0.01040163729339838, + 0.013727548532187939, + 0.024249712005257607, + 0.011293306946754456, + 0.007640160620212555, + 0.03854893147945404, + 0.03857908397912979, + 0.011029445566236973, + 0.014771704562008381, + 0.01627388224005699, + 0.005899658426642418, + 0.0032452046871185303, + 0.004794255364686251, + 0.006508320104330778, + 0.003752793651074171, + 0.01509137824177742, + 0.0036110759247094393, + 0.018316080793738365, + 0.017911704257130623, + 0.026422059163451195, + 0.02980107069015503, + 0.03396092727780342, + 0.04278085380792618, + 0.13443906605243683, + 0.028449242934584618, + 0.12984445691108704, + 0.0036043734289705753, + 0.004308596253395081, + 0.003930485341697931, + 0.004408606328070164, + 0.008190424181520939, + 0.014044726267457008, + 0.009871719405055046, + 0.01101981196552515, + 0.003191279945895076, + 0.005423546768724918, + 0.004612124990671873, + 0.00476804468780756, + 0.0032700803130865097, + 0.009828954935073853, + 0.005416201893240213, + 0.013591057620942593, + 0.0031742537394165993, + 0.0029701218008995056, + 0.002213166793808341, + 0.001676742685958743, + 0.004999087192118168, + 0.003245339961722493, + 0.004466115031391382, + 0.003626281162723899, + 0.003377834102138877, + 0.0057998597621917725, + 0.004826297517865896, + 0.003356542671099305, + 0.006999183911830187, + 0.010061556473374367, + 0.0033211498521268368, + 0.004068348556756973, + 0.0009343032143078744, + 0.0016728724585846066, + 0.0010375215206295252, + 0.0017700609751045704, + 0.0018290693406015635, + 0.0021550015080720186, + 0.002139918738976121, + 0.00276189181022346, + 0.003225124441087246, + 0.006228375248610973, + 0.004853800870478153, + 0.006055973470211029, + 0.004402237478643656, + 0.005402673501521349, + 0.003552499460056424, + 0.005620315205305815 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.0487685464322567 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038138724863529205 + ], + "opposite_logprob_tokens": [ + -3.2881386280059814 + ], + "logprob_reward": -0.038138724863529205, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038138724863529205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14712119102478027, + 0.04267771169543266, + 0.06894082576036453, + 0.4258686602115631, + 0.09454429894685745, + 0.02371877245604992, + 0.05525815114378929, + 0.09278304874897003, + 0.013904436491429806, + 0.035182882100343704 + ], + "credit_map": [ + 0.01804691180586815, + 0.016927145421504974, + 0.014028385281562805, + 0.03938528522849083, + 0.011753606610000134, + 0.012627029791474342, + 0.0127493841573596, + 0.02160344272851944, + 0.006647964008152485, + 0.003738973056897521, + 0.004224390257149935, + 0.009813087992370129, + 0.006037522107362747, + 0.0029592588543891907, + 0.0054471297189593315, + 0.0038093896582722664, + 0.007101066410541534, + 0.006227319128811359, + 0.004947077017277479, + 0.008129328489303589, + 0.008532519452273846, + 0.01225138921290636, + 0.006112005561590195, + 0.015640124678611755, + 0.014016548171639442, + 0.013329954817891121, + 0.013585327193140984, + 0.03182299807667732, + 0.04088406637310982, + 0.14847417175769806, + 0.01747601479291916, + 0.14627957344055176, + 0.007472171448171139, + 0.008304741233587265, + 0.005964233539998531, + 0.006201339885592461, + 0.007903211750090122, + 0.027343690395355225, + 0.00675501162186265, + 0.024599900469183922, + 0.0011038505472242832, + 0.001921186107210815, + 0.0021372917108237743, + 0.0021014681551605463, + 0.0026914027985185385, + 0.0042134965769946575, + 0.004607602953910828, + 0.004942473489791155, + 0.004502671305090189, + 0.006323353853076696, + 0.005852764472365379, + 0.004877129103988409, + 0.0047039431519806385, + 0.006486988626420498, + 0.012011177837848663, + 0.010500122793018818, + 0.00965596828609705, + 0.010969074442982674, + 0.009193585254251957, + 0.008705548010766506, + 0.013068711385130882, + 0.01156807690858841, + 0.011862357147037983, + 0.017759723588824272, + 0.001294006360694766, + 0.0014299957547336817, + 0.0009771054610610008, + 0.0017779263434931636, + 0.001430191216059029, + 0.003216259879991412, + 0.0014156841207295656, + 0.0023632675874978304, + 0.0025139099452644587, + 0.005638669710606337, + 0.004892089869827032, + 0.004583859816193581, + 0.00425271550193429, + 0.004702048841863871, + 0.003205610206350684, + 0.005393976345658302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.038138724863529205 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620991475880146 + ], + "opposite_logprob_tokens": [ + -2.8120992183685303 + ], + "logprob_reward": -0.0620991475880146, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620991475880146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051737964153289795, + 0.09389182180166245, + 0.042952947318553925, + 0.5142517685890198, + 0.10613387823104858, + 0.01765231415629387, + 0.017274923622608185, + 0.1184382513165474, + 0.010232629254460335, + 0.027433540672063828 + ], + "credit_map": [ + 0.0067114620469510555, + 0.005379397887736559, + 0.005261223763227463, + 0.012854590080678463, + 0.004444894380867481, + 0.004360582213848829, + 0.005150825250893831, + 0.007574988529086113, + 0.014419492334127426, + 0.007176555693149567, + 0.004364728461951017, + 0.01913747377693653, + 0.019855473190546036, + 0.008341070264577866, + 0.007898306474089622, + 0.012698713690042496, + 0.005088088568300009, + 0.00396188348531723, + 0.0033052631188184023, + 0.0060669248923659325, + 0.008009707555174828, + 0.0077320002019405365, + 0.004845624789595604, + 0.003943454474210739, + 0.01617421954870224, + 0.01456938125193119, + 0.012365011498332024, + 0.03911346569657326, + 0.04807119071483612, + 0.18392933905124664, + 0.01946885883808136, + 0.18056026101112366, + 0.0046823034062981606, + 0.004574666265398264, + 0.0047370512038469315, + 0.005603914614766836, + 0.00600874749943614, + 0.03888319432735443, + 0.006608452647924423, + 0.03503554314374924, + 0.0009739113738760352, + 0.0013602119870483875, + 0.0012379396939650178, + 0.001546571496874094, + 0.002908775582909584, + 0.0031761950813233852, + 0.003062863601371646, + 0.003385844174772501, + 0.0017082368722185493, + 0.00155072461348027, + 0.0016087517142295837, + 0.001205985783599317, + 0.0033294125460088253, + 0.0021716782357543707, + 0.0028654213529080153, + 0.0028347133193165064, + 0.01325235515832901, + 0.013267332687973976, + 0.011707361787557602, + 0.010616063140332699, + 0.026658212766051292, + 0.018084486946463585, + 0.01180356927216053, + 0.013048873282968998, + 0.0007447241805493832, + 0.0011525661684572697, + 0.0008828258141875267, + 0.0012820077827200294, + 0.0012509416555985808, + 0.0018424259033054113, + 0.0013204850256443024, + 0.0017566526075825095, + 0.002561420202255249, + 0.004004900809377432, + 0.0039766207337379456, + 0.0037177428603172302, + 0.0032383515499532223, + 0.004066435620188713, + 0.0024959673173725605, + 0.0033721018116921186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.0620991475880146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera perform a fluid 360-degree tracking shot that orbits the central figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the central figure wearing a thick matte black robe with a wide hood drawn down to obscure their features?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bloodless, sickly pale hand visible emerging from the figure's sleeve and curling into a fist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the hand strike a massive dark weathered oak door three times with distinct dull thuds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated primarily by a flickering torch casting warm orange-yellow light and dancing shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07903669774532318 + ], + "opposite_logprob_tokens": [ + -2.5790367126464844 + ], + "logprob_reward": -0.07903669774532318, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07903669774532318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29075080156326294, + 0.1213453859090805, + 0.04225677624344826, + 0.1743835061788559, + 0.09034503996372223, + 0.14498655498027802, + 0.039293963462114334, + 0.04048752039670944, + 0.014819900505244732, + 0.041330598294734955 + ], + "credit_map": [ + 0.03485983982682228, + 0.034660302102565765, + 0.029976801946759224, + 0.07851376384496689, + 0.026469087228178978, + 0.02432815358042717, + 0.029507678002119064, + 0.03243517130613327, + 0.015505677089095116, + 0.007753669749945402, + 0.006236052606254816, + 0.03167544677853584, + 0.030126674100756645, + 0.007690643426030874, + 0.009869429282844067, + 0.012487796135246754, + 0.005953110288828611, + 0.0025010909885168076, + 0.0019718254916369915, + 0.004165020305663347, + 0.0030885133892297745, + 0.008651909418404102, + 0.003177032805979252, + 0.012748274020850658, + 0.007173134479671717, + 0.006132432725280523, + 0.007907652296125889, + 0.01323715876787901, + 0.01531541720032692, + 0.05793622508645058, + 0.007855969481170177, + 0.05882549658417702, + 0.008394615724682808, + 0.00669724028557539, + 0.005332089029252529, + 0.005150056444108486, + 0.009309652261435986, + 0.024052128195762634, + 0.0076071047224104404, + 0.023802151903510094, + 0.007153925485908985, + 0.010638639330863953, + 0.009439289569854736, + 0.011968391947448254, + 0.02580728381872177, + 0.027451172471046448, + 0.024757709354162216, + 0.027770135551691055, + 0.003195917932316661, + 0.003782637184485793, + 0.003801613813266158, + 0.0029476401396095753, + 0.005541051272302866, + 0.0036978931166231632, + 0.009416243061423302, + 0.006910967640578747, + 0.00479466887190938, + 0.0055252802558243275, + 0.003863563761115074, + 0.003401960013434291, + 0.009003495797514915, + 0.005999076645821333, + 0.0034748457837849855, + 0.00442463019862771, + 0.0011653293622657657, + 0.002031721407547593, + 0.0012170168338343501, + 0.0017543509602546692, + 0.002109383698552847, + 0.0016933645820245147, + 0.002666083863005042, + 0.002182650612667203, + 0.0038566598668694496, + 0.007535942830145359, + 0.006307997275143862, + 0.006554427091032267, + 0.004774572793394327, + 0.004624203313142061, + 0.0031684297136962414, + 0.004508366342633963 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.07903669774532318 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_tensor([3172, 3172])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_tensor([3172, 3172])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d3e99814321abffe330f3435ee94f0b6f250987a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_tensor([3172, 3172])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning close-up, completely out of focus at first, reveals a round, anthropomorphic cream cake resting peacefully in a soft, fairytale-like halo. Dreamy morning sunlight filters through the window, gilding the scene with a warm golden hue and illuminating the delicate, pastel-colored ceramic plate beneath. As it slowly awakens, the thin blanket of pink frosting gently slips away, revealing a fluffy, pure white cream body beneath. At this moment, accompanied by a subtle, magical flicker on the screen, the lens's focus smoothly tightens from blur to focus sharply on the cake's face. Finally, focus settles precisely on its two dark chocolate shavings, which serve as eyes. They shimmer faintly in the soft light, then, with a hint of sleep, magically and slowly open, breathing the first glimmer of consciousness into this sweet creature. The entire image is permeated with a tranquil, miraculous awakening.", + "seed": 17181, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17181_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.22719374522566796, + "vlm_logprob": -0.22719374522566796, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.22719374522566796, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0776464119553566 + ], + "opposite_logprob_tokens": [ + -2.702646493911743 + ], + "logprob_reward": -0.0776464119553566, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0776464119553566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46179139614105225, + 0.0577399916946888, + 0.02588762901723385, + 0.07422515749931335, + 0.057445138692855835, + 0.0948350578546524, + 0.0767996683716774, + 0.052813198417425156, + 0.06843268126249313, + 0.030030108988285065 + ], + "credit_map": [ + 0.032591674476861954, + 0.045796945691108704, + 0.07001269608736038, + 0.09353134781122208, + 0.04698293283581734, + 0.03587111458182335, + 0.097272127866745, + 0.039732590317726135, + 0.005559570621699095, + 0.007324496284127235, + 0.007050209678709507, + 0.010960311628878117, + 0.005762552376836538, + 0.007573050446808338, + 0.007759367115795612, + 0.005750433541834354, + 0.0025810040533542633, + 0.002482052193954587, + 0.0034855983685702085, + 0.0033740780781954527, + 0.0023991414345800877, + 0.00504731573164463, + 0.00323084881529212, + 0.003287589643150568, + 0.004718539770692587, + 0.011994619853794575, + 0.006768514402210712, + 0.006263438146561384, + 0.018560688942670822, + 0.010865399613976479, + 0.007929078303277493, + 0.007124873343855143, + 0.00456067081540823, + 0.005273581016808748, + 0.005749646574258804, + 0.007091208826750517, + 0.008798398077487946, + 0.008116684854030609, + 0.011750594712793827, + 0.006104359868913889, + 0.008557644672691822, + 0.009224061854183674, + 0.011112508364021778, + 0.011186440475285053, + 0.011171557940542698, + 0.013336044736206532, + 0.01771398074924946, + 0.01253281719982624, + 0.007618306670337915, + 0.008180114440619946, + 0.009481407701969147, + 0.007038386072963476, + 0.007842513732612133, + 0.012091746553778648, + 0.010960129089653492, + 0.013587068766355515, + 0.0048555913381278515, + 0.005796683952212334, + 0.00625527324154973, + 0.005818487145006657, + 0.006024492904543877, + 0.008037542924284935, + 0.009723441675305367, + 0.006301683373749256, + 0.006749975495040417, + 0.005444823298603296, + 0.0049742176197469234, + 0.006602384150028229, + 0.011020516976714134, + 0.010324168018996716, + 0.01617235504090786, + 0.007144241128116846, + 0.0027081130538135767, + 0.003203583648428321, + 0.004141164943575859, + 0.002932168310508132, + 0.0042479438707232475, + 0.004782119765877724, + 0.005105470772832632, + 0.002909542992711067 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.0776464119553566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8263663053512573 + ], + "opposite_logprob_tokens": [ + -0.5763663053512573 + ], + "logprob_reward": -0.8263663053512573, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8263663053512573, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12638869881629944, + 0.12604056298732758, + 0.13945236802101135, + 0.06049719452857971, + 0.13257919251918793, + 0.08133704960346222, + 0.07990124076604843, + 0.1001366525888443, + 0.09519141167402267, + 0.05847562104463577 + ], + "credit_map": [ + 0.011913142167031765, + 0.013768902979791164, + 0.014053902588784695, + 0.01956811547279358, + 0.01458236388862133, + 0.011309834197163582, + 0.021779507398605347, + 0.019412923604249954, + 0.013334274291992188, + 0.014500418677926064, + 0.03303859382867813, + 0.012006117030978203, + 0.015787938609719276, + 0.015577639453113079, + 0.011259235441684723, + 0.010536350309848785, + 0.017930323258042336, + 0.013281364925205708, + 0.017956800758838654, + 0.015984302386641502, + 0.012337972410023212, + 0.021465787664055824, + 0.01395539753139019, + 0.02654041163623333, + 0.005906614009290934, + 0.008631926961243153, + 0.006833926308900118, + 0.007135598920285702, + 0.006480757147073746, + 0.00881786085665226, + 0.008094770833849907, + 0.008595740422606468, + 0.015591977164149284, + 0.013546598143875599, + 0.01310151256620884, + 0.01309388317167759, + 0.015921184793114662, + 0.02764888107776642, + 0.015539208427071571, + 0.018135949969291687, + 0.008151100017130375, + 0.008852314203977585, + 0.010239480063319206, + 0.008636804297566414, + 0.009265304543077946, + 0.016269084066152573, + 0.011166424490511417, + 0.008756536059081554, + 0.008284411393105984, + 0.00946560874581337, + 0.00785504374653101, + 0.009096698835492134, + 0.007376005873084068, + 0.017480604350566864, + 0.010466886684298515, + 0.009875978343188763, + 0.009195046499371529, + 0.013252560980618, + 0.01289951428771019, + 0.011428232304751873, + 0.011159316636621952, + 0.01547168754041195, + 0.01486382819712162, + 0.01186645869165659, + 0.009404162876307964, + 0.01254918985068798, + 0.010077309794723988, + 0.009883389808237553, + 0.010059382766485214, + 0.017375687137246132, + 0.014498092234134674, + 0.011344194412231445, + 0.0055927494540810585, + 0.006622794549912214, + 0.005480871070176363, + 0.006796965375542641, + 0.00819220207631588, + 0.008458687923848629, + 0.011017912067472935, + 0.006313439458608627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.8263663053512573 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129949539899826 + ], + "opposite_logprob_tokens": [ + -2.237994909286499 + ], + "logprob_reward": -0.1129949539899826, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129949539899826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10435696691274643, + 0.11124296486377716, + 0.11355956643819809, + 0.0543738529086113, + 0.19304344058036804, + 0.09104595333337784, + 0.1715446561574936, + 0.03530683368444443, + 0.0329989418387413, + 0.09252681583166122 + ], + "credit_map": [ + 0.011382934637367725, + 0.01399965025484562, + 0.013066163286566734, + 0.019398408010601997, + 0.009489315561950207, + 0.01134263165295124, + 0.012360959313809872, + 0.013316898606717587, + 0.008940666913986206, + 0.011525739915668964, + 0.02945391647517681, + 0.012653090991079807, + 0.011299471370875835, + 0.015021268278360367, + 0.011846938170492649, + 0.010501882061362267, + 0.01114162988960743, + 0.013452021405100822, + 0.01067221350967884, + 0.00968764815479517, + 0.022342372685670853, + 0.016297468915581703, + 0.013966063037514687, + 0.01600014604628086, + 0.005444774404168129, + 0.005580638535320759, + 0.005270815920084715, + 0.006505112629383802, + 0.00858995784074068, + 0.007971053011715412, + 0.008429880253970623, + 0.006581617519259453, + 0.014020954258739948, + 0.02021293342113495, + 0.018514761701226234, + 0.019444845616817474, + 0.03655429929494858, + 0.029171789065003395, + 0.03858141601085663, + 0.016542455181479454, + 0.009713496081531048, + 0.00957212969660759, + 0.009604633785784245, + 0.011229882948100567, + 0.01112863328307867, + 0.014087704941630363, + 0.01066186185926199, + 0.015047614462673664, + 0.016611620783805847, + 0.015760812908411026, + 0.018486687913537025, + 0.01689782924950123, + 0.017468035221099854, + 0.039137665182352066, + 0.02372700162231922, + 0.023454997688531876, + 0.0037933546118438244, + 0.003924766089767218, + 0.0045575364492833614, + 0.003662972478196025, + 0.004826786927878857, + 0.005768990144133568, + 0.004936716984957457, + 0.0038357095327228308, + 0.003640962764620781, + 0.0037377572152763605, + 0.003969963174313307, + 0.002857818966731429, + 0.004040144849568605, + 0.006376619450747967, + 0.004174095578491688, + 0.0042015789076685905, + 0.009108980186283588, + 0.009824092499911785, + 0.01220215205103159, + 0.00994096789509058, + 0.012392193078994751, + 0.015155413188040257, + 0.01451878808438778, + 0.00938422791659832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.1129949539899826 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11300240457057953 + ], + "opposite_logprob_tokens": [ + -2.238002300262451 + ], + "logprob_reward": -0.11300240457057953, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11300240457057953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11933824419975281, + 0.06604964286088943, + 0.030052006244659424, + 0.14263825118541718, + 0.2305447906255722, + 0.1744840443134308, + 0.09354407340288162, + 0.028530221432447433, + 0.030822576954960823, + 0.08399614691734314 + ], + "credit_map": [ + 0.013729674741625786, + 0.014340836554765701, + 0.017343852669000626, + 0.022287612780928612, + 0.011289650574326515, + 0.010943647474050522, + 0.016207588836550713, + 0.013195378705859184, + 0.0053103286772966385, + 0.007484113797545433, + 0.019801463931798935, + 0.0052747586742043495, + 0.008944853208959103, + 0.009015565738081932, + 0.005477721802890301, + 0.004740833770483732, + 0.002377773867920041, + 0.0034492698032408953, + 0.003837047843262553, + 0.0047487602569162846, + 0.0029100081883370876, + 0.004265065770596266, + 0.004341376945376396, + 0.004122704733163118, + 0.008192811161279678, + 0.02301616407930851, + 0.011929175816476345, + 0.010907965712249279, + 0.038029588758945465, + 0.021561238914728165, + 0.017195086926221848, + 0.011806218884885311, + 0.01792098954319954, + 0.023959774523973465, + 0.022858235985040665, + 0.023870548233389854, + 0.043592892587184906, + 0.03575956076383591, + 0.04144641011953354, + 0.02113637886941433, + 0.01512044109404087, + 0.016113340854644775, + 0.020516926422715187, + 0.01965600810945034, + 0.021922526881098747, + 0.02792917937040329, + 0.03086124174296856, + 0.02236437052488327, + 0.00924165453761816, + 0.00870310328900814, + 0.01110103353857994, + 0.009796738624572754, + 0.015659935772418976, + 0.014068174175918102, + 0.013392697088420391, + 0.011580734513700008, + 0.002418642630800605, + 0.002480860333889723, + 0.0027217124588787556, + 0.003380438080057502, + 0.003473254619166255, + 0.004254728555679321, + 0.006644358858466148, + 0.0031562254298478365, + 0.0025353129021823406, + 0.0031230587046593428, + 0.0032163874711841345, + 0.002584936562925577, + 0.004705504514276981, + 0.005070890765637159, + 0.00667259655892849, + 0.002913889242336154, + 0.00812794640660286, + 0.006451787427067757, + 0.008574762381613255, + 0.00885593518614769, + 0.014180018566548824, + 0.016289129853248596, + 0.013729107566177845, + 0.00778745673596859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.11300240457057953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059586502611637115 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.0059586502611637115, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059586502611637115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26348304748535156, + 0.20890817046165466, + 0.09346059709787369, + 0.09775837510824203, + 0.07885661721229553, + 0.07425420731306076, + 0.07857441157102585, + 0.036532483994960785, + 0.020171577110886574, + 0.04800049960613251 + ], + "credit_map": [ + 0.025208614766597748, + 0.0276639387011528, + 0.03656400740146637, + 0.05387454107403755, + 0.024566756561398506, + 0.021311042830348015, + 0.04003423824906349, + 0.034259893000125885, + 0.019934801384806633, + 0.024358827620744705, + 0.06202925369143486, + 0.015858013182878494, + 0.02193555235862732, + 0.03074071556329727, + 0.014440324157476425, + 0.019610702991485596, + 0.012822063639760017, + 0.008983094245195389, + 0.009340347722172737, + 0.009227003902196884, + 0.014479178003966808, + 0.010866030119359493, + 0.009517540223896503, + 0.01822534017264843, + 0.00673714792355895, + 0.015971142798662186, + 0.009449511766433716, + 0.008182169869542122, + 0.02384822815656662, + 0.01214250922203064, + 0.011011040769517422, + 0.010416635312139988, + 0.006043529137969017, + 0.008597975596785545, + 0.007092182524502277, + 0.007621574681252241, + 0.01670537143945694, + 0.015524173155426979, + 0.010952778160572052, + 0.006319035775959492, + 0.007828409783542156, + 0.006950477138161659, + 0.006655218079686165, + 0.006110948510468006, + 0.011401363648474216, + 0.011899488046765327, + 0.01073999423533678, + 0.012668315321207047, + 0.007798967882990837, + 0.006965602748095989, + 0.008405596949160099, + 0.007667244412004948, + 0.008875093422830105, + 0.017420323565602303, + 0.012790605425834656, + 0.008650985546410084, + 0.0038662166334688663, + 0.003814137075096369, + 0.004545283503830433, + 0.0033234599977731705, + 0.005039791576564312, + 0.006099286023527384, + 0.0064887492917478085, + 0.003355560591444373, + 0.001809667912311852, + 0.0016562326345592737, + 0.0022389967925846577, + 0.0021407362073659897, + 0.002496682805940509, + 0.0032287114299833775, + 0.004247092641890049, + 0.0023534563370049, + 0.004688843619078398, + 0.0035192188806831837, + 0.004317052196711302, + 0.005576371680945158, + 0.006844499614089727, + 0.010736214928328991, + 0.007733218837529421, + 0.004585085436701775 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0059586502611637115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, completely out of focus at first, reveals a round, anthropomorphic cream cake resting peacefully in a soft, fairytale-like halo. Dreamy morning sunlight filters through the window, gilding the scene with a warm golden hue and illuminating the delicate, pastel-colored ceramic plate beneath. As it slowly awakens, the thin blanket of pink frosting gently slips away, revealing a fluffy, pure white cream body beneath. At this moment, accompanied by a subtle, magical flicker on the screen, the lens's focus smoothly tightens from blur to focus sharply on the cake's face. Finally, focus settles precisely on its two dark chocolate shavings, which serve as eyes. They shimmer faintly in the soft light, then, with a hint of sleep, magically and slowly open, breathing the first glimmer of consciousness into this sweet creature. The entire image is permeated with a tranquil, miraculous awakening.", + "seed": 17182, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17182_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.19302804372273386, + "vlm_logprob": -0.19302804372273386, + "vlm_yes_no_margin": 2.874999976158142, + "vlm_token_credit": -0.19302804372273386, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01282608974725008 + ], + "opposite_logprob_tokens": [ + -4.512825965881348 + ], + "logprob_reward": -0.01282608974725008, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01282608974725008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22519804537296295, + 0.05745413154363632, + 0.029790185391902924, + 0.09358896315097809, + 0.06829490512609482, + 0.05593699961900711, + 0.0583350844681263, + 0.10570771992206573, + 0.10760512202978134, + 0.19808882474899292 + ], + "credit_map": [ + 0.02537740208208561, + 0.020714865997433662, + 0.039471060037612915, + 0.03559624031186104, + 0.017027094960212708, + 0.02526969090104103, + 0.03471698611974716, + 0.027024708688259125, + 0.006837328430265188, + 0.00599392456933856, + 0.010220210067927837, + 0.007368609309196472, + 0.00599021278321743, + 0.00687339948490262, + 0.006111565511673689, + 0.008058883249759674, + 0.003970440477132797, + 0.0035110819153487682, + 0.003325577825307846, + 0.0038682816084474325, + 0.0029132836498320103, + 0.003466523252427578, + 0.0037895177956670523, + 0.00494548212736845, + 0.011528539471328259, + 0.009467864409089088, + 0.012833095155656338, + 0.010469521395862103, + 0.012593488208949566, + 0.010071427561342716, + 0.009930376894772053, + 0.016694651916623116, + 0.00888908002525568, + 0.007277664262801409, + 0.008976547047495842, + 0.00847978238016367, + 0.007046844810247421, + 0.005417400971055031, + 0.010214654728770256, + 0.011992939747869968, + 0.00697483541443944, + 0.006994915194809437, + 0.006345079280436039, + 0.0070262569934129715, + 0.00859302468597889, + 0.004179289564490318, + 0.00833930354565382, + 0.0074842991307377815, + 0.009978881105780602, + 0.005722164176404476, + 0.00723697105422616, + 0.006410226691514254, + 0.005465359892696142, + 0.008093957789242268, + 0.005199560895562172, + 0.010227966122329235, + 0.015563315711915493, + 0.012628764845430851, + 0.011080892756581306, + 0.018855344504117966, + 0.0102513637393713, + 0.008421376347541809, + 0.013767867349088192, + 0.015138796530663967, + 0.013285022228956223, + 0.014427428133785725, + 0.01193353533744812, + 0.020566202700138092, + 0.009154225699603558, + 0.007641188334673643, + 0.0175856314599514, + 0.01301189512014389, + 0.025032851845026016, + 0.03218548372387886, + 0.019524551928043365, + 0.031250037252902985, + 0.017804209142923355, + 0.01578250527381897, + 0.025092337280511856, + 0.03141685575246811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.01282608974725008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6936253309249878 + ], + "opposite_logprob_tokens": [ + -0.6936253309249878 + ], + "logprob_reward": -0.6936253309249878, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6936253309249878, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11260764300823212, + 0.08894149959087372, + 0.06595222651958466, + 0.05819094181060791, + 0.0869823768734932, + 0.1183079406619072, + 0.0863732397556305, + 0.21302534639835358, + 0.05828714370727539, + 0.11133165657520294 + ], + "credit_map": [ + 0.01030811294913292, + 0.011656509712338448, + 0.011941744014620781, + 0.018351947888731956, + 0.01600424386560917, + 0.011699925176799297, + 0.018192986026406288, + 0.014452168717980385, + 0.011290512047708035, + 0.007433190941810608, + 0.019743362441658974, + 0.01281873881816864, + 0.005610879044979811, + 0.008452708832919598, + 0.009619617834687233, + 0.013972483575344086, + 0.008511215448379517, + 0.008613746613264084, + 0.01010209135711193, + 0.009327505715191364, + 0.00553095992654562, + 0.00602327985689044, + 0.00890104379504919, + 0.008942381478846073, + 0.006990273483097553, + 0.005670481361448765, + 0.008161762729287148, + 0.009595036506652832, + 0.006447379011660814, + 0.006144262384623289, + 0.006623069755733013, + 0.008558670058846474, + 0.010896356776356697, + 0.008137893863022327, + 0.013646796345710754, + 0.010853433981537819, + 0.008251668885350227, + 0.007022525649517775, + 0.013601764105260372, + 0.014571929350495338, + 0.013605416752398014, + 0.011824062094092369, + 0.01614750176668167, + 0.018114222213625908, + 0.012174955569207668, + 0.010966953821480274, + 0.018692439422011375, + 0.016782376915216446, + 0.011756506748497486, + 0.010438773781061172, + 0.008356871083378792, + 0.014667104929685593, + 0.010281229391694069, + 0.009566384367644787, + 0.011033025570213795, + 0.0102733438834548, + 0.027983330190181732, + 0.030618708580732346, + 0.02239609695971012, + 0.03578603267669678, + 0.0200903732329607, + 0.018815940245985985, + 0.02636566385626793, + 0.030969182029366493, + 0.00926151778548956, + 0.006696681026369333, + 0.007677576970309019, + 0.00848427601158619, + 0.004814575891941786, + 0.0041314950212836266, + 0.007098947186022997, + 0.010122066363692284, + 0.012775203213095665, + 0.010483423247933388, + 0.014087053947150707, + 0.01666697859764099, + 0.011551662348210812, + 0.007888199761509895, + 0.02166222594678402, + 0.016216900199651718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.6936253309249878 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11298462748527527 + ], + "opposite_logprob_tokens": [ + -2.2379846572875977 + ], + "logprob_reward": -0.11298462748527527, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11298462748527527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13401691615581512, + 0.0706590861082077, + 0.19304227828979492, + 0.1825675368309021, + 0.07557559758424759, + 0.07169869542121887, + 0.07688526809215546, + 0.06958580762147903, + 0.030898094177246094, + 0.09507065266370773 + ], + "credit_map": [ + 0.015294970013201237, + 0.011125950142741203, + 0.014712806791067123, + 0.020550798624753952, + 0.013012166135013103, + 0.024775661528110504, + 0.018374579027295113, + 0.016169998794794083, + 0.00754773523658514, + 0.010498602874577045, + 0.010920185595750809, + 0.01313487347215414, + 0.006262792740017176, + 0.006180519238114357, + 0.0071324012242257595, + 0.008981983177363873, + 0.024308426305651665, + 0.021406518295407295, + 0.023706261068582535, + 0.019891882315278053, + 0.022275513038039207, + 0.027692174538969994, + 0.024786077439785004, + 0.028975436463952065, + 0.020233143121004105, + 0.02340952679514885, + 0.022770991548895836, + 0.027179259806871414, + 0.017254943028092384, + 0.01826021447777748, + 0.0259290412068367, + 0.027530431747436523, + 0.007091911043971777, + 0.008444618433713913, + 0.010311945341527462, + 0.009933914989233017, + 0.011932259425520897, + 0.006631545256823301, + 0.012806829996407032, + 0.008422578684985638, + 0.008700321428477764, + 0.007830549962818623, + 0.007939656265079975, + 0.011439434252679348, + 0.007379607297480106, + 0.006540762260556221, + 0.01106374990195036, + 0.010804624296724796, + 0.010839656926691532, + 0.008319764398038387, + 0.007943320088088512, + 0.01012212410569191, + 0.009377006441354752, + 0.008827543817460537, + 0.008590838871896267, + 0.012865021824836731, + 0.00863433163613081, + 0.009056727401912212, + 0.007266823668032885, + 0.013172897510230541, + 0.006615910213440657, + 0.005782401654869318, + 0.009774573147296906, + 0.00928215030580759, + 0.004484738223254681, + 0.0036370055750012398, + 0.0034290954936295748, + 0.0058246785774827, + 0.003102911403402686, + 0.0022556050680577755, + 0.004309105221182108, + 0.0038549592718482018, + 0.010860891081392765, + 0.014879299327731133, + 0.009672025218605995, + 0.016743363812565804, + 0.008881053887307644, + 0.007175407372415066, + 0.012551703490316868, + 0.014306924305856228 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.11298462748527527 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14287599921226501 + ], + "opposite_logprob_tokens": [ + -2.017875909805298 + ], + "logprob_reward": -0.14287599921226501, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14287599921226501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14503566920757294, + 0.03944072499871254, + 0.053292643278837204, + 0.19983096420764923, + 0.09892044216394424, + 0.13746897876262665, + 0.09924478083848953, + 0.11672153323888779, + 0.03892304748296738, + 0.07112123817205429 + ], + "credit_map": [ + 0.010816188529133797, + 0.01584864966571331, + 0.021899303421378136, + 0.024594992399215698, + 0.01286208163946867, + 0.013365918770432472, + 0.029613645747303963, + 0.016034886240959167, + 0.0038060233928263187, + 0.004386949818581343, + 0.006178771145641804, + 0.006867627147585154, + 0.00445352541282773, + 0.004421373829245567, + 0.0048662517219781876, + 0.004460205789655447, + 0.006228703539818525, + 0.0057239714078605175, + 0.006409387569874525, + 0.00806046836078167, + 0.005692897830158472, + 0.006541706621646881, + 0.007377675734460354, + 0.0072578308172523975, + 0.0277982447296381, + 0.02369157411158085, + 0.023844201117753983, + 0.03214925527572632, + 0.02647000551223755, + 0.018426120281219482, + 0.02699350006878376, + 0.020458057522773743, + 0.009310093708336353, + 0.00963866338133812, + 0.012523534707725048, + 0.016105391085147858, + 0.011469615623354912, + 0.006816274486482143, + 0.01917637139558792, + 0.013880498707294464, + 0.01295782346278429, + 0.017815465107560158, + 0.014964955858886242, + 0.018211625516414642, + 0.026775160804390907, + 0.012576750479638577, + 0.018816184252500534, + 0.015351011417806149, + 0.013729890808463097, + 0.01060516107827425, + 0.013766240328550339, + 0.012453346513211727, + 0.011467596516013145, + 0.007024381775408983, + 0.01477009803056717, + 0.015428059734404087, + 0.013835906982421875, + 0.014676470309495926, + 0.012548457831144333, + 0.022909700870513916, + 0.009976287372410297, + 0.0097180912271142, + 0.017548348754644394, + 0.015508272685110569, + 0.005154139827936888, + 0.003379099303856492, + 0.004502537194639444, + 0.005562285427004099, + 0.0030535452533513308, + 0.003129353281110525, + 0.007195969112217426, + 0.006946120411157608, + 0.006997395772486925, + 0.00842767208814621, + 0.009113507345318794, + 0.011067050509154797, + 0.00810515321791172, + 0.0054532005451619625, + 0.01357070542871952, + 0.008386553265154362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.14287599921226501 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00282817124389112 + ], + "opposite_logprob_tokens": [ + -5.877828121185303 + ], + "logprob_reward": -0.00282817124389112, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.00282817124389112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24082139134407043, + 0.20705285668373108, + 0.12117363512516022, + 0.08099310100078583, + 0.09636596590280533, + 0.05471505969762802, + 0.019968660548329353, + 0.061468541622161865, + 0.084455206990242, + 0.032985568046569824 + ], + "credit_map": [ + 0.029202040284872055, + 0.022958997637033463, + 0.059756021946668625, + 0.028060605749487877, + 0.02197209931910038, + 0.01895618811249733, + 0.04009747505187988, + 0.019817959517240524, + 0.0264873206615448, + 0.022468017414212227, + 0.03505128622055054, + 0.030702468007802963, + 0.015042698942124844, + 0.02429598569869995, + 0.02054051123559475, + 0.032464560121297836, + 0.017439885064959526, + 0.014054694212973118, + 0.022261667996644974, + 0.010267629288136959, + 0.0111640440300107, + 0.013187465257942677, + 0.01384056732058525, + 0.01895768754184246, + 0.012132760137319565, + 0.008291074074804783, + 0.01032888051122427, + 0.012396031059324741, + 0.008000288158655167, + 0.010485964827239513, + 0.010949297808110714, + 0.008408804424107075, + 0.014764605090022087, + 0.00991196557879448, + 0.012171116657555103, + 0.014593389816582203, + 0.009016294963657856, + 0.009581017307937145, + 0.013546318747103214, + 0.01278125960379839, + 0.005919292103499174, + 0.005620946176350117, + 0.007211524527519941, + 0.00989400316029787, + 0.005291962530463934, + 0.003821147605776787, + 0.010598208755254745, + 0.006357980892062187, + 0.0022957047913223505, + 0.002246982418000698, + 0.002124849706888199, + 0.0031600387301295996, + 0.002269991673529148, + 0.0012072594836354256, + 0.003065698780119419, + 0.003598135896027088, + 0.007503907661885023, + 0.005753605626523495, + 0.006532988045364618, + 0.011064648628234863, + 0.005866059102118015, + 0.003204427659511566, + 0.016417745500802994, + 0.00512516126036644, + 0.010603826493024826, + 0.01175787765532732, + 0.009454125538468361, + 0.015819882974028587, + 0.00779358297586441, + 0.006179878953844309, + 0.013724450953304768, + 0.009121585637331009, + 0.004078452009707689, + 0.0038259506691247225, + 0.004503745585680008, + 0.004472471307963133, + 0.003797319484874606, + 0.002948644571006298, + 0.00518683809787035, + 0.00417214585468173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00282817124389112 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, completely out of focus at first, reveals a round, anthropomorphic cream cake resting peacefully in a soft, fairytale-like halo. Dreamy morning sunlight filters through the window, gilding the scene with a warm golden hue and illuminating the delicate, pastel-colored ceramic plate beneath. As it slowly awakens, the thin blanket of pink frosting gently slips away, revealing a fluffy, pure white cream body beneath. At this moment, accompanied by a subtle, magical flicker on the screen, the lens's focus smoothly tightens from blur to focus sharply on the cake's face. Finally, focus settles precisely on its two dark chocolate shavings, which serve as eyes. They shimmer faintly in the soft light, then, with a hint of sleep, magically and slowly open, breathing the first glimmer of consciousness into this sweet creature. The entire image is permeated with a tranquil, miraculous awakening.", + "seed": 17183, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17183_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10981204304844142, + "vlm_logprob": -0.10981204304844142, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.10981204304844142, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05074063688516617 + ], + "opposite_logprob_tokens": [ + -3.1757407188415527 + ], + "logprob_reward": -0.05074063688516617, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05074063688516617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20939473807811737, + 0.04931363835930824, + 0.06991581618785858, + 0.24450433254241943, + 0.10178432613611221, + 0.043889112770557404, + 0.036527398973703384, + 0.033366311341524124, + 0.057750895619392395, + 0.15355345606803894 + ], + "credit_map": [ + 0.02492903172969818, + 0.023175615817308426, + 0.023765241727232933, + 0.03389199078083038, + 0.028850052505731583, + 0.03519270941615105, + 0.019808752462267876, + 0.01978130079805851, + 0.009148898534476757, + 0.006252883467823267, + 0.0055890376679599285, + 0.005828987807035446, + 0.005113773513585329, + 0.009148910641670227, + 0.004168043844401836, + 0.004063095431774855, + 0.014001077972352505, + 0.009182531386613846, + 0.008880337700247765, + 0.00721179461106658, + 0.005898099392652512, + 0.013690641149878502, + 0.0053793578408658504, + 0.005671964026987553, + 0.07149461656808853, + 0.022550294175744057, + 0.016458114609122276, + 0.014868886210024357, + 0.020388958975672722, + 0.07275741547346115, + 0.0091806435957551, + 0.016805380582809448, + 0.028227224946022034, + 0.0071542710065841675, + 0.007593984715640545, + 0.0073719872161746025, + 0.00775071419775486, + 0.02699374035000801, + 0.009132914245128632, + 0.0075594764202833176, + 0.005895040463656187, + 0.0035021700896322727, + 0.00481809489428997, + 0.004003733396530151, + 0.004769050516188145, + 0.008219799026846886, + 0.006172162014991045, + 0.006509060971438885, + 0.004972024355083704, + 0.0034204560797661543, + 0.004656435921788216, + 0.0038698508869856596, + 0.004837356507778168, + 0.00448194332420826, + 0.003126374911516905, + 0.007162952329963446, + 0.010601050220429897, + 0.0025190783198922873, + 0.007218638900667429, + 0.002840943168848753, + 0.001953311264514923, + 0.0025632972829043865, + 0.0016742857405915856, + 0.0039957016706466675, + 0.010550695471465588, + 0.004895553924143314, + 0.010079232975840569, + 0.0039964355528354645, + 0.004141305107623339, + 0.0141331497579813, + 0.004307079128921032, + 0.0056474413722753525, + 0.018413212150335312, + 0.06588031351566315, + 0.03602658957242966, + 0.004625056870281696, + 0.010939043946564198, + 0.010851976461708546, + 0.002653061877936125, + 0.004164180718362331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 9 + ], + "selected_reward": -0.05074063688516617 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12719407677650452 + ], + "opposite_logprob_tokens": [ + -2.1271941661834717 + ], + "logprob_reward": -0.12719407677650452, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12719407677650452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6573852896690369, + 0.03983057662844658, + 0.04608792066574097, + 0.037520620971918106, + 0.0296478271484375, + 0.012505985796451569, + 0.04688693955540657, + 0.015919426456093788, + 0.02602635696530342, + 0.08818906545639038 + ], + "credit_map": [ + 0.11528906226158142, + 0.08271977305412292, + 0.08261547237634659, + 0.08093411475419998, + 0.05954565480351448, + 0.12544286251068115, + 0.059061069041490555, + 0.05177726596593857, + 0.005267334170639515, + 0.004641199018806219, + 0.0058489106595516205, + 0.00550431851297617, + 0.004135630093514919, + 0.004232628736644983, + 0.00546244066208601, + 0.0047381166368722916, + 0.008253946900367737, + 0.0056479922495782375, + 0.006614176090806723, + 0.0043022106401622295, + 0.00455345818772912, + 0.0066988300532102585, + 0.0053167943842709064, + 0.0047005158849060535, + 0.006339173763990402, + 0.004799799062311649, + 0.0034471265971660614, + 0.003900518175214529, + 0.003970800433307886, + 0.008880180306732655, + 0.002506003249436617, + 0.003677023109048605, + 0.006631712429225445, + 0.002715896815061569, + 0.002488793106749654, + 0.0027561024762690067, + 0.0028611088637262583, + 0.007052368484437466, + 0.0023234242107719183, + 0.002818422857671976, + 0.0025341659784317017, + 0.000817144988104701, + 0.0016544087557122111, + 0.0009452561498619616, + 0.001168358139693737, + 0.001757321530021727, + 0.0010370324598625302, + 0.0025922979693859816, + 0.0061807152815163136, + 0.004794432315975428, + 0.0050266501493752, + 0.004434256814420223, + 0.008845359086990356, + 0.004565071780234575, + 0.004871487617492676, + 0.008168968372046947, + 0.0033104883041232824, + 0.0016133099561557174, + 0.0022633704356849194, + 0.0015997139271348715, + 0.0023493259213864803, + 0.001920119277201593, + 0.0010360927553847432, + 0.0018270070431753993, + 0.004408548586070538, + 0.0024640951305627823, + 0.003352224128320813, + 0.0018783645937219262, + 0.002442959463223815, + 0.005702209193259478, + 0.0020374569576233625, + 0.0037405004259198904, + 0.009768009185791016, + 0.02666127122938633, + 0.034984324127435684, + 0.00293366564437747, + 0.00414282875135541, + 0.005814336705952883, + 0.0014632990350946784, + 0.002421335084363818 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.12719407677650452 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129763200879097 + ], + "opposite_logprob_tokens": [ + -2.237976312637329 + ], + "logprob_reward": -0.1129763200879097, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129763200879097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18163934350013733, + 0.06936097145080566, + 0.0820508599281311, + 0.12143915891647339, + 0.06937266886234283, + 0.08357542753219604, + 0.09448004513978958, + 0.03871236369013786, + 0.12013878673315048, + 0.13923045992851257 + ], + "credit_map": [ + 0.03493598848581314, + 0.017957346513867378, + 0.01824650540947914, + 0.02314644865691662, + 0.017439505085349083, + 0.036783862859010696, + 0.017633643001317978, + 0.015496045351028442, + 0.009811626747250557, + 0.012313652783632278, + 0.008353481069207191, + 0.00956625398248434, + 0.006205381825566292, + 0.009638680145144463, + 0.006574058439582586, + 0.006897834129631519, + 0.010929159820079803, + 0.014350582845509052, + 0.011791247874498367, + 0.009051145985722542, + 0.0078111570328474045, + 0.012538931332528591, + 0.007445497903972864, + 0.008133137598633766, + 0.036112230271101, + 0.01076128613203764, + 0.007799732498824596, + 0.005919183138757944, + 0.010689047165215015, + 0.03629973903298378, + 0.0052168103866279125, + 0.008641121909022331, + 0.01230609230697155, + 0.010191605426371098, + 0.007252801675349474, + 0.008292648941278458, + 0.0066601610742509365, + 0.011107835918664932, + 0.006766301114112139, + 0.006795221474021673, + 0.009043036960065365, + 0.007518500089645386, + 0.013821631669998169, + 0.0093365628272295, + 0.008700293488800526, + 0.013548106886446476, + 0.007653472013771534, + 0.013953824527561665, + 0.012216610834002495, + 0.01151325460523367, + 0.010383623652160168, + 0.012062281370162964, + 0.016346313059329987, + 0.009840494021773338, + 0.010835684835910797, + 0.011281780898571014, + 0.0047658286057412624, + 0.007896517403423786, + 0.00420027831569314, + 0.0032780508045107126, + 0.008341706357896328, + 0.003425926435738802, + 0.0026679958682507277, + 0.004136059898883104, + 0.024562444537878036, + 0.010619350709021091, + 0.0255444198846817, + 0.006528384052217007, + 0.00991163868457079, + 0.02558964304625988, + 0.00688745453953743, + 0.010495462454855442, + 0.011196399107575417, + 0.05587032064795494, + 0.041540224105119705, + 0.004252128303050995, + 0.009276293218135834, + 0.010373225435614586, + 0.002776841167360544, + 0.003945034928619862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.1129763200879097 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521856725215912 + ], + "opposite_logprob_tokens": [ + -1.5021857023239136 + ], + "logprob_reward": -0.2521856725215912, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521856725215912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3875584304332733, + 0.052926789969205856, + 0.03295661509037018, + 0.07193640619516373, + 0.03897034749388695, + 0.11515696346759796, + 0.021208452060818672, + 0.03045748732984066, + 0.09611722826957703, + 0.15271122753620148 + ], + "credit_map": [ + 0.06462490558624268, + 0.04798167198896408, + 0.0414990559220314, + 0.05244208872318268, + 0.040235865861177444, + 0.07778318226337433, + 0.030947698280215263, + 0.03204397112131119, + 0.011185718700289726, + 0.005977929569780827, + 0.004104457795619965, + 0.006713171023875475, + 0.003942685667425394, + 0.010650642216205597, + 0.006485324818640947, + 0.003866857849061489, + 0.007113068364560604, + 0.004137356765568256, + 0.002479221671819687, + 0.004116889089345932, + 0.002913933014497161, + 0.005914006382226944, + 0.003742047818377614, + 0.0025400924496352673, + 0.023870324715971947, + 0.005308416206389666, + 0.004613835830241442, + 0.004552016966044903, + 0.0047933124005794525, + 0.01960846595466137, + 0.0032330555841326714, + 0.005956979002803564, + 0.004722972866147757, + 0.003508046269416809, + 0.005812869872897863, + 0.005230019800364971, + 0.004613223019987345, + 0.005875027272850275, + 0.003964669071137905, + 0.005243518389761448, + 0.018376024439930916, + 0.009803432039916515, + 0.01583131216466427, + 0.010761240497231483, + 0.011438089422881603, + 0.01799842156469822, + 0.01029218826442957, + 0.020656250417232513, + 0.0024716495536267757, + 0.0019816963467746973, + 0.0021773057524114847, + 0.0017844837857410312, + 0.0037892882246524096, + 0.0018972090911120176, + 0.0018382556736469269, + 0.005268564447760582, + 0.007304297294467688, + 0.004656694363802671, + 0.005250599700957537, + 0.0019473277498036623, + 0.005052579566836357, + 0.0022246004082262516, + 0.0016221627593040466, + 0.0023992243222892284, + 0.01564755104482174, + 0.010875117033720016, + 0.011775245890021324, + 0.01158301904797554, + 0.008045714348554611, + 0.01693381741642952, + 0.007492457050830126, + 0.013764308765530586, + 0.03164216876029968, + 0.060133207589387894, + 0.008844037540256977, + 0.006877020467072725, + 0.012756559066474438, + 0.023077985271811485, + 0.004138628486543894, + 0.005241623148322105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.2521856725215912 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059635089710354805 + ], + "opposite_logprob_tokens": [ + -5.130963325500488 + ], + "logprob_reward": -0.0059635089710354805, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059635089710354805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.500446617603302, + 0.05241119861602783, + 0.015291832387447357, + 0.08225516974925995, + 0.08263634890317917, + 0.04042685404419899, + 0.04610688239336014, + 0.018114972859621048, + 0.09031838178634644, + 0.07199173420667648 + ], + "credit_map": [ + 0.07398055493831635, + 0.06164541095495224, + 0.05188377574086189, + 0.07120528817176819, + 0.0571574829518795, + 0.09527508169412613, + 0.04489197954535484, + 0.04440706968307495, + 0.0088412556797266, + 0.006621863692998886, + 0.006318710278719664, + 0.008070315234363079, + 0.004982120357453823, + 0.006546322256326675, + 0.005686054937541485, + 0.0053445580415427685, + 0.002543375128880143, + 0.0019941057544201612, + 0.0020686958450824022, + 0.0018102200701832771, + 0.0015921028098091483, + 0.001717834617011249, + 0.001646741176955402, + 0.0019187562866136432, + 0.02433033287525177, + 0.0073098232969641685, + 0.005456503015011549, + 0.0047887531109154224, + 0.007254555821418762, + 0.02438705414533615, + 0.003243854967877269, + 0.005484290421009064, + 0.010686607100069523, + 0.011412359774112701, + 0.008478078059852123, + 0.011376782320439816, + 0.010935612954199314, + 0.011734949424862862, + 0.007732993923127651, + 0.010278962552547455, + 0.004380719270557165, + 0.00328838755376637, + 0.0076391492038965225, + 0.004182031378149986, + 0.004453515633940697, + 0.0069929067976772785, + 0.0033245296217501163, + 0.006165617611259222, + 0.005982625763863325, + 0.004637658130377531, + 0.005565370433032513, + 0.0056807491928339005, + 0.006946123205125332, + 0.004478655755519867, + 0.005627728067338467, + 0.007187970448285341, + 0.004572473466396332, + 0.001987356459721923, + 0.0036145001649856567, + 0.0015419593546539545, + 0.001844943966716528, + 0.0017642676830291748, + 0.0010509527055546641, + 0.0017385189421474934, + 0.014179329387843609, + 0.010357950814068317, + 0.010006234981119633, + 0.012455230578780174, + 0.008061271160840988, + 0.014213445596396923, + 0.008756416849792004, + 0.012288493104279041, + 0.01289465930312872, + 0.030823610723018646, + 0.010122456587851048, + 0.0024830177426338196, + 0.00589714152738452, + 0.006053429562598467, + 0.0017484138952568173, + 0.0019690068438649178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.0059635089710354805 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning close-up, completely out of focus at first, reveals a round, anthropomorphic cream cake resting peacefully in a soft, fairytale-like halo. Dreamy morning sunlight filters through the window, gilding the scene with a warm golden hue and illuminating the delicate, pastel-colored ceramic plate beneath. As it slowly awakens, the thin blanket of pink frosting gently slips away, revealing a fluffy, pure white cream body beneath. At this moment, accompanied by a subtle, magical flicker on the screen, the lens's focus smoothly tightens from blur to focus sharply on the cake's face. Finally, focus settles precisely on its two dark chocolate shavings, which serve as eyes. They shimmer faintly in the soft light, then, with a hint of sleep, magically and slowly open, breathing the first glimmer of consciousness into this sweet creature. The entire image is permeated with a tranquil, miraculous awakening.", + "seed": 17184, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_6/index_3172_seed_17184_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.17537467144429683, + "vlm_logprob": -0.17537467144429683, + "vlm_yes_no_margin": 2.875, + "vlm_token_credit": -0.17537467144429683, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033457640558481216 + ], + "opposite_logprob_tokens": [ + -3.5334577560424805 + ], + "logprob_reward": -0.033457640558481216, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.033457640558481216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13660316169261932, + 0.07142101973295212, + 0.08605315536260605, + 0.13000358641147614, + 0.08574407547712326, + 0.05808235704898834, + 0.08481039851903915, + 0.12062226235866547, + 0.0816860944032669, + 0.14497387409210205 + ], + "credit_map": [ + 0.02025979943573475, + 0.012403929606080055, + 0.022008953616023064, + 0.016091816127300262, + 0.011495168320834637, + 0.014509974047541618, + 0.019839448854327202, + 0.019994067028164864, + 0.0068903034552931786, + 0.012691532261669636, + 0.011667958460748196, + 0.008421140722930431, + 0.008574803359806538, + 0.006860893685370684, + 0.00821380503475666, + 0.008100581355392933, + 0.010622515343129635, + 0.006918570026755333, + 0.012827885337173939, + 0.011862716637551785, + 0.005427115131169558, + 0.01076505146920681, + 0.009956181980669498, + 0.01767311431467533, + 0.015115409158170223, + 0.020259428769350052, + 0.019549261778593063, + 0.01741304248571396, + 0.01161085907369852, + 0.014532915316522121, + 0.01520908810198307, + 0.0163135826587677, + 0.011269721202552319, + 0.00953582301735878, + 0.015467654913663864, + 0.009652252309024334, + 0.007025782018899918, + 0.015529648400843143, + 0.009001764468848705, + 0.008261434733867645, + 0.01027646567672491, + 0.0058937896974384785, + 0.0076368036679923534, + 0.0059714289382100105, + 0.005091200582683086, + 0.007258087862282991, + 0.008746663108468056, + 0.007207919377833605, + 0.010332431644201279, + 0.0099822161719203, + 0.012260088697075844, + 0.009924370795488358, + 0.007974195294082165, + 0.008387355133891106, + 0.012496358714997768, + 0.013453378342092037, + 0.015712972730398178, + 0.009670360013842583, + 0.01935642585158348, + 0.006372987292706966, + 0.009688770398497581, + 0.02362244762480259, + 0.009487592615187168, + 0.02671070583164692, + 0.011669810861349106, + 0.010079555213451385, + 0.010236461646854877, + 0.0072627379558980465, + 0.00842333398759365, + 0.016235630959272385, + 0.008706264197826385, + 0.009072299115359783, + 0.021999429911375046, + 0.012404417619109154, + 0.018235046416521072, + 0.007352497894316912, + 0.008410206064581871, + 0.035165850073099136, + 0.009261169470846653, + 0.03214527294039726 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.033457640558481216 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579432129859924 + ], + "opposite_logprob_tokens": [ + -0.6329432129859924 + ], + "logprob_reward": -0.7579432129859924, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579432129859924, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3638746440410614, + 0.057950910180807114, + 0.09406189620494843, + 0.05126151069998741, + 0.03407427668571472, + 0.05134233459830284, + 0.08102792501449585, + 0.10098176449537277, + 0.06051861122250557, + 0.1049060970544815 + ], + "credit_map": [ + 0.06656070053577423, + 0.036081571131944656, + 0.051207538694143295, + 0.041203320026397705, + 0.03302011638879776, + 0.03843868151307106, + 0.041623879224061966, + 0.055738843977451324, + 0.006479428615421057, + 0.006893174257129431, + 0.00777788320556283, + 0.008790269494056702, + 0.0047520785592496395, + 0.010040405206382275, + 0.007251200266182423, + 0.005966469645500183, + 0.00897719245404005, + 0.010176009498536587, + 0.01307816430926323, + 0.019217897206544876, + 0.00813734158873558, + 0.007967483252286911, + 0.017311139032244682, + 0.009196673519909382, + 0.006935075391083956, + 0.00629477808251977, + 0.007628684397786856, + 0.006335049867630005, + 0.004111040849238634, + 0.006557981483638287, + 0.006709962617605925, + 0.00668893801048398, + 0.003963402472436428, + 0.003955456428229809, + 0.005738033447414637, + 0.004332102835178375, + 0.0030264384113252163, + 0.00459137512370944, + 0.004534008447080851, + 0.003933460917323828, + 0.006582722067832947, + 0.005847749300301075, + 0.007423610892146826, + 0.006012704689055681, + 0.004707543645054102, + 0.00768042728304863, + 0.005564764607697725, + 0.007522813975811005, + 0.012007959187030792, + 0.012666753493249416, + 0.009940482676029205, + 0.004417512100189924, + 0.006146642845124006, + 0.01590123400092125, + 0.0057536945678293705, + 0.014193642884492874, + 0.014308660291135311, + 0.013491055928170681, + 0.012897627428174019, + 0.006479127332568169, + 0.01100087445229292, + 0.021339507773518562, + 0.007062200922518969, + 0.014402705244719982, + 0.006060568615794182, + 0.008528236299753189, + 0.009781571105122566, + 0.005893408320844173, + 0.006992652080953121, + 0.008518890477716923, + 0.006455262191593647, + 0.008288022130727768, + 0.01489153504371643, + 0.02051769383251667, + 0.013364024460315704, + 0.006498525384813547, + 0.01111531164497137, + 0.014095067046582699, + 0.005501894745975733, + 0.018922047689557076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.7579432129859924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043103136122226715 + ], + "opposite_logprob_tokens": [ + -3.1681032180786133 + ], + "logprob_reward": -0.043103136122226715, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043103136122226715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15222088992595673, + 0.02794656902551651, + 0.05549982190132141, + 0.03435337170958519, + 0.04392459616065025, + 0.058980487287044525, + 0.11331865191459656, + 0.2254405915737152, + 0.1985945701599121, + 0.0897204726934433 + ], + "credit_map": [ + 0.016133109107613564, + 0.016151199117302895, + 0.02329491451382637, + 0.02292042225599289, + 0.01703544147312641, + 0.01603359542787075, + 0.02194029651582241, + 0.018711907789111137, + 0.0029948498122394085, + 0.0038676548283547163, + 0.0036241200286895037, + 0.003124590264633298, + 0.0024891579523682594, + 0.004293945152312517, + 0.0035210242494940758, + 0.004031226504594088, + 0.006045108661055565, + 0.007579922676086426, + 0.007595357950776815, + 0.007333376910537481, + 0.0040442622266709805, + 0.005301088560372591, + 0.006404974032193422, + 0.011195734143257141, + 0.004428492859005928, + 0.004030696582049131, + 0.004860751796513796, + 0.004504628479480743, + 0.003132604993879795, + 0.0036385641433298588, + 0.005016234703361988, + 0.004741399548947811, + 0.005687595810741186, + 0.004613760858774185, + 0.006535540334880352, + 0.005489401053637266, + 0.004112750757485628, + 0.004785760771483183, + 0.004950168076902628, + 0.007749619893729687, + 0.007403015159070492, + 0.005199273582547903, + 0.008200027979910374, + 0.006288912612944841, + 0.00472501665353775, + 0.010186009109020233, + 0.006321295630186796, + 0.010656935162842274, + 0.019553475081920624, + 0.01225264836102724, + 0.01416211947798729, + 0.009137860499322414, + 0.008115466684103012, + 0.025125151500105858, + 0.011099221184849739, + 0.013872707262635231, + 0.030455652624368668, + 0.025597434490919113, + 0.032078780233860016, + 0.011749906465411186, + 0.016543814912438393, + 0.043725837022066116, + 0.014702517539262772, + 0.05058666318655014, + 0.02572895586490631, + 0.01637939177453518, + 0.02568286843597889, + 0.014811564236879349, + 0.013516832143068314, + 0.05132386460900307, + 0.014528082683682442, + 0.03662301227450371, + 0.012515898793935776, + 0.008538106456398964, + 0.011262942105531693, + 0.004211351741105318, + 0.00555839017033577, + 0.020481763407588005, + 0.005369140300899744, + 0.021782875061035156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.043103136122226715 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255878061056137 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255878061056137, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255878061056137, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07394225895404816, + 0.06251057237386703, + 0.039455343037843704, + 0.06937835365533829, + 0.04459420591592789, + 0.0322820208966732, + 0.09189476072788239, + 0.19912207126617432, + 0.24675385653972626, + 0.14006660878658295 + ], + "credit_map": [ + 0.0072426083497703075, + 0.006750750821083784, + 0.009011623449623585, + 0.01283060573041439, + 0.0083773173391819, + 0.006848201155662537, + 0.01647087372839451, + 0.006410268135368824, + 0.007130562327802181, + 0.005876047536730766, + 0.0066922502592206, + 0.010678858496248722, + 0.006097985897213221, + 0.005668823607265949, + 0.010132006369531155, + 0.010234036482870579, + 0.004630230832844973, + 0.004246784374117851, + 0.005660244729369879, + 0.007161022629588842, + 0.003562329337000847, + 0.003437834559008479, + 0.007355219218879938, + 0.0034016731660813093, + 0.00880370195955038, + 0.006257132161408663, + 0.01018684171140194, + 0.007541828788816929, + 0.0047132764011621475, + 0.00971041526645422, + 0.009444794617593288, + 0.012720359489321709, + 0.006006124895066023, + 0.004407753236591816, + 0.0067258733324706554, + 0.005906815640628338, + 0.004458962939679623, + 0.0054733287543058395, + 0.0061966292560100555, + 0.005418713670223951, + 0.003949729260057211, + 0.0036307789850980043, + 0.004280034452676773, + 0.004349490161985159, + 0.0033606202341616154, + 0.0033733746968209743, + 0.004902257118374109, + 0.004435733892023563, + 0.013249058276414871, + 0.010950570926070213, + 0.013505237177014351, + 0.007907803170382977, + 0.0074880076572299, + 0.014749926514923573, + 0.009895198978483677, + 0.014148948714137077, + 0.026668235659599304, + 0.025029264390468597, + 0.026508420705795288, + 0.013257610611617565, + 0.014786044135689735, + 0.03715537488460541, + 0.014314338564872742, + 0.04140276834368706, + 0.03885018825531006, + 0.029347527772188187, + 0.03430188074707985, + 0.013023053295910358, + 0.02072983793914318, + 0.056020159274339676, + 0.01693885773420334, + 0.03754232078790665, + 0.017298897728323936, + 0.014107385650277138, + 0.01841600239276886, + 0.008695174939930439, + 0.009797834791243076, + 0.02785453386604786, + 0.011085192672908306, + 0.03281158208847046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.01255878061056137 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a round, anthropomorphic cream cake with a white cream body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cake rest on a delicate, pastel-colored ceramic plate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a thin blanket of pink frosting slip away to reveal the cake's body?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the cake's eyes, made of dark chocolate shavings, slowly open?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera focus smoothly tighten from completely out of focus to sharp focus on the cake's face?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981058694422245 + ], + "opposite_logprob_tokens": [ + -3.529810667037964 + ], + "logprob_reward": -0.02981058694422245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981058694422245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4239683449268341, + 0.045153070241212845, + 0.03432588651776314, + 0.037049803882837296, + 0.03131439536809921, + 0.04652022942900658, + 0.07261548936367035, + 0.08328951895236969, + 0.11075817793607712, + 0.11500503122806549 + ], + "credit_map": [ + 0.05924971029162407, + 0.05039747431874275, + 0.07424046099185944, + 0.04240815341472626, + 0.03615470230579376, + 0.04872080683708191, + 0.049940720200538635, + 0.06285633146762848, + 0.004641480278223753, + 0.004737545736134052, + 0.0050925323739647865, + 0.007526636589318514, + 0.005068871658295393, + 0.005593591835349798, + 0.008351611904799938, + 0.004140801262110472, + 0.0037214504554867744, + 0.005065282806754112, + 0.005564992316067219, + 0.004086518660187721, + 0.0033676563762128353, + 0.0036002390552312136, + 0.0037191766314208508, + 0.005200569983571768, + 0.00438361382111907, + 0.0048665148206055164, + 0.00495940400287509, + 0.004943526815623045, + 0.00340921594761312, + 0.0041335029527544975, + 0.005058047827333212, + 0.005295976065099239, + 0.0036948611959815025, + 0.002944683888927102, + 0.005695013795047998, + 0.003661811351776123, + 0.0028596112970262766, + 0.004481058567762375, + 0.003814146388322115, + 0.004163210280239582, + 0.005711630918085575, + 0.004099830985069275, + 0.006191571708768606, + 0.004608163144439459, + 0.0032760242465883493, + 0.008679506368935108, + 0.006428624968975782, + 0.007524876389652491, + 0.01201039832085371, + 0.0070999907329678535, + 0.00949909258633852, + 0.005600086413323879, + 0.004656596574932337, + 0.014317983761429787, + 0.00725018884986639, + 0.012181155383586884, + 0.013085268437862396, + 0.010697754099965096, + 0.011184931732714176, + 0.0054438128136098385, + 0.00648523960262537, + 0.015158135443925858, + 0.00645441422238946, + 0.014779963530600071, + 0.012497497722506523, + 0.010440579615533352, + 0.01782228611409664, + 0.009515425190329552, + 0.009726839140057564, + 0.019147302955389023, + 0.01142212189733982, + 0.02018612064421177, + 0.014455589465796947, + 0.011033937335014343, + 0.014957322739064693, + 0.0064003863371908665, + 0.006527002435177565, + 0.02611519955098629, + 0.0076109254732728004, + 0.027904663234949112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.02981058694422245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_tensor([9813, 9813])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_tensor([9813, 9813])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e847c28705791b87ba2c88b1d3b0347a17b3decd --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_tensor([9813, 9813])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up shot, tilting slowly upward from an extremely low angle, focuses on a set of ancient bluestone steps, drenched in the scent of dampness. Countless slender, emerald green vines cling tenaciously upwards, their tender green tendrils like tiny tentacles, cautiously exploring and clinging to the rough surface of the stone. Glistening water droplets hang from their leaves. These vines have formed a dense green carpet, almost completely covering the steps, originally a dark blue-gray hue due to moisture. Only in the gaps between the interwoven vines do the mottled marks of age-worn stone appear. Soft, diffused light streams down from above, revealing the veins of each leaf and causing the water droplets to shimmer with a faint, cool glow. The entire scene, framed by a cool palette, exudes an atmosphere of profound tranquility and unyielding vitality.", + "seed": 26727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26727_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017142238398082554, + "vlm_logprob": -0.017142238398082554, + "vlm_yes_no_margin": 4.725, + "vlm_token_credit": -0.017142238398082554, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02283422276377678 + ], + "opposite_logprob_tokens": [ + -4.022834300994873 + ], + "logprob_reward": -0.02283422276377678, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.02283422276377678, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23601849377155304, + 0.3730751872062683, + 0.037407469004392624, + 0.06953079998493195, + 0.03998139873147011, + 0.029856441542506218, + 0.05107330158352852, + 0.026995310559868813, + 0.10411514341831207, + 0.03194652497768402 + ], + "credit_map": [ + 0.024701140820980072, + 0.017540737986564636, + 0.046426527202129364, + 0.03251327574253082, + 0.013917016796767712, + 0.013678262941539288, + 0.03334801271557808, + 0.053893495351076126, + 0.053990717977285385, + 0.04124894365668297, + 0.06345278024673462, + 0.0902729332447052, + 0.0312091875821352, + 0.028686197474598885, + 0.03253830969333649, + 0.03167606145143509, + 0.0043304734863340855, + 0.006721366662532091, + 0.0043912907131016254, + 0.006057937163859606, + 0.004308931063860655, + 0.003810221329331398, + 0.00483887642621994, + 0.00294837006367743, + 0.006487567909061909, + 0.004357459954917431, + 0.01488599181175232, + 0.0033772096503525972, + 0.023214440792798996, + 0.008923540823161602, + 0.0052620600908994675, + 0.003022521734237671, + 0.004819116555154324, + 0.004812164697796106, + 0.004217312205582857, + 0.005379874724894762, + 0.005487666931003332, + 0.005871105007827282, + 0.006023702677339315, + 0.003370449412614107, + 0.004437165334820747, + 0.004410667810589075, + 0.004888370167464018, + 0.003181885462254286, + 0.003422203939408064, + 0.0032568564638495445, + 0.0032034555915743113, + 0.0030558337457478046, + 0.005959928035736084, + 0.007751571945846081, + 0.009223481640219688, + 0.009041845798492432, + 0.0037153323646634817, + 0.005946330726146698, + 0.005791566334664822, + 0.0036432400811463594, + 0.003973783925175667, + 0.0054837241768836975, + 0.004008663818240166, + 0.00396063132211566, + 0.0018907277844846249, + 0.0025584266986697912, + 0.0029824774246662855, + 0.0021368733141571283, + 0.007766578812152147, + 0.02510509267449379, + 0.00807327963411808, + 0.004827824886888266, + 0.006611649412661791, + 0.038850314915180206, + 0.007249440066516399, + 0.005630945786833763, + 0.0037481633480638266, + 0.004665077663958073, + 0.003957487642765045, + 0.003626330755650997, + 0.004618672654032707, + 0.004465619567781687, + 0.00342967314645648, + 0.003435496473684907 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.02283422276377678 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013522299705073237 + ], + "opposite_logprob_tokens": [ + -6.626352310180664 + ], + "logprob_reward": -0.0013522299705073237, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013522299705073237, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14356154203414917, + 0.23112444579601288, + 0.040441807359457016, + 0.10322462767362595, + 0.06985767185688019, + 0.08086491376161575, + 0.05346071347594261, + 0.05788498371839523, + 0.18533940613269806, + 0.034239958971738815 + ], + "credit_map": [ + 0.019153617322444916, + 0.01607516221702099, + 0.015998778864741325, + 0.01874873787164688, + 0.013849332928657532, + 0.018232233822345734, + 0.02083185315132141, + 0.02067181095480919, + 0.03128302842378616, + 0.018052702769637108, + 0.05619436129927635, + 0.04529745504260063, + 0.018172895535826683, + 0.02194930799305439, + 0.020551882684230804, + 0.019622797146439552, + 0.004534902982413769, + 0.0047230650670826435, + 0.004254112020134926, + 0.005028874147683382, + 0.005979635752737522, + 0.005134949926286936, + 0.006972770672291517, + 0.003813492599874735, + 0.007727969903498888, + 0.006198431830853224, + 0.023373255506157875, + 0.005765627138316631, + 0.03273595869541168, + 0.014404179528355598, + 0.008130578324198723, + 0.004888612776994705, + 0.005662259180098772, + 0.008609787560999393, + 0.006380891893059015, + 0.014033869840204716, + 0.010321523062884808, + 0.01036104653030634, + 0.009279008023440838, + 0.005209276452660561, + 0.011006666347384453, + 0.012698904611170292, + 0.010786321014165878, + 0.00868260208517313, + 0.011255746707320213, + 0.009674977511167526, + 0.008308710530400276, + 0.008450976572930813, + 0.0059668743051588535, + 0.01086471788585186, + 0.006265214644372463, + 0.0061379666440188885, + 0.005454433150589466, + 0.005853496491909027, + 0.005382657516747713, + 0.007535346783697605, + 0.007235564757138491, + 0.009578749537467957, + 0.007174806669354439, + 0.008668162859976292, + 0.005566990934312344, + 0.006879879627376795, + 0.005550938658416271, + 0.0072298841550946236, + 0.03097163513302803, + 0.01594839058816433, + 0.01169163454324007, + 0.009169191122055054, + 0.015555577352643013, + 0.06334587931632996, + 0.019098855555057526, + 0.019558219239115715, + 0.0033395010977983475, + 0.003666122443974018, + 0.0063644889742136, + 0.0036667610984295607, + 0.00342589826323092, + 0.005745396483689547, + 0.0032192228827625513, + 0.004812566097825766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0013522299705073237 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020602159202098846 + ], + "opposite_logprob_tokens": [ + -3.895602226257324 + ], + "logprob_reward": -0.020602159202098846, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020602159202098846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19602283835411072, + 0.1465049833059311, + 0.07286754250526428, + 0.10496192425489426, + 0.06705910712480545, + 0.042834293097257614, + 0.038192521780729294, + 0.13981465995311737, + 0.12121298164129257, + 0.07052914053201675 + ], + "credit_map": [ + 0.024109072983264923, + 0.021508650854229927, + 0.02801772952079773, + 0.029491793364286423, + 0.015605391934514046, + 0.01789765991270542, + 0.026416493579745293, + 0.032976046204566956, + 0.017586659640073776, + 0.018671074882149696, + 0.022542456164956093, + 0.02838299609720707, + 0.015650801360607147, + 0.013084596022963524, + 0.012561803683638573, + 0.018024593591690063, + 0.006645208224654198, + 0.008203827776014805, + 0.008350907824933529, + 0.011279581114649773, + 0.009744013659656048, + 0.009256801567971706, + 0.011251593008637428, + 0.008135611191391945, + 0.01415051706135273, + 0.008336623199284077, + 0.012985499575734138, + 0.004935064353048801, + 0.037027377635240555, + 0.01394214853644371, + 0.008775320835411549, + 0.0048093716613948345, + 0.004884858150035143, + 0.007902336306869984, + 0.004874942824244499, + 0.020496530458331108, + 0.010776156559586525, + 0.007105785422027111, + 0.007251107133924961, + 0.003767387242987752, + 0.006445653736591339, + 0.006665444932878017, + 0.00565582700073719, + 0.0045290011912584305, + 0.005402158480137587, + 0.0060860528610646725, + 0.0042149946093559265, + 0.0038351588882505894, + 0.0033025010488927364, + 0.00637884670868516, + 0.0042525045573711395, + 0.005007066763937473, + 0.005226676817983389, + 0.005022094585001469, + 0.003526311833411455, + 0.005476519465446472, + 0.015488808043301105, + 0.020823443308472633, + 0.017221471294760704, + 0.01654125563800335, + 0.01967398077249527, + 0.015294382348656654, + 0.016368845477700233, + 0.01840248890221119, + 0.011757253669202328, + 0.02523029036819935, + 0.014214206486940384, + 0.007934230379760265, + 0.007521153427660465, + 0.03754446282982826, + 0.008635219186544418, + 0.008376169949769974, + 0.00742629449814558, + 0.015822267159819603, + 0.008695663884282112, + 0.0065738423727452755, + 0.006981448736041784, + 0.00933463778346777, + 0.006909848656505346, + 0.00878513790667057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020602159202098846 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809833154082298 + ], + "opposite_logprob_tokens": [ + -3.2880983352661133 + ], + "logprob_reward": -0.03809833154082298, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809833154082298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14449861645698547, + 0.17734748125076294, + 0.022231224924325943, + 0.06675251573324203, + 0.03902595490217209, + 0.01772272400557995, + 0.08832868188619614, + 0.03975934162735939, + 0.19291044771671295, + 0.21142297983169556 + ], + "credit_map": [ + 0.01704546809196472, + 0.02090127021074295, + 0.017385749146342278, + 0.01834162324666977, + 0.013782861642539501, + 0.015508368611335754, + 0.020557986572384834, + 0.02097529172897339, + 0.020568380132317543, + 0.018628759309649467, + 0.026829944923520088, + 0.039482783526182175, + 0.023367075249552727, + 0.009298553690314293, + 0.016336515545845032, + 0.02283545956015587, + 0.003991207107901573, + 0.0031639363151043653, + 0.0031530719716101885, + 0.0023844409734010696, + 0.002707300242036581, + 0.002307879040017724, + 0.0025097487960010767, + 0.002013640245422721, + 0.005804082378745079, + 0.004954170901328325, + 0.01499889511615038, + 0.005319491028785706, + 0.020967567339539528, + 0.0070466347970068455, + 0.004776030778884888, + 0.0028856440912932158, + 0.005663439631462097, + 0.005826091859489679, + 0.0052893501706421375, + 0.004971937742084265, + 0.004467810969799757, + 0.0043418207205832005, + 0.0049670846201479435, + 0.0034984201192855835, + 0.001735122874379158, + 0.002075683558359742, + 0.0022939706686884165, + 0.0022194136399775743, + 0.002038710517808795, + 0.0025643601547926664, + 0.0024300741497427225, + 0.0023653891403228045, + 0.011599874123930931, + 0.01004659105092287, + 0.013885785825550556, + 0.01755775697529316, + 0.007411833852529526, + 0.011421669274568558, + 0.007758606690913439, + 0.008646568283438683, + 0.004910374991595745, + 0.006700798869132996, + 0.004815602675080299, + 0.006149789784103632, + 0.0035057803615927696, + 0.0036469190381467342, + 0.00466720387339592, + 0.005362872965633869, + 0.029816562309861183, + 0.021079670637845993, + 0.019259905442595482, + 0.013898610137403011, + 0.01820669136941433, + 0.047037020325660706, + 0.022482039406895638, + 0.021129947155714035, + 0.024325987324118614, + 0.07066608220338821, + 0.0203841645270586, + 0.016015207394957542, + 0.017953598871827126, + 0.02468869648873806, + 0.02444235235452652, + 0.012946903705596924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.03809833154082298 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028242485132068396 + ], + "opposite_logprob_tokens": [ + -5.877824306488037 + ], + "logprob_reward": -0.0028242485132068396, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028242485132068396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22050334513187408, + 0.11159609258174896, + 0.06356725841760635, + 0.06172134354710579, + 0.05981108546257019, + 0.04937439784407616, + 0.030316093936562538, + 0.08716734498739243, + 0.15236696600914001, + 0.16357606649398804 + ], + "credit_map": [ + 0.030451228842139244, + 0.020793065428733826, + 0.030004194006323814, + 0.03374672681093216, + 0.014898994006216526, + 0.018133100122213364, + 0.03320543095469475, + 0.039270590990781784, + 0.01250537484884262, + 0.009816469624638557, + 0.020377840846776962, + 0.015390065498650074, + 0.021226244047284126, + 0.009805052541196346, + 0.008291549049317837, + 0.014183487743139267, + 0.00877802912145853, + 0.008138268254697323, + 0.006095650140196085, + 0.010371868498623371, + 0.007761859335005283, + 0.0076818219386041164, + 0.00863491278141737, + 0.006104841362684965, + 0.005080122966319323, + 0.003652840619906783, + 0.01673288084566593, + 0.004694978706538677, + 0.018105527386069298, + 0.006012463942170143, + 0.004489265847951174, + 0.0029532562475651503, + 0.008300972171127796, + 0.006128013599663973, + 0.0059291222132742405, + 0.01205811370164156, + 0.007939075119793415, + 0.007821372710168362, + 0.007398663554340601, + 0.004235747270286083, + 0.006881998851895332, + 0.006927831564098597, + 0.007718055974692106, + 0.005804726853966713, + 0.006214946508407593, + 0.005902198608964682, + 0.00503920204937458, + 0.00488543463870883, + 0.0041680471040308475, + 0.005192638374865055, + 0.004401949234306812, + 0.003786795772612095, + 0.0032008688431233168, + 0.0025872711557894945, + 0.0028279051184654236, + 0.004150615073740482, + 0.01023498922586441, + 0.014388985000550747, + 0.010510602034628391, + 0.011783435940742493, + 0.010424081236124039, + 0.008326775394380093, + 0.010127303190529346, + 0.011371171101927757, + 0.020982790738344193, + 0.023239795118570328, + 0.008975893259048462, + 0.006949799135327339, + 0.00921877846121788, + 0.058154281228780746, + 0.01231739204376936, + 0.01252821832895279, + 0.0228583961725235, + 0.05712791159749031, + 0.013639788143336773, + 0.012953273020684719, + 0.01257109735161066, + 0.013701045885682106, + 0.01183390524238348, + 0.018890632316470146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0028242485132068396 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly upward from an extremely low angle, focuses on a set of ancient bluestone steps, drenched in the scent of dampness. Countless slender, emerald green vines cling tenaciously upwards, their tender green tendrils like tiny tentacles, cautiously exploring and clinging to the rough surface of the stone. Glistening water droplets hang from their leaves. These vines have formed a dense green carpet, almost completely covering the steps, originally a dark blue-gray hue due to moisture. Only in the gaps between the interwoven vines do the mottled marks of age-worn stone appear. Soft, diffused light streams down from above, revealing the veins of each leaf and causing the water droplets to shimmer with a faint, cool glow. The entire scene, framed by a cool palette, exudes an atmosphere of profound tranquility and unyielding vitality.", + "seed": 26728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26728_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.028904745320323853, + "vlm_logprob": -0.028904745320323853, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.028904745320323853, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02853061631321907 + ], + "opposite_logprob_tokens": [ + -3.7785305976867676 + ], + "logprob_reward": -0.02853061631321907, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02853061631321907, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10416178405284882, + 0.07236713171005249, + 0.04965256154537201, + 0.05254475772380829, + 0.017065370455384254, + 0.02566823549568653, + 0.06032811105251312, + 0.24156896770000458, + 0.06663033366203308, + 0.31001272797584534 + ], + "credit_map": [ + 0.011323116719722748, + 0.011996562592685223, + 0.01457169372588396, + 0.011954489164054394, + 0.014264944940805435, + 0.0100501524284482, + 0.016325324773788452, + 0.013675505295395851, + 0.010644644498825073, + 0.014412867836654186, + 0.011773607693612576, + 0.006478377617895603, + 0.006357405334711075, + 0.006662122905254364, + 0.005369854159653187, + 0.010668253526091576, + 0.005646880716085434, + 0.012309134937822819, + 0.005672604776918888, + 0.006850350648164749, + 0.005103942006826401, + 0.0025677874218672514, + 0.00717926537618041, + 0.004322598688304424, + 0.008265084587037563, + 0.006113545503467321, + 0.0061693619936704636, + 0.005105490796267986, + 0.007189603988081217, + 0.005785508546978235, + 0.007155401166528463, + 0.006760766729712486, + 0.002241322537884116, + 0.003009287640452385, + 0.0019560314249247313, + 0.0016501895152032375, + 0.0020926434081047773, + 0.0018247717525810003, + 0.0015683089150115848, + 0.0027228170074522495, + 0.003278580727055669, + 0.00386029202491045, + 0.0037398941349238157, + 0.002458068775013089, + 0.0031979382038116455, + 0.0027943113818764687, + 0.00269197765737772, + 0.003647172823548317, + 0.010535449720919132, + 0.007649459410458803, + 0.008674118667840958, + 0.007558399811387062, + 0.0074780844151973724, + 0.006321607157588005, + 0.00636229058727622, + 0.005748704541474581, + 0.011677805334329605, + 0.05023001506924629, + 0.014047444798052311, + 0.008167610503733158, + 0.09817543625831604, + 0.04664352536201477, + 0.00665722880512476, + 0.005969902500510216, + 0.005888517014682293, + 0.010228187777101994, + 0.00939071923494339, + 0.006734877824783325, + 0.012567763216793537, + 0.008976026438176632, + 0.005331347696483135, + 0.007512898184359074, + 0.027990208938717842, + 0.0696331262588501, + 0.0203109048306942, + 0.029878679662942886, + 0.07308245450258255, + 0.04653081297874451, + 0.025087833404541016, + 0.017498726025223732 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.02853061631321907 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009386900928802788 + ], + "opposite_logprob_tokens": [ + -7.000938892364502 + ], + "logprob_reward": -0.0009386900928802788, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009386900928802788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21414485573768616, + 0.053337302058935165, + 0.07743383944034576, + 0.032504674047231674, + 0.03466411679983139, + 0.06845688819885254, + 0.09911100566387177, + 0.12437418848276138, + 0.11504027247428894, + 0.18093284964561462 + ], + "credit_map": [ + 0.026354048401117325, + 0.03549939766526222, + 0.023671306669712067, + 0.02452836185693741, + 0.02512812241911888, + 0.020380299538373947, + 0.03108035773038864, + 0.02750297077000141, + 0.006840023212134838, + 0.01107198279350996, + 0.008149866946041584, + 0.006519573274999857, + 0.006162351928651333, + 0.005553704686462879, + 0.005278521683067083, + 0.0037612768355757, + 0.007997497916221619, + 0.014968744479119778, + 0.009962961077690125, + 0.009511088021099567, + 0.011107766069471836, + 0.008464429527521133, + 0.00988839566707611, + 0.005532960407435894, + 0.00420649629086256, + 0.005376686342060566, + 0.004107987973839045, + 0.0035547183360904455, + 0.0038269779179245234, + 0.0032754146959632635, + 0.003826012136414647, + 0.004330379422754049, + 0.004511937964707613, + 0.00531116733327508, + 0.0046070232056081295, + 0.0034373041708022356, + 0.006934365723282099, + 0.003335107583552599, + 0.0029078556690365076, + 0.0036193549167364836, + 0.009114193730056286, + 0.0088542141020298, + 0.009772306308150291, + 0.00743141258135438, + 0.007869657129049301, + 0.007549261208623648, + 0.007636910304427147, + 0.010228932835161686, + 0.011126524768769741, + 0.015776703134179115, + 0.013725467957556248, + 0.011612513102591038, + 0.010156030766665936, + 0.009766241535544395, + 0.012110983952879906, + 0.014836544170975685, + 0.008446489460766315, + 0.028085999190807343, + 0.008376896381378174, + 0.0066408938728272915, + 0.04895015060901642, + 0.015608948655426502, + 0.004977609496563673, + 0.003287198254838586, + 0.007747963070869446, + 0.018180105835199356, + 0.007563041057437658, + 0.008478637784719467, + 0.030461611226201057, + 0.023624058812856674, + 0.006483230274170637, + 0.012501630000770092, + 0.0169638954102993, + 0.03745751082897186, + 0.014642834663391113, + 0.020429952070116997, + 0.03423985838890076, + 0.023881761357188225, + 0.017597896978259087, + 0.015719136223196983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0009386900928802788 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550421990454197 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550421990454197, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550421990454197, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14324794709682465, + 0.10617976635694504, + 0.035089824348688126, + 0.070396289229393, + 0.1251344382762909, + 0.04742443189024925, + 0.11236986517906189, + 0.05026973411440849, + 0.24312254786491394, + 0.06676512956619263 + ], + "credit_map": [ + 0.01931370422244072, + 0.0318666473031044, + 0.019096557050943375, + 0.015295973047614098, + 0.013605979271233082, + 0.01083386316895485, + 0.016232846304774284, + 0.01700238138437271, + 0.012186224572360516, + 0.01930093951523304, + 0.011690056882798672, + 0.011269651353359222, + 0.018743794411420822, + 0.011781658977270126, + 0.012150214985013008, + 0.009057225659489632, + 0.00437021953985095, + 0.007216848898679018, + 0.004885726608335972, + 0.00508423475548625, + 0.004260444547981024, + 0.002799223642796278, + 0.00395868718624115, + 0.002514439867809415, + 0.007879151962697506, + 0.009748313575983047, + 0.007754143793135881, + 0.007885001599788666, + 0.01280552614480257, + 0.009023559279739857, + 0.007188169751316309, + 0.008112430572509766, + 0.01597713865339756, + 0.020487887784838676, + 0.01603192649781704, + 0.013137051835656166, + 0.014233922585844994, + 0.012692662887275219, + 0.015673743560910225, + 0.01690009981393814, + 0.005012548994272947, + 0.005891567096114159, + 0.00914401188492775, + 0.0043434458784759045, + 0.006625938229262829, + 0.005890734028071165, + 0.005474427714943886, + 0.005041757598519325, + 0.010420078411698341, + 0.019262786954641342, + 0.012121389620006084, + 0.010391202755272388, + 0.01779487542808056, + 0.011485207825899124, + 0.013524096459150314, + 0.017370233312249184, + 0.003056990448385477, + 0.0070877387188375, + 0.004463426768779755, + 0.0031191534362733364, + 0.019565412774682045, + 0.008011813275516033, + 0.002907128306105733, + 0.0020580729469656944, + 0.025744430720806122, + 0.03308548405766487, + 0.04063083976507187, + 0.02595333196222782, + 0.027331780642271042, + 0.025812353938817978, + 0.02557341940701008, + 0.03899092227220535, + 0.007802210282534361, + 0.012227307073771954, + 0.005533114541321993, + 0.007356396876275539, + 0.014987130649387836, + 0.007875852286815643, + 0.006018067244440317, + 0.0049650464206933975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.012550421990454197 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030331462621689 + ], + "opposite_logprob_tokens": [ + -2.3503034114837646 + ], + "logprob_reward": -0.10030331462621689, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030331462621689, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062018897384405136, + 0.05970671772956848, + 0.13453678786754608, + 0.12528173625469208, + 0.04018548130989075, + 0.0626571923494339, + 0.07694067806005478, + 0.1324082911014557, + 0.23761513829231262, + 0.06864908337593079 + ], + "credit_map": [ + 0.008251311257481575, + 0.008289316669106483, + 0.00851757824420929, + 0.010746010579168797, + 0.005384616553783417, + 0.005880977492779493, + 0.0076790135353803635, + 0.007270076312124729, + 0.007563471328467131, + 0.013316766358911991, + 0.008761502802371979, + 0.0055117132142186165, + 0.006864793598651886, + 0.004944812040776014, + 0.0045075975358486176, + 0.008236064575612545, + 0.0186636820435524, + 0.022400468587875366, + 0.017712289467453957, + 0.02480470947921276, + 0.015143471769988537, + 0.01237419992685318, + 0.01444216724485159, + 0.008995812386274338, + 0.020213469862937927, + 0.014292971231043339, + 0.014720595441758633, + 0.011275322176516056, + 0.015341997146606445, + 0.012898145243525505, + 0.01682521589100361, + 0.019714023917913437, + 0.005565315950661898, + 0.005439466796815395, + 0.004676732700318098, + 0.004176295828074217, + 0.006976280827075243, + 0.0034121498465538025, + 0.00465727224946022, + 0.005281968507915735, + 0.005998949985951185, + 0.007466047070920467, + 0.0140262795612216, + 0.006462005898356438, + 0.007407889235764742, + 0.007666212506592274, + 0.007274115923792124, + 0.006355692632496357, + 0.013042340986430645, + 0.01023035403341055, + 0.014673632569611073, + 0.007118691224604845, + 0.009240647777915001, + 0.007301143370568752, + 0.007511398289352655, + 0.00782248005270958, + 0.00669808080419898, + 0.016206853091716766, + 0.008464445360004902, + 0.0060739838518202305, + 0.05209175497293472, + 0.032572053372859955, + 0.0051095131784677505, + 0.005191607866436243, + 0.015476731583476067, + 0.03174114227294922, + 0.020600197836756706, + 0.017132237553596497, + 0.06730905920267105, + 0.055512964725494385, + 0.012891153804957867, + 0.01695169508457184, + 0.010130250826478004, + 0.01420363038778305, + 0.007021216209977865, + 0.008229022845625877, + 0.014262842014431953, + 0.008589399047195911, + 0.0037423993926495314, + 0.0024703240487724543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 7 + ], + "selected_reward": -0.10030331462621689 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002200683578848839 + ], + "opposite_logprob_tokens": [ + -6.127200603485107 + ], + "logprob_reward": -0.002200683578848839, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002200683578848839, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18327301740646362, + 0.10087920725345612, + 0.030345693230628967, + 0.07996869087219238, + 0.10273610800504684, + 0.0661722719669342, + 0.10860743373632431, + 0.0826990082859993, + 0.15685723721981049, + 0.08846132457256317 + ], + "credit_map": [ + 0.02365434728562832, + 0.040091514587402344, + 0.023986799642443657, + 0.025860199704766273, + 0.017437724396586418, + 0.011730045080184937, + 0.025064248591661453, + 0.015448144637048244, + 0.013391321524977684, + 0.01415073499083519, + 0.013950668275356293, + 0.01155952550470829, + 0.01612316630780697, + 0.01116392482072115, + 0.010600370354950428, + 0.009939488023519516, + 0.003718869062140584, + 0.0045416029170155525, + 0.0056735011748969555, + 0.0040123905055224895, + 0.003410740289837122, + 0.0023490465246140957, + 0.004458313807845116, + 0.002181227086111903, + 0.01046814490109682, + 0.008517822250723839, + 0.01307191513478756, + 0.009066279977560043, + 0.010760964825749397, + 0.008594070561230183, + 0.010389912873506546, + 0.009099583141505718, + 0.01393305603414774, + 0.015149313025176525, + 0.013234022073447704, + 0.012077546678483486, + 0.014858970418572426, + 0.00884257908910513, + 0.012490898370742798, + 0.01214972510933876, + 0.006608121562749147, + 0.010151121765375137, + 0.011213642545044422, + 0.006850504316389561, + 0.008721083402633667, + 0.007719954941421747, + 0.007895473390817642, + 0.00701237004250288, + 0.010798938572406769, + 0.020698029547929764, + 0.0126090869307518, + 0.01141708716750145, + 0.016190147027373314, + 0.010468072257936, + 0.011984914541244507, + 0.014441164210438728, + 0.004613298922777176, + 0.009444709867238998, + 0.006911433767527342, + 0.004009997937828302, + 0.03194306045770645, + 0.018290482461452484, + 0.004329637158662081, + 0.0031563881784677505, + 0.0155167942866683, + 0.018974119797348976, + 0.02849516086280346, + 0.016122952103614807, + 0.01804273948073387, + 0.01805007830262184, + 0.01549673080444336, + 0.026158658787608147, + 0.011849405243992805, + 0.019985930994153023, + 0.007200969848781824, + 0.007314086426049471, + 0.016604652628302574, + 0.009936835616827011, + 0.0072286901995539665, + 0.008340760134160519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.002200683578848839 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly upward from an extremely low angle, focuses on a set of ancient bluestone steps, drenched in the scent of dampness. Countless slender, emerald green vines cling tenaciously upwards, their tender green tendrils like tiny tentacles, cautiously exploring and clinging to the rough surface of the stone. Glistening water droplets hang from their leaves. These vines have formed a dense green carpet, almost completely covering the steps, originally a dark blue-gray hue due to moisture. Only in the gaps between the interwoven vines do the mottled marks of age-worn stone appear. Soft, diffused light streams down from above, revealing the veins of each leaf and causing the water droplets to shimmer with a faint, cool glow. The entire scene, framed by a cool palette, exudes an atmosphere of profound tranquility and unyielding vitality.", + "seed": 26729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26729_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01590347443707287, + "vlm_logprob": -0.01590347443707287, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.01590347443707287, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01574794575572014 + ], + "opposite_logprob_tokens": [ + -4.640748023986816 + ], + "logprob_reward": -0.01574794575572014, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01574794575572014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1606423407793045, + 0.13261252641677856, + 0.040609508752822876, + 0.03488357365131378, + 0.021699102595448494, + 0.07783329486846924, + 0.18515975773334503, + 0.21997156739234924, + 0.02338012307882309, + 0.1032082736492157 + ], + "credit_map": [ + 0.013167948462069035, + 0.021012920886278152, + 0.020350001752376556, + 0.0327729806303978, + 0.01630052924156189, + 0.016300398856401443, + 0.022640392184257507, + 0.018097147345542908, + 0.010515940375626087, + 0.024332182481884956, + 0.022322123870253563, + 0.013998649083077908, + 0.016288096085190773, + 0.015891803428530693, + 0.014818505384027958, + 0.014445221982896328, + 0.003470425494015217, + 0.004841513000428677, + 0.004294823855161667, + 0.008264059200882912, + 0.004403444938361645, + 0.005017799325287342, + 0.004999659955501556, + 0.005317780654877424, + 0.003566693514585495, + 0.004725409671664238, + 0.0049727666191756725, + 0.005907844752073288, + 0.004272906109690666, + 0.003370526246726513, + 0.004445001482963562, + 0.0036224229261279106, + 0.001414802740328014, + 0.0014647517818957567, + 0.001954170409590006, + 0.006306769791990519, + 0.003177030012011528, + 0.0026270172093063593, + 0.0031558391638100147, + 0.0015987188089638948, + 0.003202305408194661, + 0.004875132814049721, + 0.0040467786602675915, + 0.02548143081367016, + 0.007694634143263102, + 0.004506298806518316, + 0.022278355434536934, + 0.005748351104557514, + 0.006070856936275959, + 0.006299776956439018, + 0.006314804311841726, + 0.06571799516677856, + 0.00798666663467884, + 0.00661581801250577, + 0.07439266890287399, + 0.011761137284338474, + 0.024060947820544243, + 0.022025564685463905, + 0.02573271282017231, + 0.02494659274816513, + 0.034486934542655945, + 0.035306449979543686, + 0.03678932413458824, + 0.016623033210635185, + 0.0022467924281954765, + 0.002382509410381317, + 0.0027043367736041546, + 0.0042489743791520596, + 0.0037033301778137684, + 0.0028352688532322645, + 0.003115317551419139, + 0.0021435925737023354, + 0.006415336858481169, + 0.007830772548913956, + 0.005909763742238283, + 0.025912731885910034, + 0.028619932010769844, + 0.00899139791727066, + 0.012551537714898586, + 0.00697678467258811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.01574794575572014 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001347706071101129 + ], + "opposite_logprob_tokens": [ + -6.626347541809082 + ], + "logprob_reward": -0.001347706071101129, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001347706071101129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15161935985088348, + 0.15681855380535126, + 0.02787138521671295, + 0.06896982342004776, + 0.033342406153678894, + 0.27613094449043274, + 0.034858305007219315, + 0.17377550899982452, + 0.041880182921886444, + 0.03473354130983353 + ], + "credit_map": [ + 0.010296769440174103, + 0.019762329757213593, + 0.01784561760723591, + 0.043707262724637985, + 0.012937751598656178, + 0.012206695973873138, + 0.021618157625198364, + 0.013244780711829662, + 0.008789910934865475, + 0.015436160378158092, + 0.023932648822665215, + 0.034564465284347534, + 0.017102980986237526, + 0.013818723149597645, + 0.028879353776574135, + 0.014294304884970188, + 0.0019942354410886765, + 0.00217195856384933, + 0.0029635971877723932, + 0.008218443021178246, + 0.0028884191997349262, + 0.0021584422793239355, + 0.0037091770209372044, + 0.0037671115715056658, + 0.00611285911872983, + 0.010527091100811958, + 0.009139866568148136, + 0.01131486427038908, + 0.008985187858343124, + 0.006590113043785095, + 0.008797749876976013, + 0.0075020925141870975, + 0.002539062174037099, + 0.002959718694910407, + 0.004618857987225056, + 0.006478453520685434, + 0.005228187423199415, + 0.0039161378517746925, + 0.004695267416536808, + 0.0029067215509712696, + 0.008160962723195553, + 0.014270869083702564, + 0.009219799190759659, + 0.09976176917552948, + 0.02853979356586933, + 0.01120061706751585, + 0.09095115959644318, + 0.014025957323610783, + 0.0024236307945102453, + 0.0032775942236185074, + 0.006178340408951044, + 0.006069905590265989, + 0.004700527526438236, + 0.004514056723564863, + 0.004207453690469265, + 0.0034867958165705204, + 0.019161393865942955, + 0.021817263215780258, + 0.020012011751532555, + 0.018731093034148216, + 0.02761019580066204, + 0.024802569299936295, + 0.029253624379634857, + 0.01238736417144537, + 0.005286265630275011, + 0.003967189695686102, + 0.004540397785604, + 0.0060173762030899525, + 0.0072750793769955635, + 0.00536651024594903, + 0.005857583601027727, + 0.0035697848070412874, + 0.0036057191900908947, + 0.0027397701051086187, + 0.003472090233117342, + 0.004929997492581606, + 0.00496155908331275, + 0.004710727836936712, + 0.007552788592875004, + 0.0027608892414718866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.001347706071101129 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005957583896815777 + ], + "opposite_logprob_tokens": [ + -5.13095760345459 + ], + "logprob_reward": -0.005957583896815777, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005957583896815777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14482946693897247, + 0.09541145712137222, + 0.05947825685143471, + 0.05722695216536522, + 0.019730746746063232, + 0.06538095325231552, + 0.43847864866256714, + 0.035483017563819885, + 0.03733411803841591, + 0.04664643853902817 + ], + "credit_map": [ + 0.0116129694506526, + 0.018013397231698036, + 0.018492434173822403, + 0.03206178918480873, + 0.014627792872488499, + 0.013629834167659283, + 0.01941049098968506, + 0.016980746760964394, + 0.00649229483678937, + 0.01325723435729742, + 0.015861457213759422, + 0.01952545903623104, + 0.010342581197619438, + 0.010111304000020027, + 0.011203357949852943, + 0.008617759682238102, + 0.005552496295422316, + 0.00547074805945158, + 0.010196974501013756, + 0.006327860523015261, + 0.00902214739471674, + 0.005474473815411329, + 0.011545470915734768, + 0.005888079758733511, + 0.004546336829662323, + 0.004519761074334383, + 0.005900641903281212, + 0.013465886935591698, + 0.007090109866112471, + 0.005569337401539087, + 0.00984206423163414, + 0.0062928106635808945, + 0.0015849071787670255, + 0.001328564016148448, + 0.0022588553838431835, + 0.0028790452051907778, + 0.0030372124165296555, + 0.0026311136316508055, + 0.003949104808270931, + 0.002061943057924509, + 0.003021777141839266, + 0.0037721842527389526, + 0.005731065291911364, + 0.019890055060386658, + 0.004422427620738745, + 0.005291562993079424, + 0.01793050765991211, + 0.00532136857509613, + 0.014270932413637638, + 0.014778507873415947, + 0.013581840321421623, + 0.1560283750295639, + 0.01999557763338089, + 0.01494529563933611, + 0.17799530923366547, + 0.026882760226726532, + 0.0035372914280742407, + 0.0035102665424346924, + 0.0030393924098461866, + 0.0032932644244283438, + 0.007839287631213665, + 0.0047402880154550076, + 0.006355822551995516, + 0.003167401999235153, + 0.0054650562815368176, + 0.00368238496594131, + 0.003793898969888687, + 0.004178335890173912, + 0.007219658698886633, + 0.004401091486215591, + 0.005456864833831787, + 0.0031368229538202286, + 0.0023297301959246397, + 0.0038315339479595423, + 0.002569271018728614, + 0.013526737689971924, + 0.010966148227453232, + 0.0033667851239442825, + 0.005866363178938627, + 0.004189868923276663 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.005957583896815777 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494775250554085 + ], + "opposite_logprob_tokens": [ + -2.929947853088379 + ], + "logprob_reward": -0.05494775250554085, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494775250554085, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23506078124046326, + 0.05617999657988548, + 0.016578208655118942, + 0.021256187930703163, + 0.024068579077720642, + 0.06430638581514359, + 0.19939659535884857, + 0.11204274743795395, + 0.039261385798454285, + 0.23184913396835327 + ], + "credit_map": [ + 0.013996601104736328, + 0.03388674184679985, + 0.031710777431726456, + 0.07440569996833801, + 0.021674586459994316, + 0.012103085406124592, + 0.03080633096396923, + 0.016476962715387344, + 0.0040184804238379, + 0.00908062793314457, + 0.01125936210155487, + 0.0074277943931519985, + 0.00649999501183629, + 0.005790781695395708, + 0.007007778622210026, + 0.005095177795737982, + 0.0010294297244399786, + 0.0014658438740298152, + 0.001724948175251484, + 0.005200135055929422, + 0.0020652825478464365, + 0.001289510284550488, + 0.0020844493992626667, + 0.0017186099430546165, + 0.0017768783727660775, + 0.0020379102788865566, + 0.0020830724388360977, + 0.004354352597147226, + 0.0026318703312426805, + 0.0024165003560483456, + 0.0030392552725970745, + 0.0029163502622395754, + 0.002058576326817274, + 0.0019312346121296287, + 0.002221807138994336, + 0.005214405711740255, + 0.003020199714228511, + 0.002910386538133025, + 0.004782525356858969, + 0.0019294441444799304, + 0.002324308967217803, + 0.003573081223294139, + 0.003201250685378909, + 0.02349252440035343, + 0.0052349623292684555, + 0.0035977556835860014, + 0.01861489564180374, + 0.004267605487257242, + 0.006413387134671211, + 0.00687654223293066, + 0.006038106977939606, + 0.07565860450267792, + 0.009363831952214241, + 0.006359072867780924, + 0.07890032231807709, + 0.00978672131896019, + 0.014393688179552555, + 0.011093663983047009, + 0.010173327289521694, + 0.01081791426986456, + 0.016931775957345963, + 0.02068968303501606, + 0.013523440808057785, + 0.01441926322877407, + 0.0030910326167941093, + 0.0032799358014017344, + 0.0032478352077305317, + 0.008608956821262836, + 0.006765431724488735, + 0.004155457951128483, + 0.006646315101534128, + 0.00346641824580729, + 0.01380635891109705, + 0.018532440066337585, + 0.010932238772511482, + 0.06168974190950394, + 0.0741664469242096, + 0.020867520943284035, + 0.019937925040721893, + 0.011916454881429672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.05494775250554085 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015163839561864734 + ], + "opposite_logprob_tokens": [ + -6.501516342163086 + ], + "logprob_reward": -0.0015163839561864734, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015163839561864734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17966920137405396, + 0.10627668350934982, + 0.0678638368844986, + 0.051848579198122025, + 0.018054679036140442, + 0.1098988726735115, + 0.33868226408958435, + 0.058834608644247055, + 0.03555597737431526, + 0.03331533074378967 + ], + "credit_map": [ + 0.013668992556631565, + 0.022794337943196297, + 0.024021057412028313, + 0.0426078736782074, + 0.018128184601664543, + 0.016810715198516846, + 0.02394847758114338, + 0.017689546570181847, + 0.006571955978870392, + 0.015672592446208, + 0.015733404085040092, + 0.023762013763189316, + 0.011355102993547916, + 0.01068180799484253, + 0.012559191323816776, + 0.009940612129867077, + 0.004118351731449366, + 0.007483604829758406, + 0.009486515074968338, + 0.017830688506364822, + 0.006975722964853048, + 0.005252677947282791, + 0.009532374329864979, + 0.007183899637311697, + 0.004148927982896566, + 0.006425897125154734, + 0.006435773801058531, + 0.010030296631157398, + 0.00780677143484354, + 0.0046440851874649525, + 0.0067769624292850494, + 0.005579865537583828, + 0.0013695689849555492, + 0.0015920348232612014, + 0.0017532500205561519, + 0.0053446730598807335, + 0.0017454351764172316, + 0.0022914125584065914, + 0.002223692834377289, + 0.0017346126260235906, + 0.004484034609049559, + 0.005391980055719614, + 0.006359336897730827, + 0.036299072206020355, + 0.009686320088803768, + 0.0072355130687355995, + 0.033112671226263046, + 0.007329938001930714, + 0.011195200495421886, + 0.011282680556178093, + 0.010802937671542168, + 0.1213202029466629, + 0.014963493682444096, + 0.011376282200217247, + 0.13746733963489532, + 0.020274147391319275, + 0.005748586263507605, + 0.006678127683699131, + 0.006489146500825882, + 0.004787086509168148, + 0.006550267338752747, + 0.011309555731713772, + 0.008076776750385761, + 0.009195062331855297, + 0.005604004953056574, + 0.003244631690904498, + 0.0034223850816488266, + 0.0047063701786100864, + 0.005554026458412409, + 0.0042929756455123425, + 0.005583673715591431, + 0.003147908952087164, + 0.0025031063705682755, + 0.0030572537798434496, + 0.0030436827801167965, + 0.006546052172780037, + 0.008426510728895664, + 0.002941821003332734, + 0.004641599953174591, + 0.0021553041879087687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.0015163839561864734 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up shot, tilting slowly upward from an extremely low angle, focuses on a set of ancient bluestone steps, drenched in the scent of dampness. Countless slender, emerald green vines cling tenaciously upwards, their tender green tendrils like tiny tentacles, cautiously exploring and clinging to the rough surface of the stone. Glistening water droplets hang from their leaves. These vines have formed a dense green carpet, almost completely covering the steps, originally a dark blue-gray hue due to moisture. Only in the gaps between the interwoven vines do the mottled marks of age-worn stone appear. Soft, diffused light streams down from above, revealing the veins of each leaf and causing the water droplets to shimmer with a faint, cool glow. The entire scene, framed by a cool palette, exudes an atmosphere of profound tranquility and unyielding vitality.", + "seed": 26730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_7/index_9813_seed_26730_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02457710758317262, + "vlm_logprob": -0.02457710758317262, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.02457710758317262, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029216017574071884 + ], + "opposite_logprob_tokens": [ + -3.9042160511016846 + ], + "logprob_reward": -0.029216017574071884, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.029216017574071884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09892217069864273, + 0.034183744341135025, + 0.05490701273083687, + 0.02401699125766754, + 0.02688726969063282, + 0.14692294597625732, + 0.1231750026345253, + 0.29208803176879883, + 0.1421641856431961, + 0.05673259124159813 + ], + "credit_map": [ + 0.009633846580982208, + 0.013999070972204208, + 0.01138609554618597, + 0.015265622176229954, + 0.00958668440580368, + 0.009920867159962654, + 0.013994788751006126, + 0.015135197900235653, + 0.003907063975930214, + 0.0038586079608649015, + 0.004618872422724962, + 0.00523280119523406, + 0.00410154415294528, + 0.004517792258411646, + 0.004485212732106447, + 0.0034618484787642956, + 0.007126207463443279, + 0.010512442328035831, + 0.008371752686798573, + 0.007956034503877163, + 0.0035596529487520456, + 0.0066922614350914955, + 0.004395932424813509, + 0.0062927259132266045, + 0.005099725443869829, + 0.0036603000480681658, + 0.002892474178224802, + 0.0028240010142326355, + 0.0030119335278868675, + 0.0020162847358733416, + 0.00219790730625391, + 0.0023143666330724955, + 0.0037258199881762266, + 0.0033244614023715258, + 0.004035553894937038, + 0.003328286111354828, + 0.003223713720217347, + 0.003852984867990017, + 0.0020186491310596466, + 0.0033778012730181217, + 0.013853849843144417, + 0.009697708301246166, + 0.04959350824356079, + 0.021314026787877083, + 0.01065229345113039, + 0.025302529335021973, + 0.007281491998583078, + 0.009227534756064415, + 0.019552629441022873, + 0.014687950722873211, + 0.024531692266464233, + 0.013622112572193146, + 0.009947375394403934, + 0.021482041105628014, + 0.007065227255225182, + 0.012285972014069557, + 0.0786418467760086, + 0.014658884145319462, + 0.04590796306729317, + 0.018697230145335197, + 0.017131350934505463, + 0.07100798189640045, + 0.012761917896568775, + 0.0332808755338192, + 0.016144677996635437, + 0.021308530122041702, + 0.02709309570491314, + 0.0232235137373209, + 0.014827614650130272, + 0.017765125259757042, + 0.010711990296840668, + 0.011089641600847244, + 0.007565996143966913, + 0.011280824430286884, + 0.007496250793337822, + 0.009370985440909863, + 0.005185060203075409, + 0.005571614485234022, + 0.006306749768555164, + 0.003955111373215914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.029216017574071884 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013642538106068969 + ], + "opposite_logprob_tokens": [ + -6.626364231109619 + ], + "logprob_reward": -0.0013642538106068969, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013642538106068969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.083732470870018, + 0.06926938146352768, + 0.08614562451839447, + 0.03812636807560921, + 0.04249952360987663, + 0.04750828444957733, + 0.21481359004974365, + 0.25302767753601074, + 0.11692772805690765, + 0.04794936627149582 + ], + "credit_map": [ + 0.007009817752987146, + 0.0093158520758152, + 0.009253740310668945, + 0.01610674522817135, + 0.00611085956916213, + 0.007142499089241028, + 0.012950746342539787, + 0.01584220491349697, + 0.0075005763210356236, + 0.005814430303871632, + 0.008127995766699314, + 0.010263766162097454, + 0.0067096431739628315, + 0.010072838515043259, + 0.010381647385656834, + 0.010398483835160732, + 0.008860321715474129, + 0.016644615679979324, + 0.011863145977258682, + 0.010427850298583508, + 0.008412625640630722, + 0.012667355127632618, + 0.008409371599555016, + 0.008860337547957897, + 0.005859889090061188, + 0.00455244816839695, + 0.005356420762836933, + 0.0043839761056005955, + 0.004630511160939932, + 0.004359778948128223, + 0.004138005897402763, + 0.0048453365452587605, + 0.006016156170517206, + 0.005348941311240196, + 0.006238610949367285, + 0.0037967355456203222, + 0.005349006969481707, + 0.006661714054644108, + 0.004434186965227127, + 0.004654168616980314, + 0.005030288826674223, + 0.005451945122331381, + 0.0070572649128735065, + 0.007564152125269175, + 0.004655254073441029, + 0.00853382982313633, + 0.004613159224390984, + 0.004602390341460705, + 0.03794092312455177, + 0.01893327571451664, + 0.027924636378884315, + 0.016472069546580315, + 0.01620054617524147, + 0.026167985051870346, + 0.013651990331709385, + 0.05752215161919594, + 0.07030611485242844, + 0.013979349285364151, + 0.0366077795624733, + 0.011279918253421783, + 0.016353141516447067, + 0.06425073742866516, + 0.01115462090820074, + 0.02909601852297783, + 0.0126575892791152, + 0.011259774677455425, + 0.010565354488790035, + 0.020161448046565056, + 0.014682179316878319, + 0.01848328299820423, + 0.01485876739025116, + 0.014259321615099907, + 0.0074332039803266525, + 0.008591091260313988, + 0.006051888223737478, + 0.0060980967245996, + 0.006437276024371386, + 0.005421662703156471, + 0.0038852498400956392, + 0.004030897747725248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.0013642538106068969 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372899070382118 + ], + "opposite_logprob_tokens": [ + -3.408729076385498 + ], + "logprob_reward": -0.03372899070382118, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372899070382118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1448061764240265, + 0.06182444468140602, + 0.10159944742918015, + 0.08095898479223251, + 0.053797874599695206, + 0.059196487069129944, + 0.08226139098405838, + 0.22519628703594208, + 0.13119177520275116, + 0.05916709452867508 + ], + "credit_map": [ + 0.012775862589478493, + 0.01646122895181179, + 0.016119210049510002, + 0.027402393519878387, + 0.012043504975736141, + 0.015343149192631245, + 0.024754934012889862, + 0.019905904307961464, + 0.005954070016741753, + 0.00934539269655943, + 0.008076531812548637, + 0.008456695824861526, + 0.00642143189907074, + 0.007207391783595085, + 0.0077162254601716995, + 0.008646709844470024, + 0.01148285437375307, + 0.019492454826831818, + 0.012125534005463123, + 0.013890109956264496, + 0.010134031064808369, + 0.016156231984496117, + 0.008541583083570004, + 0.009776651859283447, + 0.007737182546406984, + 0.006055639125406742, + 0.01418196503072977, + 0.009100854396820068, + 0.012043370865285397, + 0.008375082165002823, + 0.009259924292564392, + 0.014204966835677624, + 0.0052580819465219975, + 0.008262979798018932, + 0.008164724335074425, + 0.006623420398682356, + 0.00748082622885704, + 0.0070686666294932365, + 0.004622318781912327, + 0.006316857412457466, + 0.006102883722633123, + 0.0061204941011965275, + 0.00668085552752018, + 0.005774331744760275, + 0.007758351508527994, + 0.01261982787400484, + 0.006873234640806913, + 0.007266506087034941, + 0.010557319968938828, + 0.007768240291625261, + 0.0157997477799654, + 0.00830887258052826, + 0.007264779880642891, + 0.011045577004551888, + 0.005831462796777487, + 0.015685394406318665, + 0.03647191822528839, + 0.015714768320322037, + 0.026936454698443413, + 0.02373170480132103, + 0.016490880399942398, + 0.03064514324069023, + 0.021112220361828804, + 0.054093215614557266, + 0.01497493963688612, + 0.017267677932977676, + 0.03885384276509285, + 0.013511705212295055, + 0.01203903928399086, + 0.016569070518016815, + 0.008415688760578632, + 0.009559824131429195, + 0.00862369779497385, + 0.007159662898629904, + 0.011276968754827976, + 0.00751807214692235, + 0.006470352411270142, + 0.006850035395473242, + 0.00564880995079875, + 0.005619497504085302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.03372899070382118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495553836226463 + ], + "opposite_logprob_tokens": [ + -2.92995548248291 + ], + "logprob_reward": -0.05495553836226463, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495553836226463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11687694489955902, + 0.050234854221343994, + 0.03456643596291542, + 0.019613521173596382, + 0.058113597333431244, + 0.0985647588968277, + 0.13797561824321747, + 0.17794770002365112, + 0.28502964973449707, + 0.021076995879411697 + ], + "credit_map": [ + 0.008955428376793861, + 0.022763660177588463, + 0.012568718753755093, + 0.018036771565675735, + 0.012564393691718578, + 0.011650113388895988, + 0.01763809658586979, + 0.012699758633971214, + 0.003741994732990861, + 0.005512668285518885, + 0.006003180518746376, + 0.010605493560433388, + 0.005170042160898447, + 0.006176366005092859, + 0.008093740791082382, + 0.004931366536766291, + 0.0041197845712304115, + 0.005337173584848642, + 0.004256957210600376, + 0.005573592614382505, + 0.0038070129230618477, + 0.004169281106442213, + 0.0033491672948002815, + 0.003953465726226568, + 0.0032912190072238445, + 0.0026148981414735317, + 0.0023513806518167257, + 0.0018436209065839648, + 0.0024706684052944183, + 0.002184554934501648, + 0.0024609402753412724, + 0.002396238502115011, + 0.007796320598572493, + 0.005924246273934841, + 0.008598838932812214, + 0.004851736593991518, + 0.007570247631520033, + 0.01452236995100975, + 0.0039400882087647915, + 0.004909751005470753, + 0.010352647863328457, + 0.00975329801440239, + 0.02713601291179657, + 0.011603405699133873, + 0.008436831645667553, + 0.016215363517403603, + 0.007528606802225113, + 0.007538595702499151, + 0.027000071480870247, + 0.01107939425855875, + 0.014672111719846725, + 0.006579390726983547, + 0.008899901062250137, + 0.044465210288763046, + 0.004404769279062748, + 0.020874759182333946, + 0.05161195248365402, + 0.014079015702009201, + 0.04805130511522293, + 0.016461778432130814, + 0.0145990876480937, + 0.016055991873145103, + 0.007702241651713848, + 0.009386320598423481, + 0.0321279838681221, + 0.03620786964893341, + 0.08780880272388458, + 0.03158320114016533, + 0.02276819571852684, + 0.03414588421583176, + 0.018906544893980026, + 0.021481167525053024, + 0.0028423878829926252, + 0.0030224472284317017, + 0.004598279017955065, + 0.0024374746717512608, + 0.002103448146954179, + 0.0022342558950185776, + 0.0021661289501935244, + 0.0016725739696994424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.05495553836226463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of ancient bluestone steps?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the steps covered by a dense carpet of slender emerald green vines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glistening water droplets visible hanging from the vine leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera tilting slowly upward from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, diffused light causing water droplets to shimmer?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036207374650985003 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036207374650985003, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036207374650985003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14135292172431946, + 0.0495186485350132, + 0.10756935179233551, + 0.039580464363098145, + 0.07127807289361954, + 0.10904726386070251, + 0.08385656774044037, + 0.2875961661338806, + 0.05328772962093353, + 0.0569128654897213 + ], + "credit_map": [ + 0.013674674555659294, + 0.01828884705901146, + 0.018164565786719322, + 0.027727309614419937, + 0.013539373874664307, + 0.015226401388645172, + 0.020926231518387794, + 0.013805519789457321, + 0.005565039813518524, + 0.007077919319272041, + 0.006744204089045525, + 0.0060164895839989185, + 0.004875322803854942, + 0.006586041767150164, + 0.006497123744338751, + 0.006156506948173046, + 0.01401242334395647, + 0.02186778374016285, + 0.013065245933830738, + 0.015083195641636848, + 0.008864865638315678, + 0.014180343598127365, + 0.00937099289149046, + 0.011124503798782825, + 0.005584521684795618, + 0.0037585017271339893, + 0.006793169304728508, + 0.005553172901272774, + 0.004605964757502079, + 0.003949373494833708, + 0.004723632242530584, + 0.004612125921994448, + 0.00669091148301959, + 0.00841903779655695, + 0.011557627469301224, + 0.006672759540379047, + 0.007610068190842867, + 0.01615086942911148, + 0.006430596578866243, + 0.007746199145913124, + 0.009793617762625217, + 0.008493424393236637, + 0.021040918305516243, + 0.00940933171659708, + 0.01071006990969181, + 0.03612840175628662, + 0.005089887417852879, + 0.008381607010960579, + 0.01648050919175148, + 0.007855025120079517, + 0.011250877752900124, + 0.006434385199099779, + 0.0077951084822416306, + 0.01988937333226204, + 0.0046533625572919846, + 0.00949792843312025, + 0.08561324328184128, + 0.01647314615547657, + 0.024460894986987114, + 0.01686886139214039, + 0.023211169987916946, + 0.09014371037483215, + 0.012293634936213493, + 0.01853150874376297, + 0.004998377524316311, + 0.00628310302272439, + 0.01259420346468687, + 0.008418749086558819, + 0.00675763376057148, + 0.006024210713803768, + 0.0039045969024300575, + 0.004306858871132135, + 0.007711710408329964, + 0.005582910496741533, + 0.010800499469041824, + 0.008506424725055695, + 0.005478091537952423, + 0.00655412720516324, + 0.0073866466991603374, + 0.004892456810921431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.0036207374650985003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_tensor([9288, 9288])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_tensor([9288, 9288])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a49e7a77911e2fcee9018f62148bffd81489d91e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_tensor([9288, 9288])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In an extreme close-up shot, against a backdrop of pure, bottomless darkness, a highly focused, sharp-edged cold-white spotlight shines diagonally from above, precisely illuminating a corner of a recently excavated ancient bronze sacrificial vessel. The vessel's surface is covered in uneven bluish-green patina, intricately carved with mythical beasts, and crevices of dried earth are embedded in its crevices. Shot in extreme slow motion, countless tiny particles of dust and sand are gently awakened by an invisible air current, slowly rising from the intricate veins of the bronze vessel. Each tiny particle is clearly captured in the beam, shimmering with a fleeting, warm golden glow. As if imbued with life, they spiral upward with an almost hypnotic grace, dispersing and ultimately dissolving into the surrounding darkness. The entire scene evokes a sense of tranquility and solemnity, as if frozen in time, silently telling the story of a long-forgotten history.", + "seed": 24727, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_9288_seed_24727_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.13305808156728743, + "vlm_logprob": -0.13305808156728743, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.13305808156728743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2590576708316803 + ], + "opposite_logprob_tokens": [ + -1.509057641029358 + ], + "logprob_reward": -0.2590576708316803, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2590576708316803, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3040142357349396, + 0.09863289445638657, + 0.2429480403661728, + 0.03895973786711693, + 0.05551345273852348, + 0.024846825748682022, + 0.04468237981200218, + 0.10279692709445953, + 0.04409298300743103, + 0.0435124933719635 + ], + "credit_map": [ + 0.06136526167392731, + 0.057761918753385544, + 0.04400506243109703, + 0.024400707334280014, + 0.023416364565491676, + 0.028389563784003258, + 0.037504930049180984, + 0.02717043086886406, + 0.01350334007292986, + 0.015287256799638271, + 0.012712076306343079, + 0.008514588698744774, + 0.010843419469892979, + 0.015570675022900105, + 0.008845129981637001, + 0.0133564043790102, + 0.049393411725759506, + 0.04418930038809776, + 0.049058523029088974, + 0.021803997457027435, + 0.01439894549548626, + 0.01660473272204399, + 0.024268271401524544, + 0.023230845108628273, + 0.007032216992229223, + 0.005255917552858591, + 0.008569140918552876, + 0.0033889871556311846, + 0.003660581773146987, + 0.002692067064344883, + 0.00519524933770299, + 0.003165577072650194, + 0.01031565759330988, + 0.009859086945652962, + 0.01542652491480112, + 0.004190814681351185, + 0.003821809310466051, + 0.0034501636400818825, + 0.004556796047836542, + 0.003892596112564206, + 0.00424688495695591, + 0.0031095074955374002, + 0.006234055384993553, + 0.0020930415485054255, + 0.002203194657340646, + 0.0016593736363574862, + 0.0027323123067617416, + 0.002568455645814538, + 0.00796421803534031, + 0.004995899274945259, + 0.009762085974216461, + 0.004991760011762381, + 0.004291407763957977, + 0.003722475143149495, + 0.005486250389367342, + 0.003468284150585532, + 0.017939062789082527, + 0.010785220190882683, + 0.02406422235071659, + 0.00847594439983368, + 0.00796709768474102, + 0.008251466788351536, + 0.009999916888773441, + 0.01531400065869093, + 0.007029748521745205, + 0.00462079793214798, + 0.008247720077633858, + 0.003874386427924037, + 0.004048474133014679, + 0.004750692285597324, + 0.005801180377602577, + 0.0057199858129024506, + 0.006124494131654501, + 0.00428434694185853, + 0.006970731075853109, + 0.004252301529049873, + 0.003824737388640642, + 0.004241520073264837, + 0.005468158051371574, + 0.008346201851963997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.2590576708316803 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134867250919342 + ], + "opposite_logprob_tokens": [ + -1.3134866952896118 + ], + "logprob_reward": -0.3134867250919342, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134867250919342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14920462667942047, + 0.23038850724697113, + 0.057259030640125275, + 0.027744488790631294, + 0.054368842393159866, + 0.11473292857408524, + 0.0893038958311081, + 0.12428802996873856, + 0.07254572212696075, + 0.08016391098499298 + ], + "credit_map": [ + 0.022707056254148483, + 0.025037987157702446, + 0.02097306028008461, + 0.016115492209792137, + 0.011402725242078304, + 0.014272374100983143, + 0.019389355555176735, + 0.019306570291519165, + 0.026791535317897797, + 0.050906576216220856, + 0.02997307851910591, + 0.018949013203382492, + 0.01845967210829258, + 0.026639150455594063, + 0.026144998148083687, + 0.032524485141038895, + 0.008847521618008614, + 0.01066755410283804, + 0.00959651917219162, + 0.005501262377947569, + 0.004613408353179693, + 0.003238230710849166, + 0.006522125098854303, + 0.008272412233054638, + 0.00417326670140028, + 0.00519450893625617, + 0.003724224166944623, + 0.0018702782690525055, + 0.003803626401349902, + 0.003591758431866765, + 0.0028199574444442987, + 0.002566866809502244, + 0.007350282743573189, + 0.005084861535578966, + 0.009607826359570026, + 0.0032962518744170666, + 0.00936079677194357, + 0.009393545798957348, + 0.005409449338912964, + 0.004865830764174461, + 0.02115366794168949, + 0.013206874951720238, + 0.035052862018346786, + 0.007692077662795782, + 0.013028981164097786, + 0.00778536032885313, + 0.008365795947611332, + 0.00844730343669653, + 0.01404562033712864, + 0.012935703620314598, + 0.013766657561063766, + 0.010374024510383606, + 0.010209341533482075, + 0.01042281836271286, + 0.007923041470348835, + 0.00962669588625431, + 0.022303873673081398, + 0.012319214642047882, + 0.029339464381337166, + 0.008574370294809341, + 0.00888162199407816, + 0.011974303051829338, + 0.014094693586230278, + 0.016800494864583015, + 0.013794055208563805, + 0.01041001733392477, + 0.022449074313044548, + 0.0040365285240113735, + 0.005230603739619255, + 0.006219654809683561, + 0.004585400223731995, + 0.005820393096655607, + 0.015668969601392746, + 0.006196859758347273, + 0.024759814143180847, + 0.004680288024246693, + 0.0048651122488081455, + 0.004972807597368956, + 0.006229718215763569, + 0.012790349312126637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.3134867250919342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043096739798784256 + ], + "opposite_logprob_tokens": [ + -3.1680967807769775 + ], + "logprob_reward": -0.043096739798784256, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043096739798784256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15695936977863312, + 0.17387068271636963, + 0.09396395087242126, + 0.3207775950431824, + 0.026799317449331284, + 0.05827254056930542, + 0.02420627512037754, + 0.057651322335004807, + 0.03704243525862694, + 0.05045653507113457 + ], + "credit_map": [ + 0.03184664249420166, + 0.02365974895656109, + 0.02378823608160019, + 0.019012346863746643, + 0.012595828622579575, + 0.012172653339803219, + 0.01988566666841507, + 0.013998247683048248, + 0.024299055337905884, + 0.05002306401729584, + 0.02679869532585144, + 0.012086069211363792, + 0.012060588225722313, + 0.012518509291112423, + 0.01947183534502983, + 0.016612859442830086, + 0.011503845453262329, + 0.013104937970638275, + 0.01289174985140562, + 0.01475051324814558, + 0.008907413110136986, + 0.006406725384294987, + 0.01368387509137392, + 0.01271488144993782, + 0.08564453572034836, + 0.04933905601501465, + 0.09066605567932129, + 0.017287693917751312, + 0.023280056193470955, + 0.01816548779606819, + 0.01848503388464451, + 0.01790970377624035, + 0.0040256790816783905, + 0.0036772009916603565, + 0.005248187575489283, + 0.002102937549352646, + 0.003130165860056877, + 0.0021329987794160843, + 0.0034973553847521544, + 0.002984790364280343, + 0.010215593501925468, + 0.007026249077171087, + 0.015802614390850067, + 0.005219298414885998, + 0.0059202793054282665, + 0.0038339935708791018, + 0.006047314032912254, + 0.0042071957141160965, + 0.0035588692408055067, + 0.0029759674798697233, + 0.00441989628598094, + 0.003168255090713501, + 0.0024604799691587687, + 0.0024662872310727835, + 0.0027186896186321974, + 0.0024378306698054075, + 0.009591917507350445, + 0.006307217758148909, + 0.011683777906000614, + 0.004601452499628067, + 0.004534426610916853, + 0.007066871039569378, + 0.00721470732241869, + 0.006650949828326702, + 0.006497814320027828, + 0.00517748948186636, + 0.009561472572386265, + 0.00391122791916132, + 0.003081804607063532, + 0.0024520792067050934, + 0.0034331358037889004, + 0.002927408553659916, + 0.007713325787335634, + 0.0049962750636041164, + 0.009857208468019962, + 0.003837241791188717, + 0.004459666088223457, + 0.005410825368016958, + 0.004897569306194782, + 0.009284423664212227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.043096739798784256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023296110332012177 + ], + "opposite_logprob_tokens": [ + -3.7732961177825928 + ], + "logprob_reward": -0.023296110332012177, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023296110332012177, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06642358750104904, + 0.0958784967660904, + 0.03304369002580643, + 0.07118920236825943, + 0.029797472059726715, + 0.05051141232252121, + 0.12738458812236786, + 0.4814405143260956, + 0.02936435304582119, + 0.014966667629778385 + ], + "credit_map": [ + 0.01585400104522705, + 0.013851100578904152, + 0.011057894676923752, + 0.004698052536696196, + 0.004244765732437372, + 0.003751295618712902, + 0.007254692725837231, + 0.005711792036890984, + 0.011843251064419746, + 0.024060208350419998, + 0.013207834213972092, + 0.008521093986928463, + 0.008829742670059204, + 0.010474218055605888, + 0.009029952809214592, + 0.00991220772266388, + 0.005662404466420412, + 0.006476706825196743, + 0.00535241886973381, + 0.0038050918374210596, + 0.0027465978637337685, + 0.0016113778110593557, + 0.004146518651396036, + 0.0032425764948129654, + 0.016177373006939888, + 0.01488229539245367, + 0.015029492788016796, + 0.004836956039071083, + 0.00595594709739089, + 0.004437423311173916, + 0.0056290049105882645, + 0.004240715876221657, + 0.004398596938699484, + 0.004997581243515015, + 0.004016126040369272, + 0.0023499575909227133, + 0.004502874333411455, + 0.0037162292283028364, + 0.002662725979462266, + 0.0031533821020275354, + 0.008626200258731842, + 0.006503689102828503, + 0.012461937963962555, + 0.004084798973053694, + 0.005805118475109339, + 0.003983747214078903, + 0.004866211209446192, + 0.004179711919277906, + 0.018258454278111458, + 0.018046937882900238, + 0.02045610174536705, + 0.017919789999723434, + 0.01609979197382927, + 0.011415157467126846, + 0.013863932341337204, + 0.011324428021907806, + 0.0840519517660141, + 0.044055040925741196, + 0.09064595401287079, + 0.038202814757823944, + 0.0366511270403862, + 0.04397105798125267, + 0.061938729137182236, + 0.08192386478185654, + 0.004574213642627001, + 0.00360804027877748, + 0.005926403217017651, + 0.0023383372463285923, + 0.0026074659544974566, + 0.003831716952845454, + 0.0025454892311245203, + 0.003932686522603035, + 0.0018730072770267725, + 0.0017859466606751084, + 0.0030752727761864662, + 0.0013517761835828424, + 0.001629954669624567, + 0.0014459037920460105, + 0.0019749386701732874, + 0.0018298691138625145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.023296110332012177 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635316178202629 + ], + "opposite_logprob_tokens": [ + -3.651353120803833 + ], + "logprob_reward": -0.02635316178202629, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635316178202629, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21590495109558105, + 0.09923816472291946, + 0.13335870206356049, + 0.06380867958068848, + 0.044229473918676376, + 0.05432053655385971, + 0.106113500893116, + 0.14347653090953827, + 0.03761386126279831, + 0.10193555802106857 + ], + "credit_map": [ + 0.039049502462148666, + 0.029580507427453995, + 0.03038586862385273, + 0.02715836837887764, + 0.01770535111427307, + 0.022206498309969902, + 0.028886007145047188, + 0.020932869985699654, + 0.014825896359980106, + 0.029561836272478104, + 0.017328981310129166, + 0.006227836944162846, + 0.007843637838959694, + 0.004502299707382917, + 0.0094331419095397, + 0.009514535777270794, + 0.022008944302797318, + 0.022037692368030548, + 0.023209761828184128, + 0.018497956916689873, + 0.011048688553273678, + 0.008535020984709263, + 0.01636635698378086, + 0.011654291301965714, + 0.007321711163967848, + 0.012666229158639908, + 0.00887947715818882, + 0.006458637770265341, + 0.008500284515321255, + 0.008183608762919903, + 0.005965793505311012, + 0.005832941737025976, + 0.007464409805834293, + 0.007530871778726578, + 0.009356792084872723, + 0.0033198660239577293, + 0.004307607654482126, + 0.003949063364416361, + 0.0033146971836686134, + 0.004986169748008251, + 0.007698293309658766, + 0.005954545456916094, + 0.00921715423464775, + 0.006802487187087536, + 0.00492594251409173, + 0.004647544119507074, + 0.007893653586506844, + 0.007180920336395502, + 0.016477081924676895, + 0.015984894707798958, + 0.01503633800894022, + 0.012634415179491043, + 0.011556588113307953, + 0.014429589733481407, + 0.009071801789104939, + 0.010922791436314583, + 0.022463956847786903, + 0.014058683067560196, + 0.019520673900842667, + 0.009162974543869495, + 0.01234894897788763, + 0.02067357301712036, + 0.01959383673965931, + 0.025653906166553497, + 0.006574329920113087, + 0.004104685503989458, + 0.008944700472056866, + 0.0029807540122419596, + 0.0029311066027730703, + 0.0029574711807072163, + 0.004078574012964964, + 0.005042240954935551, + 0.02078213542699814, + 0.007639285642653704, + 0.03444063663482666, + 0.006450706627219915, + 0.006697328295558691, + 0.007261279970407486, + 0.005774109624326229, + 0.012890082783997059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.02635316178202629 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up shot, against a backdrop of pure, bottomless darkness, a highly focused, sharp-edged cold-white spotlight shines diagonally from above, precisely illuminating a corner of a recently excavated ancient bronze sacrificial vessel. The vessel's surface is covered in uneven bluish-green patina, intricately carved with mythical beasts, and crevices of dried earth are embedded in its crevices. Shot in extreme slow motion, countless tiny particles of dust and sand are gently awakened by an invisible air current, slowly rising from the intricate veins of the bronze vessel. Each tiny particle is clearly captured in the beam, shimmering with a fleeting, warm golden glow. As if imbued with life, they spiral upward with an almost hypnotic grace, dispersing and ultimately dissolving into the surrounding darkness. The entire scene evokes a sense of tranquility and solemnity, as if frozen in time, silently telling the story of a long-forgotten history.", + "seed": 24728, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_9288_seed_24728_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06472040982916952, + "vlm_logprob": -0.06472040982916952, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.06472040982916952, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20946186780929565 + ], + "opposite_logprob_tokens": [ + -1.7094619274139404 + ], + "logprob_reward": -0.20946186780929565, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20946186780929565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4960244596004486, + 0.08035464584827423, + 0.032269470393657684, + 0.021904995664954185, + 0.020707428455352783, + 0.06965681165456772, + 0.1346903145313263, + 0.11477462947368622, + 0.020708510652184486, + 0.008908724412322044 + ], + "credit_map": [ + 0.057577986270189285, + 0.09374253451824188, + 0.020479079335927963, + 0.02499467134475708, + 0.04461811110377312, + 0.15322916209697723, + 0.019578421488404274, + 0.08180446177721024, + 0.007826322689652443, + 0.0162957813590765, + 0.004588000942021608, + 0.005269784014672041, + 0.00887625478208065, + 0.0207682978361845, + 0.004702629055827856, + 0.01202757004648447, + 0.0037784718442708254, + 0.003440747270360589, + 0.003806520253419876, + 0.0036038148682564497, + 0.0034621069207787514, + 0.007075691595673561, + 0.002761593321338296, + 0.004340520594269037, + 0.0018248609267175198, + 0.001871728920377791, + 0.004533227067440748, + 0.002800069283694029, + 0.002386941108852625, + 0.0030784467235207558, + 0.002601762767881155, + 0.002807957585901022, + 0.001891849096864462, + 0.0019339710706844926, + 0.003647120436653495, + 0.0029898863285779953, + 0.0025235351640731096, + 0.0028258785605430603, + 0.0026659704744815826, + 0.002229216741397977, + 0.0063432869501411915, + 0.005207592621445656, + 0.01005211379379034, + 0.012863855808973312, + 0.007761639542877674, + 0.00853673368692398, + 0.011566314846277237, + 0.007325270678848028, + 0.011400280520319939, + 0.01207028329372406, + 0.018765032291412354, + 0.021601224318146706, + 0.012896835803985596, + 0.022068249061703682, + 0.020692480728030205, + 0.015195924788713455, + 0.011041532270610332, + 0.008115502074360847, + 0.02463131584227085, + 0.014071421697735786, + 0.011083393357694149, + 0.016645170748233795, + 0.015387771651148796, + 0.013798515312373638, + 0.0015766900032758713, + 0.0014375869650393724, + 0.0030813291668891907, + 0.00231187860481441, + 0.001098460634239018, + 0.004535612650215626, + 0.002597809536382556, + 0.004069141577929258, + 0.0007682322175242007, + 0.0007719947607256472, + 0.0015552821569144726, + 0.0010652898345142603, + 0.0006865921895951033, + 0.0017623145831748843, + 0.0009683755342848599, + 0.0013306433102115989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.20946186780929565 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869326949119568 + ], + "opposite_logprob_tokens": [ + -3.0486931800842285 + ], + "logprob_reward": -0.04869326949119568, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869326949119568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10182962566614151, + 0.0748709961771965, + 0.029207222163677216, + 0.05781489610671997, + 0.053096771240234375, + 0.13734064996242523, + 0.25316688418388367, + 0.22281700372695923, + 0.04313703253865242, + 0.026718905195593834 + ], + "credit_map": [ + 0.009458164684474468, + 0.009548977017402649, + 0.009388569742441177, + 0.0137501060962677, + 0.01260738167911768, + 0.017601795494556427, + 0.011620673350989819, + 0.017853960394859314, + 0.00681740278378129, + 0.009761105291545391, + 0.00612562196329236, + 0.01386901643127203, + 0.006466330960392952, + 0.013956602662801743, + 0.007933856919407845, + 0.00994105450809002, + 0.003065490862354636, + 0.0026707309298217297, + 0.00388204213231802, + 0.004333582241088152, + 0.0027998341247439384, + 0.00476354593411088, + 0.003805245505645871, + 0.003886751364916563, + 0.0042608994990587234, + 0.005291125271469355, + 0.010405294597148895, + 0.007034459616988897, + 0.006427607964724302, + 0.008670712821185589, + 0.008019756525754929, + 0.007705038879066706, + 0.004688233137130737, + 0.0039038751274347305, + 0.007023641373962164, + 0.006187614984810352, + 0.007458842825144529, + 0.009239215403795242, + 0.006442824378609657, + 0.008152525871992111, + 0.01347413007169962, + 0.010844992473721504, + 0.021781230345368385, + 0.023799115791916847, + 0.01634589023888111, + 0.016910335049033165, + 0.02165444940328598, + 0.012530510313808918, + 0.02094237692654133, + 0.022690748795866966, + 0.03901592642068863, + 0.039103373885154724, + 0.02328304760158062, + 0.04108972102403641, + 0.03852786868810654, + 0.028513815253973007, + 0.021733136847615242, + 0.014841373078525066, + 0.046299975365400314, + 0.026703625917434692, + 0.02033986896276474, + 0.035069409757852554, + 0.029821328818798065, + 0.02800828590989113, + 0.003859501099213958, + 0.0035397824831306934, + 0.008268581703305244, + 0.006344453431665897, + 0.003979804459959269, + 0.005955401808023453, + 0.006826438941061497, + 0.004363072104752064, + 0.002210343023762107, + 0.0022357741836458445, + 0.005143216345459223, + 0.003991005010902882, + 0.0023456949274986982, + 0.004078810103237629, + 0.003955930005759001, + 0.0027581299655139446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.04869326949119568 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00675205560401082 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.00675205560401082, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00675205560401082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20611239969730377, + 0.1750337928533554, + 0.056235119700431824, + 0.04908084496855736, + 0.09041331708431244, + 0.07777183502912521, + 0.13911859691143036, + 0.11048492044210434, + 0.04278700053691864, + 0.05296223610639572 + ], + "credit_map": [ + 0.016139904037117958, + 0.02272689715027809, + 0.03645981475710869, + 0.02520667016506195, + 0.02021184004843235, + 0.0328046940267086, + 0.023154202848672867, + 0.029408369213342667, + 0.013567492365837097, + 0.021715806797146797, + 0.019742680713534355, + 0.024242641404271126, + 0.016145411878824234, + 0.030684154480695724, + 0.023446012288331985, + 0.02548958733677864, + 0.005026704631745815, + 0.0063339583575725555, + 0.0054873013868927956, + 0.007069762796163559, + 0.007600819692015648, + 0.009175139479339123, + 0.006703765131533146, + 0.008837672881782055, + 0.0038105593994259834, + 0.005920137278735638, + 0.00804058089852333, + 0.005419138353317976, + 0.007016465999186039, + 0.0065756626427173615, + 0.005299712065607309, + 0.00699858833104372, + 0.009362387470901012, + 0.005199286621063948, + 0.008230999112129211, + 0.015332378447055817, + 0.0068653482012450695, + 0.016432229429483414, + 0.014705886133015156, + 0.014284803532063961, + 0.0076281349174678326, + 0.006448489148169756, + 0.016387484967708588, + 0.010011745616793633, + 0.008569562807679176, + 0.012316692620515823, + 0.008603567257523537, + 0.007806154899299145, + 0.011378349736332893, + 0.01209165621548891, + 0.02309713326394558, + 0.020987486466765404, + 0.013493070378899574, + 0.020405299961566925, + 0.0215108972042799, + 0.016154693439602852, + 0.011199763044714928, + 0.007639974355697632, + 0.02294546365737915, + 0.013396484777331352, + 0.009737006388604641, + 0.017269909381866455, + 0.01414340641349554, + 0.014152915216982365, + 0.004412002395838499, + 0.0035351093392819166, + 0.006991671863943338, + 0.006460933014750481, + 0.004261259455233812, + 0.00526068452745676, + 0.0073150815442204475, + 0.004550258629024029, + 0.005230990704149008, + 0.004244126379489899, + 0.012035325169563293, + 0.008012166246771812, + 0.0039609638042747974, + 0.006139978300780058, + 0.007854191586375237, + 0.005484494846314192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.00675205560401082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811348229646683 + ], + "opposite_logprob_tokens": [ + -3.288113594055176 + ], + "logprob_reward": -0.03811348229646683, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811348229646683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5311939120292664, + 0.09124518185853958, + 0.08995871245861053, + 0.033304572105407715, + 0.06610998511314392, + 0.037673819810152054, + 0.05316157639026642, + 0.048232097178697586, + 0.02826959639787674, + 0.020850572735071182 + ], + "credit_map": [ + 0.06327859312295914, + 0.09192202985286713, + 0.03878352418541908, + 0.03243587166070938, + 0.043546512722969055, + 0.1511007696390152, + 0.025996005162596703, + 0.08413060754537582, + 0.007674520369619131, + 0.01548414584249258, + 0.00953995157033205, + 0.012031720951199532, + 0.010376455262303352, + 0.017413850873708725, + 0.0076860468834638596, + 0.011038493365049362, + 0.006537879351526499, + 0.009376255795359612, + 0.011183217167854309, + 0.01624407432973385, + 0.011342842131853104, + 0.01213453896343708, + 0.013242986053228378, + 0.009896915405988693, + 0.0028294299263507128, + 0.0024132903199642897, + 0.00492780702188611, + 0.005566296633332968, + 0.0028042292688041925, + 0.0054667918011546135, + 0.005183709319680929, + 0.004113015253096819, + 0.00654846103861928, + 0.004956423304975033, + 0.013630329631268978, + 0.008877812884747982, + 0.0072747464291751385, + 0.008219501934945583, + 0.007686818018555641, + 0.008915897458791733, + 0.004532793071120977, + 0.0033404759597033262, + 0.005179357714951038, + 0.005472367163747549, + 0.004534575622528791, + 0.006249119061976671, + 0.004741927608847618, + 0.003623203607276082, + 0.004527799319475889, + 0.004941806197166443, + 0.009634753689169884, + 0.007559769321233034, + 0.004833196755498648, + 0.007908452302217484, + 0.007699552923440933, + 0.0060562449507415295, + 0.00403290381655097, + 0.0040354616940021515, + 0.010687737725675106, + 0.005908991675823927, + 0.004624686669558287, + 0.0067329565063118935, + 0.0064383624121546745, + 0.005770993418991566, + 0.002087322762235999, + 0.0025366845075041056, + 0.004852165933698416, + 0.004238880705088377, + 0.0027185052167624235, + 0.004532580263912678, + 0.0037419768050312996, + 0.003561479039490223, + 0.001980732660740614, + 0.0018373171333223581, + 0.004274704027920961, + 0.0030961979646235704, + 0.0019527820404618979, + 0.0027669325936585665, + 0.0029725530184805393, + 0.0019693546928465366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03811348229646683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581373944878578 + ], + "opposite_logprob_tokens": [ + -3.8955814838409424 + ], + "logprob_reward": -0.020581373944878578, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581373944878578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5499573945999146, + 0.200143501162529, + 0.04496987909078598, + 0.012078848667442799, + 0.033853188157081604, + 0.02458755113184452, + 0.0259088221937418, + 0.06681261956691742, + 0.019468778744339943, + 0.022219400852918625 + ], + "credit_map": [ + 0.06204855814576149, + 0.09536126255989075, + 0.04099530726671219, + 0.04079750180244446, + 0.04578336328268051, + 0.15165975689888, + 0.03097187541425228, + 0.08233977109193802, + 0.021014004945755005, + 0.045409195125103, + 0.006740464363247156, + 0.010762177407741547, + 0.021748675033450127, + 0.05754990503191948, + 0.0063668047077953815, + 0.030552301555871964, + 0.0036306215915828943, + 0.004921832121908665, + 0.005229735281318426, + 0.008842570707201958, + 0.004954901989549398, + 0.006590580567717552, + 0.005968634504824877, + 0.004831002559512854, + 0.0011208115611225367, + 0.0008970772032625973, + 0.0020893102046102285, + 0.002429744927212596, + 0.000905097636859864, + 0.00176254624966532, + 0.0015559461899101734, + 0.001318314578384161, + 0.0033632172271609306, + 0.0028425329364836216, + 0.006418857257813215, + 0.005853971932083368, + 0.0037468811497092247, + 0.004146073944866657, + 0.004224888049066067, + 0.0032567649614065886, + 0.00206793867982924, + 0.0019692545756697655, + 0.004486276302486658, + 0.0034431028179824352, + 0.0026779272593557835, + 0.002931752474978566, + 0.003966444171965122, + 0.003044852986931801, + 0.002624188084155321, + 0.0017921815160661936, + 0.00434655137360096, + 0.004552171099931002, + 0.0024554915726184845, + 0.0032272550743073225, + 0.004296121187508106, + 0.002614862285554409, + 0.0054296222515404224, + 0.005571859888732433, + 0.01456107385456562, + 0.009542636573314667, + 0.0053558130748569965, + 0.009794739075005054, + 0.009922265075147152, + 0.006634610239416361, + 0.0016706119058653712, + 0.0018049038480967283, + 0.004024664405733347, + 0.0026207370683550835, + 0.001900549978017807, + 0.0025955575983971357, + 0.002594485878944397, + 0.0022572672460228205, + 0.0022582586389034986, + 0.0018309593433514237, + 0.004128067754209042, + 0.0034682287368923426, + 0.0018657423788681626, + 0.0026621457654982805, + 0.0036780242808163166, + 0.0023279725573956966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.020581373944878578 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up shot, against a backdrop of pure, bottomless darkness, a highly focused, sharp-edged cold-white spotlight shines diagonally from above, precisely illuminating a corner of a recently excavated ancient bronze sacrificial vessel. The vessel's surface is covered in uneven bluish-green patina, intricately carved with mythical beasts, and crevices of dried earth are embedded in its crevices. Shot in extreme slow motion, countless tiny particles of dust and sand are gently awakened by an invisible air current, slowly rising from the intricate veins of the bronze vessel. Each tiny particle is clearly captured in the beam, shimmering with a fleeting, warm golden glow. As if imbued with life, they spiral upward with an almost hypnotic grace, dispersing and ultimately dissolving into the surrounding darkness. The entire scene evokes a sense of tranquility and solemnity, as if frozen in time, silently telling the story of a long-forgotten history.", + "seed": 24729, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_9288_seed_24729_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17382717728614808, + "vlm_logprob": -0.17382717728614808, + "vlm_yes_no_margin": 2.425, + "vlm_token_credit": -0.17382717728614808, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06722187250852585 + ], + "opposite_logprob_tokens": [ + -2.8172218799591064 + ], + "logprob_reward": -0.06722187250852585, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06722187250852585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3366168141365051, + 0.09800896048545837, + 0.040704309940338135, + 0.11677129566669464, + 0.1705406904220581, + 0.03808828815817833, + 0.06312579661607742, + 0.021604478359222412, + 0.07972090691328049, + 0.034818392246961594 + ], + "credit_map": [ + 0.022408269345760345, + 0.026070602238178253, + 0.045587897300720215, + 0.019064830616116524, + 0.01516628172248602, + 0.08460845798254013, + 0.016402119770646095, + 0.10730836540460587, + 0.006722180172801018, + 0.007611272390931845, + 0.008412584662437439, + 0.006015409715473652, + 0.005605967715382576, + 0.027919892221689224, + 0.006766151636838913, + 0.028955509886145592, + 0.003661554306745529, + 0.00995325855910778, + 0.006391961593180895, + 0.00274827191606164, + 0.0023180765565484762, + 0.006312548648566008, + 0.001813191338442266, + 0.007505450397729874, + 0.01297531183809042, + 0.022182166576385498, + 0.024790529161691666, + 0.009502985514700413, + 0.006982891820371151, + 0.011199473403394222, + 0.007098134607076645, + 0.022039813920855522, + 0.00832508597522974, + 0.008671075105667114, + 0.05016019195318222, + 0.021585850045084953, + 0.008393838070333004, + 0.04313542693853378, + 0.010226486250758171, + 0.020042747259140015, + 0.006204182282090187, + 0.004852318670600653, + 0.009852534160017967, + 0.0033610162790864706, + 0.002968647051602602, + 0.005531474016606808, + 0.0021335603669285774, + 0.0031845574267208576, + 0.008251280523836613, + 0.007286444306373596, + 0.01079599093645811, + 0.01201930083334446, + 0.005809747613966465, + 0.005252208560705185, + 0.006271280813962221, + 0.007439547684043646, + 0.003033173969015479, + 0.003908054903149605, + 0.003713163547217846, + 0.003264338942244649, + 0.0018915547989308834, + 0.0019102379446849227, + 0.00179675139952451, + 0.0020872042514383793, + 0.006713473238050938, + 0.009240834042429924, + 0.010210118256509304, + 0.014373845420777798, + 0.006836226675659418, + 0.00878841895610094, + 0.014109103009104729, + 0.009448894299566746, + 0.004021761938929558, + 0.006857424974441528, + 0.00675453944131732, + 0.006015479099005461, + 0.003407512092962861, + 0.0029638248961418867, + 0.002449987456202507, + 0.002347867004573345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06722187250852585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.63277667760849 + ], + "opposite_logprob_tokens": [ + -0.75777667760849 + ], + "logprob_reward": -0.63277667760849, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.63277667760849, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06860195100307465, + 0.06951817870140076, + 0.02410082332789898, + 0.06628263741731644, + 0.1992621272802353, + 0.22681495547294617, + 0.11701815575361252, + 0.038165803998708725, + 0.018246697261929512, + 0.17198871076107025 + ], + "credit_map": [ + 0.004973117262125015, + 0.008596613071858883, + 0.008614198304712772, + 0.0049758609384298325, + 0.004147025290876627, + 0.01700304076075554, + 0.00868096575140953, + 0.011611135676503181, + 0.004872797057032585, + 0.004922256805002689, + 0.011290138587355614, + 0.00904129445552826, + 0.005702857859432697, + 0.010885573923587799, + 0.0056585893034935, + 0.017144668847322464, + 0.0016665196744725108, + 0.0030566419009119272, + 0.001959976740181446, + 0.0031266305595636368, + 0.0017216021660715342, + 0.005554180592298508, + 0.001913422136567533, + 0.005101851187646389, + 0.007122401613742113, + 0.014183112420141697, + 0.008717456832528114, + 0.006235379725694656, + 0.005869369022548199, + 0.010729015804827213, + 0.005639656446874142, + 0.007786245550960302, + 0.016090013086795807, + 0.03120359592139721, + 0.031199445948004723, + 0.02748139016330242, + 0.011630626395344734, + 0.044833701103925705, + 0.01295917946845293, + 0.02386416122317314, + 0.051079656928777695, + 0.052597761154174805, + 0.02270209975540638, + 0.02004690282046795, + 0.012761075049638748, + 0.029758550226688385, + 0.008015329018235207, + 0.02985355444252491, + 0.012824689038097858, + 0.01654765009880066, + 0.011131898500025272, + 0.02306380681693554, + 0.013735413551330566, + 0.010356026701629162, + 0.013783717527985573, + 0.015574954450130463, + 0.0043471334502100945, + 0.003866144921630621, + 0.0069086565636098385, + 0.007177100516855717, + 0.004027672111988068, + 0.003141460008919239, + 0.003680094378069043, + 0.005017539020627737, + 0.002230615122243762, + 0.002064373577013612, + 0.0027659551706165075, + 0.004096470773220062, + 0.0015402563149109483, + 0.0015218477929010987, + 0.0019095523748546839, + 0.002117625204846263, + 0.0222567617893219, + 0.037601977586746216, + 0.032377470284700394, + 0.02508164569735527, + 0.011682301759719849, + 0.016103222966194153, + 0.010450595989823341, + 0.01643473096191883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.63277667760849 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297845095396042 + ], + "opposite_logprob_tokens": [ + -2.237978458404541 + ], + "logprob_reward": -0.11297845095396042, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297845095396042, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21542736887931824, + 0.10746537148952484, + 0.07990161329507828, + 0.016520118340849876, + 0.2822897136211395, + 0.0560699999332428, + 0.0600549578666687, + 0.01693500392138958, + 0.12531119585037231, + 0.04002467915415764 + ], + "credit_map": [ + 0.00966761726886034, + 0.010265693068504333, + 0.012342709116637707, + 0.008580543100833893, + 0.010454871691763401, + 0.07462425529956818, + 0.011294791474938393, + 0.07819685339927673, + 0.007710983976721764, + 0.006445920094847679, + 0.013781552202999592, + 0.007874742150306702, + 0.0074342889711260796, + 0.028738124296069145, + 0.00976244080811739, + 0.025717303156852722, + 0.006160834804177284, + 0.014006918296217918, + 0.004403851926326752, + 0.006168600171804428, + 0.0039282189682126045, + 0.02584586665034294, + 0.003926070407032967, + 0.015461241826415062, + 0.0011568929767236114, + 0.0013023715000599623, + 0.003852937137708068, + 0.002148885279893875, + 0.0013841765467077494, + 0.0022425795905292034, + 0.0012269600993022323, + 0.0032053126487880945, + 0.011651474051177502, + 0.015276587568223476, + 0.09928055107593536, + 0.03563065826892853, + 0.012447510845959187, + 0.06481248140335083, + 0.016488829627633095, + 0.026701603084802628, + 0.014056208543479443, + 0.012166917324066162, + 0.006232522893697023, + 0.0061524491757154465, + 0.0036568723153322935, + 0.0053739603608846664, + 0.0034323360305279493, + 0.004998727701604366, + 0.007458788342773914, + 0.006664463318884373, + 0.00952286645770073, + 0.011283430270850658, + 0.005055208690464497, + 0.006164227612316608, + 0.0052971914410591125, + 0.00860877986997366, + 0.002130447421222925, + 0.0016710995696485043, + 0.003305599559098482, + 0.0015039079589769244, + 0.0017318004975095391, + 0.002691781148314476, + 0.0017444462282583117, + 0.002155918860808015, + 0.011347428895533085, + 0.014335498213768005, + 0.016998693346977234, + 0.02435411512851715, + 0.010262428782880306, + 0.013050965033471584, + 0.020481599494814873, + 0.014480458572506905, + 0.004017522092908621, + 0.009437063708901405, + 0.0076788924634456635, + 0.005186653230339289, + 0.003264760598540306, + 0.00405623996630311, + 0.003002536715939641, + 0.0033810085151344538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.11297845095396042 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980630472302437 + ], + "opposite_logprob_tokens": [ + -3.52980637550354 + ], + "logprob_reward": -0.02980630472302437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980630472302437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05559571459889412, + 0.045669134706258774, + 0.05562237650156021, + 0.0328533798456192, + 0.39170780777931213, + 0.1627129465341568, + 0.10414452105760574, + 0.027912989258766174, + 0.08062025904655457, + 0.043160878121852875 + ], + "credit_map": [ + 0.00463527487590909, + 0.005630976054817438, + 0.010655143298208714, + 0.004241469316184521, + 0.0038848333060741425, + 0.008238865062594414, + 0.004849940538406372, + 0.013459213078022003, + 0.006210614927113056, + 0.00604685302823782, + 0.009138686582446098, + 0.003921380266547203, + 0.004264906048774719, + 0.00601150980219245, + 0.004693364258855581, + 0.00538182258605957, + 0.00282512279227376, + 0.004431774374097586, + 0.004058578982949257, + 0.005082177463918924, + 0.0024948602076619864, + 0.017060937359929085, + 0.002548136981204152, + 0.017120786011219025, + 0.0020128258038312197, + 0.0026584789156913757, + 0.009235345758497715, + 0.004502090159803629, + 0.0028756449464708567, + 0.005131750833243132, + 0.002746226964518428, + 0.003691015299409628, + 0.021422602236270905, + 0.027681197971105576, + 0.1075180247426033, + 0.05610606446862221, + 0.020191047340631485, + 0.09265272319316864, + 0.02438484877347946, + 0.04175129532814026, + 0.009930516593158245, + 0.01639307290315628, + 0.05833595246076584, + 0.028585189953446388, + 0.006530401296913624, + 0.015394100919365883, + 0.010327338241040707, + 0.01721636950969696, + 0.01298609096556902, + 0.012866370379924774, + 0.015573345124721527, + 0.020343266427516937, + 0.008984727784991264, + 0.010887868702411652, + 0.008956929668784142, + 0.013545922935009003, + 0.0035494535695761442, + 0.004361325409263372, + 0.005672022234648466, + 0.003472954500466585, + 0.0018878537230193615, + 0.0033496280666440725, + 0.0026467344723641872, + 0.0029730189125984907, + 0.006884255912154913, + 0.010329494252800941, + 0.010648163966834545, + 0.014629420824348927, + 0.0069730766117572784, + 0.008590429089963436, + 0.013001314364373684, + 0.009564111940562725, + 0.006008241791278124, + 0.008768714964389801, + 0.005376605316996574, + 0.007851675152778625, + 0.003646659664809704, + 0.0035509623121470213, + 0.002951696515083313, + 0.005006324965506792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 6 + ], + "selected_reward": -0.02980630472302437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635258063673973 + ], + "opposite_logprob_tokens": [ + -3.651352643966675 + ], + "logprob_reward": -0.02635258063673973, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635258063673973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31159496307373047, + 0.047784533351659775, + 0.18403418362140656, + 0.034127939492464066, + 0.16152021288871765, + 0.06665798276662827, + 0.05779215693473816, + 0.017492905259132385, + 0.03216562420129776, + 0.08682957291603088 + ], + "credit_map": [ + 0.014114855788648129, + 0.013302725739777088, + 0.011629723943769932, + 0.01315518282353878, + 0.018682993948459625, + 0.11084023863077164, + 0.015918822959065437, + 0.11395040154457092, + 0.0047930520959198475, + 0.0039964038878679276, + 0.01057012751698494, + 0.005027687177062035, + 0.003327148500829935, + 0.005966912489384413, + 0.0036525866016745567, + 0.010450609028339386, + 0.008014368824660778, + 0.020496966317296028, + 0.006457164417952299, + 0.007144547998905182, + 0.005825628526508808, + 0.07310383021831512, + 0.004992892500013113, + 0.057998765259981155, + 0.004120481666177511, + 0.009143590927124023, + 0.008424344472587109, + 0.0030894929077476263, + 0.002121338853612542, + 0.0021361601538956165, + 0.0021224638912826777, + 0.0029700614977627993, + 0.014795429073274136, + 0.02792043425142765, + 0.02138272114098072, + 0.027696803212165833, + 0.010439976118505001, + 0.030102208256721497, + 0.011317440308630466, + 0.01786518283188343, + 0.014092431403696537, + 0.008851863443851471, + 0.010357985273003578, + 0.005831751972436905, + 0.0038954427000135183, + 0.009041808545589447, + 0.004702453035861254, + 0.009884237311780453, + 0.007006769999861717, + 0.0062098195776343346, + 0.008402850478887558, + 0.011307517997920513, + 0.005156498868018389, + 0.006335482932627201, + 0.005548502318561077, + 0.007824708707630634, + 0.002127847634255886, + 0.001564436941407621, + 0.0027731070294976234, + 0.003106652060523629, + 0.0013564948458224535, + 0.0023904494009912014, + 0.0020384322851896286, + 0.002135481685400009, + 0.003307559760287404, + 0.0030882139690220356, + 0.003338194452226162, + 0.006470264866948128, + 0.0031909295357763767, + 0.0033711432479321957, + 0.006444247905164957, + 0.0029550681356340647, + 0.009932609274983406, + 0.02080414444208145, + 0.0148426853120327, + 0.012512669898569584, + 0.006475386675447226, + 0.008664385415613651, + 0.005330579821020365, + 0.008267106488347054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.02635258063673973 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In an extreme close-up shot, against a backdrop of pure, bottomless darkness, a highly focused, sharp-edged cold-white spotlight shines diagonally from above, precisely illuminating a corner of a recently excavated ancient bronze sacrificial vessel. The vessel's surface is covered in uneven bluish-green patina, intricately carved with mythical beasts, and crevices of dried earth are embedded in its crevices. Shot in extreme slow motion, countless tiny particles of dust and sand are gently awakened by an invisible air current, slowly rising from the intricate veins of the bronze vessel. Each tiny particle is clearly captured in the beam, shimmering with a fleeting, warm golden glow. As if imbued with life, they spiral upward with an almost hypnotic grace, dispersing and ultimately dissolving into the surrounding darkness. The entire scene evokes a sense of tranquility and solemnity, as if frozen in time, silently telling the story of a long-forgotten history.", + "seed": 24730, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_8/index_9288_seed_24730_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20464124996215105, + "vlm_logprob": -0.20464124996215105, + "vlm_yes_no_margin": 2.8500000119209288, + "vlm_token_credit": -0.20464124996215105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3531438708305359 + ], + "opposite_logprob_tokens": [ + -1.2281439304351807 + ], + "logprob_reward": -0.3531438708305359, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3531438708305359, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3462391197681427, + 0.23179209232330322, + 0.10789643973112106, + 0.050980888307094574, + 0.03834086284041405, + 0.051477186381816864, + 0.04967164620757103, + 0.03699226304888725, + 0.03499665856361389, + 0.05161289498209953 + ], + "credit_map": [ + 0.02799605019390583, + 0.029837124049663544, + 0.03631431236863136, + 0.025037573650479317, + 0.05481882393360138, + 0.09222301840782166, + 0.028588566929101944, + 0.05142366141080856, + 0.016137180849909782, + 0.014264529570937157, + 0.022624360397458076, + 0.015219475142657757, + 0.024678796529769897, + 0.07751935720443726, + 0.014324383810162544, + 0.04702400788664818, + 0.007419858127832413, + 0.010409007780253887, + 0.011730260215699673, + 0.006849530152976513, + 0.027858415618538857, + 0.020239772275090218, + 0.006721249781548977, + 0.016668347641825676, + 0.003058589994907379, + 0.0037627415731549263, + 0.010174949653446674, + 0.0035044532269239426, + 0.005132423248142004, + 0.012450071983039379, + 0.004096883349120617, + 0.008800777606666088, + 0.002666821703314781, + 0.0031194197945296764, + 0.003990427125245333, + 0.003593089524656534, + 0.007189963944256306, + 0.008455445989966393, + 0.0028879514429718256, + 0.006437742151319981, + 0.0038736157584935427, + 0.004867084324359894, + 0.007062744349241257, + 0.0036456799134612083, + 0.009141235612332821, + 0.010095307603478432, + 0.004286463838070631, + 0.00850505381822586, + 0.004902367014437914, + 0.004517235327512026, + 0.007248386275023222, + 0.00595428328961134, + 0.008599565364420414, + 0.006756279151886702, + 0.00537045719102025, + 0.00632307305932045, + 0.003635939210653305, + 0.0032085131388157606, + 0.003936019726097584, + 0.003017348237335682, + 0.007746623829007149, + 0.005889206659048796, + 0.004078187979757786, + 0.005480424501001835, + 0.003895623842254281, + 0.002659748075529933, + 0.00580567168071866, + 0.003129528136923909, + 0.003049988066777587, + 0.007534250151365995, + 0.0033568127546459436, + 0.005565036553889513, + 0.005530881229788065, + 0.003946550656110048, + 0.008998324163258076, + 0.004392519127577543, + 0.005601576529443264, + 0.008818689733743668, + 0.004848218057304621, + 0.009476135484874249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.3531438708305359 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327847838401794 + ], + "opposite_logprob_tokens": [ + -0.7577847838401794 + ], + "logprob_reward": -0.6327847838401794, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327847838401794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20025967061519623, + 0.12853428721427917, + 0.07057800889015198, + 0.0836564302444458, + 0.04684305563569069, + 0.10336045920848846, + 0.22986595332622528, + 0.06603817641735077, + 0.04500734806060791, + 0.025856604799628258 + ], + "credit_map": [ + 0.018673134967684746, + 0.013312211260199547, + 0.025868333876132965, + 0.016320830211043358, + 0.03163694962859154, + 0.046085257083177567, + 0.017633002251386642, + 0.030729951336979866, + 0.010511409491300583, + 0.018304787576198578, + 0.013577951118350029, + 0.010656017810106277, + 0.0319356806576252, + 0.01683083549141884, + 0.01061411015689373, + 0.01610349677503109, + 0.005942839663475752, + 0.0055389502085745335, + 0.009857823140919209, + 0.005271679721772671, + 0.009809937328100204, + 0.017262214794754982, + 0.005772041156888008, + 0.011122524738311768, + 0.005712334997951984, + 0.007699917070567608, + 0.012651844881474972, + 0.007268710061907768, + 0.017739664763212204, + 0.013860691338777542, + 0.007228486239910126, + 0.011494779027998447, + 0.0033178841695189476, + 0.004940965678542852, + 0.006303488742560148, + 0.004380236379802227, + 0.009400537237524986, + 0.007115338928997517, + 0.004349174909293652, + 0.007035430520772934, + 0.009295015595853329, + 0.010695523582398891, + 0.013158713467419147, + 0.007711455225944519, + 0.014820828102529049, + 0.017225684598088264, + 0.006255849730223417, + 0.02419738657772541, + 0.018848106265068054, + 0.02068338729441166, + 0.024688929319381714, + 0.01838870719075203, + 0.048008453100919724, + 0.041313741356134415, + 0.022582460194826126, + 0.03535216301679611, + 0.006779506802558899, + 0.006819562055170536, + 0.00618870509788394, + 0.005362303461879492, + 0.01046459749341011, + 0.01397953275591135, + 0.005555608309805393, + 0.010888359509408474, + 0.0038472923915833235, + 0.003936543595045805, + 0.007280031219124794, + 0.0044061243534088135, + 0.005641604773700237, + 0.00851427111774683, + 0.0036974167451262474, + 0.007684062700718641, + 0.002669732552021742, + 0.0025572930462658405, + 0.003991104196757078, + 0.002280999906361103, + 0.0039375778287649155, + 0.0036913608200848103, + 0.002732247579842806, + 0.003996289800852537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.6327847838401794 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016092650592327118 + ], + "opposite_logprob_tokens": [ + -4.141092777252197 + ], + "logprob_reward": -0.016092650592327118, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016092650592327118, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10574736446142197, + 0.2426893562078476, + 0.30569592118263245, + 0.04392512887716293, + 0.034219011664390564, + 0.17143802344799042, + 0.023965105414390564, + 0.02761773392558098, + 0.014485505409538746, + 0.030216844752430916 + ], + "credit_map": [ + 0.008564203977584839, + 0.00812514964491129, + 0.012606808915734291, + 0.009012540802359581, + 0.013957435265183449, + 0.028194449841976166, + 0.008158275857567787, + 0.01712850108742714, + 0.016176756471395493, + 0.02386995404958725, + 0.02851899340748787, + 0.01752622053027153, + 0.05526351183652878, + 0.0490388385951519, + 0.016634685918688774, + 0.03566037118434906, + 0.02567591331899166, + 0.0323379822075367, + 0.03969339653849602, + 0.020563313737511635, + 0.06987559050321579, + 0.05454153195023537, + 0.01931382715702057, + 0.04369436576962471, + 0.0027964895125478506, + 0.00349981477484107, + 0.0067461696453392506, + 0.0032960399985313416, + 0.0060639167204499245, + 0.011363943107426167, + 0.003224044106900692, + 0.006934710778295994, + 0.002667160239070654, + 0.0024856762029230595, + 0.004905427806079388, + 0.0037432806566357613, + 0.005336224101483822, + 0.006362958811223507, + 0.0030327874701470137, + 0.005685495212674141, + 0.012895732186734676, + 0.0200502946972847, + 0.018974797800183296, + 0.013384227640926838, + 0.023126086220145226, + 0.03465051203966141, + 0.011145171709358692, + 0.03721121326088905, + 0.001923697884194553, + 0.002732129069045186, + 0.0025291668716818094, + 0.002268379321321845, + 0.004461455624550581, + 0.004014099016785622, + 0.0027954091783612967, + 0.0032407683320343494, + 0.002652534982189536, + 0.002447480335831642, + 0.003738746978342533, + 0.0027665020897984505, + 0.0033542721066623926, + 0.004300692584365606, + 0.0029548967722803354, + 0.005402610637247562, + 0.001285047736018896, + 0.0012970310635864735, + 0.0014214678667485714, + 0.0011283011408522725, + 0.0030758422799408436, + 0.002757970942184329, + 0.001370889600366354, + 0.0021489541977643967, + 0.002880754182115197, + 0.002695887116715312, + 0.004626247566193342, + 0.0022369250655174255, + 0.004738247487694025, + 0.0053528184071183205, + 0.002998191164806485, + 0.004687773995101452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 5 + ], + "selected_reward": -0.016092650592327118 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543358840048313 + ], + "opposite_logprob_tokens": [ + -4.387543201446533 + ], + "logprob_reward": -0.012543358840048313, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543358840048313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1396416872739792, + 0.3725263178348541, + 0.2815909683704376, + 0.03915972635149956, + 0.024818530306220055, + 0.04320592060685158, + 0.03896082937717438, + 0.019911088049411774, + 0.023956889286637306, + 0.016227994114160538 + ], + "credit_map": [ + 0.011256679892539978, + 0.012754681520164013, + 0.01528976857662201, + 0.00919931661337614, + 0.026087334379553795, + 0.034456171095371246, + 0.011480100452899933, + 0.01911764219403267, + 0.024549204856157303, + 0.02891303412616253, + 0.028701167553663254, + 0.01725296676158905, + 0.056218162178993225, + 0.12329713255167007, + 0.021085625514388084, + 0.07250902056694031, + 0.015629401430487633, + 0.01948045752942562, + 0.02173253707587719, + 0.01080558355897665, + 0.03838541731238365, + 0.1032540574669838, + 0.015239497646689415, + 0.05706401914358139, + 0.0031964234076440334, + 0.002737724222242832, + 0.008854356594383717, + 0.0038758767768740654, + 0.003804606618359685, + 0.0071002584882080555, + 0.003238800447434187, + 0.006351680960506201, + 0.0018537717405706644, + 0.0020462223328649998, + 0.003138791536912322, + 0.002473230008035898, + 0.004044339992105961, + 0.005047227721661329, + 0.0025255107320845127, + 0.0036894388031214476, + 0.003657426219433546, + 0.0046384865418076515, + 0.005453640595078468, + 0.004067361354827881, + 0.006405150983482599, + 0.007708118297159672, + 0.002500806236639619, + 0.008774926885962486, + 0.003487402107566595, + 0.002786915050819516, + 0.004178697243332863, + 0.002726026577875018, + 0.007130081299692392, + 0.008322502486407757, + 0.0034800341818481684, + 0.006849170196801424, + 0.0025987846311181784, + 0.0015894565731287003, + 0.0030037560500204563, + 0.0014286048244684935, + 0.003340406809002161, + 0.003264585044234991, + 0.0017699659802019596, + 0.0029155283700674772, + 0.003016202710568905, + 0.0018637411994859576, + 0.003915087319910526, + 0.002415106166154146, + 0.0037175766192376614, + 0.003609933890402317, + 0.0020696327555924654, + 0.003349610371515155, + 0.002160344272851944, + 0.0013870502589270473, + 0.0030464837327599525, + 0.0013669014442712069, + 0.001295948983170092, + 0.0020777250174432993, + 0.0016000716714188457, + 0.0032934683840721846 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.012543358840048313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot framed as an extreme close-up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the background pure and bottomless darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a cold-white spotlight shining diagonally from above to illuminate an ancient bronze sacrificial vessel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny particles of dust and sand rising from the vessel's crevices in extreme slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the rising particles shimmer with a warm golden glow within the spotlight beam?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864158570766449 + ], + "opposite_logprob_tokens": [ + -4.758641719818115 + ], + "logprob_reward": -0.00864158570766449, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864158570766449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1005309522151947, + 0.31302616000175476, + 0.3608764708042145, + 0.0313880629837513, + 0.0657237321138382, + 0.03861520066857338, + 0.011056223884224892, + 0.0172687079757452, + 0.021719878539443016, + 0.03979455307126045 + ], + "credit_map": [ + 0.010051586665213108, + 0.010623062960803509, + 0.013151814229786396, + 0.007243549916893244, + 0.020722048357129097, + 0.017658647149801254, + 0.00863045547157526, + 0.01244980376213789, + 0.01992275007069111, + 0.019037261605262756, + 0.025543102994561195, + 0.012939173728227615, + 0.04664497822523117, + 0.10631987452507019, + 0.021350745111703873, + 0.061268310993909836, + 0.02207980491220951, + 0.027166204527020454, + 0.03283239156007767, + 0.01161145232617855, + 0.04162266477942467, + 0.13653354346752167, + 0.015435071662068367, + 0.07359535992145538, + 0.0025243896525353193, + 0.002938238438218832, + 0.0046385508030653, + 0.0024572403635829687, + 0.0061607263050973415, + 0.005927016958594322, + 0.0022860204335302114, + 0.004455880261957645, + 0.004202583339065313, + 0.007229957729578018, + 0.007119947113096714, + 0.00656399130821228, + 0.012626751326024532, + 0.012544253841042519, + 0.005360681563615799, + 0.010075573809444904, + 0.003383910981938243, + 0.004364524967968464, + 0.004886160138994455, + 0.0035860773641616106, + 0.00632068794220686, + 0.00583532452583313, + 0.0023385563399642706, + 0.007899965159595013, + 0.0010598114458844066, + 0.0010460457997396588, + 0.0013147579738870263, + 0.001043028081767261, + 0.001709036179818213, + 0.0020396218169480562, + 0.0007961962837725878, + 0.0020477271173149347, + 0.00188344600610435, + 0.001492423820309341, + 0.0025230178143829107, + 0.0019966005347669125, + 0.0023435112088918686, + 0.0029727232176810503, + 0.001643490744754672, + 0.0024134949781000614, + 0.0032037897035479546, + 0.0019815026316791773, + 0.0035032860469073057, + 0.0022535237949341536, + 0.002735838992521167, + 0.00293072871863842, + 0.0022459879983216524, + 0.002865221118554473, + 0.004848039709031582, + 0.0033001794945448637, + 0.00621951837092638, + 0.0025684216525405645, + 0.006037399172782898, + 0.006045592948794365, + 0.0038658101111650467, + 0.006909594871103764 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00864158570766449 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_tensor([6504, 6504])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_tensor([6504, 6504])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f1207f1af3207ae1adebe8cdacac2c519ab10f86 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_tensor([6504, 6504])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "fixed, horizontal wide-angle shot captures a vibrant 8-bit pixel art concert. Center stage, an anthropomorphic USB drive composed of bright blue and gray pixels takes center stage, acting as conductor. With two simple white blocks for eyes, it stands atop a small, pixelated podium. Its body sways gently to an imaginary electronic beat. One pixelated arm holds a gleaming white baton, waving it up and down decisively and rhythmically. Another arm reaches to the side of the stage, cuing a pixelated band. This band is composed entirely of anthropomorphic 8-bit instruments: a red pixelated keyboard, its keys pulsing and glowing in time with the melody; a yellow pixelated drum set, its drumheads and cymbals emitting tiny white pixelated starbursts when struck by invisible sticks; and a green pixelated electric guitar, bobbing up and down to the rhythm of a passionate solo. The entire scene is bathed in a simulated pixelated spotlight. Deep purple beams cast sharp geometric shapes against a deep black background, dotted with slowly blinking pixels reminiscent of stars from early arcade games. The primary color palette of electric blue, magenta, and bright yellow creates a strong visual contrast, and all elements are rendered in crisp, non-anti-aliased pixel blocks, creating a nostalgic, dynamic, and playful atmosphere reminiscent of a retro gaming concert.", + "seed": 33509, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_6504_seed_33509_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.024852001294493675, + "vlm_logprob": -0.024852001294493675, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.024852001294493675, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02401234209537506 + ], + "opposite_logprob_tokens": [ + -3.774012327194214 + ], + "logprob_reward": -0.02401234209537506, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02401234209537506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33085715770721436, + 0.06273617595434189, + 0.05083904415369034, + 0.09436438232660294, + 0.09456797689199448, + 0.04229341074824333, + 0.034772053360939026, + 0.18298649787902832, + 0.05407530069351196, + 0.05250801891088486 + ], + "credit_map": [ + 0.03639288619160652, + 0.04312748834490776, + 0.02884593978524208, + 0.04093165695667267, + 0.047629598528146744, + 0.043418072164058685, + 0.06276927888393402, + 0.027742214500904083, + 0.00689469650387764, + 0.010950969532132149, + 0.006594319827854633, + 0.004994474351406097, + 0.009171769954264164, + 0.008776437491178513, + 0.006872282363474369, + 0.008481220342218876, + 0.0047654579393565655, + 0.005763535387814045, + 0.003438045736402273, + 0.004079658072441816, + 0.014161408878862858, + 0.008411026559770107, + 0.005869325250387192, + 0.00435057794675231, + 0.010021130554378033, + 0.015292285941541195, + 0.007215386256575584, + 0.009718367829918861, + 0.0148550346493721, + 0.012968883849680424, + 0.018333835527300835, + 0.005959441419690847, + 0.016053596511483192, + 0.014998584985733032, + 0.010178294032812119, + 0.00909503549337387, + 0.015625039115548134, + 0.010578880086541176, + 0.012433796189725399, + 0.005604736041277647, + 0.004538137000054121, + 0.005675457883626223, + 0.004125635605305433, + 0.005128254182636738, + 0.008042413741350174, + 0.005950205959379673, + 0.00526887271553278, + 0.0035644304007291794, + 0.0050992174074053764, + 0.0048883333802223206, + 0.003047959879040718, + 0.0031933682039380074, + 0.006367921829223633, + 0.004544723313301802, + 0.003742053871974349, + 0.003888470120728016, + 0.020481420680880547, + 0.03615208715200424, + 0.01363411545753479, + 0.01469060592353344, + 0.029703978449106216, + 0.023841597139835358, + 0.021286990493535995, + 0.023195680230855942, + 0.014742537401616573, + 0.007469731383025646, + 0.00433068722486496, + 0.00426948256790638, + 0.006851366721093655, + 0.008044527843594551, + 0.004625077825039625, + 0.0037418862339109182, + 0.0056839860044419765, + 0.010545259341597557, + 0.005750683601945639, + 0.004550976678729057, + 0.008157715201377869, + 0.005295701790601015, + 0.005364729557186365, + 0.007158961147069931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.02401234209537506 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078395880758762 + ], + "opposite_logprob_tokens": [ + -4.511078357696533 + ], + "logprob_reward": -0.011078395880758762, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078395880758762, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23595380783081055, + 0.07835885137319565, + 0.13352395594120026, + 0.10920432955026627, + 0.04955168813467026, + 0.043429531157016754, + 0.042185403406620026, + 0.14185357093811035, + 0.07957777380943298, + 0.08636108785867691 + ], + "credit_map": [ + 0.020075565204024315, + 0.02651316300034523, + 0.01900755614042282, + 0.028609558939933777, + 0.035780344158411026, + 0.02663162164390087, + 0.05741423740983009, + 0.021921761333942413, + 0.007037732284516096, + 0.012625355273485184, + 0.008153911679983139, + 0.008343875408172607, + 0.01273299939930439, + 0.009588688611984253, + 0.013709836639463902, + 0.006166450679302216, + 0.011906623840332031, + 0.015965482220053673, + 0.006864365190267563, + 0.0073896292597055435, + 0.03252732753753662, + 0.03695274889469147, + 0.012769205495715141, + 0.009148572571575642, + 0.011804021894931793, + 0.01941043511033058, + 0.010330140590667725, + 0.008525203913450241, + 0.01746480166912079, + 0.015234011225402355, + 0.01784035563468933, + 0.008595366962254047, + 0.0048528388142585754, + 0.005545671097934246, + 0.0047446079552173615, + 0.004163494799286127, + 0.009214756079018116, + 0.007153852377086878, + 0.008294902741909027, + 0.005581563804298639, + 0.005360724404454231, + 0.0058668917044997215, + 0.004027406685054302, + 0.005628217477351427, + 0.008087310008704662, + 0.006343713961541653, + 0.00482217688113451, + 0.003293090732768178, + 0.00748397083953023, + 0.007824706844985485, + 0.0054501756094396114, + 0.0027418883983045816, + 0.004819763824343681, + 0.0052051842212677, + 0.0044778501614928246, + 0.004181861877441406, + 0.01682153344154358, + 0.029445774853229523, + 0.012320879846811295, + 0.011831173673272133, + 0.020966237410902977, + 0.01776490919291973, + 0.01641603372991085, + 0.01628703437745571, + 0.016403770074248314, + 0.009618623182177544, + 0.007110001519322395, + 0.0066881002858281136, + 0.010962549597024918, + 0.011352214962244034, + 0.01023738831281662, + 0.007205127738416195, + 0.009725946933031082, + 0.011053663678467274, + 0.006159581243991852, + 0.007203350309282541, + 0.012973012402653694, + 0.009927121922373772, + 0.009744580835103989, + 0.01957383193075657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.011078395880758762 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072383262217045 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072383262217045, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072383262217045, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11209352314472198, + 0.03906415402889252, + 0.1958964318037033, + 0.11542460322380066, + 0.06077784299850464, + 0.04641076922416687, + 0.049009718000888824, + 0.20470842719078064, + 0.07081856578588486, + 0.1057959645986557 + ], + "credit_map": [ + 0.016223495826125145, + 0.015005059540271759, + 0.011173798702657223, + 0.009740586392581463, + 0.011858006939291954, + 0.012913845479488373, + 0.018967952579259872, + 0.016210779547691345, + 0.004921543411910534, + 0.004499271046370268, + 0.0041052429005503654, + 0.005006785038858652, + 0.005955887027084827, + 0.004372236784547567, + 0.005729886703193188, + 0.0044733029790222645, + 0.01861056312918663, + 0.016900306567549706, + 0.009231199510395527, + 0.008996762335300446, + 0.0489974319934845, + 0.06429266184568405, + 0.01796574704349041, + 0.01090175099670887, + 0.009509441442787647, + 0.018682263791561127, + 0.0105803357437253, + 0.010565550997853279, + 0.021086305379867554, + 0.015566697344183922, + 0.021588293835520744, + 0.007845723070204258, + 0.009457997977733612, + 0.008229979313910007, + 0.007196485996246338, + 0.004951789975166321, + 0.009409086778759956, + 0.007421470712870359, + 0.009994248859584332, + 0.0041167885065078735, + 0.007682349067181349, + 0.007075062952935696, + 0.004415961913764477, + 0.00435154652222991, + 0.008369873277842999, + 0.005360259674489498, + 0.005521555431187153, + 0.0036341617815196514, + 0.006543252617120743, + 0.007774633821099997, + 0.004430047236382961, + 0.005319226998835802, + 0.007063708268105984, + 0.006136866752058268, + 0.005634016823023558, + 0.0061079696752130985, + 0.021548137068748474, + 0.03840000182390213, + 0.017089108005166054, + 0.017869362607598305, + 0.031495220959186554, + 0.02664889767765999, + 0.025413330644369125, + 0.02624436840415001, + 0.01577036827802658, + 0.008582005277276039, + 0.006381649523973465, + 0.006686734035611153, + 0.0105089470744133, + 0.010413674637675285, + 0.006681023631244898, + 0.005794168915599585, + 0.01190178468823433, + 0.010724158957600594, + 0.005871423054486513, + 0.009340943768620491, + 0.009156310930848122, + 0.012823865748941898, + 0.014991674572229385, + 0.030985815450549126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.011072383262217045 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062034618109464645 + ], + "opposite_logprob_tokens": [ + -2.8120346069335938 + ], + "logprob_reward": -0.062034618109464645, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062034618109464645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19556045532226562, + 0.08329515159130096, + 0.27331414818763733, + 0.10293160378932953, + 0.055207595229148865, + 0.04167347401380539, + 0.037544868886470795, + 0.09212150424718857, + 0.0840558260679245, + 0.034295447170734406 + ], + "credit_map": [ + 0.031355470418930054, + 0.03189961984753609, + 0.02770490199327469, + 0.021974477916955948, + 0.022250743582844734, + 0.024114808067679405, + 0.020991429686546326, + 0.01526898704469204, + 0.010802830569446087, + 0.00837110634893179, + 0.009153860621154308, + 0.009671679697930813, + 0.010528124868869781, + 0.010394714772701263, + 0.014265256002545357, + 0.010107564739882946, + 0.025765670463442802, + 0.023673228919506073, + 0.01336683426052332, + 0.014440787956118584, + 0.06465743482112885, + 0.08796921372413635, + 0.02767951227724552, + 0.015761444345116615, + 0.007145989220589399, + 0.015195329673588276, + 0.009669331833720207, + 0.010840889066457748, + 0.01703120954334736, + 0.014696221798658371, + 0.020173607394099236, + 0.008179020136594772, + 0.0060876901261508465, + 0.006647411268204451, + 0.006615799386054277, + 0.00645051896572113, + 0.008404775522649288, + 0.008486527018249035, + 0.008075474761426449, + 0.004439391661435366, + 0.005273118149489164, + 0.006133507937192917, + 0.0034736567176878452, + 0.004277256783097982, + 0.009202249348163605, + 0.005906970240175724, + 0.004242717754095793, + 0.003163993824273348, + 0.006349822040647268, + 0.006273102015256882, + 0.0044179512187838554, + 0.0033408876042813063, + 0.004851127043366432, + 0.004670694470405579, + 0.0038465880788862705, + 0.0037946945521980524, + 0.008570948615670204, + 0.01556082721799612, + 0.0077865538187325, + 0.008256012573838234, + 0.015640510246157646, + 0.012225372716784477, + 0.012385725975036621, + 0.011695555411279202, + 0.01734526827931404, + 0.009881882928311825, + 0.005486310925334692, + 0.007299337070435286, + 0.012027760967612267, + 0.011464249342679977, + 0.012320838868618011, + 0.008230172097682953, + 0.003793054958805442, + 0.005081665236502886, + 0.003411776153370738, + 0.0046173036098480225, + 0.004045695532113314, + 0.003620967036113143, + 0.00276213139295578, + 0.00696285255253315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.062034618109464645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016062267124652863 + ], + "opposite_logprob_tokens": [ + -4.141062259674072 + ], + "logprob_reward": -0.016062267124652863, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016062267124652863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24173326790332794, + 0.10189110040664673, + 0.11061076074838638, + 0.15698128938674927, + 0.13911184668540955, + 0.06296635419130325, + 0.03925561532378197, + 0.03617320582270622, + 0.046013448387384415, + 0.06526319682598114 + ], + "credit_map": [ + 0.031099708750844002, + 0.035843901336193085, + 0.02976139262318611, + 0.029614847153425217, + 0.030623339116573334, + 0.028276441618800163, + 0.03940242901444435, + 0.017111174762248993, + 0.012673457153141499, + 0.013377909548580647, + 0.011732824146747589, + 0.009484907612204552, + 0.01555595826357603, + 0.012482233345508575, + 0.014747189357876778, + 0.011836611665785313, + 0.010693310759961605, + 0.014298207126557827, + 0.00690728472545743, + 0.00699630007147789, + 0.02674994058907032, + 0.026001831516623497, + 0.012372658587992191, + 0.00659121572971344, + 0.013081658631563187, + 0.02347959578037262, + 0.014922264963388443, + 0.012454980984330177, + 0.035037893801927567, + 0.02258204109966755, + 0.023711923509836197, + 0.01171091664582491, + 0.021815991029143333, + 0.01748870685696602, + 0.009941808879375458, + 0.013635210692882538, + 0.027555549517273903, + 0.020083963871002197, + 0.016587253659963608, + 0.012003342621028423, + 0.009536673314869404, + 0.006794070824980736, + 0.0063914707861840725, + 0.006346386391669512, + 0.01000922080129385, + 0.007901943288743496, + 0.006906750611960888, + 0.009079834446310997, + 0.00486102607101202, + 0.005198294762521982, + 0.005327221471816301, + 0.003517450764775276, + 0.004791388288140297, + 0.005202956963330507, + 0.004845776129513979, + 0.005511495750397444, + 0.004913533106446266, + 0.004372991621494293, + 0.005411752499639988, + 0.004714198876172304, + 0.004017378203570843, + 0.004197970032691956, + 0.004859249573200941, + 0.0036861274857074022, + 0.008565300144255161, + 0.00570228835567832, + 0.003241368569433689, + 0.004513056483119726, + 0.0080032330006361, + 0.006058010272681713, + 0.005102540832012892, + 0.004827646538615227, + 0.005775542464107275, + 0.010964342392981052, + 0.0077727483585476875, + 0.006398616824299097, + 0.006346260197460651, + 0.007329676300287247, + 0.00807653833180666, + 0.01259946171194315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.016062267124652863 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, horizontal wide-angle shot captures a vibrant 8-bit pixel art concert. Center stage, an anthropomorphic USB drive composed of bright blue and gray pixels takes center stage, acting as conductor. With two simple white blocks for eyes, it stands atop a small, pixelated podium. Its body sways gently to an imaginary electronic beat. One pixelated arm holds a gleaming white baton, waving it up and down decisively and rhythmically. Another arm reaches to the side of the stage, cuing a pixelated band. This band is composed entirely of anthropomorphic 8-bit instruments: a red pixelated keyboard, its keys pulsing and glowing in time with the melody; a yellow pixelated drum set, its drumheads and cymbals emitting tiny white pixelated starbursts when struck by invisible sticks; and a green pixelated electric guitar, bobbing up and down to the rhythm of a passionate solo. The entire scene is bathed in a simulated pixelated spotlight. Deep purple beams cast sharp geometric shapes against a deep black background, dotted with slowly blinking pixels reminiscent of stars from early arcade games. The primary color palette of electric blue, magenta, and bright yellow creates a strong visual contrast, and all elements are rendered in crisp, non-anti-aliased pixel blocks, creating a nostalgic, dynamic, and playful atmosphere reminiscent of a retro gaming concert.", + "seed": 33510, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_6504_seed_33510_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.009488896932452917, + "vlm_logprob": -0.009488896932452917, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.009488896932452917, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020916828885674477 + ], + "opposite_logprob_tokens": [ + -3.8959169387817383 + ], + "logprob_reward": -0.020916828885674477, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020916828885674477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13385772705078125, + 0.131138414144516, + 0.12445012480020523, + 0.04157505929470062, + 0.06787334382534027, + 0.1051778495311737, + 0.07872699946165085, + 0.22613461315631866, + 0.03116648457944393, + 0.059899408370256424 + ], + "credit_map": [ + 0.014152877032756805, + 0.014646094292402267, + 0.025348741561174393, + 0.010272222571074963, + 0.01888303831219673, + 0.018513744696974754, + 0.015030975453555584, + 0.017010025680065155, + 0.013767003081738949, + 0.026530161499977112, + 0.01041637733578682, + 0.022737620398402214, + 0.015746474266052246, + 0.016986504197120667, + 0.016610048711299896, + 0.008344228379428387, + 0.01368241012096405, + 0.019065622240304947, + 0.007853097282350063, + 0.019956344738602638, + 0.01626921258866787, + 0.013077357783913612, + 0.022906919941306114, + 0.011639149859547615, + 0.006378963589668274, + 0.008085494861006737, + 0.003012478118762374, + 0.004864122252911329, + 0.007152331992983818, + 0.005439084954559803, + 0.00413554348051548, + 0.0025070386473089457, + 0.006402785889804363, + 0.009026273153722286, + 0.006011639721691608, + 0.00959721952676773, + 0.014364541508257389, + 0.008790512569248676, + 0.007963203825056553, + 0.005717165768146515, + 0.007650517858564854, + 0.014402617700397968, + 0.006759136915206909, + 0.009363550692796707, + 0.025822876021265984, + 0.012749279849231243, + 0.019176453351974487, + 0.009253417141735554, + 0.008386601693928242, + 0.015485715121030807, + 0.005153625272214413, + 0.011604110710322857, + 0.014071903191506863, + 0.009035979397594929, + 0.009141686372458935, + 0.005847376771271229, + 0.01902078092098236, + 0.031255804002285004, + 0.013138850219547749, + 0.04234141483902931, + 0.03681647777557373, + 0.02573087252676487, + 0.03515081852674484, + 0.022679587826132774, + 0.003177615348249674, + 0.005282524041831493, + 0.0028511718846857548, + 0.005007960833609104, + 0.0028833013493567705, + 0.0038774926215410233, + 0.00378279946744442, + 0.004303618799895048, + 0.00618345569819212, + 0.008028553798794746, + 0.007647266611456871, + 0.007390650454908609, + 0.007605079561471939, + 0.006080395542085171, + 0.009213943965733051, + 0.007750063668936491 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.020916828885674477 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046372017823159695 + ], + "opposite_logprob_tokens": [ + -5.379637241363525 + ], + "logprob_reward": -0.0046372017823159695, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046372017823159695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.062496207654476166, + 0.15899083018302917, + 0.3375806510448456, + 0.07914473116397858, + 0.03968539088964462, + 0.04005436226725578, + 0.09540258347988129, + 0.040471505373716354, + 0.02837802655994892, + 0.11779573559761047 + ], + "credit_map": [ + 0.007869471795856953, + 0.008727206848561764, + 0.006806531455367804, + 0.008755892515182495, + 0.010851548053324223, + 0.009354246780276299, + 0.0059607550501823425, + 0.004170553293079138, + 0.017872167751193047, + 0.027008702978491783, + 0.011035308241844177, + 0.022068900987505913, + 0.029205571860074997, + 0.02172982506453991, + 0.017069270834326744, + 0.01300108153373003, + 0.03271859139204025, + 0.07119505107402802, + 0.016319604590535164, + 0.05325379967689514, + 0.051001131534576416, + 0.04161752387881279, + 0.0505499430000782, + 0.02092498168349266, + 0.012770040892064571, + 0.014135783538222313, + 0.006158466450870037, + 0.011457209475338459, + 0.011186830699443817, + 0.010487315244972706, + 0.006924946326762438, + 0.006024139001965523, + 0.00432589603587985, + 0.005193053744733334, + 0.0032216443214565516, + 0.004931668285280466, + 0.007559966761618853, + 0.005489567760378122, + 0.00532944779843092, + 0.0036341489758342505, + 0.003471379866823554, + 0.007041006349027157, + 0.002620923798531294, + 0.003594817593693733, + 0.007379896938800812, + 0.005166212562471628, + 0.007250342983752489, + 0.0035297845024615526, + 0.009122849442064762, + 0.017984412610530853, + 0.005752891767770052, + 0.014670614153146744, + 0.015827922150492668, + 0.010909968055784702, + 0.013356898911297321, + 0.007777026854455471, + 0.005205466412007809, + 0.004734243731945753, + 0.004502380266785622, + 0.004861314315348864, + 0.005019196309149265, + 0.005514815915375948, + 0.005102698691189289, + 0.005531391128897667, + 0.0033773097675293684, + 0.004450608976185322, + 0.002818872220814228, + 0.0045737214386463165, + 0.0036743180826306343, + 0.0033323573879897594, + 0.0031962827779352665, + 0.002954554744064808, + 0.011198538355529308, + 0.01726916991174221, + 0.016466759145259857, + 0.013834931887686253, + 0.01361851580440998, + 0.011896808631718159, + 0.017824510112404823, + 0.01568649709224701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 9 + ], + "selected_reward": -0.0046372017823159695 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006732518319040537 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.006732518319040537, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006732518319040537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24608999490737915, + 0.15122437477111816, + 0.0991506427526474, + 0.12302462756633759, + 0.049195338040590286, + 0.05490516871213913, + 0.05553203076124191, + 0.09858471155166626, + 0.053943876177072525, + 0.0683492049574852 + ], + "credit_map": [ + 0.031168699264526367, + 0.029543926939368248, + 0.029546018689870834, + 0.01789679378271103, + 0.03396356850862503, + 0.04558657109737396, + 0.025055861100554466, + 0.033328570425510406, + 0.011795997619628906, + 0.04105651378631592, + 0.010855286382138729, + 0.015536447055637836, + 0.020129969343543053, + 0.022030366584658623, + 0.016887923702597618, + 0.012931873090565205, + 0.010887132957577705, + 0.011849873699247837, + 0.018113425001502037, + 0.010680470615625381, + 0.012364957481622696, + 0.010578860528767109, + 0.012440857477486134, + 0.012235069647431374, + 0.01606762409210205, + 0.016186930239200592, + 0.016306335106492043, + 0.012812202796339989, + 0.014644427224993706, + 0.013802715577185154, + 0.014623296447098255, + 0.018581097945570946, + 0.006616190075874329, + 0.005556581076234579, + 0.0067868721671402454, + 0.005597818177193403, + 0.008277659304440022, + 0.005638516042381525, + 0.006374745164066553, + 0.0043469578959047794, + 0.0054753851145505905, + 0.007867054082453251, + 0.0051104119047522545, + 0.005464824382215738, + 0.01312969159334898, + 0.006681749597191811, + 0.006526370532810688, + 0.004649683833122253, + 0.0053438451141119, + 0.01097444910556078, + 0.004109447821974754, + 0.007058709394186735, + 0.009273067116737366, + 0.006555275525897741, + 0.007418121676892042, + 0.004799117799848318, + 0.010195363312959671, + 0.012830853462219238, + 0.011828761547803879, + 0.011454516090452671, + 0.017483267933130264, + 0.01099811028689146, + 0.010827744379639626, + 0.012966100126504898, + 0.0074127898551523685, + 0.006952222436666489, + 0.006240707356482744, + 0.008952154777944088, + 0.004207417834550142, + 0.006216987501829863, + 0.007398634683340788, + 0.006562965456396341, + 0.006464759353548288, + 0.009544743224978447, + 0.008530562743544579, + 0.007670792285352945, + 0.009652405977249146, + 0.006938292644917965, + 0.012111666612327099, + 0.007435990031808615 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.006732518319040537 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004088971298187971 + ], + "opposite_logprob_tokens": [ + -5.504088878631592 + ], + "logprob_reward": -0.004088971298187971, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004088971298187971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17133614420890808, + 0.21509380638599396, + 0.035532012581825256, + 0.2086504101753235, + 0.054004184901714325, + 0.05676150694489479, + 0.07267174869775772, + 0.07986150681972504, + 0.0675545409321785, + 0.03853423893451691 + ], + "credit_map": [ + 0.015525667928159237, + 0.02485971711575985, + 0.011725416406989098, + 0.016707893460989, + 0.04230808466672897, + 0.022106116637587547, + 0.021844711154699326, + 0.01625850424170494, + 0.01665828749537468, + 0.05924741178750992, + 0.011041688732802868, + 0.024633919820189476, + 0.03719928860664368, + 0.032410625368356705, + 0.01964350789785385, + 0.014259028248488903, + 0.003926753997802734, + 0.004430161323398352, + 0.004077174700796604, + 0.005585876759141684, + 0.0030961541924625635, + 0.004131941590458155, + 0.00492888456210494, + 0.005355058237910271, + 0.024643518030643463, + 0.029302671551704407, + 0.021816648542881012, + 0.03092854470014572, + 0.03205293416976929, + 0.025986528024077415, + 0.022871216759085655, + 0.021048322319984436, + 0.0066670626401901245, + 0.009753605350852013, + 0.005506436340510845, + 0.0039725881069898605, + 0.007013122085481882, + 0.01175444945693016, + 0.004086099099367857, + 0.005250807851552963, + 0.0038922473322600126, + 0.007353158202022314, + 0.003314356319606304, + 0.004683010280132294, + 0.015566397458314896, + 0.006940907333046198, + 0.009257380850613117, + 0.00575404055416584, + 0.00846270564943552, + 0.011732683517038822, + 0.004563606809824705, + 0.012355483137071133, + 0.00966303888708353, + 0.009039518423378468, + 0.011101929470896721, + 0.005752767901867628, + 0.008709433488547802, + 0.012729999609291553, + 0.0064545609056949615, + 0.008283525705337524, + 0.015042527578771114, + 0.010118731297552586, + 0.011502773500978947, + 0.0070199426263570786, + 0.008619029074907303, + 0.008809619583189487, + 0.005691763013601303, + 0.009687318466603756, + 0.00950744841247797, + 0.007685897406190634, + 0.009660369716584682, + 0.007893086411058903, + 0.004867467563599348, + 0.005107926670461893, + 0.005987357348203659, + 0.005398770794272423, + 0.004443483427166939, + 0.0041984631679952145, + 0.003834748873487115, + 0.004696012008935213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.004088971298187971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011068964377045631 + ], + "opposite_logprob_tokens": [ + -4.511068820953369 + ], + "logprob_reward": -0.011068964377045631, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011068964377045631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14850564301013947, + 0.21340757608413696, + 0.051195450127124786, + 0.05316270515322685, + 0.05663493648171425, + 0.1157098338007927, + 0.06923678517341614, + 0.16685646772384644, + 0.061066947877407074, + 0.06422362476587296 + ], + "credit_map": [ + 0.01768222264945507, + 0.020490320399403572, + 0.015367209911346436, + 0.012029689736664295, + 0.024385327473282814, + 0.03253184258937836, + 0.013033353723585606, + 0.012985673733055592, + 0.017138266935944557, + 0.04969162866473198, + 0.011534339748322964, + 0.028654396533966064, + 0.034020379185676575, + 0.02919348143041134, + 0.027335818856954575, + 0.015839271247386932, + 0.005134515929967165, + 0.007235165685415268, + 0.005516935605555773, + 0.005533290095627308, + 0.005965143907815218, + 0.006006899755448103, + 0.007459278218448162, + 0.008344219997525215, + 0.006019027438014746, + 0.008276364766061306, + 0.006131527479737997, + 0.00681719183921814, + 0.006911833770573139, + 0.0061007309705019, + 0.006480156444013119, + 0.00642587523907423, + 0.0072845956310629845, + 0.006801223382353783, + 0.005217459984123707, + 0.007799721322953701, + 0.008950851857662201, + 0.008293144404888153, + 0.007410218007862568, + 0.0048777200281620026, + 0.009962544776499271, + 0.015414346009492874, + 0.007060615345835686, + 0.009582677856087685, + 0.029661457985639572, + 0.01393177267163992, + 0.02055203728377819, + 0.009544377215206623, + 0.0074409376829862595, + 0.009602837264537811, + 0.005352722480893135, + 0.01187277864664793, + 0.010056114755570889, + 0.0071961344219744205, + 0.010886340402066708, + 0.006828917656093836, + 0.015003637410700321, + 0.025428729131817818, + 0.013937139883637428, + 0.027767013758420944, + 0.03368353471159935, + 0.018352041020989418, + 0.02099584974348545, + 0.011688528582453728, + 0.00816724169999361, + 0.008238528855144978, + 0.005695645231753588, + 0.012007689103484154, + 0.00605690898373723, + 0.007184662856161594, + 0.008291083388030529, + 0.005425189156085253, + 0.007116922177374363, + 0.006791068706661463, + 0.007965140976011753, + 0.00871003046631813, + 0.007859312929213047, + 0.006268810946494341, + 0.011237923987209797, + 0.008274409919977188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.011068964377045631 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, horizontal wide-angle shot captures a vibrant 8-bit pixel art concert. Center stage, an anthropomorphic USB drive composed of bright blue and gray pixels takes center stage, acting as conductor. With two simple white blocks for eyes, it stands atop a small, pixelated podium. Its body sways gently to an imaginary electronic beat. One pixelated arm holds a gleaming white baton, waving it up and down decisively and rhythmically. Another arm reaches to the side of the stage, cuing a pixelated band. This band is composed entirely of anthropomorphic 8-bit instruments: a red pixelated keyboard, its keys pulsing and glowing in time with the melody; a yellow pixelated drum set, its drumheads and cymbals emitting tiny white pixelated starbursts when struck by invisible sticks; and a green pixelated electric guitar, bobbing up and down to the rhythm of a passionate solo. The entire scene is bathed in a simulated pixelated spotlight. Deep purple beams cast sharp geometric shapes against a deep black background, dotted with slowly blinking pixels reminiscent of stars from early arcade games. The primary color palette of electric blue, magenta, and bright yellow creates a strong visual contrast, and all elements are rendered in crisp, non-anti-aliased pixel blocks, creating a nostalgic, dynamic, and playful atmosphere reminiscent of a retro gaming concert.", + "seed": 33511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_6504_seed_33511_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.007531025353819132, + "vlm_logprob": -0.007531025353819132, + "vlm_yes_no_margin": 4.95, + "vlm_token_credit": -0.007531025353819132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009277093224227428 + ], + "opposite_logprob_tokens": [ + -4.759276866912842 + ], + "logprob_reward": -0.009277093224227428, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009277093224227428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20419247448444366, + 0.09057363867759705, + 0.0560523085296154, + 0.030160585418343544, + 0.07955528795719147, + 0.06946436315774918, + 0.05151316896080971, + 0.06045423820614815, + 0.09793119132518768, + 0.2601027190685272 + ], + "credit_map": [ + 0.02883298322558403, + 0.02706053853034973, + 0.018817581236362457, + 0.023470785468816757, + 0.02906731888651848, + 0.026000039651989937, + 0.026000913232564926, + 0.02494230307638645, + 0.012613033875823021, + 0.011979921720921993, + 0.008058628998696804, + 0.006503189913928509, + 0.019188549369573593, + 0.010869713500142097, + 0.012071498669683933, + 0.009289104491472244, + 0.007765437476336956, + 0.007076961454004049, + 0.004325813613831997, + 0.005970469675958157, + 0.008464375510811806, + 0.006990015972405672, + 0.005218503996729851, + 0.01024072989821434, + 0.0043011438101530075, + 0.005127614829689264, + 0.0020400758367031813, + 0.002989653730764985, + 0.0050043086521327496, + 0.003928654361516237, + 0.004011304117739201, + 0.0027578312437981367, + 0.014915992505848408, + 0.00967923179268837, + 0.008937958627939224, + 0.009080767631530762, + 0.008270910941064358, + 0.012883796356618404, + 0.007565970532596111, + 0.008220654912292957, + 0.010478414595127106, + 0.007772074546664953, + 0.00922021921724081, + 0.006260359659790993, + 0.00794337596744299, + 0.010861446149647236, + 0.006187181919813156, + 0.010741294361650944, + 0.009205130860209465, + 0.005971607286483049, + 0.007033213973045349, + 0.0040330588817596436, + 0.006023795809596777, + 0.006717827171087265, + 0.005235407501459122, + 0.0072931284084916115, + 0.006372475065290928, + 0.0066560558043420315, + 0.003933427389711142, + 0.0061424202285707, + 0.01013433188199997, + 0.006846296600997448, + 0.01141703687608242, + 0.008952193893492222, + 0.010838082991540432, + 0.014131573960185051, + 0.007540274411439896, + 0.010976886376738548, + 0.01923011988401413, + 0.012497974559664726, + 0.013364422135055065, + 0.009351866319775581, + 0.03507286310195923, + 0.0333482027053833, + 0.02416669763624668, + 0.02556261047720909, + 0.039438918232917786, + 0.03500976786017418, + 0.04272375628352165, + 0.024779902771115303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.009277093224227428 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637676291167736 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637676291167736, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637676291167736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11653205007314682, + 0.2969878315925598, + 0.0989883542060852, + 0.04485606029629707, + 0.14139074087142944, + 0.09629110991954803, + 0.0590301975607872, + 0.04536225646734238, + 0.03843630477786064, + 0.06212504208087921 + ], + "credit_map": [ + 0.01578320376574993, + 0.016480181366205215, + 0.00899240467697382, + 0.012705575674772263, + 0.019404089078307152, + 0.015664517879486084, + 0.013355353847146034, + 0.014146735891699791, + 0.025804921984672546, + 0.052920375019311905, + 0.01090741902589798, + 0.02001214772462845, + 0.09789232164621353, + 0.040367674082517624, + 0.032854970544576645, + 0.016227995976805687, + 0.012135877273976803, + 0.012001702561974525, + 0.00793531909584999, + 0.011962011456489563, + 0.015990743413567543, + 0.012109814211726189, + 0.00989861786365509, + 0.016954269260168076, + 0.0068725827150046825, + 0.005874243099242449, + 0.004626515321433544, + 0.004960780031979084, + 0.007429861463606358, + 0.0062257759273052216, + 0.004296082071959972, + 0.004570221062749624, + 0.02094440720975399, + 0.0156665351241827, + 0.018335893750190735, + 0.01711585558950901, + 0.018536416813731194, + 0.01476328819990158, + 0.01583002507686615, + 0.020198317244648933, + 0.013988601043820381, + 0.011701955460011959, + 0.012941299937665462, + 0.009047691710293293, + 0.013177228160202503, + 0.011358283460140228, + 0.009925815276801586, + 0.014150230213999748, + 0.010701881721615791, + 0.0072462852112948895, + 0.00826474279165268, + 0.006144295912235975, + 0.006459573749452829, + 0.008449306711554527, + 0.005617738701403141, + 0.006146377418190241, + 0.007030772976577282, + 0.005062842275947332, + 0.004879542626440525, + 0.00395806273445487, + 0.004389848094433546, + 0.006065314169973135, + 0.005235324613749981, + 0.008740548975765705, + 0.005615570116788149, + 0.006577241700142622, + 0.0027756113559007645, + 0.004036950878798962, + 0.005586782470345497, + 0.005026084836572409, + 0.005104065872728825, + 0.003713997546583414, + 0.00908448826521635, + 0.00875044334679842, + 0.005145270377397537, + 0.00581867340952158, + 0.007906555198132992, + 0.008255641907453537, + 0.009928643703460693, + 0.007235329132527113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.004637676291167736 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006731215864419937 + ], + "opposite_logprob_tokens": [ + -5.006731033325195 + ], + "logprob_reward": -0.006731215864419937, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006731215864419937, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21124017238616943, + 0.2866453230381012, + 0.10086621344089508, + 0.0640021339058876, + 0.07307252287864685, + 0.07358568161725998, + 0.053427159786224365, + 0.04252830147743225, + 0.028974026441574097, + 0.06565845012664795 + ], + "credit_map": [ + 0.03255101665854454, + 0.02988366410136223, + 0.020342305302619934, + 0.022869355976581573, + 0.031965941190719604, + 0.0270216204226017, + 0.02363426797091961, + 0.0229719839990139, + 0.02314799837768078, + 0.053458623588085175, + 0.009779367595911026, + 0.01923873834311962, + 0.0963178351521492, + 0.03865963593125343, + 0.03141744062304497, + 0.014625667594373226, + 0.014176483266055584, + 0.014795581810176373, + 0.01357888150960207, + 0.014209370128810406, + 0.012607548385858536, + 0.012852119281888008, + 0.008506481535732746, + 0.010139752179384232, + 0.009031699039041996, + 0.009168858639895916, + 0.006493367720395327, + 0.0077704149298369884, + 0.0066289640963077545, + 0.00880292896181345, + 0.006203672382980585, + 0.009902226738631725, + 0.009827697649598122, + 0.009675124660134315, + 0.004321001935750246, + 0.007351497188210487, + 0.014976751059293747, + 0.009851031936705112, + 0.007190720643848181, + 0.009878695011138916, + 0.009958931244909763, + 0.010065093636512756, + 0.008612507954239845, + 0.006140592973679304, + 0.012351298704743385, + 0.009032688103616238, + 0.0071339113637804985, + 0.01029065903276205, + 0.008979622274637222, + 0.005919747054576874, + 0.006064207758754492, + 0.00615059956908226, + 0.005684021860361099, + 0.007207926828414202, + 0.006750956643372774, + 0.006670079194009304, + 0.005998194683343172, + 0.0051549519412219524, + 0.004407104104757309, + 0.004593400750309229, + 0.005417484790086746, + 0.005737639497965574, + 0.004520792979747057, + 0.006698729936033487, + 0.005205964669585228, + 0.003995383623987436, + 0.0023797666653990746, + 0.00258611049503088, + 0.005798215512186289, + 0.0038217108231037855, + 0.0030604503117501736, + 0.0021264252718538046, + 0.00897270254790783, + 0.007959260605275631, + 0.006877738516777754, + 0.00701497495174408, + 0.007447185926139355, + 0.008987580426037312, + 0.010557107627391815, + 0.00784190371632576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.006731215864419937 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942652467638254 + ], + "opposite_logprob_tokens": [ + -5.1309428215026855 + ], + "logprob_reward": -0.005942652467638254, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942652467638254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38418149948120117, + 0.04798726364970207, + 0.06592008471488953, + 0.04383258521556854, + 0.09647676348686218, + 0.1818811595439911, + 0.05850474536418915, + 0.05274171382188797, + 0.018436508253216743, + 0.050037648528814316 + ], + "credit_map": [ + 0.04685180261731148, + 0.03935940936207771, + 0.026739148423075676, + 0.03319860249757767, + 0.05745415389537811, + 0.054787527769804, + 0.04950960353016853, + 0.07628127187490463, + 0.005916391499340534, + 0.005857039708644152, + 0.0034137964248657227, + 0.0037948309909552336, + 0.010782771743834019, + 0.005801775027066469, + 0.005000090226531029, + 0.007420565467327833, + 0.007945381104946136, + 0.007650791201740503, + 0.005810871254652739, + 0.005906730890274048, + 0.011670326814055443, + 0.008488430641591549, + 0.007300340570509434, + 0.0111472113057971, + 0.004936643410474062, + 0.006580875720828772, + 0.003613366512581706, + 0.004532919265329838, + 0.007723244372755289, + 0.005390864796936512, + 0.0050048744305968285, + 0.006049799267202616, + 0.016487974673509598, + 0.013777299784123898, + 0.009691664017736912, + 0.012080498039722443, + 0.013700294308364391, + 0.012478766962885857, + 0.011695247143507004, + 0.006565021350979805, + 0.02571684680879116, + 0.023130735382437706, + 0.021749107167124748, + 0.017208507284522057, + 0.03147061541676521, + 0.020213576033711433, + 0.019177163019776344, + 0.02321459911763668, + 0.010078039951622486, + 0.0064735813066363335, + 0.009652803651988506, + 0.00663426797837019, + 0.006534097250550985, + 0.00794832780957222, + 0.005125706549733877, + 0.006057924125343561, + 0.006494642700999975, + 0.006319775711745024, + 0.005529888905584812, + 0.005154816433787346, + 0.008511456660926342, + 0.007317122537642717, + 0.005535413511097431, + 0.007878598757088184, + 0.0030506423208862543, + 0.0020343579817563295, + 0.002235892228782177, + 0.0015647293766960502, + 0.00307614728808403, + 0.0025100933853536844, + 0.001610195031389594, + 0.0023544500581920147, + 0.006592956837266684, + 0.005459619220346212, + 0.005717128049582243, + 0.005415111314505339, + 0.0062640090472996235, + 0.006596260704100132, + 0.0080111650750041, + 0.005981397815048695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.005942652467638254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011066488921642303 + ], + "opposite_logprob_tokens": [ + -4.511066436767578 + ], + "logprob_reward": -0.011066488921642303, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011066488921642303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13068781793117523, + 0.27940380573272705, + 0.04678688198328018, + 0.030528821051120758, + 0.14467516541481018, + 0.1228189542889595, + 0.07269566506147385, + 0.05234315246343613, + 0.06674099713563919, + 0.05331874266266823 + ], + "credit_map": [ + 0.019896214827895164, + 0.019141720607876778, + 0.011205584742128849, + 0.014302926138043404, + 0.01818978413939476, + 0.016606757417321205, + 0.018064701929688454, + 0.013280129991471767, + 0.025731967762112617, + 0.048217616975307465, + 0.014137663878500462, + 0.017961585894227028, + 0.08977289497852325, + 0.03598640486598015, + 0.03161550313234329, + 0.015980148687958717, + 0.0067522237077355385, + 0.006573650520294905, + 0.004691699054092169, + 0.0050574131309986115, + 0.007073663175106049, + 0.006370088085532188, + 0.005116929300129414, + 0.005151213612407446, + 0.0037354878149926662, + 0.004227850586175919, + 0.002739547984674573, + 0.0032567332964390516, + 0.005332828499376774, + 0.0038209909107536077, + 0.0030253352597355843, + 0.00439004460349679, + 0.023615457117557526, + 0.013976583257317543, + 0.017731906846165657, + 0.015597032383084297, + 0.01866491138935089, + 0.019834144040942192, + 0.014916171319782734, + 0.020338956266641617, + 0.01724327728152275, + 0.013365847989916801, + 0.015281536616384983, + 0.012474414892494678, + 0.015172306448221207, + 0.015002517029643059, + 0.012952112592756748, + 0.021326936781406403, + 0.013308431021869183, + 0.00769534008577466, + 0.011220797896385193, + 0.008277567103505135, + 0.007404669187963009, + 0.009680233895778656, + 0.007601373828947544, + 0.007507241331040859, + 0.006329371593892574, + 0.006228845566511154, + 0.006162590812891722, + 0.005336920730769634, + 0.008862648159265518, + 0.006741813383996487, + 0.0053443145006895065, + 0.007336647715419531, + 0.007864034734666348, + 0.008337653242051601, + 0.0065709794871509075, + 0.007011702284216881, + 0.012793956324458122, + 0.00894435029476881, + 0.008744115941226482, + 0.006474202964454889, + 0.007046878337860107, + 0.006806905847042799, + 0.006154102273285389, + 0.005606719758361578, + 0.0068529145792126656, + 0.006876739207655191, + 0.008355846628546715, + 0.005618635565042496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.011066488921642303 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "fixed, horizontal wide-angle shot captures a vibrant 8-bit pixel art concert. Center stage, an anthropomorphic USB drive composed of bright blue and gray pixels takes center stage, acting as conductor. With two simple white blocks for eyes, it stands atop a small, pixelated podium. Its body sways gently to an imaginary electronic beat. One pixelated arm holds a gleaming white baton, waving it up and down decisively and rhythmically. Another arm reaches to the side of the stage, cuing a pixelated band. This band is composed entirely of anthropomorphic 8-bit instruments: a red pixelated keyboard, its keys pulsing and glowing in time with the melody; a yellow pixelated drum set, its drumheads and cymbals emitting tiny white pixelated starbursts when struck by invisible sticks; and a green pixelated electric guitar, bobbing up and down to the rhythm of a passionate solo. The entire scene is bathed in a simulated pixelated spotlight. Deep purple beams cast sharp geometric shapes against a deep black background, dotted with slowly blinking pixels reminiscent of stars from early arcade games. The primary color palette of electric blue, magenta, and bright yellow creates a strong visual contrast, and all elements are rendered in crisp, non-anti-aliased pixel blocks, creating a nostalgic, dynamic, and playful atmosphere reminiscent of a retro gaming concert.", + "seed": 33512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000009/group_9/index_6504_seed_33512_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011303276661783457, + "vlm_logprob": -0.011303276661783457, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.011303276661783457, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03035808354616165 + ], + "opposite_logprob_tokens": [ + -3.530358076095581 + ], + "logprob_reward": -0.03035808354616165, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03035808354616165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04594717547297478, + 0.17628192901611328, + 0.08198730647563934, + 0.06521179527044296, + 0.3231965899467468, + 0.0870765820145607, + 0.047992561012506485, + 0.10505597293376923, + 0.03845743089914322, + 0.02879263646900654 + ], + "credit_map": [ + 0.0066686286590993404, + 0.004659825470298529, + 0.007447300013154745, + 0.004687288776040077, + 0.0030465200543403625, + 0.0097896046936512, + 0.0035204878076910973, + 0.006127523723989725, + 0.013532182201743126, + 0.017816711217164993, + 0.017039915546774864, + 0.0197296179831028, + 0.014697151258587837, + 0.05239449813961983, + 0.013250737451016903, + 0.02782113291323185, + 0.0077413045801222324, + 0.011585465632379055, + 0.011205757036805153, + 0.01207306981086731, + 0.007443586364388466, + 0.011614153161644936, + 0.006680828053504229, + 0.013643152080476284, + 0.006443272810429335, + 0.008444070816040039, + 0.006719257682561874, + 0.0056561012752354145, + 0.005577814299613237, + 0.00930807925760746, + 0.010981027968227863, + 0.012082173489034176, + 0.036947328597307205, + 0.03389029577374458, + 0.07149402052164078, + 0.02850125916302204, + 0.035730112344026566, + 0.03278370574116707, + 0.029660597443580627, + 0.054189298301935196, + 0.007920531556010246, + 0.01220838725566864, + 0.01486140862107277, + 0.007280864752829075, + 0.009375148452818394, + 0.009549735113978386, + 0.010214543901383877, + 0.015665968880057335, + 0.005815765354782343, + 0.005280448589473963, + 0.008178568445146084, + 0.006120413076132536, + 0.0040734331123530865, + 0.006588303484022617, + 0.004652820061892271, + 0.007282811217010021, + 0.01140841655433178, + 0.022339366376399994, + 0.01416940800845623, + 0.013880472630262375, + 0.012529058381915092, + 0.006936102174222469, + 0.010091920383274555, + 0.013701235875487328, + 0.004800201393663883, + 0.0064163170754909515, + 0.0038721715100109577, + 0.008221466094255447, + 0.0049377381801605225, + 0.002653641626238823, + 0.004884994123131037, + 0.00267090299166739, + 0.00264529325067997, + 0.003906450234353542, + 0.004114343319088221, + 0.004696189891546965, + 0.0041131312027573586, + 0.002195220673456788, + 0.004350751638412476, + 0.002771260216832161 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 7 + ], + "selected_reward": -0.03035808354616165 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627174723893404 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627174723893404, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627174723893404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07536584883928299, + 0.22093196213245392, + 0.121298648416996, + 0.05829751864075661, + 0.1852114200592041, + 0.11346528679132462, + 0.053105004131793976, + 0.07956641167402267, + 0.04328818619251251, + 0.049469660967588425 + ], + "credit_map": [ + 0.010568833909928799, + 0.007454318460077047, + 0.012858396396040916, + 0.007410162594169378, + 0.006371090654283762, + 0.013338753022253513, + 0.0077242180705070496, + 0.009640073403716087, + 0.018818357959389687, + 0.0266160536557436, + 0.02763022482395172, + 0.02935902774333954, + 0.01854567788541317, + 0.052449993789196014, + 0.014693559147417545, + 0.03281905874609947, + 0.010626737028360367, + 0.018602725118398666, + 0.01696656458079815, + 0.021841365844011307, + 0.011267158202826977, + 0.012542922049760818, + 0.011843958869576454, + 0.01760721206665039, + 0.005386296659708023, + 0.006910759024322033, + 0.008098926395177841, + 0.00590429687872529, + 0.005023669451475143, + 0.009482715278863907, + 0.008204957470297813, + 0.00928589515388012, + 0.02021424099802971, + 0.02037724293768406, + 0.036875177174806595, + 0.02226104773581028, + 0.021427670493721962, + 0.01505495049059391, + 0.018637631088495255, + 0.03036346286535263, + 0.01029215008020401, + 0.01632741466164589, + 0.014596379362046719, + 0.01547717209905386, + 0.014076486229896545, + 0.014956234954297543, + 0.010621684603393078, + 0.017117762938141823, + 0.00623727310448885, + 0.004486536607146263, + 0.0074540902860462666, + 0.0060376799665391445, + 0.004703710786998272, + 0.007904878817498684, + 0.00832631066441536, + 0.007954525761306286, + 0.008638138882815838, + 0.014488429762423038, + 0.010630618780851364, + 0.011557966470718384, + 0.010124142281711102, + 0.005462207831442356, + 0.008626051247119904, + 0.010038855485618114, + 0.007253043819218874, + 0.004691120702773333, + 0.005041317082941532, + 0.005208222661167383, + 0.005678214598447084, + 0.006316613405942917, + 0.004222341813147068, + 0.004877311643213034, + 0.0051383450627326965, + 0.006138663738965988, + 0.007552260998636484, + 0.006366149988025427, + 0.006191405933350325, + 0.004930236376821995, + 0.006908559240400791, + 0.006244039162993431 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.007627174723893404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028152139857411385 + ], + "opposite_logprob_tokens": [ + -5.877815246582031 + ], + "logprob_reward": -0.0028152139857411385, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028152139857411385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08756188303232193, + 0.1530185341835022, + 0.10714849084615707, + 0.07510410249233246, + 0.13679440319538116, + 0.09619204699993134, + 0.11397199332714081, + 0.13325068354606628, + 0.05018388107419014, + 0.046773914247751236 + ], + "credit_map": [ + 0.008151981048285961, + 0.011141891591250896, + 0.008840300142765045, + 0.008208761923015118, + 0.007122804876416922, + 0.024584000930190086, + 0.00723104365170002, + 0.012281106784939766, + 0.013602737337350845, + 0.020115969702601433, + 0.022182278335094452, + 0.020535478368401527, + 0.011555450037121773, + 0.03180268034338951, + 0.01092025637626648, + 0.022303692996501923, + 0.013982667587697506, + 0.01718166284263134, + 0.015948720276355743, + 0.011573411524295807, + 0.01114854495972395, + 0.013526263646781445, + 0.010993658564984798, + 0.012793570756912231, + 0.010381000116467476, + 0.01435553003102541, + 0.006033990532159805, + 0.008940392173826694, + 0.006726237945258617, + 0.00862135924398899, + 0.013719222508370876, + 0.006326373666524887, + 0.017123309895396233, + 0.013354276306927204, + 0.025008046999573708, + 0.010473829694092274, + 0.014321183785796165, + 0.019927306100726128, + 0.012236899696290493, + 0.024349557235836983, + 0.011018353514373302, + 0.01225739624351263, + 0.01658990979194641, + 0.012562851421535015, + 0.008638150990009308, + 0.010691137053072453, + 0.009393352083861828, + 0.015040897764265537, + 0.010705032385885715, + 0.01101098395884037, + 0.015107354149222374, + 0.012890267185866833, + 0.010901556350290775, + 0.011459547095000744, + 0.023392120376229286, + 0.01850513368844986, + 0.016663001850247383, + 0.014999127015471458, + 0.0180276557803154, + 0.0147856455296278, + 0.015196049585938454, + 0.009944300167262554, + 0.018604610115289688, + 0.025030313059687614, + 0.0068330299109220505, + 0.004396955017000437, + 0.006348578725010157, + 0.007894809357821941, + 0.007280405610799789, + 0.004784310702234507, + 0.008050046861171722, + 0.0045957500115036964, + 0.004617298021912575, + 0.0073378002271056175, + 0.0060681249015033245, + 0.007186566479504108, + 0.004630923271179199, + 0.005368216894567013, + 0.006168518215417862, + 0.005396468099206686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.0028152139857411385 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005941704846918583 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.005941704846918583, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005941704846918583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20988206565380096, + 0.16415734589099884, + 0.09828804433345795, + 0.058806031942367554, + 0.1328769326210022, + 0.0750105157494545, + 0.12752489745616913, + 0.04945390298962593, + 0.03439199551939964, + 0.04960828647017479 + ], + "credit_map": [ + 0.020110759884119034, + 0.028790878131985664, + 0.0173452477902174, + 0.02589787170290947, + 0.019411107525229454, + 0.05260837450623512, + 0.019187219440937042, + 0.026530591771006584, + 0.014279353432357311, + 0.025885485112667084, + 0.0223756842315197, + 0.022247420623898506, + 0.01202763244509697, + 0.03123082034289837, + 0.010543249547481537, + 0.025567710399627686, + 0.008949825540184975, + 0.00916728563606739, + 0.01652466505765915, + 0.021799342706799507, + 0.00917229987680912, + 0.010611731559038162, + 0.008184335194528103, + 0.013878554105758667, + 0.0074852826073765755, + 0.00766747584566474, + 0.008352570235729218, + 0.0069678667932748795, + 0.007074503228068352, + 0.008009197190403938, + 0.006961291190236807, + 0.00628784392029047, + 0.01838086172938347, + 0.011305619962513447, + 0.027165880426764488, + 0.010734130628407001, + 0.01578638143837452, + 0.013513568788766861, + 0.012988478876650333, + 0.023002013564109802, + 0.010937492363154888, + 0.012476886622607708, + 0.009872877970337868, + 0.006724860519170761, + 0.00766869867220521, + 0.009153238497674465, + 0.007431201636791229, + 0.010745259933173656, + 0.016809241846203804, + 0.027175182476639748, + 0.016767332330346107, + 0.01967478170990944, + 0.012288698926568031, + 0.01001118216663599, + 0.011984109878540039, + 0.012814364396035671, + 0.007152946200221777, + 0.00714111328125, + 0.006209219340234995, + 0.005146730691194534, + 0.004914320074021816, + 0.004407596308737993, + 0.005583610385656357, + 0.008898364380002022, + 0.00426233746111393, + 0.004430119879543781, + 0.0037023744080215693, + 0.003321309806779027, + 0.004226319026201963, + 0.00680988235399127, + 0.0039313482120633125, + 0.003708304837346077, + 0.003993363585323095, + 0.006274805404245853, + 0.005549405235797167, + 0.01012664195150137, + 0.005848248954862356, + 0.004346011206507683, + 0.008024639450013638, + 0.005445173010230064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.005941704846918583 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an anthropomorphic USB drive composed of bright blue and gray pixels acting as a conductor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the USB drive standing on a small pixelated podium while swaying gently and waving a white baton?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the band composed of a red pixelated keyboard, a yellow pixelated drum set emitting starbursts, and a green pixelated electric guitar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene rendered in a fixed, horizontal wide-angle shot with a deep black background and deep purple geometric light beams?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are all visual elements rendered in crisp, non-anti-aliased pixel blocks with a primary palette of electric blue, magenta, and bright yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774206206202507 + ], + "opposite_logprob_tokens": [ + -4.634774208068848 + ], + "logprob_reward": -0.009774206206202507, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774206206202507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14825654029846191, + 0.17569094896316528, + 0.1299230456352234, + 0.09571313112974167, + 0.11413156241178513, + 0.07644318789243698, + 0.05924888327717781, + 0.07243747264146805, + 0.05111170560121536, + 0.0770435705780983 + ], + "credit_map": [ + 0.016622135415673256, + 0.017070697620511055, + 0.02002076804637909, + 0.02444550395011902, + 0.014960472472012043, + 0.020387638360261917, + 0.01378138829022646, + 0.020967934280633926, + 0.017747778445482254, + 0.022310808300971985, + 0.028360631316900253, + 0.025166407227516174, + 0.014510051347315311, + 0.03300347551703453, + 0.011538860388100147, + 0.023052936419844627, + 0.014986332505941391, + 0.02216327376663685, + 0.01312017347663641, + 0.019798891618847847, + 0.011641092598438263, + 0.01586209610104561, + 0.01193018164485693, + 0.0204209815710783, + 0.012951552867889404, + 0.017189985141158104, + 0.015499899163842201, + 0.008726734668016434, + 0.008136995136737823, + 0.010389745235443115, + 0.01320726703852415, + 0.009610944427549839, + 0.01589198224246502, + 0.009904731065034866, + 0.02519945241510868, + 0.009226751513779163, + 0.010791691020131111, + 0.015201817266643047, + 0.01639832742512226, + 0.011516807600855827, + 0.010848071426153183, + 0.011476992629468441, + 0.011573274619877338, + 0.0076569910161197186, + 0.008955929428339005, + 0.009598908014595509, + 0.007214079145342112, + 0.009118933230638504, + 0.007262130733579397, + 0.009538998827338219, + 0.007677008863538504, + 0.008420050144195557, + 0.006337263155728579, + 0.005998064763844013, + 0.005794398952275515, + 0.008220961317420006, + 0.009499505162239075, + 0.009288234636187553, + 0.010635455138981342, + 0.008483430370688438, + 0.0080943051725626, + 0.005827243905514479, + 0.010865566320717335, + 0.009743721224367619, + 0.006045385263860226, + 0.008609003387391567, + 0.005667690187692642, + 0.009913245216012001, + 0.006552477832883596, + 0.004205265082418919, + 0.006061866879463196, + 0.004056765232235193, + 0.006414969451725483, + 0.009878180921077728, + 0.009031059220433235, + 0.015661628916859627, + 0.007441706955432892, + 0.007054045796394348, + 0.014098207466304302, + 0.007463761139661074 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.009774206206202507 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_tensor([5045, 5045])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_tensor([5045, 5045])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2591405dc4c41f4d2c561d226f4245fba170b595 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_tensor([5045, 5045])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme macro shot begins a slow, deliberate advance, focusing with hyper-sharp clarity on a single candle flame desperately clinging to a charred, black wick. The flame, a vibrant drop of light, dances and sways against a backdrop of absolute velvety darkness. Its core burns with a brilliant, almost incandescent intensity, transitioning into a cascade of fiery orange and warm yellow that, at its base, meets the glowing red burr of the wick. Below, a shimmering pool of translucent melted wax reflects the flame's every flicker, its surface trembling with heat. The camera's steady, relentless advance magnifies the flame's struggle, its wavering form casting a soft, warm, and profoundly lonely glow that barely holds back the oppressive, infinite darkness. This slow, mesmerizing movement creates an atmosphere of poignancy and tension, underscoring the raw, fragile beauty of light fighting for survival against the encroaching void, until the flame's dynamic textures and radiant energy fill the frame.", + "seed": 7068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7068_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031043379532638936, + "vlm_logprob": -0.031043379532638936, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.031043379532638936, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013576318509876728 + ], + "opposite_logprob_tokens": [ + -4.888576507568359 + ], + "logprob_reward": -0.013576318509876728, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.013576318509876728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21609032154083252, + 0.36346402764320374, + 0.09900778532028198, + 0.05364798754453659, + 0.07158146798610687, + 0.023870425298810005, + 0.042650312185287476, + 0.08165508508682251, + 0.03424637019634247, + 0.01378621906042099 + ], + "credit_map": [ + 0.07861962169408798, + 0.007870426401495934, + 0.03650304302573204, + 0.014333425089716911, + 0.007139840628951788, + 0.013289710506796837, + 0.01189943216741085, + 0.04643482342362404, + 0.10305655002593994, + 0.01635599695146084, + 0.0644284039735794, + 0.023217447102069855, + 0.017201872542500496, + 0.041769836097955704, + 0.03598233312368393, + 0.06145157665014267, + 0.03131649270653725, + 0.003987077623605728, + 0.008081161417067051, + 0.007178241852670908, + 0.015555710531771183, + 0.006843337789177895, + 0.0044702449813485146, + 0.02157551422715187, + 0.005137403029948473, + 0.004106678068637848, + 0.006252482999116182, + 0.0041986447758972645, + 0.006911605130881071, + 0.007512480951845646, + 0.011519787833094597, + 0.008008903823792934, + 0.006999866105616093, + 0.004963267594575882, + 0.005967346485704184, + 0.006514846812933683, + 0.010362495668232441, + 0.017884859815239906, + 0.009163182228803635, + 0.009725604206323624, + 0.002357013523578644, + 0.0023761263582855463, + 0.002054518321529031, + 0.0032410321291536093, + 0.0037469612434506416, + 0.004442781675606966, + 0.002063601976260543, + 0.003588389605283737, + 0.009541560895740986, + 0.0023218393325805664, + 0.0026637373957782984, + 0.0028813269454985857, + 0.01451892964541912, + 0.003983275033533573, + 0.002503227675333619, + 0.004236414562910795, + 0.030354388058185577, + 0.003016974776983261, + 0.0022396703716367483, + 0.0016038293251767755, + 0.012341785244643688, + 0.005383813288062811, + 0.0045950827188789845, + 0.022119540721178055, + 0.0038636066019535065, + 0.002293285680934787, + 0.002218798268586397, + 0.003426424227654934, + 0.007366021163761616, + 0.004055581521242857, + 0.004590009804815054, + 0.0064326440915465355, + 0.0017234861152246594, + 0.0009537689620628953, + 0.0013462839415296912, + 0.0023329288233071566, + 0.001341355498880148, + 0.0014073685742914677, + 0.0025919044855982065, + 0.002089122775942087 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.013576318509876728 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899880409240723 + ], + "opposite_logprob_tokens": [ + -2.5789988040924072 + ], + "logprob_reward": -0.07899880409240723, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899880409240723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04481581971049309, + 0.020379941910505295, + 0.18639063835144043, + 0.0585334487259388, + 0.045077793300151825, + 0.012654997408390045, + 0.015202861279249191, + 0.5902093052864075, + 0.013720469549298286, + 0.013014773838222027 + ], + "credit_map": [ + 0.012313193641602993, + 0.002325312700122595, + 0.00662996107712388, + 0.007664962206035852, + 0.0021052237134426832, + 0.0026368750259280205, + 0.004268471151590347, + 0.006871819030493498, + 0.005633041728287935, + 0.0011095530353486538, + 0.0033109409268945456, + 0.002271634293720126, + 0.0007809389499016106, + 0.0021958411671221256, + 0.0014805934624746442, + 0.0035973992198705673, + 0.07063060253858566, + 0.004880879539996386, + 0.012910651043057442, + 0.005780617240816355, + 0.03338661789894104, + 0.01138183195143938, + 0.006284292321652174, + 0.041135143488645554, + 0.008271866478025913, + 0.002903703600168228, + 0.0044854446314275265, + 0.004393889103084803, + 0.01004233118146658, + 0.010738774202764034, + 0.008525452576577663, + 0.009171993471682072, + 0.004818993154913187, + 0.004477606620639563, + 0.004022927954792976, + 0.005057911388576031, + 0.006459847092628479, + 0.008983534760773182, + 0.0035631642676889896, + 0.007693808525800705, + 0.0013777670683339238, + 0.0011883750557899475, + 0.0017792261205613613, + 0.001497961929999292, + 0.0017499775858595967, + 0.002205499680712819, + 0.0013313509989529848, + 0.0015248393174260855, + 0.0015734413173049688, + 0.0012762598926201463, + 0.0014079801039770246, + 0.0015402331482619047, + 0.002454344416037202, + 0.0020888042636215687, + 0.0013589112786576152, + 0.0035028879065066576, + 0.21880708634853363, + 0.02152116969227791, + 0.01949133910238743, + 0.012491567991673946, + 0.09185030311346054, + 0.03619171306490898, + 0.031457576900720596, + 0.15839853882789612, + 0.0013193882768973708, + 0.0011438692454248667, + 0.0012190137058496475, + 0.0021440826822072268, + 0.002099372213706374, + 0.001205087872222066, + 0.0014233289984986186, + 0.0031663260888308287, + 0.0012508053332567215, + 0.0012803951976820827, + 0.0012137523153796792, + 0.0026171559002250433, + 0.0010422575287520885, + 0.001356812659651041, + 0.00242983759380877, + 0.0018237581243738532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.07899880409240723 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054980359971523285 + ], + "opposite_logprob_tokens": [ + -2.9299802780151367 + ], + "logprob_reward": -0.054980359971523285, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054980359971523285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2548782527446747, + 0.04617878049612045, + 0.15757381916046143, + 0.2262091487646103, + 0.12114041298627853, + 0.02983012981712818, + 0.010090773925185204, + 0.11845730990171432, + 0.018756236881017685, + 0.016885101795196533 + ], + "credit_map": [ + 0.0996914654970169, + 0.009088418446481228, + 0.056292157620191574, + 0.00932824146002531, + 0.007720164023339748, + 0.011976887471973896, + 0.016879985108971596, + 0.043900955468416214, + 0.0072081726975739, + 0.002250277204439044, + 0.005831771995872259, + 0.004642250947654247, + 0.0024073857348412275, + 0.012767776846885681, + 0.005704981740564108, + 0.005366169381886721, + 0.04545852541923523, + 0.005567071493715048, + 0.01886536367237568, + 0.01144210621714592, + 0.02453930489718914, + 0.013562262989580631, + 0.009773004800081253, + 0.02836618758738041, + 0.023264411836862564, + 0.011437514796853065, + 0.019084451720118523, + 0.020022353157401085, + 0.037094224244356155, + 0.032016973942518234, + 0.04580450803041458, + 0.03748471662402153, + 0.010702786967158318, + 0.007892648689448833, + 0.007331965956836939, + 0.010707875713706017, + 0.018944574519991875, + 0.02266775630414486, + 0.02068503573536873, + 0.022207777947187424, + 0.0029457106720656157, + 0.0016766292974352837, + 0.0028337740805000067, + 0.002953980350866914, + 0.005659813527017832, + 0.005544177256524563, + 0.00358188571408391, + 0.004634163808077574, + 0.0018177243182435632, + 0.0010472236899659038, + 0.0011133024236187339, + 0.0010477248579263687, + 0.0019075148738920689, + 0.0012266715057194233, + 0.0008290655096061528, + 0.0011015472700819373, + 0.03511640802025795, + 0.0063033513724803925, + 0.008999014273285866, + 0.006266376469284296, + 0.015571625903248787, + 0.011453079991042614, + 0.008541533723473549, + 0.02620592713356018, + 0.002522009424865246, + 0.0015817383537068963, + 0.0015098147559911013, + 0.001653243787586689, + 0.002188768470659852, + 0.0026671946980059147, + 0.0030993972904980183, + 0.0035340713802725077, + 0.0015264259418472648, + 0.001187145127914846, + 0.001208204892463982, + 0.002899640239775181, + 0.00151598802767694, + 0.0024347431026399136, + 0.0036524676252156496, + 0.0024604895152151585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.054980359971523285 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737255025655031 + ], + "opposite_logprob_tokens": [ + -5.006737232208252 + ], + "logprob_reward": -0.006737255025655031, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737255025655031, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1591637283563614, + 0.12219620496034622, + 0.2685970664024353, + 0.06458796560764313, + 0.07690733671188354, + 0.07580353319644928, + 0.05493354797363281, + 0.02839180827140808, + 0.08793504536151886, + 0.06148369610309601 + ], + "credit_map": [ + 0.04116261377930641, + 0.01050942949950695, + 0.024807600304484367, + 0.021066058427095413, + 0.007967069745063782, + 0.014107530005276203, + 0.019757840782403946, + 0.019785607233643532, + 0.02325134538114071, + 0.007853196002542973, + 0.019405696541070938, + 0.011550547555088997, + 0.008084126748144627, + 0.020422257483005524, + 0.014574339613318443, + 0.01705469749867916, + 0.10397665202617645, + 0.007767184637486935, + 0.017146781086921692, + 0.006290394347161055, + 0.05012143775820732, + 0.018249185755848885, + 0.00754856551066041, + 0.05749688297510147, + 0.00733449961990118, + 0.00345493177883327, + 0.005025460850447416, + 0.006272715516388416, + 0.008750980719923973, + 0.01078975573182106, + 0.012562482617795467, + 0.010397150181233883, + 0.008154795505106449, + 0.0049050007946789265, + 0.0037169591523706913, + 0.00561614939942956, + 0.012767462059855461, + 0.017391208559274673, + 0.008455850183963776, + 0.015899917110800743, + 0.006084301508963108, + 0.004046478308737278, + 0.0053470502607524395, + 0.01084501575678587, + 0.010425952263176441, + 0.01671910472214222, + 0.006778263486921787, + 0.01555737666785717, + 0.008202745579183102, + 0.004349820781499147, + 0.0033843833953142166, + 0.0031080073677003384, + 0.01730753853917122, + 0.004470322746783495, + 0.003228557063266635, + 0.010882184840738773, + 0.003847141982987523, + 0.0024420530535280704, + 0.0038590787444263697, + 0.0028549334965646267, + 0.0030670783016830683, + 0.00474415672942996, + 0.003627636469900608, + 0.0039497301913797855, + 0.009534705430269241, + 0.007851655595004559, + 0.008871174417436123, + 0.011347073130309582, + 0.010426799766719341, + 0.011328428983688354, + 0.01600184477865696, + 0.012573372572660446, + 0.0051427423022687435, + 0.0035945316776633263, + 0.005367581266909838, + 0.009196859784424305, + 0.006113257724791765, + 0.009760190732777119, + 0.013726828619837761, + 0.008581710048019886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.006737255025655031 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009241600637324154 + ], + "opposite_logprob_tokens": [ + -7.000924110412598 + ], + "logprob_reward": -0.0009241600637324154, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009241600637324154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5119133591651917, + 0.12788529694080353, + 0.0724119246006012, + 0.08624511957168579, + 0.033473894000053406, + 0.02421891875565052, + 0.042547110468149185, + 0.06153488531708717, + 0.023763440549373627, + 0.016006136313080788 + ], + "credit_map": [ + 0.2069072276353836, + 0.01645282655954361, + 0.10316717624664307, + 0.0138081731274724, + 0.013616750948131084, + 0.025023646652698517, + 0.024328619241714478, + 0.1086089164018631, + 0.03406368941068649, + 0.006030292250216007, + 0.022530460730195045, + 0.00938459299504757, + 0.006536154542118311, + 0.01561151072382927, + 0.012971177697181702, + 0.020757416263222694, + 0.018892420455813408, + 0.002725516678765416, + 0.012369843199849129, + 0.004675709642469883, + 0.011344111524522305, + 0.0067783016711473465, + 0.0048048002645373344, + 0.010821221396327019, + 0.01038973219692707, + 0.005277985241264105, + 0.008121898397803307, + 0.007322965655475855, + 0.015334578230977058, + 0.015364331193268299, + 0.011867313645780087, + 0.012566318735480309, + 0.003240160411223769, + 0.0021067066118121147, + 0.0026217957492917776, + 0.0028190973680466413, + 0.004527604673057795, + 0.0066646067425608635, + 0.004557072650641203, + 0.0069368500262498856, + 0.0020284547936171293, + 0.0014595102984458208, + 0.0019784781616181135, + 0.0031710464973002672, + 0.003579929005354643, + 0.005106842610985041, + 0.0025188890285789967, + 0.004375767894089222, + 0.006861713249236345, + 0.002945313462987542, + 0.002927707973867655, + 0.0035219434648752213, + 0.012585129588842392, + 0.005175155587494373, + 0.003014261368662119, + 0.005515883211046457, + 0.018594389781355858, + 0.004243453498929739, + 0.005396856460720301, + 0.003729866351932287, + 0.00790347158908844, + 0.005005969200283289, + 0.003499064827337861, + 0.013161812908947468, + 0.00234692869707942, + 0.0027122125029563904, + 0.0031132979784160852, + 0.002867867238819599, + 0.00267333653755486, + 0.0031612038146704435, + 0.003400007728487253, + 0.0034885869827121496, + 0.0017970639746636152, + 0.0011525895679369569, + 0.001471798517741263, + 0.0029431143775582314, + 0.0014636068372055888, + 0.0021642721258103848, + 0.0030433202628046274, + 0.0019703726284205914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0009241600637324154 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, deliberate advance, focusing with hyper-sharp clarity on a single candle flame desperately clinging to a charred, black wick. The flame, a vibrant drop of light, dances and sways against a backdrop of absolute velvety darkness. Its core burns with a brilliant, almost incandescent intensity, transitioning into a cascade of fiery orange and warm yellow that, at its base, meets the glowing red burr of the wick. Below, a shimmering pool of translucent melted wax reflects the flame's every flicker, its surface trembling with heat. The camera's steady, relentless advance magnifies the flame's struggle, its wavering form casting a soft, warm, and profoundly lonely glow that barely holds back the oppressive, infinite darkness. This slow, mesmerizing movement creates an atmosphere of poignancy and tension, underscoring the raw, fragile beauty of light fighting for survival against the encroaching void, until the flame's dynamic textures and radiant energy fill the frame.", + "seed": 7069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7069_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04075280351098627, + "vlm_logprob": -0.04075280351098627, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.04075280351098627, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016336852684617043 + ], + "opposite_logprob_tokens": [ + -4.516336917877197 + ], + "logprob_reward": -0.016336852684617043, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.016336852684617043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2776215970516205, + 0.06641293317079544, + 0.1145833432674408, + 0.08135537803173065, + 0.05415620654821396, + 0.09823045879602432, + 0.05787041038274765, + 0.05841920152306557, + 0.1733223795890808, + 0.018028052523732185 + ], + "credit_map": [ + 0.017113933339715004, + 0.012477796524763107, + 0.02877720817923546, + 0.022354794666171074, + 0.020870981737971306, + 0.0751754641532898, + 0.04581204801797867, + 0.05503939092159271, + 0.005771540571004152, + 0.0044018360786139965, + 0.008810318075120449, + 0.007219418883323669, + 0.006835657171905041, + 0.009958680719137192, + 0.013077100738883018, + 0.010338391177356243, + 0.009750463999807835, + 0.00550388777628541, + 0.011643212288618088, + 0.00845438614487648, + 0.012702921405434608, + 0.018230566754937172, + 0.019747380167245865, + 0.028550537303090096, + 0.006927006877958775, + 0.00691668875515461, + 0.00919474195688963, + 0.008583947084844112, + 0.012510581873357296, + 0.014737337827682495, + 0.011848623864352703, + 0.010636452585458755, + 0.0051789237186312675, + 0.004494139924645424, + 0.005909941624850035, + 0.007729971781373024, + 0.004662195686250925, + 0.010567297227680683, + 0.007072561886161566, + 0.008541177026927471, + 0.008711032569408417, + 0.003202421125024557, + 0.011187620460987091, + 0.010848418809473515, + 0.007532753050327301, + 0.015389700420200825, + 0.01558698806911707, + 0.025771530345082283, + 0.006673004478216171, + 0.003207672154530883, + 0.0066481647081673145, + 0.005484999623149633, + 0.004015432205051184, + 0.010174836032092571, + 0.005222464445978403, + 0.01644383743405342, + 0.00642641494050622, + 0.006242113187909126, + 0.005097155924886465, + 0.005527546629309654, + 0.01369984820485115, + 0.009481417015194893, + 0.005226777400821447, + 0.006717934273183346, + 0.02165067009627819, + 0.006969156675040722, + 0.021459447219967842, + 0.017412718385457993, + 0.017126666381955147, + 0.025346148759126663, + 0.020642807707190514, + 0.042714767158031464, + 0.002550941426306963, + 0.0010848779929801822, + 0.002442376222461462, + 0.00225959625095129, + 0.0022180003579705954, + 0.00282964576035738, + 0.0018472297815605998, + 0.0027953863609582186 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.016336852684617043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270817518234253 + ], + "opposite_logprob_tokens": [ + -2.127081871032715 + ], + "logprob_reward": -0.1270817518234253, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270817518234253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1421521157026291, + 0.07053706794977188, + 0.08220408111810684, + 0.02638380602002144, + 0.06564339995384216, + 0.12820619344711304, + 0.13376757502555847, + 0.04714234173297882, + 0.2662532925605774, + 0.03771017864346504 + ], + "credit_map": [ + 0.007606326602399349, + 0.0043677338398993015, + 0.016835220158100128, + 0.008237906731665134, + 0.006378931924700737, + 0.053286850452423096, + 0.012685471214354038, + 0.03275367617607117, + 0.00570972403511405, + 0.005727700423449278, + 0.00875635165721178, + 0.009304951876401901, + 0.007301416248083115, + 0.013295138254761696, + 0.010534350760281086, + 0.009907438419759274, + 0.007187690585851669, + 0.006339170970022678, + 0.010856921784579754, + 0.01225458737462759, + 0.00980307999998331, + 0.012292357161641121, + 0.01319204829633236, + 0.010278222151100636, + 0.0028753133956342936, + 0.0015948353102430701, + 0.0024685480166226625, + 0.00267101195640862, + 0.004495964851230383, + 0.004061669576913118, + 0.004445546306669712, + 0.0037709162570536137, + 0.006570710800588131, + 0.003947157878428698, + 0.006997657474130392, + 0.006320719141513109, + 0.007396596949547529, + 0.01124069094657898, + 0.006496092304587364, + 0.016673769801855087, + 0.010499770753085613, + 0.007321990095078945, + 0.013052704744040966, + 0.019035156816244125, + 0.013914225623011589, + 0.019706744700670242, + 0.023968858644366264, + 0.020706741139292717, + 0.012690747156739235, + 0.004339901264756918, + 0.016428891569375992, + 0.016295326873660088, + 0.006602894049137831, + 0.02331106923520565, + 0.011824208311736584, + 0.04227454215288162, + 0.006005594506859779, + 0.004225637298077345, + 0.0053335498087108135, + 0.007412956561893225, + 0.0054990206845104694, + 0.006697137374430895, + 0.004154138267040253, + 0.00781430583447218, + 0.026381921023130417, + 0.01410746481269598, + 0.03127063438296318, + 0.02379368618130684, + 0.025948967784643173, + 0.04281065985560417, + 0.03398071974515915, + 0.06795921921730042, + 0.005477528087794781, + 0.0020561176352202892, + 0.005774732679128647, + 0.00592830590903759, + 0.002546793781220913, + 0.005241053178906441, + 0.004531841725111008, + 0.006153808906674385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.1270817518234253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043087493628263474 + ], + "opposite_logprob_tokens": [ + -3.1680874824523926 + ], + "logprob_reward": -0.043087493628263474, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043087493628263474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0633411779999733, + 0.03765998035669327, + 0.027143515646457672, + 0.06277000904083252, + 0.0948069840669632, + 0.15858669579029083, + 0.0222887322306633, + 0.03178718313574791, + 0.4520840048789978, + 0.049531739205121994 + ], + "credit_map": [ + 0.004102007485926151, + 0.003293305868282914, + 0.00943843461573124, + 0.008002659305930138, + 0.005044532008469105, + 0.011726410128176212, + 0.011708113364875317, + 0.010025708936154842, + 0.0029615170788019896, + 0.002585330279543996, + 0.005512883421033621, + 0.004207619931548834, + 0.003952939063310623, + 0.005518716759979725, + 0.0076434719376266, + 0.005277502350509167, + 0.0025373429525643587, + 0.002638423815369606, + 0.002620663959532976, + 0.0035269511863589287, + 0.0028512629214674234, + 0.00423183711245656, + 0.004093604627996683, + 0.004643427673727274, + 0.006970715243369341, + 0.0033365515992045403, + 0.007018595468252897, + 0.009212164208292961, + 0.006138226017355919, + 0.010323556140065193, + 0.010011888109147549, + 0.009758311323821545, + 0.009297718293964863, + 0.005693699698895216, + 0.008839116431772709, + 0.008863040246069431, + 0.008290125988423824, + 0.016219496726989746, + 0.013011916540563107, + 0.024591868743300438, + 0.017091115936636925, + 0.009702091105282307, + 0.021030493080615997, + 0.02516605705022812, + 0.015152468346059322, + 0.020770763978362083, + 0.02572084404528141, + 0.02395285665988922, + 0.0028334760572761297, + 0.0012959347805008292, + 0.003111473983153701, + 0.00333184702321887, + 0.0016028285026550293, + 0.0036262741778045893, + 0.0020028743892908096, + 0.004484023433178663, + 0.003809518413618207, + 0.002735320944339037, + 0.003949165344238281, + 0.005756860598921776, + 0.0031862531322985888, + 0.004375613294541836, + 0.0032911126036196947, + 0.004683338571339846, + 0.04981634393334389, + 0.017131473869085312, + 0.05674320086836815, + 0.04254413768649101, + 0.036736857146024704, + 0.07329963892698288, + 0.055430978536605835, + 0.1203814148902893, + 0.0074710482731461525, + 0.0024899118579924107, + 0.005303175654262304, + 0.007048584520816803, + 0.0043721385300159454, + 0.007538067176938057, + 0.00898145791143179, + 0.006327353417873383 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.043087493628263474 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071418300271034 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071418300271034, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071418300271034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10477229952812195, + 0.0842650979757309, + 0.06094631925225258, + 0.1352732926607132, + 0.08189481496810913, + 0.14582912623882294, + 0.06352943927049637, + 0.10039643198251724, + 0.19994327425956726, + 0.023149827495217323 + ], + "credit_map": [ + 0.007838760502636433, + 0.004787652753293514, + 0.010874270461499691, + 0.008536618202924728, + 0.007838970050215721, + 0.03096352331340313, + 0.01260448805987835, + 0.021328022703528404, + 0.007775063626468182, + 0.004246942698955536, + 0.009985092096030712, + 0.009340287186205387, + 0.008818238042294979, + 0.011526450514793396, + 0.017125649377703667, + 0.015447372570633888, + 0.006210969761013985, + 0.003395648906007409, + 0.008861330337822437, + 0.008241952396929264, + 0.0058173518627882, + 0.008648996241390705, + 0.009857621043920517, + 0.009912450797855854, + 0.014297306537628174, + 0.008418699726462364, + 0.01656218245625496, + 0.017931533977389336, + 0.012983138673007488, + 0.022756164893507957, + 0.018706105649471283, + 0.023618172854185104, + 0.009217820130288601, + 0.008833060041069984, + 0.008434797637164593, + 0.009241553954780102, + 0.007694312836974859, + 0.014181986451148987, + 0.009492072276771069, + 0.014799214899539948, + 0.014837990514934063, + 0.003991043660789728, + 0.01759246736764908, + 0.013205604627728462, + 0.007242459803819656, + 0.024402786046266556, + 0.016989709809422493, + 0.04756706580519676, + 0.0070778243243694305, + 0.0036915764212608337, + 0.0071903374046087265, + 0.007188914809376001, + 0.0030800483655184507, + 0.010423288680613041, + 0.005658166483044624, + 0.01921928860247135, + 0.012127195484936237, + 0.00656299851834774, + 0.008953147567808628, + 0.01075021829456091, + 0.01965724304318428, + 0.016099147498607635, + 0.010757640935480595, + 0.015488838776946068, + 0.01973334513604641, + 0.008441098034381866, + 0.024450205266475677, + 0.01962420716881752, + 0.01846078783273697, + 0.0317951999604702, + 0.024262061342597008, + 0.05317636951804161, + 0.0029086393769830465, + 0.0019819133449345827, + 0.0030815880745649338, + 0.0025025918148458004, + 0.0025398165453225374, + 0.0035955924540758133, + 0.0033020414412021637, + 0.0032376449089497328 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.016071418300271034 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011865011183544993 + ], + "opposite_logprob_tokens": [ + -6.751186370849609 + ], + "logprob_reward": -0.0011865011183544993, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011865011183544993, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09008845686912537, + 0.037751492112874985, + 0.025513380765914917, + 0.04176938906311989, + 0.042930908501148224, + 0.06432723253965378, + 0.16889210045337677, + 0.41041073203086853, + 0.06792386621236801, + 0.05039243400096893 + ], + "credit_map": [ + 0.006833676714450121, + 0.004902644082903862, + 0.011550957337021828, + 0.007430907338857651, + 0.005922377575188875, + 0.02646699920296669, + 0.012005803175270557, + 0.014975098893046379, + 0.003916982561349869, + 0.002313045784831047, + 0.00520674791187048, + 0.004236327018588781, + 0.003705444047227502, + 0.0055327704176306725, + 0.007741363253444433, + 0.005098814610391855, + 0.0020022550597786903, + 0.002206919016316533, + 0.0025965815875679255, + 0.004406509920954704, + 0.002786730881780386, + 0.003472747281193733, + 0.004620333202183247, + 0.0034213054459542036, + 0.0037772003561258316, + 0.003773287171497941, + 0.005282576195895672, + 0.005360176786780357, + 0.005017201881855726, + 0.007943426258862019, + 0.004912225995212793, + 0.005703297443687916, + 0.003998131025582552, + 0.001951513346284628, + 0.004160026088356972, + 0.003095893422141671, + 0.0031497152522206306, + 0.007848605513572693, + 0.006201898213475943, + 0.01252512913197279, + 0.007144161034375429, + 0.0029909813310950994, + 0.007866168394684792, + 0.005589361302554607, + 0.0033073772210627794, + 0.010332188569009304, + 0.006333019584417343, + 0.020763982087373734, + 0.018493929877877235, + 0.011239411309361458, + 0.02109449729323387, + 0.01772647351026535, + 0.013759946450591087, + 0.024096643552184105, + 0.019498147070407867, + 0.04298307001590729, + 0.05571988597512245, + 0.03485952690243721, + 0.035781379789114, + 0.029685519635677338, + 0.10687974840402603, + 0.06761403381824493, + 0.030659008771181107, + 0.04921167343854904, + 0.00919556524604559, + 0.0027380052488297224, + 0.008714988827705383, + 0.007956898771226406, + 0.006036004051566124, + 0.01132707204669714, + 0.008802956901490688, + 0.013152382336556911, + 0.006835302338004112, + 0.0023910091258585453, + 0.006141263991594315, + 0.0062680370174348354, + 0.005303944461047649, + 0.00841252226382494, + 0.004976368974894285, + 0.010063990950584412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0011865011183544993 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, deliberate advance, focusing with hyper-sharp clarity on a single candle flame desperately clinging to a charred, black wick. The flame, a vibrant drop of light, dances and sways against a backdrop of absolute velvety darkness. Its core burns with a brilliant, almost incandescent intensity, transitioning into a cascade of fiery orange and warm yellow that, at its base, meets the glowing red burr of the wick. Below, a shimmering pool of translucent melted wax reflects the flame's every flicker, its surface trembling with heat. The camera's steady, relentless advance magnifies the flame's struggle, its wavering form casting a soft, warm, and profoundly lonely glow that barely holds back the oppressive, infinite darkness. This slow, mesmerizing movement creates an atmosphere of poignancy and tension, underscoring the raw, fragile beauty of light fighting for survival against the encroaching void, until the flame's dynamic textures and radiant energy fill the frame.", + "seed": 7070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7070_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1143938876921311, + "vlm_logprob": -0.1143938876921311, + "vlm_yes_no_margin": 3.599999952316284, + "vlm_token_credit": -0.1143938876921311, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019696366041898727 + ], + "opposite_logprob_tokens": [ + -4.144696235656738 + ], + "logprob_reward": -0.019696366041898727, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019696366041898727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1318705528974533, + 0.08855453878641129, + 0.14287647604942322, + 0.16634777188301086, + 0.06275864690542221, + 0.06561250984668732, + 0.08017576485872269, + 0.09615181386470795, + 0.08863964676856995, + 0.07701224833726883 + ], + "credit_map": [ + 0.012187283486127853, + 0.0063262744806706905, + 0.019132860004901886, + 0.020465180277824402, + 0.01150625292211771, + 0.014345599338412285, + 0.02226634882390499, + 0.02564074844121933, + 0.01179614569991827, + 0.01192406378686428, + 0.014016433618962765, + 0.01464921422302723, + 0.0047193472273647785, + 0.006055728066712618, + 0.012029056437313557, + 0.013364541344344616, + 0.017454739660024643, + 0.011038064956665039, + 0.02498871088027954, + 0.035632774233818054, + 0.008286584168672562, + 0.00582863949239254, + 0.024623621255159378, + 0.015023340471088886, + 0.02626875974237919, + 0.014026440680027008, + 0.024707091972231865, + 0.022477978840470314, + 0.023680660873651505, + 0.010647876188158989, + 0.027180947363376617, + 0.01735803298652172, + 0.008009807206690311, + 0.005822828039526939, + 0.010418638586997986, + 0.010504537262022495, + 0.004309933632612228, + 0.004389965906739235, + 0.009447384625673294, + 0.009855556301772594, + 0.009761323221027851, + 0.0048812925815582275, + 0.010782331228256226, + 0.013324408791959286, + 0.0063804215751588345, + 0.004441116005182266, + 0.008494026958942413, + 0.007547587621957064, + 0.011735565960407257, + 0.0064030964858829975, + 0.013745659030973911, + 0.013027955777943134, + 0.006765001453459263, + 0.005322874989360571, + 0.014009530656039715, + 0.009166086092591286, + 0.01222329493612051, + 0.0062078265473246574, + 0.01820230856537819, + 0.010970945470035076, + 0.009402723051607609, + 0.007651094347238541, + 0.0154843395575881, + 0.01600928045809269, + 0.0123319486156106, + 0.00858399085700512, + 0.01383671723306179, + 0.010128908790647984, + 0.007567929569631815, + 0.005270130001008511, + 0.013969273306429386, + 0.016950754448771477, + 0.009570916183292866, + 0.00835037138313055, + 0.0114108482375741, + 0.010297724045813084, + 0.01101697888225317, + 0.004161006771028042, + 0.009826187044382095, + 0.012378218583762646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.019696366041898727 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521228790283203 + ], + "opposite_logprob_tokens": [ + -1.5021228790283203 + ], + "logprob_reward": -0.2521228790283203, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521228790283203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12951859831809998, + 0.07406928390264511, + 0.11586248874664307, + 0.10338329523801804, + 0.10434512794017792, + 0.14671088755130768, + 0.10266996175050735, + 0.09694918245077133, + 0.062314849346876144, + 0.06417630612850189 + ], + "credit_map": [ + 0.015733830630779266, + 0.007191339507699013, + 0.021430637687444687, + 0.021807782351970673, + 0.00979651976376772, + 0.011582132428884506, + 0.024168122559785843, + 0.01780823804438114, + 0.007587730418890715, + 0.007640455383807421, + 0.01041712611913681, + 0.010812688618898392, + 0.005864650942385197, + 0.011183207854628563, + 0.00985217746347189, + 0.010711245238780975, + 0.011246184818446636, + 0.011935347691178322, + 0.017639746889472008, + 0.01520467083901167, + 0.010821979492902756, + 0.010668015107512474, + 0.02431136555969715, + 0.014035174623131752, + 0.011758649721741676, + 0.007098806090652943, + 0.017905617132782936, + 0.022314932197332382, + 0.009549281559884548, + 0.007265439257025719, + 0.017086004838347435, + 0.010404564440250397, + 0.014875893481075764, + 0.009235192090272903, + 0.01792720891535282, + 0.027392005547881126, + 0.006402569357305765, + 0.004033372271806002, + 0.01173806469887495, + 0.012740821577608585, + 0.019225915893912315, + 0.012084838934242725, + 0.027068261057138443, + 0.023861581459641457, + 0.012688477523624897, + 0.008151921443641186, + 0.022915111854672432, + 0.020714785903692245, + 0.014603910967707634, + 0.00775359570980072, + 0.016921192407608032, + 0.011642091907560825, + 0.010592485778033733, + 0.0075057270005345345, + 0.021800249814987183, + 0.011850716546177864, + 0.01442197896540165, + 0.0063949646428227425, + 0.0165257528424263, + 0.012859326787292957, + 0.006716293282806873, + 0.0051446883007884026, + 0.012489449232816696, + 0.022396733984351158, + 0.00941245723515749, + 0.006010386627167463, + 0.011246219277381897, + 0.006911187432706356, + 0.007264485117048025, + 0.0041567487642169, + 0.007711880840361118, + 0.009601487778127193, + 0.006782660726457834, + 0.006181030999869108, + 0.009028616361320019, + 0.01277488935738802, + 0.004928345791995525, + 0.004970610607415438, + 0.010571540333330631, + 0.008938612416386604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.2521228790283203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2813938856124878 + ], + "opposite_logprob_tokens": [ + -1.4063938856124878 + ], + "logprob_reward": -0.2813938856124878, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2813938856124878, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2131224125623703, + 0.2676403820514679, + 0.07914501428604126, + 0.08541502058506012, + 0.13874249160289764, + 0.04301207512617111, + 0.038026776164770126, + 0.09238151460886002, + 0.029040733352303505, + 0.01347355730831623 + ], + "credit_map": [ + 0.024702822789549828, + 0.022477641701698303, + 0.03136396408081055, + 0.04046657308936119, + 0.012994852848351002, + 0.017251843586564064, + 0.03533482551574707, + 0.02852988988161087, + 0.03729591518640518, + 0.045686982572078705, + 0.04292502999305725, + 0.047603756189346313, + 0.012206126004457474, + 0.01199116837233305, + 0.035631291568279266, + 0.034300126135349274, + 0.008022027090191841, + 0.005141127854585648, + 0.011774230748414993, + 0.01419140212237835, + 0.0040495553985238075, + 0.0060975500382483006, + 0.017382243648171425, + 0.012486868537962437, + 0.011863495223224163, + 0.005509607493877411, + 0.015881186351180077, + 0.017920786514878273, + 0.00875006802380085, + 0.004438989330083132, + 0.012849611230194569, + 0.008201275952160358, + 0.02473399043083191, + 0.015721505507826805, + 0.02677033096551895, + 0.022166073322296143, + 0.007560062687844038, + 0.008535342290997505, + 0.013867868110537529, + 0.01938732899725437, + 0.0052702222019433975, + 0.002693498507142067, + 0.006687339395284653, + 0.010423015803098679, + 0.0032724810298532248, + 0.0024659172631800175, + 0.005452880170196295, + 0.006746720988303423, + 0.0073705739341676235, + 0.001988140167668462, + 0.0068131317384541035, + 0.00450542289763689, + 0.00335314660333097, + 0.0029208804480731487, + 0.005343949422240257, + 0.005731531418859959, + 0.015008913353085518, + 0.007408732548356056, + 0.018243372440338135, + 0.010175500065088272, + 0.008497401140630245, + 0.005380656104534864, + 0.012010028585791588, + 0.015656907111406326, + 0.003989468794316053, + 0.002811670769006014, + 0.005233271978795528, + 0.0030992187093943357, + 0.002327623078599572, + 0.0020299351308494806, + 0.00352258188650012, + 0.006026962772011757, + 0.0027572421822696924, + 0.0011673448607325554, + 0.0025115415919572115, + 0.0017312477575615048, + 0.001155666308477521, + 0.0008358375052921474, + 0.001417322433553636, + 0.0018973549595102668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.2813938856124878 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819058135151863 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.01819058135151863, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819058135151863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5480855107307434, + 0.051074083894491196, + 0.1671379953622818, + 0.036139462143182755, + 0.049561597406864166, + 0.0736471489071846, + 0.008473431691527367, + 0.023661384359002113, + 0.014896239154040813, + 0.027323130518198013 + ], + "credit_map": [ + 0.05971516668796539, + 0.036948852241039276, + 0.08420026302337646, + 0.07148567587137222, + 0.03062615357339382, + 0.03773517906665802, + 0.12246627360582352, + 0.10490793734788895, + 0.005387379787862301, + 0.006101703736931086, + 0.0072892955504357815, + 0.007391469087451696, + 0.0026687108911573887, + 0.0029184771701693535, + 0.00856424868106842, + 0.010752800852060318, + 0.01729835942387581, + 0.010321272537112236, + 0.026587823405861855, + 0.02581086941063404, + 0.009879009798169136, + 0.01202600821852684, + 0.03974352404475212, + 0.025471126660704613, + 0.004935831762850285, + 0.003048849757760763, + 0.006451005581766367, + 0.004717322066426277, + 0.004968683701008558, + 0.002162894466891885, + 0.005999544635415077, + 0.0038553322665393353, + 0.0069499872624874115, + 0.0027861176058650017, + 0.006872047204524279, + 0.008645176887512207, + 0.0027881592977792025, + 0.0026125581935048103, + 0.010173034854233265, + 0.008734517730772495, + 0.008874685503542423, + 0.007140246219933033, + 0.009574864991009235, + 0.010258185677230358, + 0.011286506429314613, + 0.005490222480148077, + 0.013521326705813408, + 0.007501114159822464, + 0.0011850587325170636, + 0.0007350894156843424, + 0.001471966621465981, + 0.0016333931125700474, + 0.0005774780875071883, + 0.0004923415835946798, + 0.001099039101973176, + 0.0012790648033842444, + 0.0032962060067802668, + 0.0018588625825941563, + 0.003858352079987526, + 0.0030884752050042152, + 0.002080327132716775, + 0.0019038650207221508, + 0.003273247741162777, + 0.004302049055695534, + 0.0022657644003629684, + 0.0008229897357523441, + 0.002647640183568001, + 0.002006090711802244, + 0.00099539989605546, + 0.0009989740792661905, + 0.0023162257857620716, + 0.002843153430148959, + 0.0034998340997844934, + 0.004268126096576452, + 0.004181812983006239, + 0.0033658817410469055, + 0.0029783803038299084, + 0.0015160497277975082, + 0.003808059496805072, + 0.0037049870006740093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.01819058135151863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005657264264300466 + ], + "opposite_logprob_tokens": [ + -7.500565528869629 + ], + "logprob_reward": -0.0005657264264300466, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005657264264300466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1575922966003418, + 0.11414002627134323, + 0.12827154994010925, + 0.0759122297167778, + 0.09369944036006927, + 0.11372938007116318, + 0.08901039510965347, + 0.06792095303535461, + 0.06916392594575882, + 0.09055984765291214 + ], + "credit_map": [ + 0.015782658010721207, + 0.013110481202602386, + 0.022848255932331085, + 0.023356780409812927, + 0.009490886703133583, + 0.01618230901658535, + 0.028595125302672386, + 0.028225796297192574, + 0.011874726042151451, + 0.0100211501121521, + 0.015245876275002956, + 0.01875581033527851, + 0.00750420056283474, + 0.00907928217202425, + 0.02204238995909691, + 0.019616592675447464, + 0.014404579997062683, + 0.009288269095122814, + 0.02017313614487648, + 0.037890948355197906, + 0.0071138180792331696, + 0.006117264274507761, + 0.02240348979830742, + 0.010880045592784882, + 0.010159565135836601, + 0.005241334438323975, + 0.013845655135810375, + 0.011537179350852966, + 0.009628500789403915, + 0.003959342837333679, + 0.013561478815972805, + 0.007979165762662888, + 0.013233932666480541, + 0.004813906271010637, + 0.014185662381350994, + 0.018073853105306625, + 0.004579231608659029, + 0.003939824178814888, + 0.02105647698044777, + 0.01381655689328909, + 0.01605021394789219, + 0.008428867906332016, + 0.017754437401890755, + 0.018755923956632614, + 0.012926326133310795, + 0.0071571823209524155, + 0.019401054829359055, + 0.01325538381934166, + 0.01329671312123537, + 0.005383040755987167, + 0.014810225926339626, + 0.014140089973807335, + 0.007669465150684118, + 0.006219285074621439, + 0.015736866742372513, + 0.01175470557063818, + 0.013218846172094345, + 0.004694798495620489, + 0.013296827673912048, + 0.007847276516258717, + 0.006940843071788549, + 0.0044390298426151276, + 0.00927784014493227, + 0.008205488324165344, + 0.010747565887868404, + 0.004752807319164276, + 0.012213435024023056, + 0.010064025409519672, + 0.0062650274485349655, + 0.005224619526416063, + 0.008955721743404865, + 0.010940724052488804, + 0.010787835344672203, + 0.007947707548737526, + 0.013721469789743423, + 0.019831670448184013, + 0.007936147972941399, + 0.00601439643651247, + 0.011040696874260902, + 0.013279926963150501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.0005657264264300466 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme macro shot begins a slow, deliberate advance, focusing with hyper-sharp clarity on a single candle flame desperately clinging to a charred, black wick. The flame, a vibrant drop of light, dances and sways against a backdrop of absolute velvety darkness. Its core burns with a brilliant, almost incandescent intensity, transitioning into a cascade of fiery orange and warm yellow that, at its base, meets the glowing red burr of the wick. Below, a shimmering pool of translucent melted wax reflects the flame's every flicker, its surface trembling with heat. The camera's steady, relentless advance magnifies the flame's struggle, its wavering form casting a soft, warm, and profoundly lonely glow that barely holds back the oppressive, infinite darkness. This slow, mesmerizing movement creates an atmosphere of poignancy and tension, underscoring the raw, fragile beauty of light fighting for survival against the encroaching void, until the flame's dynamic textures and radiant energy fill the frame.", + "seed": 7071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_1/index_5045_seed_7071_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.033649641834199426, + "vlm_logprob": -0.033649641834199426, + "vlm_yes_no_margin": 4.374999952316284, + "vlm_token_credit": -0.033649641834199426, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008262617513537407 + ], + "opposite_logprob_tokens": [ + -5.383262634277344 + ], + "logprob_reward": -0.008262617513537407, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008262617513537407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051286932080984116, + 0.08666978776454926, + 0.031047100201249123, + 0.028361419215798378, + 0.07298436015844345, + 0.01717122085392475, + 0.5923290848731995, + 0.05038747936487198, + 0.027332961559295654, + 0.042429663240909576 + ], + "credit_map": [ + 0.00505437608808279, + 0.00843041855841875, + 0.009888380765914917, + 0.010797162540256977, + 0.002527856733649969, + 0.0020199704449623823, + 0.007439112290740013, + 0.005129656288772821, + 0.005459317937493324, + 0.03444773703813553, + 0.030269838869571686, + 0.006562823895365, + 0.0022567776031792164, + 0.001676102983765304, + 0.003404695075005293, + 0.0025925012305378914, + 0.0035466940607875586, + 0.005971397738903761, + 0.005390182603150606, + 0.006868927739560604, + 0.001968029886484146, + 0.0018851312343031168, + 0.003060080111026764, + 0.0023566579911857843, + 0.0020442528184503317, + 0.005215339362621307, + 0.006993200164288282, + 0.0039054485969245434, + 0.002364126034080982, + 0.0016543909441679716, + 0.00343228573910892, + 0.002752376953139901, + 0.0037843091413378716, + 0.030169891193509102, + 0.030056266114115715, + 0.0023403398226946592, + 0.0021014795638620853, + 0.001127609284594655, + 0.002100863493978977, + 0.0013035997981205583, + 0.0015691767912358046, + 0.0030919252894818783, + 0.00331138470210135, + 0.002035751473158598, + 0.0017009401926770806, + 0.0011050156317651272, + 0.00255375518463552, + 0.0018032718217000365, + 0.03349035605788231, + 0.2707301676273346, + 0.24478404223918915, + 0.015253504738211632, + 0.011898986995220184, + 0.005733228754252195, + 0.004707812797278166, + 0.005730984732508659, + 0.004322560038417578, + 0.015830526128411293, + 0.012101146392524242, + 0.0037215573247522116, + 0.005054297391325235, + 0.003138534491881728, + 0.0039036059752106667, + 0.002315252786502242, + 0.0032210680656135082, + 0.0061608003452420235, + 0.005284016486257315, + 0.0029681939631700516, + 0.002592699835076928, + 0.0011484733549878001, + 0.00438557518646121, + 0.0015721331583335996, + 0.004486538469791412, + 0.013884395360946655, + 0.010900617577135563, + 0.0023160516284406185, + 0.00406902888789773, + 0.0019338083220645785, + 0.0029931548051536083, + 0.0018460673745721579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.008262617513537407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062060609459877014 + ], + "opposite_logprob_tokens": [ + -2.812060594558716 + ], + "logprob_reward": -0.062060609459877014, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062060609459877014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.017795581370592117, + 0.07847699522972107, + 0.025253979489207268, + 0.017576763406395912, + 0.13050711154937744, + 0.037123631685972214, + 0.435407817363739, + 0.06250092387199402, + 0.049171991646289825, + 0.1461852788925171 + ], + "credit_map": [ + 0.002233098493888974, + 0.0030944945756345987, + 0.003263831837102771, + 0.0036708267871290445, + 0.0008522305870428681, + 0.001313009299337864, + 0.0018629051046445966, + 0.0015051825903356075, + 0.005188336130231619, + 0.03203723579645157, + 0.027457432821393013, + 0.007089347578585148, + 0.001777462661266327, + 0.00100427633151412, + 0.0020796137396246195, + 0.0018432863289490342, + 0.002390461741015315, + 0.004673285409808159, + 0.004868506453931332, + 0.004431389272212982, + 0.0019032526761293411, + 0.0016423685010522604, + 0.0023155072703957558, + 0.0030292070005089045, + 0.0014515670482069254, + 0.004101133439689875, + 0.0040613929741084576, + 0.002714947098866105, + 0.0009011935908347368, + 0.0008686339715495706, + 0.0022207763977348804, + 0.0012571184197440743, + 0.006630008574575186, + 0.04985301196575165, + 0.048534493893384933, + 0.0048138839192688465, + 0.004726224113255739, + 0.003208993235602975, + 0.008975023403763771, + 0.0037654684856534004, + 0.004247758071869612, + 0.00578617537394166, + 0.006691609974950552, + 0.005910778883844614, + 0.0032552366610616446, + 0.0021913277450948954, + 0.005053454544395208, + 0.003987286705523729, + 0.024821823462843895, + 0.19857057929039001, + 0.17891930043697357, + 0.011173398233950138, + 0.008723502047359943, + 0.004071176517754793, + 0.00529166916385293, + 0.003836308605968952, + 0.006041070446372032, + 0.01513577438890934, + 0.013722209259867668, + 0.005849302280694246, + 0.006794390734285116, + 0.003416606690734625, + 0.007910863496363163, + 0.003630702616646886, + 0.006144389510154724, + 0.007385252974927425, + 0.00710025429725647, + 0.006541453767567873, + 0.005294728558510542, + 0.0029735390562564135, + 0.010141578502953053, + 0.0035907921846956015, + 0.015697840601205826, + 0.047958847135305405, + 0.037635914981365204, + 0.009543885476887226, + 0.01089374627918005, + 0.004890196491032839, + 0.011029182933270931, + 0.008535653352737427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 4 + ], + "selected_reward": -0.062060609459877014 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901356369256973 + ], + "opposite_logprob_tokens": [ + -2.5790135860443115 + ], + "logprob_reward": -0.07901356369256973, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901356369256973, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022572627291083336, + 0.1579248458147049, + 0.018010806292295456, + 0.031146371737122536, + 0.13334886729717255, + 0.022963769733905792, + 0.5877232551574707, + 0.011894458904862404, + 0.006908481940627098, + 0.007506540510803461 + ], + "credit_map": [ + 0.0027872894424945116, + 0.00432414747774601, + 0.005932579282671213, + 0.004586309194564819, + 0.0008481263648718596, + 0.0008074533543549478, + 0.002033248310908675, + 0.0012534736888483167, + 0.011535941623151302, + 0.05505063012242317, + 0.05382148176431656, + 0.02387915551662445, + 0.003543847007676959, + 0.002267421456053853, + 0.004292304627597332, + 0.0035340450704097748, + 0.0025429241359233856, + 0.003014815039932728, + 0.003654609899967909, + 0.003764888271689415, + 0.0010589637095108628, + 0.0010308194905519485, + 0.0016959602944552898, + 0.0012478232383728027, + 0.0022605000995099545, + 0.007021344732493162, + 0.0075086988508701324, + 0.006735316477715969, + 0.0012148763053119183, + 0.0017475588247179985, + 0.0025459458120167255, + 0.0021121290046721697, + 0.00816696509718895, + 0.05557875707745552, + 0.05313417315483093, + 0.004923256114125252, + 0.00294137978926301, + 0.0022749495692551136, + 0.0031655223574489355, + 0.0031638548243790865, + 0.0025341762229800224, + 0.004713779781013727, + 0.00644068093970418, + 0.002895050449296832, + 0.0013713042717427015, + 0.0010887132957577705, + 0.0018638119800016284, + 0.0020562505815178156, + 0.03233170136809349, + 0.2675754725933075, + 0.2434222549200058, + 0.015296690165996552, + 0.011800399981439114, + 0.005606585647910833, + 0.006547007709741592, + 0.0051430766470730305, + 0.0013788811629638076, + 0.0024295179173350334, + 0.0022039655596017838, + 0.0019450550898909569, + 0.0012114208657294512, + 0.0005653012194670737, + 0.0013161363312974572, + 0.0008441802929155529, + 0.0006485378253273666, + 0.0015323094557970762, + 0.0012660722713917494, + 0.0007553726900368929, + 0.000523491355124861, + 0.0005708247190341353, + 0.0011845543049275875, + 0.0004273182712495327, + 0.0008327913819812238, + 0.0013020129408687353, + 0.0011231127427890897, + 0.00088428397430107, + 0.0010667189490050077, + 0.00040615079342387617, + 0.0011192770907655358, + 0.0007721918518655002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 4 + ], + "selected_reward": -0.07901356369256973 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819058135151863 + ], + "opposite_logprob_tokens": [ + -4.018190383911133 + ], + "logprob_reward": -0.01819058135151863, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01819058135151863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042402543127536774, + 0.06259013712406158, + 0.022569740191102028, + 0.017604051157832146, + 0.04928888380527496, + 0.028988106176257133, + 0.5582038164138794, + 0.06784971058368683, + 0.07612752169370651, + 0.07437542080879211 + ], + "credit_map": [ + 0.004235892090946436, + 0.009329825639724731, + 0.011934276670217514, + 0.007865285500884056, + 0.0019083484075963497, + 0.0015959269367158413, + 0.0028297773096710443, + 0.0027032154612243176, + 0.003948595374822617, + 0.02307003177702427, + 0.020625002682209015, + 0.006594562903046608, + 0.0017035298515111208, + 0.0012627913383767009, + 0.0024001693818718195, + 0.0029854553285986185, + 0.0018289604922756553, + 0.003951087594032288, + 0.004631260875612497, + 0.005343953613191843, + 0.0009820668492466211, + 0.0007406248478218913, + 0.003193374490365386, + 0.0018984132912009954, + 0.0021215868182480335, + 0.0038035064935684204, + 0.003292677691206336, + 0.004084888845682144, + 0.0009244629763998091, + 0.0008177799172699451, + 0.001448279363103211, + 0.0011108702747151256, + 0.002541362075135112, + 0.018935324624180794, + 0.01861751824617386, + 0.00235480978153646, + 0.0017549543408676982, + 0.0011540886480361223, + 0.002335108583793044, + 0.0015957214636728168, + 0.002812034683302045, + 0.005933292675763369, + 0.006136225536465645, + 0.0035568480379879475, + 0.0020304664503782988, + 0.001413431135006249, + 0.003999871667474508, + 0.0031059382017701864, + 0.030828891322016716, + 0.253894180059433, + 0.23066844046115875, + 0.014843880198895931, + 0.011345287784934044, + 0.005344062577933073, + 0.005909655708819628, + 0.005369476042687893, + 0.0061965384520590305, + 0.02325473353266716, + 0.01912146806716919, + 0.004100354854017496, + 0.004932312294840813, + 0.0031717356760054827, + 0.004768991842865944, + 0.00230358331464231, + 0.00977766141295433, + 0.015536991879343987, + 0.012897438369691372, + 0.008197168819606304, + 0.008919622749090195, + 0.004202656447887421, + 0.011542268097400665, + 0.0050537255592644215, + 0.009928511455655098, + 0.011798989959061146, + 0.012746961787343025, + 0.008326748386025429, + 0.009202388115227222, + 0.0050546349957585335, + 0.010880119167268276, + 0.0064370776526629925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 9 + ], + "selected_reward": -0.01819058135151863 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme macro view focusing on a single candle flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera perform a slow, deliberate advance towards the flame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background depicted as absolute velvety darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a pool of translucent melted wax visible below the flame reflecting its light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the flame display dynamic textures with a core of brilliant intensity transitioning to orange and yellow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007208371534943581 + ], + "opposite_logprob_tokens": [ + -7.250720977783203 + ], + "logprob_reward": -0.0007208371534943581, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007208371534943581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07647562026977539, + 0.07043104618787766, + 0.04298529028892517, + 0.025707531720399857, + 0.24899376928806305, + 0.030898217111825943, + 0.3528108596801758, + 0.06135823205113411, + 0.0290194284170866, + 0.061320025473833084 + ], + "credit_map": [ + 0.006879118271172047, + 0.010560830123722553, + 0.012138931080698967, + 0.018419452011585236, + 0.005853509996086359, + 0.006243565585464239, + 0.009656506590545177, + 0.006723708473145962, + 0.006429694127291441, + 0.015091861598193645, + 0.012910461984574795, + 0.012816138565540314, + 0.00422343984246254, + 0.0034683654084801674, + 0.007033206522464752, + 0.008457877673208714, + 0.0049659619107842445, + 0.0067235445603728294, + 0.009147894568741322, + 0.008814750239253044, + 0.0023624494206160307, + 0.002060716273263097, + 0.005089576356112957, + 0.0038203990552574396, + 0.002583252266049385, + 0.0040196385234594345, + 0.004346909001469612, + 0.004611570853739977, + 0.0024249055422842503, + 0.0025235749781131744, + 0.003374959109351039, + 0.0018227208638563752, + 0.012953641824424267, + 0.10282594710588455, + 0.09980759769678116, + 0.007710999809205532, + 0.006003185641020536, + 0.004221669863909483, + 0.009517579339444637, + 0.005953140091150999, + 0.003105788491666317, + 0.00759151391685009, + 0.006904583424329758, + 0.0034824442118406296, + 0.002277104649692774, + 0.0014207946369424462, + 0.004228025209158659, + 0.0018879628041759133, + 0.019630778580904007, + 0.16004197299480438, + 0.14510256052017212, + 0.009389203041791916, + 0.00733086047694087, + 0.00341921322979033, + 0.00454602250829339, + 0.0033502611331641674, + 0.007270015310496092, + 0.017789708450436592, + 0.012723620049655437, + 0.006108063738793135, + 0.005575425922870636, + 0.00286543695256114, + 0.005849364213645458, + 0.003176596947014332, + 0.0034851697273552418, + 0.004786306526511908, + 0.004679492674767971, + 0.0036443599965423346, + 0.003205783199518919, + 0.001535547780804336, + 0.005122251342982054, + 0.0025605205446481705, + 0.007484354544430971, + 0.009491072036325932, + 0.009312217123806477, + 0.007440095767378807, + 0.00844703521579504, + 0.004730734042823315, + 0.009423801675438881, + 0.0049907187931239605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.0007208371534943581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_tensor([7831, 7831])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_tensor([7831, 7831])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6c11bd32e954163d1684cbd2920986c604a5b649 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_tensor([7831, 7831])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "wide-angle shot, filmed from a slightly upward underwater angle and panning slowly to the right, captures a massive, 3D, cartoon-like blue whale, swimming with majestic composure through the deep, dark, and tranquil depths of the ocean. The whale's skin is velvety midnight blue, transitioning to a soft, pale azure beneath its belly. Its enormous eyes gleam with a gentle, intelligent light, and its massive tail fin gracefully sways in a slow, hypnotic rhythm. Ahead of it, a dense school of hundreds of jellyfish hangs suspended in the water, their translucent bell-shaped bodies and slender tentacles emitting a soft, pulsating bioluminescence that paints the surrounding darkness with ethereal shades of pink, electric blue, and pastel green. As the whale approaches, the side of its massive body delicately and gently brushes the outer edge of the swarm. This gesture is incredibly gentle, like a loving touch. Instead of scattering, the jellyfish gracefully part and swirl in the current created by the whale, creating a vibrant, shimmering curtain of light. The light from the jellyfish casts a shifting, dreamlike shadow on the whale's smooth skin, emphasizing its massive silhouette. The entire scene is imbued with tranquility, wonder, and awe, presenting a magnificent spectacle of life dancing harmoniously in the deep.", + "seed": 10102, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_7831_seed_10102_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02428389899432659, + "vlm_logprob": -0.02428389899432659, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.02428389899432659, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012990372255444527 + ], + "opposite_logprob_tokens": [ + -4.637990474700928 + ], + "logprob_reward": -0.012990372255444527, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.012990372255444527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1363886296749115, + 0.0825018361210823, + 0.05499858036637306, + 0.16569505631923676, + 0.109177365899086, + 0.0413341298699379, + 0.05042766407132149, + 0.23902474343776703, + 0.03979140892624855, + 0.08066056668758392 + ], + "credit_map": [ + 0.019847039133310318, + 0.017376909032464027, + 0.015133033506572247, + 0.0197598934173584, + 0.014176953583955765, + 0.01662217453122139, + 0.014108872972428799, + 0.0193637665361166, + 0.013056771829724312, + 0.011136377230286598, + 0.00820776168256998, + 0.010831941850483418, + 0.00816523376852274, + 0.010670989751815796, + 0.009105069562792778, + 0.011327695101499557, + 0.007144742179661989, + 0.007648383732885122, + 0.0071923802606761456, + 0.011826187372207642, + 0.006307629868388176, + 0.004443109501153231, + 0.0044791679829359055, + 0.005956983659416437, + 0.014880623668432236, + 0.04849494993686676, + 0.04470599815249443, + 0.015674402937293053, + 0.012461892329156399, + 0.007692379876971245, + 0.0129160825163126, + 0.008868718519806862, + 0.011824236251413822, + 0.016048414632678032, + 0.011239362880587578, + 0.008169177919626236, + 0.019118215888738632, + 0.0115133598446846, + 0.016759876161813736, + 0.014504731632769108, + 0.005061861593276262, + 0.00801923032850027, + 0.005876419134438038, + 0.0041611031629145145, + 0.004159220494329929, + 0.004053879529237747, + 0.004754246212542057, + 0.005248173605650663, + 0.007994363084435463, + 0.009370696730911732, + 0.005176310893148184, + 0.004962109960615635, + 0.005164747126400471, + 0.006901503540575504, + 0.005411902442574501, + 0.0054460312239825726, + 0.042815446853637695, + 0.07366953790187836, + 0.022687263786792755, + 0.010399485006928444, + 0.030859289690852165, + 0.018289074301719666, + 0.016801444813609123, + 0.023503225296735764, + 0.004386789631098509, + 0.008593648672103882, + 0.005104480776935816, + 0.0036545100156217813, + 0.005271504633128643, + 0.00382643542252481, + 0.004644957836717367, + 0.00430908240377903, + 0.010208760388195515, + 0.01694577746093273, + 0.009203517809510231, + 0.009375382214784622, + 0.015231652185320854, + 0.005331162363290787, + 0.005937313660979271, + 0.008427002467215061 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.012990372255444527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011092189699411392 + ], + "opposite_logprob_tokens": [ + -4.511092185974121 + ], + "logprob_reward": -0.011092189699411392, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092189699411392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.041277412325143814, + 0.008767134509980679, + 0.042009495198726654, + 0.08688750863075256, + 0.02646302431821823, + 0.08487560600042343, + 0.028156761080026627, + 0.5813165903091431, + 0.03006286919116974, + 0.07018361240625381 + ], + "credit_map": [ + 0.006026103161275387, + 0.005303573329001665, + 0.0045107086189091206, + 0.00586623465642333, + 0.004243253264576197, + 0.004490984603762627, + 0.0035344769712537527, + 0.00730207609012723, + 0.0012113655684515834, + 0.0012280202936381102, + 0.0016846833750605583, + 0.0008234598208218813, + 0.0007721416186541319, + 0.001005986356176436, + 0.0006778960814699531, + 0.0013635812792927027, + 0.005552821327000856, + 0.006123773287981749, + 0.006592977792024612, + 0.00392167316749692, + 0.00433405302464962, + 0.003843877697363496, + 0.004845743998885155, + 0.006794576067477465, + 0.00477480748668313, + 0.03086954914033413, + 0.02541925571858883, + 0.005537872668355703, + 0.007813832722604275, + 0.00323486584238708, + 0.005724478978663683, + 0.0035128453746438026, + 0.0022597911302000284, + 0.007940051145851612, + 0.00494301225990057, + 0.0023094043135643005, + 0.003190410090610385, + 0.00155547633767128, + 0.0018098818836733699, + 0.0024549972731620073, + 0.0063539305701851845, + 0.026728808879852295, + 0.026785247027873993, + 0.005315741989761591, + 0.0060011716559529305, + 0.00413138885051012, + 0.004478007089346647, + 0.005081309005618095, + 0.0033277557231485844, + 0.0056921024806797504, + 0.0035214705858379602, + 0.0028976656030863523, + 0.0029567244928330183, + 0.004232246428728104, + 0.0026923352852463722, + 0.002836461178958416, + 0.13727109134197235, + 0.2520930767059326, + 0.047450173646211624, + 0.014986342750489712, + 0.08142226934432983, + 0.014898831956088543, + 0.00992030743509531, + 0.023274490609765053, + 0.0033749095164239407, + 0.006990163121372461, + 0.0036025438457727432, + 0.0033235575538128614, + 0.00432661734521389, + 0.001964510651305318, + 0.0027552107349038124, + 0.003725355723872781, + 0.009231019765138626, + 0.019572431221604347, + 0.005731891840696335, + 0.006226933561265469, + 0.013311173766851425, + 0.004453625995665789, + 0.005735212005674839, + 0.005921319592744112 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 5 + ], + "selected_reward": -0.011092189699411392 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020583707839250565 + ], + "opposite_logprob_tokens": [ + -3.8955836296081543 + ], + "logprob_reward": -0.020583707839250565, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020583707839250565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028994161635637283, + 0.010675347410142422, + 0.07462617009878159, + 0.10245799273252487, + 0.06310882419347763, + 0.07070819288492203, + 0.0333300344645977, + 0.5644357800483704, + 0.018355490639805794, + 0.03330792486667633 + ], + "credit_map": [ + 0.003344716038554907, + 0.003070143749937415, + 0.0033846055157482624, + 0.003361135022714734, + 0.002842293120920658, + 0.0037958757020533085, + 0.002592161064967513, + 0.006603230722248554, + 0.0014683374902233481, + 0.0012808502651751041, + 0.0018336387583985925, + 0.0013891684357076883, + 0.0010874273721128702, + 0.0010973102180287242, + 0.0010208096355199814, + 0.0014978052349761128, + 0.01734532043337822, + 0.013468264602124691, + 0.008502122946083546, + 0.008185763843357563, + 0.008035518229007721, + 0.006694552954286337, + 0.004654422868043184, + 0.007740207016468048, + 0.006499383132904768, + 0.036985836923122406, + 0.029879052191972733, + 0.007019164506345987, + 0.009078185074031353, + 0.0032875454053282738, + 0.005346199963241816, + 0.0043626208789646626, + 0.006448548752814531, + 0.008878663182258606, + 0.0067211780697107315, + 0.004931414965540171, + 0.01202479936182499, + 0.006313816644251347, + 0.00971269141882658, + 0.0080777108669281, + 0.007202941924333572, + 0.013343594036996365, + 0.012673042714595795, + 0.005727038718760014, + 0.005774928256869316, + 0.008715030737221241, + 0.007762482389807701, + 0.0095091313123703, + 0.004052329808473587, + 0.007294883485883474, + 0.0037131612189114094, + 0.0036905964370816946, + 0.0035461278166621923, + 0.004468071274459362, + 0.003113995771855116, + 0.003450870979577303, + 0.13021230697631836, + 0.23840346932411194, + 0.04712122678756714, + 0.015241701155900955, + 0.07843508571386337, + 0.017988378182053566, + 0.013199102133512497, + 0.023834533989429474, + 0.002127415267750621, + 0.004298850428313017, + 0.0020226447377353907, + 0.0014358189655467868, + 0.0032730225939303637, + 0.0012658796040341258, + 0.0017543777357786894, + 0.002177480375394225, + 0.004974937532097101, + 0.008141573518514633, + 0.0023096196819096804, + 0.002356188604608178, + 0.00819123163819313, + 0.001930127153173089, + 0.00263004912994802, + 0.0027741994708776474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.020583707839250565 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033704325556755066 + ], + "opposite_logprob_tokens": [ + -3.4087042808532715 + ], + "logprob_reward": -0.033704325556755066, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033704325556755066, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04432307928800583, + 0.02002646028995514, + 0.18550188839435577, + 0.056915540248155594, + 0.03466275706887245, + 0.08711395412683487, + 0.03792750462889671, + 0.4295806586742401, + 0.03963366150856018, + 0.06431453675031662 + ], + "credit_map": [ + 0.005749694537371397, + 0.004599909298121929, + 0.005713280290365219, + 0.006459048017859459, + 0.004300642758607864, + 0.005579038057476282, + 0.005687474738806486, + 0.006233986001461744, + 0.0031239318195730448, + 0.0030884025618433952, + 0.0030090815853327513, + 0.002028609625995159, + 0.0020531965419650078, + 0.002464044140651822, + 0.0018133220728486776, + 0.0024458698462694883, + 0.02304934151470661, + 0.0250370055437088, + 0.024135196581482887, + 0.021184910088777542, + 0.023742688819766045, + 0.016792045906186104, + 0.021012157201766968, + 0.03054853156208992, + 0.007643384393304586, + 0.01084892451763153, + 0.009079807437956333, + 0.008001628331840038, + 0.005049779545515776, + 0.004528775345534086, + 0.0072660744190216064, + 0.004497160669416189, + 0.004464158322662115, + 0.006176639813929796, + 0.005052520427852869, + 0.0042073545046150684, + 0.0044451444409787655, + 0.0032857940532267094, + 0.0034170683939009905, + 0.0036140726879239082, + 0.009296135976910591, + 0.022430425509810448, + 0.022145232185721397, + 0.007984393276274204, + 0.0065080709755420685, + 0.00562813738361001, + 0.00628596032038331, + 0.006835592910647392, + 0.004583314061164856, + 0.009134388528764248, + 0.0038883681409060955, + 0.0036412409972399473, + 0.004136281553655863, + 0.004394028801470995, + 0.004101753234863281, + 0.004048124887049198, + 0.08926403522491455, + 0.16084988415241241, + 0.0367431640625, + 0.014750258065760136, + 0.05773479864001274, + 0.023469427600502968, + 0.019704129546880722, + 0.027064945548772812, + 0.005158967804163694, + 0.010203170590102673, + 0.00489417789503932, + 0.0034784937743097544, + 0.005942382849752903, + 0.0027214137371629477, + 0.003483777167275548, + 0.003751275362446904, + 0.008089415729045868, + 0.017621638253331184, + 0.0038065940607339144, + 0.006766451057046652, + 0.012073920108377934, + 0.005854603834450245, + 0.004590877331793308, + 0.0055110277608036995 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 5 + ], + "selected_reward": -0.033704325556755066 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304889962077141 + ], + "opposite_logprob_tokens": [ + -3.168048858642578 + ], + "logprob_reward": -0.04304889962077141, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304889962077141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.024327129125595093, + 0.025747111067175865, + 0.03292662650346756, + 0.040999237447977066, + 0.025755414739251137, + 0.03158058598637581, + 0.025910312309861183, + 0.7608458399772644, + 0.006969210226088762, + 0.024938512593507767 + ], + "credit_map": [ + 0.0024814067874103785, + 0.003246284555643797, + 0.0025448526721447706, + 0.002957892371341586, + 0.002379968063905835, + 0.004136319272220135, + 0.003303440287709236, + 0.0032769644167274237, + 0.004701064433902502, + 0.0030114809051156044, + 0.0029307070653885603, + 0.0033243235666304827, + 0.002148081548511982, + 0.0033820054959505796, + 0.0021698579657822847, + 0.004079590085893869, + 0.00605654064565897, + 0.005061691161245108, + 0.0030767156276851892, + 0.004442671313881874, + 0.0036111727822571993, + 0.0024482139851897955, + 0.002844512462615967, + 0.005385108292102814, + 0.004469455685466528, + 0.013693531975150108, + 0.009926475584506989, + 0.0027059230487793684, + 0.0038357486482709646, + 0.0015832119388505816, + 0.0025648546870797873, + 0.0022200369276106358, + 0.004410595633089542, + 0.0036861365661025047, + 0.0033696836326271296, + 0.0035578468814492226, + 0.0026382417418062687, + 0.002161844866350293, + 0.0027915684040635824, + 0.0031394953839480877, + 0.0036056002136319876, + 0.0052282256074249744, + 0.004529597237706184, + 0.0035993994679301977, + 0.0029325212817639112, + 0.0035442111548036337, + 0.0032597624231129885, + 0.004881265107542276, + 0.002953509334474802, + 0.005043426528573036, + 0.0029533610213547945, + 0.0029699576552957296, + 0.002747963648289442, + 0.0032744226045906544, + 0.0029428289271891117, + 0.0030248428229242563, + 0.17771954834461212, + 0.32463428378105164, + 0.05684678256511688, + 0.022791434079408646, + 0.10234708338975906, + 0.022666655480861664, + 0.01683361828327179, + 0.0370064415037632, + 0.0008170318324118853, + 0.001324090058915317, + 0.0008478163508698344, + 0.0007390165119431913, + 0.0009677666239440441, + 0.0006927113863639534, + 0.0008578794659115374, + 0.0007228984031826258, + 0.00391710689291358, + 0.006259779445827007, + 0.0014472706243395805, + 0.0016918075270950794, + 0.006162493489682674, + 0.0016389352967962623, + 0.002073387149721384, + 0.0017477328656241298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.04304889962077141 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, filmed from a slightly upward underwater angle and panning slowly to the right, captures a massive, 3D, cartoon-like blue whale, swimming with majestic composure through the deep, dark, and tranquil depths of the ocean. The whale's skin is velvety midnight blue, transitioning to a soft, pale azure beneath its belly. Its enormous eyes gleam with a gentle, intelligent light, and its massive tail fin gracefully sways in a slow, hypnotic rhythm. Ahead of it, a dense school of hundreds of jellyfish hangs suspended in the water, their translucent bell-shaped bodies and slender tentacles emitting a soft, pulsating bioluminescence that paints the surrounding darkness with ethereal shades of pink, electric blue, and pastel green. As the whale approaches, the side of its massive body delicately and gently brushes the outer edge of the swarm. This gesture is incredibly gentle, like a loving touch. Instead of scattering, the jellyfish gracefully part and swirl in the current created by the whale, creating a vibrant, shimmering curtain of light. The light from the jellyfish casts a shifting, dreamlike shadow on the whale's smooth skin, emphasizing its massive silhouette. The entire scene is imbued with tranquility, wonder, and awe, presenting a magnificent spectacle of life dancing harmoniously in the deep.", + "seed": 10103, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_7831_seed_10103_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.022091750334948302, + "vlm_logprob": -0.022091750334948302, + "vlm_yes_no_margin": 4.0749999523162845, + "vlm_token_credit": -0.022091750334948302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006300941575318575 + ], + "opposite_logprob_tokens": [ + -5.506300926208496 + ], + "logprob_reward": -0.006300941575318575, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.006300941575318575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12756134569644928, + 0.13215962052345276, + 0.08692716807126999, + 0.24472256004810333, + 0.08276720345020294, + 0.05574100464582443, + 0.06432527303695679, + 0.053464632481336594, + 0.07637669146060944, + 0.07595449686050415 + ], + "credit_map": [ + 0.019328707829117775, + 0.02582409977912903, + 0.016903545707464218, + 0.012286282144486904, + 0.018490498885512352, + 0.015282915905117989, + 0.008120561018586159, + 0.011324731633067131, + 0.014930123463273048, + 0.022302351891994476, + 0.01584877073764801, + 0.019895559176802635, + 0.018510332331061363, + 0.016113486140966415, + 0.012080034241080284, + 0.012478968128561974, + 0.013954277150332928, + 0.009238648228347301, + 0.016112033277750015, + 0.009488306939601898, + 0.00810176506638527, + 0.01318441703915596, + 0.007874143309891224, + 0.00897357426583767, + 0.025637375190854073, + 0.06195084750652313, + 0.013237136416137218, + 0.014342299662530422, + 0.03191186860203743, + 0.03647993132472038, + 0.03457285091280937, + 0.0265902541577816, + 0.009420407935976982, + 0.010699616745114326, + 0.008902668952941895, + 0.00537640368565917, + 0.013466172851622105, + 0.013184433802962303, + 0.008406106382608414, + 0.013311396352946758, + 0.003974891733378172, + 0.004645901266485453, + 0.002798314206302166, + 0.0029063972178846598, + 0.016100376844406128, + 0.005746361333876848, + 0.00264522316865623, + 0.016923541203141212, + 0.006273017730563879, + 0.007538997568190098, + 0.008359025232493877, + 0.007951421663165092, + 0.010050185956060886, + 0.00880552176386118, + 0.007778832223266363, + 0.007568273693323135, + 0.005473433528095484, + 0.01136463601142168, + 0.004794502165168524, + 0.003996680956333876, + 0.013444014824926853, + 0.00732223829254508, + 0.003444599686190486, + 0.003624530741944909, + 0.011937360279262066, + 0.011808757670223713, + 0.006809648592025042, + 0.007296810857951641, + 0.010108640417456627, + 0.013506466522812843, + 0.008084879256784916, + 0.006824133452028036, + 0.012723026797175407, + 0.012491772882640362, + 0.007906780578196049, + 0.006263604387640953, + 0.013003427535295486, + 0.010485664941370487, + 0.007436850573867559, + 0.005643371492624283 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.006300941575318575 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018192455172538757 + ], + "opposite_logprob_tokens": [ + -4.018192291259766 + ], + "logprob_reward": -0.018192455172538757, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018192455172538757, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1816147416830063, + 0.18424396216869354, + 0.05009560286998749, + 0.16007471084594727, + 0.08068009465932846, + 0.07952602207660675, + 0.060977958142757416, + 0.07987067103385925, + 0.09514498710632324, + 0.027771301567554474 + ], + "credit_map": [ + 0.022958951070904732, + 0.03966597840189934, + 0.02091287449002266, + 0.018284866586327553, + 0.027580561116337776, + 0.02481103502213955, + 0.012681497260928154, + 0.014718972146511078, + 0.024879662320017815, + 0.04559081420302391, + 0.021312154829502106, + 0.021147869527339935, + 0.024243878200650215, + 0.016720103099942207, + 0.010863701812922955, + 0.019485758617520332, + 0.00649260263890028, + 0.008353596553206444, + 0.006598582956939936, + 0.004428659100085497, + 0.0051095071248710155, + 0.006175088696181774, + 0.005398632492870092, + 0.007538928184658289, + 0.018406901508569717, + 0.04274618253111839, + 0.008130280300974846, + 0.008865616284310818, + 0.018571017310023308, + 0.024382444098591805, + 0.020918937399983406, + 0.0180533267557621, + 0.0058508748188614845, + 0.01044539175927639, + 0.007906551472842693, + 0.007089745253324509, + 0.012854300439357758, + 0.012016276828944683, + 0.010202402248978615, + 0.014314546249806881, + 0.009528394788503647, + 0.010701609775424004, + 0.00880449265241623, + 0.006632387172430754, + 0.009200469590723515, + 0.01586066745221615, + 0.007181460037827492, + 0.011616530828177929, + 0.008009283803403378, + 0.00750465365126729, + 0.007440307643264532, + 0.007133514620363712, + 0.008479993790388107, + 0.008422845974564552, + 0.006734943948686123, + 0.007252405397593975, + 0.008492576889693737, + 0.017802877351641655, + 0.006975136697292328, + 0.006436273455619812, + 0.01597614213824272, + 0.012683545239269733, + 0.00586994644254446, + 0.005634164437651634, + 0.013392637483775616, + 0.01640472561120987, + 0.012160036712884903, + 0.009249979630112648, + 0.01648918353021145, + 0.010104107670485973, + 0.012149635702371597, + 0.0051946742460131645, + 0.003242968115955591, + 0.0036635813303291798, + 0.0033102419693022966, + 0.0033074363600462675, + 0.004082037601619959, + 0.004025042988359928, + 0.003176541067659855, + 0.002963449340313673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.018192455172538757 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07813503593206406, + 0.03480902314186096, + 0.035873595625162125, + 0.34299004077911377, + 0.03316250070929527, + 0.1273285448551178, + 0.13097582757472992, + 0.06267701834440231, + 0.10209763795137405, + 0.05195075646042824 + ], + "credit_map": [ + 0.013193693943321705, + 0.014472130686044693, + 0.011004764586687088, + 0.008579355664551258, + 0.00965183787047863, + 0.010964123532176018, + 0.004278477746993303, + 0.005990657024085522, + 0.0041666897013783455, + 0.008095086552202702, + 0.004854774102568626, + 0.003994565457105637, + 0.004668313078582287, + 0.00351409032009542, + 0.0025077161844819784, + 0.003007790306583047, + 0.005410592537373304, + 0.005340167786926031, + 0.004350945353507996, + 0.003762071253731847, + 0.0034317222889512777, + 0.0047438484616577625, + 0.003758486593142152, + 0.00507576297968626, + 0.03628000617027283, + 0.08326996862888336, + 0.019877271726727486, + 0.021725615486502647, + 0.04297072812914848, + 0.04862691089510918, + 0.052426207810640335, + 0.03781336173415184, + 0.0037615145556628704, + 0.0038174286019057035, + 0.0028307014144957066, + 0.0021234930027276278, + 0.005730184260755777, + 0.0038275686092674732, + 0.004473537672311068, + 0.006598076317459345, + 0.011079926043748856, + 0.013413789682090282, + 0.016459889709949493, + 0.010334042832255363, + 0.01716339774429798, + 0.019280724227428436, + 0.00838758610188961, + 0.031209208071231842, + 0.016462288796901703, + 0.01712026447057724, + 0.01163306925445795, + 0.011737318709492683, + 0.02158014103770256, + 0.026254694908857346, + 0.013922931626439095, + 0.01226513646543026, + 0.0073011224158108234, + 0.01412109099328518, + 0.006174457725137472, + 0.005326421931385994, + 0.01313685905188322, + 0.006889374926686287, + 0.004656004719436169, + 0.005071684252470732, + 0.011503336019814014, + 0.01970711350440979, + 0.009676462039351463, + 0.005960832349956036, + 0.02567303366959095, + 0.011142987757921219, + 0.013024658896028996, + 0.005409230012446642, + 0.005591737572103739, + 0.007548075169324875, + 0.0046019889414310455, + 0.004301281180232763, + 0.013468645513057709, + 0.007256713230162859, + 0.005167012102901936, + 0.004015310201793909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369026258587837 + ], + "opposite_logprob_tokens": [ + -3.4086902141571045 + ], + "logprob_reward": -0.03369026258587837, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369026258587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0818474143743515, + 0.0821838304400444, + 0.06381505727767944, + 0.31578150391578674, + 0.0640435516834259, + 0.10401009768247604, + 0.06095568835735321, + 0.1058289110660553, + 0.07158461958169937, + 0.0499492846429348 + ], + "credit_map": [ + 0.014050183817744255, + 0.014113218523561954, + 0.01000497117638588, + 0.007597592659294605, + 0.009476619772613049, + 0.011726920492947102, + 0.006939595099538565, + 0.007938320748507977, + 0.010912254452705383, + 0.013201676309108734, + 0.011024088598787785, + 0.010602510534226894, + 0.012762479484081268, + 0.009690402075648308, + 0.005652058403939009, + 0.008338365703821182, + 0.010986046865582466, + 0.0080069275572896, + 0.01190256793051958, + 0.006460640113800764, + 0.005237745586782694, + 0.007821575738489628, + 0.0068360501900315285, + 0.00656350702047348, + 0.033204689621925354, + 0.0857657864689827, + 0.01579531468451023, + 0.018730852752923965, + 0.03840511292219162, + 0.04838048666715622, + 0.04178231582045555, + 0.03371695056557655, + 0.007496644277125597, + 0.007411588449031115, + 0.007734762504696846, + 0.0038751582615077496, + 0.010766424238681793, + 0.00898654293268919, + 0.008198758587241173, + 0.0095736775547266, + 0.010931292548775673, + 0.009179037064313889, + 0.01597258634865284, + 0.010109676979482174, + 0.01753254607319832, + 0.011747634038329124, + 0.006129049696028233, + 0.02240827865898609, + 0.009232118725776672, + 0.0058119879104197025, + 0.01005824375897646, + 0.011720542795956135, + 0.007210899610072374, + 0.007230755873024464, + 0.004540636669844389, + 0.005150508135557175, + 0.00806874968111515, + 0.02526857517659664, + 0.006802376359701157, + 0.004496081732213497, + 0.02875879406929016, + 0.019202766939997673, + 0.007575321942567825, + 0.005656249355524778, + 0.011120899580419064, + 0.014129837974905968, + 0.006843107286840677, + 0.0057170335203409195, + 0.012793606147170067, + 0.008012360893189907, + 0.008798089809715748, + 0.0041696904227137566, + 0.007663355208933353, + 0.007084304932504892, + 0.005696907639503479, + 0.005467972252517939, + 0.00812173169106245, + 0.00674896314740181, + 0.005068700760602951, + 0.004097351338714361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.03369026258587837 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808249533176422 + ], + "opposite_logprob_tokens": [ + -3.2880825996398926 + ], + "logprob_reward": -0.03808249533176422, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808249533176422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051274459809064865, + 0.08243966102600098, + 0.07528562843799591, + 0.264751136302948, + 0.04563077539205551, + 0.03737705200910568, + 0.05908295512199402, + 0.15555576980113983, + 0.06773605197668076, + 0.16086651384830475 + ], + "credit_map": [ + 0.008667750284075737, + 0.007466320879757404, + 0.006870147306472063, + 0.0036473500076681376, + 0.00771202240139246, + 0.007872410118579865, + 0.004465143661946058, + 0.0045733158476650715, + 0.01154955755919218, + 0.014663610607385635, + 0.010130779817700386, + 0.011016905307769775, + 0.010259868577122688, + 0.009395255707204342, + 0.006143977399915457, + 0.0092797065153718, + 0.010113143362104893, + 0.009871854446828365, + 0.00924141239374876, + 0.011403342708945274, + 0.008619561791419983, + 0.012071751058101654, + 0.006259317509829998, + 0.007705241907387972, + 0.030369989573955536, + 0.06593151390552521, + 0.016169553622603416, + 0.015053470619022846, + 0.03228537738323212, + 0.038694899529218674, + 0.04003550857305527, + 0.02621079608798027, + 0.004166249185800552, + 0.005807277746498585, + 0.005453700665384531, + 0.0038380296900868416, + 0.0066704051569104195, + 0.005959421396255493, + 0.00657739769667387, + 0.007158288732171059, + 0.005635088309645653, + 0.004560443107038736, + 0.004991894122213125, + 0.0036894422955811024, + 0.0047189462929964066, + 0.006094593089073896, + 0.002820934634655714, + 0.004865708760917187, + 0.006953001953661442, + 0.007244393229484558, + 0.006239017006009817, + 0.004616176709532738, + 0.010115692391991615, + 0.010448862798511982, + 0.004545833449810743, + 0.00891997292637825, + 0.01545125525444746, + 0.032704923301935196, + 0.009570599533617496, + 0.008811559528112411, + 0.03418610990047455, + 0.033915892243385315, + 0.012037607841193676, + 0.008877820335328579, + 0.005830548238009214, + 0.015323265455663204, + 0.00409013032913208, + 0.006651381496340036, + 0.016384581103920937, + 0.007953397929668427, + 0.008030909113585949, + 0.00347183202393353, + 0.016537675634026527, + 0.01746346801519394, + 0.020661665126681328, + 0.01515225786715746, + 0.019613226875662804, + 0.025578301399946213, + 0.018529269844293594, + 0.027330651879310608 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 7 + ], + "selected_reward": -0.03808249533176422 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, filmed from a slightly upward underwater angle and panning slowly to the right, captures a massive, 3D, cartoon-like blue whale, swimming with majestic composure through the deep, dark, and tranquil depths of the ocean. The whale's skin is velvety midnight blue, transitioning to a soft, pale azure beneath its belly. Its enormous eyes gleam with a gentle, intelligent light, and its massive tail fin gracefully sways in a slow, hypnotic rhythm. Ahead of it, a dense school of hundreds of jellyfish hangs suspended in the water, their translucent bell-shaped bodies and slender tentacles emitting a soft, pulsating bioluminescence that paints the surrounding darkness with ethereal shades of pink, electric blue, and pastel green. As the whale approaches, the side of its massive body delicately and gently brushes the outer edge of the swarm. This gesture is incredibly gentle, like a loving touch. Instead of scattering, the jellyfish gracefully part and swirl in the current created by the whale, creating a vibrant, shimmering curtain of light. The light from the jellyfish casts a shifting, dreamlike shadow on the whale's smooth skin, emphasizing its massive silhouette. The entire scene is imbued with tranquility, wonder, and awe, presenting a magnificent spectacle of life dancing harmoniously in the deep.", + "seed": 10104, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_7831_seed_10104_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.057057302072644235, + "vlm_logprob": -0.057057302072644235, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.057057302072644235, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015036923810839653 + ], + "opposite_logprob_tokens": [ + -4.3900370597839355 + ], + "logprob_reward": -0.015036923810839653, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.015036923810839653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2637064754962921, + 0.10994873195886612, + 0.03878980875015259, + 0.14482519030570984, + 0.1663704365491867, + 0.027904681861400604, + 0.08898592740297318, + 0.016263684257864952, + 0.02511201612651348, + 0.11809302866458893 + ], + "credit_map": [ + 0.015762094408273697, + 0.08044077455997467, + 0.09330057352781296, + 0.017099715769290924, + 0.021251890808343887, + 0.01224428229033947, + 0.012568667531013489, + 0.011038469150662422, + 0.008082270622253418, + 0.02106451243162155, + 0.015035944990813732, + 0.019093569368124008, + 0.009707081131637096, + 0.01546721626073122, + 0.012711833231151104, + 0.008786310441792011, + 0.003447715425863862, + 0.005547270644456148, + 0.005446556489914656, + 0.003921941388398409, + 0.005489557981491089, + 0.005583786405622959, + 0.004140803590416908, + 0.0052121784538030624, + 0.011820952408015728, + 0.05219035595655441, + 0.006297031883150339, + 0.04059194028377533, + 0.009541084058582783, + 0.00960454624146223, + 0.00815898459404707, + 0.006620291154831648, + 0.03236382454633713, + 0.06002483889460564, + 0.012399719096720219, + 0.026718154549598694, + 0.010287598706781864, + 0.01010844949632883, + 0.007858245633542538, + 0.00660961028188467, + 0.0030620142351835966, + 0.002269021701067686, + 0.005049421917647123, + 0.00347237940877676, + 0.002768479287624359, + 0.0037029723171144724, + 0.0038566517177969217, + 0.0037237415090203285, + 0.009123936295509338, + 0.006774777080863714, + 0.013153993524610996, + 0.007986640557646751, + 0.006856046617031097, + 0.01625082828104496, + 0.014548735693097115, + 0.014290967956185341, + 0.0015381172997877002, + 0.0011456129141151905, + 0.0024232263676822186, + 0.002511808881536126, + 0.00166355783585459, + 0.003012833883985877, + 0.001859043724834919, + 0.002109482418745756, + 0.002057528356090188, + 0.002629418158903718, + 0.002822689712047577, + 0.004214002750813961, + 0.0029916020575910807, + 0.00365529116243124, + 0.003991239704191685, + 0.002750242594629526, + 0.008156191557645798, + 0.013854890130460262, + 0.014331855811178684, + 0.013944552280008793, + 0.010254872031509876, + 0.014841997995972633, + 0.019355494529008865, + 0.023353174328804016 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.015036923810839653 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197137862443924 + ], + "opposite_logprob_tokens": [ + -4.018197059631348 + ], + "logprob_reward": -0.018197137862443924, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197137862443924, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17485058307647705, + 0.18576225638389587, + 0.09707373380661011, + 0.08000187575817108, + 0.21389709413051605, + 0.05718039721250534, + 0.0778338611125946, + 0.02199719473719597, + 0.030562203377485275, + 0.06084074079990387 + ], + "credit_map": [ + 0.010043368674814701, + 0.0503116250038147, + 0.058488134294748306, + 0.012836971320211887, + 0.014963888563215733, + 0.007954065687954426, + 0.010938365012407303, + 0.009314149618148804, + 0.009098583832383156, + 0.0551011897623539, + 0.040098175406455994, + 0.020535752177238464, + 0.02079005353152752, + 0.015171726234257221, + 0.011547237634658813, + 0.013419553637504578, + 0.010851218365132809, + 0.014166559092700481, + 0.010797286406159401, + 0.014675174839794636, + 0.013725493103265762, + 0.012363141402602196, + 0.008894972503185272, + 0.011599891819059849, + 0.006608346477150917, + 0.01906440407037735, + 0.005329613108187914, + 0.01793886162340641, + 0.0069410111755132675, + 0.009923107922077179, + 0.006927318871021271, + 0.007269206922501326, + 0.01464757788926363, + 0.05742207169532776, + 0.03448547422885895, + 0.07929351925849915, + 0.008611371740698814, + 0.00807590689510107, + 0.007726251147687435, + 0.003634916851297021, + 0.005187598522752523, + 0.004293940495699644, + 0.008865601383149624, + 0.006361821666359901, + 0.006251445040106773, + 0.007883607409894466, + 0.010001025162637234, + 0.008335361257195473, + 0.0070508867502212524, + 0.004606977570801973, + 0.022044634446501732, + 0.013969427905976772, + 0.005206417758017778, + 0.009972388856112957, + 0.008472214452922344, + 0.006510921288281679, + 0.001956034218892455, + 0.0012053587706759572, + 0.00299119227565825, + 0.0035501474048942327, + 0.002385970437899232, + 0.004716912750154734, + 0.003161277621984482, + 0.0020303018391132355, + 0.0023432818707078695, + 0.0030537229031324387, + 0.0032841397915035486, + 0.0050634779036045074, + 0.0036582795437425375, + 0.00535200210288167, + 0.004468581173568964, + 0.0033387187868356705, + 0.00351702724583447, + 0.0060879564844071865, + 0.007311348803341389, + 0.008259065449237823, + 0.005112701095640659, + 0.007284361869096756, + 0.009662522934377193, + 0.013605755753815174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.018197137862443924 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035276651382446 + ], + "opposite_logprob_tokens": [ + -1.9103527069091797 + ], + "logprob_reward": -0.16035276651382446, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035276651382446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0896814689040184, + 0.10360574722290039, + 0.044331006705760956, + 0.06800790876150131, + 0.28906598687171936, + 0.0691385492682457, + 0.14112266898155212, + 0.036100633442401886, + 0.03839637339115143, + 0.12054961174726486 + ], + "credit_map": [ + 0.0051689110696315765, + 0.027644935995340347, + 0.0309565681964159, + 0.005536871962249279, + 0.007238027639687061, + 0.004114643204957247, + 0.004684245679527521, + 0.004337270278483629, + 0.00514551205560565, + 0.0326058454811573, + 0.02304116077721119, + 0.006735336035490036, + 0.01140922587364912, + 0.010811559855937958, + 0.005918882321566343, + 0.007938232272863388, + 0.0038620654959231615, + 0.006198024842888117, + 0.00492776557803154, + 0.006172287277877331, + 0.005896358285099268, + 0.008157208561897278, + 0.004293588921427727, + 0.004823709838092327, + 0.006131431087851524, + 0.015888087451457977, + 0.005965460557490587, + 0.01145612820982933, + 0.005999200977385044, + 0.010345854796469212, + 0.0061170621775090694, + 0.006104685831815004, + 0.017077051103115082, + 0.07083601504564285, + 0.0512082576751709, + 0.11360590904951096, + 0.010471106506884098, + 0.010073193348944187, + 0.010683635249733925, + 0.005110837984830141, + 0.006418255157768726, + 0.005257898010313511, + 0.009598092176020145, + 0.009621177799999714, + 0.007087480742484331, + 0.009268094785511494, + 0.012358974665403366, + 0.009528576396405697, + 0.011585821397602558, + 0.00998023059219122, + 0.023974595591425896, + 0.019205350428819656, + 0.011134681291878223, + 0.02497989870607853, + 0.02136608026921749, + 0.018896011635661125, + 0.003234574105590582, + 0.0019817929714918137, + 0.00563003309071064, + 0.005800432525575161, + 0.0036274250596761703, + 0.007507901173084974, + 0.005142188630998135, + 0.003176288679242134, + 0.0025367012713104486, + 0.0029754117131233215, + 0.003918068017810583, + 0.006426042411476374, + 0.004217825829982758, + 0.007349939085543156, + 0.0068635218776762486, + 0.004108866676688194, + 0.007087612058967352, + 0.012627137824892998, + 0.015423111617565155, + 0.014640961773693562, + 0.010050502605736256, + 0.014651400968432426, + 0.02047327533364296, + 0.025595620274543762 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 9 + ], + "selected_reward": -0.16035276651382446 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654209822416306 + ], + "opposite_logprob_tokens": [ + -3.048654317855835 + ], + "logprob_reward": -0.048654209822416306, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654209822416306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22705650329589844, + 0.08628379553556442, + 0.10954036563634872, + 0.16514427959918976, + 0.1583411693572998, + 0.05392148345708847, + 0.07642452418804169, + 0.02737238071858883, + 0.05019364133477211, + 0.04572185501456261 + ], + "credit_map": [ + 0.015024290420114994, + 0.051042262464761734, + 0.06369967758655548, + 0.023896601051092148, + 0.02115609124302864, + 0.020020412281155586, + 0.020359041169285774, + 0.01185813918709755, + 0.006351116579025984, + 0.011598305776715279, + 0.009563435800373554, + 0.016237758100032806, + 0.00954289361834526, + 0.01133083738386631, + 0.01153471227735281, + 0.010124740190804005, + 0.009597145952284336, + 0.013730429112911224, + 0.014461458660662174, + 0.016766319051384926, + 0.014454279094934464, + 0.016702987253665924, + 0.00878109596669674, + 0.015046658925712109, + 0.016147002577781677, + 0.045522917062044144, + 0.013202578760683537, + 0.03193475678563118, + 0.014282084070146084, + 0.017061514779925346, + 0.013180619105696678, + 0.013812809251248837, + 0.016446588560938835, + 0.013586144894361496, + 0.03799756243824959, + 0.06548859179019928, + 0.005462342407554388, + 0.006446424871683121, + 0.0077052549459040165, + 0.0052082533948123455, + 0.005530407652258873, + 0.004247531294822693, + 0.006639606785029173, + 0.009590486995875835, + 0.005967228207737207, + 0.00785883516073227, + 0.006995866075158119, + 0.007091522216796875, + 0.00958473701030016, + 0.0049462527967989445, + 0.0195439662784338, + 0.008081811480224133, + 0.004401543643325567, + 0.010578042827546597, + 0.008832240477204323, + 0.010455932468175888, + 0.0026008333079516888, + 0.0018313858890905976, + 0.004651433788239956, + 0.0030781461391597986, + 0.0035197227261960506, + 0.0068437280133366585, + 0.002433470217511058, + 0.0024136602878570557, + 0.005358199123293161, + 0.007152002304792404, + 0.005594958085566759, + 0.00875891000032425, + 0.005378784146159887, + 0.006266563665121794, + 0.007205836940556765, + 0.004478387534618378, + 0.0037720894906669855, + 0.005080512724816799, + 0.007265836000442505, + 0.005906837061047554, + 0.004590311553329229, + 0.006655762437731028, + 0.00710884016007185, + 0.005341666750609875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.048654209822416306 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304547235369682 + ], + "opposite_logprob_tokens": [ + -3.1680455207824707 + ], + "logprob_reward": -0.04304547235369682, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304547235369682, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16593044996261597, + 0.0829055905342102, + 0.053137220442295074, + 0.23387116193771362, + 0.19414594769477844, + 0.07623415440320969, + 0.06161818280816078, + 0.03861690312623978, + 0.06165279448032379, + 0.031887564808130264 + ], + "credit_map": [ + 0.011019120924174786, + 0.04737474024295807, + 0.05379931628704071, + 0.010479208081960678, + 0.011691113002598286, + 0.010788686573505402, + 0.009529859758913517, + 0.011248419061303139, + 0.009350246749818325, + 0.013378643430769444, + 0.012447202578186989, + 0.016721120104193687, + 0.008391467854380608, + 0.008524924516677856, + 0.007148474920541048, + 0.006943510379642248, + 0.005019560921937227, + 0.00979077722877264, + 0.008814230561256409, + 0.005889714229851961, + 0.007177382707595825, + 0.005315676331520081, + 0.005684134550392628, + 0.005445745773613453, + 0.02025415562093258, + 0.04021759703755379, + 0.029015690088272095, + 0.03325233235955238, + 0.02252727933228016, + 0.035163916647434235, + 0.026097089052200317, + 0.027343114838004112, + 0.03992977365851402, + 0.05903877317905426, + 0.023776862770318985, + 0.018036792054772377, + 0.013858954422175884, + 0.014993743039667606, + 0.013425249606370926, + 0.01108578685671091, + 0.007301469333469868, + 0.006354436744004488, + 0.009543562307953835, + 0.013426987454295158, + 0.00791215244680643, + 0.011449459940195084, + 0.011835139244794846, + 0.008410952985286713, + 0.009355461224913597, + 0.0035767110530287027, + 0.016091249883174896, + 0.008783449418842793, + 0.004326161928474903, + 0.007357336580753326, + 0.005806287284940481, + 0.006321520544588566, + 0.0036635783035308123, + 0.002838896820321679, + 0.0039590466767549515, + 0.004117676056921482, + 0.00411489512771368, + 0.00761497812345624, + 0.006558175198733807, + 0.005749658215790987, + 0.005163409747183323, + 0.005245665088295937, + 0.0063297138549387455, + 0.010269896127283573, + 0.005637764930725098, + 0.010652821511030197, + 0.011193012818694115, + 0.007160511799156666, + 0.0031646641436964273, + 0.003506903536617756, + 0.004633177537471056, + 0.005674127023667097, + 0.0027407011948525906, + 0.00395264383405447, + 0.003514332929626107, + 0.004701012745499611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.04304547235369682 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "wide-angle shot, filmed from a slightly upward underwater angle and panning slowly to the right, captures a massive, 3D, cartoon-like blue whale, swimming with majestic composure through the deep, dark, and tranquil depths of the ocean. The whale's skin is velvety midnight blue, transitioning to a soft, pale azure beneath its belly. Its enormous eyes gleam with a gentle, intelligent light, and its massive tail fin gracefully sways in a slow, hypnotic rhythm. Ahead of it, a dense school of hundreds of jellyfish hangs suspended in the water, their translucent bell-shaped bodies and slender tentacles emitting a soft, pulsating bioluminescence that paints the surrounding darkness with ethereal shades of pink, electric blue, and pastel green. As the whale approaches, the side of its massive body delicately and gently brushes the outer edge of the swarm. This gesture is incredibly gentle, like a loving touch. Instead of scattering, the jellyfish gracefully part and swirl in the current created by the whale, creating a vibrant, shimmering curtain of light. The light from the jellyfish casts a shifting, dreamlike shadow on the whale's smooth skin, emphasizing its massive silhouette. The entire scene is imbued with tranquility, wonder, and awe, presenting a magnificent spectacle of life dancing harmoniously in the deep.", + "seed": 10105, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_11/index_7831_seed_10105_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02044827356003225, + "vlm_logprob": -0.02044827356003225, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.02044827356003225, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008738137781620026 + ], + "opposite_logprob_tokens": [ + -5.383738040924072 + ], + "logprob_reward": -0.008738137781620026, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.008738137781620026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22120067477226257, + 0.1143345981836319, + 0.06347974389791489, + 0.026332005858421326, + 0.08923874795436859, + 0.03598836809396744, + 0.010800021700561047, + 0.07351122051477432, + 0.30351722240448, + 0.061597321182489395 + ], + "credit_map": [ + 0.02048996090888977, + 0.026994971558451653, + 0.03236555680632591, + 0.015774384140968323, + 0.024440791457891464, + 0.029148248955607414, + 0.022533737123012543, + 0.04945303499698639, + 0.012512505054473877, + 0.00982677936553955, + 0.027762481942772865, + 0.008590438403189182, + 0.010515722446143627, + 0.013955038040876389, + 0.013672414235770702, + 0.01749921403825283, + 0.009921116754412651, + 0.00590881984680891, + 0.008772151544690132, + 0.005266269203275442, + 0.006932777352631092, + 0.009427924640476704, + 0.006734434515237808, + 0.01051624771207571, + 0.002727044280618429, + 0.005688476841896772, + 0.003723267000168562, + 0.002862220862880349, + 0.002324535045772791, + 0.003554749535396695, + 0.002531006233766675, + 0.002920708851888776, + 0.006410202477127314, + 0.011811475269496441, + 0.006979490164667368, + 0.006197699345648289, + 0.024205299094319344, + 0.01853381097316742, + 0.006441353354603052, + 0.008659414015710354, + 0.003258773824200034, + 0.006082217674702406, + 0.004569295793771744, + 0.0032122491393238306, + 0.006756752263754606, + 0.004154372029006481, + 0.003363900352269411, + 0.004590804688632488, + 0.0010596956126391888, + 0.0015922154998406768, + 0.0013741840375587344, + 0.0015714950859546661, + 0.001522442908026278, + 0.0011548929614946246, + 0.001160253887064755, + 0.0013648418243974447, + 0.007638224400579929, + 0.008408110588788986, + 0.007891541346907616, + 0.014980441890656948, + 0.011597731150686741, + 0.009846913628280163, + 0.006648671813309193, + 0.006499579641968012, + 0.018736649304628372, + 0.11604408174753189, + 0.08455585688352585, + 0.014038479886949062, + 0.037396758794784546, + 0.01516627799719572, + 0.008681771345436573, + 0.008897338062524796, + 0.019060920923948288, + 0.00402078265324235, + 0.01785423420369625, + 0.005377511493861675, + 0.004126362968236208, + 0.0024727177806198597, + 0.004847472533583641, + 0.0038373165298253298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.008738137781620026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024951535742729902 + ], + "opposite_logprob_tokens": [ + -6.002495288848877 + ], + "logprob_reward": -0.0024951535742729902, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024951535742729902, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10532169044017792, + 0.09997982531785965, + 0.0787447988986969, + 0.06688573956489563, + 0.031631242483854294, + 0.025550995022058487, + 0.04747654125094414, + 0.059649281203746796, + 0.27866727113723755, + 0.20609256625175476 + ], + "credit_map": [ + 0.00902512576431036, + 0.012619421817362309, + 0.011962275952100754, + 0.010124160908162594, + 0.01629640720784664, + 0.01403326727449894, + 0.011149355210363865, + 0.0201116856187582, + 0.011468975804746151, + 0.008667852729558945, + 0.02447020821273327, + 0.007109513971954584, + 0.008447270840406418, + 0.011740935035049915, + 0.012105943635106087, + 0.015969138592481613, + 0.009307917207479477, + 0.01026294194161892, + 0.008008211851119995, + 0.008081316016614437, + 0.00974222645163536, + 0.010817194357514381, + 0.009995326399803162, + 0.012529673986136913, + 0.007694776635617018, + 0.009601249359548092, + 0.011776653118431568, + 0.006372024770826101, + 0.007553607691079378, + 0.009481831453740597, + 0.006616135127842426, + 0.007789472118020058, + 0.00320296548306942, + 0.0041721477173268795, + 0.0029797686729580164, + 0.002653944306075573, + 0.005944403354078531, + 0.005149682518094778, + 0.0027530291117727757, + 0.004775304347276688, + 0.0029693699907511473, + 0.00583586934953928, + 0.001996599370613694, + 0.0022542071528732777, + 0.005351932719349861, + 0.0022113581653684378, + 0.002487597754225135, + 0.002444064011797309, + 0.004332555923610926, + 0.007020363584160805, + 0.005100776441395283, + 0.005945295561105013, + 0.009451393969357014, + 0.005142472218722105, + 0.0061641596257686615, + 0.004319531377404928, + 0.006387160625308752, + 0.007489883806556463, + 0.005606777034699917, + 0.010194254107773304, + 0.010677602142095566, + 0.00944208912551403, + 0.005843435879796743, + 0.004008082207292318, + 0.02002483606338501, + 0.10664714872837067, + 0.061219871044158936, + 0.015290874056518078, + 0.039499372243881226, + 0.014856074936687946, + 0.013129763305187225, + 0.00799938291311264, + 0.05143977701663971, + 0.021027717739343643, + 0.019530044868588448, + 0.014469489455223083, + 0.04700499027967453, + 0.021954819560050964, + 0.013727033510804176, + 0.016938718035817146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.0024951535742729902 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018180865794420242 + ], + "opposite_logprob_tokens": [ + -4.018180847167969 + ], + "logprob_reward": -0.018180865794420242, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018180865794420242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4605677127838135, + 0.15969660878181458, + 0.05087505653500557, + 0.028901010751724243, + 0.057110317051410675, + 0.028068428859114647, + 0.02416745387017727, + 0.056617461144924164, + 0.07703430950641632, + 0.056961655616760254 + ], + "credit_map": [ + 0.033343683928251266, + 0.03347771242260933, + 0.056013140827417374, + 0.026239212602376938, + 0.05726217105984688, + 0.07615566998720169, + 0.06280063092708588, + 0.11527550220489502, + 0.01645106077194214, + 0.010246288031339645, + 0.03703797236084938, + 0.010893508791923523, + 0.01140753086656332, + 0.01965741440653801, + 0.02339591272175312, + 0.030606912449002266, + 0.006407071370631456, + 0.004848229233175516, + 0.005919077433645725, + 0.003663500538095832, + 0.0057053775526583195, + 0.0071168397553265095, + 0.007018766365945339, + 0.010196193121373653, + 0.0028542003128677607, + 0.003950560465455055, + 0.005131429061293602, + 0.002755069639533758, + 0.0031332839280366898, + 0.004371798597276211, + 0.003243454732000828, + 0.0034612128511071205, + 0.0032786799129098654, + 0.007935059256851673, + 0.002654916839674115, + 0.0025314809754490852, + 0.018938254565000534, + 0.015590127557516098, + 0.002713252557441592, + 0.0034685463178902864, + 0.0028854459524154663, + 0.004106272477656603, + 0.00283409864641726, + 0.002664354629814625, + 0.004917583893984556, + 0.003431235207244754, + 0.003094607498496771, + 0.004134829621762037, + 0.002401640173047781, + 0.0033443677239120007, + 0.0031004708725959063, + 0.0028831851668655872, + 0.003903347533196211, + 0.00264050904661417, + 0.0030032387003302574, + 0.0028906946536153555, + 0.005703320261090994, + 0.007877593860030174, + 0.005197563208639622, + 0.007639668881893158, + 0.010041284374892712, + 0.010206994600594044, + 0.005194545723497868, + 0.004756491631269455, + 0.005422284360975027, + 0.026435868814587593, + 0.018878420814871788, + 0.005088691599667072, + 0.0092160077765584, + 0.00540638342499733, + 0.0036748270504176617, + 0.002911824034526944, + 0.015908578410744667, + 0.006036198232322931, + 0.004275549668818712, + 0.0030562416650354862, + 0.014723291620612144, + 0.0038838665932416916, + 0.005458992440253496, + 0.0036189353559166193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.018180865794420242 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787560924887657 + ], + "opposite_logprob_tokens": [ + -3.529787540435791 + ], + "logprob_reward": -0.029787560924887657, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787560924887657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1614522635936737, + 0.05682950094342232, + 0.09038042277097702, + 0.032385583966970444, + 0.07191100716590881, + 0.044870346784591675, + 0.026607541367411613, + 0.0959620401263237, + 0.33574187755584717, + 0.08385945111513138 + ], + "credit_map": [ + 0.015726890414953232, + 0.020768344402313232, + 0.01996419206261635, + 0.01296983566135168, + 0.018689271062612534, + 0.02161053754389286, + 0.01671331375837326, + 0.03500987961888313, + 0.006820351351052523, + 0.00973794236779213, + 0.008267072029411793, + 0.006720850709825754, + 0.006885777693241835, + 0.007654694374650717, + 0.00479918671771884, + 0.005943621974438429, + 0.009887042455375195, + 0.017986604943871498, + 0.009264453314244747, + 0.00866105780005455, + 0.017022810876369476, + 0.01349756307899952, + 0.006618389394134283, + 0.007442501839250326, + 0.0030617055017501116, + 0.005038475152105093, + 0.002997490344569087, + 0.003259564284235239, + 0.004839859902858734, + 0.004197277594357729, + 0.004856079816818237, + 0.004135127644985914, + 0.007458149455487728, + 0.009955928660929203, + 0.006073444150388241, + 0.007494787219911814, + 0.011880808509886265, + 0.0120362788438797, + 0.00865630991756916, + 0.008355303667485714, + 0.005351129453629255, + 0.01039805356413126, + 0.00378053798340261, + 0.004548124969005585, + 0.007137604057788849, + 0.0047172559425234795, + 0.0032036222983151674, + 0.005734018981456757, + 0.00257323682308197, + 0.004312735516577959, + 0.0028998644556850195, + 0.003898847382515669, + 0.004298132378607988, + 0.002806424628943205, + 0.0029007329139858484, + 0.0029175670351833105, + 0.009160393849015236, + 0.01450591254979372, + 0.01077186968177557, + 0.013924729079008102, + 0.018727190792560577, + 0.011076899245381355, + 0.00976539496332407, + 0.008029651828110218, + 0.02163516730070114, + 0.12724444270133972, + 0.09035130590200424, + 0.015948643907904625, + 0.04095624014735222, + 0.016117461025714874, + 0.010329211130738258, + 0.013159393332898617, + 0.024754302576184273, + 0.005493710748851299, + 0.024668002501130104, + 0.006030986085534096, + 0.006041622254997492, + 0.006826255936175585, + 0.005685201846063137, + 0.004359368700534105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.029787560924887657 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set underwater in a deep, dark ocean environment?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive blue whale visible with velvety midnight blue skin and pale azure underbelly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dense school of hundreds of jellyfish emitting pink, electric blue, and pastel green bioluminescence visible ahead of the whale?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the whale's tail fin sway in a slow rhythm while the jellyfish part and swirl around it without scattering?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot wide-angle, filmed from a slightly upward underwater angle, and panning slowly to the right?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04303964972496033 + ], + "opposite_logprob_tokens": [ + -3.168039560317993 + ], + "logprob_reward": -0.04303964972496033, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04303964972496033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10344566404819489, + 0.047530315816402435, + 0.03502565994858742, + 0.05920937657356262, + 0.05310565233230591, + 0.032726891338825226, + 0.03273404389619827, + 0.13430553674697876, + 0.4105543792247772, + 0.09136246889829636 + ], + "credit_map": [ + 0.010540462099015713, + 0.010446353815495968, + 0.013289342634379864, + 0.007773590739816427, + 0.011659209616482258, + 0.014272050932049751, + 0.011192390695214272, + 0.024272259324789047, + 0.006005209870636463, + 0.006247266195714474, + 0.010125800967216492, + 0.003006682265549898, + 0.004885378759354353, + 0.006114024203270674, + 0.0049300417304039, + 0.006215913221240044, + 0.003989365417510271, + 0.004382984712719917, + 0.004013444297015667, + 0.002981595927849412, + 0.006310359109193087, + 0.005220931954681873, + 0.0038632701616734266, + 0.004263709299266338, + 0.006275458261370659, + 0.009542018175125122, + 0.007886090315878391, + 0.006093838717788458, + 0.005875278729945421, + 0.008429168723523617, + 0.006690661422908306, + 0.008416865020990372, + 0.005188065581023693, + 0.009755515493452549, + 0.005278467200696468, + 0.006331580691039562, + 0.0066780392080545425, + 0.007302476093173027, + 0.0051497891545295715, + 0.00742172310128808, + 0.0039037158712744713, + 0.00492452597245574, + 0.002955225994810462, + 0.0034993442241102457, + 0.005641436669975519, + 0.004500662907958031, + 0.002696207957342267, + 0.0046057733707129955, + 0.0027951630763709545, + 0.004768225364387035, + 0.003426855430006981, + 0.004107850603759289, + 0.006373479031026363, + 0.0033847703598439693, + 0.003499558661133051, + 0.00437813950702548, + 0.01391120720654726, + 0.014733581803739071, + 0.017531057819724083, + 0.02761857584118843, + 0.019104324281215668, + 0.01719801500439644, + 0.01263403706252575, + 0.011574744246900082, + 0.025703778490424156, + 0.1429426372051239, + 0.1187627837061882, + 0.02651257999241352, + 0.042412761598825455, + 0.019531646743416786, + 0.015912558883428574, + 0.018775638192892075, + 0.016186436638236046, + 0.010336739011108875, + 0.0259891115128994, + 0.005749691277742386, + 0.01672566495835781, + 0.006169163156300783, + 0.005924862809479237, + 0.004280800931155682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.04303964972496033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_tensor([6204, 6204])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_tensor([6204, 6204])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1be72807d8429f5b073d1198a6689a6c88b933a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_tensor([6204, 6204])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "handheld medium close-up shot from a slightly upward angle, the camera whips back violently, following the subject's movements, finally framing the face of an East Asian woman, her face pale with fear. Her long, damp, black hair, a few strands tangled and plastered to her pale cheeks and forehead. Her eyes are wide open with utter terror, pupils constricted, her mouth drawn into a silent, screaming O, her jaw clenched with tension, as if she's exerting all her strength but unable to utter a single sound. She's in a narrow alley, squeezed oppressively by high, crumbling walls on either side. The only source of light is a flickering, poorly connected ceiling light. Each flicker casts a twitching, rapidly alternating light and shadow on her distorted features and the damp brick wall behind her, reflecting this sickly light off the damp ground. She's wearing a dark gray hoodie, and her entire body freezes as she suddenly turns back, her shoulders clenched, creating a suffocating atmosphere of intense tension and suspense, as if time has frozen in time, leaving only her silent terror.", + "seed": 95, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_6204_seed_95_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0592898559756577, + "vlm_logprob": -0.0592898559756577, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.0592898559756577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11253297328948975 + ], + "opposite_logprob_tokens": [ + -3.4875330924987793 + ], + "logprob_reward": -0.11253297328948975, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.11253297328948975, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12195240706205368, + 0.10569903254508972, + 0.15338659286499023, + 0.05215638130903244, + 0.05388545244932175, + 0.27578121423721313, + 0.057456955313682556, + 0.05806855857372284, + 0.05141915753483772, + 0.07019416987895966 + ], + "credit_map": [ + 0.01853020489215851, + 0.008384013548493385, + 0.013442211784422398, + 0.01730496995151043, + 0.014054493978619576, + 0.016522977501153946, + 0.01812274381518364, + 0.015590794384479523, + 0.009297197684645653, + 0.01339649222791195, + 0.010174705646932125, + 0.02167714387178421, + 0.010084672830998898, + 0.009870056062936783, + 0.016836611554026604, + 0.014362151734530926, + 0.009833094663918018, + 0.019672591239213943, + 0.016026439145207405, + 0.03662774711847305, + 0.026921598240733147, + 0.008382193744182587, + 0.02680308185517788, + 0.009119854308664799, + 0.0035229744389653206, + 0.004690834321081638, + 0.005223474465310574, + 0.009359634481370449, + 0.003987574949860573, + 0.011562583968043327, + 0.004595225211232901, + 0.00921408087015152, + 0.006297072395682335, + 0.005338076502084732, + 0.00994005799293518, + 0.008614207617938519, + 0.005021088756620884, + 0.007226572837680578, + 0.0068505629897117615, + 0.004597814287990332, + 0.027092168107628822, + 0.0059179943054914474, + 0.1002454161643982, + 0.014987480826675892, + 0.004177364986389875, + 0.02002105489373207, + 0.009671725332736969, + 0.09366800636053085, + 0.007508744020015001, + 0.0057181986048817635, + 0.0076333098113536835, + 0.008564908057451248, + 0.004435110371559858, + 0.008622197434306145, + 0.007023625075817108, + 0.007950861938297749, + 0.008166537620127201, + 0.006452958099544048, + 0.01000839751213789, + 0.005781387910246849, + 0.005665873177349567, + 0.005524428561329842, + 0.01114483829587698, + 0.005324138794094324, + 0.005770781543105841, + 0.0033471188507974148, + 0.006994988303631544, + 0.004120719153434038, + 0.005860261153429747, + 0.008991437032818794, + 0.00867210142314434, + 0.0076617528684437275, + 0.013836626894772053, + 0.0036981620360165834, + 0.01093695405870676, + 0.004414898343384266, + 0.008245406672358513, + 0.01127303671091795, + 0.010898429900407791, + 0.006890660151839256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.11253297328948975 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038439929485321 + ], + "opposite_logprob_tokens": [ + -2.350384473800659 + ], + "logprob_reward": -0.10038439929485321, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038439929485321, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07377323508262634, + 0.0420796163380146, + 0.07040099054574966, + 0.05583294481039047, + 0.01310187578201294, + 0.612922728061676, + 0.01352720707654953, + 0.04198122397065163, + 0.008989324793219566, + 0.06739083677530289 + ], + "credit_map": [ + 0.0059994906187057495, + 0.006825594697147608, + 0.006796351633965969, + 0.011897539719939232, + 0.012131120078265667, + 0.006334053818136454, + 0.015711918473243713, + 0.008077165111899376, + 0.0037545585073530674, + 0.0042317030020058155, + 0.0054244534112513065, + 0.0077476478181779385, + 0.00437560211867094, + 0.006683625280857086, + 0.004703231155872345, + 0.005158799234777689, + 0.003992489073425531, + 0.008203377015888691, + 0.006933462340384722, + 0.016724755987524986, + 0.01319370698183775, + 0.004026781301945448, + 0.013199886307120323, + 0.0041265347972512245, + 0.004905387759208679, + 0.0056106215342879295, + 0.004536891356110573, + 0.010051421821117401, + 0.002472967142239213, + 0.013377261348068714, + 0.0040337517857551575, + 0.010844642296433449, + 0.0012830342166125774, + 0.0016445734072476625, + 0.002226336160674691, + 0.0027763741090893745, + 0.0013743682065978646, + 0.0013396089198067784, + 0.0013059116899967194, + 0.0011516694212332368, + 0.060784678906202316, + 0.013232437893748283, + 0.22537924349308014, + 0.03264440596103668, + 0.009676745161414146, + 0.04267125204205513, + 0.020436599850654602, + 0.20809733867645264, + 0.0012295001652091742, + 0.0014750385889783502, + 0.0016774098621681333, + 0.0014020326780155301, + 0.0015971900429576635, + 0.002704082289710641, + 0.0014863767428323627, + 0.001955576939508319, + 0.00563321728259325, + 0.0024452682118862867, + 0.005731963086873293, + 0.002380362246185541, + 0.006868390366435051, + 0.0031534943263977766, + 0.012292629107832909, + 0.003475899575278163, + 0.000890067545697093, + 0.0006653487798757851, + 0.0011332548456266522, + 0.0008056452497839928, + 0.0013217099476605654, + 0.00107957620639354, + 0.002056621480733156, + 0.0010371002135798335, + 0.006603785790503025, + 0.002765767741948366, + 0.010531298816204071, + 0.0039011589251458645, + 0.010627438314259052, + 0.016100851818919182, + 0.009535948745906353, + 0.007324583828449249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.10038439929485321 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067501612938940525 + ], + "opposite_logprob_tokens": [ + -5.006750106811523 + ], + "logprob_reward": -0.0067501612938940525, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067501612938940525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07759988307952881, + 0.11700570583343506, + 0.023221494629979134, + 0.2458019107580185, + 0.17494551837444305, + 0.06324224174022675, + 0.03768949583172798, + 0.03587988391518593, + 0.04470347985625267, + 0.179910346865654 + ], + "credit_map": [ + 0.010585798881947994, + 0.00585668720304966, + 0.007748029660433531, + 0.008812552317976952, + 0.007421555928885937, + 0.0154016874730587, + 0.00990265142172575, + 0.011870923452079296, + 0.009093944914638996, + 0.013262797147035599, + 0.015174681320786476, + 0.02564200572669506, + 0.012285860255360603, + 0.012027395889163017, + 0.016916753724217415, + 0.012602269649505615, + 0.0017896619392558932, + 0.003027267986908555, + 0.0022093215957283974, + 0.003976965788751841, + 0.002807869575917721, + 0.0015971171669661999, + 0.003288882551714778, + 0.004524406045675278, + 0.01686021126806736, + 0.0250917449593544, + 0.01976240612566471, + 0.028557220473885536, + 0.014338796958327293, + 0.06434046477079391, + 0.019570033997297287, + 0.05728102847933769, + 0.024214208126068115, + 0.01627640798687935, + 0.04473578929901123, + 0.027957994490861893, + 0.008369295857846737, + 0.02689896896481514, + 0.016322489827871323, + 0.010170373134315014, + 0.0062018632888793945, + 0.0017145925667136908, + 0.021741842851042747, + 0.003948120400309563, + 0.0014749013353139162, + 0.004799732472747564, + 0.002713921945542097, + 0.020647265017032623, + 0.005280676763504744, + 0.0026353937573730946, + 0.006045905407518148, + 0.0031169666908681393, + 0.004452167544513941, + 0.008174443617463112, + 0.003386928467079997, + 0.004597012884914875, + 0.004031924065202475, + 0.0029325908981263638, + 0.00570973614230752, + 0.004043930675834417, + 0.004566488321870565, + 0.003217701567336917, + 0.008169925771653652, + 0.0032075862400233746, + 0.0050441124476492405, + 0.002733831061050296, + 0.006753216963261366, + 0.004614740144461393, + 0.006029525771737099, + 0.005423219874501228, + 0.009595240466296673, + 0.004509592894464731, + 0.03355657309293747, + 0.013960327953100204, + 0.028740813955664635, + 0.010333950631320477, + 0.018152538686990738, + 0.018408605828881264, + 0.04072291776537895, + 0.016034606844186783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 4 + ], + "selected_reward": -0.0067501612938940525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067444778978824615 + ], + "opposite_logprob_tokens": [ + -5.006744384765625 + ], + "logprob_reward": -0.0067444778978824615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067444778978824615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15097281336784363, + 0.18229961395263672, + 0.15022575855255127, + 0.09674783796072006, + 0.08293531090021133, + 0.04500925913453102, + 0.11053454130887985, + 0.057703062891960144, + 0.03356526419520378, + 0.09000654518604279 + ], + "credit_map": [ + 0.016414204612374306, + 0.01983422413468361, + 0.015522299334406853, + 0.023968050256371498, + 0.01737295649945736, + 0.01680091582238674, + 0.027102146297693253, + 0.013958006165921688, + 0.014184962958097458, + 0.017433729022741318, + 0.01837354339659214, + 0.037924833595752716, + 0.0190921351313591, + 0.029665067791938782, + 0.02568822354078293, + 0.019937124103307724, + 0.006893958430737257, + 0.0160970538854599, + 0.013537885621190071, + 0.030968239530920982, + 0.03455740213394165, + 0.008824247866868973, + 0.026210607960820198, + 0.013136358000338078, + 0.007743932772427797, + 0.006541465874761343, + 0.007857821881771088, + 0.009436112828552723, + 0.011471851728856564, + 0.026065440848469734, + 0.006891404744237661, + 0.020739808678627014, + 0.011247620917856693, + 0.004322742577642202, + 0.02424272894859314, + 0.006307818926870823, + 0.0071223434060812, + 0.014102159067988396, + 0.008525047451257706, + 0.0070648398250341415, + 0.00911199115216732, + 0.0022106643300503492, + 0.007080406881868839, + 0.004601934924721718, + 0.002172573236748576, + 0.008014305494725704, + 0.0047773090191185474, + 0.007040073163807392, + 0.013651265762746334, + 0.006886592600494623, + 0.013598854653537273, + 0.010598424822092056, + 0.010424645617604256, + 0.02866242453455925, + 0.010725164785981178, + 0.015987172722816467, + 0.009641140699386597, + 0.0037843554746359587, + 0.00960933230817318, + 0.0048886071890592575, + 0.006447352934628725, + 0.006882513407617807, + 0.010166109539568424, + 0.006283652037382126, + 0.003228971501812339, + 0.002086916007101536, + 0.004598014522343874, + 0.003367304801940918, + 0.0038991516921669245, + 0.005103607662022114, + 0.007017393596470356, + 0.004263903945684433, + 0.011095268651843071, + 0.004499009344726801, + 0.013706427998840809, + 0.005785036366432905, + 0.011488948948681355, + 0.01971415989100933, + 0.013319290243089199, + 0.010398403741419315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.0067444778978824615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003726810216904 + ], + "opposite_logprob_tokens": [ + -2.695037364959717 + ], + "logprob_reward": -0.07003726810216904, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003726810216904, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08149165660142899, + 0.06320498883724213, + 0.07278148829936981, + 0.09804045408964157, + 0.18336662650108337, + 0.27634719014167786, + 0.02945864014327526, + 0.040348269045352936, + 0.04401560127735138, + 0.11094511300325394 + ], + "credit_map": [ + 0.01268141157925129, + 0.0071227471344172955, + 0.009625299833714962, + 0.009321081452071667, + 0.012775078415870667, + 0.010069659911096096, + 0.011415814980864525, + 0.00848056748509407, + 0.006983309984207153, + 0.010839312337338924, + 0.00782663282006979, + 0.01223619282245636, + 0.004744409117847681, + 0.005940856412053108, + 0.007366918958723545, + 0.0072673531249165535, + 0.004869073163717985, + 0.009357698261737823, + 0.008404738269746304, + 0.016816111281514168, + 0.011631043627858162, + 0.004206052049994469, + 0.01151561550796032, + 0.005981155205518007, + 0.0096375597640872, + 0.006484590470790863, + 0.007542741484940052, + 0.01794939674437046, + 0.005462913773953915, + 0.025042925029993057, + 0.007059402298182249, + 0.018860923126339912, + 0.023723885416984558, + 0.008993919938802719, + 0.056754231452941895, + 0.01695379801094532, + 0.008392563089728355, + 0.04237374663352966, + 0.016194667667150497, + 0.009979824535548687, + 0.02773120254278183, + 0.005945340730249882, + 0.09953074157238007, + 0.015487407334148884, + 0.004441104829311371, + 0.01945328153669834, + 0.009705782867968082, + 0.09405231475830078, + 0.004353851545602083, + 0.0018753391923382878, + 0.004310297314077616, + 0.003020418109372258, + 0.0024347142316401005, + 0.0046593742445111275, + 0.0037702741101384163, + 0.0050343722105026245, + 0.009099678136408329, + 0.002968869172036648, + 0.006796208210289478, + 0.003237241180613637, + 0.003732769750058651, + 0.004156819079071283, + 0.006414898205548525, + 0.003941785544157028, + 0.004537604749202728, + 0.003138309810310602, + 0.005252387374639511, + 0.004136743023991585, + 0.007859665900468826, + 0.004283391870558262, + 0.010052992962300777, + 0.0047545055858790874, + 0.010394958779215813, + 0.006280712317675352, + 0.015863370150327682, + 0.007585577666759491, + 0.016993554309010506, + 0.018121831119060516, + 0.023134147748351097, + 0.012570962309837341 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 9 + ], + "selected_reward": -0.07003726810216904 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld medium close-up shot from a slightly upward angle, the camera whips back violently, following the subject's movements, finally framing the face of an East Asian woman, her face pale with fear. Her long, damp, black hair, a few strands tangled and plastered to her pale cheeks and forehead. Her eyes are wide open with utter terror, pupils constricted, her mouth drawn into a silent, screaming O, her jaw clenched with tension, as if she's exerting all her strength but unable to utter a single sound. She's in a narrow alley, squeezed oppressively by high, crumbling walls on either side. The only source of light is a flickering, poorly connected ceiling light. Each flicker casts a twitching, rapidly alternating light and shadow on her distorted features and the damp brick wall behind her, reflecting this sickly light off the damp ground. She's wearing a dark gray hoodie, and her entire body freezes as she suddenly turns back, her shoulders clenched, creating a suffocating atmosphere of intense tension and suspense, as if time has frozen in time, leaving only her silent terror.", + "seed": 96, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_6204_seed_96_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09398513860069216, + "vlm_logprob": -0.09398513860069216, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.09398513860069216, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12399922311306 + ], + "opposite_logprob_tokens": [ + -2.9989991188049316 + ], + "logprob_reward": -0.12399922311306, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.12399922311306, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08861195296049118, + 0.18977473676204681, + 0.06084507331252098, + 0.04483223706483841, + 0.05447591841220856, + 0.163956880569458, + 0.08938047289848328, + 0.1985619068145752, + 0.05820790305733681, + 0.05135291442275047 + ], + "credit_map": [ + 0.005403126124292612, + 0.005539200734347105, + 0.010992113500833511, + 0.01186264120042324, + 0.014215742237865925, + 0.013756399974226952, + 0.014536784961819649, + 0.012305940501391888, + 0.018617598339915276, + 0.006036400329321623, + 0.01352296955883503, + 0.013525028713047504, + 0.013735591433942318, + 0.046568386256694794, + 0.06655523180961609, + 0.011213536374270916, + 0.008160642348229885, + 0.004393896088004112, + 0.0064093065448105335, + 0.006788432132452726, + 0.011302828788757324, + 0.007402726449072361, + 0.010403846390545368, + 0.0059833950363099575, + 0.004951138515025377, + 0.003070379141718149, + 0.007659115828573704, + 0.004541903268545866, + 0.005937391892075539, + 0.006155672017484903, + 0.00628791656345129, + 0.0062287189066410065, + 0.00734972907230258, + 0.003556976793333888, + 0.007359200622886419, + 0.00432802876457572, + 0.007329954300075769, + 0.009251149371266365, + 0.006827528588473797, + 0.008473348803818226, + 0.020783303305506706, + 0.02408309280872345, + 0.025500567629933357, + 0.017649825662374496, + 0.018787559121847153, + 0.022110510617494583, + 0.01650814712047577, + 0.018533864989876747, + 0.0138962147757411, + 0.00458644749596715, + 0.009733770042657852, + 0.016510020941495895, + 0.0068543702363967896, + 0.022352326661348343, + 0.007285850588232279, + 0.008161473087966442, + 0.015542054548859596, + 0.009042874909937382, + 0.06494584679603577, + 0.020973239094018936, + 0.009910613298416138, + 0.060828469693660736, + 0.00853653158992529, + 0.008782287128269672, + 0.0058360472321510315, + 0.004158809781074524, + 0.013274072669446468, + 0.007841266691684723, + 0.0061456612311303616, + 0.008669452741742134, + 0.005232555326074362, + 0.007050040643662214, + 0.005141218658536673, + 0.0035508559085428715, + 0.0046904925256967545, + 0.0040184869430959225, + 0.0055006579495966434, + 0.012197470292448997, + 0.0059290602803230286, + 0.010324670001864433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.12399922311306 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.225677490234375 + ], + "opposite_logprob_tokens": [ + -1.600677490234375 + ], + "logprob_reward": -0.225677490234375, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.225677490234375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12043912708759308, + 0.45656314492225647, + 0.11833921074867249, + 0.03819975629448891, + 0.05009719729423523, + 0.08569002896547318, + 0.023484906181693077, + 0.04610356315970421, + 0.031396038830280304, + 0.02968708425760269 + ], + "credit_map": [ + 0.012085244059562683, + 0.005649327766150236, + 0.020126348361372948, + 0.011047828011214733, + 0.024579087272286415, + 0.019529210403561592, + 0.012662270106375217, + 0.014759812504053116, + 0.04690515622496605, + 0.016476791352033615, + 0.028528064489364624, + 0.03309663385152817, + 0.03320702910423279, + 0.10985159128904343, + 0.16270257532596588, + 0.02579532004892826, + 0.010657070204615593, + 0.007409021258354187, + 0.015491434372961521, + 0.014905803836882114, + 0.019761117175221443, + 0.013457447290420532, + 0.024575520306825638, + 0.012081802822649479, + 0.005497091915458441, + 0.003158000996336341, + 0.004971527494490147, + 0.0038801373448222876, + 0.0065382798202335835, + 0.0036089906934648752, + 0.005717989522963762, + 0.004827740136533976, + 0.004529376979917288, + 0.0029963774140924215, + 0.005769095383584499, + 0.007035226095467806, + 0.010681268759071827, + 0.005685616284608841, + 0.006089506205171347, + 0.007310733199119568, + 0.010706573724746704, + 0.011307356879115105, + 0.011667273938655853, + 0.008841380476951599, + 0.010531878098845482, + 0.014155423268675804, + 0.008956464007496834, + 0.009523680433630943, + 0.0029616064857691526, + 0.0019512856379151344, + 0.003410888835787773, + 0.003724922426044941, + 0.0028756391257047653, + 0.004037045408040285, + 0.0024026294704526663, + 0.0021208885591477156, + 0.0033643320202827454, + 0.00231537246145308, + 0.006638692691922188, + 0.005913515109568834, + 0.0049267360009253025, + 0.015222795307636261, + 0.004751382861286402, + 0.002970735542476177, + 0.004898850340396166, + 0.002770721912384033, + 0.0034966953098773956, + 0.005175705999135971, + 0.004526230040937662, + 0.0027983877807855606, + 0.003722775960341096, + 0.004006669856607914, + 0.0027161103207618, + 0.0025274762883782387, + 0.0036182596813887358, + 0.003787312190979719, + 0.004469358827918768, + 0.0060274493880569935, + 0.003276657545939088, + 0.0032644595485180616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.225677490234375 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00320504535920918 + ], + "opposite_logprob_tokens": [ + -5.753204822540283 + ], + "logprob_reward": -0.00320504535920918, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.00320504535920918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15117718279361725, + 0.236888125538826, + 0.08980293571949005, + 0.032607126981019974, + 0.13218529522418976, + 0.04695644974708557, + 0.1156616359949112, + 0.07234315574169159, + 0.08098148554563522, + 0.041396614164114 + ], + "credit_map": [ + 0.011200882494449615, + 0.009091774933040142, + 0.020637771114706993, + 0.015767553821206093, + 0.020999668166041374, + 0.0230095237493515, + 0.01732797361910343, + 0.03314204141497612, + 0.026056688278913498, + 0.011012248694896698, + 0.017797324806451797, + 0.016060898080468178, + 0.01735336147248745, + 0.0528247132897377, + 0.07663875073194504, + 0.01914413832128048, + 0.007962473668158054, + 0.008435668423771858, + 0.012068500742316246, + 0.010214881040155888, + 0.015824805945158005, + 0.01114480197429657, + 0.014448956586420536, + 0.009702840819954872, + 0.003016301430761814, + 0.003149236086755991, + 0.004023306537419558, + 0.006210387218743563, + 0.004017202649265528, + 0.0031500759068876505, + 0.004192569758743048, + 0.004848048090934753, + 0.01403867732733488, + 0.008358423598110676, + 0.016510050743818283, + 0.016464367508888245, + 0.02193206176161766, + 0.01656978949904442, + 0.018529394641518593, + 0.01978253200650215, + 0.005357487127184868, + 0.003926156088709831, + 0.005478831008076668, + 0.005190900526940823, + 0.009218751452863216, + 0.004720696248114109, + 0.005367549601942301, + 0.00769607350230217, + 0.014379948377609253, + 0.006032947450876236, + 0.013704473152756691, + 0.023683225736021996, + 0.011984073556959629, + 0.02761959843337536, + 0.009367907419800758, + 0.00888946745544672, + 0.005316498223692179, + 0.005474809557199478, + 0.013050046749413013, + 0.012801161967217922, + 0.00882011093199253, + 0.008459595963358879, + 0.01212535984814167, + 0.00629557017236948, + 0.008076153695583344, + 0.006310479715466499, + 0.014954384416341782, + 0.013457507826387882, + 0.014580258168280125, + 0.005837025120854378, + 0.011226769536733627, + 0.0065389047376811504, + 0.0035857870243489742, + 0.0024531898088753223, + 0.007413323037326336, + 0.003965198993682861, + 0.0049743144772946835, + 0.008704781532287598, + 0.0058946143835783005, + 0.004405402112752199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.00320504535920918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0041007245890796185 + ], + "opposite_logprob_tokens": [ + -5.504100799560547 + ], + "logprob_reward": -0.0041007245890796185, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.0041007245890796185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05515559762716293, + 0.3774131238460541, + 0.15762141346931458, + 0.04169681295752525, + 0.03510395064949989, + 0.014473323710262775, + 0.03984124958515167, + 0.18933288753032684, + 0.03600805625319481, + 0.05335363745689392 + ], + "credit_map": [ + 0.004542229697108269, + 0.004366133827716112, + 0.004964192397892475, + 0.006519090849906206, + 0.00671087671071291, + 0.008798450231552124, + 0.006107699126005173, + 0.013146919198334217, + 0.0404881052672863, + 0.013218767009675503, + 0.024108702316880226, + 0.025149190798401833, + 0.02801680937409401, + 0.09006111323833466, + 0.13264867663383484, + 0.02372172847390175, + 0.02264690399169922, + 0.013910946436226368, + 0.037052400410175323, + 0.017022235319018364, + 0.018478794023394585, + 0.017867732793092728, + 0.01583288051187992, + 0.014809499494731426, + 0.005048043094575405, + 0.003182078478857875, + 0.006471318192780018, + 0.00513150030747056, + 0.005399439483880997, + 0.00563815888017416, + 0.005194641649723053, + 0.005631630774587393, + 0.0035980690736323595, + 0.0022696552332490683, + 0.004658447112888098, + 0.003143852576613426, + 0.006508217193186283, + 0.00411995267495513, + 0.005105611402541399, + 0.00570014538243413, + 0.0017278913874179125, + 0.0011769705452024937, + 0.0020566273014992476, + 0.0015718296635895967, + 0.0029117069207131863, + 0.00120827357750386, + 0.0020615251269191504, + 0.0017584989545866847, + 0.006339564453810453, + 0.0020630948711186647, + 0.0033973169047385454, + 0.00953578669577837, + 0.0023668897338211536, + 0.008110253140330315, + 0.003604417433962226, + 0.004423927050083876, + 0.014756755903363228, + 0.0065568722784519196, + 0.06285286694765091, + 0.020973172038793564, + 0.008763064630329609, + 0.06029141694307327, + 0.006186266429722309, + 0.008952444419264793, + 0.0037577981129288673, + 0.0020924427080899477, + 0.012153439223766327, + 0.008915782906115055, + 0.003019297495484352, + 0.0021442763973027468, + 0.002110143192112446, + 0.0018148724921047688, + 0.00675337016582489, + 0.003701315727084875, + 0.011232891120016575, + 0.005910111591219902, + 0.005933935754001141, + 0.009110328741371632, + 0.007038133218884468, + 0.0036735525354743004 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 2 + ], + "selected_reward": -0.0041007245890796185 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294320970773697 + ], + "opposite_logprob_tokens": [ + -2.237943172454834 + ], + "logprob_reward": -0.11294320970773697, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294320970773697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08209143579006195, + 0.14635510742664337, + 0.04834819212555885, + 0.026402831077575684, + 0.07406368106603622, + 0.044536467641592026, + 0.13389845192432404, + 0.34062328934669495, + 0.06582804024219513, + 0.03785249963402748 + ], + "credit_map": [ + 0.00959351658821106, + 0.0030369414016604424, + 0.016691070050001144, + 0.008015528321266174, + 0.006950119975954294, + 0.013218767940998077, + 0.01062952633947134, + 0.013955967500805855, + 0.014746678061783314, + 0.007624630816280842, + 0.008983040228486061, + 0.011896943673491478, + 0.011824562214314938, + 0.03334858641028404, + 0.048441387712955475, + 0.009489278309047222, + 0.004600859712809324, + 0.0033804448321461678, + 0.00567240035161376, + 0.004848930984735489, + 0.012812916189432144, + 0.006478353403508663, + 0.006944684311747551, + 0.003609603736549616, + 0.003084261203184724, + 0.0024638811592012644, + 0.002491893945261836, + 0.0030025639571249485, + 0.005004680249840021, + 0.0033984340261667967, + 0.004065750166773796, + 0.002891366370022297, + 0.005883654113858938, + 0.005315983202308416, + 0.009266084991395473, + 0.007450744975358248, + 0.014146532863378525, + 0.00877259112894535, + 0.013383644632995129, + 0.009844444692134857, + 0.003242152277380228, + 0.003201722400262952, + 0.004177352879196405, + 0.0048886751756072044, + 0.008388621732592583, + 0.006391576956957579, + 0.006918579339981079, + 0.007327787112444639, + 0.017148541286587715, + 0.007588355336338282, + 0.013538911938667297, + 0.02637953869998455, + 0.01735902763903141, + 0.0293943602591753, + 0.011494695208966732, + 0.010995023883879185, + 0.024438966065645218, + 0.013493476435542107, + 0.09866823256015778, + 0.03508439287543297, + 0.018389804288744926, + 0.11637590080499649, + 0.016981102526187897, + 0.01719142682850361, + 0.007356496062129736, + 0.0068499743938446045, + 0.0088801309466362, + 0.012715985998511314, + 0.010733472183346748, + 0.005040287505835295, + 0.009055268950760365, + 0.005196422804147005, + 0.005969036370515823, + 0.003557261312380433, + 0.008263486437499523, + 0.003679226152598858, + 0.004277701489627361, + 0.0034579355269670486, + 0.004935995675623417, + 0.0037118576001375914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 6 + ], + "selected_reward": -0.11294320970773697 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld medium close-up shot from a slightly upward angle, the camera whips back violently, following the subject's movements, finally framing the face of an East Asian woman, her face pale with fear. Her long, damp, black hair, a few strands tangled and plastered to her pale cheeks and forehead. Her eyes are wide open with utter terror, pupils constricted, her mouth drawn into a silent, screaming O, her jaw clenched with tension, as if she's exerting all her strength but unable to utter a single sound. She's in a narrow alley, squeezed oppressively by high, crumbling walls on either side. The only source of light is a flickering, poorly connected ceiling light. Each flicker casts a twitching, rapidly alternating light and shadow on her distorted features and the damp brick wall behind her, reflecting this sickly light off the damp ground. She's wearing a dark gray hoodie, and her entire body freezes as she suddenly turns back, her shoulders clenched, creating a suffocating atmosphere of intense tension and suspense, as if time has frozen in time, leaving only her silent terror.", + "seed": 97, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_6204_seed_97_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09161838374566286, + "vlm_logprob": -0.09161838374566286, + "vlm_yes_no_margin": 3.7749999880790712, + "vlm_token_credit": -0.09161838374566286, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06939007341861725 + ], + "opposite_logprob_tokens": [ + -3.194390058517456 + ], + "logprob_reward": -0.06939007341861725, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.06939007341861725, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1627821922302246, + 0.26341116428375244, + 0.05022629350423813, + 0.07373838126659393, + 0.06434042751789093, + 0.028521748259663582, + 0.1059999093413353, + 0.052537817507982254, + 0.037944894284009933, + 0.16049720346927643 + ], + "credit_map": [ + 0.02282807230949402, + 0.010794761590659618, + 0.020417548716068268, + 0.022385476157069206, + 0.01476532407104969, + 0.015345845371484756, + 0.01938106119632721, + 0.03686409443616867, + 0.013984344899654388, + 0.021585645154118538, + 0.007886216044425964, + 0.014157382771372795, + 0.10409675538539886, + 0.020631181076169014, + 0.04576091095805168, + 0.03530871868133545, + 0.005509135778993368, + 0.006676285061985254, + 0.005742480047047138, + 0.005258786957710981, + 0.008624010719358921, + 0.007051554508507252, + 0.004293013364076614, + 0.0070710256695747375, + 0.00720018707215786, + 0.004238576628267765, + 0.006147176027297974, + 0.005068262107670307, + 0.006293030455708504, + 0.017475785687565804, + 0.003789186477661133, + 0.02352616935968399, + 0.007021200377494097, + 0.00410460215061903, + 0.01692698709666729, + 0.006320907734334469, + 0.004567721392959356, + 0.004329662770032883, + 0.006534254644066095, + 0.014535089954733849, + 0.0044643511064350605, + 0.0025513593573123217, + 0.003787643276154995, + 0.002747035352513194, + 0.00345222488977015, + 0.0031198812648653984, + 0.006229640915989876, + 0.0021696118637919426, + 0.0163725595921278, + 0.006888119503855705, + 0.023183999583125114, + 0.013089722022414207, + 0.00874919444322586, + 0.013558032922446728, + 0.0076670655980706215, + 0.016491219401359558, + 0.008957319892942905, + 0.004294055979698896, + 0.0065987249836325645, + 0.006263916380703449, + 0.006522094365209341, + 0.006998175289481878, + 0.007377303671091795, + 0.005526221822947264, + 0.005056456662714481, + 0.002611057134345174, + 0.005862031597644091, + 0.004236310254782438, + 0.0040857186540961266, + 0.005387991201132536, + 0.007066016085445881, + 0.003639309899881482, + 0.031229836866259575, + 0.013230530545115471, + 0.02052282728254795, + 0.01980956643819809, + 0.01235019788146019, + 0.025205694139003754, + 0.021728213876485825, + 0.016420327126979828 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.06939007341861725 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31353265047073364 + ], + "opposite_logprob_tokens": [ + -1.3135325908660889 + ], + "logprob_reward": -0.31353265047073364, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31353265047073364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15251508355140686, + 0.30348533391952515, + 0.03372907638549805, + 0.12300664186477661, + 0.0612216517329216, + 0.057036902755498886, + 0.10239177942276001, + 0.03296123445034027, + 0.029484378173947334, + 0.10416793078184128 + ], + "credit_map": [ + 0.022098368033766747, + 0.011669404804706573, + 0.018033569678664207, + 0.019119862467050552, + 0.01577327400445938, + 0.013034380041062832, + 0.02089359611272812, + 0.031892623752355576, + 0.01598588563501835, + 0.02509726583957672, + 0.01112257782369852, + 0.01987740583717823, + 0.11921565979719162, + 0.01925547607243061, + 0.06578190624713898, + 0.027149168774485588, + 0.004424029029905796, + 0.0042794207111001015, + 0.0043152328580617905, + 0.005549825727939606, + 0.002762210089713335, + 0.004385062493383884, + 0.003320156130939722, + 0.004693138878792524, + 0.010164957493543625, + 0.014365158975124359, + 0.011865624226629734, + 0.00947452150285244, + 0.01928592659533024, + 0.02226843312382698, + 0.008771167136728764, + 0.026810863986611366, + 0.006788860075175762, + 0.003779746824875474, + 0.014341084286570549, + 0.007685651071369648, + 0.008118420839309692, + 0.004792382009327412, + 0.007720600813627243, + 0.007994904182851315, + 0.008211917243897915, + 0.004043687600642443, + 0.008607042953372002, + 0.006974759511649609, + 0.0073884776793420315, + 0.004124043975025415, + 0.01209945511072874, + 0.0055875168181955814, + 0.024278147146105766, + 0.0028931391425430775, + 0.010774871334433556, + 0.00795337837189436, + 0.005984441377222538, + 0.016363633796572685, + 0.006161899771541357, + 0.027982259169220924, + 0.004927974194288254, + 0.002256041392683983, + 0.004578334279358387, + 0.004182077012956142, + 0.003205147571861744, + 0.005129978992044926, + 0.005227597896009684, + 0.0034540854394435883, + 0.0060983020812273026, + 0.003305681049823761, + 0.005331814754754305, + 0.003676072461530566, + 0.0020020005758851767, + 0.0034148753620684147, + 0.0027313735336065292, + 0.0029242578893899918, + 0.017033791169524193, + 0.007192047778517008, + 0.016625266522169113, + 0.008996839635074139, + 0.011517214588820934, + 0.012552804313600063, + 0.01743401400744915, + 0.012815961614251137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.31353265047073364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015243588713929057 + ], + "opposite_logprob_tokens": [ + -6.501524448394775 + ], + "logprob_reward": -0.0015243588713929057, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015243588713929057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08680955320596695, + 0.12090953439474106, + 0.03584814444184303, + 0.08656594157218933, + 0.10686803609132767, + 0.06742870062589645, + 0.16495196521282196, + 0.03798677772283554, + 0.03311896696686745, + 0.2595123052597046 + ], + "credit_map": [ + 0.007333504036068916, + 0.00802552793174982, + 0.008427471853792667, + 0.012194579467177391, + 0.013272305019199848, + 0.009042446501553059, + 0.010181606747210026, + 0.018332118168473244, + 0.0053277090191841125, + 0.013241332024335861, + 0.009487982839345932, + 0.01512155495584011, + 0.012031666934490204, + 0.006893436890095472, + 0.0340755395591259, + 0.024730319157242775, + 0.0041855513118207455, + 0.004319639410823584, + 0.005841619335114956, + 0.0055908095091581345, + 0.0028604629915207624, + 0.003475970821455121, + 0.003606164362281561, + 0.005967929493635893, + 0.007041153963655233, + 0.009248301386833191, + 0.009353418834507465, + 0.013871811330318451, + 0.005356646608561277, + 0.014976635575294495, + 0.006229138933122158, + 0.020488841459155083, + 0.011851814575493336, + 0.004786457866430283, + 0.03217992186546326, + 0.010488258674740791, + 0.007943165488541126, + 0.007076111156493425, + 0.007131385616958141, + 0.02541092224419117, + 0.007981641218066216, + 0.0046429503709077835, + 0.009842315688729286, + 0.01126146875321865, + 0.008156386204063892, + 0.004460315220057964, + 0.015364260412752628, + 0.005719368811696768, + 0.030890513211488724, + 0.010856013745069504, + 0.017378102988004684, + 0.01693844608962536, + 0.013945012353360653, + 0.01643349975347519, + 0.023662718012928963, + 0.034847673028707504, + 0.0068098073825240135, + 0.0028921496123075485, + 0.004981407430022955, + 0.004943561274558306, + 0.0047882660292088985, + 0.005014336667954922, + 0.005422604735940695, + 0.003134647849947214, + 0.004931660369038582, + 0.003384773386642337, + 0.0050508687272667885, + 0.004804418887943029, + 0.002766365185379982, + 0.0043565621599555016, + 0.004416279960423708, + 0.0034080410841852427, + 0.049544092267751694, + 0.021396737545728683, + 0.034258242696523666, + 0.02575782686471939, + 0.023854903876781464, + 0.04089429974555969, + 0.03999396413564682, + 0.02381226234138012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.0015243588713929057 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036199060268700123 + ], + "opposite_logprob_tokens": [ + -5.62861967086792 + ], + "logprob_reward": -0.0036199060268700123, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036199060268700123, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11815998703241348, + 0.10591097176074982, + 0.05981690436601639, + 0.11796479672193527, + 0.09423930197954178, + 0.04330388829112053, + 0.10289275646209717, + 0.03586496412754059, + 0.05988633632659912, + 0.26196011900901794 + ], + "credit_map": [ + 0.010526196099817753, + 0.01724512130022049, + 0.011734980158507824, + 0.01568508706986904, + 0.014831884764134884, + 0.012371976859867573, + 0.013727573677897453, + 0.02203715592622757, + 0.009162932634353638, + 0.008755382150411606, + 0.010407675057649612, + 0.010477171279489994, + 0.011162027716636658, + 0.016826510429382324, + 0.020077800378203392, + 0.019041456282138824, + 0.008316285908222198, + 0.004876842722296715, + 0.008725382387638092, + 0.006271117366850376, + 0.004347317386418581, + 0.0062677594833076, + 0.004900310188531876, + 0.016111884266138077, + 0.009534093551337719, + 0.007706127595156431, + 0.013346133753657341, + 0.008403855375945568, + 0.01609491929411888, + 0.025857223197817802, + 0.006292293779551983, + 0.030730152502655983, + 0.012395943515002728, + 0.005643047858029604, + 0.03024960495531559, + 0.009655213914811611, + 0.006866620387881994, + 0.004899295512586832, + 0.006054992787539959, + 0.018474576994776726, + 0.006097820587456226, + 0.005823300685733557, + 0.006198684684932232, + 0.005460812710225582, + 0.004463511519134045, + 0.0034403568133711815, + 0.007543153595179319, + 0.00427624536678195, + 0.025124333798885345, + 0.005007054191082716, + 0.010203547775745392, + 0.008265124633908272, + 0.004711344372481108, + 0.014561968855559826, + 0.0049626464024186134, + 0.030056722462177277, + 0.005710447672754526, + 0.0025040321052074432, + 0.0055152010172605515, + 0.0044270907528698444, + 0.004461802076548338, + 0.005340770818293095, + 0.004337471444159746, + 0.003568145213648677, + 0.009832433424890041, + 0.00804110337048769, + 0.009297197684645653, + 0.008054839447140694, + 0.005602216813713312, + 0.007349851541221142, + 0.007528321351855993, + 0.004180368967354298, + 0.046211350709199905, + 0.022963983938097954, + 0.03407345712184906, + 0.02750815451145172, + 0.023411624133586884, + 0.04226348549127579, + 0.041641026735305786, + 0.023887038230895996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 3 + ], + "selected_reward": -0.0036199060268700123 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002492994070053 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002492994070053, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002492994070053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08472846448421478, + 0.20005597174167633, + 0.024669745936989784, + 0.047399915754795074, + 0.03601527959108353, + 0.03376075625419617, + 0.1369343101978302, + 0.03184831142425537, + 0.026636093854904175, + 0.37795114517211914 + ], + "credit_map": [ + 0.008795137517154217, + 0.005682957824319601, + 0.008673139847815037, + 0.013490661978721619, + 0.010611428879201412, + 0.00753403315320611, + 0.009428043849766254, + 0.02051307074725628, + 0.010360629297792912, + 0.016805225983262062, + 0.006435228977352381, + 0.012939131818711758, + 0.07825177162885666, + 0.01338737178593874, + 0.04057873785495758, + 0.021297866478562355, + 0.002481377450749278, + 0.002865780144929886, + 0.00419113552197814, + 0.003176300786435604, + 0.002865957096219063, + 0.002640513936057687, + 0.002369280904531479, + 0.00407940149307251, + 0.004911716561764479, + 0.003717347513884306, + 0.0039958953857421875, + 0.00543192820623517, + 0.004211406223475933, + 0.010005590505897999, + 0.0026068040169775486, + 0.012519228272140026, + 0.003197399666532874, + 0.0029867328703403473, + 0.00806227046996355, + 0.003862622892484069, + 0.005247737746685743, + 0.002627547597512603, + 0.005363017320632935, + 0.0046679493971168995, + 0.00546125927940011, + 0.003855663351714611, + 0.005190463736653328, + 0.003660938935354352, + 0.003763098269701004, + 0.004004180431365967, + 0.005170851945877075, + 0.002654296113178134, + 0.03881046175956726, + 0.005070263985544443, + 0.018452545627951622, + 0.008766095153987408, + 0.005173171404749155, + 0.0181130338460207, + 0.006541033275425434, + 0.036007706075906754, + 0.0057489946484565735, + 0.002113709691911936, + 0.004577262327075005, + 0.003579947631806135, + 0.004636847414076328, + 0.0027674322482198477, + 0.006426949519664049, + 0.001997163984924555, + 0.0034869504161179066, + 0.002778668189421296, + 0.003316093236207962, + 0.004433007910847664, + 0.0032815937884151936, + 0.0029245284385979176, + 0.0036580925807356834, + 0.0027571592945605516, + 0.05599534511566162, + 0.025448640808463097, + 0.04689004272222519, + 0.039274320006370544, + 0.04873095452785492, + 0.05142877623438835, + 0.07606472074985504, + 0.03411831706762314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.07002492994070053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "handheld medium close-up shot from a slightly upward angle, the camera whips back violently, following the subject's movements, finally framing the face of an East Asian woman, her face pale with fear. Her long, damp, black hair, a few strands tangled and plastered to her pale cheeks and forehead. Her eyes are wide open with utter terror, pupils constricted, her mouth drawn into a silent, screaming O, her jaw clenched with tension, as if she's exerting all her strength but unable to utter a single sound. She's in a narrow alley, squeezed oppressively by high, crumbling walls on either side. The only source of light is a flickering, poorly connected ceiling light. Each flicker casts a twitching, rapidly alternating light and shadow on her distorted features and the damp brick wall behind her, reflecting this sickly light off the damp ground. She's wearing a dark gray hoodie, and her entire body freezes as she suddenly turns back, her shoulders clenched, creating a suffocating atmosphere of intense tension and suspense, as if time has frozen in time, leaving only her silent terror.", + "seed": 98, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_12/index_6204_seed_98_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.07108449432998895, + "vlm_logprob": -0.07108449432998895, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.07108449432998895, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09972494840621948 + ], + "opposite_logprob_tokens": [ + -2.7247250080108643 + ], + "logprob_reward": -0.09972494840621948, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.09972494840621948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12469884753227234, + 0.058497458696365356, + 0.15373152494430542, + 0.03263615444302559, + 0.025043347850441933, + 0.04476213455200195, + 0.03878758102655411, + 0.4309144914150238, + 0.03870241716504097, + 0.052226029336452484 + ], + "credit_map": [ + 0.014224205166101456, + 0.008990233764052391, + 0.020889578387141228, + 0.010613986290991306, + 0.007137734908610582, + 0.02691144496202469, + 0.027178389951586723, + 0.008753270842134953, + 0.004998011980205774, + 0.0026147880125790834, + 0.016585391014814377, + 0.0045920019038021564, + 0.003780188038945198, + 0.003987020347267389, + 0.007930450141429901, + 0.014009607024490833, + 0.025011353194713593, + 0.010145191103219986, + 0.03659798949956894, + 0.02571377344429493, + 0.013100081123411655, + 0.010279227048158646, + 0.021768011152744293, + 0.011115899309515953, + 0.004902007523924112, + 0.005621659103780985, + 0.005522369407117367, + 0.0034326883032917976, + 0.004551291931420565, + 0.0017177853733301163, + 0.004486407153308392, + 0.002401943551376462, + 0.0025794340763241053, + 0.0019689989276230335, + 0.002637391211465001, + 0.0037049083039164543, + 0.0032008453272283077, + 0.001835460658185184, + 0.006359445862472057, + 0.0027568647637963295, + 0.0028451704420149326, + 0.00521733658388257, + 0.0032836764585226774, + 0.010589315555989742, + 0.002556927502155304, + 0.005479632876813412, + 0.003434559563174844, + 0.011355512775480747, + 0.0036156068090349436, + 0.0059643955901265144, + 0.004429750610142946, + 0.004650507587939501, + 0.004423072095960379, + 0.003412471152842045, + 0.006318974308669567, + 0.005972801707684994, + 0.08461447805166245, + 0.019774802029132843, + 0.1720074713230133, + 0.03923989459872246, + 0.01942860707640648, + 0.037121888250112534, + 0.02650585025548935, + 0.03222150728106499, + 0.004778213798999786, + 0.004316598176956177, + 0.004634341225028038, + 0.0044254218228161335, + 0.0036497849505394697, + 0.006789874285459518, + 0.004376731812953949, + 0.005731452722102404, + 0.008293138816952705, + 0.003764732275158167, + 0.006996059790253639, + 0.005119942128658295, + 0.002444413723424077, + 0.007405460812151432, + 0.0033184122294187546, + 0.014883868396282196 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.09972494840621948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17996911704540253 + ], + "opposite_logprob_tokens": [ + -1.804969072341919 + ], + "logprob_reward": -0.17996911704540253, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17996911704540253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2606773376464844, + 0.16880302131175995, + 0.08242020010948181, + 0.04560767859220505, + 0.03049313649535179, + 0.06612058728933334, + 0.049983441829681396, + 0.11516325920820236, + 0.05821079760789871, + 0.12252054363489151 + ], + "credit_map": [ + 0.03081493265926838, + 0.014391710050404072, + 0.07360006123781204, + 0.026503929868340492, + 0.021116331219673157, + 0.025962717831134796, + 0.05238259956240654, + 0.015905052423477173, + 0.010584999807178974, + 0.01185030397027731, + 0.02048138529062271, + 0.048127926886081696, + 0.017666397616267204, + 0.01069438923150301, + 0.021518360823392868, + 0.0278792604804039, + 0.012466886080801487, + 0.005561660043895245, + 0.01702159270644188, + 0.01612619124352932, + 0.007836428470909595, + 0.005246893502771854, + 0.01310078427195549, + 0.005059763323515654, + 0.005794479511678219, + 0.007605044171214104, + 0.006511877756565809, + 0.004583542700856924, + 0.006086255423724651, + 0.0036390714813023806, + 0.008271362632513046, + 0.003116049338132143, + 0.0027451294008642435, + 0.002415178809314966, + 0.003606505459174514, + 0.004808213096112013, + 0.003306064521893859, + 0.002227812772616744, + 0.008125750347971916, + 0.003258482087403536, + 0.004874181002378464, + 0.009305238723754883, + 0.0056060985662043095, + 0.00929009635001421, + 0.008000348694622517, + 0.014651050791144371, + 0.007708325050771236, + 0.00668525043874979, + 0.010294740088284016, + 0.0036474321968853474, + 0.015950018540024757, + 0.003731869626790285, + 0.0028963033109903336, + 0.0041633835062384605, + 0.00448228232562542, + 0.004817408975213766, + 0.020465970039367676, + 0.00772778270766139, + 0.03786282613873482, + 0.010483904741704464, + 0.007642828393727541, + 0.010474870912730694, + 0.007880846969783306, + 0.012624234892427921, + 0.006563579197973013, + 0.0040503269992768764, + 0.009501507505774498, + 0.006068097427487373, + 0.003088445169851184, + 0.01545217726379633, + 0.004998311400413513, + 0.008488353341817856, + 0.014441605657339096, + 0.014212852343916893, + 0.015429673716425896, + 0.015607559122145176, + 0.012316391803324223, + 0.01594184711575508, + 0.014803293161094189, + 0.019767316058278084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.17996911704540253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002498126355931163 + ], + "opposite_logprob_tokens": [ + -6.002498149871826 + ], + "logprob_reward": -0.002498126355931163, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002498126355931163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20328477025032043, + 0.13387496769428253, + 0.06622952967882156, + 0.11614200472831726, + 0.02734096348285675, + 0.05472801253199577, + 0.02073962427675724, + 0.13067665696144104, + 0.0759190246462822, + 0.17106451094150543 + ], + "credit_map": [ + 0.02159665897488594, + 0.010236937552690506, + 0.07696651667356491, + 0.01260283961892128, + 0.009870854206383228, + 0.025346213951706886, + 0.03220276162028313, + 0.014461977407336235, + 0.004827791824936867, + 0.008735684677958488, + 0.011110682971775532, + 0.03700326383113861, + 0.007747546769678593, + 0.008401075378060341, + 0.01711644046008587, + 0.03893246874213219, + 0.00954294390976429, + 0.00533899525180459, + 0.01228378713130951, + 0.012154231779277325, + 0.00920193362981081, + 0.003733985126018524, + 0.00957232341170311, + 0.004401325713843107, + 0.01615472324192524, + 0.01593550480902195, + 0.01837826520204544, + 0.01168548408895731, + 0.01964622177183628, + 0.007719276938587427, + 0.01815912500023842, + 0.008463400416076183, + 0.004087028093636036, + 0.002411953406408429, + 0.004043924622237682, + 0.0036065883468836546, + 0.0027118546422570944, + 0.0020156146492809057, + 0.005727474112063646, + 0.00273652421310544, + 0.004199847113341093, + 0.009317883290350437, + 0.004940374288707972, + 0.012648277916014194, + 0.003610801650211215, + 0.0044642179273068905, + 0.0054329815320670605, + 0.010113624855875969, + 0.0031731671188026667, + 0.001546804211102426, + 0.00517254089936614, + 0.001968634780496359, + 0.001364933792501688, + 0.002321175765246153, + 0.002649588044732809, + 0.002542778616771102, + 0.025116343051195145, + 0.00586965074762702, + 0.052257832139730453, + 0.01470947079360485, + 0.0049620941281318665, + 0.010280318558216095, + 0.008175673894584179, + 0.009305273182690144, + 0.008254899643361568, + 0.0073832254856824875, + 0.009431292302906513, + 0.011528071947395802, + 0.004957735538482666, + 0.015783265233039856, + 0.008047940209507942, + 0.010532589629292488, + 0.025382043793797493, + 0.0100218141451478, + 0.03383984789252281, + 0.014331121928989887, + 0.0069931442849338055, + 0.018545866012573242, + 0.01247507892549038, + 0.049475573003292084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.002498126355931163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031960143242031336 + ], + "opposite_logprob_tokens": [ + -5.7531962394714355 + ], + "logprob_reward": -0.0031960143242031336, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031960143242031336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21064932644367218, + 0.08618064969778061, + 0.1243571788072586, + 0.15265150368213654, + 0.10093566030263901, + 0.03295525163412094, + 0.022023670375347137, + 0.13639605045318604, + 0.03660713881254196, + 0.09724350273609161 + ], + "credit_map": [ + 0.026991529390215874, + 0.011903064325451851, + 0.07155458629131317, + 0.014458592049777508, + 0.008395249955356121, + 0.021228637546300888, + 0.04341680929064751, + 0.012700877152383327, + 0.004972119815647602, + 0.004162140190601349, + 0.01300475187599659, + 0.014441060833632946, + 0.004465225152671337, + 0.0060441577807068825, + 0.011125660501420498, + 0.027965541929006577, + 0.01823711022734642, + 0.009124618023633957, + 0.025914106518030167, + 0.021809915080666542, + 0.015641095116734505, + 0.007550152949988842, + 0.01742972992360592, + 0.008650467731058598, + 0.034091804176568985, + 0.016296787187457085, + 0.041159652173519135, + 0.015248360112309456, + 0.010874691419303417, + 0.009985825046896935, + 0.017601080238819122, + 0.007393313571810722, + 0.01809193380177021, + 0.00852795597165823, + 0.017306344583630562, + 0.012270933948457241, + 0.006937769241631031, + 0.011220126412808895, + 0.011907589621841908, + 0.014673012308776379, + 0.00462186336517334, + 0.005250184331089258, + 0.002921047620475292, + 0.006379012484103441, + 0.002066292567178607, + 0.0037441204767674208, + 0.003992853686213493, + 0.003979882691055536, + 0.0029443928506225348, + 0.0019724147859960794, + 0.002960559446364641, + 0.0026097611989825964, + 0.0017589183989912271, + 0.0029701986350119114, + 0.0032458193600177765, + 0.0035616084933280945, + 0.026130756363272667, + 0.007291059009730816, + 0.04961838573217392, + 0.010759948752820492, + 0.0068799760192632675, + 0.011904495768249035, + 0.010933485813438892, + 0.012877951376140118, + 0.007139978464692831, + 0.0031449561938643456, + 0.0055013904348015785, + 0.0038615968078374863, + 0.0029792708810418844, + 0.007133425213396549, + 0.002816362539306283, + 0.004030162934213877, + 0.012183001264929771, + 0.007233758922666311, + 0.015517566353082657, + 0.009141426533460617, + 0.005444090347737074, + 0.01559643354266882, + 0.006689161527901888, + 0.025438077747821808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.0031960143242031336 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a handheld medium close-up from a slightly upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera whip back violently to frame the face of an East Asian woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the woman's face pale with fear, featuring long damp black hair plastered to her cheeks and forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the woman wearing a dark gray hoodie in a narrow alley with high crumbling walls and a flickering ceiling light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman's body freeze as she turns back with clenched shoulders, creating a static moment of intense tension?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07003426551818848 + ], + "opposite_logprob_tokens": [ + -2.6950342655181885 + ], + "logprob_reward": -0.07003426551818848, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07003426551818848, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24016635119915009, + 0.06323873996734619, + 0.08977541327476501, + 0.07765839248895645, + 0.05064767599105835, + 0.023317154496908188, + 0.04378573223948479, + 0.285958856344223, + 0.03407437354326248, + 0.09137731790542603 + ], + "credit_map": [ + 0.023572400212287903, + 0.013944446109235287, + 0.06943638622760773, + 0.03065396659076214, + 0.012275177985429764, + 0.05330260097980499, + 0.024094242602586746, + 0.012887141667306423, + 0.0049007912166416645, + 0.004720607306808233, + 0.007740276865661144, + 0.014839383773505688, + 0.004936940502375364, + 0.005394256208091974, + 0.009522572159767151, + 0.0111839072778821, + 0.006926883477717638, + 0.004691531881690025, + 0.01100753154605627, + 0.027568940073251724, + 0.0073980228044092655, + 0.004630963318049908, + 0.023447033017873764, + 0.00410451041534543, + 0.014395189471542835, + 0.008029443211853504, + 0.018822424113750458, + 0.008185641840100288, + 0.007089919876307249, + 0.00615270808339119, + 0.009477639570832253, + 0.0055054230615496635, + 0.008028478361666203, + 0.004153293091803789, + 0.00946118962019682, + 0.006112842820584774, + 0.004478611517697573, + 0.00391039066016674, + 0.008780588395893574, + 0.005722283385694027, + 0.0025453518610447645, + 0.0029950514435768127, + 0.0020340168848633766, + 0.0039165448397397995, + 0.002006707713007927, + 0.003221039427444339, + 0.0028178554493933916, + 0.003780584316700697, + 0.008507986553013325, + 0.003416908672079444, + 0.009124196134507656, + 0.004764790181070566, + 0.0033766706474125385, + 0.003945846576243639, + 0.0056313141249120235, + 0.005018019583076239, + 0.05590396374464035, + 0.01103918720036745, + 0.11562816053628922, + 0.03284573182463646, + 0.009823794476687908, + 0.021801240742206573, + 0.018890390172600746, + 0.020026402547955513, + 0.0053032501600682735, + 0.004732840694487095, + 0.0035069112200289965, + 0.006303715985268354, + 0.0024076614063233137, + 0.0037385341711342335, + 0.004539770539849997, + 0.003541688434779644, + 0.011016790755093098, + 0.00689314678311348, + 0.02039109542965889, + 0.007965822704136372, + 0.0037084314972162247, + 0.010785169899463654, + 0.005253123585134745, + 0.025363733991980553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.07003426551818848 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_tensor([9652, 9652])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_tensor([9652, 9652])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2edfb6a49102e845afc5698f8502e3200631b002 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_tensor([9652, 9652])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, high-angle, wide-angle shot gazes down upon a sleepy, peaceful farm nestled in a valley beneath a vast night sky. A brilliant full moon hangs high, radiating a soft, silvery glow that gently illuminates the scene, creating long, softly edged shadows. Below, a classic two-story white farmhouse with a dark, gabled roof rests peacefully. A single upstairs window emits a warm, inviting golden light that pulses with a slow, subtle rhythm like a soft heartbeat. Beside it stands a large, traditional red barn, its silhouette a crimson crimson in the cool moonlight. The air is silent and filled with magic; countless fireflies lazily drift across the quiet pasture, their tiny lanterns flickering in slow, hypnotic synchronization. Pale, shimmering, ethereal points of light slowly descend from the deep blue sky, capturing the drifting moonlight and enveloping the landscape in a shimmering, magical dust, creating a profoundly tranquil, calming, bedtime atmosphere. Absolutely peaceful.", + "seed": 25842, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_9652_seed_25842_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03420449490658939, + "vlm_logprob": -0.03420449490658939, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.03420449490658939, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05146896466612816 + ], + "opposite_logprob_tokens": [ + -3.051468849182129 + ], + "logprob_reward": -0.05146896466612816, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05146896466612816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08616430312395096, + 0.06984283775091171, + 0.25262364745140076, + 0.03973052650690079, + 0.06594083458185196, + 0.28678423166275024, + 0.09124460071325302, + 0.018022218719124794, + 0.06579490751028061, + 0.02385178953409195 + ], + "credit_map": [ + 0.0099785216152668, + 0.011590154841542244, + 0.012350548058748245, + 0.012300686910748482, + 0.005481763277202845, + 0.01351314689964056, + 0.010318565182387829, + 0.010630929842591286, + 0.00847756676375866, + 0.009459682740271091, + 0.009156789630651474, + 0.011178916320204735, + 0.004147960804402828, + 0.009853088296949863, + 0.008375339210033417, + 0.009193508885800838, + 0.06439374387264252, + 0.10993904620409012, + 0.005157523788511753, + 0.006883756257593632, + 0.008439897559583187, + 0.044595442712306976, + 0.006852481979876757, + 0.006361803039908409, + 0.005217468831688166, + 0.010015200823545456, + 0.004704590421169996, + 0.004651769530028105, + 0.0026548278983682394, + 0.0050437068566679955, + 0.003509181085973978, + 0.003933788277208805, + 0.00662922952324152, + 0.011890876106917858, + 0.007681612856686115, + 0.009921845979988575, + 0.003051674459129572, + 0.008316275663673878, + 0.0067062233574688435, + 0.011743106879293919, + 0.038931019604206085, + 0.10552235692739487, + 0.015549601055681705, + 0.016883280128240585, + 0.013718921691179276, + 0.07024358212947845, + 0.013301429338753223, + 0.012634092941880226, + 0.009830528870224953, + 0.0223542470484972, + 0.007214517332613468, + 0.010573597624897957, + 0.007868709042668343, + 0.017457786947488785, + 0.007116339169442654, + 0.008828887715935707, + 0.0023045188281685114, + 0.0034057674929499626, + 0.0028664974961429834, + 0.002165053738281131, + 0.0012028903001919389, + 0.0014863492688164115, + 0.0020343284122645855, + 0.002556814346462488, + 0.005760790314525366, + 0.008603747002780437, + 0.009883610531687737, + 0.011763157323002815, + 0.005587800405919552, + 0.0065674628131091595, + 0.007788660004734993, + 0.009839686565101147, + 0.002663844730705023, + 0.0033317890483886003, + 0.002983757993206382, + 0.004181980155408382, + 0.002829317469149828, + 0.0027072622906416655, + 0.002801312366500497, + 0.0023525303695350885 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.05146896466612816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032043324317783117 + ], + "opposite_logprob_tokens": [ + -5.753204345703125 + ], + "logprob_reward": -0.0032043324317783117, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032043324317783117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06631489098072052, + 0.08785764873027802, + 0.38784459233283997, + 0.017489956691861153, + 0.3645398020744324, + 0.019692573696374893, + 0.014889970421791077, + 0.015697207301855087, + 0.015252316370606422, + 0.010421027429401875 + ], + "credit_map": [ + 0.016817335039377213, + 0.007106997538357973, + 0.0035578887909650803, + 0.00396305974572897, + 0.0043303524143993855, + 0.021480610594153404, + 0.00517750671133399, + 0.0038811401464045048, + 0.009318992495536804, + 0.011070074513554573, + 0.010912632569670677, + 0.013463815674185753, + 0.005086957011371851, + 0.009006928652524948, + 0.014051113277673721, + 0.014947133138775826, + 0.08447524905204773, + 0.17095781862735748, + 0.0059874639846384525, + 0.007593084592372179, + 0.012816479429602623, + 0.08697980642318726, + 0.00964941456913948, + 0.009385276585817337, + 0.001958196982741356, + 0.002081439597532153, + 0.0020417161285877228, + 0.0029213158413767815, + 0.0015880652936175466, + 0.002018972532823682, + 0.0022674966603517532, + 0.002612753538414836, + 0.06138201430439949, + 0.1512303203344345, + 0.007744841277599335, + 0.01043308898806572, + 0.012487546540796757, + 0.10084040462970734, + 0.007187434937804937, + 0.013234150595963001, + 0.0016963202506303787, + 0.0029656277038156986, + 0.0018815501825883985, + 0.0026513012126088142, + 0.0015496446285396814, + 0.0029641343280673027, + 0.002392408438026905, + 0.0035915866028517485, + 0.0015642479993402958, + 0.0027560838498175144, + 0.001620618044398725, + 0.0022921788040548563, + 0.0012489842483773828, + 0.002413590904325247, + 0.0012590866535902023, + 0.0017351804999634624, + 0.0017438028007745743, + 0.0023954221978783607, + 0.0017576566897332668, + 0.003010608023032546, + 0.001240378012880683, + 0.001460172003135085, + 0.001761072431690991, + 0.0023280938621610403, + 0.001321951625868678, + 0.0022816378623247147, + 0.0018166840309277177, + 0.002262681722640991, + 0.001586981932632625, + 0.001395663246512413, + 0.0019345356849953532, + 0.0026521803811192513, + 0.001373876235447824, + 0.0016979981446638703, + 0.0013466546079143882, + 0.001391490688547492, + 0.0007315272814594209, + 0.001084613730199635, + 0.001692508114501834, + 0.0011023591505363584 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 1 + ], + "selected_reward": -0.0032043324317783117 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023304728791117668 + ], + "opposite_logprob_tokens": [ + -3.7733047008514404 + ], + "logprob_reward": -0.023304728791117668, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023304728791117668, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06379653513431549, + 0.03551951423287392, + 0.4934879243373871, + 0.04023095220327377, + 0.10931558907032013, + 0.036828748881816864, + 0.10212807357311249, + 0.048833202570676804, + 0.02656647004187107, + 0.04329294338822365 + ], + "credit_map": [ + 0.008228668011724949, + 0.009406987577676773, + 0.007675105705857277, + 0.009126930497586727, + 0.006373864132910967, + 0.005568076856434345, + 0.008199673146009445, + 0.00921722687780857, + 0.004702277015894651, + 0.00487276166677475, + 0.0064607467502355576, + 0.004988074768334627, + 0.0033875745721161366, + 0.003455821191892028, + 0.003823797684162855, + 0.0038284631446003914, + 0.11367980390787125, + 0.22039178013801575, + 0.006152194458991289, + 0.009252957999706268, + 0.01618598960340023, + 0.10656466335058212, + 0.01216292753815651, + 0.009097627364099026, + 0.0074362787418067455, + 0.013857834972441196, + 0.0021137166768312454, + 0.0029627683106809855, + 0.002235863357782364, + 0.006365972571074963, + 0.0023190565407276154, + 0.002939461963251233, + 0.01889042928814888, + 0.04054074361920357, + 0.0047688912600278854, + 0.005284757819026709, + 0.0040512727573513985, + 0.024188842624425888, + 0.003939314279705286, + 0.007651344873011112, + 0.003714499296620488, + 0.006582918576896191, + 0.0029121246188879013, + 0.004057780373841524, + 0.00519939698278904, + 0.007019845303148031, + 0.003089022124186158, + 0.004253161605447531, + 0.01659882813692093, + 0.03457087278366089, + 0.004627906251698732, + 0.007730409968644381, + 0.007283893879503012, + 0.020831571891903877, + 0.005053751636296511, + 0.005430843681097031, + 0.007513081654906273, + 0.007434470113366842, + 0.005478186067193747, + 0.0058513120748102665, + 0.00625515915453434, + 0.005480328109115362, + 0.004949474707245827, + 0.005871186964213848, + 0.003131014993414283, + 0.004332070238888264, + 0.003209220478311181, + 0.0035277248825877905, + 0.003530931193381548, + 0.0026299820747226477, + 0.003086225362494588, + 0.003119299653917551, + 0.0056812213733792305, + 0.006948282942175865, + 0.004735554568469524, + 0.004924359731376171, + 0.005274463444948196, + 0.005509880371391773, + 0.005901807453483343, + 0.004317374434322119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.023304728791117668 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494944378733635 + ], + "opposite_logprob_tokens": [ + -2.9299495220184326 + ], + "logprob_reward": -0.05494944378733635, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494944378733635, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06396850943565369, + 0.02863958850502968, + 0.4510978162288666, + 0.03432171419262886, + 0.1264619082212448, + 0.18556295335292816, + 0.04117341712117195, + 0.021365782245993614, + 0.015222229063510895, + 0.03218603879213333 + ], + "credit_map": [ + 0.010661615058779716, + 0.007626259699463844, + 0.00829313788563013, + 0.007908359169960022, + 0.005898169707506895, + 0.011524348519742489, + 0.005594383459538221, + 0.006462239194661379, + 0.003932585008442402, + 0.004224528092890978, + 0.0047997282817959785, + 0.0034492856357246637, + 0.0025676563382148743, + 0.0030965812038630247, + 0.0028370851650834084, + 0.003732138080522418, + 0.11143329739570618, + 0.19879041612148285, + 0.007013015914708376, + 0.010517962276935577, + 0.014725384302437305, + 0.08623922616243362, + 0.011849816888570786, + 0.01052873395383358, + 0.0031987170223146677, + 0.006176111754029989, + 0.003767681075260043, + 0.005053211469203234, + 0.0019596717320382595, + 0.0050965361297130585, + 0.0033776562195271254, + 0.0056921313516795635, + 0.021672600880265236, + 0.05010346695780754, + 0.004059860948473215, + 0.004265035502612591, + 0.004482201766222715, + 0.03186073154211044, + 0.00330392736941576, + 0.006714080926030874, + 0.026890059933066368, + 0.06434118002653122, + 0.011315923184156418, + 0.013004396110773087, + 0.011140905320644379, + 0.03864287957549095, + 0.010726132430136204, + 0.009501495398581028, + 0.006388927344232798, + 0.01342091429978609, + 0.0020473788026720285, + 0.0033005177974700928, + 0.003059899900108576, + 0.008342844434082508, + 0.002210147911682725, + 0.0024027859326452017, + 0.002476160414516926, + 0.0031868659425526857, + 0.0025970847345888615, + 0.0026934552006423473, + 0.0023277821019291878, + 0.002481592120602727, + 0.002536394167691469, + 0.0030664480291306973, + 0.0018921545706689358, + 0.0021517248824238777, + 0.0015338925877586007, + 0.0019283887231722474, + 0.00234187301248312, + 0.0016588702565059066, + 0.0016325990436598659, + 0.0020827259868383408, + 0.004304085858166218, + 0.005349760875105858, + 0.0030166280921548605, + 0.002864107023924589, + 0.005259290337562561, + 0.004295974038541317, + 0.003928899299353361, + 0.0031672935001552105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 4 + ], + "selected_reward": -0.05494944378733635 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038095004856586456 + ], + "opposite_logprob_tokens": [ + -3.288094997406006 + ], + "logprob_reward": -0.038095004856586456, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038095004856586456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07796242833137512, + 0.02540065161883831, + 0.535897433757782, + 0.03130703791975975, + 0.04040443152189255, + 0.029768941923975945, + 0.0782383382320404, + 0.045619476586580276, + 0.02910807356238365, + 0.10629317164421082 + ], + "credit_map": [ + 0.011860146187245846, + 0.013423401862382889, + 0.009845247492194176, + 0.009397787973284721, + 0.009184910915791988, + 0.00875950139015913, + 0.008687283843755722, + 0.006804145872592926, + 0.003739151870831847, + 0.0029859067872166634, + 0.0023120983969420195, + 0.002811093581840396, + 0.0023559893015772104, + 0.005464298650622368, + 0.002301182132214308, + 0.003430931130424142, + 0.12745827436447144, + 0.23932667076587677, + 0.007076292764395475, + 0.010640637949109077, + 0.01742713898420334, + 0.11085116863250732, + 0.013471717946231365, + 0.009645492769777775, + 0.0033727744594216347, + 0.006202246993780136, + 0.0036151595413684845, + 0.0035800489131361246, + 0.002922121435403824, + 0.004579637665301561, + 0.0028291514609009027, + 0.004205898381769657, + 0.007192801218479872, + 0.012899762950837612, + 0.0023065723944455385, + 0.0028521402273327112, + 0.0022601124364882708, + 0.006691882386803627, + 0.003126927185803652, + 0.003074233653023839, + 0.003890089690685272, + 0.004200557246804237, + 0.0038036832120269537, + 0.0034265012945979834, + 0.0040145209059119225, + 0.003436545841395855, + 0.003517600242048502, + 0.0034794455859810114, + 0.010128329508006573, + 0.023720193654298782, + 0.004553616978228092, + 0.007483766879886389, + 0.006802880670875311, + 0.01632956974208355, + 0.004634964745491743, + 0.004585009999573231, + 0.005951638799160719, + 0.006754834670573473, + 0.005515091586858034, + 0.005522442515939474, + 0.006345957517623901, + 0.004773596301674843, + 0.00437272759154439, + 0.006383191794157028, + 0.003762196283787489, + 0.0037320845294743776, + 0.003907283302396536, + 0.004172009415924549, + 0.0037351036444306374, + 0.003163991030305624, + 0.0032350511755794287, + 0.003400354413315654, + 0.012513581663370132, + 0.019533341750502586, + 0.008694198913872242, + 0.008983842097222805, + 0.019671425223350525, + 0.01448677759617567, + 0.012450167909264565, + 0.009959839284420013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 6 + ], + "selected_reward": -0.038095004856586456 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot gazes down upon a sleepy, peaceful farm nestled in a valley beneath a vast night sky. A brilliant full moon hangs high, radiating a soft, silvery glow that gently illuminates the scene, creating long, softly edged shadows. Below, a classic two-story white farmhouse with a dark, gabled roof rests peacefully. A single upstairs window emits a warm, inviting golden light that pulses with a slow, subtle rhythm like a soft heartbeat. Beside it stands a large, traditional red barn, its silhouette a crimson crimson in the cool moonlight. The air is silent and filled with magic; countless fireflies lazily drift across the quiet pasture, their tiny lanterns flickering in slow, hypnotic synchronization. Pale, shimmering, ethereal points of light slowly descend from the deep blue sky, capturing the drifting moonlight and enveloping the landscape in a shimmering, magical dust, creating a profoundly tranquil, calming, bedtime atmosphere. Absolutely peaceful.", + "seed": 25843, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_9652_seed_25843_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -1.0664606288075447, + "vlm_logprob": -1.0664606288075447, + "vlm_yes_no_margin": 0.5, + "vlm_token_credit": -1.0664606288075447, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -4.635336875915527 + ], + "opposite_logprob_tokens": [ + -0.010337027721107006 + ], + "logprob_reward": -4.635336875915527, + "yes_no_margin_reward": -4.625, + "token_credit_reward": -4.635336875915527, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2916894555091858, + 0.06316650658845901, + 0.054102543741464615, + 0.09709852933883667, + 0.06987207382917404, + 0.0895448550581932, + 0.1273755133152008, + 0.12155693024396896, + 0.04742421582341194, + 0.03816939517855644 + ], + "credit_map": [ + 0.04428181052207947, + 0.024531595408916473, + 0.03799460828304291, + 0.04854942858219147, + 0.029780885204672813, + 0.03352400287985802, + 0.039919473230838776, + 0.03310764580965042, + 0.008912757970392704, + 0.004649654030799866, + 0.008993319235742092, + 0.006850281730294228, + 0.014545331709086895, + 0.008056450635194778, + 0.005553606431931257, + 0.005605105776339769, + 0.006669586058706045, + 0.005208627320826054, + 0.005712863523513079, + 0.011959153227508068, + 0.004789224360138178, + 0.006877667270600796, + 0.007510841824114323, + 0.005374579690396786, + 0.014329791069030762, + 0.006581752561032772, + 0.012753164395689964, + 0.01290332991629839, + 0.017600378021597862, + 0.014546753838658333, + 0.010553395375609398, + 0.007829965092241764, + 0.008696979843080044, + 0.003381655551493168, + 0.007879314944148064, + 0.01105539221316576, + 0.013376956805586815, + 0.009007139131426811, + 0.00897773914039135, + 0.0074968961998820305, + 0.0071064792573452, + 0.0034170029684901237, + 0.008352474309504032, + 0.019485177472233772, + 0.009278617799282074, + 0.012759777717292309, + 0.016950780525803566, + 0.012194548733532429, + 0.010917427949607372, + 0.005544029176235199, + 0.0125948591157794, + 0.030780712142586708, + 0.009493454359471798, + 0.015081578865647316, + 0.028096983209252357, + 0.014866472221910954, + 0.009592510759830475, + 0.0064688981510698795, + 0.015317645855247974, + 0.027361763641238213, + 0.01081605814397335, + 0.01620521955192089, + 0.022440176457166672, + 0.013354661874473095, + 0.006430072709918022, + 0.006111914291977882, + 0.005526745691895485, + 0.004874676000326872, + 0.008842970244586468, + 0.008039524778723717, + 0.0038476865738630295, + 0.0037506225053220987, + 0.003192577278241515, + 0.002169149462133646, + 0.0039880345575511456, + 0.00855211354792118, + 0.004900509025901556, + 0.005797394551336765, + 0.005424996372312307, + 0.004144621081650257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -4.635336875915527 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28162965178489685 + ], + "opposite_logprob_tokens": [ + -1.4066296815872192 + ], + "logprob_reward": -0.28162965178489685, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28162965178489685, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08365270495414734, + 0.08317753672599792, + 0.0382065549492836, + 0.06536172330379486, + 0.07340320199728012, + 0.1169300526380539, + 0.21047423779964447, + 0.26292139291763306, + 0.03463052958250046, + 0.031241996213793755 + ], + "credit_map": [ + 0.012926897034049034, + 0.006924390327185392, + 0.015073014423251152, + 0.009594660252332687, + 0.009995630010962486, + 0.009007143788039684, + 0.011037862859666348, + 0.009093107655644417, + 0.01128841657191515, + 0.005303341429680586, + 0.011165781877934933, + 0.009312713518738747, + 0.015506276860833168, + 0.010332034900784492, + 0.012158367782831192, + 0.008110611699521542, + 0.006951784715056419, + 0.003019964788109064, + 0.005273875314742327, + 0.005350537598133087, + 0.004368223249912262, + 0.0054056779481470585, + 0.004432485438883305, + 0.00340400543063879, + 0.0067351507022976875, + 0.003939428832381964, + 0.006989453919231892, + 0.012785687111318111, + 0.009085855446755886, + 0.008526141755282879, + 0.010200660675764084, + 0.00709934625774622, + 0.00885334238409996, + 0.0032519090455025434, + 0.008220325224101543, + 0.012898456305265427, + 0.010231927037239075, + 0.009976661764085293, + 0.013106507249176502, + 0.006864071357995272, + 0.011451355181634426, + 0.0061119613237679005, + 0.011752256192266941, + 0.02115892618894577, + 0.011635923758149147, + 0.01782601699233055, + 0.025110037997364998, + 0.011883572675287724, + 0.01656160317361355, + 0.009117751382291317, + 0.019885316491127014, + 0.04867994412779808, + 0.01745384931564331, + 0.02697063609957695, + 0.04951711371541023, + 0.022288024425506592, + 0.02318091318011284, + 0.012066594325006008, + 0.03672356903553009, + 0.05446572229266167, + 0.02054966241121292, + 0.03530942648649216, + 0.051979001611471176, + 0.02864651568233967, + 0.004288611933588982, + 0.004223569296300411, + 0.0033518909476697445, + 0.0036596350837498903, + 0.006706522777676582, + 0.00540394801646471, + 0.004125664941966534, + 0.002870685188099742, + 0.0027854046784341335, + 0.0017013625474646688, + 0.002944118110463023, + 0.00616172282025218, + 0.00440173177048564, + 0.004553585313260555, + 0.005036840680986643, + 0.0036572308745235205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 5 + ], + "selected_reward": -0.28162965178489685 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08923519402742386 + ], + "opposite_logprob_tokens": [ + -2.464235305786133 + ], + "logprob_reward": -0.08923519402742386, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08923519402742386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11747896671295166, + 0.08283742517232895, + 0.037139393389225006, + 0.14470018446445465, + 0.11267703026533127, + 0.11315345019102097, + 0.17047671973705292, + 0.1317254602909088, + 0.05008122697472572, + 0.03973012417554855 + ], + "credit_map": [ + 0.01692781038582325, + 0.006639714352786541, + 0.02150576561689377, + 0.01397365890443325, + 0.01086634211242199, + 0.01476375199854374, + 0.015884265303611755, + 0.016917653381824493, + 0.010147386230528355, + 0.007486472837626934, + 0.013017705641686916, + 0.010141669772565365, + 0.017010731622576714, + 0.010225039906799793, + 0.008940545842051506, + 0.005867875646799803, + 0.005521263461560011, + 0.003227036213502288, + 0.005492584314197302, + 0.003642660565674305, + 0.003825064515694976, + 0.005467129871249199, + 0.005076572299003601, + 0.00488708121702075, + 0.024036414921283722, + 0.009979739785194397, + 0.019935710355639458, + 0.014913042075932026, + 0.028728235512971878, + 0.023039687424898148, + 0.013191958889365196, + 0.010875411331653595, + 0.011332209222018719, + 0.008587023243308067, + 0.015141302719712257, + 0.016042135655879974, + 0.018108878284692764, + 0.015325978398323059, + 0.017163660377264023, + 0.01097584143280983, + 0.009451491758227348, + 0.0040894546546041965, + 0.013228479772806168, + 0.021277572959661484, + 0.011249136179685593, + 0.01347452774643898, + 0.024687016382813454, + 0.01569577120244503, + 0.01764805242419243, + 0.00800721999257803, + 0.017306141555309296, + 0.03272715210914612, + 0.01722608134150505, + 0.02345748245716095, + 0.036292701959609985, + 0.017811881378293037, + 0.011930271051824093, + 0.006371675990521908, + 0.017488088458776474, + 0.022697748616337776, + 0.017116455361247063, + 0.018824107944965363, + 0.025245271623134613, + 0.012051844969391823, + 0.004420221783220768, + 0.004680706188082695, + 0.005877919495105743, + 0.008657426573336124, + 0.006031181663274765, + 0.007080371025949717, + 0.008547747507691383, + 0.004785653669387102, + 0.004671711474657059, + 0.0030727635603398085, + 0.00532310688868165, + 0.006258749403059483, + 0.006000025197863579, + 0.0058825197629630566, + 0.004121462348848581, + 0.004399784374982119 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 7 + ], + "selected_reward": -0.08923519402742386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2257205843925476 + ], + "opposite_logprob_tokens": [ + -1.6007206439971924 + ], + "logprob_reward": -0.2257205843925476, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2257205843925476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16952329874038696, + 0.11013675481081009, + 0.05322239547967911, + 0.16910406947135925, + 0.07932686060667038, + 0.1477147787809372, + 0.1205543801188469, + 0.0716191753745079, + 0.05217987298965454, + 0.026618458330631256 + ], + "credit_map": [ + 0.02109064720571041, + 0.009666508063673973, + 0.023209577426314354, + 0.031932417303323746, + 0.014205167070031166, + 0.023399224504828453, + 0.022590328007936478, + 0.02342941425740719, + 0.015031881630420685, + 0.016203848645091057, + 0.012205113656818867, + 0.01653500460088253, + 0.01735527440905571, + 0.012019742280244827, + 0.011055400595068932, + 0.009730495512485504, + 0.0066607799381017685, + 0.006715086754411459, + 0.006410026922821999, + 0.009158393368124962, + 0.006860875524580479, + 0.006374543532729149, + 0.006525382399559021, + 0.004517304711043835, + 0.027785541489720345, + 0.01992022804915905, + 0.021838055923581123, + 0.019131096079945564, + 0.02293144166469574, + 0.026532458141446114, + 0.015600976534187794, + 0.015364271588623524, + 0.009747346863150597, + 0.005490646231919527, + 0.010958113707602024, + 0.00860331580042839, + 0.0162651464343071, + 0.012070922181010246, + 0.008334603160619736, + 0.007856763899326324, + 0.015776561573147774, + 0.0114900516346097, + 0.013552282936871052, + 0.025113508105278015, + 0.02252843976020813, + 0.018776671960949898, + 0.027426207438111305, + 0.013051056303083897, + 0.012085577473044395, + 0.005480449181050062, + 0.014261515811085701, + 0.022295741364359856, + 0.013258291408419609, + 0.017067039385437965, + 0.024540681391954422, + 0.01156508270651102, + 0.008383549749851227, + 0.00325580476783216, + 0.008841799572110176, + 0.01007501594722271, + 0.010984384454786777, + 0.011798369698226452, + 0.01204086933284998, + 0.006239383947104216, + 0.0043805744498968124, + 0.002670752350240946, + 0.0040991040877997875, + 0.011317308992147446, + 0.007432939019054174, + 0.006557277869433165, + 0.01067862007766962, + 0.0050432924181222916, + 0.0029185134917497635, + 0.001816516276448965, + 0.0027796283829957247, + 0.0038798179011791945, + 0.004728397820144892, + 0.0038883674424141645, + 0.003776655299589038, + 0.0028305640444159508 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.2257205843925476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10038083791732788 + ], + "opposite_logprob_tokens": [ + -2.3503808975219727 + ], + "logprob_reward": -0.10038083791732788, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10038083791732788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.334381639957428, + 0.09860414266586304, + 0.04705620929598808, + 0.11054684966802597, + 0.07056601345539093, + 0.0775270015001297, + 0.09683650732040405, + 0.07444778829813004, + 0.059183090925216675, + 0.030850782990455627 + ], + "credit_map": [ + 0.05164548382163048, + 0.025893189013004303, + 0.046904273331165314, + 0.05661512911319733, + 0.03573538362979889, + 0.04096579924225807, + 0.036682914942502975, + 0.039939459413290024, + 0.01330634020268917, + 0.010134893469512463, + 0.012355023995041847, + 0.01131510455161333, + 0.021681955084204674, + 0.012185454368591309, + 0.010314291343092918, + 0.007311083376407623, + 0.004542024340480566, + 0.005656464025378227, + 0.004734197165817022, + 0.009922233410179615, + 0.004585827235132456, + 0.005581605713814497, + 0.007903221063315868, + 0.004130636341869831, + 0.014888275414705276, + 0.010016431100666523, + 0.014820553362369537, + 0.012635021470487118, + 0.017231736332178116, + 0.016242478042840958, + 0.01303472463041544, + 0.011677629314363003, + 0.009430444799363613, + 0.004730343818664551, + 0.0074708894826471806, + 0.011030049063265324, + 0.010079178027808666, + 0.011495831422507763, + 0.009203632362186909, + 0.007125647272914648, + 0.009077335707843304, + 0.006054941099137068, + 0.005637240130454302, + 0.012313829734921455, + 0.011483034119009972, + 0.010791021399199963, + 0.014080931432545185, + 0.008088666945695877, + 0.00972050242125988, + 0.005497874692082405, + 0.010299322195351124, + 0.019519835710525513, + 0.01212319079786539, + 0.014483513310551643, + 0.015230645425617695, + 0.009961622767150402, + 0.007936162874102592, + 0.0035427638795226812, + 0.00825587660074234, + 0.01665540598332882, + 0.005649416707456112, + 0.01000973116606474, + 0.013878493569791317, + 0.008519935421645641, + 0.0049078697338700294, + 0.003322661854326725, + 0.0054391720332205296, + 0.010444961488246918, + 0.00792966689914465, + 0.007133914157748222, + 0.013067532330751419, + 0.006937310099601746, + 0.003222169354557991, + 0.00258883205242455, + 0.003291459521278739, + 0.0041739377193152905, + 0.006956680677831173, + 0.005051447544246912, + 0.0029149807523936033, + 0.0026512749027460814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.10038083791732788 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot gazes down upon a sleepy, peaceful farm nestled in a valley beneath a vast night sky. A brilliant full moon hangs high, radiating a soft, silvery glow that gently illuminates the scene, creating long, softly edged shadows. Below, a classic two-story white farmhouse with a dark, gabled roof rests peacefully. A single upstairs window emits a warm, inviting golden light that pulses with a slow, subtle rhythm like a soft heartbeat. Beside it stands a large, traditional red barn, its silhouette a crimson crimson in the cool moonlight. The air is silent and filled with magic; countless fireflies lazily drift across the quiet pasture, their tiny lanterns flickering in slow, hypnotic synchronization. Pale, shimmering, ethereal points of light slowly descend from the deep blue sky, capturing the drifting moonlight and enveloping the landscape in a shimmering, magical dust, creating a profoundly tranquil, calming, bedtime atmosphere. Absolutely peaceful.", + "seed": 25844, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_9652_seed_25844_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08131922092288732, + "vlm_logprob": -0.08131922092288732, + "vlm_yes_no_margin": 2.824999976158142, + "vlm_token_credit": -0.08131922092288732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16334232687950134 + ], + "opposite_logprob_tokens": [ + -1.9133423566818237 + ], + "logprob_reward": -0.16334232687950134, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16334232687950134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35883471369743347, + 0.09706374257802963, + 0.15635590255260468, + 0.06924742460250854, + 0.06978166103363037, + 0.06791715323925018, + 0.03174780681729317, + 0.09949403256177902, + 0.01796475611627102, + 0.031592752784490585 + ], + "credit_map": [ + 0.02930748090147972, + 0.03653514012694359, + 0.0510701984167099, + 0.07438822090625763, + 0.03892248496413231, + 0.0346311517059803, + 0.046173956245183945, + 0.047806091606616974, + 0.01497946958988905, + 0.007723520044237375, + 0.008590372279286385, + 0.013515066355466843, + 0.025811271741986275, + 0.011404084973037243, + 0.00798292737454176, + 0.007057025097310543, + 0.029004164040088654, + 0.007932704873383045, + 0.009172279387712479, + 0.006302420515567064, + 0.056433066725730896, + 0.027171967551112175, + 0.009122454561293125, + 0.011216833256185055, + 0.013337444514036179, + 0.006075975019484758, + 0.006972090806812048, + 0.010000876151025295, + 0.012958796694874763, + 0.00570311164483428, + 0.006598032079637051, + 0.007601098157465458, + 0.012121567502617836, + 0.0050236196257174015, + 0.006055285222828388, + 0.00693141482770443, + 0.0177021361887455, + 0.00802155863493681, + 0.0073773423209786415, + 0.006548744160681963, + 0.007163648959249258, + 0.006162086967378855, + 0.007782282307744026, + 0.012919544242322445, + 0.007084183860570192, + 0.008396808989346027, + 0.009987620636820793, + 0.008420977741479874, + 0.003937267232686281, + 0.002539560664445162, + 0.0038509590085595846, + 0.004495606292039156, + 0.0052630966529250145, + 0.003803206840530038, + 0.00346503802575171, + 0.0043930732645094395, + 0.016728850081562996, + 0.006168101914227009, + 0.005425058770924807, + 0.004349754191935062, + 0.03193458542227745, + 0.01732899434864521, + 0.008866003714501858, + 0.008692680858075619, + 0.0019729137420654297, + 0.0013618252705782652, + 0.0022523102816194296, + 0.0015152192208915949, + 0.0039050995837897062, + 0.0025703588034957647, + 0.0022555782925337553, + 0.0021314495243132114, + 0.0034336159005761147, + 0.002701200544834137, + 0.004137981683015823, + 0.006201628129929304, + 0.0023865452967584133, + 0.0027062236331403255, + 0.004874539095908403, + 0.005151018965989351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.16334232687950134 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088653467595577 + ], + "opposite_logprob_tokens": [ + -4.511088848114014 + ], + "logprob_reward": -0.011088653467595577, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088653467595577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25785088539123535, + 0.0551908053457737, + 0.2649826109409332, + 0.05853278934955597, + 0.046494435518980026, + 0.07793150842189789, + 0.017788115888834, + 0.12498749792575836, + 0.03481166809797287, + 0.06142966449260712 + ], + "credit_map": [ + 0.03975620120763779, + 0.01106790080666542, + 0.01993612013757229, + 0.015804698690772057, + 0.07309642434120178, + 0.036388177424669266, + 0.027283690869808197, + 0.03451767936348915, + 0.005312480032444, + 0.004886434879153967, + 0.007530392613261938, + 0.006401747930794954, + 0.006831254810094833, + 0.006368773058056831, + 0.009291419759392738, + 0.008568307384848595, + 0.04795939475297928, + 0.012548291124403477, + 0.016094183549284935, + 0.012985368259251118, + 0.09553319215774536, + 0.04729936271905899, + 0.01580321043729782, + 0.01675965078175068, + 0.010851679369807243, + 0.004381537903100252, + 0.004859842360019684, + 0.006196413654834032, + 0.014132870361208916, + 0.00672850850969553, + 0.005683744791895151, + 0.005698196589946747, + 0.00817813165485859, + 0.004061616491526365, + 0.0027091859374195337, + 0.006406156811863184, + 0.012069925665855408, + 0.005057690665125847, + 0.0038205150049179792, + 0.00419122027233243, + 0.015136848203837872, + 0.004505414981395006, + 0.005093208514153957, + 0.004725167993456125, + 0.024860957637429237, + 0.01334890816360712, + 0.004953854251652956, + 0.0053071570582687855, + 0.0016066861571744084, + 0.0012686689151450992, + 0.002925074426457286, + 0.001777732395567, + 0.001733343000523746, + 0.002301532309502363, + 0.0030818814411759377, + 0.0030931977089494467, + 0.021311180666089058, + 0.006758754141628742, + 0.00757577596232295, + 0.009303772822022438, + 0.03986712545156479, + 0.019803443923592567, + 0.010647603310644627, + 0.009719847701489925, + 0.003944553434848785, + 0.0025193020701408386, + 0.004999089520424604, + 0.006044408772140741, + 0.005257896613329649, + 0.0033345541451126337, + 0.004532631020992994, + 0.004179235547780991, + 0.006877083797007799, + 0.004511017818003893, + 0.008928700350224972, + 0.009387561120092869, + 0.007286252919584513, + 0.008523806929588318, + 0.006703471764922142, + 0.009211769327521324 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 7 + ], + "selected_reward": -0.011088653467595577 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307687282562256 + ], + "opposite_logprob_tokens": [ + -3.168076992034912 + ], + "logprob_reward": -0.04307687282562256, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307687282562256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33675646781921387, + 0.11123579740524292, + 0.09647554904222488, + 0.05270959809422493, + 0.08026324957609177, + 0.03180646151304245, + 0.06184529885649681, + 0.15495990216732025, + 0.051166508346796036, + 0.022781167179346085 + ], + "credit_map": [ + 0.04018528014421463, + 0.030875016003847122, + 0.052520912140607834, + 0.04755654186010361, + 0.0631917268037796, + 0.04161837324500084, + 0.024831514805555344, + 0.03597712889313698, + 0.010447915643453598, + 0.010680877603590488, + 0.01843918114900589, + 0.022287975996732712, + 0.013238170184195042, + 0.00971070397645235, + 0.014513385482132435, + 0.011917591094970703, + 0.01455607358366251, + 0.005932468920946121, + 0.007274052128195763, + 0.008875982835888863, + 0.0292299073189497, + 0.01646970584988594, + 0.007444658316671848, + 0.006692706607282162, + 0.008603970520198345, + 0.005261268466711044, + 0.004866255447268486, + 0.006933043710887432, + 0.009975586086511612, + 0.005482405424118042, + 0.005796709563583136, + 0.005790357477962971, + 0.01483385544270277, + 0.006814191117882729, + 0.005911270622164011, + 0.007643102668225765, + 0.02163342386484146, + 0.008972126990556717, + 0.006410921458154917, + 0.008044353686273098, + 0.00428696908056736, + 0.0022702908609062433, + 0.002749767154455185, + 0.0044648414477705956, + 0.006837791297584772, + 0.004674070049077272, + 0.0034593725576996803, + 0.003063361858949065, + 0.010741675272583961, + 0.005634145811200142, + 0.00439411448314786, + 0.006744008045643568, + 0.0157296285033226, + 0.008333683013916016, + 0.005550907924771309, + 0.0047171334736049175, + 0.03132099285721779, + 0.010109142400324345, + 0.006101660430431366, + 0.005070332437753677, + 0.05876372754573822, + 0.02814721129834652, + 0.007594422437250614, + 0.007852411828935146, + 0.004535905551165342, + 0.0036023431457579136, + 0.008577445521950722, + 0.007307135500013828, + 0.005712945479899645, + 0.004324624314904213, + 0.0072262464091181755, + 0.009879866614937782, + 0.0021747578866779804, + 0.0012361378176137805, + 0.0038545546121895313, + 0.0020434632897377014, + 0.001975076273083687, + 0.0025219658855348825, + 0.004041305277496576, + 0.004933908116072416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.04307687282562256 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704668939113617 + ], + "opposite_logprob_tokens": [ + -2.127046585083008 + ], + "logprob_reward": -0.12704668939113617, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12704668939113617, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27659207582473755, + 0.09197036176919937, + 0.1358741968870163, + 0.04519738256931305, + 0.06534364074468613, + 0.05022582784295082, + 0.07561149448156357, + 0.1975993812084198, + 0.03242643177509308, + 0.02915921062231064 + ], + "credit_map": [ + 0.028987599536776543, + 0.035722486674785614, + 0.0491141639649868, + 0.021223751828074455, + 0.040576446801424026, + 0.03852074593305588, + 0.028109321370720863, + 0.03433755040168762, + 0.00997631810605526, + 0.01109201367944479, + 0.014999253675341606, + 0.015746820718050003, + 0.012279198504984379, + 0.007796141318976879, + 0.01084131933748722, + 0.00923929363489151, + 0.019216343760490417, + 0.007997936569154263, + 0.010899915359914303, + 0.013670427724719048, + 0.0377664677798748, + 0.020861882716417313, + 0.013453524559736252, + 0.012007695622742176, + 0.004916967824101448, + 0.004127738066017628, + 0.004217179957777262, + 0.009822606109082699, + 0.0037361590657383204, + 0.004367793444544077, + 0.007099553477019072, + 0.006909381132572889, + 0.01093260757625103, + 0.007336254697293043, + 0.004453521221876144, + 0.00818764790892601, + 0.01627657562494278, + 0.007370700128376484, + 0.0046746754087507725, + 0.006111653987318277, + 0.008539142087101936, + 0.003032990265637636, + 0.003556253621354699, + 0.004398211371153593, + 0.014687811955809593, + 0.008093987591564655, + 0.004072589334100485, + 0.003844841383397579, + 0.011989676393568516, + 0.004364559426903725, + 0.011449944227933884, + 0.006117234006524086, + 0.01424052007496357, + 0.007916942238807678, + 0.007951767183840275, + 0.01158085372298956, + 0.03334299102425575, + 0.010960090905427933, + 0.011770526878535748, + 0.015288310125470161, + 0.06189405918121338, + 0.03175922855734825, + 0.016270315274596214, + 0.016313860192894936, + 0.0037961776833981276, + 0.0018410459160804749, + 0.005817375611513853, + 0.0029078160878270864, + 0.005148364696651697, + 0.003949148580431938, + 0.0034577150363475084, + 0.005508788395673037, + 0.002925430191680789, + 0.0020548668690025806, + 0.004519534297287464, + 0.0037565568927675486, + 0.0020803262013942003, + 0.0029612442012876272, + 0.005246107466518879, + 0.005615146365016699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.12704668939113617 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204156205058098 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.06204156205058098, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204156205058098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24342124164104462, + 0.23606441915035248, + 0.06190824508666992, + 0.06588391959667206, + 0.06858114898204803, + 0.05573815107345581, + 0.055408407002687454, + 0.07423616200685501, + 0.07636284083127975, + 0.06239544600248337 + ], + "credit_map": [ + 0.03536628931760788, + 0.02079804800450802, + 0.03501594439148903, + 0.021909330040216446, + 0.055486150085926056, + 0.029636317864060402, + 0.018796775490045547, + 0.026412397623062134, + 0.039437759667634964, + 0.02603565715253353, + 0.02382403425872326, + 0.023195432499051094, + 0.06555131822824478, + 0.027712609618902206, + 0.01577458344399929, + 0.014533012174069881, + 0.006708814762532711, + 0.00898111704736948, + 0.007710323669016361, + 0.007175957318395376, + 0.008391408249735832, + 0.0074457088485360146, + 0.00683924974873662, + 0.008655667304992676, + 0.011210640892386436, + 0.004450756590813398, + 0.005755257792770863, + 0.007592252921313047, + 0.009719454683363438, + 0.006371764931827784, + 0.010629965923726559, + 0.010153829120099545, + 0.00825929082930088, + 0.009885305538773537, + 0.00727328285574913, + 0.011328503489494324, + 0.007767244707792997, + 0.008375845849514008, + 0.006044394336640835, + 0.009647284634411335, + 0.00920557789504528, + 0.002808244898915291, + 0.003606179030612111, + 0.0034682522527873516, + 0.01632745750248432, + 0.010586793534457684, + 0.004573279991745949, + 0.005162366665899754, + 0.009220228530466557, + 0.0031984238885343075, + 0.007242538966238499, + 0.0035569067113101482, + 0.012208141386508942, + 0.006538657937198877, + 0.006467963103204966, + 0.006975547410547733, + 0.011049406602978706, + 0.00403862539678812, + 0.0051907747983932495, + 0.006758982315659523, + 0.020577101036906242, + 0.010455302894115448, + 0.007235778495669365, + 0.00893019512295723, + 0.01156027801334858, + 0.005226777866482735, + 0.00771962758153677, + 0.004327370319515467, + 0.022647274658083916, + 0.012845009565353394, + 0.005411574617028236, + 0.006624926347285509, + 0.006155503448098898, + 0.0024500801227986813, + 0.009860705584287643, + 0.00379119161516428, + 0.011038611643016338, + 0.008708723820745945, + 0.006531865336000919, + 0.013858761638402939 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.06204156205058098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, high-angle, wide-angle shot gazes down upon a sleepy, peaceful farm nestled in a valley beneath a vast night sky. A brilliant full moon hangs high, radiating a soft, silvery glow that gently illuminates the scene, creating long, softly edged shadows. Below, a classic two-story white farmhouse with a dark, gabled roof rests peacefully. A single upstairs window emits a warm, inviting golden light that pulses with a slow, subtle rhythm like a soft heartbeat. Beside it stands a large, traditional red barn, its silhouette a crimson crimson in the cool moonlight. The air is silent and filled with magic; countless fireflies lazily drift across the quiet pasture, their tiny lanterns flickering in slow, hypnotic synchronization. Pale, shimmering, ethereal points of light slowly descend from the deep blue sky, capturing the drifting moonlight and enveloping the landscape in a shimmering, magical dust, creating a profoundly tranquil, calming, bedtime atmosphere. Absolutely peaceful.", + "seed": 25845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_5/index_9652_seed_25845_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19492256455123425, + "vlm_logprob": -0.19492256455123425, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.19492256455123425, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7592383623123169 + ], + "opposite_logprob_tokens": [ + -0.6342383623123169 + ], + "logprob_reward": -0.7592383623123169, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7592383623123169, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.434950053691864, + 0.06945858150720596, + 0.07412410527467728, + 0.02835662290453911, + 0.15115375816822052, + 0.03647821024060249, + 0.08812420070171356, + 0.04873348027467728, + 0.02954561822116375, + 0.03907538205385208 + ], + "credit_map": [ + 0.04093843325972557, + 0.027028463780879974, + 0.0624883770942688, + 0.0738535150885582, + 0.05047610402107239, + 0.05896356329321861, + 0.044541001319885254, + 0.07666060328483582, + 0.007775047328323126, + 0.004751979373395443, + 0.011653109453618526, + 0.01203578244894743, + 0.005453647579997778, + 0.008246012032032013, + 0.007403661962598562, + 0.012139342725276947, + 0.009035159833729267, + 0.0071336147375404835, + 0.01076157484203577, + 0.0069644576869904995, + 0.012133331969380379, + 0.011865218169987202, + 0.006331521086394787, + 0.009899226017296314, + 0.004391455557197332, + 0.0022949487902224064, + 0.0030003110878169537, + 0.0030671078711748123, + 0.003808120032772422, + 0.0044233305379748344, + 0.0035706344060599804, + 0.0038007127586752176, + 0.01027480699121952, + 0.006998125929385424, + 0.009389422833919525, + 0.010315665043890476, + 0.04490446299314499, + 0.049720942974090576, + 0.009462039917707443, + 0.010088291019201279, + 0.0030734860338270664, + 0.00221634772606194, + 0.005220708902925253, + 0.007084984332323074, + 0.003578688483685255, + 0.004607710521668196, + 0.006226005032658577, + 0.004470281768590212, + 0.010007860139012337, + 0.006291916128247976, + 0.010452566668391228, + 0.016109507530927658, + 0.009288661181926727, + 0.013362028636038303, + 0.011364881880581379, + 0.01124678086489439, + 0.003889335785061121, + 0.003389087039977312, + 0.004741073120385408, + 0.010029242374002934, + 0.004755871370434761, + 0.005377184599637985, + 0.008903423324227333, + 0.00764826126396656, + 0.0026338265743106604, + 0.002322058193385601, + 0.0020862261299043894, + 0.0038220242131501436, + 0.006092512980103493, + 0.00472738454118371, + 0.004318824037909508, + 0.003542762016877532, + 0.003570303786545992, + 0.0033224013168364763, + 0.0037350852508097887, + 0.003561164252460003, + 0.010088524781167507, + 0.007844941690564156, + 0.0028471630066633224, + 0.004105799365788698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.7592383623123169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332126535475254 + ], + "opposite_logprob_tokens": [ + -3.7733211517333984 + ], + "logprob_reward": -0.02332126535475254, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332126535475254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36302000284194946, + 0.10185068100690842, + 0.05680672451853752, + 0.03635043278336525, + 0.16776436567306519, + 0.043337006121873856, + 0.06641456484794617, + 0.08600813150405884, + 0.034429632127285004, + 0.0440184623003006 + ], + "credit_map": [ + 0.02936013974249363, + 0.017149293795228004, + 0.04007948935031891, + 0.05645247921347618, + 0.06262126564979553, + 0.0766001045703888, + 0.03613820672035217, + 0.04461906477808952, + 0.008691519498825073, + 0.006548379547894001, + 0.010949299670755863, + 0.017604108899831772, + 0.014498194679617882, + 0.021343722939491272, + 0.010701697319746017, + 0.01151376310735941, + 0.0056085772812366486, + 0.006065433379262686, + 0.007367903832346201, + 0.008194110356271267, + 0.00788150168955326, + 0.008956561796367168, + 0.0063370605930686, + 0.006395580247044563, + 0.003983158618211746, + 0.0037296803202480078, + 0.005462584085762501, + 0.003948940895497799, + 0.004688769578933716, + 0.00475516589358449, + 0.004738642834126949, + 0.005043493118137121, + 0.010387708432972431, + 0.009448456577956676, + 0.010523635894060135, + 0.014365799725055695, + 0.04771996662020683, + 0.051940832287073135, + 0.014683792367577553, + 0.008694179356098175, + 0.003675581654533744, + 0.005362485535442829, + 0.006163109093904495, + 0.00780281238257885, + 0.005400032736361027, + 0.005772764328867197, + 0.005627369973808527, + 0.0035328541416674852, + 0.006881895940750837, + 0.006124784238636494, + 0.005612457171082497, + 0.0091583002358675, + 0.01237290259450674, + 0.009799771942198277, + 0.008939425460994244, + 0.007525033317506313, + 0.009675105102360249, + 0.00800107792019844, + 0.008987518958747387, + 0.011184310540556908, + 0.014507889747619629, + 0.014332116581499577, + 0.008789703249931335, + 0.010530410334467888, + 0.003668868914246559, + 0.00433077709749341, + 0.0020252650137990713, + 0.0032455092296004295, + 0.00841293204575777, + 0.00644593546167016, + 0.0028523877263069153, + 0.0034479631576687098, + 0.003881704295054078, + 0.004567095544189215, + 0.00520154507830739, + 0.006335223559290171, + 0.008553871884942055, + 0.006601125933229923, + 0.004497678950428963, + 0.0043802205473184586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02332126535475254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867987334728241 + ], + "opposite_logprob_tokens": [ + -3.048679828643799 + ], + "logprob_reward": -0.04867987334728241, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867987334728241, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23235385119915009, + 0.06311966478824615, + 0.12347988039255142, + 0.07043240964412689, + 0.24925722181797028, + 0.04432852566242218, + 0.0765235498547554, + 0.037631191313266754, + 0.08797110617160797, + 0.014902600087225437 + ], + "credit_map": [ + 0.02154008112847805, + 0.016164008527994156, + 0.025572728365659714, + 0.023453541100025177, + 0.05118172988295555, + 0.05271216854453087, + 0.016317611560225487, + 0.025412002578377724, + 0.007995794527232647, + 0.006327536888420582, + 0.00979312602430582, + 0.008243937976658344, + 0.007708415389060974, + 0.008086959831416607, + 0.0060156118124723434, + 0.008948281407356262, + 0.019266752526164055, + 0.010724679566919804, + 0.011710443533957005, + 0.007561801467090845, + 0.02694665640592575, + 0.0314214788377285, + 0.006731226108968258, + 0.009116848930716515, + 0.009052070789039135, + 0.006401583086699247, + 0.011271536350250244, + 0.006041940301656723, + 0.009581845253705978, + 0.00969678070396185, + 0.007340929936617613, + 0.011045721359550953, + 0.017639316618442535, + 0.014029753394424915, + 0.012073489837348461, + 0.008779187686741352, + 0.08446795493364334, + 0.09263882040977478, + 0.008492711000144482, + 0.011135982349514961, + 0.004404581151902676, + 0.007054314482957125, + 0.006048022303730249, + 0.004844362847507, + 0.0053776828572154045, + 0.006182109005749226, + 0.00514443963766098, + 0.005273015238344669, + 0.009548504836857319, + 0.006400784477591515, + 0.009407705627381802, + 0.00772347766906023, + 0.012885486707091331, + 0.015559105202555656, + 0.0075106434524059296, + 0.007487847004085779, + 0.0035207809414714575, + 0.004272816702723503, + 0.0044095758348703384, + 0.004039690364152193, + 0.0064916545525193214, + 0.007540292572230101, + 0.003754587844014168, + 0.003601795295253396, + 0.007894221693277359, + 0.005154504906386137, + 0.0056876325979828835, + 0.006364369764924049, + 0.02925463765859604, + 0.021337172016501427, + 0.00565630616620183, + 0.0066222697496414185, + 0.0019624673295766115, + 0.0017746640369296074, + 0.0024850827176123857, + 0.0015661179786548018, + 0.001920123933814466, + 0.00196444452740252, + 0.0014047864824533463, + 0.0018249140121042728 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.04867987334728241 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1003093495965004 + ], + "opposite_logprob_tokens": [ + -2.350309371948242 + ], + "logprob_reward": -0.1003093495965004, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.1003093495965004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2737499475479126, + 0.08126252889633179, + 0.22553564608097076, + 0.037756308913230896, + 0.10299073159694672, + 0.03215135261416435, + 0.14147400856018066, + 0.02062237448990345, + 0.0705876424908638, + 0.013869508169591427 + ], + "credit_map": [ + 0.02807754836976528, + 0.019073130562901497, + 0.02957029640674591, + 0.02057502605021, + 0.06322770565748215, + 0.0695265606045723, + 0.015994926914572716, + 0.027704700827598572, + 0.011538897641003132, + 0.0073143173940479755, + 0.01153171993792057, + 0.009081228636205196, + 0.01049415860325098, + 0.013412969186902046, + 0.007473287638276815, + 0.01041593961417675, + 0.03781348839402199, + 0.019504046067595482, + 0.022080285474658012, + 0.013022756204009056, + 0.04659458249807358, + 0.05715559050440788, + 0.009483519941568375, + 0.019881362095475197, + 0.004656501580029726, + 0.004260899033397436, + 0.006241531111299992, + 0.0038138525560498238, + 0.005037114955484867, + 0.005779069382697344, + 0.003150914330035448, + 0.0048164245672523975, + 0.007114045321941376, + 0.0067451111972332, + 0.007827211171388626, + 0.004892611410468817, + 0.03202996030449867, + 0.03237275779247284, + 0.004544274415820837, + 0.007464749738574028, + 0.003373290877789259, + 0.00597754679620266, + 0.004473349545150995, + 0.0024490319192409515, + 0.0046086241491138935, + 0.005660600494593382, + 0.002197178080677986, + 0.0034117286559194326, + 0.01596236415207386, + 0.014682065695524216, + 0.017071539536118507, + 0.011707707308232784, + 0.025537915527820587, + 0.029943840578198433, + 0.011058172211050987, + 0.015510400757193565, + 0.002035647165030241, + 0.002182696247473359, + 0.002439247677102685, + 0.0019959970377385616, + 0.003317553084343672, + 0.004140560049563646, + 0.001917214598506689, + 0.0025934576988220215, + 0.005859080236405134, + 0.0041653006337583065, + 0.004305427893996239, + 0.0038487983401864767, + 0.025086645036935806, + 0.01901973970234394, + 0.003944241441786289, + 0.004358398262411356, + 0.0019925341475754976, + 0.001486776047386229, + 0.0018588289385661483, + 0.001587224774993956, + 0.001989344833418727, + 0.002585787558928132, + 0.0009965798817574978, + 0.0013724302407354116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.1003093495965004 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene a classic two-story white farmhouse with a dark gabled roof?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a large traditional red barn visible beside the farmhouse?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a single upstairs window emit a warm golden light that pulses with a slow rhythm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are countless fireflies drifting across the pasture with flickering lights?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do pale shimmering points of light descend from the sky to envelop the landscape?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043063972145318985 + ], + "opposite_logprob_tokens": [ + -3.1680638790130615 + ], + "logprob_reward": -0.043063972145318985, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043063972145318985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26582029461860657, + 0.06598635017871857, + 0.1581583470106125, + 0.026493581011891365, + 0.2004225254058838, + 0.04033273458480835, + 0.06547525525093079, + 0.023360515013337135, + 0.11921797692775726, + 0.03473241627216339 + ], + "credit_map": [ + 0.02816477231681347, + 0.02712702751159668, + 0.023442614823579788, + 0.020700640976428986, + 0.06200471520423889, + 0.06851261109113693, + 0.01597725972533226, + 0.01989065669476986, + 0.008203374221920967, + 0.008489524014294147, + 0.006591833662241697, + 0.00955413281917572, + 0.009343023411929607, + 0.01015844102948904, + 0.005880488082766533, + 0.007765533868223429, + 0.027091313153505325, + 0.014336619526147842, + 0.01238152664154768, + 0.007164157927036285, + 0.0351342149078846, + 0.04519442096352577, + 0.005892858374863863, + 0.010963235050439835, + 0.002701158169656992, + 0.0028546855319291353, + 0.0038896584883332253, + 0.0026069616433233023, + 0.0033371103927493095, + 0.0038225611206144094, + 0.0032582897692918777, + 0.00402315566316247, + 0.01204669289290905, + 0.007306965533643961, + 0.009673514403402805, + 0.007483983412384987, + 0.07068727910518646, + 0.07518204301595688, + 0.007730297744274139, + 0.010311750695109367, + 0.0044571952894330025, + 0.005271072033792734, + 0.004630363546311855, + 0.0042935325764119625, + 0.007496279664337635, + 0.00722457142546773, + 0.0032600590493530035, + 0.003699658904224634, + 0.0081540048122406, + 0.00603882409632206, + 0.008911440148949623, + 0.006476485636085272, + 0.011984985321760178, + 0.010913465172052383, + 0.005906144622713327, + 0.007089908700436354, + 0.0029259929433465004, + 0.0028517472092062235, + 0.0037323348224163055, + 0.002016856800764799, + 0.003128127660602331, + 0.0035014820750802755, + 0.002052886877208948, + 0.003151086624711752, + 0.011994794011116028, + 0.016790231689810753, + 0.009136687964200974, + 0.005857209675014019, + 0.03285173699259758, + 0.02847808040678501, + 0.00512767443433404, + 0.00898156501352787, + 0.003897323738783598, + 0.002885763766244054, + 0.0029199188575148582, + 0.002421795856207609, + 0.009551194496452808, + 0.007529116701334715, + 0.0026988366153091192, + 0.002828466473147273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 2 + ], + "selected_reward": -0.043063972145318985 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_tensor([2529, 2529])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_tensor([2529, 2529])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..da4d06281fa9f24305e213b5b124941330127514 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_tensor([2529, 2529])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "high-angle, wide-angle shot, rendered in a high-definition, surreal anime-style with backgrounds reminiscent of the works of Hayao Miyazaki, slowly pans right, tracking the catastrophic aftermath of a land-based earthquake. The camera gazes down upon a shattered valley, where a massive mud avalanche—a sticky, roiling mass of liquefied earth and dark, muddy debris—seeps relentlessly forward. Its surface is a chaotic tapestry of thick brown mud and gray silt, shimmering with a damp sheen under the sweltering, sun-bleached sky, a clear sign of intense surface warming. The mudslide's front, slow but unstoppable, its jagged front deliberately engulfs majestic pine trees, bending, splitting, and then swallowing them whole. The shattered remains of wooden structures, their timbers splintered and damaged, are devoured by the relentless current. Wisps of steam rise continuously from the sticky, roiling mass, which bubbles and moves as if alive, creating a terrifying and awe-inspiring atmosphere that showcases the raw, destructive power of nature.", + "seed": 16085, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16085_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08543903799727559, + "vlm_logprob": -0.08543903799727559, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.08543903799727559, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12651847302913666 + ], + "opposite_logprob_tokens": [ + -2.626518487930298 + ], + "logprob_reward": -0.12651847302913666, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.12651847302913666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1090184822678566, + 0.0529111847281456, + 0.15524961054325104, + 0.034761734306812286, + 0.08738412708044052, + 0.04973824694752693, + 0.03164559602737427, + 0.03496384993195534, + 0.397566020488739, + 0.046761106699705124 + ], + "credit_map": [ + 0.01729581132531166, + 0.010343031957745552, + 0.011978860013186932, + 0.009485218673944473, + 0.022171707823872566, + 0.02024879679083824, + 0.009061249904334545, + 0.008433803915977478, + 0.008900822140276432, + 0.005153718404471874, + 0.004581058397889137, + 0.00744983134791255, + 0.0057267011143267155, + 0.006234512664377689, + 0.006998930126428604, + 0.007865610532462597, + 0.013591698370873928, + 0.007110310718417168, + 0.0029655257239937782, + 0.007895514369010925, + 0.008217732422053814, + 0.059896320104599, + 0.005396381486207247, + 0.05017613619565964, + 0.004068918526172638, + 0.0043017505668103695, + 0.0038909653667360544, + 0.005617208778858185, + 0.003945549018681049, + 0.0037361495196819305, + 0.005606199614703655, + 0.00359499454498291, + 0.007486050482839346, + 0.008373996242880821, + 0.01036582887172699, + 0.01481853611767292, + 0.011456371285021305, + 0.01212327741086483, + 0.014321702532470226, + 0.008438358083367348, + 0.004448447376489639, + 0.0076320054940879345, + 0.006572021637111902, + 0.004114450886845589, + 0.006684804335236549, + 0.005623531527817249, + 0.0072663817554712296, + 0.00739660020917654, + 0.0032207434996962547, + 0.0036589186638593674, + 0.0036048279143869877, + 0.00429678987711668, + 0.00437289709225297, + 0.004334882367402315, + 0.004546046257019043, + 0.0036104884929955006, + 0.005085499957203865, + 0.004263935144990683, + 0.0034836474806070328, + 0.005488472059369087, + 0.003563082544133067, + 0.004075687378644943, + 0.0047289966605603695, + 0.00427452614530921, + 0.015122747048735619, + 0.006803319323807955, + 0.006582571659237146, + 0.015875816345214844, + 0.017221717163920403, + 0.01558275893330574, + 0.17206430435180664, + 0.1483127921819687, + 0.004546486306935549, + 0.004372281022369862, + 0.0029656984843313694, + 0.010874011553823948, + 0.006189075298607349, + 0.003748158924281597, + 0.007452024612575769, + 0.006613373290747404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.12651847302913666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038252413272858 + ], + "opposite_logprob_tokens": [ + -1.9103825092315674 + ], + "logprob_reward": -0.16038252413272858, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038252413272858, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2216261476278305, + 0.0578317791223526, + 0.2854805588722229, + 0.07630860060453415, + 0.10616646707057953, + 0.02502809651196003, + 0.02029525861144066, + 0.028600402176380157, + 0.10315060615539551, + 0.07551205158233643 + ], + "credit_map": [ + 0.03525923565030098, + 0.020263012498617172, + 0.028596602380275726, + 0.048451051115989685, + 0.017624279484152794, + 0.017411019653081894, + 0.026124505326151848, + 0.027896439656615257, + 0.011026527732610703, + 0.005916704423725605, + 0.004334405995905399, + 0.0042998599819839, + 0.006063222885131836, + 0.007544675376266241, + 0.005485753994435072, + 0.01316063292324543, + 0.03209211304783821, + 0.031503621488809586, + 0.011562278494238853, + 0.012915944680571556, + 0.016792690381407738, + 0.10254895687103271, + 0.014229809865355492, + 0.06383515894412994, + 0.006504544056952, + 0.007681012153625488, + 0.006656331941485405, + 0.007912348955869675, + 0.007764633744955063, + 0.007656139321625233, + 0.022103620693087578, + 0.010029972530901432, + 0.010345378890633583, + 0.00896189734339714, + 0.01344060618430376, + 0.01419183798134327, + 0.016892630606889725, + 0.011771521531045437, + 0.020950041711330414, + 0.00961255468428135, + 0.0022568784188479185, + 0.0022966377437114716, + 0.0036342910025268793, + 0.002606486901640892, + 0.005104737356305122, + 0.002428169595077634, + 0.0034070087131112814, + 0.0032938867807388306, + 0.002432694658637047, + 0.002222304930910468, + 0.0025953026488423347, + 0.002103628357872367, + 0.002103769686073065, + 0.0025939333718270063, + 0.003590781707316637, + 0.0026528439484536648, + 0.0037211300805211067, + 0.004227004945278168, + 0.003484501736238599, + 0.004809703212231398, + 0.0035370790865272284, + 0.002385759027674794, + 0.002961284015327692, + 0.0034739405382424593, + 0.004834514576941729, + 0.005919551942497492, + 0.00215062964707613, + 0.004634312354028225, + 0.004875401966273785, + 0.004554969724267721, + 0.0406312420964241, + 0.035549990832805634, + 0.013426839374005795, + 0.009042088873684406, + 0.0065293204970657825, + 0.01341331098228693, + 0.009019395336508751, + 0.006273035425692797, + 0.00865914486348629, + 0.009148913435637951 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.16038252413272858 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08066005259752274, + 0.016572751104831696, + 0.12493348866701126, + 0.022575894370675087, + 0.04886658489704132, + 0.09447789937257767, + 0.028146151453256607, + 0.018932251259684563, + 0.5325443148612976, + 0.032290633767843246 + ], + "credit_map": [ + 0.01205041166394949, + 0.008497077040374279, + 0.006737954448908567, + 0.006236362271010876, + 0.012621133588254452, + 0.016022948548197746, + 0.006621379870921373, + 0.01187279261648655, + 0.002219386165961623, + 0.001286263344809413, + 0.0023968741297721863, + 0.002145929727703333, + 0.0015658184420317411, + 0.002296577673405409, + 0.00241908710449934, + 0.0022428142838180065, + 0.011635255999863148, + 0.005961198825389147, + 0.005835290998220444, + 0.006587803363800049, + 0.006560355890542269, + 0.042580537497997284, + 0.006231843959540129, + 0.039541199803352356, + 0.002090246882289648, + 0.00210904935374856, + 0.002514165360480547, + 0.00384891452267766, + 0.0017662806203588843, + 0.0021011470817029476, + 0.005430118180811405, + 0.002715972950682044, + 0.004455601796507835, + 0.0062554883770644665, + 0.00561285438016057, + 0.008173233829438686, + 0.006436550058424473, + 0.004988668020814657, + 0.007369592785835266, + 0.00557459844276309, + 0.008359326981008053, + 0.014890887774527073, + 0.010594340041279793, + 0.008810240775346756, + 0.01683136075735092, + 0.009195289574563503, + 0.013611851260066032, + 0.01218460500240326, + 0.00243938947096467, + 0.0038269790820777416, + 0.0023533448111265898, + 0.0035721920430660248, + 0.0028464780189096928, + 0.0027983500622212887, + 0.006674216128885746, + 0.0036352009046822786, + 0.0024134970735758543, + 0.0026655523106455803, + 0.0021902360022068024, + 0.0021542301401495934, + 0.0018780373502522707, + 0.0017975273076444864, + 0.0032702984753996134, + 0.0025628714356571436, + 0.02025008201599121, + 0.011388192884624004, + 0.007865524850785732, + 0.022151634097099304, + 0.02313624694943428, + 0.02179444395005703, + 0.2289777398109436, + 0.19698046147823334, + 0.004248855169862509, + 0.003964602015912533, + 0.0024286950938403606, + 0.006247582379728556, + 0.003930631559342146, + 0.0027055994141846895, + 0.005645124241709709, + 0.003119546687230468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 5 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008639576844871044 + ], + "opposite_logprob_tokens": [ + -4.758639812469482 + ], + "logprob_reward": -0.008639576844871044, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008639576844871044, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042626939713954926, + 0.02573373168706894, + 0.24080203473567963, + 0.016085414215922356, + 0.026779169216752052, + 0.3190261721611023, + 0.02449915185570717, + 0.013959914445877075, + 0.26794806122779846, + 0.022539479658007622 + ], + "credit_map": [ + 0.007081679534167051, + 0.004247174598276615, + 0.0092817647382617, + 0.005708483979105949, + 0.003327391343191266, + 0.0038201750721782446, + 0.005127863492816687, + 0.004032401833683252, + 0.004166821017861366, + 0.001657211105339229, + 0.0038832665886729956, + 0.003920526709407568, + 0.0018719900399446487, + 0.0027651493437588215, + 0.0029234199319034815, + 0.004545347299426794, + 0.022665077820420265, + 0.01774604059755802, + 0.009463366121053696, + 0.010302991606295109, + 0.01376095786690712, + 0.09294254332780838, + 0.010934458114206791, + 0.0629865825176239, + 0.001985431183129549, + 0.0018611494451761246, + 0.001804050407372415, + 0.0015687348786741495, + 0.001972000114619732, + 0.0017987322062253952, + 0.003216648241505027, + 0.0018786671571433544, + 0.0025691301561892033, + 0.0022116994950920343, + 0.0037375858519226313, + 0.0030557201243937016, + 0.0039049764163792133, + 0.0030584416817873716, + 0.005220857448875904, + 0.0030207554809749126, + 0.0314696729183197, + 0.03996409475803375, + 0.034566257148981094, + 0.024176282808184624, + 0.0832018107175827, + 0.03644115850329399, + 0.03428883105516434, + 0.03491806983947754, + 0.0022462180349975824, + 0.003659686306491494, + 0.0023640866857022047, + 0.002962562022730708, + 0.0024551854003220797, + 0.0027397763915359974, + 0.005133653525263071, + 0.0029379823245108128, + 0.001721312291920185, + 0.002668938133865595, + 0.0013105215039104223, + 0.0019320358987897635, + 0.0012279603397473693, + 0.0012327485019341111, + 0.0022847082000225782, + 0.0015816889936104417, + 0.010062589310109615, + 0.004160420037806034, + 0.004833471029996872, + 0.011880097910761833, + 0.011637411080300808, + 0.010814699344336987, + 0.11496315151453018, + 0.09959621727466583, + 0.0027519697323441505, + 0.0018633638974279165, + 0.0016070306301116943, + 0.005302022211253643, + 0.003138904692605138, + 0.0021150107495486736, + 0.0034813638776540756, + 0.0022798110730946064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 2 + ], + "selected_reward": -0.008639576844871044 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270153969526291 + ], + "opposite_logprob_tokens": [ + -2.1270153522491455 + ], + "logprob_reward": -0.1270153969526291, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270153969526291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19417938590049744, + 0.07434811443090439, + 0.08094695955514908, + 0.0524965263903141, + 0.12152833491563797, + 0.0583941675722599, + 0.09398350864648819, + 0.1654917150735855, + 0.09501785784959793, + 0.0636134073138237 + ], + "credit_map": [ + 0.02332833968102932, + 0.017822327092289925, + 0.02504105120897293, + 0.03877156972885132, + 0.014170032925903797, + 0.020748648792505264, + 0.02863406389951706, + 0.025663353502750397, + 0.011399922892451286, + 0.004720813129097223, + 0.008042802102863789, + 0.008553351275622845, + 0.007897992618381977, + 0.01157365646213293, + 0.006796186789870262, + 0.015363388694822788, + 0.010510814376175404, + 0.013136530295014381, + 0.008766492828726768, + 0.007867292501032352, + 0.005407882388681173, + 0.018237311393022537, + 0.008319167420268059, + 0.008701478131115437, + 0.006001187488436699, + 0.007625902071595192, + 0.0037297923117876053, + 0.004922113381326199, + 0.006016251631081104, + 0.005672334227710962, + 0.012428967282176018, + 0.006099978927522898, + 0.010315431281924248, + 0.013020996935665607, + 0.011660097166895866, + 0.013240884989500046, + 0.022488823160529137, + 0.013853057287633419, + 0.024166090413928032, + 0.01278295461088419, + 0.0058617014437913895, + 0.006558190565556288, + 0.007858896628022194, + 0.005896636284887791, + 0.013362809084355831, + 0.00610311096534133, + 0.005870661698281765, + 0.006882160902023315, + 0.007935913279652596, + 0.009374091401696205, + 0.010391839779913425, + 0.014003930613398552, + 0.011801246553659439, + 0.016209255903959274, + 0.013527289032936096, + 0.01073994766920805, + 0.01763075590133667, + 0.023959165439009666, + 0.022509967908263206, + 0.028022393584251404, + 0.015629924833774567, + 0.015003914944827557, + 0.023924928158521652, + 0.018810652196407318, + 0.007736800238490105, + 0.011659386567771435, + 0.00450042262673378, + 0.007990673184394836, + 0.0050870925188064575, + 0.005791292991489172, + 0.028844676911830902, + 0.023407520726323128, + 0.006983238272368908, + 0.00923230405896902, + 0.005151801276952028, + 0.00929256621748209, + 0.007868314161896706, + 0.0062764957547187805, + 0.010537280701100826, + 0.008271410129964352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.1270153969526291 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot, rendered in a high-definition, surreal anime-style with backgrounds reminiscent of the works of Hayao Miyazaki, slowly pans right, tracking the catastrophic aftermath of a land-based earthquake. The camera gazes down upon a shattered valley, where a massive mud avalanche—a sticky, roiling mass of liquefied earth and dark, muddy debris—seeps relentlessly forward. Its surface is a chaotic tapestry of thick brown mud and gray silt, shimmering with a damp sheen under the sweltering, sun-bleached sky, a clear sign of intense surface warming. The mudslide's front, slow but unstoppable, its jagged front deliberately engulfs majestic pine trees, bending, splitting, and then swallowing them whole. The shattered remains of wooden structures, their timbers splintered and damaged, are devoured by the relentless current. Wisps of steam rise continuously from the sticky, roiling mass, which bubbles and moves as if alive, creating a terrifying and awe-inspiring atmosphere that showcases the raw, destructive power of nature.", + "seed": 16086, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16086_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0731430706102401, + "vlm_logprob": -0.0731430706102401, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.0731430706102401, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11787988245487213 + ], + "opposite_logprob_tokens": [ + -2.492879867553711 + ], + "logprob_reward": -0.11787988245487213, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.11787988245487213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16815853118896484, + 0.06986197084188461, + 0.07932911068201065, + 0.042498402297496796, + 0.05568060651421547, + 0.07531630247831345, + 0.21375980973243713, + 0.11862394958734512, + 0.1493694931268692, + 0.02740180492401123 + ], + "credit_map": [ + 0.024804716929793358, + 0.039422258734703064, + 0.014514986425638199, + 0.013712603598833084, + 0.030688468366861343, + 0.021539971232414246, + 0.013066297397017479, + 0.010409231297671795, + 0.007752655074000359, + 0.012291953898966312, + 0.008707822300493717, + 0.008834154345095158, + 0.009457514621317387, + 0.00875060260295868, + 0.007954458706080914, + 0.006112807895988226, + 0.007962812669575214, + 0.012035226449370384, + 0.005150972865521908, + 0.009193800389766693, + 0.01953118108212948, + 0.010697231628000736, + 0.008891071192920208, + 0.005866811145097017, + 0.003711194032803178, + 0.004433916416019201, + 0.006870510522276163, + 0.0054746284149587154, + 0.0046209716238081455, + 0.0067705800756812096, + 0.005027547478675842, + 0.005589051637798548, + 0.006289704702794552, + 0.00633825920522213, + 0.006667166016995907, + 0.007759351283311844, + 0.007334774825721979, + 0.008325794711709023, + 0.007935973815619946, + 0.005029584281146526, + 0.012348750606179237, + 0.013608487322926521, + 0.0088731050491333, + 0.006243634037673473, + 0.012170357629656792, + 0.007407095283269882, + 0.007790285162627697, + 0.00687458785250783, + 0.037468865513801575, + 0.014407969079911709, + 0.011733303777873516, + 0.008995650336146355, + 0.08695309609174728, + 0.01833087019622326, + 0.010553303174674511, + 0.025316748768091202, + 0.012401733547449112, + 0.011193829588592052, + 0.024671323597431183, + 0.01206221804022789, + 0.02731257863342762, + 0.01286963652819395, + 0.009271604008972645, + 0.008841026574373245, + 0.020708946511149406, + 0.03425362706184387, + 0.01910492032766342, + 0.012226286344230175, + 0.012872499413788319, + 0.016376757994294167, + 0.022063039243221283, + 0.011763427406549454, + 0.003232878167182207, + 0.0024674555752426386, + 0.003253965638577938, + 0.004214020445942879, + 0.003162842011079192, + 0.004501470364630222, + 0.0034974708687514067, + 0.0030717020854353905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.11787988245487213 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297706514596939 + ], + "opposite_logprob_tokens": [ + -2.2379770278930664 + ], + "logprob_reward": -0.11297706514596939, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297706514596939, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30462646484375, + 0.14446094632148743, + 0.06509897112846375, + 0.017268721014261246, + 0.08966129273176193, + 0.05575849115848541, + 0.1046748012304306, + 0.09322814643383026, + 0.08154060691595078, + 0.04368158057332039 + ], + "credit_map": [ + 0.04645560309290886, + 0.06862066686153412, + 0.022598639130592346, + 0.024557380005717278, + 0.05828719213604927, + 0.03146262839436531, + 0.03451763093471527, + 0.018126701936125755, + 0.01879345066845417, + 0.028090180829167366, + 0.012490167282521725, + 0.013433529064059258, + 0.03098723106086254, + 0.017115192487835884, + 0.01297163125127554, + 0.010579561814665794, + 0.0066145481541752815, + 0.010425460524857044, + 0.006442658603191376, + 0.007506653666496277, + 0.009402144700288773, + 0.013283924199640751, + 0.004169366788119078, + 0.007254213560372591, + 0.001715975347906351, + 0.002859283471480012, + 0.0016616034554317594, + 0.0026390941347926855, + 0.0018640607595443726, + 0.0018315364141017199, + 0.0019472303101792932, + 0.002749938052147627, + 0.00988792721182108, + 0.009797092527151108, + 0.008114761672914028, + 0.01121531892567873, + 0.01396517176181078, + 0.018660632893443108, + 0.011150804348289967, + 0.006869582924991846, + 0.011629919521510601, + 0.00714480597525835, + 0.005383965093642473, + 0.0038084592670202255, + 0.014182503335177898, + 0.006038027349859476, + 0.003927329555153847, + 0.0036434815265238285, + 0.015765849500894547, + 0.008535400032997131, + 0.011174413375556469, + 0.00505642918869853, + 0.036571718752384186, + 0.01106202881783247, + 0.005934169981628656, + 0.01057478878647089, + 0.008972895331680775, + 0.005980762653052807, + 0.023146791383624077, + 0.007702493574470282, + 0.027699220925569534, + 0.010548337362706661, + 0.0046683712862432, + 0.0045092725194990635, + 0.008788704872131348, + 0.012138407677412033, + 0.008079941384494305, + 0.009817587211728096, + 0.010744859464466572, + 0.012234492227435112, + 0.012796963565051556, + 0.0069396463222801685, + 0.007329156156629324, + 0.004801430739462376, + 0.0027653018478304148, + 0.006835305597633123, + 0.005453609861433506, + 0.006039980333298445, + 0.00448651984333992, + 0.0059702736325562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.11297706514596939 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003193875541910529 + ], + "opposite_logprob_tokens": [ + -5.7531938552856445 + ], + "logprob_reward": -0.003193875541910529, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003193875541910529, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1673724204301834, + 0.16886432468891144, + 0.11613266915082932, + 0.04944489151239395, + 0.12338551878929138, + 0.04028157889842987, + 0.17725515365600586, + 0.05225738510489464, + 0.04660676419734955, + 0.058399349451065063 + ], + "credit_map": [ + 0.01870208978652954, + 0.030487138777971268, + 0.01491005439311266, + 0.018559062853455544, + 0.026565607637166977, + 0.02304929681122303, + 0.01883259415626526, + 0.01626656763255596, + 0.01791047304868698, + 0.02436801791191101, + 0.010619589127600193, + 0.013188326731324196, + 0.0454280786216259, + 0.025783337652683258, + 0.017756694927811623, + 0.013809808529913425, + 0.011758989654481411, + 0.012561442330479622, + 0.012628435157239437, + 0.011707432568073273, + 0.01623566262423992, + 0.023983007296919823, + 0.013946467079222202, + 0.013311225920915604, + 0.004698673263192177, + 0.008631122298538685, + 0.007408583536744118, + 0.005528626963496208, + 0.006201664451509714, + 0.005223194602876902, + 0.00542459124699235, + 0.006328434683382511, + 0.013922832906246185, + 0.01232907921075821, + 0.008143334649503231, + 0.017916537821292877, + 0.015837492421269417, + 0.01959719881415367, + 0.020552407950162888, + 0.015086641535162926, + 0.009229214861989021, + 0.0063203987665474415, + 0.0041335104033350945, + 0.0032052542082965374, + 0.0056989118456840515, + 0.004103910643607378, + 0.004179865587502718, + 0.003410513512790203, + 0.028860384598374367, + 0.011595992371439934, + 0.010767052881419659, + 0.0076482840813696384, + 0.0695110633969307, + 0.016151754185557365, + 0.010239562019705772, + 0.022481059655547142, + 0.0055279056541621685, + 0.007724678609520197, + 0.009232587181031704, + 0.005137184169143438, + 0.0095650814473629, + 0.00808973703533411, + 0.0037079304456710815, + 0.003272279165685177, + 0.0042539979331195354, + 0.004081238992512226, + 0.006211558356881142, + 0.0029358877800405025, + 0.014596497640013695, + 0.00764026353135705, + 0.003825693391263485, + 0.003061625873669982, + 0.0066422224044799805, + 0.006529598496854305, + 0.004463281948119402, + 0.013004555366933346, + 0.007032900117337704, + 0.00890157651156187, + 0.00546952523291111, + 0.006355691701173782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.003193875541910529 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464135454967618 + ], + "opposite_logprob_tokens": [ + -5.379641532897949 + ], + "logprob_reward": -0.00464135454967618, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464135454967618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2544398605823517, + 0.09300469607114792, + 0.2257830947637558, + 0.04103250801563263, + 0.10425131022930145, + 0.06381609290838242, + 0.07064032554626465, + 0.03749465197324753, + 0.05752613767981529, + 0.05201132595539093 + ], + "credit_map": [ + 0.03076101467013359, + 0.04184882342815399, + 0.027665818110108376, + 0.024611255154013634, + 0.05165223777294159, + 0.0380285307765007, + 0.021998286247253418, + 0.017873913049697876, + 0.009665830992162228, + 0.013608753681182861, + 0.012167598120868206, + 0.01164009515196085, + 0.020288491621613503, + 0.012020130641758442, + 0.008210059255361557, + 0.005403734743595123, + 0.0215285774320364, + 0.04031231254339218, + 0.024149268865585327, + 0.0338384248316288, + 0.030555976554751396, + 0.03336873650550842, + 0.024678394198417664, + 0.017351413145661354, + 0.0037166280671954155, + 0.007791958749294281, + 0.0027674725279212, + 0.0044799502938985825, + 0.005863508675247431, + 0.007377886679023504, + 0.0034194914624094963, + 0.005615612026304007, + 0.013445036485791206, + 0.010434168390929699, + 0.008978678844869137, + 0.009419912472367287, + 0.018759362399578094, + 0.025036698207259178, + 0.010879699140787125, + 0.007297754753381014, + 0.00639123423025012, + 0.0075934468768537045, + 0.005291416309773922, + 0.0048774052411317825, + 0.01792287267744541, + 0.00868260208517313, + 0.008372600190341473, + 0.00468450877815485, + 0.006547840777784586, + 0.00523285660892725, + 0.005236660595983267, + 0.004088836256414652, + 0.021280616521835327, + 0.01287358533591032, + 0.007704034447669983, + 0.007675897795706987, + 0.004892803728580475, + 0.0056004710495471954, + 0.0033184101339429617, + 0.0044282954186201096, + 0.005143746733665466, + 0.007678698282688856, + 0.003195497440174222, + 0.003236731281504035, + 0.010803655721247196, + 0.0132181067019701, + 0.005181053187698126, + 0.005176128353923559, + 0.0073493667878210545, + 0.005300724413245916, + 0.00587939890101552, + 0.004617704078555107, + 0.0057743024080991745, + 0.004572039470076561, + 0.004240768495947123, + 0.008578150533139706, + 0.010121926665306091, + 0.006952183786779642, + 0.005600034724920988, + 0.006171920336782932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00464135454967618 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12702317535877228 + ], + "opposite_logprob_tokens": [ + -2.127023220062256 + ], + "logprob_reward": -0.12702317535877228, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12702317535877228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12389307469129562, + 0.043455395847558975, + 0.06871096044778824, + 0.0372002087533474, + 0.11630155891180038, + 0.12219642102718353, + 0.10422461479902267, + 0.1955278217792511, + 0.1734970360994339, + 0.014992890879511833 + ], + "credit_map": [ + 0.017815571278333664, + 0.024313030764460564, + 0.010417018085718155, + 0.010850644670426846, + 0.024566486477851868, + 0.017547428607940674, + 0.008661055006086826, + 0.00972183421254158, + 0.005319464486092329, + 0.008434942923486233, + 0.004233186598867178, + 0.004304284695535898, + 0.009088763035833836, + 0.004553024657070637, + 0.004207230173051357, + 0.003314499044790864, + 0.007358708418905735, + 0.010070659220218658, + 0.0063735865987837315, + 0.008494794368743896, + 0.012062820605933666, + 0.01299236062914133, + 0.006404081359505653, + 0.004953947849571705, + 0.0033770152367651463, + 0.003776772879064083, + 0.007085446733981371, + 0.005917063448578119, + 0.0045771244913339615, + 0.0050514815375208855, + 0.0036385362036526203, + 0.0037767684552818537, + 0.010357619263231754, + 0.015041578561067581, + 0.0088846730068326, + 0.016102943569421768, + 0.018722545355558395, + 0.020790070295333862, + 0.01614244095981121, + 0.010259686037898064, + 0.011465031653642654, + 0.013708421029150486, + 0.009538454003632069, + 0.009228319860994816, + 0.03479861095547676, + 0.017960190773010254, + 0.015956198796629906, + 0.009541190229356289, + 0.009257557801902294, + 0.01019842829555273, + 0.012715975753962994, + 0.006513221189379692, + 0.014629553072154522, + 0.022042641416192055, + 0.012633630074560642, + 0.01623360812664032, + 0.018329255282878876, + 0.01639462448656559, + 0.03998840972781181, + 0.01873726397752762, + 0.04835439473390579, + 0.026034614071249962, + 0.014414386823773384, + 0.01327486615628004, + 0.030886055901646614, + 0.02505382150411606, + 0.020973894745111465, + 0.01795087940990925, + 0.027991699054837227, + 0.016756068915128708, + 0.018352415412664413, + 0.015532201156020164, + 0.0018690815195441246, + 0.0016676918603479862, + 0.0009890412911772728, + 0.002167881466448307, + 0.0020882696844637394, + 0.0033283429220318794, + 0.0014481040416285396, + 0.0014344787923619151 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.12702317535877228 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot, rendered in a high-definition, surreal anime-style with backgrounds reminiscent of the works of Hayao Miyazaki, slowly pans right, tracking the catastrophic aftermath of a land-based earthquake. The camera gazes down upon a shattered valley, where a massive mud avalanche—a sticky, roiling mass of liquefied earth and dark, muddy debris—seeps relentlessly forward. Its surface is a chaotic tapestry of thick brown mud and gray silt, shimmering with a damp sheen under the sweltering, sun-bleached sky, a clear sign of intense surface warming. The mudslide's front, slow but unstoppable, its jagged front deliberately engulfs majestic pine trees, bending, splitting, and then swallowing them whole. The shattered remains of wooden structures, their timbers splintered and damaged, are devoured by the relentless current. Wisps of steam rise continuously from the sticky, roiling mass, which bubbles and moves as if alive, creating a terrifying and awe-inspiring atmosphere that showcases the raw, destructive power of nature.", + "seed": 16087, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16087_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1033052958548069, + "vlm_logprob": -0.1033052958548069, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.1033052958548069, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1178724616765976 + ], + "opposite_logprob_tokens": [ + -2.492872476577759 + ], + "logprob_reward": -0.1178724616765976, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.1178724616765976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17684438824653625, + 0.04335297271609306, + 0.14393316209316254, + 0.155532568693161, + 0.10262855887413025, + 0.01975524052977562, + 0.043707896023988724, + 0.09811723977327347, + 0.10241583734750748, + 0.11371217668056488 + ], + "credit_map": [ + 0.0327356718480587, + 0.01797361858189106, + 0.026814712211489677, + 0.02230319380760193, + 0.015803182497620583, + 0.021203728392720222, + 0.023699544370174408, + 0.01631072908639908, + 0.004180253483355045, + 0.004457766655832529, + 0.006748886778950691, + 0.005584889557212591, + 0.0041518546640872955, + 0.006091642193496227, + 0.006554532330483198, + 0.005583145655691624, + 0.011884355917572975, + 0.0163424052298069, + 0.02091916836798191, + 0.024894747883081436, + 0.013980804942548275, + 0.019534720107913017, + 0.022364482283592224, + 0.014012472704052925, + 0.017757227644324303, + 0.014197391457855701, + 0.01526354718953371, + 0.012957602739334106, + 0.024007519707083702, + 0.01837827078998089, + 0.037871796637773514, + 0.015099200420081615, + 0.011853708885610104, + 0.012955778278410435, + 0.011482981033623219, + 0.011830905452370644, + 0.010230453684926033, + 0.010983000509440899, + 0.023678964003920555, + 0.009612766094505787, + 0.0031235008500516415, + 0.002427790779620409, + 0.0023856007028371096, + 0.002588293282315135, + 0.0019252067431807518, + 0.002509546931833029, + 0.003010454820469022, + 0.001784846419468522, + 0.005443357862532139, + 0.0074934265576303005, + 0.004268543794751167, + 0.00529760867357254, + 0.006112277507781982, + 0.004875963553786278, + 0.006851476151496172, + 0.0033652428537607193, + 0.019306134432554245, + 0.012361729517579079, + 0.01237423438578844, + 0.011893593706190586, + 0.012228942476212978, + 0.01034263800829649, + 0.013182176277041435, + 0.006427781190723181, + 0.007094521541148424, + 0.005542335100471973, + 0.013197322376072407, + 0.014854131266474724, + 0.010530155152082443, + 0.016260860487818718, + 0.015647467225790024, + 0.019289037212729454, + 0.014707248657941818, + 0.008714480325579643, + 0.010502425022423267, + 0.011113740503787994, + 0.028628090396523476, + 0.012719832360744476, + 0.010816790163516998, + 0.016509560868144035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.1178724616765976 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428326964378357 + ], + "opposite_logprob_tokens": [ + -2.0178327560424805 + ], + "logprob_reward": -0.1428326964378357, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428326964378357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29868969321250916, + 0.10164296627044678, + 0.12296617031097412, + 0.061301760375499725, + 0.06831126660108566, + 0.07690851390361786, + 0.039032500237226486, + 0.07057564705610275, + 0.1299264132976532, + 0.030645053833723068 + ], + "credit_map": [ + 0.04013143107295036, + 0.02163718082010746, + 0.0401732437312603, + 0.04040710628032684, + 0.030580729246139526, + 0.03733690828084946, + 0.03961004316806793, + 0.048813045024871826, + 0.014513836242258549, + 0.010916351340711117, + 0.01584227755665779, + 0.011109843850135803, + 0.011098796501755714, + 0.014710393734276295, + 0.009312584064900875, + 0.01413888018578291, + 0.010346508584916592, + 0.010613245889544487, + 0.020700709894299507, + 0.019605670124292374, + 0.009029330685734749, + 0.017934903502464294, + 0.021026847884058952, + 0.013708957470953465, + 0.006599111948162317, + 0.003988415002822876, + 0.006522671785205603, + 0.005736073944717646, + 0.007975693792104721, + 0.007118048146367073, + 0.016555309295654297, + 0.006806431338191032, + 0.005101451184600592, + 0.008697932586073875, + 0.005275142844766378, + 0.007918525487184525, + 0.011943693272769451, + 0.006796360481530428, + 0.016466451808810234, + 0.006111710332334042, + 0.009705237112939358, + 0.007873867638409138, + 0.008393355645239353, + 0.01076053362339735, + 0.007695371750742197, + 0.007543756626546383, + 0.01678590476512909, + 0.00815048348158598, + 0.00519748218357563, + 0.007882271893322468, + 0.0031738660763949156, + 0.005546256434172392, + 0.005728463176637888, + 0.004237764049321413, + 0.004325344692915678, + 0.0029410466086119413, + 0.01546521671116352, + 0.009404378943145275, + 0.008950336836278439, + 0.007532132789492607, + 0.010274529457092285, + 0.006290638353675604, + 0.007542608771473169, + 0.005115802399814129, + 0.011687424033880234, + 0.008991795592010021, + 0.01828509010374546, + 0.018329689279198647, + 0.01436564326286316, + 0.01822102814912796, + 0.02027968503534794, + 0.01976606994867325, + 0.0029458312783390284, + 0.0030586812645196915, + 0.004068552982062101, + 0.00507992971688509, + 0.005364110227674246, + 0.003572908928617835, + 0.0035953305196017027, + 0.0029597098473459482 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.1428326964378357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033692799508571625 + ], + "opposite_logprob_tokens": [ + -3.4086928367614746 + ], + "logprob_reward": -0.033692799508571625, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033692799508571625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27935755252838135, + 0.08485373109579086, + 0.09943556785583496, + 0.08832249045372009, + 0.04669320583343506, + 0.0318174883723259, + 0.087833933532238, + 0.13356134295463562, + 0.0717015489935875, + 0.07642313838005066 + ], + "credit_map": [ + 0.03586086630821228, + 0.016630634665489197, + 0.03824076056480408, + 0.03544694185256958, + 0.021801410242915154, + 0.03339291736483574, + 0.05188607797026634, + 0.046097949147224426, + 0.008065585047006607, + 0.0073966155759990215, + 0.013585892505943775, + 0.010035822167992592, + 0.010210839100182056, + 0.009863612242043018, + 0.009850706905126572, + 0.015844661742448807, + 0.008415007032454014, + 0.011067071929574013, + 0.015623102895915508, + 0.015539593994617462, + 0.009096754714846611, + 0.012696029618382454, + 0.01801423169672489, + 0.008983773179352283, + 0.009250779636204243, + 0.008518812246620655, + 0.00875069946050644, + 0.00846562348306179, + 0.015414412133395672, + 0.010494700632989407, + 0.01871383935213089, + 0.008713616989552975, + 0.0034949269611388445, + 0.006753592751920223, + 0.004266656935214996, + 0.0055962433107197285, + 0.01031106524169445, + 0.00550618814304471, + 0.007137977052479982, + 0.0036265552043914795, + 0.003775195684283972, + 0.003046976402401924, + 0.0032643007580190897, + 0.0040131681598722935, + 0.0032391492277383804, + 0.002873323392122984, + 0.008339655585587025, + 0.0032657189294695854, + 0.007361458148807287, + 0.007918118499219418, + 0.007205124944448471, + 0.013812728226184845, + 0.011274061165750027, + 0.011426229029893875, + 0.01966959238052368, + 0.009166623465716839, + 0.014511927962303162, + 0.011630428023636341, + 0.014242559671401978, + 0.020349305123090744, + 0.013996852561831474, + 0.015116195194423199, + 0.029069017618894577, + 0.014645049348473549, + 0.004657148849219084, + 0.00456551369279623, + 0.010604454204440117, + 0.010102077387273312, + 0.00903890561312437, + 0.00910933967679739, + 0.013608613051474094, + 0.010015493258833885, + 0.007206559181213379, + 0.007645868230611086, + 0.010238492861390114, + 0.013851611874997616, + 0.011174964718520641, + 0.008717510849237442, + 0.011684667319059372, + 0.005903461016714573 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.033692799508571625 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585577934980392 + ], + "opposite_logprob_tokens": [ + -3.895585536956787 + ], + "logprob_reward": -0.020585577934980392, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585577934980392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1991281658411026, + 0.08658262342214584, + 0.07368834316730499, + 0.09892869740724564, + 0.06772586703300476, + 0.0783018171787262, + 0.054933950304985046, + 0.08825164288282394, + 0.18242605030536652, + 0.07003285735845566 + ], + "credit_map": [ + 0.02807755209505558, + 0.017349980771541595, + 0.02476361207664013, + 0.024302590638399124, + 0.016791950911283493, + 0.0172230526804924, + 0.03368782624602318, + 0.036931585520505905, + 0.00987828616052866, + 0.013900572434067726, + 0.011042138561606407, + 0.007144523318856955, + 0.01480917539447546, + 0.010651113465428352, + 0.007199431303888559, + 0.01195738185197115, + 0.010064571164548397, + 0.012864286080002785, + 0.007841711863875389, + 0.008510256186127663, + 0.007831793278455734, + 0.008490335196256638, + 0.01078876107931137, + 0.007296631578356028, + 0.011201293207705021, + 0.006124167703092098, + 0.011535339057445526, + 0.009955067187547684, + 0.011186366900801659, + 0.012417039833962917, + 0.02600840851664543, + 0.010501024313271046, + 0.007862504571676254, + 0.009137850254774094, + 0.007659936789423227, + 0.006747521925717592, + 0.006530234590172768, + 0.006803417112678289, + 0.013867419213056564, + 0.009116982109844685, + 0.009422986768186092, + 0.010415405966341496, + 0.00801751110702753, + 0.009902498684823513, + 0.008036598563194275, + 0.007194347679615021, + 0.017918268218636513, + 0.007394199725240469, + 0.00908166915178299, + 0.00789982546120882, + 0.005794495344161987, + 0.00571100227534771, + 0.006741715129464865, + 0.00759311905130744, + 0.006633565295487642, + 0.005478563252836466, + 0.011528689414262772, + 0.010009081102907658, + 0.008486047387123108, + 0.01374292653053999, + 0.006980860140174627, + 0.00945597980171442, + 0.019529545679688454, + 0.008518515154719353, + 0.011331351473927498, + 0.011151066981256008, + 0.02426208183169365, + 0.027204200625419617, + 0.018415922299027443, + 0.027262631803750992, + 0.0305384062230587, + 0.03226037696003914, + 0.0068636746145784855, + 0.004706752952188253, + 0.006862631533294916, + 0.006892944686114788, + 0.01581147126853466, + 0.008901112712919712, + 0.008277220651507378, + 0.011717048473656178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.020585577934980392 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015429437160492 + ], + "opposite_logprob_tokens": [ + -1.7015429735183716 + ], + "logprob_reward": -0.2015429437160492, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015429437160492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06173742935061455, + 0.04089174419641495, + 0.16154472529888153, + 0.06302394717931747, + 0.049486905336380005, + 0.023029126226902008, + 0.05129201337695122, + 0.17161472141742706, + 0.10132261365652084, + 0.27605676651000977 + ], + "credit_map": [ + 0.011810442432761192, + 0.00464101554825902, + 0.007636348716914654, + 0.008661499246954918, + 0.005405825562775135, + 0.00570636847987771, + 0.008960843086242676, + 0.008915082551538944, + 0.005331866908818483, + 0.0032668670173734426, + 0.005129189230501652, + 0.005269261077046394, + 0.004705847706645727, + 0.0062493435107171535, + 0.004219695460051298, + 0.006719671655446291, + 0.013889472931623459, + 0.018504435196518898, + 0.019849205389618874, + 0.023468539118766785, + 0.020249849185347557, + 0.021655771881341934, + 0.028306739404797554, + 0.015620705671608448, + 0.007982530631124973, + 0.005621724762022495, + 0.006119345780462027, + 0.00595172168686986, + 0.010570614598691463, + 0.0076226117089390755, + 0.012941863387823105, + 0.00621353555470705, + 0.005476159509271383, + 0.004567616153508425, + 0.006489613093435764, + 0.004979702178388834, + 0.006364763714373112, + 0.005690917372703552, + 0.011099007911980152, + 0.004819128662347794, + 0.0032502631656825542, + 0.0025448224041610956, + 0.002519250148907304, + 0.003212542040273547, + 0.002109292894601822, + 0.002714031608775258, + 0.004561993293464184, + 0.0021169306710362434, + 0.008360497653484344, + 0.009032740257680416, + 0.004816785920411348, + 0.00550086097791791, + 0.005997329950332642, + 0.006142703350633383, + 0.007533201947808266, + 0.0039078909903764725, + 0.0388215072453022, + 0.024135317653417587, + 0.020235570147633553, + 0.014853633008897305, + 0.023981915786862373, + 0.01740669459104538, + 0.021161243319511414, + 0.01101885549724102, + 0.012825561687350273, + 0.008496670052409172, + 0.013550304807722569, + 0.013197331689298153, + 0.012702863663434982, + 0.012528207153081894, + 0.013281974010169506, + 0.014739698730409145, + 0.028955310583114624, + 0.017632700502872467, + 0.0289069302380085, + 0.03264227136969566, + 0.06402310729026794, + 0.03264070302248001, + 0.028863927349448204, + 0.0423918180167675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 2 + ], + "selected_reward": -0.2015429437160492 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "high-angle, wide-angle shot, rendered in a high-definition, surreal anime-style with backgrounds reminiscent of the works of Hayao Miyazaki, slowly pans right, tracking the catastrophic aftermath of a land-based earthquake. The camera gazes down upon a shattered valley, where a massive mud avalanche—a sticky, roiling mass of liquefied earth and dark, muddy debris—seeps relentlessly forward. Its surface is a chaotic tapestry of thick brown mud and gray silt, shimmering with a damp sheen under the sweltering, sun-bleached sky, a clear sign of intense surface warming. The mudslide's front, slow but unstoppable, its jagged front deliberately engulfs majestic pine trees, bending, splitting, and then swallowing them whole. The shattered remains of wooden structures, their timbers splintered and damaged, are devoured by the relentless current. Wisps of steam rise continuously from the sticky, roiling mass, which bubbles and moves as if alive, creating a terrifying and awe-inspiring atmosphere that showcases the raw, destructive power of nature.", + "seed": 16088, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000048/group_6/index_2529_seed_16088_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.16832789704203605, + "vlm_logprob": -0.16832789704203605, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.16832789704203605, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12639114260673523 + ], + "opposite_logprob_tokens": [ + -2.3763911724090576 + ], + "logprob_reward": -0.12639114260673523, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.12639114260673523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10648375004529953, + 0.15580451488494873, + 0.05615473538637161, + 0.2610991895198822, + 0.11915348470211029, + 0.042863328009843826, + 0.054498884826898575, + 0.06586438417434692, + 0.08428186178207397, + 0.053795844316482544 + ], + "credit_map": [ + 0.008088422007858753, + 0.013185961171984673, + 0.01024945080280304, + 0.018166016787290573, + 0.009109559468925, + 0.008593116886913776, + 0.02231471613049507, + 0.01677650772035122, + 0.008230840787291527, + 0.008096879348158836, + 0.007141757290810347, + 0.015216243453323841, + 0.006413452327251434, + 0.005957032088190317, + 0.059827160090208054, + 0.0449211485683918, + 0.005675519350916147, + 0.005980037152767181, + 0.005042333155870438, + 0.012518329545855522, + 0.003856998635455966, + 0.004627471789717674, + 0.010254100896418095, + 0.008199941366910934, + 0.018680403009057045, + 0.025749120861291885, + 0.0260437224060297, + 0.06852088868618011, + 0.01094056200236082, + 0.014866462908685207, + 0.054644759744405746, + 0.04165324941277504, + 0.008615796454250813, + 0.010574783198535442, + 0.011099820956587791, + 0.028349842876195908, + 0.0069577437825500965, + 0.009124230593442917, + 0.022045232355594635, + 0.02238602377474308, + 0.0029726887587457895, + 0.0021477900445461273, + 0.007035478018224239, + 0.011242210865020752, + 0.002096035750582814, + 0.0024934750981628895, + 0.005032345652580261, + 0.009843301959335804, + 0.0037607834674417973, + 0.0028238047380000353, + 0.009019323624670506, + 0.010054442100226879, + 0.00403629383072257, + 0.002910671290010214, + 0.014902439899742603, + 0.006991129368543625, + 0.0042839557863771915, + 0.0048417081125080585, + 0.007671704515814781, + 0.016777504235506058, + 0.004252083133906126, + 0.0036039857659488916, + 0.010446433909237385, + 0.013987011276185513, + 0.006049488205462694, + 0.006966379005461931, + 0.01483766920864582, + 0.01549217663705349, + 0.010259614326059818, + 0.004577263258397579, + 0.01605222001671791, + 0.010047052055597305, + 0.004683258011937141, + 0.0029898607172071934, + 0.011128261685371399, + 0.005636474583297968, + 0.0032864718232303858, + 0.003010163549333811, + 0.018957069143652916, + 0.004104285500943661 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.12639114260673523 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.160392165184021 + ], + "opposite_logprob_tokens": [ + -1.910392165184021 + ], + "logprob_reward": -0.160392165184021, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.160392165184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.46549534797668457, + 0.05781250819563866, + 0.16413502395153046, + 0.10710857808589935, + 0.034652937203645706, + 0.023488745093345642, + 0.02658774144947529, + 0.04817180708050728, + 0.05203460156917572, + 0.020512783899903297 + ], + "credit_map": [ + 0.027815092355012894, + 0.01215303223580122, + 0.08029972016811371, + 0.04179377481341362, + 0.018131785094738007, + 0.02175384946167469, + 0.19445161521434784, + 0.06909643858671188, + 0.0032922872342169285, + 0.00209264294244349, + 0.0036778065841645002, + 0.005219826940447092, + 0.0031258745584636927, + 0.0023007767740637064, + 0.022478070110082626, + 0.01562521792948246, + 0.011942273937165737, + 0.012651142664253712, + 0.013226774521172047, + 0.026862330734729767, + 0.00989765115082264, + 0.009516910649836063, + 0.054841235280036926, + 0.025196686387062073, + 0.008250266313552856, + 0.011803863570094109, + 0.009723411872982979, + 0.023781100288033485, + 0.006479327101260424, + 0.005917894188314676, + 0.02378370240330696, + 0.017369011417031288, + 0.0030387865845113993, + 0.0029004127718508244, + 0.0037038191221654415, + 0.00772695941850543, + 0.0032430985011160374, + 0.003135599195957184, + 0.005409351550042629, + 0.00549490749835968, + 0.0028984148520976305, + 0.0022271533962339163, + 0.0025791239459067583, + 0.003611229592934251, + 0.0011708368547260761, + 0.0022650458849966526, + 0.0047944169491529465, + 0.003942520823329687, + 0.0018292138120159507, + 0.0021316546481102705, + 0.0035843600053340197, + 0.004976990167051554, + 0.001870938460342586, + 0.0021720142103731632, + 0.006271157413721085, + 0.00375141273252666, + 0.0052781677804887295, + 0.0033925906755030155, + 0.006443677470088005, + 0.011539713479578495, + 0.003127017291262746, + 0.0035620853304862976, + 0.008037708699703217, + 0.0067908428609371185, + 0.0027618701569736004, + 0.003569632302969694, + 0.01019282266497612, + 0.00900519173592329, + 0.006882050074636936, + 0.0032971887849271297, + 0.00923064537346363, + 0.007095197215676308, + 0.0016161783132702112, + 0.0012791218468919396, + 0.004007375333458185, + 0.0024333635810762644, + 0.0014680558815598488, + 0.0013824767665937543, + 0.005725067108869553, + 0.0026011448353528976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.160392165184021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34862133860588074 + ], + "opposite_logprob_tokens": [ + -1.2236213684082031 + ], + "logprob_reward": -0.34862133860588074, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.34862133860588074, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09750574082136154, + 0.2001408040523529, + 0.057146843522787094, + 0.22999416291713715, + 0.05978349968791008, + 0.029338961467146873, + 0.02966083399951458, + 0.0905335545539856, + 0.16762211918830872, + 0.038273464888334274 + ], + "credit_map": [ + 0.006853642873466015, + 0.007975751534104347, + 0.00890421587973833, + 0.013995500281453133, + 0.006390917114913464, + 0.006027453579008579, + 0.03025544434785843, + 0.017102820798754692, + 0.010838394053280354, + 0.009715083986520767, + 0.013828867115080357, + 0.027088167145848274, + 0.008312349207699299, + 0.006234389264136553, + 0.07682079076766968, + 0.04730275273323059, + 0.004581810906529427, + 0.004165987018495798, + 0.007766943424940109, + 0.012406298890709877, + 0.00394233176484704, + 0.00361534277908504, + 0.011241337284445763, + 0.009426791220903397, + 0.02174334041774273, + 0.02250954508781433, + 0.024430107325315475, + 0.05925385653972626, + 0.010638878680765629, + 0.016247998923063278, + 0.03764510899782181, + 0.03752531856298447, + 0.004790273495018482, + 0.0037865042686462402, + 0.00656467117369175, + 0.015405936166644096, + 0.0027101750019937754, + 0.00418228842318058, + 0.012983097694814205, + 0.009360556490719318, + 0.0024557861033827066, + 0.0018424751469865441, + 0.003988423850387335, + 0.007242915220558643, + 0.0011577334953472018, + 0.0015475749969482422, + 0.005066493060439825, + 0.006037558428943157, + 0.0025585670955479145, + 0.001740779960528016, + 0.002990411128848791, + 0.007144944276660681, + 0.0016406060894951224, + 0.0014259369345381856, + 0.006858242675662041, + 0.0053013465367257595, + 0.01174174901098013, + 0.0069615719839930534, + 0.012983830645680428, + 0.013579079881310463, + 0.006552777253091335, + 0.007889149710536003, + 0.01738636940717697, + 0.013439026661217213, + 0.010180529206991196, + 0.012150966562330723, + 0.033275503665208817, + 0.03113824501633644, + 0.020586872473359108, + 0.01150494534522295, + 0.027429915964603424, + 0.021355148404836655, + 0.004507557954639196, + 0.003792723175138235, + 0.005527401342988014, + 0.00459287641569972, + 0.0036226154770702124, + 0.003231025068089366, + 0.007929105311632156, + 0.005070157814770937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 8 + ], + "selected_reward": -0.34862133860588074 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635374292731285 + ], + "opposite_logprob_tokens": [ + -3.6513538360595703 + ], + "logprob_reward": -0.02635374292731285, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635374292731285, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21353371441364288, + 0.13655336201190948, + 0.04881897568702698, + 0.19875779747962952, + 0.06984482705593109, + 0.03358529508113861, + 0.043094322085380554, + 0.07860114425420761, + 0.14214181900024414, + 0.03506875038146973 + ], + "credit_map": [ + 0.012175977230072021, + 0.010120694525539875, + 0.03186937794089317, + 0.021395694464445114, + 0.009566344320774078, + 0.011321349069476128, + 0.08306335657835007, + 0.03402091562747955, + 0.006984110455960035, + 0.00399967934936285, + 0.012777559459209442, + 0.013068881817162037, + 0.0055068242363631725, + 0.0038298561703413725, + 0.05655117332935333, + 0.0338352732360363, + 0.00421835295855999, + 0.004572274629026651, + 0.005347349215298891, + 0.009254738688468933, + 0.0038832903373986483, + 0.0034345609601587057, + 0.010944085195660591, + 0.007164327427744865, + 0.017385564744472504, + 0.020283348858356476, + 0.02054518833756447, + 0.05129729211330414, + 0.009190171025693417, + 0.013127611018717289, + 0.03452550619840622, + 0.0324031226336956, + 0.004933518823236227, + 0.0040879319421947, + 0.00877891480922699, + 0.01881541684269905, + 0.0029041036032140255, + 0.003976908512413502, + 0.01464628241956234, + 0.011701756156980991, + 0.0043082027696073055, + 0.0021952707320451736, + 0.004414456896483898, + 0.006833360064774752, + 0.0015480249421671033, + 0.0021202489733695984, + 0.004532207269221544, + 0.007633523549884558, + 0.002589806681498885, + 0.0017236876301467419, + 0.006148903630673885, + 0.009381726384162903, + 0.002746900310739875, + 0.001970262499526143, + 0.012831032276153564, + 0.005702000577002764, + 0.010403313674032688, + 0.006188467610627413, + 0.01108589582145214, + 0.01208438165485859, + 0.005854888819158077, + 0.006683186627924442, + 0.015201045200228691, + 0.01109996810555458, + 0.008168541826307774, + 0.009088103659451008, + 0.029207825660705566, + 0.02656901814043522, + 0.017264073714613914, + 0.00834010262042284, + 0.024302346631884575, + 0.01920180581510067, + 0.003006923943758011, + 0.002824830822646618, + 0.006192953325808048, + 0.0038261879235506058, + 0.002755938097834587, + 0.0022326880134642124, + 0.011420018970966339, + 0.002809209516271949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.02635374292731285 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a high-angle, wide-angle perspective looking down upon the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style a high-definition surreal anime style with backgrounds reminiscent of Hayao Miyazaki's works?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene depict a massive mud avalanche composed of sticky, roiling liquefied earth and dark muddy debris?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are wisps of steam rising continuously from the mudslide surface while it engulfs pine trees and wooden structures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera slowly pan right while tracking the forward movement of the mud avalanche?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17988109588623047 + ], + "opposite_logprob_tokens": [ + -1.8048810958862305 + ], + "logprob_reward": -0.17988109588623047, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17988109588623047, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08110792934894562, + 0.14073871076107025, + 0.054076045751571655, + 0.40552371740341187, + 0.1271016001701355, + 0.04073706269264221, + 0.016138076782226562, + 0.026507357135415077, + 0.04605232924222946, + 0.06201718747615814 + ], + "credit_map": [ + 0.005629414692521095, + 0.007815943099558353, + 0.015422863885760307, + 0.01148890983313322, + 0.003984829410910606, + 0.003332675900310278, + 0.024566302075982094, + 0.008866995573043823, + 0.00795306172221899, + 0.005035869311541319, + 0.007544614374637604, + 0.015366993844509125, + 0.00730533804744482, + 0.004382363986223936, + 0.054830800741910934, + 0.0383196696639061, + 0.005074379965662956, + 0.0040795025415718555, + 0.006563308648765087, + 0.012153297662734985, + 0.003185550682246685, + 0.002904697321355343, + 0.01222517341375351, + 0.007890134118497372, + 0.029755540192127228, + 0.040151771157979965, + 0.0409240759909153, + 0.10581182688474655, + 0.017685692757368088, + 0.023157693445682526, + 0.08403992652893066, + 0.06399720907211304, + 0.009459328837692738, + 0.015686046332120895, + 0.013162510469555855, + 0.021756663918495178, + 0.009041479788720608, + 0.01043081283569336, + 0.02657460793852806, + 0.020990146324038506, + 0.003607917344197631, + 0.002288720104843378, + 0.0059927464462816715, + 0.00975052546709776, + 0.0015839627012610435, + 0.0023916871286928654, + 0.005614015273749828, + 0.00950748659670353, + 0.0026985672302544117, + 0.001375014428049326, + 0.0018115012207999825, + 0.002860358450561762, + 0.0014039125526323915, + 0.001122920773923397, + 0.0029526627622544765, + 0.0019131411099806428, + 0.003502386622130871, + 0.0022818271536380053, + 0.0034821457229554653, + 0.003917488735169172, + 0.001950978534296155, + 0.0022414876148104668, + 0.005143634509295225, + 0.003987408243119717, + 0.002943650586530566, + 0.003400675253942609, + 0.007564040366560221, + 0.007883057929575443, + 0.006222840864211321, + 0.0031348788179457188, + 0.009172691032290459, + 0.005730495788156986, + 0.005158740095794201, + 0.0032769532408565283, + 0.014675159938633442, + 0.008735030889511108, + 0.003996286541223526, + 0.003209484973922372, + 0.018183041363954544, + 0.0047824932262301445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 4 + ], + "selected_reward": -0.17988109588623047 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_tensor([5473, 5473])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_tensor([5473, 5473])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4078fd53bf0d097c5d9ce879aada485f20b9a2ea --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_tensor([5473, 5473])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping panoramic shot, the camera slowly pulling back to reveal the panorama of a precipitous canyon, etched in the style of Tsui Hark's martial arts fantasy. In the dim light of early dusk, jagged, knife-cut cliffs rise like the skeletons of a giant beast. A damp mist lingers in the bottomless chasm, creating an atmosphere of mystery and danger. Within this silence, the dense forest of blue devil irises, covering the entire valley floor and portions of the sheer rock faces, their tightly closed deep blue, almost black buds suddenly open at the same moment, as if receiving a silent command. Each petal, as thin as a cicada's wing, burst open with a visible speed, its edges gleaming with an eerie electric blue light. A core of pure energy pulsed in unison, releasing a silent roar of energy—a force that transformed into a transparent, visible shockwave that rapidly spread from the center of the sea of flowers. Wherever it passed, the air violently twisted and rippled, even sending tiny rocks tumbling from the cliffs. This strange blue light instantly illuminated the entire valley, tinting the outlines of every rock a deep blue and casting long, twisting black shadows. It perfectly captured the spectacular scene of the valley being instantly engulfed by a supernatural force, creating a fantastical atmosphere that was both magnificent and oppressive.", + "seed": 18408, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_5473_seed_18408_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3324581887573004, + "vlm_logprob": -0.3324581887573004, + "vlm_yes_no_margin": 1.9, + "vlm_token_credit": -0.3324581887573004, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9073560833930969 + ], + "opposite_logprob_tokens": [ + -0.5323560833930969 + ], + "logprob_reward": -0.9073560833930969, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9073560833930969, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15088629722595215, + 0.22216898202896118, + 0.026302224025130272, + 0.21208947896957397, + 0.04343952238559723, + 0.08519642055034637, + 0.03931105136871338, + 0.1022261455655098, + 0.054456181824207306, + 0.063923679292202 + ], + "credit_map": [ + 0.0173968356102705, + 0.02879011258482933, + 0.011797969229519367, + 0.019844306632876396, + 0.032023828476667404, + 0.017111951485276222, + 0.013706274330615997, + 0.010215021669864655, + 0.02565084956586361, + 0.03546756133437157, + 0.028151337057352066, + 0.028995666652917862, + 0.03936466947197914, + 0.024173593148589134, + 0.023972665891051292, + 0.016392625868320465, + 0.002621866064146161, + 0.004248238168656826, + 0.0037344421725720167, + 0.0025584930554032326, + 0.004089561756700277, + 0.002703359816223383, + 0.004109413828700781, + 0.0022368503268808126, + 0.021500328555703163, + 0.028702840209007263, + 0.02545049972832203, + 0.02163895033299923, + 0.03745567426085472, + 0.019059330224990845, + 0.02674335427582264, + 0.03153849020600319, + 0.004228001460433006, + 0.00558874849230051, + 0.005462699569761753, + 0.00505836121737957, + 0.0067879511043429375, + 0.003968140576034784, + 0.007826116867363453, + 0.0045194970443844795, + 0.007302907761186361, + 0.01535994652658701, + 0.008737410418689251, + 0.010910902172327042, + 0.01344460342079401, + 0.007798510603606701, + 0.013858940452337265, + 0.007783193606883287, + 0.00368199753575027, + 0.007059273310005665, + 0.005130918230861425, + 0.00626466004177928, + 0.003357161534950137, + 0.005502966232597828, + 0.004611540585756302, + 0.0037025304045528173, + 0.010367870330810547, + 0.020293772220611572, + 0.010718284174799919, + 0.013307299464941025, + 0.009680190123617649, + 0.015494518913328648, + 0.01186361163854599, + 0.010500594973564148, + 0.003589428961277008, + 0.010543763637542725, + 0.013429486192762852, + 0.004576913546770811, + 0.007452038582414389, + 0.005104478914290667, + 0.004764091223478317, + 0.004995976574718952, + 0.006918064784258604, + 0.007022842299193144, + 0.006191012915223837, + 0.005589108914136887, + 0.00889529101550579, + 0.01192699745297432, + 0.005129522643983364, + 0.012250836938619614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.9073560833930969 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026381375268101692 + ], + "opposite_logprob_tokens": [ + -3.651381492614746 + ], + "logprob_reward": -0.026381375268101692, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026381375268101692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14099077880382538, + 0.1254642754793167, + 0.04052046686410904, + 0.10427439212799072, + 0.03597986325621605, + 0.06165983900427818, + 0.03007439337670803, + 0.19589225947856903, + 0.0770779550075531, + 0.188065767288208 + ], + "credit_map": [ + 0.012949683703482151, + 0.02499973028898239, + 0.010899286717176437, + 0.016683368012309074, + 0.03232642263174057, + 0.017585396766662598, + 0.013959514908492565, + 0.011587383225560188, + 0.013934774324297905, + 0.021916180849075317, + 0.017209863290190697, + 0.015495887957513332, + 0.02233070321381092, + 0.013536037877202034, + 0.013116644695401192, + 0.007924186997115612, + 0.004303957801312208, + 0.007396102882921696, + 0.004358110949397087, + 0.004291942343115807, + 0.006202257703989744, + 0.004764180164784193, + 0.005440465174615383, + 0.0037634491454809904, + 0.012148403562605381, + 0.016877461224794388, + 0.00953515525907278, + 0.011091790162026882, + 0.019357871264219284, + 0.01004354003816843, + 0.012105016969144344, + 0.01311515737324953, + 0.003799054306000471, + 0.004330628551542759, + 0.004327498376369476, + 0.003787783905863762, + 0.005919553339481354, + 0.00397499231621623, + 0.006065409630537033, + 0.0037749414332211018, + 0.006039479281753302, + 0.011309102177619934, + 0.005637546069920063, + 0.006434015929698944, + 0.012295101769268513, + 0.007038874085992575, + 0.00823060143738985, + 0.004675120115280151, + 0.002408065367490053, + 0.004344839137047529, + 0.005464211571961641, + 0.005812498740851879, + 0.00248245382681489, + 0.003328053280711174, + 0.0034032503608614206, + 0.0028310201596468687, + 0.019085299223661423, + 0.03374067693948746, + 0.028334030881524086, + 0.02508452907204628, + 0.02076921984553337, + 0.029626596719026566, + 0.020067453384399414, + 0.019184453412890434, + 0.004080530721694231, + 0.019118182361125946, + 0.018977507948875427, + 0.005566904321312904, + 0.012784981168806553, + 0.005766558460891247, + 0.00582516472786665, + 0.0049581266939640045, + 0.02523058094084263, + 0.023971188813447952, + 0.01856289803981781, + 0.01664067432284355, + 0.02235683985054493, + 0.034005746245384216, + 0.01339185331016779, + 0.0339059941470623 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.026381375268101692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371354565024376 + ], + "opposite_logprob_tokens": [ + -3.4087135791778564 + ], + "logprob_reward": -0.03371354565024376, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371354565024376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22324059903621674, + 0.337327778339386, + 0.05868598818778992, + 0.05325014144182205, + 0.04455574229359627, + 0.03255569934844971, + 0.02498292550444603, + 0.09205412119626999, + 0.07871716469526291, + 0.05462978780269623 + ], + "credit_map": [ + 0.01798602007329464, + 0.024622421711683273, + 0.02308369241654873, + 0.04651883989572525, + 0.0322384350001812, + 0.023333052173256874, + 0.026130951941013336, + 0.02932719886302948, + 0.0449172742664814, + 0.05161271616816521, + 0.05201040208339691, + 0.04625628888607025, + 0.04639330878853798, + 0.03249962255358696, + 0.036340612918138504, + 0.027297567576169968, + 0.007266400847584009, + 0.008511388674378395, + 0.005482747685164213, + 0.00638555595651269, + 0.007796252146363258, + 0.008329859934747219, + 0.00944550521671772, + 0.00546828331425786, + 0.0037643411196768284, + 0.007530618458986282, + 0.007246510591357946, + 0.005892390850931406, + 0.008128954097628593, + 0.0051153539679944515, + 0.0061035193502902985, + 0.00946845393627882, + 0.004683611914515495, + 0.006430133245885372, + 0.005171401426196098, + 0.0043001482263207436, + 0.008956165984272957, + 0.004908387549221516, + 0.005684876814484596, + 0.004421020392328501, + 0.005044898483902216, + 0.004828840494155884, + 0.0039698276668787, + 0.0034804821480065584, + 0.0025624807458370924, + 0.0040696049109101295, + 0.003482244908809662, + 0.005117322783917189, + 0.003163909539580345, + 0.0037963250651955605, + 0.0038071894086897373, + 0.002907544607296586, + 0.0027939672581851482, + 0.0031134190503507853, + 0.002699202625080943, + 0.002701370744034648, + 0.00899769738316536, + 0.018617160618305206, + 0.00979562383145094, + 0.012154779396951199, + 0.010428321547806263, + 0.014917494729161263, + 0.009036383591592312, + 0.00810666847974062, + 0.003617936745285988, + 0.021611548960208893, + 0.004903252236545086, + 0.006730349734425545, + 0.024814752861857414, + 0.005625640973448753, + 0.006698643788695335, + 0.004715043120086193, + 0.008277497254312038, + 0.006897119339555502, + 0.00562315946444869, + 0.003998309373855591, + 0.007367787882685661, + 0.008627683855593204, + 0.0043690381571650505, + 0.009469196200370789 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.03371354565024376 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205175817012787 + ], + "opposite_logprob_tokens": [ + -2.812051773071289 + ], + "logprob_reward": -0.06205175817012787, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205175817012787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21204018592834473, + 0.0619017519056797, + 0.0786796510219574, + 0.10104278475046158, + 0.05964933708310127, + 0.09877562522888184, + 0.052784889936447144, + 0.1047491803765297, + 0.17358680069446564, + 0.056789811700582504 + ], + "credit_map": [ + 0.01740266941487789, + 0.023829525336623192, + 0.02191588655114174, + 0.03845430910587311, + 0.03464464843273163, + 0.021154770627617836, + 0.02182086557149887, + 0.03281748294830322, + 0.009586308151483536, + 0.00844574999064207, + 0.005247884895652533, + 0.00809821393340826, + 0.009533455595374107, + 0.006182072218507528, + 0.007282862439751625, + 0.007525202352553606, + 0.008430403657257557, + 0.015748532488942146, + 0.007945341058075428, + 0.009722011163830757, + 0.015063001774251461, + 0.007900123484432697, + 0.008192967623472214, + 0.005677261855453253, + 0.01056699175387621, + 0.01723840832710266, + 0.008759496733546257, + 0.009751972742378712, + 0.022544121369719505, + 0.009119647555053234, + 0.012110980227589607, + 0.010951162315905094, + 0.005654219072312117, + 0.009590515866875648, + 0.007441038265824318, + 0.006292878184467554, + 0.011039569973945618, + 0.005571715999394655, + 0.008903204463422298, + 0.00515619246289134, + 0.013516124337911606, + 0.017349446192383766, + 0.008696557022631168, + 0.01211111806333065, + 0.011507228948175907, + 0.013967087492346764, + 0.010960256680846214, + 0.010667801834642887, + 0.004596461076289415, + 0.006295344326645136, + 0.009715239517390728, + 0.007779636885970831, + 0.00632711173966527, + 0.005471597425639629, + 0.00777832604944706, + 0.00482116499915719, + 0.008810487575829029, + 0.015540232881903648, + 0.01345133502036333, + 0.01202460192143917, + 0.017204489558935165, + 0.014513429254293442, + 0.01343823503702879, + 0.009766362607479095, + 0.013069765642285347, + 0.027603115886449814, + 0.01856900379061699, + 0.016288774088025093, + 0.04521607980132103, + 0.01516568660736084, + 0.02159498631954193, + 0.016079377382993698, + 0.005889670457690954, + 0.007722637616097927, + 0.00606390368193388, + 0.007244820240885019, + 0.00886402279138565, + 0.006370610557496548, + 0.005046226549893618, + 0.009587915614247322 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.06205175817012787 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6327881813049316 + ], + "opposite_logprob_tokens": [ + -0.7577881813049316 + ], + "logprob_reward": -0.6327881813049316, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6327881813049316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08622805029153824, + 0.12297042459249496, + 0.07255253195762634, + 0.12203994393348694, + 0.032442037016153336, + 0.060766082257032394, + 0.07208501547574997, + 0.10342176258563995, + 0.21130794286727905, + 0.11618619412183762 + ], + "credit_map": [ + 0.009112500585615635, + 0.013408282771706581, + 0.007851568050682545, + 0.010908400639891624, + 0.01629427820444107, + 0.010447927750647068, + 0.008651881478726864, + 0.00955321453511715, + 0.015287909656763077, + 0.017629507929086685, + 0.01772531308233738, + 0.01596890203654766, + 0.016713107004761696, + 0.011628914624452591, + 0.017490576952695847, + 0.010526211932301521, + 0.008474637754261494, + 0.013571645133197308, + 0.008779257535934448, + 0.007082945201545954, + 0.013103185221552849, + 0.005694589111953974, + 0.010420246981084347, + 0.005426025949418545, + 0.012008054181933403, + 0.019663089886307716, + 0.012759010307490826, + 0.014057131484150887, + 0.02437136322259903, + 0.01091789361089468, + 0.012956536374986172, + 0.015306871384382248, + 0.0027189545799046755, + 0.0051875910721719265, + 0.0036894192453473806, + 0.003449220210313797, + 0.006335013546049595, + 0.0032011924777179956, + 0.003799128346145153, + 0.004061518702656031, + 0.00626563373953104, + 0.008524340577423573, + 0.0087353540584445, + 0.00789865292608738, + 0.00679820915684104, + 0.0052392752841115, + 0.008899091742932796, + 0.008405525237321854, + 0.009066252037882805, + 0.01161080040037632, + 0.007780393585562706, + 0.011064327321946621, + 0.006678255740553141, + 0.010990024544298649, + 0.007560407742857933, + 0.007334562949836254, + 0.009395609609782696, + 0.017635781317949295, + 0.012582601979374886, + 0.01659315451979637, + 0.010615288279950619, + 0.013655003160238266, + 0.011297592893242836, + 0.011646734550595284, + 0.015591220930218697, + 0.01978970505297184, + 0.04018925130367279, + 0.01894094981253147, + 0.04136897623538971, + 0.024985915049910545, + 0.022470606490969658, + 0.027971334755420685, + 0.016315532848238945, + 0.014021278358995914, + 0.011638334020972252, + 0.008742393925786018, + 0.010631055571138859, + 0.023172156885266304, + 0.007816808298230171, + 0.02384864166378975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 3 + ], + "selected_reward": -0.6327881813049316 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping panoramic shot, the camera slowly pulling back to reveal the panorama of a precipitous canyon, etched in the style of Tsui Hark's martial arts fantasy. In the dim light of early dusk, jagged, knife-cut cliffs rise like the skeletons of a giant beast. A damp mist lingers in the bottomless chasm, creating an atmosphere of mystery and danger. Within this silence, the dense forest of blue devil irises, covering the entire valley floor and portions of the sheer rock faces, their tightly closed deep blue, almost black buds suddenly open at the same moment, as if receiving a silent command. Each petal, as thin as a cicada's wing, burst open with a visible speed, its edges gleaming with an eerie electric blue light. A core of pure energy pulsed in unison, releasing a silent roar of energy—a force that transformed into a transparent, visible shockwave that rapidly spread from the center of the sea of flowers. Wherever it passed, the air violently twisted and rippled, even sending tiny rocks tumbling from the cliffs. This strange blue light instantly illuminated the entire valley, tinting the outlines of every rock a deep blue and casting long, twisting black shadows. It perfectly captured the spectacular scene of the valley being instantly engulfed by a supernatural force, creating a fantastical atmosphere that was both magnificent and oppressive.", + "seed": 18409, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_5473_seed_18409_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2030729878693819, + "vlm_logprob": -0.2030729878693819, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.2030729878693819, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2339628040790558 + ], + "opposite_logprob_tokens": [ + -1.6089627742767334 + ], + "logprob_reward": -0.2339628040790558, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2339628040790558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17795045673847198, + 0.15362338721752167, + 0.26918619871139526, + 0.032548245042562485, + 0.1200995221734047, + 0.02271367236971855, + 0.032982129603624344, + 0.0524384081363678, + 0.09737316519021988, + 0.04108482226729393 + ], + "credit_map": [ + 0.03587261587381363, + 0.018330004066228867, + 0.013650014065206051, + 0.019586807116866112, + 0.02012084797024727, + 0.0413186177611351, + 0.018051164224743843, + 0.011020383797585964, + 0.015183358453214169, + 0.02563263662159443, + 0.0177774615585804, + 0.011028188280761242, + 0.020696187391877174, + 0.02849309705197811, + 0.023927245289087296, + 0.010885220021009445, + 0.040716324001550674, + 0.021788625046610832, + 0.03322137892246246, + 0.01934579387307167, + 0.05248330160975456, + 0.03290906921029091, + 0.0361507274210453, + 0.03257095068693161, + 0.003996122628450394, + 0.002792306011542678, + 0.004915718920528889, + 0.0034349500201642513, + 0.0036003771238029003, + 0.0060502453707158566, + 0.004056171979755163, + 0.003702352289110422, + 0.010284253396093845, + 0.01038030069321394, + 0.022470826283097267, + 0.013603932224214077, + 0.014776351861655712, + 0.01577489636838436, + 0.019724901765584946, + 0.01308405864983797, + 0.0026692021638154984, + 0.0028735476080328226, + 0.002082533435896039, + 0.0022028645034879446, + 0.0032121227122843266, + 0.0029600798152387142, + 0.002960460726171732, + 0.0037528607062995434, + 0.004172117915004492, + 0.006581225898116827, + 0.0026737565640360117, + 0.0025906688533723354, + 0.005196664948016405, + 0.0040800971910357475, + 0.004464633297175169, + 0.0032229649368673563, + 0.005402375012636185, + 0.006115203723311424, + 0.0071775647811591625, + 0.007281750440597534, + 0.005602754186838865, + 0.009750462137162685, + 0.006317722145467997, + 0.00479057477787137, + 0.006837987340986729, + 0.00949282106012106, + 0.00913052074611187, + 0.008481566794216633, + 0.01253848522901535, + 0.031093932688236237, + 0.008440230041742325, + 0.011357619427144527, + 0.004492383450269699, + 0.003274250775575638, + 0.009723911993205547, + 0.003579249605536461, + 0.002670734189450741, + 0.004878564737737179, + 0.002311916323378682, + 0.010153812356293201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.2339628040790558 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02332301251590252 + ], + "opposite_logprob_tokens": [ + -3.7733230590820312 + ], + "logprob_reward": -0.02332301251590252, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02332301251590252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17613133788108826, + 0.10781552642583847, + 0.12009473890066147, + 0.04643694683909416, + 0.1932867169380188, + 0.03677845001220703, + 0.12811695039272308, + 0.0428677573800087, + 0.03201957046985626, + 0.11645196378231049 + ], + "credit_map": [ + 0.02317175641655922, + 0.02038305625319481, + 0.01780872605741024, + 0.025750773027539253, + 0.02406929060816765, + 0.03201713413000107, + 0.019919514656066895, + 0.013011086732149124, + 0.009752225130796432, + 0.019194256514310837, + 0.010842172428965569, + 0.008721002377569675, + 0.011750854551792145, + 0.01961522549390793, + 0.020719731226563454, + 0.007220054976642132, + 0.014377934858202934, + 0.012021421454846859, + 0.016225239261984825, + 0.01133202575147152, + 0.01925688609480858, + 0.014893726445734501, + 0.017941519618034363, + 0.01404599193483591, + 0.005676445085555315, + 0.005340945441275835, + 0.007012385409325361, + 0.004543965682387352, + 0.005305437371134758, + 0.007001110352575779, + 0.0059313164092600346, + 0.005625339690595865, + 0.017262184992432594, + 0.01756812259554863, + 0.0331231951713562, + 0.022957520559430122, + 0.02077353000640869, + 0.021257027983665466, + 0.03813082352280617, + 0.02221430279314518, + 0.0035083992406725883, + 0.0033164452761411667, + 0.004414535593241453, + 0.003419192275032401, + 0.005259588360786438, + 0.005789242219179869, + 0.00508810905739665, + 0.005982937291264534, + 0.019664743915200233, + 0.02463029883801937, + 0.013173597864806652, + 0.01014159806072712, + 0.01745324581861496, + 0.013452163897454739, + 0.015247511677443981, + 0.014353793114423752, + 0.005243160296231508, + 0.007088561076670885, + 0.003822401398792863, + 0.003982442896813154, + 0.004874782171100378, + 0.007206853944808245, + 0.005334150046110153, + 0.005315405782312155, + 0.0033628176897764206, + 0.0038788991514593363, + 0.0031866428907960653, + 0.003599872114136815, + 0.005055351182818413, + 0.005023668985813856, + 0.0038500376977026463, + 0.004062276799231768, + 0.012889225035905838, + 0.007814685814082623, + 0.03196856006979942, + 0.009662440046668053, + 0.005450189579278231, + 0.008725222200155258, + 0.004255755338817835, + 0.03568588197231293 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.02332301251590252 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079511493444443 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079511493444443, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079511493444443, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10638890415430069, + 0.07973434031009674, + 0.33055567741394043, + 0.051220353692770004, + 0.08178581297397614, + 0.05995969474315643, + 0.04451463744044304, + 0.16205643117427826, + 0.03891540691256523, + 0.044868677854537964 + ], + "credit_map": [ + 0.01798888109624386, + 0.008749774657189846, + 0.01012993324548006, + 0.009455817751586437, + 0.019794179126620293, + 0.021901527419686317, + 0.009711580350995064, + 0.008657226338982582, + 0.007826592773199081, + 0.010986870154738426, + 0.012851017527282238, + 0.007626756560057402, + 0.011424621567130089, + 0.01101681124418974, + 0.012023151852190495, + 0.005978526547551155, + 0.0479150116443634, + 0.029042614623904228, + 0.03939281404018402, + 0.025339879095554352, + 0.05710896849632263, + 0.03928731009364128, + 0.053005654364824295, + 0.03946346417069435, + 0.004356944467872381, + 0.006766004487872124, + 0.006162350997328758, + 0.005480883177369833, + 0.007180459331721067, + 0.006566979456692934, + 0.008623733185231686, + 0.006083001848310232, + 0.006877767853438854, + 0.007642664480954409, + 0.015067021362483501, + 0.009823882021009922, + 0.008863856084644794, + 0.009600549936294556, + 0.014985686168074608, + 0.008924396708607674, + 0.006789205130189657, + 0.0071883974596858025, + 0.006440069992095232, + 0.005358725320547819, + 0.011437488719820976, + 0.011261454783380032, + 0.005551887676119804, + 0.005932471249252558, + 0.007070655468851328, + 0.00892072357237339, + 0.004866772796958685, + 0.004195569548755884, + 0.0052749705500900745, + 0.004028995521366596, + 0.005107460543513298, + 0.005049495492130518, + 0.022553076967597008, + 0.023797642439603806, + 0.015565826557576656, + 0.015099380165338516, + 0.020452454686164856, + 0.025206875056028366, + 0.021959638223052025, + 0.0174215417355299, + 0.003340634051710367, + 0.005460366141051054, + 0.0029934931080788374, + 0.0048825410194695, + 0.006827828474342823, + 0.0061857993714511395, + 0.004062052350491285, + 0.0051626963540911674, + 0.005114516243338585, + 0.006459889933466911, + 0.009203620254993439, + 0.00445469981059432, + 0.0038402238860726357, + 0.0048015424981713295, + 0.0027255485765635967, + 0.00826864130795002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 0 + ], + "selected_reward": -0.016079511493444443 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866431653499603 + ], + "opposite_logprob_tokens": [ + -3.0486643314361572 + ], + "logprob_reward": -0.04866431653499603, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866431653499603, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27091625332832336, + 0.08192701637744904, + 0.06364590674638748, + 0.029333719983696938, + 0.023604217916727066, + 0.01382360141724348, + 0.039807405322790146, + 0.0513526014983654, + 0.0679142102599144, + 0.35767507553100586 + ], + "credit_map": [ + 0.04786624759435654, + 0.022843612357974052, + 0.022251177579164505, + 0.03500876948237419, + 0.026814861223101616, + 0.05186337232589722, + 0.02765076421201229, + 0.036617446690797806, + 0.008086895570158958, + 0.011849211528897285, + 0.009333537891507149, + 0.006079733371734619, + 0.00909107830375433, + 0.015572857111692429, + 0.016027674078941345, + 0.00588603038340807, + 0.007929036393761635, + 0.007328659296035767, + 0.0072258408181369305, + 0.006837454158812761, + 0.00907325092703104, + 0.010509377345442772, + 0.008965174667537212, + 0.005777118261903524, + 0.0026911143213510513, + 0.0033373828046023846, + 0.0033593245316296816, + 0.003773922799155116, + 0.003163894871249795, + 0.004384471569210291, + 0.004480435512959957, + 0.004143171478062868, + 0.0020003400277346373, + 0.002762248506769538, + 0.0029597096145153046, + 0.0024865155573934317, + 0.0031449163798242807, + 0.0037583382800221443, + 0.003952170256525278, + 0.002539978129789233, + 0.001384423696435988, + 0.0014162507141008973, + 0.0014366568066179752, + 0.001173884142190218, + 0.0021405320148915052, + 0.002335710683837533, + 0.0013674574438482523, + 0.002568686380982399, + 0.0036904478911310434, + 0.004023431334644556, + 0.003238111501559615, + 0.004518183413892984, + 0.006809515412896872, + 0.004292635712772608, + 0.0072428639978170395, + 0.005992213264107704, + 0.005368656944483519, + 0.00497707212343812, + 0.007628633640706539, + 0.007281584199517965, + 0.006267060991376638, + 0.00780018325895071, + 0.0068220882676541805, + 0.005207323003560305, + 0.006281259935349226, + 0.006356153637170792, + 0.007861573249101639, + 0.0056172083131968975, + 0.009695167653262615, + 0.01882627047598362, + 0.005811954848468304, + 0.007464625872671604, + 0.03973039612174034, + 0.023846810683608055, + 0.09948387742042542, + 0.032199714332818985, + 0.016495756804943085, + 0.02456257864832878, + 0.020897550508379936, + 0.10045841336250305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.04866431653499603 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933352947235107 + ], + "opposite_logprob_tokens": [ + -0.6933352947235107 + ], + "logprob_reward": -0.6933352947235107, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933352947235107, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.26325252652168274, + 0.041941992938518524, + 0.029377851635217667, + 0.05011908710002899, + 0.12085450440645218, + 0.01757032983005047, + 0.036364298313856125, + 0.034242577850818634, + 0.048753395676612854, + 0.35752347111701965 + ], + "credit_map": [ + 0.05313803628087044, + 0.02657775767147541, + 0.020034560933709145, + 0.0274722371250391, + 0.025387447327375412, + 0.06177052482962608, + 0.022992482408881187, + 0.02587945945560932, + 0.0038914817851036787, + 0.00574279110878706, + 0.006019930355250835, + 0.0034221080131828785, + 0.005453568417578936, + 0.006293093785643578, + 0.007811829913407564, + 0.0033071856014430523, + 0.0031914603896439075, + 0.003969921730458736, + 0.0032313261181116104, + 0.002951989881694317, + 0.0042019314132630825, + 0.005932260770350695, + 0.0034763296134769917, + 0.0024226289242506027, + 0.004631850868463516, + 0.0056899054907262325, + 0.007560194469988346, + 0.006690004840493202, + 0.006597653031349182, + 0.005510313436388969, + 0.0064374483190476894, + 0.007001710589975119, + 0.010431359522044659, + 0.010349181480705738, + 0.02094179391860962, + 0.014000441879034042, + 0.013588429428637028, + 0.013950117863714695, + 0.0238129124045372, + 0.013780252076685429, + 0.0012354128994047642, + 0.0019103578524664044, + 0.001601057592779398, + 0.0018112262478098273, + 0.0035542724654078484, + 0.0023624077439308167, + 0.0031544126104563475, + 0.0019411819521337748, + 0.004236974753439426, + 0.004446008242666721, + 0.0038119154050946236, + 0.002947769593447447, + 0.004967548418790102, + 0.00627316627651453, + 0.004757101647555828, + 0.0049238125793635845, + 0.004200729541480541, + 0.004832044243812561, + 0.003019861178472638, + 0.003979810513556004, + 0.003863742807880044, + 0.00444153742864728, + 0.004930997267365456, + 0.0049738516099750996, + 0.0047503141686320305, + 0.005658736918121576, + 0.005793610122054815, + 0.006168809253722429, + 0.00645596394315362, + 0.008720656856894493, + 0.005862286314368248, + 0.005343012511730194, + 0.03916875645518303, + 0.014684522524476051, + 0.11121495813131332, + 0.02856440097093582, + 0.015093915164470673, + 0.02577502466738224, + 0.012945052236318588, + 0.11007681488990784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.6933352947235107 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "sweeping panoramic shot, the camera slowly pulling back to reveal the panorama of a precipitous canyon, etched in the style of Tsui Hark's martial arts fantasy. In the dim light of early dusk, jagged, knife-cut cliffs rise like the skeletons of a giant beast. A damp mist lingers in the bottomless chasm, creating an atmosphere of mystery and danger. Within this silence, the dense forest of blue devil irises, covering the entire valley floor and portions of the sheer rock faces, their tightly closed deep blue, almost black buds suddenly open at the same moment, as if receiving a silent command. Each petal, as thin as a cicada's wing, burst open with a visible speed, its edges gleaming with an eerie electric blue light. A core of pure energy pulsed in unison, releasing a silent roar of energy—a force that transformed into a transparent, visible shockwave that rapidly spread from the center of the sea of flowers. Wherever it passed, the air violently twisted and rippled, even sending tiny rocks tumbling from the cliffs. This strange blue light instantly illuminated the entire valley, tinting the outlines of every rock a deep blue and casting long, twisting black shadows. It perfectly captured the spectacular scene of the valley being instantly engulfed by a supernatural force, creating a fantastical atmosphere that was both magnificent and oppressive.", + "seed": 18410, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_5473_seed_18410_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.21560589876025915, + "vlm_logprob": -0.21560589876025915, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.21560589876025915, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32519587874412537 + ], + "opposite_logprob_tokens": [ + -1.3251959085464478 + ], + "logprob_reward": -0.32519587874412537, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32519587874412537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09623467177152634, + 0.08191823214292526, + 0.066536083817482, + 0.0579497329890728, + 0.07595990598201752, + 0.3912694752216339, + 0.0693550556898117, + 0.09601336717605591, + 0.050587646663188934, + 0.01417587324976921 + ], + "credit_map": [ + 0.014736498706042767, + 0.015202412381768227, + 0.008439946919679642, + 0.007301344070583582, + 0.011973824352025986, + 0.012704244814813137, + 0.014687665738165379, + 0.011188735254108906, + 0.008538440801203251, + 0.010664163157343864, + 0.009958920069038868, + 0.009039639495313168, + 0.00847108568996191, + 0.011336242780089378, + 0.015686288475990295, + 0.008223457261919975, + 0.007210346404463053, + 0.006460919044911861, + 0.007856310345232487, + 0.005582910496741533, + 0.007108551450073719, + 0.005525072570890188, + 0.017816215753555298, + 0.008975756354629993, + 0.00832721870392561, + 0.008664339780807495, + 0.005065483041107655, + 0.005391373764723539, + 0.011245771311223507, + 0.010465226136147976, + 0.004273426719009876, + 0.00451689213514328, + 0.0061239395290613174, + 0.008288882672786713, + 0.0058900341391563416, + 0.006293462589383125, + 0.01868588477373123, + 0.015683375298976898, + 0.007988976314663887, + 0.007005354389548302, + 0.022057872265577316, + 0.16618454456329346, + 0.015592440031468868, + 0.01637018658220768, + 0.026950698345899582, + 0.11101993173360825, + 0.01184073369950056, + 0.02125304937362671, + 0.006913323421031237, + 0.008824796415865421, + 0.010929109528660774, + 0.005856793373823166, + 0.006394432857632637, + 0.015681955963373184, + 0.00527452165260911, + 0.009480123408138752, + 0.00911848247051239, + 0.010346469469368458, + 0.015251652337610722, + 0.013002954423427582, + 0.012417733669281006, + 0.02041720412671566, + 0.0064827315509319305, + 0.008976142853498459, + 0.0047588106244802475, + 0.01472645252943039, + 0.007748201955109835, + 0.0041534751653671265, + 0.00553804449737072, + 0.005560444667935371, + 0.0032928914297372103, + 0.004809327889233828, + 0.0012976732105016708, + 0.00186199601739645, + 0.0021782240364700556, + 0.0018110743258148432, + 0.001945652300491929, + 0.0025548769626766443, + 0.001249512075446546, + 0.0012768652522936463 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.32519587874412537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023318471387028694 + ], + "opposite_logprob_tokens": [ + -3.7733185291290283 + ], + "logprob_reward": -0.023318471387028694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023318471387028694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21893510222434998, + 0.3007708787918091, + 0.05002213642001152, + 0.05607570707798004, + 0.08153706043958664, + 0.09812136739492416, + 0.04289169982075691, + 0.07512879371643066, + 0.04248431324958801, + 0.034032952040433884 + ], + "credit_map": [ + 0.022271281108260155, + 0.030247312039136887, + 0.01946403831243515, + 0.013506537303328514, + 0.02377413772046566, + 0.03955669701099396, + 0.01980499178171158, + 0.050310082733631134, + 0.03629465773701668, + 0.05084233731031418, + 0.02049913816154003, + 0.0242458488792181, + 0.03232860937714577, + 0.05548853799700737, + 0.05121181160211563, + 0.02985992841422558, + 0.00580205861479044, + 0.006789751350879669, + 0.006799571681767702, + 0.005456557963043451, + 0.00502143194898963, + 0.004431563895195723, + 0.010143272578716278, + 0.005577924661338329, + 0.006877128034830093, + 0.008556213229894638, + 0.004434660077095032, + 0.004995404276996851, + 0.012932748533785343, + 0.008557697758078575, + 0.004312042612582445, + 0.005409806966781616, + 0.006139552220702171, + 0.012961316853761673, + 0.0037935434374958277, + 0.00455947732552886, + 0.02498655393719673, + 0.01702756993472576, + 0.006400299724191427, + 0.005668742582201958, + 0.0059252954088151455, + 0.0289031770080328, + 0.00736684026196599, + 0.008852103725075722, + 0.008064609952270985, + 0.02136925421655178, + 0.010732350870966911, + 0.0069077350199222565, + 0.00460963835939765, + 0.004583762492984533, + 0.005230261012911797, + 0.004029917065054178, + 0.004930141847580671, + 0.008701196871697903, + 0.004787936341017485, + 0.00601884163916111, + 0.005823098588734865, + 0.012173101305961609, + 0.007861679419875145, + 0.006629801355302334, + 0.01127924956381321, + 0.01727774739265442, + 0.005742567125707865, + 0.008341539651155472, + 0.005127197131514549, + 0.014453437179327011, + 0.00496635353192687, + 0.002730146050453186, + 0.004703728947788477, + 0.0055107721127569675, + 0.002070257905870676, + 0.0029224189929664135, + 0.003657223656773567, + 0.005600450560450554, + 0.004376756958663464, + 0.003525619162246585, + 0.0036960358265787363, + 0.007194836158305407, + 0.0025533000007271767, + 0.0034287311136722565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.023318471387028694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009795336984097958 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009795336984097958, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009795336984097958, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23430383205413818, + 0.11581411212682724, + 0.07867782562971115, + 0.0677134245634079, + 0.05875968933105469, + 0.23220153152942657, + 0.07038947194814682, + 0.09575584530830383, + 0.03130360320210457, + 0.015080668963491917 + ], + "credit_map": [ + 0.015016823075711727, + 0.025305677205324173, + 0.026805954053997993, + 0.023919519037008286, + 0.0396585613489151, + 0.025804413482546806, + 0.03748040273785591, + 0.0403125025331974, + 0.010072953999042511, + 0.014972196891903877, + 0.019572611898183823, + 0.01261433120816946, + 0.016574595123529434, + 0.01364297978579998, + 0.013446150347590446, + 0.014918305911123753, + 0.006888361647725105, + 0.009184732101857662, + 0.011952072381973267, + 0.007180908694863319, + 0.011028698645532131, + 0.010112756863236427, + 0.011500292457640171, + 0.010830004699528217, + 0.006876442581415176, + 0.007832469418644905, + 0.009365113452076912, + 0.007525425869971514, + 0.012844743207097054, + 0.010455705225467682, + 0.0054336548782885075, + 0.007379879709333181, + 0.0060018133372068405, + 0.008663051761686802, + 0.005278323777019978, + 0.005032426677644253, + 0.007874790579080582, + 0.012132171541452408, + 0.007555749733000994, + 0.006221367511898279, + 0.01142213772982359, + 0.08614812791347504, + 0.011702357791364193, + 0.01480566244572401, + 0.014927615411579609, + 0.06258358061313629, + 0.016727369278669357, + 0.0138847129419446, + 0.0065457383170723915, + 0.007609867490828037, + 0.009328042156994343, + 0.0074283950962126255, + 0.008848179131746292, + 0.011699382215738297, + 0.007992840372025967, + 0.010937029495835304, + 0.007410676218569279, + 0.010164760053157806, + 0.014064931310713291, + 0.012667378410696983, + 0.011027499102056026, + 0.01308189146220684, + 0.01341995969414711, + 0.013918749988079071, + 0.0031441878527402878, + 0.004253605846315622, + 0.006718536838889122, + 0.0033575622364878654, + 0.003461581887677312, + 0.005214712582528591, + 0.0027973719406872988, + 0.002356044016778469, + 0.0017913675401359797, + 0.002264541108161211, + 0.0028653594199568033, + 0.002125304425135255, + 0.0014948693569749594, + 0.0020981894340366125, + 0.001148501643911004, + 0.0012925381306558847 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.009795336984097958 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02635768987238407 + ], + "opposite_logprob_tokens": [ + -3.651357650756836 + ], + "logprob_reward": -0.02635768987238407, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02635768987238407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1416090726852417, + 0.06785125285387039, + 0.12360890209674835, + 0.37916260957717896, + 0.08997366577386856, + 0.038174450397491455, + 0.03911852091550827, + 0.03584672883152962, + 0.05902409553527832, + 0.025630755349993706 + ], + "credit_map": [ + 0.01951189897954464, + 0.016130659729242325, + 0.01675228588283062, + 0.01126403920352459, + 0.014869928359985352, + 0.02246742695569992, + 0.012766695581376553, + 0.02784612961113453, + 0.005923338234424591, + 0.007839662954211235, + 0.005557472817599773, + 0.005382999312132597, + 0.013157038018107414, + 0.008704946376383305, + 0.009921771474182606, + 0.01136402040719986, + 0.013031103648245335, + 0.014672627672553062, + 0.016537196934223175, + 0.010647748596966267, + 0.017727956175804138, + 0.010112875141203403, + 0.027881890535354614, + 0.012997507117688656, + 0.049766656011343, + 0.05746461823582649, + 0.03937152773141861, + 0.03878459334373474, + 0.06445770710706711, + 0.05982052534818649, + 0.030149327591061592, + 0.039347678422927856, + 0.007172748912125826, + 0.008193453773856163, + 0.005568006541579962, + 0.0066873240284621716, + 0.028656218200922012, + 0.018534185364842415, + 0.007957526482641697, + 0.007204211782664061, + 0.0027459762059152126, + 0.01384322252124548, + 0.0022512308787554502, + 0.0022541119251400232, + 0.0033772271126508713, + 0.008362898603081703, + 0.001994085032492876, + 0.0033456990495324135, + 0.003630958730354905, + 0.004185852594673634, + 0.004745940677821636, + 0.004663811530917883, + 0.00437640305608511, + 0.00723894452676177, + 0.0038105242419987917, + 0.006466085556894541, + 0.00325387972407043, + 0.007388059049844742, + 0.004473998676985502, + 0.0038376052398234606, + 0.004315417259931564, + 0.005514959804713726, + 0.004036193247884512, + 0.0030266144312918186, + 0.006891506724059582, + 0.019536782056093216, + 0.005687506403774023, + 0.0037304949946701527, + 0.00680040568113327, + 0.008985496126115322, + 0.002766498364508152, + 0.00462540565058589, + 0.0027061132714152336, + 0.003807635512202978, + 0.0034479449968785048, + 0.0023808814585208893, + 0.0025044740177690983, + 0.005292534362524748, + 0.0024478088598698378, + 0.003043362870812416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 2 + ], + "selected_reward": -0.02635768987238407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6933621168136597 + ], + "opposite_logprob_tokens": [ + -0.6933621168136597 + ], + "logprob_reward": -0.6933621168136597, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6933621168136597, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1342972069978714, + 0.11013677716255188, + 0.1687961220741272, + 0.08104010671377182, + 0.07523252815008163, + 0.20045794546604156, + 0.0802466869354248, + 0.0525154247879982, + 0.0670742467045784, + 0.03020288236439228 + ], + "credit_map": [ + 0.016274087131023407, + 0.021604610607028008, + 0.014360812492668629, + 0.009751624427735806, + 0.015631988644599915, + 0.020126797258853912, + 0.014502574689686298, + 0.022044703364372253, + 0.012464134022593498, + 0.011698811315000057, + 0.009421296417713165, + 0.011451216414570808, + 0.01531678345054388, + 0.01918618753552437, + 0.018595349043607712, + 0.012002997100353241, + 0.017765559256076813, + 0.01954673044383526, + 0.02091144397854805, + 0.013851821422576904, + 0.020003963261842728, + 0.011942461133003235, + 0.04969901591539383, + 0.01507512852549553, + 0.009259091690182686, + 0.012186129577457905, + 0.005976567976176739, + 0.0060251690447330475, + 0.01934254728257656, + 0.012352606281638145, + 0.00860636867582798, + 0.007291629444807768, + 0.007970563136041164, + 0.011863837018609047, + 0.007438245229423046, + 0.005442745517939329, + 0.017414681613445282, + 0.013018680736422539, + 0.005580454133450985, + 0.0065033212304115295, + 0.011309286579489708, + 0.08335286378860474, + 0.00799459870904684, + 0.008907015435397625, + 0.012834419496357441, + 0.05745869502425194, + 0.006770158652216196, + 0.011830919422209263, + 0.007127400487661362, + 0.009343565441668034, + 0.010710704140365124, + 0.010201340541243553, + 0.008318567648530006, + 0.012857959605753422, + 0.012968026101589203, + 0.008719116449356079, + 0.0038339029997587204, + 0.009123110212385654, + 0.005746930837631226, + 0.006184994243085384, + 0.006786561571061611, + 0.011770626530051231, + 0.004997043404728174, + 0.004072254523634911, + 0.0061906129121780396, + 0.01674998737871647, + 0.009406776167452335, + 0.005243884865194559, + 0.007408057805150747, + 0.009914645925164223, + 0.006065388210117817, + 0.0060948943719267845, + 0.0034882789477705956, + 0.005010395776480436, + 0.005003721918910742, + 0.0030836353544145823, + 0.00257298001088202, + 0.004917427897453308, + 0.002813789527863264, + 0.0033126529306173325 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.6933621168136597 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "sweeping panoramic shot, the camera slowly pulling back to reveal the panorama of a precipitous canyon, etched in the style of Tsui Hark's martial arts fantasy. In the dim light of early dusk, jagged, knife-cut cliffs rise like the skeletons of a giant beast. A damp mist lingers in the bottomless chasm, creating an atmosphere of mystery and danger. Within this silence, the dense forest of blue devil irises, covering the entire valley floor and portions of the sheer rock faces, their tightly closed deep blue, almost black buds suddenly open at the same moment, as if receiving a silent command. Each petal, as thin as a cicada's wing, burst open with a visible speed, its edges gleaming with an eerie electric blue light. A core of pure energy pulsed in unison, releasing a silent roar of energy—a force that transformed into a transparent, visible shockwave that rapidly spread from the center of the sea of flowers. Wherever it passed, the air violently twisted and rippled, even sending tiny rocks tumbling from the cliffs. This strange blue light instantly illuminated the entire valley, tinting the outlines of every rock a deep blue and casting long, twisting black shadows. It perfectly captured the spectacular scene of the valley being instantly engulfed by a supernatural force, creating a fantastical atmosphere that was both magnificent and oppressive.", + "seed": 18411, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_0/index_5473_seed_18411_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20310870539397002, + "vlm_logprob": -0.20310870539397002, + "vlm_yes_no_margin": 2.324999976158142, + "vlm_token_credit": -0.20310870539397002, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1511579304933548 + ], + "opposite_logprob_tokens": [ + -2.026157855987549 + ], + "logprob_reward": -0.1511579304933548, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.1511579304933548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2932223975658417, + 0.049478836357593536, + 0.11373628675937653, + 0.09514417499303818, + 0.08972042053937912, + 0.0659773126244545, + 0.07758812606334686, + 0.12987971305847168, + 0.05380372330546379, + 0.03144904226064682 + ], + "credit_map": [ + 0.039007749408483505, + 0.03569073975086212, + 0.028253372758626938, + 0.025239884853363037, + 0.07085508108139038, + 0.03501618281006813, + 0.02411770448088646, + 0.035041678696870804, + 0.007230591960251331, + 0.007075767032802105, + 0.005285909399390221, + 0.004387610591948032, + 0.008061207830905914, + 0.010318167507648468, + 0.003990913275629282, + 0.0031286654993891716, + 0.020371804013848305, + 0.007903525605797768, + 0.018051834776997566, + 0.009951310232281685, + 0.012060827575623989, + 0.028096791356801987, + 0.007605993654578924, + 0.009694191627204418, + 0.025123342871665955, + 0.005084543488919735, + 0.009344280697405338, + 0.006152527406811714, + 0.00823298655450344, + 0.031124036759138107, + 0.005100605543702841, + 0.004981856793165207, + 0.015729472041130066, + 0.006040152162313461, + 0.008398168720304966, + 0.007444123737514019, + 0.013191601261496544, + 0.027137180790305138, + 0.005722022149711847, + 0.0060576992109417915, + 0.009923392906785011, + 0.005748304072767496, + 0.004879339132457972, + 0.006084539461880922, + 0.005918743554502726, + 0.007229783572256565, + 0.00864618830382824, + 0.017547018826007843, + 0.0067948815412819386, + 0.008464002050459385, + 0.0073862080462276936, + 0.008878884837031364, + 0.018810857087373734, + 0.015482209622859955, + 0.006819402799010277, + 0.004951683804392815, + 0.008682030253112316, + 0.011186178773641586, + 0.013580992817878723, + 0.011899272911250591, + 0.013330545276403427, + 0.011081007309257984, + 0.047824323177337646, + 0.012295365333557129, + 0.006166848819702864, + 0.004915686324238777, + 0.005995252635329962, + 0.009186956100165844, + 0.005585153121501207, + 0.006210133899003267, + 0.010081261396408081, + 0.005662430543452501, + 0.004321670159697533, + 0.0034521904308348894, + 0.003839303972199559, + 0.0032242138404399157, + 0.005286758299916983, + 0.0064250449649989605, + 0.0026183631271123886, + 0.0022814972326159477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.1511579304933548 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026369649916887283 + ], + "opposite_logprob_tokens": [ + -3.651369571685791 + ], + "logprob_reward": -0.026369649916887283, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026369649916887283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4691561162471771, + 0.0315152145922184, + 0.06021953374147415, + 0.04621633514761925, + 0.11201347410678864, + 0.07871710509061813, + 0.022127751260995865, + 0.06217418238520622, + 0.09396360069513321, + 0.023896757513284683 + ], + "credit_map": [ + 0.05789388716220856, + 0.056222885847091675, + 0.058785371482372284, + 0.03936126083135605, + 0.10671662539243698, + 0.060272492468357086, + 0.03895711153745651, + 0.050946421921253204, + 0.003281108569353819, + 0.0036692789290100336, + 0.00420461455360055, + 0.0035061235539615154, + 0.006036675069481134, + 0.004455332178622484, + 0.0030825070571154356, + 0.0032795656006783247, + 0.005583120509982109, + 0.004754100926220417, + 0.00987344328314066, + 0.0074103595688939095, + 0.008759832940995693, + 0.007545221596956253, + 0.005762708373367786, + 0.010530740953981876, + 0.006083469372242689, + 0.004291541408747435, + 0.004541506990790367, + 0.002924797823652625, + 0.012339077889919281, + 0.008727493695914745, + 0.0038433545269072056, + 0.0034650887828320265, + 0.025968480855226517, + 0.005310336593538523, + 0.005870956927537918, + 0.0053539699874818325, + 0.01732007972896099, + 0.03977782651782036, + 0.008066472597420216, + 0.0043453373946249485, + 0.016820576041936874, + 0.005389942787587643, + 0.005380329675972462, + 0.006015700753778219, + 0.0059852320700883865, + 0.01252142433077097, + 0.008749224245548248, + 0.01785467006266117, + 0.0018772062612697482, + 0.0027566137723624706, + 0.0020801795180886984, + 0.001979772699996829, + 0.003639100817963481, + 0.002465510042384267, + 0.005271583329886198, + 0.002057783305644989, + 0.0042488910257816315, + 0.005014028865844011, + 0.009772582910954952, + 0.005766029004007578, + 0.00684661278501153, + 0.006112542003393173, + 0.0190057922154665, + 0.005407693795859814, + 0.013677453622221947, + 0.006471552420407534, + 0.009547658264636993, + 0.013547207228839397, + 0.008114592172205448, + 0.012169990688562393, + 0.008055302314460278, + 0.022379834204912186, + 0.0027184232603758574, + 0.0026334431022405624, + 0.0028040255419909954, + 0.002246052958071232, + 0.00374068645760417, + 0.004854612518101931, + 0.002254252089187503, + 0.002645260887220502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.026369649916887283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978896301239729 + ], + "opposite_logprob_tokens": [ + -4.634788990020752 + ], + "logprob_reward": -0.00978896301239729, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978896301239729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36677631735801697, + 0.10844876617193222, + 0.05979158729314804, + 0.06381602585315704, + 0.04886766895651817, + 0.03144701197743416, + 0.05257734656333923, + 0.04375072941184044, + 0.17128534615039825, + 0.05323924496769905 + ], + "credit_map": [ + 0.05413186922669411, + 0.06315763294696808, + 0.025806795805692673, + 0.028197811916470528, + 0.08149190992116928, + 0.06797489523887634, + 0.024018889293074608, + 0.02199648693203926, + 0.012327504344284534, + 0.011561795137822628, + 0.01820509321987629, + 0.011572507210075855, + 0.015414985828101635, + 0.015948906540870667, + 0.012131970375776291, + 0.011285992339253426, + 0.006562729366123676, + 0.007473915349692106, + 0.007340413052588701, + 0.00776352034881711, + 0.009009119123220444, + 0.008725693449378014, + 0.005202307365834713, + 0.007713884115219116, + 0.011416175402700901, + 0.005117717664688826, + 0.007553258910775185, + 0.0054387059062719345, + 0.010616803541779518, + 0.009372193366289139, + 0.0060914428904652596, + 0.008209723979234695, + 0.007818478159606457, + 0.005344314966350794, + 0.00509988609701395, + 0.00391525449231267, + 0.005931729916483164, + 0.010455580428242683, + 0.0042714765295386314, + 0.006030940916389227, + 0.00444811349734664, + 0.002692625392228365, + 0.004345590714365244, + 0.0030041011050343513, + 0.004080495331436396, + 0.0036073653027415276, + 0.005561457015573978, + 0.00370725872926414, + 0.005185614340007305, + 0.004282879643142223, + 0.009283607825636864, + 0.004107157699763775, + 0.010168887674808502, + 0.008239410817623138, + 0.0076051140204072, + 0.0037046661600470543, + 0.004243187140673399, + 0.0035904760006815195, + 0.006837137974798679, + 0.0038252065423876047, + 0.0037148925475776196, + 0.003615474095568061, + 0.014095366932451725, + 0.003828981891274452, + 0.030480360612273216, + 0.012974310666322708, + 0.017277196049690247, + 0.015634890645742416, + 0.014181227423250675, + 0.022480648010969162, + 0.02322656475007534, + 0.03503013774752617, + 0.0070931026712059975, + 0.003975631669163704, + 0.004678184166550636, + 0.004196071997284889, + 0.017786294221878052, + 0.009398314170539379, + 0.003233540104702115, + 0.0028781003784388304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00978896301239729 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520797848701477 + ], + "opposite_logprob_tokens": [ + -1.502079725265503 + ], + "logprob_reward": -0.2520797848701477, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520797848701477, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08879534900188446, + 0.05667679384350777, + 0.07213567197322845, + 0.0951845645904541, + 0.49001815915107727, + 0.0661424845457077, + 0.052888114005327225, + 0.020428350195288658, + 0.04086849093437195, + 0.01686197705566883 + ], + "credit_map": [ + 0.010386282578110695, + 0.008315772749483585, + 0.013577541336417198, + 0.008765032514929771, + 0.017231743782758713, + 0.012357259169220924, + 0.008266204968094826, + 0.009895507246255875, + 0.009745646268129349, + 0.0037307734601199627, + 0.006699522025883198, + 0.005776551552116871, + 0.007911461405456066, + 0.01271018572151661, + 0.004965243395417929, + 0.005137407686561346, + 0.014789756387472153, + 0.0053564743138849735, + 0.009000126272439957, + 0.00644112890586257, + 0.0078047337010502815, + 0.018430976197123528, + 0.003978497348725796, + 0.00633397838100791, + 0.026101911440491676, + 0.006463476922363043, + 0.007528504356741905, + 0.00417120847851038, + 0.009812607429921627, + 0.03175262734293938, + 0.004406708292663097, + 0.004947517532855272, + 0.12850426137447357, + 0.018400803208351135, + 0.017347432672977448, + 0.018194660544395447, + 0.07024143636226654, + 0.18918530642986298, + 0.029049599543213844, + 0.019094660878181458, + 0.01376692857593298, + 0.007221410516649485, + 0.004656754434108734, + 0.004341759718954563, + 0.007276495452970266, + 0.014936202205717564, + 0.007290743291378021, + 0.0066521926783025265, + 0.00620394479483366, + 0.004216712433844805, + 0.010859479196369648, + 0.004972805269062519, + 0.009962175041437149, + 0.0072961472906172276, + 0.005720478482544422, + 0.003656371496617794, + 0.002013841411098838, + 0.0016790474765002728, + 0.005625950638204813, + 0.002093291375786066, + 0.0019609546288847923, + 0.0016650826437398791, + 0.003918931819498539, + 0.0014712491538375616, + 0.007858870550990105, + 0.003372122999280691, + 0.00422583008185029, + 0.005300424061715603, + 0.0034772497601807117, + 0.006065627560019493, + 0.005113624501973391, + 0.005454739090055227, + 0.0019489320693537593, + 0.0018580385949462652, + 0.0017999042756855488, + 0.0011828412534669042, + 0.004084393847733736, + 0.0034429102670401335, + 0.0012154701398685575, + 0.0013294873060658574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.2520797848701477 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a sweeping panorama that slowly pulls back to reveal a precipitous canyon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are jagged, knife-cut cliffs visible rising from the canyon floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a dense forest of blue devil irises cover the valley floor and portions of the rock faces?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the deep blue buds of the irises open simultaneously, emitting an eerie electric blue light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is a transparent, visible shockwave shown spreading from the flowers and causing rocks to tumble from the cliffs?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576147198677063 + ], + "opposite_logprob_tokens": [ + -0.826147198677063 + ], + "logprob_reward": -0.576147198677063, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576147198677063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16224142909049988, + 0.05762116611003876, + 0.05479849502444267, + 0.07066460698843002, + 0.13264666497707367, + 0.05208142474293709, + 0.15308450162410736, + 0.07448334991931915, + 0.20475877821445465, + 0.03761957958340645 + ], + "credit_map": [ + 0.0218947920948267, + 0.025562986731529236, + 0.013784592971205711, + 0.010691300965845585, + 0.046174291521310806, + 0.020833605900406837, + 0.011032708920538425, + 0.012267145328223705, + 0.007060189265757799, + 0.005301722325384617, + 0.006477035116404295, + 0.006434018258005381, + 0.01332481112331152, + 0.009735038504004478, + 0.004687685985118151, + 0.004600663669407368, + 0.0111018605530262, + 0.005072630941867828, + 0.00662032887339592, + 0.0037242870312184095, + 0.006677008233964443, + 0.013760014437139034, + 0.0034674680791795254, + 0.004374898504465818, + 0.016963720321655273, + 0.006312962155789137, + 0.005587913561612368, + 0.004086107015609741, + 0.009510992094874382, + 0.018871286883950233, + 0.0036785919219255447, + 0.005653022788465023, + 0.028003664687275887, + 0.00727656576782465, + 0.009041819721460342, + 0.00636201212182641, + 0.021249735727906227, + 0.04316113889217377, + 0.01077908556908369, + 0.006772648077458143, + 0.0073901754803955555, + 0.006619041785597801, + 0.007114374078810215, + 0.00596745265647769, + 0.00565623864531517, + 0.005556800868362188, + 0.008205409161746502, + 0.00557193486019969, + 0.01721203327178955, + 0.01172809861600399, + 0.03191431239247322, + 0.014051408506929874, + 0.02626129984855652, + 0.018535401672124863, + 0.02290918119251728, + 0.010472767986357212, + 0.007301046513020992, + 0.007518223486840725, + 0.008357777260243893, + 0.007366807200014591, + 0.00816965103149414, + 0.008095336146652699, + 0.02109609544277191, + 0.006578408181667328, + 0.04199874773621559, + 0.014638698659837246, + 0.01900067739188671, + 0.022671110928058624, + 0.022710198536515236, + 0.04392356798052788, + 0.02238130196928978, + 0.017434479668736458, + 0.005229060538113117, + 0.0038180737756192684, + 0.003324114019051194, + 0.002632804913446307, + 0.011385878548026085, + 0.006511773448437452, + 0.002561671193689108, + 0.0021562043111771345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.576147198677063 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_tensor([8100, 8100])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_tensor([8100, 8100])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5e4bbe842cd06d7b9b8d9791d23d50961eaf9ede --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_tensor([8100, 8100])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a fluid, slowly rotating 360-degree camera, a dense vortex of countless flying sawdust and fine dust particles hovers in the center of a craftsman's workshop. Warm, golden afternoon sunlight slants through dusty windows, illuminating each suspended particle like a shimmering swarm of golden insects. A softly blurred backdrop features a wall covered in tools and piles of lumber. As the camera continues to pan, the initially chaotic, fluttering debris begins to be drawn by an invisible force, rapidly and orderly converging toward the center. They begin to weave and intertwine, larger pieces decisively forming the outlines of the chair's legs and backrest, while finer pieces of sawdust and dust seamlessly fill and compact the space, forming themselves as if possessing a life of their own. Finally, the rotation came to an abrupt halt, and the flying mass of matter had completely solidified, transforming into a wooden chair with smooth lines and solid texture. Its surface was smooth, and the warm oak grain was clearly revealed under the light. It stood quietly in place. The whole process was full of a focused and magical sense of creation, turning disorder into order and presenting a fascinating visual spectacle.", + "seed": 333, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_333_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0496982529759407, + "vlm_logprob": -0.0496982529759407, + "vlm_yes_no_margin": 3.375000023841858, + "vlm_token_credit": -0.0496982529759407, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019672170281410217 + ], + "opposite_logprob_tokens": [ + -4.144672393798828 + ], + "logprob_reward": -0.019672170281410217, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019672170281410217, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09184223413467407, + 0.07739730924367905, + 0.10027714818716049, + 0.03122054785490036, + 0.06865094602108002, + 0.10319560021162033, + 0.2530958354473114, + 0.03975798934698105, + 0.20244960486888885, + 0.03211287409067154 + ], + "credit_map": [ + 0.01475268043577671, + 0.01108148880302906, + 0.01203281432390213, + 0.011122907511889935, + 0.009155974723398685, + 0.011342101730406284, + 0.01458880677819252, + 0.00776545749977231, + 0.007013383321464062, + 0.012377197854220867, + 0.006586072035133839, + 0.011185630224645138, + 0.010200811550021172, + 0.012632281519472599, + 0.01000791136175394, + 0.007394020911306143, + 0.007989170029759407, + 0.01333316694945097, + 0.007002479396760464, + 0.014290871098637581, + 0.023339472711086273, + 0.010622642002999783, + 0.01590903289616108, + 0.007790305186063051, + 0.0021743702236562967, + 0.003719268599525094, + 0.0021386081352829933, + 0.004855518229305744, + 0.006261219270527363, + 0.0050953347235918045, + 0.004157942719757557, + 0.00281828292645514, + 0.004780000075697899, + 0.00894985068589449, + 0.004815311636775732, + 0.011233926750719547, + 0.009846789762377739, + 0.008368728682398796, + 0.01291304174810648, + 0.007743288762867451, + 0.020574385300278664, + 0.019181516021490097, + 0.008139466866850853, + 0.006486908067017794, + 0.013172421604394913, + 0.010437035001814365, + 0.006809213664382696, + 0.018394645303487778, + 0.04077444598078728, + 0.022282756865024567, + 0.036107782274484634, + 0.01513436809182167, + 0.06706944108009338, + 0.0257977694272995, + 0.019058775156736374, + 0.02687044069170952, + 0.004650925751775503, + 0.0025176997296512127, + 0.006567124743014574, + 0.0034915367141366005, + 0.006516281049698591, + 0.0062897298485040665, + 0.004159322939813137, + 0.005565362051129341, + 0.07723341137170792, + 0.00690539600327611, + 0.06005735695362091, + 0.008135293610394001, + 0.012032714672386646, + 0.008664419874548912, + 0.017969008535146713, + 0.011451980099081993, + 0.003672113409265876, + 0.0020393820013850927, + 0.005656953435391188, + 0.004071780014783144, + 0.002219205489382148, + 0.005470856558531523, + 0.0038258207496255636, + 0.005156756844371557 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 5 + ], + "selected_reward": -0.019672170281410217 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811095654964447 + ], + "opposite_logprob_tokens": [ + -3.2881109714508057 + ], + "logprob_reward": -0.03811095654964447, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811095654964447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11550109833478928, + 0.1209636926651001, + 0.3045238256454468, + 0.07271766662597656, + 0.02920575626194477, + 0.0909361019730568, + 0.10871696472167969, + 0.05973472073674202, + 0.07362352311611176, + 0.024076633155345917 + ], + "credit_map": [ + 0.015128805302083492, + 0.01712035946547985, + 0.01278769038617611, + 0.008728383108973503, + 0.016808195039629936, + 0.01772790402173996, + 0.015839941799640656, + 0.011359824799001217, + 0.01612682268023491, + 0.02061900869011879, + 0.013307017274200916, + 0.015093275345861912, + 0.013875247910618782, + 0.013425718992948532, + 0.018324045464396477, + 0.010192560032010078, + 0.031125545501708984, + 0.04262812063097954, + 0.021531933918595314, + 0.04535364732146263, + 0.06671624630689621, + 0.02823605388402939, + 0.04705209657549858, + 0.02188018709421158, + 0.007801728323101997, + 0.01374048087745905, + 0.006089941132813692, + 0.009546330198645592, + 0.0074301594868302345, + 0.009214035235345364, + 0.008483213372528553, + 0.010411773808300495, + 0.002530731027945876, + 0.00392858125269413, + 0.0037093672435730696, + 0.004705488216131926, + 0.004539976362138987, + 0.0032658101990818977, + 0.004175372887402773, + 0.002350429305806756, + 0.00783916749060154, + 0.016553416848182678, + 0.017626261338591576, + 0.011994566768407822, + 0.01085079088807106, + 0.007702153641730547, + 0.007865789346396923, + 0.010503954254090786, + 0.015667300671339035, + 0.009196129627525806, + 0.018320992588996887, + 0.006746075581759214, + 0.025879377499222755, + 0.013182290829718113, + 0.008291538804769516, + 0.011433255858719349, + 0.012317984364926815, + 0.004615600686520338, + 0.012250127270817757, + 0.004137109965085983, + 0.006825543940067291, + 0.006687966641038656, + 0.005871591158211231, + 0.007028798572719097, + 0.028315797448158264, + 0.0031064501963555813, + 0.022463202476501465, + 0.0032903568353503942, + 0.0046485913917422295, + 0.002491485793143511, + 0.006767933256924152, + 0.0025397061835974455, + 0.0031041400507092476, + 0.001421747263520956, + 0.003819811623543501, + 0.0058179558254778385, + 0.0017725315410643816, + 0.003062805626541376, + 0.0033585461787879467, + 0.0017190960934385657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.03811095654964447 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280717074871063 + ], + "opposite_logprob_tokens": [ + -2.0178072452545166 + ], + "logprob_reward": -0.14280717074871063, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14280717074871063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1452781707048416, + 0.060442544519901276, + 0.1802065223455429, + 0.07182770222425461, + 0.04426199570298195, + 0.1319865882396698, + 0.06408126652240753, + 0.03851678967475891, + 0.195078045129776, + 0.0683203935623169 + ], + "credit_map": [ + 0.022216705605387688, + 0.025220902636647224, + 0.015874145552515984, + 0.009271039627492428, + 0.018804501742124557, + 0.02062518522143364, + 0.020244883373379707, + 0.013020804151892662, + 0.006845629774034023, + 0.0072039044462144375, + 0.006198767572641373, + 0.006259467918425798, + 0.01011040061712265, + 0.009912077337503433, + 0.008541639894247055, + 0.005370656028389931, + 0.015285369008779526, + 0.024096082895994186, + 0.013955725356936455, + 0.02274409867823124, + 0.04271351173520088, + 0.01902761124074459, + 0.029242176562547684, + 0.013141961768269539, + 0.004372516181319952, + 0.00880441814661026, + 0.006115734577178955, + 0.008393913507461548, + 0.0126485051587224, + 0.010477878153324127, + 0.010330266319215298, + 0.01068447157740593, + 0.005925768055021763, + 0.008785114623606205, + 0.005668765399605036, + 0.0046509732492268085, + 0.005613826680928469, + 0.004626672249287367, + 0.00437410781159997, + 0.004616765771061182, + 0.013518407940864563, + 0.023784048855304718, + 0.026771754026412964, + 0.018591349944472313, + 0.015191995538771152, + 0.011810440570116043, + 0.012978103011846542, + 0.009340478107333183, + 0.01142430491745472, + 0.007607735227793455, + 0.008870811201632023, + 0.006260057911276817, + 0.014435793273150921, + 0.005688142497092485, + 0.004904454108327627, + 0.004889971110969782, + 0.007092300336807966, + 0.0029615729581564665, + 0.01043596863746643, + 0.0038403389044106007, + 0.0041616992093622684, + 0.0028689634054899216, + 0.004475349094718695, + 0.0026805975940078497, + 0.07364650815725327, + 0.009275930933654308, + 0.06159212812781334, + 0.011080792173743248, + 0.011335683055222034, + 0.005195985548198223, + 0.017486702650785446, + 0.005464315414428711, + 0.008453281596302986, + 0.008382683619856834, + 0.013595717959105968, + 0.010148327797651291, + 0.006668305955827236, + 0.007461050525307655, + 0.00871865451335907, + 0.004892371129244566 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.14280717074871063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370778262615204 + ], + "opposite_logprob_tokens": [ + -3.408707857131958 + ], + "logprob_reward": -0.03370778262615204, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370778262615204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14263613522052765, + 0.09125005453824997, + 0.07400455325841904, + 0.078868567943573, + 0.061132874339818954, + 0.20900876820087433, + 0.09842237085103989, + 0.0549345500767231, + 0.14297299087047577, + 0.046769123524427414 + ], + "credit_map": [ + 0.01745114102959633, + 0.02026045136153698, + 0.014877072535455227, + 0.011175570078194141, + 0.015494294464588165, + 0.018343467265367508, + 0.030590444803237915, + 0.014443689957261086, + 0.011676803231239319, + 0.012562684714794159, + 0.008292914368212223, + 0.00827874056994915, + 0.01929554156959057, + 0.01475759781897068, + 0.00973441544920206, + 0.00665135495364666, + 0.00716804526746273, + 0.008686797693371773, + 0.0062036975286901, + 0.00782522652298212, + 0.015187060460448265, + 0.008713371120393276, + 0.013216407969594002, + 0.007003949489444494, + 0.0075394208543002605, + 0.007887734100222588, + 0.007961063645780087, + 0.0112199317663908, + 0.010304047726094723, + 0.011027634143829346, + 0.015138736926019192, + 0.007789990399032831, + 0.00775066576898098, + 0.015499201603233814, + 0.005995404906570911, + 0.005266846623271704, + 0.0069830818101763725, + 0.007633083034306765, + 0.005754012614488602, + 0.006250573787838221, + 0.037627846002578735, + 0.04068392887711525, + 0.030992869287729263, + 0.013135253451764584, + 0.019680900499224663, + 0.017591916024684906, + 0.012125037610530853, + 0.037171024829149246, + 0.01735110580921173, + 0.00791498739272356, + 0.013956896029412746, + 0.005546618718653917, + 0.025168441236019135, + 0.009831354953348637, + 0.006927340291440487, + 0.011725617572665215, + 0.010161587037146091, + 0.003651775885373354, + 0.016736287623643875, + 0.0057246810756623745, + 0.0049195545725524426, + 0.003589906031265855, + 0.006615507882088423, + 0.003535252995789051, + 0.05501769855618477, + 0.006262072827666998, + 0.046755269169807434, + 0.007664647419005632, + 0.007708479184657335, + 0.004000284243375063, + 0.012412221170961857, + 0.003152313409373164, + 0.006199045572429895, + 0.003879773197695613, + 0.008190426975488663, + 0.0079912468791008, + 0.003941391594707966, + 0.006144138518720865, + 0.006021607201546431, + 0.004401494283229113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.03370778262615204 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17515432834625244, + 0.044252555817365646, + 0.06081439182162285, + 0.10398619621992111, + 0.10665738582611084, + 0.13175179064273834, + 0.049908727407455444, + 0.09196814894676208, + 0.173695906996727, + 0.06181056424975395 + ], + "credit_map": [ + 0.01667105220258236, + 0.02287178300321102, + 0.022788994014263153, + 0.02471182867884636, + 0.022629067301750183, + 0.01890782080590725, + 0.02871660329401493, + 0.01785719022154808, + 0.004015470389276743, + 0.006058001425117254, + 0.00522000202909112, + 0.005893846973776817, + 0.0057940310798585415, + 0.005553023423999548, + 0.007102790754288435, + 0.00461539626121521, + 0.004825101234018803, + 0.010278805159032345, + 0.004183770623058081, + 0.0059893387369811535, + 0.010063376277685165, + 0.007312885485589504, + 0.01096678338944912, + 0.00719434255734086, + 0.007539835758507252, + 0.01174101047217846, + 0.007106797304004431, + 0.016623852774500847, + 0.016347769647836685, + 0.017549565061926842, + 0.01848623715341091, + 0.008591130375862122, + 0.006764615885913372, + 0.014712370000779629, + 0.011182696558535099, + 0.016270646825432777, + 0.01863272860646248, + 0.011911657638847828, + 0.01856769435107708, + 0.008614988066256046, + 0.024939587339758873, + 0.018353812396526337, + 0.012250004336237907, + 0.017191505059599876, + 0.011567874811589718, + 0.012958202511072159, + 0.013717427849769592, + 0.020773380994796753, + 0.0051428512670099735, + 0.005847106222063303, + 0.00789165310561657, + 0.003201930783689022, + 0.011012844741344452, + 0.006776057183742523, + 0.0039529744535684586, + 0.006083312910050154, + 0.016259346157312393, + 0.0068636732175946236, + 0.027552897110581398, + 0.00942222960293293, + 0.008447474800050259, + 0.006686343811452389, + 0.01074746623635292, + 0.005988729186356068, + 0.0652923658490181, + 0.007365310098975897, + 0.050230856984853745, + 0.012376699596643448, + 0.008933266624808311, + 0.006648852489888668, + 0.015174840576946735, + 0.007673709187656641, + 0.007805260363966227, + 0.008091750554740429, + 0.01749417558312416, + 0.006010398268699646, + 0.00441774632781744, + 0.007014075759798288, + 0.006404344458132982, + 0.004572822712361813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.014193184673786163 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid, slowly rotating 360-degree camera, a dense vortex of countless flying sawdust and fine dust particles hovers in the center of a craftsman's workshop. Warm, golden afternoon sunlight slants through dusty windows, illuminating each suspended particle like a shimmering swarm of golden insects. A softly blurred backdrop features a wall covered in tools and piles of lumber. As the camera continues to pan, the initially chaotic, fluttering debris begins to be drawn by an invisible force, rapidly and orderly converging toward the center. They begin to weave and intertwine, larger pieces decisively forming the outlines of the chair's legs and backrest, while finer pieces of sawdust and dust seamlessly fill and compact the space, forming themselves as if possessing a life of their own. Finally, the rotation came to an abrupt halt, and the flying mass of matter had completely solidified, transforming into a wooden chair with smooth lines and solid texture. Its surface was smooth, and the warm oak grain was clearly revealed under the light. It stood quietly in place. The whole process was full of a focused and magical sense of creation, turning disorder into order and presenting a fascinating visual spectacle.", + "seed": 334, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_334_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.6139189941808582, + "vlm_logprob": -0.6139189941808582, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.6139189941808582, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -2.9309017658233643 + ], + "opposite_logprob_tokens": [ + -0.0559016615152359 + ], + "logprob_reward": -2.9309017658233643, + "yes_no_margin_reward": -2.875, + "token_credit_reward": -2.9309017658233643, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.12677313387393951, + 0.042540937662124634, + 0.03530441224575043, + 0.1142413541674614, + 0.04516732320189476, + 0.017748679965734482, + 0.010482437908649445, + 0.01180464681237936, + 0.5566702485084534, + 0.039266906678676605 + ], + "credit_map": [ + 0.00803566724061966, + 0.010686060413718224, + 0.009317235089838505, + 0.023423703387379646, + 0.012840759009122849, + 0.021291300654411316, + 0.014180521480739117, + 0.02699786238372326, + 0.005945988930761814, + 0.003717780578881502, + 0.005434362683445215, + 0.00706517556682229, + 0.004253765568137169, + 0.00581336347386241, + 0.004765075631439686, + 0.005545423831790686, + 0.0029327610973268747, + 0.00404714560136199, + 0.004302630666643381, + 0.0035569712053984404, + 0.0036090081557631493, + 0.006473328918218613, + 0.0025733066722750664, + 0.007809256669133902, + 0.009598765522241592, + 0.013839353807270527, + 0.0099132489413023, + 0.005490109324455261, + 0.013020824640989304, + 0.033409491181373596, + 0.0032678567804396152, + 0.02570168673992157, + 0.005419633816927671, + 0.0035904948599636555, + 0.009747235104441643, + 0.005503491032868624, + 0.004269604571163654, + 0.0028317372780293226, + 0.004510559141635895, + 0.009294562041759491, + 0.002687149215489626, + 0.002398333279415965, + 0.0031405859626829624, + 0.0019331102957949042, + 0.0019589639268815517, + 0.002180516952648759, + 0.0010932611767202616, + 0.0023567560128867626, + 0.0025372153613716364, + 0.0011245881905779243, + 0.001270674867555499, + 0.0006483044708147645, + 0.0007331712404266, + 0.0006849644123576581, + 0.0006631019059568644, + 0.002820416586473584, + 0.0018838627729564905, + 0.0017764684744179249, + 0.0016717021353542805, + 0.0015944303013384342, + 0.0011456774082034826, + 0.0013584234984591603, + 0.0010102817323058844, + 0.0013637999072670937, + 0.02507844939827919, + 0.02828933298587799, + 0.03510698303580284, + 0.022059621289372444, + 0.025156350806355476, + 0.20132997632026672, + 0.02071341685950756, + 0.1989360898733139, + 0.004985298030078411, + 0.0072380234487354755, + 0.0033273883163928986, + 0.0034771047066897154, + 0.004936452955007553, + 0.0060204570181667805, + 0.004644346889108419, + 0.00463782949373126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -2.9309017658233643 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05505584180355072 + ], + "opposite_logprob_tokens": [ + -2.930055856704712 + ], + "logprob_reward": -0.05505584180355072, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05505584180355072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1404653787612915, + 0.019215654581785202, + 0.04298168793320656, + 0.3723118007183075, + 0.011273505166172981, + 0.01740122027695179, + 0.01214069128036499, + 0.023000100627541542, + 0.3342571258544922, + 0.026952875778079033 + ], + "credit_map": [ + 0.010122798383235931, + 0.012781591154634953, + 0.022113213315606117, + 0.024510927498340607, + 0.014687308110296726, + 0.017712002620100975, + 0.014641450718045235, + 0.023896094411611557, + 0.003268146188929677, + 0.0016939580673351884, + 0.0031778523698449135, + 0.0022568462882190943, + 0.00229668989777565, + 0.00238470989279449, + 0.0016360045410692692, + 0.00250144861638546, + 0.0025035955477505922, + 0.0029078545048832893, + 0.003918414004147053, + 0.002950008725747466, + 0.0031847271602600813, + 0.014108952134847641, + 0.0021157334558665752, + 0.01129240170121193, + 0.0327657125890255, + 0.04013613238930702, + 0.0325079970061779, + 0.016977349296212196, + 0.03933977335691452, + 0.11533881723880768, + 0.010118940845131874, + 0.0851270779967308, + 0.00105065549723804, + 0.0018884685123339295, + 0.001613035798072815, + 0.001115031074732542, + 0.0017912989715114236, + 0.0012591916602104902, + 0.0009202264482155442, + 0.0016355975531041622, + 0.002883029170334339, + 0.0023584922309964895, + 0.0034523787908256054, + 0.0023014280013740063, + 0.0015184762887656689, + 0.0014111010823398829, + 0.0012866697506979108, + 0.002189644845202565, + 0.002786168595775962, + 0.0016544164391234517, + 0.001976832514628768, + 0.0008363475790247321, + 0.0008838765206746757, + 0.0009242641972377896, + 0.0007849163375794888, + 0.0022938696201890707, + 0.002863689325749874, + 0.006381096784025431, + 0.003102813381701708, + 0.0019221034599468112, + 0.001956036314368248, + 0.002838629763573408, + 0.0013903520302847028, + 0.0025453800335526466, + 0.01233628485351801, + 0.01522745005786419, + 0.01570124179124832, + 0.010228368453681469, + 0.012584621086716652, + 0.13091018795967102, + 0.009859232231974602, + 0.12740972638130188, + 0.004018161911517382, + 0.005057977046817541, + 0.0034572521690279245, + 0.0022207440342754126, + 0.0028362355660647154, + 0.00446281349286437, + 0.0019515500171110034, + 0.0029481425881385803 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.05505584180355072 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05501105263829231 + ], + "opposite_logprob_tokens": [ + -2.930011034011841 + ], + "logprob_reward": -0.05501105263829231, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05501105263829231, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03683578968048096, + 0.01601596735417843, + 0.05675609037280083, + 0.44515928626060486, + 0.011917666532099247, + 0.013232531026005745, + 0.023589299991726875, + 0.019276857376098633, + 0.35992419719696045, + 0.01729237101972103 + ], + "credit_map": [ + 0.002087369794026017, + 0.003635169705376029, + 0.0029760354664176702, + 0.0033557226415723562, + 0.003691050922498107, + 0.006634061224758625, + 0.0029185230378061533, + 0.011537853628396988, + 0.002700376557186246, + 0.0012179662007838488, + 0.002199245383962989, + 0.002108711050823331, + 0.0016837941948324442, + 0.002407860942184925, + 0.0013105093967169523, + 0.002387501997873187, + 0.002079552970826626, + 0.00299374433234334, + 0.0049290708266198635, + 0.003939082380384207, + 0.003241707105189562, + 0.02024243026971817, + 0.0019175964407622814, + 0.017412900924682617, + 0.03982028737664223, + 0.043193913996219635, + 0.03982842341065407, + 0.020453711971640587, + 0.0418068952858448, + 0.14353710412979126, + 0.011841650120913982, + 0.10467725247144699, + 0.0012599820038303733, + 0.0011046442668884993, + 0.002267614472657442, + 0.0017086623702198267, + 0.0012804048601537943, + 0.0008423984982073307, + 0.001367237651720643, + 0.002086720196530223, + 0.0013528127456083894, + 0.0019596735946834087, + 0.002018081257119775, + 0.0013579942751675844, + 0.0018768898444250226, + 0.0018079852452501655, + 0.000920455961022526, + 0.0019386374624446034, + 0.005995508749037981, + 0.0029899384826421738, + 0.004041982349008322, + 0.0027285171672701836, + 0.0012429915368556976, + 0.001419849693775177, + 0.0023912438191473484, + 0.002779264934360981, + 0.002495590830221772, + 0.002922887448221445, + 0.002390258014202118, + 0.002481241011992097, + 0.0022061532363295555, + 0.003092332510277629, + 0.0014885285636410117, + 0.002199862850829959, + 0.012036526575684547, + 0.015281853266060352, + 0.01333286426961422, + 0.008904444985091686, + 0.013207321986556053, + 0.1466268002986908, + 0.009637757204473019, + 0.14089660346508026, + 0.0019537569023668766, + 0.004334019497036934, + 0.0020499159581959248, + 0.001203315332531929, + 0.0014568144688382745, + 0.003234655363485217, + 0.001187867484986782, + 0.0018720256630331278 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.05501105263829231 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255407091230154 + ], + "opposite_logprob_tokens": [ + -4.387554168701172 + ], + "logprob_reward": -0.01255407091230154, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255407091230154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10023702681064606, + 0.036283910274505615, + 0.08983376622200012, + 0.2183990478515625, + 0.024720458313822746, + 0.007812450639903545, + 0.010959181003272533, + 0.0241406187415123, + 0.4585362374782562, + 0.02907729521393776 + ], + "credit_map": [ + 0.007108291145414114, + 0.008242575451731682, + 0.011480295099318027, + 0.01714678481221199, + 0.00967249646782875, + 0.01629233919084072, + 0.008961603976786137, + 0.02133263275027275, + 0.006038243882358074, + 0.0033775405026972294, + 0.004363925661891699, + 0.004711595829576254, + 0.003275061957538128, + 0.005671737249940634, + 0.003005575854331255, + 0.005840227473527193, + 0.004515682812780142, + 0.003652637591585517, + 0.008016693405807018, + 0.0045287455432116985, + 0.004027490504086018, + 0.03383916616439819, + 0.0028896403964608908, + 0.028363710269331932, + 0.01736464910209179, + 0.02182207629084587, + 0.021970972418785095, + 0.010522125288844109, + 0.023064345121383667, + 0.06343439221382141, + 0.00668753869831562, + 0.05353294685482979, + 0.0035497392527759075, + 0.002076677978038788, + 0.004915397148579359, + 0.00295541575178504, + 0.0025126985274255276, + 0.0015549535164609551, + 0.0024645121302455664, + 0.0046910629607737064, + 0.0006254444597288966, + 0.0010940711945295334, + 0.0012519452720880508, + 0.001050401828251779, + 0.0009334647329524159, + 0.0010919378837570548, + 0.0008741061901673675, + 0.0008910788455978036, + 0.002142511308193207, + 0.0013616064097732306, + 0.001975810155272484, + 0.0010831558611243963, + 0.0007678671972826123, + 0.0007689876947551966, + 0.0009928593644872308, + 0.0018663824303075671, + 0.004020615480840206, + 0.0046323551796376705, + 0.002884030109271407, + 0.0024201262276619673, + 0.0031303141731768847, + 0.003249269677326083, + 0.001908326055854559, + 0.0018955820705741644, + 0.017769737169146538, + 0.02093488536775112, + 0.016228314489126205, + 0.011931897141039371, + 0.01663973368704319, + 0.184805229306221, + 0.01234959252178669, + 0.17787685990333557, + 0.0036636637523770332, + 0.003033616580069065, + 0.003836793825030327, + 0.0036700989585369825, + 0.00493140984326601, + 0.004499692469835281, + 0.0023781072814017534, + 0.0030639139004051685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 0 + ], + "selected_reward": -0.01255407091230154 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016072239726781845 + ], + "opposite_logprob_tokens": [ + -4.1410722732543945 + ], + "logprob_reward": -0.016072239726781845, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016072239726781845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10864289849996567, + 0.04964609816670418, + 0.12718801200389862, + 0.18433339893817902, + 0.04163825511932373, + 0.01895901933312416, + 0.012429575435817242, + 0.02168194390833378, + 0.3893009424209595, + 0.046179864555597305 + ], + "credit_map": [ + 0.007468739058822393, + 0.011264261789619923, + 0.010069405660033226, + 0.0165544506162405, + 0.01110096275806427, + 0.018664216622710228, + 0.010051197372376919, + 0.02346966229379177, + 0.008622335270047188, + 0.004351666662842035, + 0.0063463798724114895, + 0.007048659957945347, + 0.003650836180895567, + 0.00756996450945735, + 0.004231288097798824, + 0.007824965752661228, + 0.006744971498847008, + 0.004804250784218311, + 0.011388027109205723, + 0.00588528998196125, + 0.005709268152713776, + 0.04836316034197807, + 0.003899256233125925, + 0.04039379581809044, + 0.014818633906543255, + 0.017573876306414604, + 0.018736526370048523, + 0.008629878982901573, + 0.019549960270524025, + 0.05397053435444832, + 0.005257192999124527, + 0.04579678922891617, + 0.004767071921378374, + 0.0035866396501660347, + 0.008936035446822643, + 0.004538761451840401, + 0.004425800871104002, + 0.002615166362375021, + 0.0035909698344767094, + 0.00917781051248312, + 0.002481994451954961, + 0.002352033043280244, + 0.00326425745151937, + 0.0019010441610589623, + 0.0024599223397672176, + 0.002705312566831708, + 0.0015436060493811965, + 0.0022508488036692142, + 0.002415788359940052, + 0.0013792648678645492, + 0.00126367446500808, + 0.001234489376656711, + 0.00117943761870265, + 0.0013950992142781615, + 0.0010801530443131924, + 0.0024816698860377073, + 0.003971001133322716, + 0.005045388359576464, + 0.003950315061956644, + 0.001885302015580237, + 0.0021764268167316914, + 0.0018734054174274206, + 0.0010273734806105494, + 0.0017527302261441946, + 0.018146788701415062, + 0.018828731030225754, + 0.016857203096151352, + 0.010890974663197994, + 0.013984699733555317, + 0.15198151767253876, + 0.01098716352134943, + 0.14762386679649353, + 0.0039389608427882195, + 0.005029703956097364, + 0.004870819393545389, + 0.005630161613225937, + 0.008212508633732796, + 0.00728793628513813, + 0.005401124712079763, + 0.005808650515973568 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 2 + ], + "selected_reward": -0.016072239726781845 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid, slowly rotating 360-degree camera, a dense vortex of countless flying sawdust and fine dust particles hovers in the center of a craftsman's workshop. Warm, golden afternoon sunlight slants through dusty windows, illuminating each suspended particle like a shimmering swarm of golden insects. A softly blurred backdrop features a wall covered in tools and piles of lumber. As the camera continues to pan, the initially chaotic, fluttering debris begins to be drawn by an invisible force, rapidly and orderly converging toward the center. They begin to weave and intertwine, larger pieces decisively forming the outlines of the chair's legs and backrest, while finer pieces of sawdust and dust seamlessly fill and compact the space, forming themselves as if possessing a life of their own. Finally, the rotation came to an abrupt halt, and the flying mass of matter had completely solidified, transforming into a wooden chair with smooth lines and solid texture. Its surface was smooth, and the warm oak grain was clearly revealed under the light. It stood quietly in place. The whole process was full of a focused and magical sense of creation, turning disorder into order and presenting a fascinating visual spectacle.", + "seed": 335, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_335_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02883811332285404, + "vlm_logprob": -0.02883811332285404, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.02883811332285404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053836606442928314 + ], + "opposite_logprob_tokens": [ + -3.0538365840911865 + ], + "logprob_reward": -0.053836606442928314, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053836606442928314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13141611218452454, + 0.04210531711578369, + 0.06475363671779633, + 0.18972891569137573, + 0.2144060879945755, + 0.03237062320113182, + 0.1053231805562973, + 0.11827438324689865, + 0.0484333373606205, + 0.05318842828273773 + ], + "credit_map": [ + 0.014463746920228004, + 0.01509361993521452, + 0.023776747286319733, + 0.016733311116695404, + 0.021706270053982735, + 0.01001039519906044, + 0.019319437444210052, + 0.010312593542039394, + 0.006482277996838093, + 0.0052908374927937984, + 0.0054488470777869225, + 0.005353832617402077, + 0.005361291114240885, + 0.0039484635926783085, + 0.004966711159795523, + 0.0052530583925545216, + 0.003563342848792672, + 0.009862761944532394, + 0.015667032450437546, + 0.00959805864840746, + 0.009911665692925453, + 0.003944805357605219, + 0.005227694753557444, + 0.006978271529078484, + 0.016334818676114082, + 0.01325145922601223, + 0.01015180442482233, + 0.018437892198562622, + 0.03563975542783737, + 0.015576310455799103, + 0.05666811019182205, + 0.023668762296438217, + 0.012605748139321804, + 0.02510868012905121, + 0.032746486365795135, + 0.0320269949734211, + 0.015250950120389462, + 0.013788992539048195, + 0.03958667069673538, + 0.04329155758023262, + 0.007965875789523125, + 0.003669202560558915, + 0.004375568591058254, + 0.002206319710239768, + 0.002627447945997119, + 0.0028712351340800524, + 0.004018183331936598, + 0.004636787809431553, + 0.021984072402119637, + 0.009342966601252556, + 0.017850929871201515, + 0.01043760497123003, + 0.01013142243027687, + 0.011114129796624184, + 0.010848750360310078, + 0.013613307848572731, + 0.014695640653371811, + 0.010975501500070095, + 0.01761830411851406, + 0.017814116552472115, + 0.009910021908581257, + 0.011282378807663918, + 0.014692054130136967, + 0.021286366507411003, + 0.005830386653542519, + 0.00481076817959547, + 0.00703859981149435, + 0.0053889513947069645, + 0.003849941538646817, + 0.005018624011427164, + 0.009580809623003006, + 0.00691525824368, + 0.0070806541480124, + 0.011789520271122456, + 0.004851804114878178, + 0.004881399217993021, + 0.009920857846736908, + 0.004948242101818323, + 0.004163225181400776, + 0.00555272214114666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.053836606442928314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865293201059103 + ], + "opposite_logprob_tokens": [ + -4.758653163909912 + ], + "logprob_reward": -0.00865293201059103, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865293201059103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4134482145309448, + 0.0402439720928669, + 0.1874106228351593, + 0.10466541349887848, + 0.08153243362903595, + 0.01842261664569378, + 0.054620709270238876, + 0.034259747713804245, + 0.03586350753903389, + 0.029532665386795998 + ], + "credit_map": [ + 0.04410935938358307, + 0.037684351205825806, + 0.07739635556936264, + 0.04597001150250435, + 0.06598391383886337, + 0.025288719683885574, + 0.08098886907100677, + 0.03602668642997742, + 0.006416084710508585, + 0.004536542110145092, + 0.005798913072794676, + 0.004843894857913256, + 0.004319393076002598, + 0.003938902635127306, + 0.005025454796850681, + 0.005364789627492428, + 0.007160662207752466, + 0.07570195198059082, + 0.005147213581949472, + 0.005419021937996149, + 0.07508478313684464, + 0.007997664622962475, + 0.005432621575891972, + 0.0054667177610099316, + 0.009186924435198307, + 0.006919921841472387, + 0.0061340248212218285, + 0.011720234528183937, + 0.018792161718010902, + 0.008602441288530827, + 0.030527645722031593, + 0.012782074511051178, + 0.004663224332034588, + 0.011512957513332367, + 0.00816702377051115, + 0.007081835065037012, + 0.005675168242305517, + 0.004229078534990549, + 0.019655665382742882, + 0.02054748311638832, + 0.0029426959808915854, + 0.0018632012652233243, + 0.002559634391218424, + 0.0022762990556657314, + 0.0018632563296705484, + 0.0016472432762384415, + 0.00293764378875494, + 0.0023326436057686806, + 0.012134292162954807, + 0.0052448054775595665, + 0.007367595098912716, + 0.005508316680788994, + 0.006072306074202061, + 0.005446468945592642, + 0.005809236317873001, + 0.0070376917719841, + 0.004095351789146662, + 0.004366347100585699, + 0.004171049688011408, + 0.004449580330401659, + 0.002480276394635439, + 0.004105879459530115, + 0.006611411459743977, + 0.003979855682700872, + 0.004122810438275337, + 0.0037131835706532, + 0.004714077804237604, + 0.003265589475631714, + 0.004380178637802601, + 0.005222001112997532, + 0.005799353122711182, + 0.004646316170692444, + 0.0035441555082798004, + 0.0033724349923431873, + 0.0034556679893285036, + 0.002878862898796797, + 0.005379454232752323, + 0.002674208488315344, + 0.0038338033482432365, + 0.004394080955535173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 3 + ], + "selected_reward": -0.00865293201059103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054969191551208496 + ], + "opposite_logprob_tokens": [ + -2.929969310760498 + ], + "logprob_reward": -0.054969191551208496, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054969191551208496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06346552819013596, + 0.04320106655359268, + 0.09976493567228317, + 0.30837342143058777, + 0.1492980718612671, + 0.032349564135074615, + 0.04332122579216957, + 0.1391443908214569, + 0.04522201418876648, + 0.07585973292589188 + ], + "credit_map": [ + 0.00687498738989234, + 0.007035669405013323, + 0.014359365217387676, + 0.00883212685585022, + 0.007498238701373339, + 0.004594157915562391, + 0.007636939641088247, + 0.006634053308516741, + 0.010565461590886116, + 0.00844753161072731, + 0.0034789922647178173, + 0.0030606070067733526, + 0.00541666429489851, + 0.0035299344453960657, + 0.0048417565412819386, + 0.003860124386847019, + 0.00470619136467576, + 0.03593821823596954, + 0.0061464980244636536, + 0.005403257440775633, + 0.034663714468479156, + 0.0038391537964344025, + 0.0050180405378341675, + 0.0040498836897313595, + 0.02729620598256588, + 0.020962141454219818, + 0.015511504374444485, + 0.03224090486764908, + 0.056097738444805145, + 0.025001339614391327, + 0.09354951232671738, + 0.037714097648859024, + 0.00412141066044569, + 0.006031224969774485, + 0.012699113227427006, + 0.012710601091384888, + 0.005509835202246904, + 0.006098022684454918, + 0.0504629947245121, + 0.051664870232343674, + 0.0024634874425828457, + 0.0024128106888383627, + 0.006991222500801086, + 0.003779670223593712, + 0.0027520807925611734, + 0.0037548549007624388, + 0.005108790937811136, + 0.005086652003228664, + 0.008235466666519642, + 0.004701187834143639, + 0.005909749306738377, + 0.003957544453442097, + 0.004753021989017725, + 0.005663731601089239, + 0.0051564862951636314, + 0.0049440436996519566, + 0.016684185713529587, + 0.013729169964790344, + 0.020830310881137848, + 0.02518760785460472, + 0.010530331172049046, + 0.01240857969969511, + 0.018286576494574547, + 0.0214876439422369, + 0.004698847886174917, + 0.004025687463581562, + 0.005214388947933912, + 0.003725397167727351, + 0.0061627766117453575, + 0.006776143796741962, + 0.006738217547535896, + 0.007880565710365772, + 0.008666491135954857, + 0.010037917643785477, + 0.009420141577720642, + 0.006858973763883114, + 0.014066058211028576, + 0.007913081906735897, + 0.008957559242844582, + 0.0099395252764225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.054969191551208496 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540650554001331 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540650554001331, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540650554001331, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23704811930656433, + 0.048194240778684616, + 0.19077938795089722, + 0.0854656770825386, + 0.09880710393190384, + 0.04172809422016144, + 0.0895182341337204, + 0.08842681348323822, + 0.05869061499834061, + 0.06134168803691864 + ], + "credit_map": [ + 0.029216082766652107, + 0.023170676082372665, + 0.05252072960138321, + 0.022631976753473282, + 0.038321927189826965, + 0.015352549031376839, + 0.03754464536905289, + 0.01828952692449093, + 0.0076097664423286915, + 0.004670224618166685, + 0.005692712496966124, + 0.004525060765445232, + 0.0047735427506268024, + 0.005859640426933765, + 0.006963490508496761, + 0.008099802769720554, + 0.013740768656134605, + 0.029345545917749405, + 0.04556179791688919, + 0.02707716077566147, + 0.030267992988228798, + 0.008734998293220997, + 0.015724021941423416, + 0.020327117294073105, + 0.009485009126365185, + 0.007152537815272808, + 0.007204109337180853, + 0.009790271520614624, + 0.010226601734757423, + 0.005522152874618769, + 0.022606195881962776, + 0.01347880158573389, + 0.00466776080429554, + 0.008920833468437195, + 0.006501699797809124, + 0.008752123452723026, + 0.005264461971819401, + 0.0062355585396289825, + 0.02804877795279026, + 0.030415888875722885, + 0.004361392930150032, + 0.00376324076205492, + 0.009616173803806305, + 0.004091281443834305, + 0.003717568702995777, + 0.003893211018294096, + 0.005757532548159361, + 0.006527692545205355, + 0.017920957878232002, + 0.009943093173205853, + 0.010230801068246365, + 0.007151367608457804, + 0.01110761146992445, + 0.012049883604049683, + 0.009669408202171326, + 0.011445112526416779, + 0.0095747672021389, + 0.010891689918935299, + 0.009807297959923744, + 0.013679005205631256, + 0.006355141289532185, + 0.00991478655487299, + 0.017375636845827103, + 0.010828486643731594, + 0.0070180767215788364, + 0.006282510701566935, + 0.007897857576608658, + 0.005750723648816347, + 0.005563534330576658, + 0.0086284214630723, + 0.010679586790502071, + 0.006869901437312365, + 0.006540870293974876, + 0.007153861224651337, + 0.007788850460201502, + 0.005858529359102249, + 0.011782578192651272, + 0.005557718686759472, + 0.007727182470262051, + 0.008932098746299744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.012540650554001331 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191186055541039 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014191186055541039, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191186055541039, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1160571426153183, + 0.06171770393848419, + 0.21314974129199982, + 0.0484592467546463, + 0.11113154888153076, + 0.06214096024632454, + 0.04201560094952583, + 0.0735090896487236, + 0.10881749540567398, + 0.1630014181137085 + ], + "credit_map": [ + 0.01153619959950447, + 0.015237539075314999, + 0.014026476070284843, + 0.020617106929421425, + 0.015310251154005527, + 0.011921732686460018, + 0.017344322055578232, + 0.01006352063268423, + 0.011307396925985813, + 0.011172622442245483, + 0.007017229218035936, + 0.00544830784201622, + 0.00580930057913065, + 0.0067145926877856255, + 0.00656247278675437, + 0.007685786113142967, + 0.01409649197012186, + 0.048551227897405624, + 0.04081602767109871, + 0.023451661691069603, + 0.0463327057659626, + 0.008364810608327389, + 0.013732749968767166, + 0.01780407316982746, + 0.005081063136458397, + 0.004103680141270161, + 0.004232985898852348, + 0.006454956252127886, + 0.007933755405247211, + 0.004484259523451328, + 0.010291658341884613, + 0.005876895040273666, + 0.008161831647157669, + 0.018813099712133408, + 0.012023615650832653, + 0.008900925517082214, + 0.008624542504549026, + 0.0061393254436552525, + 0.02214011736214161, + 0.02632809244096279, + 0.007329263258725405, + 0.0043286108411848545, + 0.015802646055817604, + 0.007836719043552876, + 0.0052437568083405495, + 0.003925568424165249, + 0.00748774129897356, + 0.01018666010349989, + 0.006679506506770849, + 0.004083015490323305, + 0.007704582996666431, + 0.0034769836347550154, + 0.005221682135015726, + 0.00481750164180994, + 0.004413813352584839, + 0.005618518218398094, + 0.007519521284848452, + 0.008544893935322762, + 0.012216970324516296, + 0.009633822366595268, + 0.005033331457525492, + 0.007389679551124573, + 0.013837862759828568, + 0.009333011694252491, + 0.008729320019483566, + 0.012948191724717617, + 0.01504515577107668, + 0.0136029664427042, + 0.008919536136090755, + 0.011123470030725002, + 0.022503217682242393, + 0.01594565063714981, + 0.022160708904266357, + 0.03929489478468895, + 0.019765064120292664, + 0.014165770262479782, + 0.02111160010099411, + 0.01787443645298481, + 0.014300747774541378, + 0.014328207820653915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.014191186055541039 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a fluid, slowly rotating 360-degree camera, a dense vortex of countless flying sawdust and fine dust particles hovers in the center of a craftsman's workshop. Warm, golden afternoon sunlight slants through dusty windows, illuminating each suspended particle like a shimmering swarm of golden insects. A softly blurred backdrop features a wall covered in tools and piles of lumber. As the camera continues to pan, the initially chaotic, fluttering debris begins to be drawn by an invisible force, rapidly and orderly converging toward the center. They begin to weave and intertwine, larger pieces decisively forming the outlines of the chair's legs and backrest, while finer pieces of sawdust and dust seamlessly fill and compact the space, forming themselves as if possessing a life of their own. Finally, the rotation came to an abrupt halt, and the flying mass of matter had completely solidified, transforming into a wooden chair with smooth lines and solid texture. Its surface was smooth, and the warm oak grain was clearly revealed under the light. It stood quietly in place. The whole process was full of a focused and magical sense of creation, turning disorder into order and presenting a fascinating visual spectacle.", + "seed": 336, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_1/index_8100_seed_336_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.026518897339701652, + "vlm_logprob": -0.026518897339701652, + "vlm_yes_no_margin": 4.1, + "vlm_token_credit": -0.026518897339701652, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03563039377331734 + ], + "opposite_logprob_tokens": [ + -3.410630464553833 + ], + "logprob_reward": -0.03563039377331734, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03563039377331734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2821410596370697, + 0.04321465641260147, + 0.05364532023668289, + 0.07804892957210541, + 0.16689132153987885, + 0.16921456158161163, + 0.05138801038265228, + 0.0922926664352417, + 0.03547082468867302, + 0.027692701667547226 + ], + "credit_map": [ + 0.02229706011712551, + 0.07407458871603012, + 0.03644450753927231, + 0.013383174315094948, + 0.03355550765991211, + 0.01566355489194393, + 0.016688449308276176, + 0.07003417611122131, + 0.006232385989278555, + 0.0041815899312496185, + 0.006166121922433376, + 0.003419853514060378, + 0.004548805765807629, + 0.0026818474289029837, + 0.009344834834337234, + 0.006639217492192984, + 0.007315341848880053, + 0.010702828876674175, + 0.005725332535803318, + 0.006452137138694525, + 0.006215536966919899, + 0.0035103510599583387, + 0.008083722554147243, + 0.005640064831823111, + 0.012948636896908283, + 0.007491292431950569, + 0.012876957654953003, + 0.008781716227531433, + 0.009922160767018795, + 0.005689459852874279, + 0.01403024047613144, + 0.006308459211140871, + 0.021219639107584953, + 0.014415194280445576, + 0.01794549636542797, + 0.024771688506007195, + 0.014198698103427887, + 0.03525036573410034, + 0.015811162069439888, + 0.023279063403606415, + 0.017411572858691216, + 0.012830087915062904, + 0.020000601187348366, + 0.020072091370821, + 0.012593019753694534, + 0.04067516326904297, + 0.01467352919280529, + 0.03095848858356476, + 0.002496316097676754, + 0.002223578281700611, + 0.01893380470573902, + 0.0028128058183938265, + 0.0018751632887870073, + 0.016459614038467407, + 0.0023801056668162346, + 0.00420661736279726, + 0.006206844002008438, + 0.007147778291255236, + 0.016499260440468788, + 0.01904149539768696, + 0.00710294721648097, + 0.013319198042154312, + 0.0055243391543626785, + 0.017450803890824318, + 0.004118339158594608, + 0.003190430114045739, + 0.006400795187801123, + 0.005118954461067915, + 0.005026740487664938, + 0.005892687942832708, + 0.0025824541226029396, + 0.003140420885756612, + 0.0019485410302877426, + 0.0021067450288683176, + 0.005077938083559275, + 0.0022119360510259867, + 0.006362117361277342, + 0.004557447507977486, + 0.002662347164005041, + 0.0027656296733766794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.03563039377331734 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206935651600361 + ], + "opposite_logprob_tokens": [ + -4.264206886291504 + ], + "logprob_reward": -0.014206935651600361, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206935651600361, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.50222247838974, + 0.07569331675767899, + 0.08532898873090744, + 0.019330520182847977, + 0.04020579159259796, + 0.05531035363674164, + 0.10310368239879608, + 0.0684296116232872, + 0.038866277784109116, + 0.011508983559906483 + ], + "credit_map": [ + 0.04361420497298241, + 0.11951106786727905, + 0.06985331326723099, + 0.018103644251823425, + 0.05569024384021759, + 0.02907506749033928, + 0.023385969921946526, + 0.14298897981643677, + 0.009174760431051254, + 0.007038956508040428, + 0.01712513342499733, + 0.006293389946222305, + 0.00582492770627141, + 0.004311066120862961, + 0.015567725524306297, + 0.01035736408084631, + 0.010649246163666248, + 0.015931865200400352, + 0.013014176860451698, + 0.010840145871043205, + 0.0089906370267272, + 0.005794495809823275, + 0.012758485972881317, + 0.007349938619881868, + 0.0031891821417957544, + 0.0020011819433420897, + 0.0029194108210504055, + 0.0023341479245573282, + 0.0026539722457528114, + 0.0016233236528933048, + 0.002562287962064147, + 0.0020470120944082737, + 0.005405241157859564, + 0.004097659140825272, + 0.0049124425277113914, + 0.007388831581920385, + 0.0032902301754802465, + 0.006695772521197796, + 0.003354700980708003, + 0.005060913506895304, + 0.006911139003932476, + 0.0025080759078264236, + 0.007188369054347277, + 0.006342200096696615, + 0.003332087304443121, + 0.0142610939219594, + 0.002248988253995776, + 0.012518403120338917, + 0.00636639166623354, + 0.0047675068490207195, + 0.034913040697574615, + 0.006813686341047287, + 0.005292633082717657, + 0.03029797412455082, + 0.007559194229543209, + 0.007093258202075958, + 0.0038329416420310736, + 0.006922543048858643, + 0.00996672734618187, + 0.010831352323293686, + 0.011478451080620289, + 0.009149863384664059, + 0.0038881967775523663, + 0.012359540909528732, + 0.004677929915487766, + 0.00424091937020421, + 0.0054210214875638485, + 0.0059886896051466465, + 0.0037795663811266422, + 0.0077052335254848, + 0.0032767984084784985, + 0.0037761193234473467, + 0.0015567532973363996, + 0.0007376792491413653, + 0.001673587248660624, + 0.00120662827976048, + 0.001741972635500133, + 0.002225315198302269, + 0.0012293412582948804, + 0.0011377071496099234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.014206935651600361 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002504169940948 + ], + "opposite_logprob_tokens": [ + -2.6950249671936035 + ], + "logprob_reward": -0.07002504169940948, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002504169940948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5952556133270264, + 0.025680791586637497, + 0.028688108548521996, + 0.019329383969306946, + 0.029780518263578415, + 0.07375596463680267, + 0.05094589665532112, + 0.044731590896844864, + 0.09203697741031647, + 0.03979514166712761 + ], + "credit_map": [ + 0.05366741865873337, + 0.13186168670654297, + 0.08552688360214233, + 0.020428989082574844, + 0.061074819415807724, + 0.0329383984208107, + 0.027900198474526405, + 0.181857168674469, + 0.0026051753666251898, + 0.0024070311337709427, + 0.003196347039192915, + 0.002121981466189027, + 0.003216368844732642, + 0.002230012323707342, + 0.006629468873143196, + 0.003274408634752035, + 0.0026868160348385572, + 0.004065125249326229, + 0.00396224670112133, + 0.0037781011778861284, + 0.003353157313540578, + 0.002282307483255863, + 0.004202366340905428, + 0.004357986617833376, + 0.0026509580202400684, + 0.0019086115062236786, + 0.002992701716721058, + 0.0021490147337317467, + 0.0020392844453454018, + 0.001726216054521501, + 0.0028045037761330605, + 0.0030580938328057528, + 0.004092795308679342, + 0.003068413119763136, + 0.0034390490036457777, + 0.0043476782739162445, + 0.002841222332790494, + 0.005758260376751423, + 0.002626639325171709, + 0.003606459591537714, + 0.008482453413307667, + 0.006421188358217478, + 0.012025771662592888, + 0.00973712932318449, + 0.007394271902740002, + 0.008981727063655853, + 0.00786796398460865, + 0.012845459394156933, + 0.002813610713928938, + 0.0021022947039455175, + 0.01946081407368183, + 0.0031755559612065554, + 0.002250277902930975, + 0.015764761716127396, + 0.0027585001662373543, + 0.0026200790889561176, + 0.0035567344166338444, + 0.0041162967681884766, + 0.005620230920612812, + 0.006446590647101402, + 0.007776379585266113, + 0.006005349103361368, + 0.0032754121348261833, + 0.007934597320854664, + 0.010464778169989586, + 0.009154016152024269, + 0.01424552034586668, + 0.01299890223890543, + 0.009310060180723667, + 0.021818500012159348, + 0.007192580029368401, + 0.0068526179529726505, + 0.004407777916640043, + 0.0033850150648504496, + 0.007154817692935467, + 0.003735022386536002, + 0.005726971197873354, + 0.006285790354013443, + 0.0038158996030688286, + 0.0052838451229035854 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.07002504169940948 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008638276718556881 + ], + "opposite_logprob_tokens": [ + -4.758638381958008 + ], + "logprob_reward": -0.008638276718556881, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008638276718556881, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41288650035858154, + 0.0678662657737732, + 0.03556782752275467, + 0.03131187707185745, + 0.05220143496990204, + 0.054171763360500336, + 0.20831835269927979, + 0.058185648173093796, + 0.042240653187036514, + 0.03724965453147888 + ], + "credit_map": [ + 0.03735022991895676, + 0.10176029801368713, + 0.052891846746206284, + 0.01565762795507908, + 0.04705338552594185, + 0.02488470822572708, + 0.023103412240743637, + 0.11018496751785278, + 0.008637393824756145, + 0.004980162251740694, + 0.009181514382362366, + 0.005283499602228403, + 0.007675331085920334, + 0.0050066374242305756, + 0.017818162217736244, + 0.009283563122153282, + 0.0045761726796627045, + 0.004837401676923037, + 0.004709369968622923, + 0.0051171425729990005, + 0.003460078965872526, + 0.0032190962228924036, + 0.006046846974641085, + 0.0036017168313264847, + 0.004764681216329336, + 0.0032170861959457397, + 0.005548608023673296, + 0.005206756293773651, + 0.00281617627479136, + 0.002429212210699916, + 0.004578172229230404, + 0.0027511841617524624, + 0.0070809475146234035, + 0.005111453123390675, + 0.007328931242227554, + 0.011766507290303707, + 0.004513823427259922, + 0.005889677442610264, + 0.00444940896704793, + 0.006060685496777296, + 0.006212552078068256, + 0.004285062663257122, + 0.008543144911527634, + 0.00998563040047884, + 0.004485964309424162, + 0.0059820255264639854, + 0.0036391443572938442, + 0.011038238182663918, + 0.01132564339786768, + 0.007648122031241655, + 0.08531209081411362, + 0.011149493977427483, + 0.006538083776831627, + 0.07170306891202927, + 0.006100723519921303, + 0.00854114256799221, + 0.004167761653661728, + 0.004483507014811039, + 0.010552267543971539, + 0.011829974129796028, + 0.0043887970969080925, + 0.00846561137586832, + 0.00344811100512743, + 0.01084961462765932, + 0.0037775004748255014, + 0.005847220309078693, + 0.00548304058611393, + 0.007416304666548967, + 0.005756849888712168, + 0.005885214079171419, + 0.004069538787007332, + 0.004004981834441423, + 0.00425442960113287, + 0.0025638232473284006, + 0.006353510543704033, + 0.0036811085883527994, + 0.005324483383446932, + 0.0064795371145009995, + 0.0033705439418554306, + 0.005222218111157417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.008638276718556881 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a craftsman's workshop with a wall covered in tools and piles of lumber?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a dense vortex of flying sawdust and fine dust particles illuminated by warm, golden afternoon sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera perform a fluid, slowly rotating 360-degree motion that eventually comes to an abrupt halt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the dust particles visibly converge, weave, and intertwine to form the outlines and solid texture of a wooden chair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the final object a wooden chair with smooth lines, solid texture, and clearly visible warm oak grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093838855624199 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093838855624199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093838855624199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36120137572288513, + 0.06721211224794388, + 0.03577727824449539, + 0.03291716426610947, + 0.06792770326137543, + 0.09711357951164246, + 0.0520264096558094, + 0.031250208616256714, + 0.16325153410434723, + 0.09132257103919983 + ], + "credit_map": [ + 0.03541282191872597, + 0.06793411821126938, + 0.052782606333494186, + 0.01331601943820715, + 0.03467137739062309, + 0.020295768976211548, + 0.023667892441153526, + 0.11312077939510345, + 0.007924078963696957, + 0.005307369865477085, + 0.0066451989114284515, + 0.0045189703814685345, + 0.007643683347851038, + 0.004925992339849472, + 0.019092850387096405, + 0.011153968051075935, + 0.004347251262515783, + 0.004917779006063938, + 0.00490307342261076, + 0.0047229016199707985, + 0.004547683522105217, + 0.0027943963650614023, + 0.005642860662192106, + 0.003901334013789892, + 0.004950211849063635, + 0.004464368335902691, + 0.005939610302448273, + 0.0037960116751492023, + 0.0043639265932142735, + 0.0021191872656345367, + 0.00442887656390667, + 0.002854973077774048, + 0.008835307322442532, + 0.007545961067080498, + 0.007333098445087671, + 0.012351815588772297, + 0.009120063856244087, + 0.008617788553237915, + 0.005665178410708904, + 0.008458493277430534, + 0.016426723450422287, + 0.008742177858948708, + 0.00961736124008894, + 0.011480134911835194, + 0.009277952834963799, + 0.02247735485434532, + 0.008045206777751446, + 0.011046666651964188, + 0.0026880877558141947, + 0.0025335524696856737, + 0.018418606370687485, + 0.004287112504243851, + 0.0026126273442059755, + 0.014751370996236801, + 0.0029982104897499084, + 0.0037368398625403643, + 0.0022349862847477198, + 0.0030861226841807365, + 0.004711045417934656, + 0.0033403579145669937, + 0.006435627583414316, + 0.005344939883798361, + 0.002372921910136938, + 0.003724206006154418, + 0.02243662253022194, + 0.015887005254626274, + 0.02367165870964527, + 0.01964842528104782, + 0.01651853509247303, + 0.0436517633497715, + 0.009627913124859333, + 0.01180961076170206, + 0.007816768251359463, + 0.006585360039025545, + 0.011352384462952614, + 0.00816124677658081, + 0.018858153373003006, + 0.014730011112987995, + 0.009978312999010086, + 0.013840340077877045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.004093838855624199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_tensor([7133, 7133])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_tensor([7133, 7133])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..050a25c931887258effe7db89f07366881eff31c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_tensor([7133, 7133])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Shot from a slightly elevated angle, the camera slowly tilts upward and gently pans, following a jet-black cat, its silky fur reflecting a silvery sheen in the cool moonlight. Its emerald eyes gleam with a mysterious, all-seeing light in the darkness, and its body is enveloped in a soft, purplish, translucent magical aura that gently pulsates with its movements. Instead of wings, it gracefully swirls through the air with a deliberate ease that defies the laws of physics. Its extended body and long tail, tracing a graceful arc behind it, imbue each glide with rhythmic grace. Against the backdrop is a deep midnight sky, dotted with countless diamond-like stars and a brilliant full moon. Below, a sleeping city, blurred and dotted with lights, looms. The entire scene, rendered in cool tones, exudes a sense of profound tranquility and dreaminess, creating an atmosphere of mystery and magic.", + "seed": 8581, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8581_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06191852688789368, + "vlm_logprob": -0.06191852688789368, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.06191852688789368, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05182025209069252 + ], + "opposite_logprob_tokens": [ + -3.0518202781677246 + ], + "logprob_reward": -0.05182025209069252, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05182025209069252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04365106299519539, + 0.06124149262905121, + 0.029614560306072235, + 0.031592465937137604, + 0.6812805533409119, + 0.009786635637283325, + 0.06249340623617172, + 0.04532342404127121, + 0.020944887772202492, + 0.014071417041122913 + ], + "credit_map": [ + 0.006533999461680651, + 0.006039941683411598, + 0.0046080625616014, + 0.004389807116240263, + 0.00617844145745039, + 0.006668874062597752, + 0.004006080329418182, + 0.0052258591167628765, + 0.011010967195034027, + 0.005035297479480505, + 0.005683375988155603, + 0.0054699634201824665, + 0.009904418140649796, + 0.009468818083405495, + 0.006910266820341349, + 0.007758385967463255, + 0.004088175017386675, + 0.003751940093934536, + 0.0033854073844850063, + 0.004715018440037966, + 0.003912664484232664, + 0.003541225101798773, + 0.0026442210655659437, + 0.003575908485800028, + 0.00446033151820302, + 0.003875863505527377, + 0.004687094129621983, + 0.0034585129469633102, + 0.0025226916186511517, + 0.00670622568577528, + 0.00275464728474617, + 0.003127100644633174, + 0.04846787452697754, + 0.013520326465368271, + 0.29277727007865906, + 0.017663417384028435, + 0.011633730493485928, + 0.03885805606842041, + 0.01744883507490158, + 0.2409111112356186, + 0.0016630031168460846, + 0.001068203360773623, + 0.0018911752849817276, + 0.001248687505722046, + 0.0007889973931014538, + 0.0009074239642359316, + 0.0008520188275724649, + 0.0013671265915036201, + 0.005277686286717653, + 0.004484820645302534, + 0.019379789009690285, + 0.005756858736276627, + 0.002901384374126792, + 0.0027120294980704784, + 0.0034573005978018045, + 0.01852353848516941, + 0.004195006098598242, + 0.00454073678702116, + 0.013377340510487556, + 0.00455294456332922, + 0.0018796685617417097, + 0.0017828939016908407, + 0.003882878227159381, + 0.011111957021057606, + 0.0019881143234670162, + 0.0016591926105320454, + 0.006086674984544516, + 0.0026098277885466814, + 0.0008899721433408558, + 0.000840089051052928, + 0.0022846353240311146, + 0.004586382303386927, + 0.0020442798268049955, + 0.0016667068703100085, + 0.0019185262499377131, + 0.0022514131851494312, + 0.0013759848661720753, + 0.0010756202973425388, + 0.0016669572796672583, + 0.002071928698569536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 1 + ], + "selected_reward": -0.05182025209069252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17993058264255524 + ], + "opposite_logprob_tokens": [ + -1.804930567741394 + ], + "logprob_reward": -0.17993058264255524, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17993058264255524, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08696296811103821, + 0.05984578654170036, + 0.04594894498586655, + 0.020343495532870293, + 0.6278685331344604, + 0.021221192553639412, + 0.0407608263194561, + 0.04512013494968414, + 0.03069237805902958, + 0.02123567834496498 + ], + "credit_map": [ + 0.00817161239683628, + 0.018448766320943832, + 0.006264440715312958, + 0.005718372296541929, + 0.015559435822069645, + 0.014701099134981632, + 0.0071015493012964725, + 0.010997701436281204, + 0.011758222244679928, + 0.0057847327552735806, + 0.009358086623251438, + 0.005126411560922861, + 0.003985022660344839, + 0.008168504573404789, + 0.006051766686141491, + 0.00961304735392332, + 0.005423432681709528, + 0.0033537657000124454, + 0.00422897981479764, + 0.005630199797451496, + 0.00590833555907011, + 0.006696170195937157, + 0.005133727099746466, + 0.009574341587722301, + 0.002926170825958252, + 0.001824337407015264, + 0.0026819880586117506, + 0.002022027736529708, + 0.0023545303847640753, + 0.004846205469220877, + 0.001279586460441351, + 0.00240865140222013, + 0.04599463939666748, + 0.012937257066369057, + 0.2685994505882263, + 0.018021905794739723, + 0.011018339544534683, + 0.035736002027988434, + 0.015693001449108124, + 0.21986795961856842, + 0.0028442745096981525, + 0.0016752511728554964, + 0.005753813311457634, + 0.0022517358884215355, + 0.0017252668039873242, + 0.001609428902156651, + 0.0019707074388861656, + 0.003390717785805464, + 0.00355701451189816, + 0.002708969172090292, + 0.0142994225025177, + 0.003004668513312936, + 0.0014229447115212679, + 0.001519631827250123, + 0.002508734120056033, + 0.01173944491893053, + 0.004528036806732416, + 0.004766819532960653, + 0.011941924691200256, + 0.004346655681729317, + 0.0019935141317546368, + 0.0020693771075457335, + 0.004619020037353039, + 0.010854791849851608, + 0.003207580652087927, + 0.0024926362093538046, + 0.007391887251287699, + 0.0032645384781062603, + 0.0016538582276552916, + 0.001104578492231667, + 0.0027929595671594143, + 0.008784342557191849, + 0.0023719947785139084, + 0.003563550068065524, + 0.0028619978111237288, + 0.004180075600743294, + 0.001710418495349586, + 0.0010430064285174012, + 0.0024919691495597363, + 0.0030126695055514574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.17993058264255524 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004642066545784473 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004642066545784473, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004642066545784473, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10045374184846878, + 0.09224093705415726, + 0.03439740464091301, + 0.05052211135625839, + 0.5481449961662292, + 0.012776191346347332, + 0.06431926786899567, + 0.03433533012866974, + 0.031823284924030304, + 0.030986744910478592 + ], + "credit_map": [ + 0.011946885846555233, + 0.021194152534008026, + 0.009445966221392155, + 0.006370833143591881, + 0.013862291350960732, + 0.015351222828030586, + 0.00886763446033001, + 0.013414759188890457, + 0.01826552301645279, + 0.00883652362972498, + 0.013215266168117523, + 0.006787608843296766, + 0.010634507983922958, + 0.014494824223220348, + 0.008475805632770061, + 0.01153087429702282, + 0.005379481241106987, + 0.005125364288687706, + 0.0028797131963074207, + 0.005076066590845585, + 0.003242030506953597, + 0.005294846836477518, + 0.002272174460813403, + 0.005127730313688517, + 0.005090117454528809, + 0.006478994153439999, + 0.006282089278101921, + 0.00660804333165288, + 0.004700060002505779, + 0.00862076599150896, + 0.0060968538746237755, + 0.006645191460847855, + 0.040087129920721054, + 0.016156962141394615, + 0.23033691942691803, + 0.017323587089776993, + 0.00995181966573, + 0.03136129677295685, + 0.013750407844781876, + 0.18917687237262726, + 0.0014672217657789588, + 0.0016876286827027798, + 0.002061625011265278, + 0.0021242632064968348, + 0.001040296396240592, + 0.0011857886565849185, + 0.0014947197632864118, + 0.001714647631160915, + 0.008908474817872047, + 0.010539982467889786, + 0.01004858035594225, + 0.009315412491559982, + 0.0053823161870241165, + 0.0048557789996266365, + 0.006632237229496241, + 0.00863648857921362, + 0.003659340552985668, + 0.0044969217851758, + 0.0056678531691432, + 0.005555920302867889, + 0.0023553785867989063, + 0.0023650783114135265, + 0.004502688068896532, + 0.005732148420065641, + 0.002991745248436928, + 0.003019624622538686, + 0.007444020826369524, + 0.0037014263216406107, + 0.0020836133044213057, + 0.0012506362982094288, + 0.0030225878581404686, + 0.008309632539749146, + 0.0039250669069588184, + 0.0045146937482059, + 0.003840018529444933, + 0.006021976936608553, + 0.0026537759695202112, + 0.0017226261552423239, + 0.003492293879389763, + 0.004816294647753239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.004642066545784473 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031911423429846764 + ], + "opposite_logprob_tokens": [ + -5.753190994262695 + ], + "logprob_reward": -0.0031911423429846764, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031911423429846764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.035082150250673294, + 0.08767131716012955, + 0.014612858183681965, + 0.02688669227063656, + 0.730115532875061, + 0.021051878109574318, + 0.022377649322152138, + 0.017674999311566353, + 0.027954883873462677, + 0.016571994870901108 + ], + "credit_map": [ + 0.00337403267621994, + 0.004035720601677895, + 0.0028569246642291546, + 0.0033966863993555307, + 0.003637878457084298, + 0.005605658050626516, + 0.0038398888427764177, + 0.008335365913808346, + 0.022181792184710503, + 0.008730651810765266, + 0.012538074515759945, + 0.006176393944770098, + 0.005520184990018606, + 0.014548673294484615, + 0.008109701797366142, + 0.009865849278867245, + 0.0018377708038315177, + 0.0011836286867037416, + 0.0011486350558698177, + 0.0022015415597707033, + 0.001624950091354549, + 0.0019257367821410298, + 0.00165943696629256, + 0.003031159518286586, + 0.003887971630319953, + 0.0031962732318788767, + 0.0030529415234923363, + 0.002710043452680111, + 0.002120691817253828, + 0.006108327768743038, + 0.002037044847384095, + 0.003773400094360113, + 0.05231441929936409, + 0.01333773136138916, + 0.31508344411849976, + 0.019475480541586876, + 0.012962202541530132, + 0.04179937765002251, + 0.018426401540637016, + 0.2567165493965149, + 0.0023296403232961893, + 0.0016818196745589375, + 0.004761048126965761, + 0.0017661354504525661, + 0.001922782976180315, + 0.0015408972976729274, + 0.001875010784715414, + 0.005174544174224138, + 0.002244888339191675, + 0.002914771670475602, + 0.004551846068352461, + 0.004347421228885651, + 0.0017311900155618787, + 0.0016112970188260078, + 0.002277785213664174, + 0.0026984529104083776, + 0.0019042446510866284, + 0.0021527945064008236, + 0.002964683109894395, + 0.0023076192010194063, + 0.0016938449116423726, + 0.001133754849433899, + 0.0025649273302406073, + 0.002953131450340152, + 0.00277168327011168, + 0.002859661355614662, + 0.005699667613953352, + 0.004480655770748854, + 0.002150068525224924, + 0.0014481213875114918, + 0.0042709168046712875, + 0.0042741126380860806, + 0.0018895246321335435, + 0.0027313579339534044, + 0.002325555542483926, + 0.0030987393110990524, + 0.0012962003238499165, + 0.0008098858525045216, + 0.0018688467098399997, + 0.0025518876500427723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.0031911423429846764 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000859081745148 + ], + "opposite_logprob_tokens": [ + -2.6950085163116455 + ], + "logprob_reward": -0.07000859081745148, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000859081745148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12379125505685806, + 0.07476756721735, + 0.09203111380338669, + 0.08266530185937881, + 0.21995016932487488, + 0.13518767058849335, + 0.09191010147333145, + 0.09327182173728943, + 0.043696943670511246, + 0.0427280031144619 + ], + "credit_map": [ + 0.011599283665418625, + 0.022184887900948524, + 0.00857725739479065, + 0.01121100690215826, + 0.021628910675644875, + 0.019974371418356895, + 0.010010804980993271, + 0.018604731187224388, + 0.010560757480561733, + 0.0073113576509058475, + 0.010112551040947437, + 0.006918321829289198, + 0.006964265368878841, + 0.011689240112900734, + 0.007529919501394033, + 0.013681151904165745, + 0.009674137458205223, + 0.0126497196033597, + 0.010922163724899292, + 0.009879174642264843, + 0.01137963030487299, + 0.01082594320178032, + 0.010951615869998932, + 0.015748728066682816, + 0.012556006200611591, + 0.010429704561829567, + 0.013404936529695988, + 0.006945380941033363, + 0.004678723867982626, + 0.014957123436033726, + 0.006247984245419502, + 0.01344543881714344, + 0.020607933402061462, + 0.011336219497025013, + 0.08015439659357071, + 0.008715467527508736, + 0.009500041604042053, + 0.013670840300619602, + 0.007365252822637558, + 0.06860002130270004, + 0.015737656503915787, + 0.0143447769805789, + 0.024554474279284477, + 0.015452216379344463, + 0.011269408278167248, + 0.010338879190385342, + 0.01532957423478365, + 0.028160683810710907, + 0.010259628295898438, + 0.012643099762499332, + 0.017551051452755928, + 0.010371795855462551, + 0.00485956110060215, + 0.005650951061397791, + 0.006842791102826595, + 0.023731227964162827, + 0.009926431812345982, + 0.010159185156226158, + 0.020823488011956215, + 0.010730906389653683, + 0.006190838757902384, + 0.00536450045183301, + 0.00819222442805767, + 0.021884242072701454, + 0.005319276358932257, + 0.003988053184002638, + 0.011332109570503235, + 0.00541003979742527, + 0.0036946656182408333, + 0.0023438811767846346, + 0.004305089823901653, + 0.007303830236196518, + 0.005176527891308069, + 0.005285479128360748, + 0.005629001185297966, + 0.005111678037792444, + 0.005980379413813353, + 0.0039792549796402454, + 0.004683471284806728, + 0.006882212124764919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.07000859081745148 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the camera slowly tilts upward and gently pans, following a jet-black cat, its silky fur reflecting a silvery sheen in the cool moonlight. Its emerald eyes gleam with a mysterious, all-seeing light in the darkness, and its body is enveloped in a soft, purplish, translucent magical aura that gently pulsates with its movements. Instead of wings, it gracefully swirls through the air with a deliberate ease that defies the laws of physics. Its extended body and long tail, tracing a graceful arc behind it, imbue each glide with rhythmic grace. Against the backdrop is a deep midnight sky, dotted with countless diamond-like stars and a brilliant full moon. Below, a sleeping city, blurred and dotted with lights, looms. The entire scene, rendered in cool tones, exudes a sense of profound tranquility and dreaminess, creating an atmosphere of mystery and magic.", + "seed": 8582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8582_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03837393103167415, + "vlm_logprob": -0.03837393103167415, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.03837393103167415, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021826298907399178 + ], + "opposite_logprob_tokens": [ + -4.021826267242432 + ], + "logprob_reward": -0.021826298907399178, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.021826298907399178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12109184265136719, + 0.5083057284355164, + 0.04572492465376854, + 0.044800225645303726, + 0.07478750497102737, + 0.01799928769469261, + 0.031982921063899994, + 0.031931884586811066, + 0.028750503435730934, + 0.09462518990039825 + ], + "credit_map": [ + 0.02299625240266323, + 0.01135226059705019, + 0.016430078074336052, + 0.013159139081835747, + 0.015709787607192993, + 0.011490397155284882, + 0.015120189636945724, + 0.014833743683993816, + 0.04724825918674469, + 0.03649229556322098, + 0.046460773795843124, + 0.08517128974199295, + 0.08798868954181671, + 0.07211633026599884, + 0.07080516964197159, + 0.062022916972637177, + 0.005191384349018335, + 0.004715110640972853, + 0.004931248724460602, + 0.006593940779566765, + 0.006516209337860346, + 0.004787651356309652, + 0.006372079718858004, + 0.006617299281060696, + 0.0028695438522845507, + 0.002970934147015214, + 0.006689933594316244, + 0.00710309250280261, + 0.007215122692286968, + 0.006283533293753862, + 0.007807073649019003, + 0.003860990982502699, + 0.010987384244799614, + 0.004439034499228001, + 0.012714767828583717, + 0.009750224649906158, + 0.011765696108341217, + 0.010686086490750313, + 0.006028423551470041, + 0.008415883406996727, + 0.0031641563400626183, + 0.0017442380776628852, + 0.002577108098194003, + 0.0020465252455323935, + 0.002101458143442869, + 0.0020269814413040876, + 0.0020318091847002506, + 0.0023070103488862514, + 0.0037041015457361937, + 0.0038012820295989513, + 0.0019418955780565739, + 0.004696614108979702, + 0.0042616515420377254, + 0.004661846905946732, + 0.004200468305498362, + 0.0047150603495538235, + 0.0027824505232274532, + 0.0026344629004597664, + 0.003394798841327429, + 0.004497697576880455, + 0.005396656692028046, + 0.004099155776202679, + 0.00456596864387393, + 0.00456069502979517, + 0.003548186272382736, + 0.002385732950642705, + 0.004658303689211607, + 0.0034518172033131123, + 0.003130802884697914, + 0.004000577609986067, + 0.0028658020310103893, + 0.004709281027317047, + 0.009567118249833584, + 0.005372601095587015, + 0.028322383761405945, + 0.005957321729511023, + 0.0057547432370483875, + 0.021565472707152367, + 0.007544850930571556, + 0.010540693067014217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.021826298907399178 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707588076591492 + ], + "opposite_logprob_tokens": [ + -2.1270759105682373 + ], + "logprob_reward": -0.12707588076591492, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707588076591492, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2608892023563385, + 0.22193019092082977, + 0.07981591671705246, + 0.1059688851237297, + 0.08088254928588867, + 0.020044619217514992, + 0.07775256782770157, + 0.02453557960689068, + 0.054755911231040955, + 0.07342453300952911 + ], + "credit_map": [ + 0.027351025491952896, + 0.018112361431121826, + 0.03239675611257553, + 0.03233399987220764, + 0.052152350544929504, + 0.036775603890419006, + 0.030501946806907654, + 0.03126515820622444, + 0.027890203520655632, + 0.013385928235948086, + 0.03013494424521923, + 0.030338065698742867, + 0.026965731754899025, + 0.019576773047447205, + 0.035265881568193436, + 0.03837268054485321, + 0.0081928800791502, + 0.008413859643042088, + 0.008405212312936783, + 0.011022550985217094, + 0.010134789161384106, + 0.008578573353588581, + 0.014177259989082813, + 0.010890793986618519, + 0.0050806221552193165, + 0.012678454630076885, + 0.00975279975682497, + 0.01839468814432621, + 0.015388702042400837, + 0.014313525520265102, + 0.02144590951502323, + 0.00891419779509306, + 0.011227790266275406, + 0.00625670375302434, + 0.008583702147006989, + 0.009544937871396542, + 0.017227917909622192, + 0.01266733929514885, + 0.008000293746590614, + 0.007373872213065624, + 0.002183834556490183, + 0.0021859013941138983, + 0.002614500466734171, + 0.0027595199644565582, + 0.0024715845938771963, + 0.0026545505970716476, + 0.0024857560638338327, + 0.0026889722794294357, + 0.007377708330750465, + 0.007804589346051216, + 0.009959032759070396, + 0.012580839917063713, + 0.010623103938996792, + 0.009418776258826256, + 0.012054787948727608, + 0.007933737710118294, + 0.0027758677024394274, + 0.001915801432915032, + 0.003595201764255762, + 0.0027761610690504313, + 0.003486220957711339, + 0.00376225751824677, + 0.0024711035657674074, + 0.003752965945750475, + 0.006545154377818108, + 0.00435728020966053, + 0.008392720483243465, + 0.006184582598507404, + 0.009568206034600735, + 0.006205731071531773, + 0.0053803506307303905, + 0.008121894672513008, + 0.007986858487129211, + 0.0059481654316186905, + 0.014068354852497578, + 0.008404911495745182, + 0.008606566116213799, + 0.011496965773403645, + 0.008095848374068737, + 0.008816873654723167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.12707588076591492 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004639219027012587 + ], + "opposite_logprob_tokens": [ + -5.379639148712158 + ], + "logprob_reward": -0.004639219027012587, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004639219027012587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18535540997982025, + 0.1788662075996399, + 0.06368819624185562, + 0.0525527149438858, + 0.079071544110775, + 0.06479240208864212, + 0.11719132959842682, + 0.06052897870540619, + 0.0393582321703434, + 0.15859492123126984 + ], + "credit_map": [ + 0.025637757033109665, + 0.014384589157998562, + 0.028519412502646446, + 0.027868585661053658, + 0.021382320672273636, + 0.022900132462382317, + 0.023260122165083885, + 0.021402498707175255, + 0.0243972335010767, + 0.009837394580245018, + 0.024883396923542023, + 0.02102779410779476, + 0.016649508848786354, + 0.013322348706424236, + 0.03018980845808983, + 0.038558728992938995, + 0.010446364991366863, + 0.0044662002474069595, + 0.007262414786964655, + 0.008181546814739704, + 0.00679486058652401, + 0.006625573616474867, + 0.011805785819888115, + 0.008105449378490448, + 0.005015936680138111, + 0.00659330515190959, + 0.008377215825021267, + 0.005772688891738653, + 0.007809571456164122, + 0.007188130170106888, + 0.0063782078213989735, + 0.005417658481746912, + 0.014316613785922527, + 0.0049397386610507965, + 0.012551900930702686, + 0.007425012532621622, + 0.012306354939937592, + 0.011933596804738045, + 0.007501349784433842, + 0.008096977137029171, + 0.00632745586335659, + 0.005647857207804918, + 0.009010417386889458, + 0.0073301843367516994, + 0.00756573211401701, + 0.008820885792374611, + 0.00877736508846283, + 0.011312506161630154, + 0.010920269414782524, + 0.015822621062397957, + 0.009848040528595448, + 0.015446803532540798, + 0.018504031002521515, + 0.015991244465112686, + 0.019355982542037964, + 0.011302333325147629, + 0.007467712741345167, + 0.005720196757465601, + 0.006873627193272114, + 0.00817809533327818, + 0.008830699138343334, + 0.008001744747161865, + 0.005814496893435717, + 0.0096424026414752, + 0.0040010204538702965, + 0.0028828762006014585, + 0.004140171222388744, + 0.005444574635475874, + 0.004821339622139931, + 0.006240546237677336, + 0.0049677356146276, + 0.006859971676021814, + 0.015011203475296497, + 0.007478801999241114, + 0.05550378933548927, + 0.008124807849526405, + 0.007931578904390335, + 0.03986208885908127, + 0.007408631034195423, + 0.01727401278913021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004639219027012587 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1352483481168747, + 0.14072465896606445, + 0.11194737255573273, + 0.053695663809776306, + 0.059254150837659836, + 0.057024996727705, + 0.0845375806093216, + 0.04802228882908821, + 0.154143825173378, + 0.15540112555027008 + ], + "credit_map": [ + 0.0278319101780653, + 0.010775355622172356, + 0.019931765273213387, + 0.015166417695581913, + 0.014714227057993412, + 0.014122280292212963, + 0.017024630680680275, + 0.015681762248277664, + 0.014744753018021584, + 0.009214140474796295, + 0.012278839945793152, + 0.022887295112013817, + 0.020595699548721313, + 0.017170436680316925, + 0.021666478365659714, + 0.022167017683386803, + 0.007887580431997776, + 0.009666940197348595, + 0.014687657356262207, + 0.01650417223572731, + 0.017206551507115364, + 0.010288912802934647, + 0.015713389962911606, + 0.019992168992757797, + 0.004244793672114611, + 0.007478177081793547, + 0.005511080846190453, + 0.006901096552610397, + 0.006169796921312809, + 0.007289932109415531, + 0.009489981457591057, + 0.00661080377176404, + 0.01073523797094822, + 0.0049633146263659, + 0.010090311989188194, + 0.007807585410773754, + 0.0059803505428135395, + 0.009030780754983425, + 0.004209074657410383, + 0.006437499076128006, + 0.006671215873211622, + 0.004212893079966307, + 0.007805026136338711, + 0.005636959336698055, + 0.009347171522676945, + 0.006646695546805859, + 0.009170535951852798, + 0.007534499280154705, + 0.010868714191019535, + 0.0088372016325593, + 0.00794132985174656, + 0.016910329461097717, + 0.010924416594207287, + 0.010320102795958519, + 0.009894355200231075, + 0.008841132745146751, + 0.004059412516653538, + 0.003472385462373495, + 0.007002638652920723, + 0.005907841958105564, + 0.0063514974899590015, + 0.006377968471497297, + 0.00612228736281395, + 0.00872825551778078, + 0.022066494449973106, + 0.010191893205046654, + 0.023367557674646378, + 0.011387323960661888, + 0.02361481264233589, + 0.018930429592728615, + 0.01582265831530094, + 0.02876264974474907, + 0.01492537371814251, + 0.0070393686182796955, + 0.0503174364566803, + 0.008285733871161938, + 0.006842235568910837, + 0.04150332137942314, + 0.009180121123790741, + 0.017307542264461517 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 1 + ], + "selected_reward": -0.004636727273464203 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369152918457985 + ], + "opposite_logprob_tokens": [ + -3.408691644668579 + ], + "logprob_reward": -0.03369152918457985, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369152918457985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056360043585300446, + 0.04325629025697708, + 0.09436021745204926, + 0.10710107535123825, + 0.06804455071687698, + 0.2087586671113968, + 0.23282738029956818, + 0.14820295572280884, + 0.01911771669983864, + 0.02197110652923584 + ], + "credit_map": [ + 0.007627914659678936, + 0.004511951003223658, + 0.008306829258799553, + 0.006472267676144838, + 0.007345887366682291, + 0.007931937463581562, + 0.007737285923212767, + 0.006425969302654266, + 0.006016144994646311, + 0.0030347593128681183, + 0.004746762104332447, + 0.006842349655926228, + 0.004796768072992563, + 0.0038184216246008873, + 0.005877270363271236, + 0.00812381599098444, + 0.011568229645490646, + 0.005369255784898996, + 0.013244610279798508, + 0.008202306926250458, + 0.008658516220748425, + 0.010788424871861935, + 0.013240990228950977, + 0.023287884891033173, + 0.005719311535358429, + 0.013166540302336216, + 0.010403760708868504, + 0.018137138336896896, + 0.015137650072574615, + 0.014135178178548813, + 0.020611125975847244, + 0.009790375828742981, + 0.010091505944728851, + 0.004676144104450941, + 0.009470128454267979, + 0.010517626069486141, + 0.011116080917418003, + 0.009554445743560791, + 0.006290901452302933, + 0.006327716168016195, + 0.026775648817420006, + 0.016331810504198074, + 0.03215919807553291, + 0.026347020640969276, + 0.026546385139226913, + 0.02465643920004368, + 0.027862537652254105, + 0.02807963825762272, + 0.027469176799058914, + 0.022295724600553513, + 0.02776428312063217, + 0.044334933161735535, + 0.026678618043661118, + 0.03025529906153679, + 0.033197443932294846, + 0.02083190158009529, + 0.014139858074486256, + 0.012565902434289455, + 0.018813198432326317, + 0.018054256215691566, + 0.02207677997648716, + 0.019097551703453064, + 0.018733270466327667, + 0.024722132831811905, + 0.0024299633223563433, + 0.0015111149987205863, + 0.0022805025801062584, + 0.002139970427379012, + 0.0028073114808648825, + 0.0025588616263121367, + 0.001962582115083933, + 0.003427409566938877, + 0.0019609183073043823, + 0.0017926006112247705, + 0.004603817127645016, + 0.0024189490359276533, + 0.002642215695232153, + 0.0031992075964808464, + 0.0024754574988037348, + 0.002877941820770502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 7 + ], + "selected_reward": -0.03369152918457985 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the camera slowly tilts upward and gently pans, following a jet-black cat, its silky fur reflecting a silvery sheen in the cool moonlight. Its emerald eyes gleam with a mysterious, all-seeing light in the darkness, and its body is enveloped in a soft, purplish, translucent magical aura that gently pulsates with its movements. Instead of wings, it gracefully swirls through the air with a deliberate ease that defies the laws of physics. Its extended body and long tail, tracing a graceful arc behind it, imbue each glide with rhythmic grace. Against the backdrop is a deep midnight sky, dotted with countless diamond-like stars and a brilliant full moon. Below, a sleeping city, blurred and dotted with lights, looms. The entire scene, rendered in cool tones, exudes a sense of profound tranquility and dreaminess, creating an atmosphere of mystery and magic.", + "seed": 8583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8583_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14384038215503098, + "vlm_logprob": -0.14384038215503098, + "vlm_yes_no_margin": 3.175, + "vlm_token_credit": -0.14384038215503098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2546905279159546 + ], + "opposite_logprob_tokens": [ + -1.5046905279159546 + ], + "logprob_reward": -0.2546905279159546, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2546905279159546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19376374781131744, + 0.3227419853210449, + 0.1488816887140274, + 0.028636861592531204, + 0.04546299949288368, + 0.021231405436992645, + 0.029099976643919945, + 0.021997125819325447, + 0.0635082796216011, + 0.12467585504055023 + ], + "credit_map": [ + 0.025656312704086304, + 0.019119231030344963, + 0.01730366051197052, + 0.02153771184384823, + 0.044553015381097794, + 0.044663552194833755, + 0.011786781251430511, + 0.00914351548999548, + 0.024913011118769646, + 0.03913101553916931, + 0.023710759356617928, + 0.027172530069947243, + 0.03889977186918259, + 0.04489133879542351, + 0.08035160601139069, + 0.04367198050022125, + 0.019366702064871788, + 0.01803627610206604, + 0.025807715952396393, + 0.0185980387032032, + 0.015163025818765163, + 0.012630085460841656, + 0.01849544234573841, + 0.020784424617886543, + 0.004087317269295454, + 0.0037909094244241714, + 0.003803510684520006, + 0.003349010832607746, + 0.0031380567234009504, + 0.0016692259814590216, + 0.003578254720196128, + 0.005220578983426094, + 0.0064374953508377075, + 0.004792790859937668, + 0.004584711045026779, + 0.004446790087968111, + 0.006565959192812443, + 0.007149323355406523, + 0.004878792446106672, + 0.0066071380861103535, + 0.002174747409299016, + 0.0025355848483741283, + 0.0029000318609178066, + 0.005825057625770569, + 0.0019029563991352916, + 0.0017837807536125183, + 0.0024540817830711603, + 0.0016551685985177755, + 0.0028905048966407776, + 0.004158289171755314, + 0.005011731293052435, + 0.004689592402428389, + 0.003108988283202052, + 0.002685777610167861, + 0.0034861131571233273, + 0.0030689837876707315, + 0.002091866685077548, + 0.004380709491670132, + 0.002980780554935336, + 0.003106181276962161, + 0.003162337699905038, + 0.002032739343121648, + 0.002345131244510412, + 0.001897382433526218, + 0.005974388215690851, + 0.010060066357254982, + 0.007723809219896793, + 0.008283763192594051, + 0.008742675185203552, + 0.006922420579940081, + 0.007489799987524748, + 0.008311360143125057, + 0.01001507043838501, + 0.012746128253638744, + 0.02400662936270237, + 0.01032126322388649, + 0.015396864153444767, + 0.027653221040964127, + 0.013301600702106953, + 0.01123509556055069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.2546905279159546 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3872075378894806 + ], + "opposite_logprob_tokens": [ + -1.1372075080871582 + ], + "logprob_reward": -0.3872075378894806, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3872075378894806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20463182032108307, + 0.1249462366104126, + 0.21299363672733307, + 0.07692427933216095, + 0.12353208661079407, + 0.02769165113568306, + 0.059022385627031326, + 0.02037712372839451, + 0.05084509029984474, + 0.09903562068939209 + ], + "credit_map": [ + 0.010339845903217793, + 0.01050031092017889, + 0.008742740377783775, + 0.020417990162968636, + 0.06262331455945969, + 0.062463127076625824, + 0.012304588221013546, + 0.017239905893802643, + 0.009896350093185902, + 0.01498772855848074, + 0.011911300010979176, + 0.011956038884818554, + 0.014890647493302822, + 0.016816819086670876, + 0.02883758582174778, + 0.01564977690577507, + 0.029818406328558922, + 0.025258537381887436, + 0.04993054270744324, + 0.021405553445219994, + 0.017593305557966232, + 0.014653104357421398, + 0.018663231283426285, + 0.03567096218466759, + 0.01354549452662468, + 0.00987870804965496, + 0.010722555220127106, + 0.0066141714341938496, + 0.008252755738794804, + 0.005182479042559862, + 0.008765020407736301, + 0.013963107019662857, + 0.02110373228788376, + 0.017178501933813095, + 0.01570841483771801, + 0.01341369841247797, + 0.018562408164143562, + 0.015428385697305202, + 0.010870540514588356, + 0.011266415938735008, + 0.003714755643159151, + 0.0035086541902273893, + 0.00420848885551095, + 0.00426974892616272, + 0.0030906065367162228, + 0.0033119861036539078, + 0.003341237548738718, + 0.00224617519415915, + 0.004471053835004568, + 0.007840883918106556, + 0.008262342773377895, + 0.013001891784369946, + 0.0066842432133853436, + 0.004975955933332443, + 0.007472589612007141, + 0.00631343200802803, + 0.0021220077760517597, + 0.0025048740208148956, + 0.0029935166239738464, + 0.0033534883987158537, + 0.003007099498063326, + 0.0021731900051236153, + 0.001998589374125004, + 0.0022243584971874952, + 0.00488627515733242, + 0.008264967240393162, + 0.005573900882154703, + 0.0076454258523881435, + 0.007955295033752918, + 0.0046967132948338985, + 0.0066305603832006454, + 0.005191953852772713, + 0.00957142561674118, + 0.01035646814852953, + 0.016585372388362885, + 0.009634713642299175, + 0.01526045985519886, + 0.017496313899755478, + 0.009342322126030922, + 0.010788549669086933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.3872075378894806 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099300131201744 + ], + "opposite_logprob_tokens": [ + -5.504099369049072 + ], + "logprob_reward": -0.004099300131201744, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099300131201744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28350502252578735, + 0.14813698828220367, + 0.2605835199356079, + 0.03352572023868561, + 0.06389279663562775, + 0.04933726787567139, + 0.030474305152893066, + 0.02653035707771778, + 0.06123077869415283, + 0.042783305048942566 + ], + "credit_map": [ + 0.027207938954234123, + 0.021832602098584175, + 0.0268668495118618, + 0.047579072415828705, + 0.04614400491118431, + 0.0463939793407917, + 0.030179796740412712, + 0.037300754338502884, + 0.013047437183558941, + 0.018415018916130066, + 0.012288551777601242, + 0.012426899746060371, + 0.017486553639173508, + 0.02000398002564907, + 0.034798312932252884, + 0.01967022195458412, + 0.032012585550546646, + 0.030048104003071785, + 0.06109875813126564, + 0.028182586655020714, + 0.026625357568264008, + 0.0197465680539608, + 0.02557060681283474, + 0.037298932671546936, + 0.004735331516712904, + 0.003473566845059395, + 0.004223987925797701, + 0.0040847184136509895, + 0.005355236120522022, + 0.0032704018522053957, + 0.00478925509378314, + 0.0035932213068008423, + 0.010456852614879608, + 0.0077621107921004295, + 0.009354444220662117, + 0.010602171532809734, + 0.007789185270667076, + 0.006167354062199593, + 0.005338453687727451, + 0.0064222184009850025, + 0.00500859972089529, + 0.0062839374877512455, + 0.006171330809593201, + 0.011533482000231743, + 0.005298371892422438, + 0.005137844476848841, + 0.005264799110591412, + 0.004638900980353355, + 0.0031006799545139074, + 0.0037451451644301414, + 0.0041724275797605515, + 0.005641613155603409, + 0.0029048072174191475, + 0.0036833540070801973, + 0.003587009385228157, + 0.0036392672918736935, + 0.0018900273134931922, + 0.003928096033632755, + 0.0036268963012844324, + 0.004489646293222904, + 0.0028412574902176857, + 0.0033531736116856337, + 0.002596158068627119, + 0.003805099055171013, + 0.00590479327365756, + 0.009742974303662777, + 0.00737809156998992, + 0.008473312482237816, + 0.008376698940992355, + 0.0063683046028018, + 0.006985480897128582, + 0.008001119829714298, + 0.004685279913246632, + 0.0038949185982346535, + 0.005974180996417999, + 0.004917805548757315, + 0.0046669431030750275, + 0.006560806650668383, + 0.005120901390910149, + 0.006962464191019535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.004099300131201744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031916177831590176 + ], + "opposite_logprob_tokens": [ + -5.7531914710998535 + ], + "logprob_reward": -0.0031916177831590176, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031916177831590176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4892498254776001, + 0.19924035668373108, + 0.016837740316987038, + 0.022096360102295876, + 0.06063811853528023, + 0.029674919322133064, + 0.07399299740791321, + 0.021755710244178772, + 0.06306084990501404, + 0.02345314994454384 + ], + "credit_map": [ + 0.021132471039891243, + 0.018947741016745567, + 0.01836707442998886, + 0.04057547450065613, + 0.12194272875785828, + 0.12032586336135864, + 0.04986707866191864, + 0.09809140861034393, + 0.015555799938738346, + 0.024411074817180634, + 0.014356452040374279, + 0.01759933866560459, + 0.023403581231832504, + 0.027606405317783356, + 0.049621857702732086, + 0.02668583206832409, + 0.0036890117917209864, + 0.00218393187969923, + 0.002368980087339878, + 0.00150138174649328, + 0.0010611932957544923, + 0.0016500413184985518, + 0.0018290812149643898, + 0.0025541181676089764, + 0.006427688058465719, + 0.003060200484469533, + 0.003305514343082905, + 0.0018554525449872017, + 0.0015537213766947389, + 0.0010295120300725102, + 0.0021299091167747974, + 0.0027343605179339647, + 0.010335218161344528, + 0.0058793737553060055, + 0.008272219449281693, + 0.008969896472990513, + 0.005203510634601116, + 0.005953994579613209, + 0.004736138507723808, + 0.01128776092082262, + 0.002945773536339402, + 0.004045210778713226, + 0.003619755618274212, + 0.007707770448178053, + 0.0028324180748313665, + 0.0028238806407898664, + 0.002996100578457117, + 0.0027040094137191772, + 0.010707023553550243, + 0.007580150384455919, + 0.011540811508893967, + 0.016681374981999397, + 0.004340998362749815, + 0.004160743672400713, + 0.010288751684129238, + 0.008693140931427479, + 0.0021083333995193243, + 0.002796705113723874, + 0.0028336101677268744, + 0.00304934149608016, + 0.002706256229430437, + 0.0031519480980932713, + 0.002344203880056739, + 0.00276531046256423, + 0.006640003994107246, + 0.00993063859641552, + 0.007704199757426977, + 0.009239139035344124, + 0.008773578330874443, + 0.0058521670289337635, + 0.007799334824085236, + 0.007121780421584845, + 0.001581778284162283, + 0.0020432774908840656, + 0.0051220678724348545, + 0.001933853141963482, + 0.0031723929569125175, + 0.0056474413722753525, + 0.0020628096535801888, + 0.0018895268440246582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0031916177831590176 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001292705535889 + ], + "opposite_logprob_tokens": [ + -2.6950130462646484 + ], + "logprob_reward": -0.07001292705535889, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001292705535889, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2065829634666443, + 0.3014437258243561, + 0.06650002300739288, + 0.09737015515565872, + 0.11612512171268463, + 0.07226092368364334, + 0.032340966165065765, + 0.0336986780166626, + 0.05204346030950546, + 0.02163395658135414 + ], + "credit_map": [ + 0.010493142530322075, + 0.010714692994952202, + 0.010606767609715462, + 0.012643405236303806, + 0.0588902086019516, + 0.05753392353653908, + 0.01713963784277439, + 0.028561174869537354, + 0.02444499358534813, + 0.038265835493803024, + 0.021926937624812126, + 0.02706875465810299, + 0.03729339316487312, + 0.04196510463953018, + 0.07253808528184891, + 0.037940628826618195, + 0.010250109247863293, + 0.007091760169714689, + 0.015624725259840488, + 0.006741960532963276, + 0.004594729747623205, + 0.005305896978825331, + 0.006046221125870943, + 0.010844622738659382, + 0.012268593534827232, + 0.012426142580807209, + 0.011126943863928318, + 0.010882633738219738, + 0.01044174563139677, + 0.007166499737650156, + 0.01477071177214384, + 0.01828688569366932, + 0.016176335513591766, + 0.014325538650155067, + 0.014265445061028004, + 0.011411238461732864, + 0.009420475922524929, + 0.008568208664655685, + 0.012664003297686577, + 0.029293879866600037, + 0.008320754393935204, + 0.007825531996786594, + 0.011672957800328732, + 0.008983191102743149, + 0.009948651306331158, + 0.00971943885087967, + 0.009409047663211823, + 0.00638135802000761, + 0.002809680998325348, + 0.0029520015232264996, + 0.0074972230941057205, + 0.006393779534846544, + 0.0022334901150316, + 0.0036407799925655127, + 0.0036666281521320343, + 0.003147382754832506, + 0.0028941016644239426, + 0.0045898654498159885, + 0.004128115251660347, + 0.006357118487358093, + 0.004140597768127918, + 0.003781907958909869, + 0.003985783085227013, + 0.003821189049631357, + 0.006139300763607025, + 0.0065567707642912865, + 0.0063701500184834, + 0.0077981301583349705, + 0.007802542299032211, + 0.005903167184442282, + 0.005998743698000908, + 0.005474653095006943, + 0.00195762375369668, + 0.0024009894113987684, + 0.0026860255748033524, + 0.002398831071332097, + 0.003831555601209402, + 0.0029957243241369724, + 0.002902642358094454, + 0.0024605640210211277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.07001292705535889 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Shot from a slightly elevated angle, the camera slowly tilts upward and gently pans, following a jet-black cat, its silky fur reflecting a silvery sheen in the cool moonlight. Its emerald eyes gleam with a mysterious, all-seeing light in the darkness, and its body is enveloped in a soft, purplish, translucent magical aura that gently pulsates with its movements. Instead of wings, it gracefully swirls through the air with a deliberate ease that defies the laws of physics. Its extended body and long tail, tracing a graceful arc behind it, imbue each glide with rhythmic grace. Against the backdrop is a deep midnight sky, dotted with countless diamond-like stars and a brilliant full moon. Below, a sleeping city, blurred and dotted with lights, looms. The entire scene, rendered in cool tones, exudes a sense of profound tranquility and dreaminess, creating an atmosphere of mystery and magic.", + "seed": 8584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_10/index_7133_seed_8584_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09568139798939228, + "vlm_logprob": -0.09568139798939228, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.09568139798939228, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11422491073608398 + ], + "opposite_logprob_tokens": [ + -2.239224910736084 + ], + "logprob_reward": -0.11422491073608398, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11422491073608398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4207073748111725, + 0.019978992640972137, + 0.028093654662370682, + 0.028045058250427246, + 0.01746510900557041, + 0.03207038715481758, + 0.023864414542913437, + 0.07377871870994568, + 0.3078330159187317, + 0.04816325008869171 + ], + "credit_map": [ + 0.01733868010342121, + 0.062130872160196304, + 0.11084018647670746, + 0.05297405272722244, + 0.023308390751481056, + 0.12171104550361633, + 0.00966330710798502, + 0.02274087257683277, + 0.0013709078775718808, + 0.0023550421465188265, + 0.0019530400168150663, + 0.003490487579256296, + 0.0024458204861730337, + 0.002722010714933276, + 0.003431914607062936, + 0.002209770493209362, + 0.001406793249770999, + 0.003175761317834258, + 0.0035394143778830767, + 0.006404544692486525, + 0.00204060529358685, + 0.004049174953252077, + 0.003149355761706829, + 0.004328005947172642, + 0.0010476375464349985, + 0.0020480093080550432, + 0.004189866129308939, + 0.005688945297151804, + 0.002457280410453677, + 0.003498365404084325, + 0.0030327378772199154, + 0.0060822186060249805, + 0.0015132513362914324, + 0.0011526911985129118, + 0.002306561917066574, + 0.0014891051687300205, + 0.0027262999210506678, + 0.00251877773553133, + 0.0025305121671408415, + 0.003227910026907921, + 0.004535532556474209, + 0.0019336645491421223, + 0.0036325498949736357, + 0.0022001483011990786, + 0.003583212150260806, + 0.0059313345700502396, + 0.005295911803841591, + 0.004958033096045256, + 0.0020312273409217596, + 0.002582866931334138, + 0.0025601203087717295, + 0.0019288097973912954, + 0.005405446980148554, + 0.0030996378045529127, + 0.003861568635329604, + 0.0023947374429553747, + 0.006616745609790087, + 0.009581121616065502, + 0.007146546617150307, + 0.009085879661142826, + 0.011785409413278103, + 0.009274495765566826, + 0.011618273332715034, + 0.008670255541801453, + 0.03441258519887924, + 0.04816336929798126, + 0.027724089100956917, + 0.03859134763479233, + 0.05722932890057564, + 0.04545735940337181, + 0.03232984617352486, + 0.023925086483359337, + 0.004853073041886091, + 0.006755937822163105, + 0.006182055454701185, + 0.008004329167306423, + 0.00526703055948019, + 0.007513046730309725, + 0.004812886472791433, + 0.00477489223703742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.11422491073608398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28143858909606934 + ], + "opposite_logprob_tokens": [ + -1.4064385890960693 + ], + "logprob_reward": -0.28143858909606934, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28143858909606934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920924127101898, + 0.19422787427902222, + 0.02741020731627941, + 0.06846409291028976, + 0.019238628447055817, + 0.023689748719334602, + 0.027681779116392136, + 0.04037221148610115, + 0.10630909353494644, + 0.20051392912864685 + ], + "credit_map": [ + 0.014873642474412918, + 0.04512877017259598, + 0.07048068940639496, + 0.033218562602996826, + 0.018612273037433624, + 0.08508408814668655, + 0.008856545202434063, + 0.015837857499718666, + 0.008577226661145687, + 0.04762163385748863, + 0.018727753311395645, + 0.01600985787808895, + 0.016720149666070938, + 0.059788912534713745, + 0.01320972666144371, + 0.013572627678513527, + 0.0017071241745725274, + 0.004366887267678976, + 0.0028089801780879498, + 0.004841456655412912, + 0.0034022442996501923, + 0.003766140667721629, + 0.003941542934626341, + 0.002575831487774849, + 0.0032026960980147123, + 0.005623556207865477, + 0.010289445519447327, + 0.014500470831990242, + 0.003995667677372694, + 0.008954761549830437, + 0.006806772667914629, + 0.015090727247297764, + 0.0014024778502061963, + 0.001343437354080379, + 0.001987022114917636, + 0.0016141716623678803, + 0.0039120321162045, + 0.0025648886803537607, + 0.003618342336267233, + 0.0027962569147348404, + 0.0034993307199329138, + 0.0017628830391913652, + 0.0029653857927769423, + 0.0016125210095196962, + 0.003054291009902954, + 0.0042207990773022175, + 0.0038645078893750906, + 0.0027100290171802044, + 0.0023611965589225292, + 0.0030687584076076746, + 0.003349123289808631, + 0.0020309085957705975, + 0.0038457780610769987, + 0.0036604944616556168, + 0.0043742964044213295, + 0.004991226363927126, + 0.003937836736440659, + 0.005527476780116558, + 0.003720217850059271, + 0.004817119333893061, + 0.006123125087469816, + 0.005489066243171692, + 0.005989566445350647, + 0.0047678034752607346, + 0.010262924246490002, + 0.022014543414115906, + 0.010892439633607864, + 0.013396669179201126, + 0.00987724494189024, + 0.013114399276673794, + 0.0166434645652771, + 0.010107425041496754, + 0.02282257005572319, + 0.024501772597432137, + 0.026613516733050346, + 0.03023851476609707, + 0.020714830607175827, + 0.025963297113776207, + 0.02630554884672165, + 0.023353906348347664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.28143858909606934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004099775105714798 + ], + "opposite_logprob_tokens": [ + -5.5040998458862305 + ], + "logprob_reward": -0.004099775105714798, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004099775105714798, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3032510578632355, + 0.11487200856208801, + 0.04682543873786926, + 0.09391762316226959, + 0.026129448786377907, + 0.02530551142990589, + 0.12482164800167084, + 0.03379005938768387, + 0.20007561147212982, + 0.031011588871479034 + ], + "credit_map": [ + 0.01517848763614893, + 0.03652271628379822, + 0.07818161696195602, + 0.04269271716475487, + 0.01847025193274021, + 0.0765828862786293, + 0.016934027895331383, + 0.018688341602683067, + 0.005110400728881359, + 0.02858133055269718, + 0.013216572813689709, + 0.01239484641700983, + 0.009060225449502468, + 0.030226390808820724, + 0.008694257587194443, + 0.007587979082018137, + 0.0019057219615206122, + 0.006882332265377045, + 0.0051908306777477264, + 0.010533735156059265, + 0.004310543183237314, + 0.0071831741370260715, + 0.005360976327210665, + 0.005458122584968805, + 0.004952356219291687, + 0.008361591957509518, + 0.010355370119214058, + 0.014001937583088875, + 0.005947268567979336, + 0.01409991830587387, + 0.009908170439302921, + 0.026290999725461006, + 0.003169833682477474, + 0.0017839819192886353, + 0.0028077049646526575, + 0.002343561267480254, + 0.004494420252740383, + 0.004272750578820705, + 0.004535234998911619, + 0.0027219608891755342, + 0.0036398579832166433, + 0.0014692316763103008, + 0.002778660273179412, + 0.0017428165301680565, + 0.0034724019933491945, + 0.00451537175104022, + 0.0042848433367908, + 0.0034023260232061148, + 0.011549404822289944, + 0.013715057633817196, + 0.017354056239128113, + 0.025129998102784157, + 0.015470581129193306, + 0.016798147931694984, + 0.0128823546692729, + 0.011922042816877365, + 0.003894842928275466, + 0.005133509635925293, + 0.0037913681007921696, + 0.00399694312363863, + 0.004287710413336754, + 0.005176906939595938, + 0.003217090852558613, + 0.0042916857637465, + 0.020171158015727997, + 0.029508894309401512, + 0.018938478082418442, + 0.022100551053881645, + 0.030874084681272507, + 0.031754348427057266, + 0.028683152049779892, + 0.018044935539364815, + 0.0030263441149145365, + 0.002804873278364539, + 0.004442129284143448, + 0.005517152603715658, + 0.0033133639954030514, + 0.004289841279387474, + 0.0038801776245236397, + 0.003737705759704113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.004099775105714798 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863579474389553 + ], + "opposite_logprob_tokens": [ + -4.758635997772217 + ], + "logprob_reward": -0.00863579474389553, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863579474389553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6435219645500183, + 0.10890580713748932, + 0.03460840508341789, + 0.027227390557527542, + 0.012640190310776234, + 0.017628982663154602, + 0.03839139640331268, + 0.019496841356158257, + 0.07810425013303757, + 0.019474796950817108 + ], + "credit_map": [ + 0.026741327717900276, + 0.09445393085479736, + 0.16950921714305878, + 0.08129021525382996, + 0.03615054860711098, + 0.18316349387168884, + 0.017020517960190773, + 0.03519264608621597, + 0.004982314072549343, + 0.028077922761440277, + 0.011673695407807827, + 0.010632449761033058, + 0.008306693285703659, + 0.033543169498443604, + 0.005378225818276405, + 0.006311324890702963, + 0.0028332078363746405, + 0.004975883290171623, + 0.004503895062953234, + 0.00594380171969533, + 0.0023609590716660023, + 0.006666107568889856, + 0.004141591489315033, + 0.0031829557847231627, + 0.0021443874575197697, + 0.002666287124156952, + 0.004779879003763199, + 0.006300836335867643, + 0.0017354206647723913, + 0.0033525924663990736, + 0.002688091015443206, + 0.0035598932299762964, + 0.0016054435400292277, + 0.0012027656193822622, + 0.00106608506757766, + 0.0012503019534051418, + 0.0019185642013326287, + 0.001965555362403393, + 0.0023977786768227816, + 0.001233694376423955, + 0.0025582981761544943, + 0.0014478943776339293, + 0.0023059207014739513, + 0.001830653054639697, + 0.0022682328708469868, + 0.0026357523165643215, + 0.002678578021004796, + 0.0019036517478525639, + 0.003536597127094865, + 0.004119496326893568, + 0.00517471507191658, + 0.006029617041349411, + 0.004815393127501011, + 0.00552636943757534, + 0.0049491338431835175, + 0.004240068141371012, + 0.0019116704352200031, + 0.002906559966504574, + 0.0016271504573523998, + 0.002968668006360531, + 0.0025067543610930443, + 0.0025927501264959574, + 0.0024557001888751984, + 0.0025275861844420433, + 0.00781967956572771, + 0.010947579517960548, + 0.008688499219715595, + 0.008475019596517086, + 0.013177521526813507, + 0.011493349447846413, + 0.009802320972084999, + 0.0077002765610814095, + 0.002253476297482848, + 0.002008026232942939, + 0.00251587713137269, + 0.0034813005477190018, + 0.0018986009526997805, + 0.0025918392930179834, + 0.0024546454660594463, + 0.0022710287012159824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.00863579474389553 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a jet-black cat with silky fur reflecting a silvery sheen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the cat have emerald eyes that are visible in the darkness?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the cat enveloped in a soft, purplish, translucent magical aura that pulses with its movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a deep midnight sky with diamond-like stars and a brilliant full moon?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start from a slightly elevated angle and slowly tilt upward while gently panning?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000792026519775 + ], + "opposite_logprob_tokens": [ + -2.695007801055908 + ], + "logprob_reward": -0.07000792026519775, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000792026519775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3154378831386566, + 0.162637397646904, + 0.09998113662004471, + 0.05399862304329872, + 0.037467122077941895, + 0.04573791101574898, + 0.07669144123792648, + 0.0745558813214302, + 0.1034775897860527, + 0.03001503460109234 + ], + "credit_map": [ + 0.04486587271094322, + 0.05122718587517738, + 0.03832760825753212, + 0.044804152101278305, + 0.026004506275057793, + 0.0492345467209816, + 0.033304452896118164, + 0.02766953967511654, + 0.007162024732679129, + 0.026515435427427292, + 0.019168339669704437, + 0.020170550793409348, + 0.022008895874023438, + 0.023061197251081467, + 0.02834910899400711, + 0.016201840713620186, + 0.005869582761079073, + 0.00977406557649374, + 0.013218584470450878, + 0.022714078426361084, + 0.009771833196282387, + 0.01026933267712593, + 0.012370433658361435, + 0.015993228182196617, + 0.0032963077537715435, + 0.0037882321048527956, + 0.004141680896282196, + 0.006654553581029177, + 0.004933216143399477, + 0.006593307945877314, + 0.011048723012208939, + 0.013542602770030499, + 0.0020031845197081566, + 0.0019664110150188208, + 0.0030073507223278284, + 0.004643274005502462, + 0.006459509022533894, + 0.004838401917368174, + 0.009272294119000435, + 0.005276698153465986, + 0.007033265195786953, + 0.0023746113292872906, + 0.004516405984759331, + 0.002784837270155549, + 0.006088543217629194, + 0.00804310292005539, + 0.009063204750418663, + 0.0058339438401162624, + 0.006892492063343525, + 0.0074077872559428215, + 0.011398048140108585, + 0.009544131346046925, + 0.011900071986019611, + 0.009840026497840881, + 0.012343961745500565, + 0.007364919409155846, + 0.0073810480535030365, + 0.011848869733512402, + 0.007519985549151897, + 0.008201424032449722, + 0.009878667071461678, + 0.008929108269512653, + 0.009903018362820148, + 0.01089375838637352, + 0.012145120650529861, + 0.016350267454981804, + 0.016932932659983635, + 0.011794346384704113, + 0.011196564882993698, + 0.012711708433926105, + 0.012033500708639622, + 0.010313150472939014, + 0.003033566288650036, + 0.004353842698037624, + 0.0035459937062114477, + 0.003735531121492386, + 0.0037373146042227745, + 0.0043495730496943, + 0.003816507291048765, + 0.003442706074565649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.07000792026519775 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_tensor([390, 390])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_tensor([390, 390])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..2dfde3b2c753b64e17de92084b7501c8f0e04844 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_tensor([390, 390])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, emotive close-up focuses on a wrinkled elderly man. The soft golden afterglow of the evening sun casts a sidelight on his face, sculpting each deep wrinkle on his skin like a dried-up riverbed, dotted with age spots. His slightly cloudy yet still deep eyes gaze unfocused beyond the camera, as if reminiscing about a distant past. His sparse, graying hair is gently ruffled by the breeze. His lips, cracked and pale, tremble almost imperceptibly before slowly opening. Finally, with a slump of his shoulders, he lets out a long, drawn-out sigh, a fleeting puff of white mist in the cool air. The scene is bathed in a warm, slightly melancholic hue, with the completely blurred outlines of the fields in the background, creating an atmosphere of profound contemplation and nostalgia, as if silently telling a long life story.", + "seed": 31885, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_390_seed_31885_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05024843527935445, + "vlm_logprob": -0.05024843527935445, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.05024843527935445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010386107489466667 + ], + "opposite_logprob_tokens": [ + -5.635385990142822 + ], + "logprob_reward": -0.010386107489466667, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.010386107489466667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14531506597995758, + 0.09983273595571518, + 0.031941164284944534, + 0.04163576290011406, + 0.027088593691587448, + 0.025794880464673042, + 0.04805425554513931, + 0.051539309322834015, + 0.13645409047603607, + 0.3923441171646118 + ], + "credit_map": [ + 0.015493868850171566, + 0.015961797907948494, + 0.013914505019783974, + 0.012086289003491402, + 0.02403019368648529, + 0.025882167741656303, + 0.018833069130778313, + 0.01911318302154541, + 0.006887286435812712, + 0.01544073037803173, + 0.009877304546535015, + 0.00768034253269434, + 0.01879078708589077, + 0.017747608944773674, + 0.011243311688303947, + 0.012165358290076256, + 0.0033434447832405567, + 0.002390250563621521, + 0.004117247182875872, + 0.002764421980828047, + 0.006004669237881899, + 0.004058707971125841, + 0.004586808383464813, + 0.004675617907196283, + 0.003995615988969803, + 0.00647116219624877, + 0.0047813234850764275, + 0.004185877740383148, + 0.006228447891771793, + 0.005865451414138079, + 0.004884300287812948, + 0.005223582964390516, + 0.0027251748833805323, + 0.002611480187624693, + 0.002826410112902522, + 0.0023330277763307095, + 0.004483901429921389, + 0.004559208173304796, + 0.0031010638922452927, + 0.004448327235877514, + 0.002237846376374364, + 0.00221504014916718, + 0.002807272132486105, + 0.002567063784226775, + 0.005777567625045776, + 0.0038245520554482937, + 0.0038618361577391624, + 0.0025037028826773167, + 0.005054242443293333, + 0.004264261107891798, + 0.0047038039192557335, + 0.004930206574499607, + 0.00731665687635541, + 0.009933088906109333, + 0.005570938810706139, + 0.006281055510044098, + 0.005197940859943628, + 0.006583620328456163, + 0.006024950183928013, + 0.005746500100940466, + 0.007919896394014359, + 0.007000813260674477, + 0.006265742238610983, + 0.0067998445592820644, + 0.012233905494213104, + 0.02248086966574192, + 0.013490427285432816, + 0.014016576111316681, + 0.0195629745721817, + 0.02404770627617836, + 0.015975184738636017, + 0.01464645192027092, + 0.019471772015094757, + 0.016486817970871925, + 0.13946789503097534, + 0.11385533213615417, + 0.017387211322784424, + 0.029195794835686684, + 0.019330797716975212, + 0.037148505449295044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.010386107489466667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014230675995349884 + ], + "opposite_logprob_tokens": [ + -4.264230728149414 + ], + "logprob_reward": -0.014230675995349884, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014230675995349884, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09043475985527039, + 0.05050702765583992, + 0.029906174167990685, + 0.03425535187125206, + 0.014714177697896957, + 0.01575458236038685, + 0.02664082869887352, + 0.03099440038204193, + 0.03512589633464813, + 0.6716668009757996 + ], + "credit_map": [ + 0.013415772467851639, + 0.008564318530261517, + 0.011228848248720169, + 0.011950509622693062, + 0.016028081998229027, + 0.007775610778480768, + 0.011499115265905857, + 0.009972496889531612, + 0.0054593998938798904, + 0.005516570061445236, + 0.006303609348833561, + 0.00964661967009306, + 0.007015750743448734, + 0.004996457137167454, + 0.0061707873828709126, + 0.005397832952439785, + 0.0030719346832484007, + 0.0027960450388491154, + 0.003717128885909915, + 0.005580361932516098, + 0.004067905712872744, + 0.002868866315111518, + 0.003757571568712592, + 0.004046357702463865, + 0.004920548759400845, + 0.0032506606075912714, + 0.004908366594463587, + 0.0060481056571006775, + 0.005659105721861124, + 0.0028147324919700623, + 0.003983575385063887, + 0.0026702540926635265, + 0.0018815939547494054, + 0.0013913939474150538, + 0.0018359047826379538, + 0.0019482813077047467, + 0.0022164322435855865, + 0.0015997759765014052, + 0.0019260502886027098, + 0.0019147437997162342, + 0.0014657616848126054, + 0.0012602589558809996, + 0.0023114685900509357, + 0.004053597338497639, + 0.001533128903247416, + 0.0012153631541877985, + 0.002404173370450735, + 0.0015108303632587194, + 0.003701155772432685, + 0.0024599297903478146, + 0.00349116581492126, + 0.005272822920233011, + 0.0035083063412457705, + 0.0023114625364542007, + 0.0032169297337532043, + 0.0026790539268404245, + 0.0028358441777527332, + 0.0021166661754250526, + 0.004707752261310816, + 0.008044539019465446, + 0.0030653064604848623, + 0.003624094184488058, + 0.003280868288129568, + 0.0033193284180015326, + 0.0028443739283829927, + 0.006425292231142521, + 0.0037282316479831934, + 0.003907349426299334, + 0.005252482835203409, + 0.005738778039813042, + 0.004042729269713163, + 0.003186658024787903, + 0.03299704194068909, + 0.02037835493683815, + 0.25583451986312866, + 0.20857246220111847, + 0.015865115448832512, + 0.04338172823190689, + 0.014939649030566216, + 0.07969792932271957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014230675995349884 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307721555233002 + ], + "opposite_logprob_tokens": [ + -3.168077230453491 + ], + "logprob_reward": -0.04307721555233002, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307721555233002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03357382118701935, + 0.051670320332050323, + 0.012239601463079453, + 0.013473178260028362, + 0.02617723122239113, + 0.015074338763952255, + 0.01924813911318779, + 0.08226679265499115, + 0.21931235492229462, + 0.5269641876220703 + ], + "credit_map": [ + 0.0030362403485924006, + 0.0037255093920975924, + 0.0035090604797005653, + 0.003025231184437871, + 0.0050894273445010185, + 0.005290290340781212, + 0.004859123378992081, + 0.005038943607360125, + 0.0037056722212582827, + 0.00738554960116744, + 0.005060436669737101, + 0.003142849076539278, + 0.011311561800539494, + 0.00977657362818718, + 0.006629878655076027, + 0.00465780682861805, + 0.0009734805789776146, + 0.0016313035739585757, + 0.0014225346967577934, + 0.001679625129327178, + 0.0016786643536761403, + 0.0015038358978927135, + 0.002163998084142804, + 0.0011861607199534774, + 0.0019652352202683687, + 0.0014194078976288438, + 0.0020373754668980837, + 0.001374230720102787, + 0.002107736188918352, + 0.001806021435186267, + 0.001135349040850997, + 0.0016278235707432032, + 0.0016420409083366394, + 0.0032408980187028646, + 0.0029100063256919384, + 0.001696395338512957, + 0.005990303587168455, + 0.004521777853369713, + 0.003003602847456932, + 0.0031722071580588818, + 0.001439491519704461, + 0.0010371137177571654, + 0.001639022259041667, + 0.0012921120505779982, + 0.0029634577222168446, + 0.002385779982432723, + 0.0024583812337368727, + 0.001858981093391776, + 0.0023899408988654613, + 0.0016356679843738675, + 0.0027823983691632748, + 0.0028035903815180063, + 0.0028009696397930384, + 0.002723441459238529, + 0.002404688159003854, + 0.0017074444331228733, + 0.006678737699985504, + 0.007241522893309593, + 0.01146197970956564, + 0.017007943242788315, + 0.01001761294901371, + 0.010475781746208668, + 0.009430673904716969, + 0.009952551685273647, + 0.027134450152516365, + 0.029924895614385605, + 0.02358715794980526, + 0.023151839151978493, + 0.029349088668823242, + 0.037719279527664185, + 0.026023127138614655, + 0.022422531619668007, + 0.02134202979505062, + 0.02078692428767681, + 0.18877862393856049, + 0.14880366623401642, + 0.017103606835007668, + 0.04100743308663368, + 0.01861388422548771, + 0.0705280601978302 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.04307721555233002 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.179927796125412 + ], + "opposite_logprob_tokens": [ + -1.8049278259277344 + ], + "logprob_reward": -0.179927796125412, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.179927796125412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02770676650106907, + 0.01411648653447628, + 0.007547642104327679, + 0.009255649521946907, + 0.009508204646408558, + 0.006665194872766733, + 0.016479574143886566, + 0.020479146391153336, + 0.1313266158103943, + 0.7569147348403931 + ], + "credit_map": [ + 0.003525064792484045, + 0.002538911998271942, + 0.003100419184193015, + 0.003160021733492613, + 0.0053933956660330296, + 0.0032914266921579838, + 0.0037809088826179504, + 0.0029166170861572027, + 0.0014360274653881788, + 0.0015505708288401365, + 0.0014861610252410173, + 0.0015891273505985737, + 0.002654861891642213, + 0.001796044409275055, + 0.0020141650456935167, + 0.0015895288670435548, + 0.0006896806298755109, + 0.0008225581259466708, + 0.0009038077550940216, + 0.0008356138132512569, + 0.001568219275213778, + 0.0012383941793814301, + 0.0007860798505134881, + 0.0007032885332591832, + 0.0010676278034225106, + 0.0009942284086719155, + 0.0013568333815783262, + 0.0008619502186775208, + 0.0018452709773555398, + 0.0011701345210894942, + 0.0011551459319889545, + 0.0008044581045396626, + 0.0009480336448177695, + 0.0007437012973241508, + 0.0012950522359460592, + 0.0005949749029241502, + 0.0021287957206368446, + 0.0015819966793060303, + 0.00129306607414037, + 0.000922584265936166, + 0.0009997845627367496, + 0.0005175663973204792, + 0.0008301006746478379, + 0.0012768710730597377, + 0.0008199728326871991, + 0.0007115546613931656, + 0.0008346399408765137, + 0.0006747049046680331, + 0.001448099035769701, + 0.0011946921003982425, + 0.0020164300221949816, + 0.0031734188087284565, + 0.0020801061764359474, + 0.0021824410650879145, + 0.002648028312250972, + 0.0017363588558509946, + 0.0023252898827195168, + 0.0018610344268381596, + 0.0025079732295125723, + 0.0031428791116923094, + 0.0028385070618242025, + 0.0028803087770938873, + 0.002923107473179698, + 0.002000048290938139, + 0.009682437404990196, + 0.022739414125680923, + 0.012560724280774593, + 0.012971397489309311, + 0.021135883405804634, + 0.02322208322584629, + 0.016998611390590668, + 0.012016051448881626, + 0.0369156077504158, + 0.028196999803185463, + 0.27816590666770935, + 0.23192551732063293, + 0.02663489617407322, + 0.05742955207824707, + 0.028796885162591934, + 0.06884940713644028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.179927796125412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362038123421371 + ], + "opposite_logprob_tokens": [ + -5.628620147705078 + ], + "logprob_reward": -0.00362038123421371, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362038123421371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10499507933855057, + 0.03668771684169769, + 0.01938237063586712, + 0.024219075217843056, + 0.021117866039276123, + 0.027416426688432693, + 0.03284180536866188, + 0.04648956283926964, + 0.033332813531160355, + 0.6535172462463379 + ], + "credit_map": [ + 0.017235800623893738, + 0.010197232477366924, + 0.013036448508501053, + 0.01394924707710743, + 0.016076045110821724, + 0.011898629367351532, + 0.012233220972120762, + 0.010368462651968002, + 0.00633616978302598, + 0.0035515367053449154, + 0.004244885873049498, + 0.006058120634406805, + 0.004382910206913948, + 0.004261078778654337, + 0.004201354458928108, + 0.0036516631953418255, + 0.0019939763005822897, + 0.0018641198985278606, + 0.0027912803925573826, + 0.0024941356386989355, + 0.003356295870617032, + 0.0020968508906662464, + 0.0027007246389985085, + 0.002084987470880151, + 0.0037046039942651987, + 0.0025889493990689516, + 0.003079513320699334, + 0.0029606581665575504, + 0.004134656395763159, + 0.0028862119652330875, + 0.0024885002057999372, + 0.002375983400270343, + 0.002075842348858714, + 0.002135191112756729, + 0.0029155598022043705, + 0.0023080878891050816, + 0.004378507845103741, + 0.0029278926085680723, + 0.002234627492725849, + 0.0021421576384454966, + 0.0036795889027416706, + 0.002597403945401311, + 0.0037585946265608072, + 0.004709532018750906, + 0.003975261002779007, + 0.003333492437377572, + 0.0032033456955105066, + 0.002159211551770568, + 0.003894147928804159, + 0.0024715785402804613, + 0.005175622180104256, + 0.00827836710959673, + 0.0036402286496013403, + 0.0031738844700157642, + 0.003484513610601425, + 0.002723465673625469, + 0.0037329494953155518, + 0.006277103442698717, + 0.006979647092521191, + 0.010904081165790558, + 0.0045684282667934895, + 0.004699708428233862, + 0.004591919016093016, + 0.004735727794468403, + 0.0041060736402869225, + 0.003768431954085827, + 0.004777329042553902, + 0.005393872037529945, + 0.004110748879611492, + 0.004969698376953602, + 0.0032416570466011763, + 0.0029650037176907063, + 0.051689449697732925, + 0.028947632759809494, + 0.20882906019687653, + 0.15780021250247955, + 0.01926286891102791, + 0.0543055459856987, + 0.019203143194317818, + 0.1134793683886528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.00362038123421371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, emotive close-up focuses on a wrinkled elderly man. The soft golden afterglow of the evening sun casts a sidelight on his face, sculpting each deep wrinkle on his skin like a dried-up riverbed, dotted with age spots. His slightly cloudy yet still deep eyes gaze unfocused beyond the camera, as if reminiscing about a distant past. His sparse, graying hair is gently ruffled by the breeze. His lips, cracked and pale, tremble almost imperceptibly before slowly opening. Finally, with a slump of his shoulders, he lets out a long, drawn-out sigh, a fleeting puff of white mist in the cool air. The scene is bathed in a warm, slightly melancholic hue, with the completely blurred outlines of the fields in the background, creating an atmosphere of profound contemplation and nostalgia, as if silently telling a long life story.", + "seed": 31886, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_390_seed_31886_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11897175768390297, + "vlm_logprob": -0.11897175768390297, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.11897175768390297, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01147562637925148 + ], + "opposite_logprob_tokens": [ + -5.886475563049316 + ], + "logprob_reward": -0.01147562637925148, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.01147562637925148, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24126344919204712, + 0.04905426502227783, + 0.031463123857975006, + 0.0499819852411747, + 0.03716916963458061, + 0.07971012592315674, + 0.15975281596183777, + 0.07366189360618591, + 0.12318737804889679, + 0.1547558605670929 + ], + "credit_map": [ + 0.026343531906604767, + 0.03806891664862633, + 0.030429445207118988, + 0.04611937329173088, + 0.022869257256388664, + 0.03127702698111534, + 0.020898493006825447, + 0.025257403030991554, + 0.006612605880945921, + 0.011155622079968452, + 0.006335382349789143, + 0.010199441574513912, + 0.0036956495605409145, + 0.0033093097154051065, + 0.002954226452857256, + 0.004792030435055494, + 0.004060645122081041, + 0.00669679744169116, + 0.004617322236299515, + 0.005879414267838001, + 0.0029208476189523935, + 0.00190462835598737, + 0.0023063758853822947, + 0.003077093977481127, + 0.008913243189454079, + 0.011447714641690254, + 0.005878510884940624, + 0.008101520128548145, + 0.0037991460412740707, + 0.004425602499395609, + 0.0029622255824506283, + 0.004454020876437426, + 0.00628800829872489, + 0.007878952659666538, + 0.005639636889100075, + 0.005869194399565458, + 0.002583218039944768, + 0.0034429004881531, + 0.0024365419521927834, + 0.0030307192355394363, + 0.01757999323308468, + 0.019469937309622765, + 0.00897946860641241, + 0.009023054502904415, + 0.005609998479485512, + 0.008713778108358383, + 0.003805429209023714, + 0.0065284608863294125, + 0.050259605050086975, + 0.027121582999825478, + 0.020785314962267876, + 0.016509512439370155, + 0.008511925116181374, + 0.012929754331707954, + 0.006982841994613409, + 0.01665227673947811, + 0.025009673088788986, + 0.010843885131180286, + 0.01030781865119934, + 0.007274923380464315, + 0.003908408340066671, + 0.006763227749615908, + 0.0031787187326699495, + 0.006375241093337536, + 0.019834322854876518, + 0.024428321048617363, + 0.01813945174217224, + 0.0172042865306139, + 0.011749421246349812, + 0.008055654354393482, + 0.008159717544913292, + 0.015616207383573055, + 0.02769411914050579, + 0.025921611115336418, + 0.016265572980046272, + 0.02438698150217533, + 0.015428843908011913, + 0.019129639491438866, + 0.011394540779292583, + 0.014534545131027699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.01147562637925148 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02333221398293972 + ], + "opposite_logprob_tokens": [ + -3.773332118988037 + ], + "logprob_reward": -0.02333221398293972, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02333221398293972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13790951669216156, + 0.05478399991989136, + 0.036115776747465134, + 0.03823593631386757, + 0.09261034429073334, + 0.12531565129756927, + 0.14210261404514313, + 0.1237073540687561, + 0.029998648911714554, + 0.21922017633914948 + ], + "credit_map": [ + 0.019229285418987274, + 0.0223378948867321, + 0.014392584562301636, + 0.029610326513648033, + 0.01569971814751625, + 0.010906869545578957, + 0.009544719010591507, + 0.016188116744160652, + 0.009606966748833656, + 0.009972263127565384, + 0.006592031102627516, + 0.010453348979353905, + 0.003409869270399213, + 0.006402942817658186, + 0.003013735171407461, + 0.00533284479752183, + 0.00836186297237873, + 0.006155576556921005, + 0.005261418875306845, + 0.005123160779476166, + 0.0029101429972797632, + 0.0023348440881818533, + 0.0025920062325894833, + 0.0033767642453312874, + 0.007888407446444035, + 0.0071944184601306915, + 0.005148522090166807, + 0.005034835543483496, + 0.0029632572550326586, + 0.0032502682879567146, + 0.002847644966095686, + 0.003908584360033274, + 0.012787681072950363, + 0.01701314188539982, + 0.016128212213516235, + 0.01670379377901554, + 0.006451852153986692, + 0.010187285952270031, + 0.005414648912847042, + 0.00792373064905405, + 0.027295732870697975, + 0.024925490841269493, + 0.019945940002799034, + 0.020410429686307907, + 0.007738019805401564, + 0.01020430400967598, + 0.00645291106775403, + 0.008342820219695568, + 0.04316769167780876, + 0.024800980463624, + 0.018688805401325226, + 0.0160756167024374, + 0.009349478408694267, + 0.011057361960411072, + 0.006240374408662319, + 0.012722288258373737, + 0.034634169191122055, + 0.021075081080198288, + 0.014515131711959839, + 0.013409283012151718, + 0.008523894473910332, + 0.01073838397860527, + 0.006262312643229961, + 0.014549092389643192, + 0.006536742206662893, + 0.004954509437084198, + 0.0034561804495751858, + 0.0055246115662157536, + 0.0018161835614591837, + 0.0028057917952537537, + 0.001604800927452743, + 0.003299828851595521, + 0.06868698447942734, + 0.03402238339185715, + 0.030390428379178047, + 0.020918823778629303, + 0.015001300722360611, + 0.017296282574534416, + 0.014383216388523579, + 0.018520765006542206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.02333221398293972 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899396121501923 + ], + "opposite_logprob_tokens": [ + -2.578994035720825 + ], + "logprob_reward": -0.07899396121501923, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899396121501923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08394929766654968, + 0.04560575634241104, + 0.045263223350048065, + 0.03266653046011925, + 0.04306252300739288, + 0.10748831182718277, + 0.12401125580072403, + 0.0579422228038311, + 0.1760130226612091, + 0.2839977741241455 + ], + "credit_map": [ + 0.011752504855394363, + 0.019012529402971268, + 0.009766043163836002, + 0.018327079713344574, + 0.007710098754614592, + 0.004288278985768557, + 0.0054474640637636185, + 0.00764530198648572, + 0.005632069427520037, + 0.010339271277189255, + 0.004663936328142881, + 0.010625422932207584, + 0.003991334233433008, + 0.003058838192373514, + 0.0031160633079707623, + 0.004178821574896574, + 0.005698958411812782, + 0.008160577155649662, + 0.005025884602218866, + 0.009868115186691284, + 0.0045420266687870026, + 0.0033642896451056004, + 0.0034138683695346117, + 0.005189503543078899, + 0.00427595991641283, + 0.0065183439292013645, + 0.003668735967949033, + 0.007026353385299444, + 0.0021676276810467243, + 0.0028530138079077005, + 0.0025627808645367622, + 0.0035937170032411814, + 0.008339213207364082, + 0.008927131071686745, + 0.0057099349796772, + 0.0068524847738444805, + 0.002276873216032982, + 0.003563677193596959, + 0.0025933110155165195, + 0.004799897316843271, + 0.017984796315431595, + 0.022063471376895905, + 0.013819425366818905, + 0.019042154774069786, + 0.006309413816779852, + 0.008354730904102325, + 0.006241629831492901, + 0.013672687113285065, + 0.03605597838759422, + 0.020223084837198257, + 0.015602403320372105, + 0.01571684144437313, + 0.006959979422390461, + 0.010951579548418522, + 0.005089220125228167, + 0.0134121747687459, + 0.01507231593132019, + 0.010528255254030228, + 0.00786960031837225, + 0.0071020303294062614, + 0.003317894646897912, + 0.004930595867335796, + 0.002568058902397752, + 0.006553471088409424, + 0.0374118909239769, + 0.03383651003241539, + 0.023680703714489937, + 0.02216169983148575, + 0.013868260197341442, + 0.01163474377244711, + 0.011823859065771103, + 0.02159535326063633, + 0.030080918222665787, + 0.04241393506526947, + 0.027173280715942383, + 0.04565541446208954, + 0.030913522467017174, + 0.04432785511016846, + 0.03085811249911785, + 0.03257475793361664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.07899396121501923 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.47431617975234985 + ], + "opposite_logprob_tokens": [ + -0.9743161797523499 + ], + "logprob_reward": -0.47431617975234985, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.47431617975234985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13010169565677643, + 0.048432089388370514, + 0.05429654195904732, + 0.07177356630563736, + 0.09823612868785858, + 0.0850587785243988, + 0.11277991533279419, + 0.08460181206464767, + 0.08861079812049866, + 0.2261086255311966 + ], + "credit_map": [ + 0.017283663153648376, + 0.02760322391986847, + 0.020119542255997658, + 0.023691916838288307, + 0.009862334467470646, + 0.008783494122326374, + 0.008501900359988213, + 0.014255616813898087, + 0.008710140362381935, + 0.008895852603018284, + 0.005914618726819754, + 0.007483776193112135, + 0.0035143119748681784, + 0.004548678174614906, + 0.002488234778866172, + 0.006876479834318161, + 0.009181491099298, + 0.00980411097407341, + 0.00578071316704154, + 0.008134971372783184, + 0.006243437062948942, + 0.004610872361809015, + 0.00374143966473639, + 0.00679950462654233, + 0.009953605011105537, + 0.018682735040783882, + 0.008641029708087444, + 0.01251618005335331, + 0.006632116623222828, + 0.0038944531697779894, + 0.005217695143073797, + 0.006235749460756779, + 0.020560942590236664, + 0.01574805937707424, + 0.013733337633311749, + 0.01503319013863802, + 0.006240549962967634, + 0.012120107188820839, + 0.004973750561475754, + 0.009826196357607841, + 0.02045060321688652, + 0.01818358711898327, + 0.010671192780137062, + 0.012724194675683975, + 0.004921725019812584, + 0.006629606243222952, + 0.004065678920596838, + 0.00741219287738204, + 0.02035333774983883, + 0.023194892331957817, + 0.01604144647717476, + 0.015584008768200874, + 0.010194032453000546, + 0.008370221592485905, + 0.006212901324033737, + 0.012829073704779148, + 0.012876760214567184, + 0.015571411699056625, + 0.010874936357140541, + 0.013057108037173748, + 0.006239235866814852, + 0.006619472987949848, + 0.00597879895940423, + 0.01338408887386322, + 0.02241445705294609, + 0.014519684948027134, + 0.010457978583872318, + 0.013834725134074688, + 0.004653970245271921, + 0.008987285196781158, + 0.005178868304938078, + 0.008563819341361523, + 0.048540402203798294, + 0.026853304356336594, + 0.025617795065045357, + 0.025160277262330055, + 0.024784931913018227, + 0.03138228878378868, + 0.022745121270418167, + 0.021024487912654877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.47431617975234985 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006740807089954615 + ], + "opposite_logprob_tokens": [ + -5.006740570068359 + ], + "logprob_reward": -0.006740807089954615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006740807089954615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08305379748344421, + 0.04225684329867363, + 0.028705287724733353, + 0.05303964391350746, + 0.042862050235271454, + 0.35771289467811584, + 0.06925559788942337, + 0.06322470307350159, + 0.050419051200151443, + 0.20947018265724182 + ], + "credit_map": [ + 0.011495424434542656, + 0.013931184075772762, + 0.013795976527035236, + 0.014272351749241352, + 0.006961329374462366, + 0.008332675322890282, + 0.00456336559727788, + 0.009701477363705635, + 0.00799267366528511, + 0.0071240817196667194, + 0.006658078171312809, + 0.00530247064307332, + 0.0023556933738291264, + 0.006027222611010075, + 0.002844818402081728, + 0.003951799124479294, + 0.004250815138220787, + 0.00629759393632412, + 0.0036855721846222878, + 0.006504527293145657, + 0.002048494992777705, + 0.0013227717718109488, + 0.002218681387603283, + 0.0023768304381519556, + 0.005663691554218531, + 0.014637498185038567, + 0.008313421159982681, + 0.011372982524335384, + 0.003534948453307152, + 0.003135210834443569, + 0.0029900928493589163, + 0.003391796024516225, + 0.008973419666290283, + 0.007850629277527332, + 0.006523472256958485, + 0.007524937391281128, + 0.002168734557926655, + 0.003629553597420454, + 0.0027316377963870764, + 0.0034596628975123167, + 0.10030052810907364, + 0.07534535229206085, + 0.03812616318464279, + 0.030808525159955025, + 0.022172264754772186, + 0.03790294751524925, + 0.017744705080986023, + 0.03531239181756973, + 0.015917479991912842, + 0.012213144451379776, + 0.009401329793035984, + 0.01058302540332079, + 0.005916336085647345, + 0.003991180565208197, + 0.004217890556901693, + 0.00701520312577486, + 0.020489230751991272, + 0.009800191037356853, + 0.007896759547293186, + 0.005297275725752115, + 0.0034150714054703712, + 0.005853765644133091, + 0.0026876849588006735, + 0.007784723304212093, + 0.016010206192731857, + 0.00826320517808199, + 0.006605679169297218, + 0.0049443477764725685, + 0.003012212924659252, + 0.0049415696412324905, + 0.001996217994019389, + 0.004645610228180885, + 0.05714801698923111, + 0.0301649272441864, + 0.026334267109632492, + 0.027371658012270927, + 0.020124871283769608, + 0.016624750569462776, + 0.013402999378740788, + 0.018298670649528503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.006740807089954615 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, emotive close-up focuses on a wrinkled elderly man. The soft golden afterglow of the evening sun casts a sidelight on his face, sculpting each deep wrinkle on his skin like a dried-up riverbed, dotted with age spots. His slightly cloudy yet still deep eyes gaze unfocused beyond the camera, as if reminiscing about a distant past. His sparse, graying hair is gently ruffled by the breeze. His lips, cracked and pale, tremble almost imperceptibly before slowly opening. Finally, with a slump of his shoulders, he lets out a long, drawn-out sigh, a fleeting puff of white mist in the cool air. The scene is bathed in a warm, slightly melancholic hue, with the completely blurred outlines of the fields in the background, creating an atmosphere of profound contemplation and nostalgia, as if silently telling a long life story.", + "seed": 31887, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_390_seed_31887_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08874524161219596, + "vlm_logprob": -0.08874524161219596, + "vlm_yes_no_margin": 3.999999988079071, + "vlm_token_credit": -0.08874524161219596, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00557285500690341 + ], + "opposite_logprob_tokens": [ + -7.00557279586792 + ], + "logprob_reward": -0.00557285500690341, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.00557285500690341, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21971336007118225, + 0.10068976879119873, + 0.10360462963581085, + 0.18933692574501038, + 0.033647093921899796, + 0.03160647675395012, + 0.05230025574564934, + 0.05639924854040146, + 0.13442842662334442, + 0.07827381789684296 + ], + "credit_map": [ + 0.022093918174505234, + 0.0171784870326519, + 0.034684013575315475, + 0.012313954532146454, + 0.019088350236415863, + 0.05426706746220589, + 0.024977827444672585, + 0.0351097397506237, + 0.006127535831183195, + 0.02655816450715065, + 0.018841862678527832, + 0.010169616900384426, + 0.00919848121702671, + 0.008219747804105282, + 0.009357383474707603, + 0.012216981500387192, + 0.011417106725275517, + 0.020513156428933144, + 0.01651749759912491, + 0.010234829038381577, + 0.01204795204102993, + 0.007111418526619673, + 0.015090540051460266, + 0.01067212875932455, + 0.024884233251214027, + 0.01115687657147646, + 0.06545478105545044, + 0.018971553072333336, + 0.01301671378314495, + 0.00987331010401249, + 0.02664930745959282, + 0.019330153241753578, + 0.002564640250056982, + 0.0034362005535513163, + 0.007492804434150457, + 0.0024193881545215845, + 0.0034543124493211508, + 0.0054241870529949665, + 0.0042983656749129295, + 0.004557195585221052, + 0.0023639299906790257, + 0.0024759077932685614, + 0.004023835528641939, + 0.0021306946873664856, + 0.004773260094225407, + 0.005061225034296513, + 0.004652949515730143, + 0.006124674342572689, + 0.006056268699467182, + 0.00427356269210577, + 0.006363474298268557, + 0.0031221096869558096, + 0.007100942078977823, + 0.009126603603363037, + 0.006453140638768673, + 0.009804151020944118, + 0.005639625713229179, + 0.004022690001875162, + 0.008730598725378513, + 0.0037189852446317673, + 0.006131004076451063, + 0.010841180570423603, + 0.008503411896526814, + 0.008811749517917633, + 0.008866913616657257, + 0.014307069592177868, + 0.007991495542228222, + 0.02007845789194107, + 0.03241478279232979, + 0.01016951259225607, + 0.022926900535821915, + 0.01767328381538391, + 0.005645054392516613, + 0.014445415697991848, + 0.007289835251867771, + 0.008331121876835823, + 0.011978900991380215, + 0.010794701054692268, + 0.010800983756780624, + 0.008987801149487495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.00557285500690341 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014222919009625912 + ], + "opposite_logprob_tokens": [ + -4.264223098754883 + ], + "logprob_reward": -0.014222919009625912, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014222919009625912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11400426179170609, + 0.1441129893064499, + 0.11030071973800659, + 0.13118860125541687, + 0.014856347814202309, + 0.014512094669044018, + 0.027166852727532387, + 0.2621985077857971, + 0.10322590917348862, + 0.07843374460935593 + ], + "credit_map": [ + 0.008452208712697029, + 0.01687224954366684, + 0.009126034565269947, + 0.01225324533879757, + 0.013963742181658745, + 0.019790878519415855, + 0.015334900468587875, + 0.018211007118225098, + 0.01319156400859356, + 0.010523301549255848, + 0.036517053842544556, + 0.009760209359228611, + 0.0130051514133811, + 0.025788169354200363, + 0.014874082058668137, + 0.020453443750739098, + 0.005650061648339033, + 0.016783716157078743, + 0.019665101543068886, + 0.007061557844281197, + 0.00876070186495781, + 0.009620890021324158, + 0.013173687271773815, + 0.029585007578134537, + 0.009008292108774185, + 0.04027349129319191, + 0.031137321144342422, + 0.012207862921059132, + 0.010251575149595737, + 0.008301555179059505, + 0.011661732569336891, + 0.008346766233444214, + 0.001125690876506269, + 0.001781061408109963, + 0.0021494445390999317, + 0.0013184320414438844, + 0.0014481351245194674, + 0.002773945452645421, + 0.001760571263730526, + 0.002499067224562168, + 0.001960139721632004, + 0.0009872738737612963, + 0.0019566393457353115, + 0.0007390002720057964, + 0.0015614207368344069, + 0.0022278730757534504, + 0.0016508395783603191, + 0.0034289085306227207, + 0.003441479289904237, + 0.001886242302134633, + 0.004902931395918131, + 0.001722018700093031, + 0.003063490381464362, + 0.003520847298204899, + 0.003006037324666977, + 0.005623804870992899, + 0.037046026438474655, + 0.022186201065778732, + 0.036823589354753494, + 0.017340851947665215, + 0.032451264560222626, + 0.06350528448820114, + 0.025033190846443176, + 0.02781210094690323, + 0.007143397815525532, + 0.012349733151495457, + 0.006527500692754984, + 0.011357979848980904, + 0.025510141626000404, + 0.009401393122971058, + 0.016168734058737755, + 0.014767028391361237, + 0.010258576832711697, + 0.008522802963852882, + 0.008366788737475872, + 0.005293237045407295, + 0.011841210536658764, + 0.01571587659418583, + 0.009541850537061691, + 0.008893399499356747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 3 + ], + "selected_reward": -0.014222919009625912 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001826167106628 + ], + "opposite_logprob_tokens": [ + -2.6950182914733887 + ], + "logprob_reward": -0.07001826167106628, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001826167106628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06407266855239868, + 0.10885260254144669, + 0.10347616672515869, + 0.369190514087677, + 0.024066248908638954, + 0.010991131886839867, + 0.053339120000600815, + 0.03909200429916382, + 0.17662642896175385, + 0.050293102860450745 + ], + "credit_map": [ + 0.005713574588298798, + 0.008224686607718468, + 0.010910608805716038, + 0.005414961371570826, + 0.005956312175840139, + 0.011621370911598206, + 0.008220714516937733, + 0.008010444231331348, + 0.00655381241813302, + 0.03250407800078392, + 0.016387123614549637, + 0.011678841896355152, + 0.010031288489699364, + 0.006274125073105097, + 0.011235150508582592, + 0.014188201166689396, + 0.01270400919020176, + 0.01728125661611557, + 0.025502821430563927, + 0.010840801522135735, + 0.0067507256753742695, + 0.005631742998957634, + 0.013898473232984543, + 0.010866350494325161, + 0.04702422395348549, + 0.08436651527881622, + 0.06976746022701263, + 0.038166165351867676, + 0.03210045024752617, + 0.025147607550024986, + 0.04290572926402092, + 0.029712384566664696, + 0.001899828901514411, + 0.0022820206359028816, + 0.00525189284235239, + 0.0021045745816081762, + 0.002161435317248106, + 0.003832726739346981, + 0.0033329331781715155, + 0.0032008385751396418, + 0.0009632095461711287, + 0.000705018057487905, + 0.002310070674866438, + 0.0006815321394242346, + 0.0013866866938769817, + 0.0010989245492964983, + 0.0012989065144211054, + 0.0025467853993177414, + 0.008560745045542717, + 0.0029021725058555603, + 0.011461636982858181, + 0.0019854663405567408, + 0.0064142122864723206, + 0.012660685926675797, + 0.0045253015123307705, + 0.004828908480703831, + 0.0032973268534988165, + 0.004337605554610491, + 0.005768223199993372, + 0.003058768343180418, + 0.004481710027903318, + 0.004500885959714651, + 0.006346135400235653, + 0.007301356177777052, + 0.011059213429689407, + 0.020214775577187538, + 0.01447862759232521, + 0.00905410572886467, + 0.04780673608183861, + 0.016039244830608368, + 0.028306758031249046, + 0.029666973277926445, + 0.004448280669748783, + 0.008257326669991016, + 0.005050248466432095, + 0.004552368074655533, + 0.008866206742823124, + 0.006170241627842188, + 0.006757880561053753, + 0.0061905537731945515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 1 + ], + "selected_reward": -0.07001826167106628 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34865742921829224 + ], + "opposite_logprob_tokens": [ + -1.2236573696136475 + ], + "logprob_reward": -0.34865742921829224, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34865742921829224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17789514362812042, + 0.09214188903570175, + 0.14093975722789764, + 0.1973153054714203, + 0.04336013272404671, + 0.010343383066356182, + 0.06040370091795921, + 0.0456693097949028, + 0.19369131326675415, + 0.03824016451835632 + ], + "credit_map": [ + 0.015471263788640499, + 0.019749002531170845, + 0.05895759165287018, + 0.017392462119460106, + 0.012727496214210987, + 0.011059166863560677, + 0.02262677066028118, + 0.019911380484700203, + 0.0076685212552547455, + 0.009730817750096321, + 0.012458575889468193, + 0.0058603971265256405, + 0.008310657925903797, + 0.019927185028791428, + 0.01104632019996643, + 0.017139405012130737, + 0.01644825004041195, + 0.01354898326098919, + 0.043987710028886795, + 0.01367658656090498, + 0.008898453786969185, + 0.007749491836875677, + 0.020887427031993866, + 0.01574285514652729, + 0.021085606887936592, + 0.06124672666192055, + 0.01898724026978016, + 0.021866820752620697, + 0.018629491329193115, + 0.01411296147853136, + 0.01995875872671604, + 0.02142767608165741, + 0.002720338525250554, + 0.0025277293752878904, + 0.006830765400081873, + 0.001774042146280408, + 0.0035724628251045942, + 0.007288945838809013, + 0.005591887515038252, + 0.013053959235548973, + 0.001119007938541472, + 0.0007873850408941507, + 0.0013520194916054606, + 0.0005846458952873945, + 0.001275936490856111, + 0.0018262643134221435, + 0.0015765978023409843, + 0.0018215251620858908, + 0.007195138838142157, + 0.0034014431294053793, + 0.015591518953442574, + 0.0028958141338080168, + 0.007749132812023163, + 0.012274888344109058, + 0.005473604425787926, + 0.005822155624628067, + 0.005650020204484463, + 0.0036874799989163876, + 0.005863133352249861, + 0.003389358287677169, + 0.007427216973155737, + 0.005311209708452225, + 0.005795484874397516, + 0.008545405231416225, + 0.01128541212528944, + 0.02819264493882656, + 0.012550580315291882, + 0.019696170464158058, + 0.0476994663476944, + 0.02136930450797081, + 0.02804492972791195, + 0.024852782487869263, + 0.003847926389425993, + 0.004640151746571064, + 0.004001040477305651, + 0.003787270048633218, + 0.007310209330171347, + 0.005196152720600367, + 0.004787972196936607, + 0.0046694399788975716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.34865742921829224 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005254743155092001 + ], + "opposite_logprob_tokens": [ + -5.255254745483398 + ], + "logprob_reward": -0.005254743155092001, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005254743155092001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.332753449678421, + 0.13100136816501617, + 0.09923011064529419, + 0.08073485642671585, + 0.039939459413290024, + 0.01911572553217411, + 0.03446199372410774, + 0.10338065773248672, + 0.11816828697919846, + 0.04121403396129608 + ], + "credit_map": [ + 0.02094520814716816, + 0.09526810050010681, + 0.06462187319993973, + 0.03344923257827759, + 0.029372982680797577, + 0.016599247232079506, + 0.03275727853178978, + 0.039739541709423065, + 0.009641796350479126, + 0.021572737023234367, + 0.0406208410859108, + 0.012204855680465698, + 0.010378861799836159, + 0.007863478735089302, + 0.012899075634777546, + 0.015819735825061798, + 0.0056444257497787476, + 0.008496000431478024, + 0.022596916183829308, + 0.006320038810372353, + 0.006533855572342873, + 0.006868792697787285, + 0.013212192803621292, + 0.029557909816503525, + 0.0061856601387262344, + 0.015397397801280022, + 0.02227763831615448, + 0.005583298392593861, + 0.0056382473558187485, + 0.004792370367795229, + 0.009200929664075375, + 0.011659332551062107, + 0.0030208141542971134, + 0.0026142930146306753, + 0.00517751881852746, + 0.0015986630460247397, + 0.0033652624115347862, + 0.005608770065009594, + 0.005150537472218275, + 0.013403606601059437, + 0.0018074745312333107, + 0.001000428106635809, + 0.00428377091884613, + 0.0007813909905962646, + 0.004494087304919958, + 0.0019739256240427494, + 0.0023769924882799387, + 0.002397656673565507, + 0.004279873799532652, + 0.0017462809337303042, + 0.005924836732447147, + 0.0012401106068864465, + 0.003610101994127035, + 0.00807552132755518, + 0.0034427836071699858, + 0.0061424910090863705, + 0.007974873296916485, + 0.008219567127525806, + 0.020276956260204315, + 0.006564054172486067, + 0.01285377237945795, + 0.01333730760961771, + 0.014313258230686188, + 0.019840888679027557, + 0.007392727769911289, + 0.01613132283091545, + 0.00756555562838912, + 0.01294547226279974, + 0.026904184371232986, + 0.017569737508893013, + 0.01773436926305294, + 0.01192492712289095, + 0.0041422490030527115, + 0.006175345275551081, + 0.0037998377811163664, + 0.003436771221458912, + 0.005811627954244614, + 0.007788319140672684, + 0.005535529926419258, + 0.004524360876530409 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.005254743155092001 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, emotive close-up focuses on a wrinkled elderly man. The soft golden afterglow of the evening sun casts a sidelight on his face, sculpting each deep wrinkle on his skin like a dried-up riverbed, dotted with age spots. His slightly cloudy yet still deep eyes gaze unfocused beyond the camera, as if reminiscing about a distant past. His sparse, graying hair is gently ruffled by the breeze. His lips, cracked and pale, tremble almost imperceptibly before slowly opening. Finally, with a slump of his shoulders, he lets out a long, drawn-out sigh, a fleeting puff of white mist in the cool air. The scene is bathed in a warm, slightly melancholic hue, with the completely blurred outlines of the fields in the background, creating an atmosphere of profound contemplation and nostalgia, as if silently telling a long life story.", + "seed": 31888, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_11/index_390_seed_31888_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08469371069222689, + "vlm_logprob": -0.08469371069222689, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.08469371069222689, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013892264105379581 + ], + "opposite_logprob_tokens": [ + -5.51389217376709 + ], + "logprob_reward": -0.013892264105379581, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.013892264105379581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34147214889526367, + 0.07732398062944412, + 0.07371680438518524, + 0.20964780449867249, + 0.08885239064693451, + 0.06477513909339905, + 0.04292675480246544, + 0.021358568221330643, + 0.0399484746158123, + 0.03997795283794403 + ], + "credit_map": [ + 0.0243872981518507, + 0.06663866341114044, + 0.07833468168973923, + 0.06887608766555786, + 0.025560777634382248, + 0.01715630665421486, + 0.023866673931479454, + 0.036651674658060074, + 0.007176677230745554, + 0.013566208072006702, + 0.016743628308176994, + 0.012024841271340847, + 0.0065872082486748695, + 0.0056165121495723724, + 0.0048654526472091675, + 0.010743455030024052, + 0.006356546655297279, + 0.015392860397696495, + 0.024838684126734734, + 0.00916522741317749, + 0.0038047803100198507, + 0.0037739709950983524, + 0.003675104584544897, + 0.006709633395075798, + 0.017730021849274635, + 0.037659913301467896, + 0.07251223921775818, + 0.023800771683454514, + 0.01723417267203331, + 0.006470465566962957, + 0.007800713647156954, + 0.026439514011144638, + 0.00893621239811182, + 0.015020614489912987, + 0.017696481198072433, + 0.009397925809025764, + 0.012489695101976395, + 0.008025669492781162, + 0.00608329102396965, + 0.01120249554514885, + 0.006008863914757967, + 0.00884003285318613, + 0.010655800811946392, + 0.0073454356752336025, + 0.005388080142438412, + 0.006533432751893997, + 0.0051047890447080135, + 0.014898702502250671, + 0.002957924734801054, + 0.0068216267973184586, + 0.00586822722107172, + 0.006336634047329426, + 0.004435286857187748, + 0.0033676112070679665, + 0.0047588287852704525, + 0.008380615152418613, + 0.0013455500593408942, + 0.0033701963257044554, + 0.0031141522340476513, + 0.00290236109867692, + 0.0026858600322157145, + 0.0019167670980095863, + 0.002289809752255678, + 0.003733871504664421, + 0.004086745902895927, + 0.007033062633126974, + 0.004722610581666231, + 0.00518120126798749, + 0.004763348028063774, + 0.00554053345695138, + 0.002981940982863307, + 0.005639030598104, + 0.002960082609206438, + 0.006756101734936237, + 0.006374081131070852, + 0.005876609124243259, + 0.003946597222238779, + 0.0037803982850164175, + 0.004061009734869003, + 0.006223072297871113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.013892264105379581 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02061488851904869 + ], + "opposite_logprob_tokens": [ + -3.8956148624420166 + ], + "logprob_reward": -0.02061488851904869, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02061488851904869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2466411292552948, + 0.17661316692829132, + 0.04320991411805153, + 0.1306019276380539, + 0.1539837270975113, + 0.03901590034365654, + 0.1035819873213768, + 0.032314568758010864, + 0.03852524980902672, + 0.03551243618130684 + ], + "credit_map": [ + 0.01951655186712742, + 0.038196779787540436, + 0.04587341099977493, + 0.05073222890496254, + 0.022542282938957214, + 0.019116509705781937, + 0.022419724613428116, + 0.02824365347623825, + 0.01683998294174671, + 0.033358875662088394, + 0.04606589674949646, + 0.02245297282934189, + 0.01594354398548603, + 0.009383043274283409, + 0.010011948645114899, + 0.022556912153959274, + 0.0032965547870844603, + 0.008207223378121853, + 0.01436704769730568, + 0.004973114002496004, + 0.003673819126561284, + 0.0017930266913026571, + 0.002412453293800354, + 0.004486674442887306, + 0.008692932315170765, + 0.01903308369219303, + 0.03651728108525276, + 0.013524706475436687, + 0.01819540560245514, + 0.0066339666955173016, + 0.007084947545081377, + 0.020919613540172577, + 0.013858524151146412, + 0.020374678075313568, + 0.028354238718748093, + 0.02204018272459507, + 0.015481136739253998, + 0.0138023616746068, + 0.01901627518236637, + 0.02105633169412613, + 0.0034064806532114744, + 0.004981736186891794, + 0.0053451829589903355, + 0.004770161118358374, + 0.00419032946228981, + 0.004644711036235094, + 0.005184895824640989, + 0.00649240193888545, + 0.006788800470530987, + 0.018011095002293587, + 0.018590029329061508, + 0.016868088394403458, + 0.010378849692642689, + 0.006630039773881435, + 0.012564042583107948, + 0.013751042075455189, + 0.0029559966642409563, + 0.005602128338068724, + 0.005548128392547369, + 0.004837458021938801, + 0.0033503284212201834, + 0.002550350036472082, + 0.003292164532467723, + 0.004178013652563095, + 0.003754225792363286, + 0.006156491581350565, + 0.006124596111476421, + 0.005893545690923929, + 0.0035632774233818054, + 0.004340836778283119, + 0.0025316625833511353, + 0.006160612218081951, + 0.003275328315794468, + 0.006211312487721443, + 0.005111597944051027, + 0.007664965000003576, + 0.00378949916921556, + 0.0030376864597201347, + 0.002925220411270857, + 0.0034968273248523474 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.02061488851904869 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705424427986145 + ], + "opposite_logprob_tokens": [ + -2.127054214477539 + ], + "logprob_reward": -0.12705424427986145, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705424427986145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1994655430316925, + 0.046265486627817154, + 0.07577692717313766, + 0.40536731481552124, + 0.14470605552196503, + 0.020469509065151215, + 0.04798988625407219, + 0.026305848732590675, + 0.021770980209112167, + 0.011882456950843334 + ], + "credit_map": [ + 0.014433314092457294, + 0.035056885331869125, + 0.055255744606256485, + 0.029829541221261024, + 0.01171368919312954, + 0.008582575246691704, + 0.012733161449432373, + 0.03186064958572388, + 0.0033917580731213093, + 0.009770328179001808, + 0.012491313740611076, + 0.00649984460324049, + 0.00428624777123332, + 0.0027076564729213715, + 0.0028235677164047956, + 0.0042947703041136265, + 0.0064354813657701015, + 0.016778334975242615, + 0.02760588936507702, + 0.007921723648905754, + 0.0035473196767270565, + 0.0028725562151521444, + 0.0029897207859903574, + 0.007625897880643606, + 0.028378210961818695, + 0.06115764006972313, + 0.11443182826042175, + 0.04480613395571709, + 0.05728230997920036, + 0.018260207027196884, + 0.02139655314385891, + 0.05965443328022957, + 0.00851865392178297, + 0.022444849833846092, + 0.020944995805621147, + 0.02258264645934105, + 0.020369891077280045, + 0.0081979651004076, + 0.02504221722483635, + 0.016604842618107796, + 0.0015814397484064102, + 0.0032945245038717985, + 0.0031119827181100845, + 0.003004610538482666, + 0.00274915574118495, + 0.0015472631203010678, + 0.0018642558716237545, + 0.003316276241093874, + 0.0028523667715489864, + 0.00842489767819643, + 0.007917704991996288, + 0.00761670945212245, + 0.005379859823733568, + 0.0036502291914075613, + 0.005393927451223135, + 0.006754190661013126, + 0.0015360681572929025, + 0.0053563425317406654, + 0.0031615763437002897, + 0.004263600334525108, + 0.0030020612757653, + 0.0017132946522906423, + 0.0038327528163790703, + 0.0034401551820337772, + 0.0012751984177157283, + 0.005561316851526499, + 0.0033778867218643427, + 0.003329111495986581, + 0.0028113063890486956, + 0.001243968028575182, + 0.002313199220225215, + 0.0018589941319078207, + 0.001031692372635007, + 0.002278560306876898, + 0.0016807238571345806, + 0.0026756012812256813, + 0.0011995058739557862, + 0.0010616659419611096, + 0.0009102404583245516, + 0.0010444667423143983 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.12705424427986145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212362408638 + ], + "opposite_logprob_tokens": [ + -1.5021235942840576 + ], + "logprob_reward": -0.25212362408638, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212362408638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13090546429157257, + 0.04554218426346779, + 0.06947703659534454, + 0.3007184863090515, + 0.32377082109451294, + 0.04328693822026253, + 0.02865750342607498, + 0.03105495311319828, + 0.008615640923380852, + 0.01797093078494072 + ], + "credit_map": [ + 0.012278088368475437, + 0.025271127000451088, + 0.022382039576768875, + 0.027590885758399963, + 0.011517670005559921, + 0.010252385400235653, + 0.008033555932343006, + 0.01357971504330635, + 0.003113755024969578, + 0.009112083353102207, + 0.009860726073384285, + 0.00600430415943265, + 0.004161639139056206, + 0.0030779873486608267, + 0.004048073664307594, + 0.006163617596030235, + 0.006219335366040468, + 0.014049523510038853, + 0.02311202883720398, + 0.007662889547646046, + 0.004427015781402588, + 0.0032300923485308886, + 0.003993633668869734, + 0.006782523822039366, + 0.02280525676906109, + 0.05010339245200157, + 0.11505093425512314, + 0.028249947354197502, + 0.021972402930259705, + 0.008004378527402878, + 0.008414439857006073, + 0.046117741614580154, + 0.020034290850162506, + 0.04876828193664551, + 0.04493563249707222, + 0.051824528723955154, + 0.047420207411050797, + 0.019079195335507393, + 0.05995338410139084, + 0.031755343079566956, + 0.0030865375883877277, + 0.007382068317383528, + 0.007799838203936815, + 0.005731575656682253, + 0.003733749035745859, + 0.003279212163761258, + 0.0058514634147286415, + 0.006422499660402536, + 0.0018223754595965147, + 0.004908805713057518, + 0.004539999179542065, + 0.004443362820893526, + 0.003678889712318778, + 0.0021472678054124117, + 0.0031022923067212105, + 0.004014513920992613, + 0.0017672841204330325, + 0.0050662546418607235, + 0.004070752765983343, + 0.005673297680914402, + 0.0035269728396087885, + 0.001739479135721922, + 0.0056215981021523476, + 0.003589315339922905, + 0.000660606543533504, + 0.00147379154805094, + 0.0012026764452457428, + 0.0014615716645494103, + 0.0007717994158156216, + 0.0009596693562343717, + 0.0006937899743206799, + 0.0013917374890297651, + 0.0012273623142391443, + 0.0034599367063492537, + 0.0028093091677874327, + 0.003040950745344162, + 0.001792156370356679, + 0.0012119291350245476, + 0.0014203216414898634, + 0.0030089665669947863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.25212362408638 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a wrinkled elderly man with sparse, graying hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man's face illuminated by soft golden sidelight from the evening sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man's mouth tremble and then slowly open while his shoulders slump?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a puff of white mist visible as the man sighs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of completely blurred outlines of fields?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783532470464706 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783532470464706, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783532470464706, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3292680084705353, + 0.02529168874025345, + 0.014036284759640694, + 0.20631085336208344, + 0.297590970993042, + 0.04090788960456848, + 0.012300643138587475, + 0.03425157070159912, + 0.01851842924952507, + 0.021523648872971535 + ], + "credit_map": [ + 0.022504733875393867, + 0.0469374805688858, + 0.07493187487125397, + 0.038636740297079086, + 0.03464015945792198, + 0.021336112171411514, + 0.03389905020594597, + 0.05638188123703003, + 0.0025581824593245983, + 0.003733207704499364, + 0.0037104361690580845, + 0.0047171711921691895, + 0.0027790481690317392, + 0.0021836156956851482, + 0.002159373601898551, + 0.0034506558440625668, + 0.001666643307544291, + 0.002227206714451313, + 0.0026954077184200287, + 0.001817540731281042, + 0.0019680315162986517, + 0.0011254248674958944, + 0.001342092757113278, + 0.001193938427604735, + 0.01655794121325016, + 0.03466648608446121, + 0.0680701956152916, + 0.024263037368655205, + 0.024899618700146675, + 0.008233589120209217, + 0.00892047118395567, + 0.020699530839920044, + 0.01926480233669281, + 0.043644100427627563, + 0.04344645515084267, + 0.04446281120181084, + 0.03996250033378601, + 0.01908285729587078, + 0.052166491746902466, + 0.03556098788976669, + 0.0036077264230698347, + 0.004689280409365892, + 0.00492642167955637, + 0.006166249979287386, + 0.0036990209482610226, + 0.0042269728146493435, + 0.004731324966996908, + 0.008860895410180092, + 0.0008197820279747248, + 0.002083537168800831, + 0.00183953158557415, + 0.001673435908742249, + 0.0015251989243552089, + 0.0006512048421427608, + 0.0016733662923797965, + 0.002034588484093547, + 0.002053845440968871, + 0.005061380099505186, + 0.006151826586574316, + 0.0040881033055484295, + 0.0026593736838549376, + 0.0024887784384191036, + 0.0037742997519671917, + 0.007973967120051384, + 0.002000529784709215, + 0.0026621336583048105, + 0.0029166010208427906, + 0.0025832979008555412, + 0.0014991655480116606, + 0.003128267824649811, + 0.0012695157201960683, + 0.0024589188396930695, + 0.002092356327921152, + 0.003247409826144576, + 0.0045905462466180325, + 0.0026483924593776464, + 0.0015277238562703133, + 0.0023425398394465446, + 0.0013472181744873524, + 0.003727465867996216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.009783532470464706 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_tensor([642, 642])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_tensor([642, 642])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a3d2586c83567bf4cc03ca4963ffaa42230a86a6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_tensor([642, 642])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slowly rotating, orbiting medium shot brings us into the core control area of a futuristic drone racing arena. Clad in explosive competitive cinematography, it focuses on two competitors battling side by side. The female competitor on the left, sporting a sleek, dark purple ponytail, wears immersive FPV (first-person view) goggles that half-cover her face, their lenses reflecting a dazzling stream of data. She wears a tight, black, high-tech racing suit with a neon blue circuit pattern. She leans forward slightly, her hands steady and decisive as she grips a complex multi-stick drone controller, her thumbs fine-tuning the joysticks with a speed barely tangible to the naked eye. Beside her, her male teammate, with slightly damp short black hair and a furrowed brow, is absorbed in analyzing a translucent holographic data screen floating before them. His fingers swiftly and deftly swipe and tap across the illuminated interface, adjusting performance parameters and flight path diagrams in real time. Behind them, on the huge circular track structure, other drones turned into blurred colorful light bands and flew by at high speed. The whole scene was illuminated by the cool light emitted by their equipment and the neon lights of the track, casting dynamic light and shadow on their focused faces, creating a tacit cooperation and highly tense competitive atmosphere in the high-speed confrontation.", + "seed": 16433, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_642_seed_16433_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01147734122350812, + "vlm_logprob": -0.01147734122350812, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.01147734122350812, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01652541756629944 + ], + "opposite_logprob_tokens": [ + -4.6415252685546875 + ], + "logprob_reward": -0.01652541756629944, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01652541756629944, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22750800848007202, + 0.06114374101161957, + 0.0746825560927391, + 0.022234611213207245, + 0.04168606549501419, + 0.05994529277086258, + 0.03174322098493576, + 0.061649300158023834, + 0.3731299340724945, + 0.04627731442451477 + ], + "credit_map": [ + 0.055956557393074036, + 0.030548419803380966, + 0.02621154487133026, + 0.021221181377768517, + 0.023192429915070534, + 0.017630446702241898, + 0.01923961564898491, + 0.033507827669382095, + 0.005708773620426655, + 0.010333728976547718, + 0.007409863173961639, + 0.01266123540699482, + 0.005087807774543762, + 0.00513522420078516, + 0.0059709083288908005, + 0.008836197666823864, + 0.0114452438428998, + 0.01350250281393528, + 0.00979456678032875, + 0.008475398644804955, + 0.008349483832716942, + 0.005577662959694862, + 0.008073920384049416, + 0.009463776834309101, + 0.002415797207504511, + 0.0021117005962878466, + 0.0033094361424446106, + 0.0020455450285226107, + 0.002464428311213851, + 0.003023100085556507, + 0.0030655809678137302, + 0.0037990219425410032, + 0.0033892334904521704, + 0.004186984617263079, + 0.004969588480889797, + 0.0037721535190939903, + 0.005493054632097483, + 0.007807030342519283, + 0.005949543323367834, + 0.006118475459516048, + 0.007097356487065554, + 0.014427323825657368, + 0.004712338093668222, + 0.0041295443661510944, + 0.004922803957015276, + 0.004611624870449305, + 0.008910726755857468, + 0.011133572086691856, + 0.0026778080500662327, + 0.004061288200318813, + 0.005623079836368561, + 0.004174729809165001, + 0.003913071006536484, + 0.00532847223803401, + 0.0026579920668154955, + 0.003306780708953738, + 0.0051913196220994, + 0.003940468654036522, + 0.0035061093512922525, + 0.013825654052197933, + 0.004325220827013254, + 0.023077214136719704, + 0.002291641663759947, + 0.005491669289767742, + 0.022064678370952606, + 0.024642761796712875, + 0.023061560466885567, + 0.1200367659330368, + 0.01877216249704361, + 0.029649803414940834, + 0.019585393369197845, + 0.11531680077314377, + 0.004411743488162756, + 0.0027341304812580347, + 0.0023368503898382187, + 0.002354659605771303, + 0.0038331609684973955, + 0.014608189463615417, + 0.003009240375831723, + 0.012989338487386703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.01652541756629944 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544536031782627 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544536031782627, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544536031782627, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25675153732299805, + 0.04751879721879959, + 0.024358781054615974, + 0.022697633132338524, + 0.03132347762584686, + 0.020912427455186844, + 0.04355637729167938, + 0.19632582366466522, + 0.30848926305770874, + 0.0480659082531929 + ], + "credit_map": [ + 0.05883697047829628, + 0.03372449055314064, + 0.0313556045293808, + 0.023449787870049477, + 0.023681988939642906, + 0.021380065008997917, + 0.02634318731725216, + 0.03797942027449608, + 0.004006481263786554, + 0.00289785279892385, + 0.005881140008568764, + 0.01259281300008297, + 0.004225750919431448, + 0.002367062261328101, + 0.004695134703069925, + 0.010852557606995106, + 0.002283445093780756, + 0.0019475342705845833, + 0.002591098891571164, + 0.002640161896124482, + 0.005103493109345436, + 0.004566033370792866, + 0.002703669248148799, + 0.0025233435444533825, + 0.0026211903896182775, + 0.0017706031212583184, + 0.003965356387197971, + 0.002366017084568739, + 0.004415508359670639, + 0.002646863926202059, + 0.00241492735221982, + 0.0024971652310341597, + 0.0027548696380108595, + 0.0029851289000362158, + 0.0033219940960407257, + 0.003163571236655116, + 0.005377914756536484, + 0.0031713133212178946, + 0.00515044154599309, + 0.0053982422687113285, + 0.0027548896614462137, + 0.005359582602977753, + 0.0034333246294409037, + 0.0020544270519167185, + 0.0028039889875799417, + 0.0013200730318203568, + 0.0015509913209825754, + 0.0016351480735465884, + 0.004963227082043886, + 0.006057182792574167, + 0.007875864394009113, + 0.005371487233787775, + 0.007277380209416151, + 0.0038527012802660465, + 0.0034510313998907804, + 0.0047074975445866585, + 0.013181326910853386, + 0.009640656411647797, + 0.010777590796351433, + 0.050572771579027176, + 0.010472210124135017, + 0.07856731116771698, + 0.007183461915701628, + 0.015930475667119026, + 0.026608027517795563, + 0.023348240181803703, + 0.030958227813243866, + 0.08083012700080872, + 0.023793978616595268, + 0.029747413471341133, + 0.017475655302405357, + 0.07572758942842484, + 0.004057408310472965, + 0.0033066794276237488, + 0.002338831778615713, + 0.0030731179285794497, + 0.004500192124396563, + 0.014324044808745384, + 0.0029531344771385193, + 0.013512494042515755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.012544536031782627 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068249940872192 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068249940872192, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068249940872192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23809058964252472, + 0.03850124031305313, + 0.023286333307623863, + 0.018994741141796112, + 0.1788196712732315, + 0.03951262682676315, + 0.03430232033133507, + 0.161632239818573, + 0.23637887835502625, + 0.030481353402137756 + ], + "credit_map": [ + 0.05461224541068077, + 0.03123471513390541, + 0.028362583369016647, + 0.022646887227892876, + 0.022452013567090034, + 0.021330878138542175, + 0.023675160482525826, + 0.033776093274354935, + 0.004066281020641327, + 0.003819277510046959, + 0.006198268849402666, + 0.010085797868669033, + 0.004050394520163536, + 0.0032599905971437693, + 0.0029664267785847187, + 0.0040548029355704784, + 0.002424505539238453, + 0.002389014232903719, + 0.0025294418446719646, + 0.0030763535760343075, + 0.003698911750689149, + 0.002749226288869977, + 0.003775512333959341, + 0.0026433661114424467, + 0.0021599242463707924, + 0.001643631374463439, + 0.0020073596388101578, + 0.0015160957118496299, + 0.002241073176264763, + 0.0036370486486703157, + 0.0021065312903374434, + 0.0036830741446465254, + 0.00954111386090517, + 0.015844203531742096, + 0.02062240242958069, + 0.01837804540991783, + 0.0272232573479414, + 0.02134333923459053, + 0.034928496927022934, + 0.03093879483640194, + 0.005178343039005995, + 0.012440156191587448, + 0.0035686383489519358, + 0.004359372425824404, + 0.004230311140418053, + 0.002477179281413555, + 0.004315702710300684, + 0.0029429213609546423, + 0.003386331256479025, + 0.0037042165640741587, + 0.00485722441226244, + 0.0041187782771885395, + 0.005119489971548319, + 0.006803667172789574, + 0.0025527742691338062, + 0.003759834449738264, + 0.01077286060899496, + 0.008424980565905571, + 0.0093580586835742, + 0.04200969263911247, + 0.009086723439395428, + 0.06331950426101685, + 0.005950357299298048, + 0.012710053473711014, + 0.014528604224324226, + 0.016760293394327164, + 0.01532235462218523, + 0.07280649989843369, + 0.012725314125418663, + 0.020132938399910927, + 0.013203970156610012, + 0.0708988755941391, + 0.001905514276586473, + 0.0019641495309770107, + 0.0012980718165636063, + 0.0016367788193747401, + 0.0022181121166795492, + 0.010985660366714, + 0.0016624443233013153, + 0.008810621686279774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.016068249940872192 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046300822868943214 + ], + "opposite_logprob_tokens": [ + -5.379630088806152 + ], + "logprob_reward": -0.0046300822868943214, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046300822868943214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24874216318130493, + 0.024675903841853142, + 0.01504075713455677, + 0.08674760162830353, + 0.08401837944984436, + 0.07266145199537277, + 0.019389357417821884, + 0.053181134164333344, + 0.3470205068588257, + 0.04852275550365448 + ], + "credit_map": [ + 0.055805690586566925, + 0.031242763623595238, + 0.03118370845913887, + 0.024046799167990685, + 0.02355271391570568, + 0.022790949791669846, + 0.02591659501194954, + 0.034202951937913895, + 0.002311509335413575, + 0.0038677644915878773, + 0.002848373493179679, + 0.004388485103845596, + 0.002367316512390971, + 0.0022983804810792208, + 0.0028707319870591164, + 0.003723343601450324, + 0.001701013301499188, + 0.0013785494957119226, + 0.0014046423602849245, + 0.0012626124080270529, + 0.0024882489815354347, + 0.0026272658724337816, + 0.002156710484996438, + 0.002021714113652706, + 0.008977902121841908, + 0.005874298512935638, + 0.007775587495416403, + 0.0075715286657214165, + 0.023204393684864044, + 0.011660757474601269, + 0.00808961596339941, + 0.0135935228317976, + 0.005402532871812582, + 0.010482227429747581, + 0.008626103401184082, + 0.009974442422389984, + 0.008609569631516933, + 0.008787302300333977, + 0.018050936982035637, + 0.01408525463193655, + 0.013811972923576832, + 0.02431458793580532, + 0.007837392389774323, + 0.004566419869661331, + 0.008510081097483635, + 0.0038036610931158066, + 0.005563016049563885, + 0.004254322033375502, + 0.003450048388913274, + 0.0021887565962970257, + 0.00165198789909482, + 0.0019403095357120037, + 0.0022989327553659678, + 0.004884851165115833, + 0.0008497221861034632, + 0.002124746795743704, + 0.003906308673322201, + 0.002947957022115588, + 0.0026268227957189083, + 0.013392120599746704, + 0.002898340579122305, + 0.020974192768335342, + 0.0018728430150076747, + 0.004562548827379942, + 0.019477015361189842, + 0.02139970101416111, + 0.01947934553027153, + 0.11394961178302765, + 0.016382692381739616, + 0.0270331222563982, + 0.018307700753211975, + 0.11099134385585785, + 0.006060968618839979, + 0.00545212347060442, + 0.0037493384443223476, + 0.005332111846655607, + 0.007509392686188221, + 0.008494389243423939, + 0.004240165930241346, + 0.007684266194701195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.0046300822868943214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076184202916920185 + ], + "opposite_logprob_tokens": [ + -4.882618427276611 + ], + "logprob_reward": -0.0076184202916920185, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076184202916920185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15262706577777863, + 0.0619388222694397, + 0.018659744411706924, + 0.059633687138557434, + 0.04463884234428406, + 0.4571397602558136, + 0.021361060440540314, + 0.09944242238998413, + 0.0630500540137291, + 0.021508513018488884 + ], + "credit_map": [ + 0.032852303236722946, + 0.01979578100144863, + 0.01623743213713169, + 0.016257118433713913, + 0.014949603006243706, + 0.013768547214567661, + 0.015075484290719032, + 0.023690801113843918, + 0.005531673319637775, + 0.007752959616482258, + 0.005904644727706909, + 0.01650194078683853, + 0.004079956561326981, + 0.0034545394591987133, + 0.005731722339987755, + 0.012981387786567211, + 0.002493745880201459, + 0.0024539653677493334, + 0.0026474352926015854, + 0.0020331323612481356, + 0.002599452156573534, + 0.001922090188600123, + 0.0017678258009254932, + 0.002742097480222583, + 0.0073965792544186115, + 0.0049767447635531425, + 0.010885308496654034, + 0.0057505168952047825, + 0.008503920398652554, + 0.007751191966235638, + 0.00597247201949358, + 0.00839694868773222, + 0.004040662199258804, + 0.005148360971361399, + 0.004673193208873272, + 0.004420184064656496, + 0.006665635854005814, + 0.005012650042772293, + 0.006962047424167395, + 0.007716109976172447, + 0.08711138367652893, + 0.17564567923545837, + 0.043111834675073624, + 0.03191152215003967, + 0.05692237243056297, + 0.01885245554149151, + 0.021925069391727448, + 0.02165944315493107, + 0.0032489716541022062, + 0.0016526603139936924, + 0.0016933631850406528, + 0.00206874031573534, + 0.0023177668917924166, + 0.006244951859116554, + 0.0010997324716299772, + 0.0030348743312060833, + 0.006659433711320162, + 0.005541972350329161, + 0.00513295317068696, + 0.02534271776676178, + 0.004830858204513788, + 0.03996920585632324, + 0.003544582286849618, + 0.008420695550739765, + 0.009523965418338776, + 0.007141515612602234, + 0.00965188443660736, + 0.0074365283362567425, + 0.007566200103610754, + 0.010069487616419792, + 0.005143078044056892, + 0.006517397239804268, + 0.0023936869110912085, + 0.0027111421804875135, + 0.0020343512296676636, + 0.0023521375842392445, + 0.003101866692304611, + 0.003575703129172325, + 0.0020941204857081175, + 0.0032455064356327057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.0076184202916920185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating, orbiting medium shot brings us into the core control area of a futuristic drone racing arena. Clad in explosive competitive cinematography, it focuses on two competitors battling side by side. The female competitor on the left, sporting a sleek, dark purple ponytail, wears immersive FPV (first-person view) goggles that half-cover her face, their lenses reflecting a dazzling stream of data. She wears a tight, black, high-tech racing suit with a neon blue circuit pattern. She leans forward slightly, her hands steady and decisive as she grips a complex multi-stick drone controller, her thumbs fine-tuning the joysticks with a speed barely tangible to the naked eye. Beside her, her male teammate, with slightly damp short black hair and a furrowed brow, is absorbed in analyzing a translucent holographic data screen floating before them. His fingers swiftly and deftly swipe and tap across the illuminated interface, adjusting performance parameters and flight path diagrams in real time. Behind them, on the huge circular track structure, other drones turned into blurred colorful light bands and flew by at high speed. The whole scene was illuminated by the cool light emitted by their equipment and the neon lights of the track, casting dynamic light and shadow on their focused faces, creating a tacit cooperation and highly tense competitive atmosphere in the high-speed confrontation.", + "seed": 16434, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_642_seed_16434_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01601222436875105, + "vlm_logprob": -0.01601222436875105, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.01601222436875105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02183598093688488 + ], + "opposite_logprob_tokens": [ + -4.271835803985596 + ], + "logprob_reward": -0.02183598093688488, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.02183598093688488, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17906488478183746, + 0.10713273286819458, + 0.07962335646152496, + 0.07237555831670761, + 0.1624009609222412, + 0.07600197941064835, + 0.07272864133119583, + 0.08009197562932968, + 0.09511775523424149, + 0.07546214759349823 + ], + "credit_map": [ + 0.01776307076215744, + 0.02588212676346302, + 0.015854453667998314, + 0.02080385573208332, + 0.03187964856624603, + 0.02439073473215103, + 0.019439280033111572, + 0.023051729425787926, + 0.01206528302282095, + 0.013681461103260517, + 0.016921961680054665, + 0.009549008682370186, + 0.012336981482803822, + 0.021062660962343216, + 0.012315226718783379, + 0.009200146421790123, + 0.010769298300147057, + 0.009477508254349232, + 0.014912968501448631, + 0.009646493010222912, + 0.009091605432331562, + 0.011373436078429222, + 0.007209631614387035, + 0.007142412476241589, + 0.009742585010826588, + 0.013274633325636387, + 0.009511192329227924, + 0.0067373584024608135, + 0.007350102532655001, + 0.009842555969953537, + 0.009642021730542183, + 0.00627511041238904, + 0.014241463504731655, + 0.010677885264158249, + 0.01295468769967556, + 0.010225892066955566, + 0.010343302972614765, + 0.04094986990094185, + 0.05233153700828552, + 0.010676315985620022, + 0.008360163308680058, + 0.011517028324306011, + 0.01191141176968813, + 0.0069035752676427364, + 0.009237030521035194, + 0.01158332359045744, + 0.010272322222590446, + 0.006217123009264469, + 0.006807177793234587, + 0.01135586854070425, + 0.009684435091912746, + 0.008379126898944378, + 0.013196580111980438, + 0.006623817142099142, + 0.01073514949530363, + 0.005946489050984383, + 0.009686208330094814, + 0.01151823066174984, + 0.009105064906179905, + 0.010779810138046741, + 0.008324488997459412, + 0.013977210968732834, + 0.010398630052804947, + 0.006302330642938614, + 0.015988215804100037, + 0.011042200028896332, + 0.009007101878523827, + 0.012311956845223904, + 0.007826531305909157, + 0.01535426452755928, + 0.011374717578291893, + 0.012212764471769333, + 0.011668997816741467, + 0.00860095676034689, + 0.005536715034395456, + 0.008672608993947506, + 0.011463279835879803, + 0.009843307547271252, + 0.007545736152678728, + 0.012130549177527428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.02183598093688488 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588263869285583 + ], + "opposite_logprob_tokens": [ + -3.8955881595611572 + ], + "logprob_reward": -0.020588263869285583, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588263869285583, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11253903061151505, + 0.15080960094928741, + 0.07410314679145813, + 0.18095745146274567, + 0.19054661691188812, + 0.04043935611844063, + 0.07468865066766739, + 0.07045213878154755, + 0.054833658039569855, + 0.05063029006123543 + ], + "credit_map": [ + 0.01271242555230856, + 0.01180021371692419, + 0.02089002914726734, + 0.010350730270147324, + 0.016030296683311462, + 0.01472585927695036, + 0.012545529752969742, + 0.013483943417668343, + 0.013718917965888977, + 0.014377102255821228, + 0.03278959169983864, + 0.010549511760473251, + 0.012818748131394386, + 0.04039302095770836, + 0.015094063244760036, + 0.011068640276789665, + 0.00836501270532608, + 0.009909110143780708, + 0.011020245961844921, + 0.007944625802338123, + 0.008900625631213188, + 0.013713888823986053, + 0.007456129416823387, + 0.00679350970312953, + 0.02917379140853882, + 0.02221526950597763, + 0.026204170659184456, + 0.0186481736600399, + 0.027943888679146767, + 0.016997888684272766, + 0.02283192239701748, + 0.016942352056503296, + 0.014950216747820377, + 0.010299707762897015, + 0.01498686708509922, + 0.01124630682170391, + 0.0150382611900568, + 0.0434446707367897, + 0.06991885602474213, + 0.010661746375262737, + 0.004549427889287472, + 0.0064174518920481205, + 0.005085913930088282, + 0.004236313048750162, + 0.00408034585416317, + 0.006429852917790413, + 0.005155032966285944, + 0.00448501855134964, + 0.00715671107172966, + 0.012097587808966637, + 0.0072648730129003525, + 0.00911991111934185, + 0.01476158294826746, + 0.009412537328898907, + 0.01076254528015852, + 0.004112903960049152, + 0.01005942840129137, + 0.014497692696750164, + 0.005559534765779972, + 0.008896195329725742, + 0.009511249139904976, + 0.007346601691097021, + 0.010610746219754219, + 0.003970691002905369, + 0.005122468341141939, + 0.006849302910268307, + 0.007024095393717289, + 0.008568470366299152, + 0.006448924075812101, + 0.0059768990613520145, + 0.009341238997876644, + 0.00550225842744112, + 0.006950902286916971, + 0.005038914270699024, + 0.003359791124239564, + 0.0050487094558775425, + 0.00880789291113615, + 0.008466264232993126, + 0.004483882803469896, + 0.008473933674395084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.020588263869285583 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288307711482048 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288307711482048, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288307711482048, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23031115531921387, + 0.13652758300304413, + 0.08592040091753006, + 0.07225292176008224, + 0.11477470397949219, + 0.10862933844327927, + 0.03833089396357536, + 0.057812463492155075, + 0.05790881812572479, + 0.09753169119358063 + ], + "credit_map": [ + 0.022580726072192192, + 0.035098880529403687, + 0.0347244031727314, + 0.025787051767110825, + 0.03674225136637688, + 0.02658781222999096, + 0.021463653072714806, + 0.027326388284564018, + 0.012082671746611595, + 0.017950324341654778, + 0.01987386867403984, + 0.023054268211126328, + 0.014383804984390736, + 0.015234080143272877, + 0.019229158759117126, + 0.014719413593411446, + 0.009859755635261536, + 0.00830414891242981, + 0.014870036393404007, + 0.008174239657819271, + 0.01227612141519785, + 0.010902639478445053, + 0.013365470804274082, + 0.00816799234598875, + 0.013371603563427925, + 0.009841141290962696, + 0.0094485804438591, + 0.007421508431434631, + 0.009276056662201881, + 0.007230632938444614, + 0.009276993572711945, + 0.0063864062540233135, + 0.01292615756392479, + 0.01013550627976656, + 0.01565517857670784, + 0.01336462702602148, + 0.022708674892783165, + 0.017403803765773773, + 0.012495525181293488, + 0.010085229761898518, + 0.01155852247029543, + 0.01783149689435959, + 0.014682098291814327, + 0.011685064993798733, + 0.012155025266110897, + 0.017120443284511566, + 0.010739478282630444, + 0.012857208028435707, + 0.002854144899174571, + 0.00604590168222785, + 0.0037463614717125893, + 0.006076149642467499, + 0.007657493464648724, + 0.004747151397168636, + 0.004282794892787933, + 0.002920895116403699, + 0.006411747541278601, + 0.010877328924834728, + 0.0047781565226614475, + 0.006892144680023193, + 0.006850711535662413, + 0.011909396387636662, + 0.007192044518887997, + 0.0029009347781538963, + 0.006086238194257021, + 0.006619629450142384, + 0.008383278734982014, + 0.009140112437307835, + 0.004996531642973423, + 0.00821373239159584, + 0.005878441501408815, + 0.008590852841734886, + 0.008912832476198673, + 0.008244844153523445, + 0.007204368244856596, + 0.009548830799758434, + 0.02120289020240307, + 0.020282389596104622, + 0.008239556103944778, + 0.01389598660171032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.023288307711482048 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067284926772117615 + ], + "opposite_logprob_tokens": [ + -5.006728649139404 + ], + "logprob_reward": -0.0067284926772117615, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067284926772117615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15452389419078827, + 0.22001869976520538, + 0.09238512814044952, + 0.19798576831817627, + 0.10445482283830643, + 0.04069436341524124, + 0.04706761986017227, + 0.0313635878264904, + 0.023251131176948547, + 0.08825497329235077 + ], + "credit_map": [ + 0.01622731238603592, + 0.026617515832185745, + 0.020054640248417854, + 0.01826203428208828, + 0.022370388731360435, + 0.020801737904548645, + 0.013835114426910877, + 0.016355153173208237, + 0.021013902500271797, + 0.02253076620399952, + 0.054648470133543015, + 0.016841618344187737, + 0.01427800115197897, + 0.05819722265005112, + 0.016501588746905327, + 0.016007130965590477, + 0.011502211913466454, + 0.00800089817494154, + 0.01553679071366787, + 0.009098486974835396, + 0.01520149689167738, + 0.0117651941254735, + 0.012305115349590778, + 0.00897493027150631, + 0.02433878369629383, + 0.0325096994638443, + 0.026914818212389946, + 0.019988300278782845, + 0.025264009833335876, + 0.024484891444444656, + 0.02476871944963932, + 0.019716549664735794, + 0.008016256615519524, + 0.005552640184760094, + 0.008758440613746643, + 0.005677790381014347, + 0.007743686903268099, + 0.02757737599313259, + 0.03466005250811577, + 0.006468576844781637, + 0.004232219886034727, + 0.007205565460026264, + 0.00567292608320713, + 0.005158009938895702, + 0.0035942031536251307, + 0.0034395887050777674, + 0.0073427301831543446, + 0.004049118608236313, + 0.004028176888823509, + 0.008557737804949284, + 0.004434765782207251, + 0.005293981172144413, + 0.009562342427670956, + 0.00582974124699831, + 0.005603265482932329, + 0.0037576095201075077, + 0.004487555008381605, + 0.0037441651802510023, + 0.0025732978247106075, + 0.0037423360627144575, + 0.003681376576423645, + 0.007926310412585735, + 0.0032435280736535788, + 0.0019650200847536325, + 0.00436050770804286, + 0.002601951127871871, + 0.002574547426775098, + 0.002850244753062725, + 0.0026450117584317923, + 0.004219910129904747, + 0.00225005391985178, + 0.0017489040037617087, + 0.007584827486425638, + 0.006610941607505083, + 0.006241167429834604, + 0.00788468960672617, + 0.019833801314234734, + 0.01863921619951725, + 0.006829659920185804, + 0.014630667865276337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.0067284926772117615 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007620076648890972 + ], + "opposite_logprob_tokens": [ + -4.882619857788086 + ], + "logprob_reward": -0.007620076648890972, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007620076648890972, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24892252683639526, + 0.35381779074668884, + 0.07365147769451141, + 0.06493367999792099, + 0.031047319993376732, + 0.032304584980010986, + 0.03498358279466629, + 0.032166171818971634, + 0.05693032965064049, + 0.07124251127243042 + ], + "credit_map": [ + 0.02469138614833355, + 0.03360661491751671, + 0.03261571750044823, + 0.03086131252348423, + 0.03146221861243248, + 0.028244152665138245, + 0.03408416733145714, + 0.03335696458816528, + 0.04099288582801819, + 0.019687622785568237, + 0.10540550202131271, + 0.017957596108317375, + 0.01433734130114317, + 0.11614418774843216, + 0.014298212714493275, + 0.024994462728500366, + 0.007087537087500095, + 0.010962366126477718, + 0.01152874156832695, + 0.00926990807056427, + 0.008467063307762146, + 0.011817628517746925, + 0.008539742790162563, + 0.005978494882583618, + 0.008838783018290997, + 0.011677210219204426, + 0.008854870684444904, + 0.008142479695379734, + 0.007416306994855404, + 0.0077533479779958725, + 0.006568567361682653, + 0.005682114977389574, + 0.003652040148153901, + 0.002561399480327964, + 0.003392915241420269, + 0.0032131881453096867, + 0.003795841708779335, + 0.004653158597648144, + 0.008093458600342274, + 0.0016853167908266187, + 0.002807165263220668, + 0.0041596596129238605, + 0.0041576167568564415, + 0.0033723448868840933, + 0.0033700906205922365, + 0.0034758311230689287, + 0.008433923125267029, + 0.002527952892705798, + 0.002896735444664955, + 0.004899699240922928, + 0.0036974074319005013, + 0.004103787709027529, + 0.006181224714964628, + 0.005089153535664082, + 0.00477048521861434, + 0.003345088567584753, + 0.004365937784314156, + 0.003836588002741337, + 0.003859573509544134, + 0.004107688087970018, + 0.0034533399157226086, + 0.007637687027454376, + 0.0027459345292299986, + 0.0021594238933175802, + 0.0069800629280507565, + 0.007095505483448505, + 0.007475059013813734, + 0.01023942418396473, + 0.005823034327477217, + 0.006196458823978901, + 0.008862530812621117, + 0.00425825547426939, + 0.009569532237946987, + 0.006845179479569197, + 0.006408564746379852, + 0.008271395228803158, + 0.007526989094913006, + 0.010059436783194542, + 0.006740430369973183, + 0.015820980072021484 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007620076648890972 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating, orbiting medium shot brings us into the core control area of a futuristic drone racing arena. Clad in explosive competitive cinematography, it focuses on two competitors battling side by side. The female competitor on the left, sporting a sleek, dark purple ponytail, wears immersive FPV (first-person view) goggles that half-cover her face, their lenses reflecting a dazzling stream of data. She wears a tight, black, high-tech racing suit with a neon blue circuit pattern. She leans forward slightly, her hands steady and decisive as she grips a complex multi-stick drone controller, her thumbs fine-tuning the joysticks with a speed barely tangible to the naked eye. Beside her, her male teammate, with slightly damp short black hair and a furrowed brow, is absorbed in analyzing a translucent holographic data screen floating before them. His fingers swiftly and deftly swipe and tap across the illuminated interface, adjusting performance parameters and flight path diagrams in real time. Behind them, on the huge circular track structure, other drones turned into blurred colorful light bands and flew by at high speed. The whole scene was illuminated by the cool light emitted by their equipment and the neon lights of the track, casting dynamic light and shadow on their focused faces, creating a tacit cooperation and highly tense competitive atmosphere in the high-speed confrontation.", + "seed": 16435, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_642_seed_16435_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012448145542293787, + "vlm_logprob": -0.012448145542293787, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.012448145542293787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01527739129960537 + ], + "opposite_logprob_tokens": [ + -4.89027738571167 + ], + "logprob_reward": -0.01527739129960537, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01527739129960537, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49959796667099, + 0.08184065669775009, + 0.05613217502832413, + 0.02878807671368122, + 0.03803158923983574, + 0.06369571387767792, + 0.059327565133571625, + 0.04476124793291092, + 0.043678008019924164, + 0.08414698392152786 + ], + "credit_map": [ + 0.025750979781150818, + 0.01904326304793358, + 0.036444734781980515, + 0.014114897698163986, + 0.03535294532775879, + 0.02281855046749115, + 0.19978590309619904, + 0.14628669619560242, + 0.009679335169494152, + 0.008308778516948223, + 0.010026915930211544, + 0.007976273074746132, + 0.011854784563183784, + 0.009062659926712513, + 0.01053325179964304, + 0.014398660510778427, + 0.0051316432654857635, + 0.0053676762618124485, + 0.006927133072167635, + 0.003342877607792616, + 0.006629875395447016, + 0.006848584860563278, + 0.016359694302082062, + 0.005524694453924894, + 0.0025679469108581543, + 0.004309955984354019, + 0.003480230923742056, + 0.002493887674063444, + 0.0048825680278241634, + 0.004556253552436829, + 0.0033103334717452526, + 0.003186902031302452, + 0.003576223272830248, + 0.009299380704760551, + 0.003996171522885561, + 0.00366103183478117, + 0.003786883782595396, + 0.005463569890707731, + 0.003931179642677307, + 0.004317150451242924, + 0.006584499962627888, + 0.0048530870117247105, + 0.007902424782514572, + 0.008233466185629368, + 0.008749251253902912, + 0.015137862414121628, + 0.005192650947719812, + 0.007042479235678911, + 0.006698470562696457, + 0.009520214051008224, + 0.005599500611424446, + 0.007065465208142996, + 0.007669572252780199, + 0.0109838442876935, + 0.0062267715111374855, + 0.005563733633607626, + 0.005062933545559645, + 0.005768103525042534, + 0.0050478409975767136, + 0.005000531207770109, + 0.007994823157787323, + 0.006476162001490593, + 0.00506211444735527, + 0.004348740912973881, + 0.005026035942137241, + 0.00878734141588211, + 0.006813012063503265, + 0.005827341228723526, + 0.005227431654930115, + 0.004726906772702932, + 0.0030688547994941473, + 0.004201087635010481, + 0.00618781428784132, + 0.00957965012639761, + 0.012304853647947311, + 0.006295708939433098, + 0.01053544320166111, + 0.0077220480889081955, + 0.026037493720650673, + 0.005483983084559441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.01527739129960537 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819467917084694 + ], + "opposite_logprob_tokens": [ + -4.018194675445557 + ], + "logprob_reward": -0.01819467917084694, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819467917084694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14106348156929016, + 0.17033103108406067, + 0.09864544123411179, + 0.06727483123540878, + 0.059440720826387405, + 0.08672462403774261, + 0.040062788873910904, + 0.15722572803497314, + 0.07620885223150253, + 0.10302254557609558 + ], + "credit_map": [ + 0.009269489906728268, + 0.007269959431141615, + 0.013082610443234444, + 0.005473007448017597, + 0.011217559687793255, + 0.010755854658782482, + 0.038589801639318466, + 0.04540519416332245, + 0.0193063635379076, + 0.019348526373505592, + 0.020533470436930656, + 0.017246777191758156, + 0.022774551063776016, + 0.014760555699467659, + 0.02204792946577072, + 0.03431284800171852, + 0.006692442111670971, + 0.006254279054701328, + 0.014560282230377197, + 0.005746576469391584, + 0.01144031248986721, + 0.010627918876707554, + 0.02038154937326908, + 0.02294207736849785, + 0.005070728249847889, + 0.011848594062030315, + 0.007331186905503273, + 0.00962368119508028, + 0.009239035658538342, + 0.009571212343871593, + 0.008344701491296291, + 0.0062456936575472355, + 0.007474231068044901, + 0.012180672027170658, + 0.006724557373672724, + 0.005198214668780565, + 0.010824030265212059, + 0.00709084328263998, + 0.005097544752061367, + 0.004850623197853565, + 0.009393710643053055, + 0.00823265127837658, + 0.011511540040373802, + 0.012049165554344654, + 0.01491724792867899, + 0.015587744303047657, + 0.006486987229436636, + 0.008545567281544209, + 0.0032338504679501057, + 0.004723165649920702, + 0.005125454626977444, + 0.006416975986212492, + 0.005045266356319189, + 0.005975616164505482, + 0.0044092885218560696, + 0.005133167374879122, + 0.024779200553894043, + 0.026052339002490044, + 0.022303303703665733, + 0.02042102813720703, + 0.021600091829895973, + 0.015200512483716011, + 0.013661521486938, + 0.013207722455263138, + 0.00855899229645729, + 0.013422840274870396, + 0.014666988514363766, + 0.010103042237460613, + 0.008855543099343777, + 0.00784676056355238, + 0.006408358458429575, + 0.006346318870782852, + 0.007740167900919914, + 0.014243709854781628, + 0.012801174074411392, + 0.015558778308331966, + 0.015859421342611313, + 0.014606219716370106, + 0.012795716524124146, + 0.009417352266609669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.01819467917084694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419647503644228 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.01419647503644228, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419647503644228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29419630765914917, + 0.061346933245658875, + 0.06386116147041321, + 0.045385535806417465, + 0.07307153195142746, + 0.05621954798698425, + 0.18895363807678223, + 0.04505998268723488, + 0.046378348022699356, + 0.1255270093679428 + ], + "credit_map": [ + 0.0195609200745821, + 0.013323687948286533, + 0.023955434560775757, + 0.01053514052182436, + 0.022445589303970337, + 0.012837214395403862, + 0.1171441376209259, + 0.07439418137073517, + 0.004553290084004402, + 0.006977716460824013, + 0.008853115141391754, + 0.005608283448964357, + 0.013737545348703861, + 0.00470645260065794, + 0.008204304613173008, + 0.008706226944923401, + 0.003424118272960186, + 0.004023950081318617, + 0.005179231986403465, + 0.0027254759334027767, + 0.004591609816998243, + 0.007048351690173149, + 0.01618625596165657, + 0.020682163536548615, + 0.0046480256132781506, + 0.010132835246622562, + 0.003715624799951911, + 0.00467329565435648, + 0.006994330324232578, + 0.007101848255842924, + 0.0044075190089643, + 0.0037120545748621225, + 0.009153861552476883, + 0.00826600007712841, + 0.0080745043233037, + 0.007099173963069916, + 0.013362549245357513, + 0.007696995045989752, + 0.0077861458994448185, + 0.011632305569946766, + 0.004343176726251841, + 0.007654561195522547, + 0.0055857133120298386, + 0.006347340531647205, + 0.0066028437577188015, + 0.009800419211387634, + 0.006654985249042511, + 0.009230505675077438, + 0.01825380139052868, + 0.0184779055416584, + 0.014084569178521633, + 0.01848101057112217, + 0.03231050446629524, + 0.04287254810333252, + 0.023737533017992973, + 0.020735763013362885, + 0.005053234752267599, + 0.005277968477457762, + 0.005798387341201305, + 0.006074278615415096, + 0.008459209464490414, + 0.005959637463092804, + 0.005150508601218462, + 0.0032867579720914364, + 0.004760469309985638, + 0.007837288081645966, + 0.008413046598434448, + 0.006326003931462765, + 0.00505866901949048, + 0.004773447290062904, + 0.005664083641022444, + 0.0035453366581350565, + 0.00834220927208662, + 0.0173620767891407, + 0.01738472282886505, + 0.010695741511881351, + 0.022700905799865723, + 0.014028904028236866, + 0.0256540235131979, + 0.009358423762023449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.01419647503644228 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942060146480799 + ], + "opposite_logprob_tokens": [ + -5.130941867828369 + ], + "logprob_reward": -0.005942060146480799, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942060146480799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07770969718694687, + 0.12591753900051117, + 0.05863449349999428, + 0.06815893203020096, + 0.10034359246492386, + 0.05666922777891159, + 0.18060171604156494, + 0.14899173378944397, + 0.0815008357167244, + 0.10147220641374588 + ], + "credit_map": [ + 0.009746059775352478, + 0.009103563614189625, + 0.012309013865888119, + 0.006380672100931406, + 0.00845441222190857, + 0.010047447867691517, + 0.013581882230937481, + 0.00808663759380579, + 0.01360677182674408, + 0.013409176841378212, + 0.014332616701722145, + 0.012935101054608822, + 0.015094693750143051, + 0.011042679660022259, + 0.01716557890176773, + 0.02833092398941517, + 0.0035836391616612673, + 0.0034180895891040564, + 0.010187476873397827, + 0.0033927219919860363, + 0.007924878038465977, + 0.005240482743829489, + 0.011685516685247421, + 0.013201693072915077, + 0.007510875351727009, + 0.017442934215068817, + 0.006392233539372683, + 0.005457913503050804, + 0.009779606945812702, + 0.009408649057149887, + 0.00581835163757205, + 0.0063483682461082935, + 0.010558477602899075, + 0.017154473811388016, + 0.00893591158092022, + 0.009709571488201618, + 0.01818441040813923, + 0.011072557419538498, + 0.011735042557120323, + 0.012993145734071732, + 0.005949261132627726, + 0.007840749807655811, + 0.004369718953967094, + 0.007257359102368355, + 0.007985519245266914, + 0.013409018516540527, + 0.004005893133580685, + 0.005851707421243191, + 0.014638346619904041, + 0.016175344586372375, + 0.016614897176623344, + 0.01696844771504402, + 0.029231345281004906, + 0.038086701184511185, + 0.021403683349490166, + 0.02748294360935688, + 0.014912835322320461, + 0.014822312630712986, + 0.012915418483316898, + 0.0178562980145216, + 0.03231475129723549, + 0.02471381425857544, + 0.01466667465865612, + 0.016789628192782402, + 0.008207757957279682, + 0.013188930228352547, + 0.01377463061362505, + 0.014874774031341076, + 0.008905341848731041, + 0.009878301993012428, + 0.00593792786821723, + 0.006733175367116928, + 0.008655566722154617, + 0.015400001779198647, + 0.013606250286102295, + 0.009221605956554413, + 0.013491127640008926, + 0.012738716788589954, + 0.018672093749046326, + 0.009686848148703575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.005942060146480799 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008630122058093548 + ], + "opposite_logprob_tokens": [ + -4.758630275726318 + ], + "logprob_reward": -0.008630122058093548, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008630122058093548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11388123780488968, + 0.10101033002138138, + 0.06602654606103897, + 0.07140152156352997, + 0.09070227295160294, + 0.04273509979248047, + 0.1402413249015808, + 0.13587860763072968, + 0.14750295877456665, + 0.09062013775110245 + ], + "credit_map": [ + 0.015079578384757042, + 0.012294268235564232, + 0.01125919446349144, + 0.0077666486613452435, + 0.009681958705186844, + 0.015544557943940163, + 0.027402088046073914, + 0.014852946624159813, + 0.010484094731509686, + 0.009896760806441307, + 0.013773971237242222, + 0.011598118580877781, + 0.014226372353732586, + 0.011648552492260933, + 0.01177589874714613, + 0.01760655641555786, + 0.00387417315505445, + 0.005208934191614389, + 0.00749641377478838, + 0.003616793779656291, + 0.007706962991505861, + 0.004020061809569597, + 0.022907402366399765, + 0.011195806786417961, + 0.005832343827933073, + 0.011328433640301228, + 0.005409030709415674, + 0.006562323309481144, + 0.00945716816931963, + 0.01574821211397648, + 0.01033130381256342, + 0.006732705980539322, + 0.009652246721088886, + 0.020250139757990837, + 0.009217514656484127, + 0.007530433591455221, + 0.012126966379582882, + 0.011058936826884747, + 0.009659452363848686, + 0.011206581257283688, + 0.004160426557064056, + 0.0038863811641931534, + 0.0037402547895908356, + 0.005292878486216068, + 0.007271347101777792, + 0.009368989616632462, + 0.0038554773200303316, + 0.005159346852451563, + 0.014884954318404198, + 0.01641843654215336, + 0.010171830654144287, + 0.013111340813338757, + 0.019658314064145088, + 0.030977336689829826, + 0.015444525517523289, + 0.01957458257675171, + 0.012157388962805271, + 0.01265299879014492, + 0.015694553032517433, + 0.018437618389725685, + 0.020666372030973434, + 0.020742252469062805, + 0.01655159331858158, + 0.01897582598030567, + 0.013073091395199299, + 0.023980535566806793, + 0.019394034519791603, + 0.023927386850118637, + 0.01895713619887829, + 0.017433349043130875, + 0.015453741885721684, + 0.015283691696822643, + 0.006722357589751482, + 0.011341859586536884, + 0.011485863476991653, + 0.01073960866779089, + 0.013196869753301144, + 0.010339067317545414, + 0.01905241422355175, + 0.0077420976012945175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.008630122058093548 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slowly rotating, orbiting medium shot brings us into the core control area of a futuristic drone racing arena. Clad in explosive competitive cinematography, it focuses on two competitors battling side by side. The female competitor on the left, sporting a sleek, dark purple ponytail, wears immersive FPV (first-person view) goggles that half-cover her face, their lenses reflecting a dazzling stream of data. She wears a tight, black, high-tech racing suit with a neon blue circuit pattern. She leans forward slightly, her hands steady and decisive as she grips a complex multi-stick drone controller, her thumbs fine-tuning the joysticks with a speed barely tangible to the naked eye. Beside her, her male teammate, with slightly damp short black hair and a furrowed brow, is absorbed in analyzing a translucent holographic data screen floating before them. His fingers swiftly and deftly swipe and tap across the illuminated interface, adjusting performance parameters and flight path diagrams in real time. Behind them, on the huge circular track structure, other drones turned into blurred colorful light bands and flew by at high speed. The whole scene was illuminated by the cool light emitted by their equipment and the neon lights of the track, casting dynamic light and shadow on their focused faces, creating a tacit cooperation and highly tense competitive atmosphere in the high-speed confrontation.", + "seed": 16436, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_12/index_642_seed_16436_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.011976826190948486, + "vlm_logprob": -0.011976826190948486, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.011976826190948486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018900418654084206 + ], + "opposite_logprob_tokens": [ + -4.143900394439697 + ], + "logprob_reward": -0.018900418654084206, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.018900418654084206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10872557014226913, + 0.0223884005099535, + 0.13276897370815277, + 0.01816616952419281, + 0.06263775378465652, + 0.045057181268930435, + 0.04687301069498062, + 0.047355420887470245, + 0.44124141335487366, + 0.07478614896535873 + ], + "credit_map": [ + 0.008321096189320087, + 0.004086184781044722, + 0.005486511625349522, + 0.004907914903014898, + 0.009087302722036839, + 0.0347665511071682, + 0.005872481968253851, + 0.03619753196835518, + 0.003910450264811516, + 0.00391558138653636, + 0.00258445693179965, + 0.0019562235102057457, + 0.0022772701922804117, + 0.002522664610296488, + 0.002663983264937997, + 0.002557771047577262, + 0.006382095627486706, + 0.05055181309580803, + 0.004640594124794006, + 0.006327676121145487, + 0.023773808032274246, + 0.015452357940375805, + 0.005248824134469032, + 0.020391810685396194, + 0.002713026013225317, + 0.0024433955550193787, + 0.0025924730580300093, + 0.002644344000145793, + 0.0022657979279756546, + 0.00176182109862566, + 0.002015917794778943, + 0.0017293940763920546, + 0.004772092681378126, + 0.00947052426636219, + 0.007116529159247875, + 0.007581598591059446, + 0.014554652385413647, + 0.005673294886946678, + 0.007316445931792259, + 0.006152612157166004, + 0.004842048976570368, + 0.014459195546805859, + 0.0037685914430767298, + 0.005814820993691683, + 0.006428238470107317, + 0.00391136109828949, + 0.0032456680200994015, + 0.0025872604455798864, + 0.003203211585059762, + 0.005455892998725176, + 0.005249794106930494, + 0.005273590795695782, + 0.007053726818412542, + 0.007924102246761322, + 0.007484863977879286, + 0.005227826070040464, + 0.005302064120769501, + 0.0034259932581335306, + 0.006908508017659187, + 0.007182934787124395, + 0.011372451670467854, + 0.006684551481157541, + 0.004294311162084341, + 0.0021846042945981026, + 0.019051313400268555, + 0.012797631323337555, + 0.1078653484582901, + 0.040952689945697784, + 0.052098099142313004, + 0.155441552400589, + 0.02512378990650177, + 0.027910949662327766, + 0.005417174194008112, + 0.006134983617812395, + 0.012319546192884445, + 0.011753596365451813, + 0.009806366637349129, + 0.0052193221636116505, + 0.013605083338916302, + 0.010530075058341026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.018900418654084206 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011079457588493824 + ], + "opposite_logprob_tokens": [ + -4.51107931137085 + ], + "logprob_reward": -0.011079457588493824, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011079457588493824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43487751483917236, + 0.04886968806385994, + 0.09453378617763519, + 0.04280227795243263, + 0.020274857059121132, + 0.03940233960747719, + 0.08489672094583511, + 0.03626682236790657, + 0.11376353353261948, + 0.08431245386600494 + ], + "credit_map": [ + 0.03107561357319355, + 0.017061715945601463, + 0.014839697629213333, + 0.016023904085159302, + 0.02237827330827713, + 0.17081888020038605, + 0.01781509257853031, + 0.14486436545848846, + 0.006690030451864004, + 0.007626236416399479, + 0.0065709506161510944, + 0.0052251010201871395, + 0.00782434456050396, + 0.0063605280593037605, + 0.0035260990262031555, + 0.00504639744758606, + 0.004457530565559864, + 0.03398166969418526, + 0.003945872187614441, + 0.004332241136580706, + 0.01843186840415001, + 0.011809369549155235, + 0.0038692806847393513, + 0.013705953024327755, + 0.007539660204201937, + 0.005493263714015484, + 0.009519845247268677, + 0.0057618990540504456, + 0.003040366806089878, + 0.0038848062977194786, + 0.003788433503359556, + 0.003774006851017475, + 0.0024851951748132706, + 0.004002141300588846, + 0.0026426566764712334, + 0.002877771155908704, + 0.0031556282192468643, + 0.0015830127522349358, + 0.0018737498903647065, + 0.0016547022387385368, + 0.003980875946581364, + 0.010445925407111645, + 0.004621578846126795, + 0.0045283944346010685, + 0.004876530729234219, + 0.003918026108294725, + 0.004882223904132843, + 0.0021487846970558167, + 0.01459554024040699, + 0.0066273389384150505, + 0.012109165079891682, + 0.011755342595279217, + 0.0116667989641428, + 0.012942919507622719, + 0.008402622304856777, + 0.006796994712203741, + 0.0029290036763995886, + 0.0028897870797663927, + 0.004603203386068344, + 0.0038355637807399035, + 0.01108709629625082, + 0.003642195137217641, + 0.00425365474075079, + 0.0030263178050518036, + 0.006221655290573835, + 0.004693933296948671, + 0.04268050566315651, + 0.01935976929962635, + 0.009296390227973461, + 0.01763199269771576, + 0.00818036962300539, + 0.005698923021554947, + 0.006032619159668684, + 0.0078024729155004025, + 0.012795648537576199, + 0.01436444278806448, + 0.012909318320453167, + 0.006001852452754974, + 0.01307298056781292, + 0.011333122849464417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.011079457588493824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192244037985802 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014192244037985802, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192244037985802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39082998037338257, + 0.02689174748957157, + 0.03915080428123474, + 0.07418493181467056, + 0.09267150610685349, + 0.035265207290649414, + 0.0606519915163517, + 0.023933490738272667, + 0.1643737554550171, + 0.0920465737581253 + ], + "credit_map": [ + 0.028305863961577415, + 0.012023266404867172, + 0.0065080877393484116, + 0.009909508749842644, + 0.020542893558740616, + 0.16022716462612152, + 0.014879101887345314, + 0.13843409717082977, + 0.003706591436639428, + 0.0038637593388557434, + 0.004438081756234169, + 0.004101624712347984, + 0.0024455832317471504, + 0.002673613140359521, + 0.0030053984373807907, + 0.0026570989284664392, + 0.0034015437122434378, + 0.005834651179611683, + 0.0031705335713922977, + 0.003998580388724804, + 0.0063787116669118404, + 0.009430186823010445, + 0.003252315567806363, + 0.0036842855624854565, + 0.01007995754480362, + 0.009526287205517292, + 0.009309090673923492, + 0.01109715923666954, + 0.007350573316216469, + 0.006682293489575386, + 0.010813523083925247, + 0.009326049126684666, + 0.00828482210636139, + 0.014166287146508694, + 0.010456365533173084, + 0.010282515548169613, + 0.023625168949365616, + 0.0077642519026994705, + 0.010774550959467888, + 0.00731755280867219, + 0.003939591348171234, + 0.00768987787887454, + 0.003948224242776632, + 0.004202221520245075, + 0.005177609622478485, + 0.003361114766448736, + 0.004327437374740839, + 0.00261913170106709, + 0.006070784293115139, + 0.00588564807549119, + 0.006817359011620283, + 0.007697199936956167, + 0.009107068181037903, + 0.013366803526878357, + 0.006217220798134804, + 0.005489914678037167, + 0.0019306832691654563, + 0.0011000566883012652, + 0.002865259302780032, + 0.002874288707971573, + 0.003058262635022402, + 0.007301465142518282, + 0.002878018422052264, + 0.001925456221215427, + 0.009106016717851162, + 0.006512490101158619, + 0.06190374493598938, + 0.028541352599859238, + 0.014341018162667751, + 0.023647723719477654, + 0.011977641843259335, + 0.00834378320723772, + 0.00712257856503129, + 0.007711127866059542, + 0.01405639760196209, + 0.014865686185657978, + 0.013277530670166016, + 0.0065438407473266125, + 0.01703479140996933, + 0.011434626765549183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.014192244037985802 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005940282717347145 + ], + "opposite_logprob_tokens": [ + -5.1309404373168945 + ], + "logprob_reward": -0.005940282717347145, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005940282717347145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7772029638290405, + 0.0201781764626503, + 0.02463315986096859, + 0.016057351604104042, + 0.012663187459111214, + 0.029685920104384422, + 0.016413848847150803, + 0.009792237542569637, + 0.016454890370368958, + 0.07691825181245804 + ], + "credit_map": [ + 0.057342205196619034, + 0.021303456276655197, + 0.012939054518938065, + 0.017537854611873627, + 0.040908023715019226, + 0.3212818205356598, + 0.02763259783387184, + 0.27825799584388733, + 0.002847186289727688, + 0.0026206662878394127, + 0.0019426975632086396, + 0.0024549770168960094, + 0.0021744449622929096, + 0.002191501669585705, + 0.0032393161673098803, + 0.0027073875535279512, + 0.0018219101475551724, + 0.005820310674607754, + 0.002157153794541955, + 0.0020634466782212257, + 0.0031521818600594997, + 0.0043365745805203915, + 0.0012127774534747005, + 0.004068804904818535, + 0.0021001347340643406, + 0.0024097629357129335, + 0.0023169280029833317, + 0.0018074932741001248, + 0.00210551917552948, + 0.001677444321103394, + 0.0018629112746566534, + 0.0017771567218005657, + 0.0009591002599336207, + 0.0016649668104946613, + 0.0018569116946309805, + 0.0015299125807359815, + 0.002317613223567605, + 0.0016176492208614945, + 0.001239210250787437, + 0.001477823476307094, + 0.0030081714503467083, + 0.003651645267382264, + 0.004002800211310387, + 0.005332164466381073, + 0.004308617208153009, + 0.00242943549528718, + 0.003751208772882819, + 0.0032018774654716253, + 0.0019451528787612915, + 0.002000721637159586, + 0.0013708336045965552, + 0.0015451682265847921, + 0.0025375685654580593, + 0.0033895003143697977, + 0.0016955184983089566, + 0.0019293844234198332, + 0.0011843054089695215, + 0.0005980449495837092, + 0.0014324333751574159, + 0.0013598602963611484, + 0.0008730513509362936, + 0.0020703189074993134, + 0.0015138533199205995, + 0.0007603698177263141, + 0.0011578291887417436, + 0.0010517173213884234, + 0.0017477247165516019, + 0.0016702734865248203, + 0.00267566810362041, + 0.004685272928327322, + 0.0014079820830374956, + 0.002058422891423106, + 0.005372012034058571, + 0.006844914518296719, + 0.011256340891122818, + 0.012751983478665352, + 0.012910949066281319, + 0.005653570871800184, + 0.011748649179935455, + 0.01037982851266861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.005940282717347145 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium shot featuring two competitors side by side in a futuristic drone racing arena?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the female competitor on the left wearing a tight black high-tech racing suit with a neon blue circuit pattern and sporting a dark purple ponytail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the male competitor on the right wearing a suit, has short black hair, and is interacting with a translucent holographic data screen?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are blurred colorful light bands visible in the background representing drones flying at high speed on a huge circular track?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by cool equipment light and neon track lights casting dynamic shadows on the competitors' faces?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009771727956831455 + ], + "opposite_logprob_tokens": [ + -4.634771823883057 + ], + "logprob_reward": -0.009771727956831455, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009771727956831455, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.45638373494148254, + 0.02345329150557518, + 0.0404624305665493, + 0.014107893221080303, + 0.02548045665025711, + 0.01841031014919281, + 0.027213700115680695, + 0.02977020852267742, + 0.2031778246164322, + 0.16154015064239502 + ], + "credit_map": [ + 0.03460371121764183, + 0.016034893691539764, + 0.009471463039517403, + 0.011750055477023125, + 0.023655205965042114, + 0.18517310917377472, + 0.01604503206908703, + 0.15965023636817932, + 0.0025082495994865894, + 0.002485834062099457, + 0.002524806186556816, + 0.0024592436384409666, + 0.004150095861405134, + 0.0031667565926909447, + 0.0032112959306687117, + 0.0029470098670572042, + 0.0036671916022896767, + 0.011261381208896637, + 0.0024462328292429447, + 0.00345055665820837, + 0.004713724367320538, + 0.003879024414345622, + 0.002873125486075878, + 0.00817119237035513, + 0.0016325062606483698, + 0.002131030894815922, + 0.001838715630583465, + 0.001968477852642536, + 0.0019173803739249706, + 0.0016252081841230392, + 0.0015787184238433838, + 0.0014158558333292603, + 0.002840133849531412, + 0.0038423375226557255, + 0.00508448900654912, + 0.004682449623942375, + 0.003062083153054118, + 0.0018719444051384926, + 0.0021847644820809364, + 0.001912255072966218, + 0.002514253370463848, + 0.0022149202413856983, + 0.002004793146625161, + 0.0021142088808119297, + 0.00254568993113935, + 0.0015964729245752096, + 0.003621620126068592, + 0.0017983498983085155, + 0.0026608684565871954, + 0.002503104042261839, + 0.004804672207683325, + 0.004521611146628857, + 0.005903587676584721, + 0.0026660978328436613, + 0.002593982731923461, + 0.0015597785823047161, + 0.002851826371625066, + 0.0018741211388260126, + 0.0035869062412530184, + 0.003236229997128248, + 0.0056278458796441555, + 0.0043963100761175156, + 0.005646070931106806, + 0.0025508992839604616, + 0.009424848482012749, + 0.005388887133449316, + 0.040703125298023224, + 0.015054875053465366, + 0.03255746141076088, + 0.07669276744127274, + 0.011181781068444252, + 0.01217408012598753, + 0.011605575680732727, + 0.015222904272377491, + 0.026795649901032448, + 0.027238138020038605, + 0.023748911917209625, + 0.010691200383007526, + 0.02634834684431553, + 0.01988941803574562 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.009771727956831455 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_tensor([989, 989])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_tensor([989, 989])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..421a36abdb961e0a3b10caf470b428ac81691e20 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_tensor([989, 989])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "camera lens, held at eye level, slowly pans forward through a dreamy forest scene bathed in a warm, ethereal glow. In the background, a soft golden light pulses rhythmically, illuminating lazy white smoke drifting across a soft-focus backdrop of ancient trees. As the camera draws closer, a delightful group of small animals gathers on a patch of lush green moss. A fluffy white bunny with twitching pink ears, a ginger kitten with bright green eyes, and a fluffy yellow chick, all lost in pure joy. In a captivatingly out-of-sync rhythm, they begin to shake their bodies; the bunny vigorously exhales a fine spray of dew from its fur, the kitten playfully wiggles from head to tail, and the chick adorably fluffs its feathers in a rapid, quivering throb. The magical golden light catches the edges of their soft fur and down, creating a gentle halo effect that casts long, soft shadows and creates a deeply serene and warm atmosphere of pure, innocent magic.", + "seed": 9322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9322_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14012389256968164, + "vlm_logprob": -0.14012389256968164, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.14012389256968164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6429944038391113 + ], + "opposite_logprob_tokens": [ + -0.7679944038391113 + ], + "logprob_reward": -0.6429944038391113, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6429944038391113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13510632514953613, + 0.06594748795032501, + 0.14317083358764648, + 0.11344368755817413, + 0.1289629340171814, + 0.07085192948579788, + 0.07223638147115707, + 0.038018979132175446, + 0.18440696597099304, + 0.0478544645011425 + ], + "credit_map": [ + 0.011736961081624031, + 0.012467860244214535, + 0.014952368102967739, + 0.01650427095592022, + 0.012378823012113571, + 0.022915804758667946, + 0.030771570280194283, + 0.013378662057220936, + 0.009117896668612957, + 0.0046492330729961395, + 0.014434002339839935, + 0.006163174752146006, + 0.008048269897699356, + 0.013966353610157967, + 0.0047503625974059105, + 0.00481819873675704, + 0.011076841503381729, + 0.012492893263697624, + 0.018545016646385193, + 0.019253646954894066, + 0.013855529949069023, + 0.015060515142977238, + 0.032508544623851776, + 0.020377852022647858, + 0.008346346206963062, + 0.0121785132214427, + 0.014046655036509037, + 0.017273642122745514, + 0.010945693589746952, + 0.013820482417941093, + 0.02597566321492195, + 0.01085670106112957, + 0.01368427649140358, + 0.011420439928770065, + 0.018218468874692917, + 0.01608310453593731, + 0.014328774996101856, + 0.014646295458078384, + 0.02274269051849842, + 0.017838889732956886, + 0.0066204858012497425, + 0.008761990815401077, + 0.009334078989923, + 0.007530652452260256, + 0.007355981506407261, + 0.01199785154312849, + 0.00886657927185297, + 0.010384314693510532, + 0.007469270844012499, + 0.010083048604428768, + 0.01319186482578516, + 0.008783329278230667, + 0.0072412737645208836, + 0.0075150057673454285, + 0.008011839352548122, + 0.009940749034285545, + 0.002848604926839471, + 0.003733317367732525, + 0.0032318029552698135, + 0.005043851211667061, + 0.00485859764739871, + 0.007372372318059206, + 0.004716258030384779, + 0.006214179564267397, + 0.00851193256676197, + 0.015072638168931007, + 0.013388006947934628, + 0.025500938296318054, + 0.017613528296351433, + 0.02224992774426937, + 0.05623466148972511, + 0.025835340842604637, + 0.006766243372112513, + 0.006002756301313639, + 0.007107936777174473, + 0.004583174362778664, + 0.00435657799243927, + 0.005604668520390987, + 0.006155633833259344, + 0.0072774747386574745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 2, + 0 + ], + "selected_reward": -0.6429944038391113 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038114167749881744 + ], + "opposite_logprob_tokens": [ + -3.288114070892334 + ], + "logprob_reward": -0.038114167749881744, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038114167749881744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11151324212551117, + 0.12633919715881348, + 0.10409189760684967, + 0.13021260499954224, + 0.10364340245723724, + 0.025453533977270126, + 0.04110148921608925, + 0.0645710751414299, + 0.2641143798828125, + 0.02895914763212204 + ], + "credit_map": [ + 0.017904380336403847, + 0.014979428611695766, + 0.014929075725376606, + 0.012824763543903828, + 0.009476889856159687, + 0.010102053172886372, + 0.015521209686994553, + 0.015775449573993683, + 0.012774217873811722, + 0.007837003096938133, + 0.011678682640194893, + 0.009273790754377842, + 0.02379557304084301, + 0.035050224512815475, + 0.012783179990947247, + 0.013146539218723774, + 0.009052244946360588, + 0.012748200446367264, + 0.018340645357966423, + 0.015247364528477192, + 0.008527792990207672, + 0.013643571175634861, + 0.01671629026532173, + 0.009815809316933155, + 0.012554960325360298, + 0.01163181196898222, + 0.015289181843400002, + 0.016988283023238182, + 0.011714915744960308, + 0.015985770151019096, + 0.029567215591669083, + 0.01648048125207424, + 0.011097468435764313, + 0.012190500274300575, + 0.015563135966658592, + 0.01289532519876957, + 0.008882577531039715, + 0.014281651936471462, + 0.015469608828425407, + 0.013263137079775333, + 0.003003472462296486, + 0.0029659969732165337, + 0.002647612476721406, + 0.0024655309971421957, + 0.002554917009547353, + 0.004411407746374607, + 0.003305371617898345, + 0.004099228885024786, + 0.003977824933826923, + 0.005715405102819204, + 0.005223389249294996, + 0.004085268825292587, + 0.00654225330799818, + 0.00458351569250226, + 0.005679560359567404, + 0.005294273141771555, + 0.007241323124617338, + 0.006388317793607712, + 0.006234089843928814, + 0.006918033584952354, + 0.005582519806921482, + 0.009000339545309544, + 0.010448072105646133, + 0.012758384458720684, + 0.01386986393481493, + 0.026171552017331123, + 0.02117353305220604, + 0.03465016558766365, + 0.024558991193771362, + 0.028251133859157562, + 0.07241831719875336, + 0.0430208183825016, + 0.00477789668366313, + 0.0033894875086843967, + 0.004544341936707497, + 0.002995688933879137, + 0.002539857989177108, + 0.0032185183372348547, + 0.003858573269098997, + 0.0036347866989672184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 1 + ], + "selected_reward": -0.038114167749881744 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0002451834443490952 + ], + "opposite_logprob_tokens": [ + -8.375245094299316 + ], + "logprob_reward": -0.0002451834443490952, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.0002451834443490952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24374903738498688, + 0.16615277528762817, + 0.07132966071367264, + 0.1775503009557724, + 0.07647721469402313, + 0.0333903506398201, + 0.031522441655397415, + 0.035069648176431656, + 0.09252214431762695, + 0.07223641872406006 + ], + "credit_map": [ + 0.018869411200284958, + 0.017411982640624046, + 0.023547621443867683, + 0.030224472284317017, + 0.027018671855330467, + 0.04524229094386101, + 0.05339754372835159, + 0.02803703211247921, + 0.016835996881127357, + 0.011374551802873611, + 0.01117630023509264, + 0.011463398113846779, + 0.029976986348628998, + 0.0462474450469017, + 0.021101035177707672, + 0.01797705702483654, + 0.007917739450931549, + 0.012950927950441837, + 0.011981770396232605, + 0.008152317255735397, + 0.00703786825761199, + 0.009902567602694035, + 0.007004992105066776, + 0.00638146698474884, + 0.014990460127592087, + 0.014647213742136955, + 0.018717186525464058, + 0.03610718995332718, + 0.01449357345700264, + 0.019583312794566154, + 0.02564278617501259, + 0.033368561416864395, + 0.006483737844973803, + 0.006079480051994324, + 0.009702540934085846, + 0.01035118568688631, + 0.0068947249092161655, + 0.009742124937474728, + 0.014766099862754345, + 0.012457314878702164, + 0.003829775843769312, + 0.00303543359041214, + 0.004033799283206463, + 0.00354002951644361, + 0.003276763716712594, + 0.006401697639375925, + 0.004331698175519705, + 0.004941146820783615, + 0.0030197836458683014, + 0.004259791690856218, + 0.005146778654307127, + 0.003969243261963129, + 0.0031594280153512955, + 0.0032122221309691668, + 0.003856220282614231, + 0.004898968618363142, + 0.004608003422617912, + 0.004858688451349735, + 0.0040995837189257145, + 0.003949017729610205, + 0.0036422840785235167, + 0.0037166441325098276, + 0.0065423049964010715, + 0.0036531188525259495, + 0.008052303455770016, + 0.010824325494468212, + 0.012960296124219894, + 0.013674605637788773, + 0.009591871872544289, + 0.01144176535308361, + 0.01704919897019863, + 0.00892777182161808, + 0.0066628167405724525, + 0.005071471910923719, + 0.008343073539435863, + 0.007498116698116064, + 0.00715832831338048, + 0.008500166237354279, + 0.012257336638867855, + 0.016745099797844887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0002451834443490952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016076110303401947 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016076110303401947, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016076110303401947, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10778740048408508, + 0.0979982316493988, + 0.061549391597509384, + 0.17260463535785675, + 0.08225134760141373, + 0.048957400023937225, + 0.2852029800415039, + 0.026166291907429695, + 0.08482422679662704, + 0.03265811502933502 + ], + "credit_map": [ + 0.014886243268847466, + 0.01884828321635723, + 0.013035290874540806, + 0.011281443759799004, + 0.011913691647350788, + 0.013883189298212528, + 0.012220860458910465, + 0.011718403548002243, + 0.01166410930454731, + 0.00890686921775341, + 0.014467697590589523, + 0.008702022023499012, + 0.01257024984806776, + 0.02168380841612816, + 0.01269285473972559, + 0.007310620974749327, + 0.0048210290260612965, + 0.005649644881486893, + 0.008338920772075653, + 0.008645954541862011, + 0.0051999781280756, + 0.007114008069038391, + 0.014235652051866055, + 0.007544201333075762, + 0.017352888360619545, + 0.0202699676156044, + 0.024093953892588615, + 0.02742161601781845, + 0.014907029457390308, + 0.02213689498603344, + 0.019259674474596977, + 0.027162613347172737, + 0.008325580507516861, + 0.01319469977170229, + 0.018395615741610527, + 0.010429088957607746, + 0.009275520220398903, + 0.008551616221666336, + 0.007035332731902599, + 0.007043895777314901, + 0.007119240704923868, + 0.009200112894177437, + 0.006206207443028688, + 0.006566060706973076, + 0.0050539374351501465, + 0.005287422798573971, + 0.005369804333895445, + 0.004154615104198456, + 0.03347556293010712, + 0.047683730721473694, + 0.05236928164958954, + 0.03600946441292763, + 0.033160049468278885, + 0.026412788778543472, + 0.02500685676932335, + 0.03108522854745388, + 0.0033649541437625885, + 0.003451016265898943, + 0.0030455151572823524, + 0.0035740116145461798, + 0.0026136680971831083, + 0.002420647069811821, + 0.004377727396786213, + 0.00331875286065042, + 0.008903495967388153, + 0.01155430544167757, + 0.007343331351876259, + 0.010419018566608429, + 0.008228501304984093, + 0.009101424366235733, + 0.017181098461151123, + 0.012093051336705685, + 0.0036656546872109175, + 0.00432567996904254, + 0.003850153647363186, + 0.00447497796267271, + 0.003025135025382042, + 0.003370553022250533, + 0.0043465932831168175, + 0.005599369294941425 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 0 + ], + "selected_reward": -0.016076110303401947 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003189597511664033 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.003189597511664033, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003189597511664033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09253717213869095, + 0.05867763236165047, + 0.05269763991236687, + 0.4407059848308563, + 0.05236087739467621, + 0.02169036865234375, + 0.06256221234798431, + 0.030301107093691826, + 0.13081872463226318, + 0.057648222893476486 + ], + "credit_map": [ + 0.01133789587765932, + 0.007771736476570368, + 0.012088148854672909, + 0.01283953059464693, + 0.007707610726356506, + 0.012814112938940525, + 0.013371079228818417, + 0.01460705790668726, + 0.006771168205887079, + 0.0060514784418046474, + 0.004246401134878397, + 0.00507788360118866, + 0.010940192267298698, + 0.013307658955454826, + 0.006556398700922728, + 0.005726449191570282, + 0.004581852816045284, + 0.0041618309915065765, + 0.0045836614444851875, + 0.005497689358890057, + 0.005530696362257004, + 0.005482158157974482, + 0.01549556851387024, + 0.007364183664321899, + 0.046607159078121185, + 0.041169729083776474, + 0.05159139260649681, + 0.07285616546869278, + 0.03805136680603027, + 0.05902562290430069, + 0.04746071994304657, + 0.08394387364387512, + 0.005301253404468298, + 0.006868027150630951, + 0.011594653129577637, + 0.00679340586066246, + 0.0046059018932282925, + 0.004730623215436935, + 0.00598002877086401, + 0.006486984435468912, + 0.0032895051408559084, + 0.0030434473883360624, + 0.003346592653542757, + 0.001948110992088914, + 0.002120934659615159, + 0.002944719512015581, + 0.002438101451843977, + 0.0025589570868760347, + 0.006646792404353619, + 0.00958135537803173, + 0.010711327195167542, + 0.008219488896429539, + 0.0070665087550878525, + 0.0061089289374649525, + 0.0063496851362288, + 0.007878128439188004, + 0.002773144282400608, + 0.003090905724093318, + 0.0024665414821356535, + 0.0042414884082973, + 0.0034750383347272873, + 0.005420644301921129, + 0.003988118376582861, + 0.004845224320888519, + 0.007427567150443792, + 0.011352317407727242, + 0.009206709451973438, + 0.019267290830612183, + 0.013236967846751213, + 0.019367599859833717, + 0.03305064141750336, + 0.017909636721014977, + 0.0061727650463581085, + 0.004538508597761393, + 0.005907971877604723, + 0.0052084531635046005, + 0.00544707803055644, + 0.007396687287837267, + 0.010478022508323193, + 0.012498737312853336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.003189597511664033 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera lens, held at eye level, slowly pans forward through a dreamy forest scene bathed in a warm, ethereal glow. In the background, a soft golden light pulses rhythmically, illuminating lazy white smoke drifting across a soft-focus backdrop of ancient trees. As the camera draws closer, a delightful group of small animals gathers on a patch of lush green moss. A fluffy white bunny with twitching pink ears, a ginger kitten with bright green eyes, and a fluffy yellow chick, all lost in pure joy. In a captivatingly out-of-sync rhythm, they begin to shake their bodies; the bunny vigorously exhales a fine spray of dew from its fur, the kitten playfully wiggles from head to tail, and the chick adorably fluffs its feathers in a rapid, quivering throb. The magical golden light catches the edges of their soft fur and down, creating a gentle halo effect that casts long, soft shadows and creates a deeply serene and warm atmosphere of pure, innocent magic.", + "seed": 9323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9323_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14609960382804274, + "vlm_logprob": -0.14609960382804274, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.14609960382804274, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6527001857757568 + ], + "opposite_logprob_tokens": [ + -0.7777001857757568 + ], + "logprob_reward": -0.6527001857757568, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6527001857757568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16831731796264648, + 0.0314265713095665, + 0.0426902137696743, + 0.07084543257951736, + 0.08463846892118454, + 0.044213514775037766, + 0.10959378629922867, + 0.0344231091439724, + 0.2315709888935089, + 0.1822805553674698 + ], + "credit_map": [ + 0.021805092692375183, + 0.014896297827363014, + 0.03742098808288574, + 0.013635761104524136, + 0.016558701172471046, + 0.013232292607426643, + 0.015438308008015156, + 0.03532987833023071, + 0.005775706376880407, + 0.003787752939388156, + 0.002866956638172269, + 0.0034278708044439554, + 0.004278811626136303, + 0.0035795685835182667, + 0.004010400269180536, + 0.003699503606185317, + 0.004692812915891409, + 0.006212365813553333, + 0.003799417754635215, + 0.005769048817455769, + 0.00434868736192584, + 0.006811202969402075, + 0.005088758189231157, + 0.005967920646071434, + 0.011181380599737167, + 0.010848984122276306, + 0.00525401858612895, + 0.005029888823628426, + 0.008312206715345383, + 0.00869708601385355, + 0.009026437997817993, + 0.012495428323745728, + 0.010378834791481495, + 0.015409808605909348, + 0.008753369562327862, + 0.01264838594943285, + 0.01570945233106613, + 0.009852086193859577, + 0.006116734351962805, + 0.005769796669483185, + 0.005818651057779789, + 0.007873279973864555, + 0.005340870004147291, + 0.005440326873213053, + 0.007160321809351444, + 0.004446597304195166, + 0.003325206460431218, + 0.004808262921869755, + 0.013042056001722813, + 0.014214768074452877, + 0.009647072292864323, + 0.00939919427037239, + 0.028842724859714508, + 0.01081649400293827, + 0.008765810169279575, + 0.014865659177303314, + 0.004296173341572285, + 0.006398467347025871, + 0.004970262758433819, + 0.0033782573882490396, + 0.005801055580377579, + 0.0026376377791166306, + 0.002945314859971404, + 0.003995941486209631, + 0.04081875458359718, + 0.033786047250032425, + 0.02835954539477825, + 0.03264845535159111, + 0.030687959864735603, + 0.024268263950943947, + 0.02357555367052555, + 0.017426403239369392, + 0.04290897399187088, + 0.022694529965519905, + 0.028968367725610733, + 0.016533704474568367, + 0.024760765954852104, + 0.01442705187946558, + 0.014430390670895576, + 0.017556777223944664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.6527001857757568 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867521673440933 + ], + "opposite_logprob_tokens": [ + -3.048675298690796 + ], + "logprob_reward": -0.04867521673440933, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867521673440933, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36088672280311584, + 0.08370164036750793, + 0.06403390318155289, + 0.12764587998390198, + 0.031951189041137695, + 0.06531477719545364, + 0.07670189440250397, + 0.05752740055322647, + 0.09580780565738678, + 0.0364287830889225 + ], + "credit_map": [ + 0.04589032381772995, + 0.0379558801651001, + 0.06469502300024033, + 0.02783612348139286, + 0.04296235740184784, + 0.02776670828461647, + 0.03283284977078438, + 0.08094743639230728, + 0.010064076632261276, + 0.00816945917904377, + 0.010989082045853138, + 0.0061417231336236, + 0.01393567118793726, + 0.009253347292542458, + 0.010960943065583706, + 0.014187341555953026, + 0.008310656994581223, + 0.006928287912160158, + 0.004804580006748438, + 0.005438879132270813, + 0.009405857883393764, + 0.012485763058066368, + 0.006421598605811596, + 0.010238279588520527, + 0.020593974739313126, + 0.017742184922099113, + 0.007900633849203587, + 0.007857505232095718, + 0.014506721869111061, + 0.01722870022058487, + 0.016038674861192703, + 0.025777487084269524, + 0.00334958010353148, + 0.00554891861975193, + 0.0034737223759293556, + 0.004741501063108444, + 0.00664918590337038, + 0.0025672116316854954, + 0.00309786107391119, + 0.0025232071056962013, + 0.008516720496118069, + 0.012666257098317146, + 0.006119874771684408, + 0.007054440211504698, + 0.012499659322202206, + 0.006950957700610161, + 0.006773679051548243, + 0.004733183886855841, + 0.00851869210600853, + 0.008230253122746944, + 0.008998822420835495, + 0.008529969491064548, + 0.010848931036889553, + 0.006055715028196573, + 0.009653608314692974, + 0.015865899622440338, + 0.0075974613428115845, + 0.008250098675489426, + 0.007907566614449024, + 0.006979034282267094, + 0.006162600591778755, + 0.00583028607070446, + 0.0054767741821706295, + 0.009323579259216785, + 0.01279469858855009, + 0.012556483037769794, + 0.014079824090003967, + 0.018329760059714317, + 0.010337875224649906, + 0.009391785599291325, + 0.008696625009179115, + 0.009620754048228264, + 0.00514255603775382, + 0.00456473883241415, + 0.004221681505441666, + 0.006632226053625345, + 0.005189144518226385, + 0.002959434874355793, + 0.003246928099542856, + 0.0044720713049173355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.04867521673440933 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005708495154976845 + ], + "opposite_logprob_tokens": [ + -7.500570774078369 + ], + "logprob_reward": -0.0005708495154976845, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005708495154976845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.181264266371727, + 0.07902441173791885, + 0.1322147250175476, + 0.04051005467772484, + 0.07859347760677338, + 0.04075747728347778, + 0.09647520631551743, + 0.1122785285115242, + 0.12278680503368378, + 0.11609499156475067 + ], + "credit_map": [ + 0.024699078872799873, + 0.01252751424908638, + 0.02672244980931282, + 0.011775282211601734, + 0.02006291225552559, + 0.020136410370469093, + 0.016198553144931793, + 0.04914206266403198, + 0.011780023574829102, + 0.01119701936841011, + 0.011681409552693367, + 0.007101324386894703, + 0.013329923152923584, + 0.007704880554229021, + 0.008319389075040817, + 0.007910436019301414, + 0.0151666896417737, + 0.011978372931480408, + 0.009762594476342201, + 0.01371566578745842, + 0.0165861863642931, + 0.02337775006890297, + 0.015189049765467644, + 0.026438428089022636, + 0.004742200020700693, + 0.005530613474547863, + 0.0030201177578419447, + 0.004445843398571014, + 0.005189828108996153, + 0.006189288571476936, + 0.005876583978533745, + 0.005515579599887133, + 0.00858776830136776, + 0.00860202219337225, + 0.007823683321475983, + 0.010181091725826263, + 0.01724103093147278, + 0.008291840553283691, + 0.008999359793961048, + 0.00886669009923935, + 0.005100567359477282, + 0.006095028016716242, + 0.004332159645855427, + 0.003690578741952777, + 0.0072553036734461784, + 0.0067280554212629795, + 0.003480014856904745, + 0.004075770266354084, + 0.012328258715569973, + 0.018010171130299568, + 0.008699371479451656, + 0.008310387842357159, + 0.022320805117487907, + 0.008090096525847912, + 0.007145155221223831, + 0.011570969596505165, + 0.011533843353390694, + 0.015225728042423725, + 0.0141072366386652, + 0.012646586634218693, + 0.015792973339557648, + 0.01229217927902937, + 0.01197420060634613, + 0.01870577782392502, + 0.014111717231571674, + 0.01641100086271763, + 0.015455846674740314, + 0.022219588980078697, + 0.018044933676719666, + 0.012959878891706467, + 0.013690836727619171, + 0.00989300012588501, + 0.02848242223262787, + 0.012308964505791664, + 0.015668608248233795, + 0.011269547045230865, + 0.0184773001819849, + 0.008377022109925747, + 0.00998908095061779, + 0.011522045359015465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0005708495154976845 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355484500527382 + ], + "opposite_logprob_tokens": [ + -3.651355504989624 + ], + "logprob_reward": -0.026355484500527382, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355484500527382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22057729959487915, + 0.03608429804444313, + 0.061636392027139664, + 0.0585806779563427, + 0.05209121108055115, + 0.05731011554598808, + 0.08833259344100952, + 0.06116821989417076, + 0.24377910792827606, + 0.12044011801481247 + ], + "credit_map": [ + 0.02473892644047737, + 0.01687423326075077, + 0.03657732531428337, + 0.017543651163578033, + 0.024628279730677605, + 0.02124323509633541, + 0.02167566493153572, + 0.05729597806930542, + 0.004724421538412571, + 0.005038864444941282, + 0.007257568649947643, + 0.0037356086540967226, + 0.0038192258216440678, + 0.002431432483717799, + 0.004183058626949787, + 0.004894119221717119, + 0.007632538676261902, + 0.007938220165669918, + 0.005567566026002169, + 0.005847583990544081, + 0.009890032932162285, + 0.009535783901810646, + 0.007095048204064369, + 0.00812961719930172, + 0.009161279536783695, + 0.0080795893445611, + 0.004895167890936136, + 0.006587023381143808, + 0.005835253745317459, + 0.006514113862067461, + 0.008128242567181587, + 0.009380008094012737, + 0.0064670974388718605, + 0.007171283476054668, + 0.00489836698397994, + 0.004879502579569817, + 0.013616216368973255, + 0.006046760827302933, + 0.004623156040906906, + 0.004388828761875629, + 0.006939057260751724, + 0.010642960667610168, + 0.004408407025039196, + 0.0066480753012001514, + 0.012469346635043621, + 0.007450071629136801, + 0.004238903056830168, + 0.004513295833021402, + 0.010401584208011627, + 0.013849334791302681, + 0.009783657267689705, + 0.011202600784599781, + 0.018813079223036766, + 0.0069406903348863125, + 0.007404657546430826, + 0.009936990216374397, + 0.007688002660870552, + 0.008553076535463333, + 0.008616486564278603, + 0.008077265694737434, + 0.0061887893825769424, + 0.005958808120340109, + 0.0056231203489005566, + 0.010462668724358082, + 0.027514623478055, + 0.034802649170160294, + 0.03876880556344986, + 0.03954269364476204, + 0.03576446697115898, + 0.0194166898727417, + 0.02432248368859291, + 0.02364669367671013, + 0.031860172748565674, + 0.014306084252893925, + 0.017756277695298195, + 0.013143890537321568, + 0.017628969624638557, + 0.00884518027305603, + 0.008814486674964428, + 0.008085053414106369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.026355484500527382 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021962826140224934 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.0021962826140224934, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021962826140224934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30883437395095825, + 0.1222107782959938, + 0.04938710108399391, + 0.07292097061872482, + 0.05354466289281845, + 0.04643210396170616, + 0.14374493062496185, + 0.057760898023843765, + 0.08054696768522263, + 0.06461716443300247 + ], + "credit_map": [ + 0.04731892794370651, + 0.026673223823308945, + 0.04084648936986923, + 0.026685984805226326, + 0.043653007596731186, + 0.03302529826760292, + 0.026638854295015335, + 0.06399261951446533, + 0.01683046668767929, + 0.017460409551858902, + 0.01871628686785698, + 0.009753587655723095, + 0.023723620921373367, + 0.011593862436711788, + 0.01108239684253931, + 0.013050162233412266, + 0.006540755275636911, + 0.007313356269150972, + 0.004104395862668753, + 0.005332426633685827, + 0.0063718403689563274, + 0.00761210173368454, + 0.005178087390959263, + 0.00693414406850934, + 0.01121711079031229, + 0.014758418314158916, + 0.005778536666184664, + 0.0062386407516896725, + 0.00750068249180913, + 0.008596997708082199, + 0.007867362350225449, + 0.01096322387456894, + 0.007849601097404957, + 0.01070138718932867, + 0.00538343982771039, + 0.008759857155382633, + 0.006454647518694401, + 0.005557587835937738, + 0.004525718744844198, + 0.004312429111450911, + 0.0051289210096001625, + 0.006661322899162769, + 0.002702118596062064, + 0.004614937584847212, + 0.011206514202058315, + 0.005835933145135641, + 0.0053874957375228405, + 0.0048948610201478004, + 0.015517337247729301, + 0.018110428005456924, + 0.017828864976763725, + 0.01620473526418209, + 0.03282945975661278, + 0.011130893602967262, + 0.014202681370079517, + 0.01792052760720253, + 0.009006493724882603, + 0.008417792618274689, + 0.008434277027845383, + 0.0067920382134616375, + 0.006382872816175222, + 0.005498319864273071, + 0.004847185220569372, + 0.008381922729313374, + 0.007487238850444555, + 0.011181667447090149, + 0.011065434664487839, + 0.017117977142333984, + 0.009657667949795723, + 0.007676632609218359, + 0.00973379984498024, + 0.006626554764807224, + 0.011318429373204708, + 0.006953191477805376, + 0.011663469485938549, + 0.006466052494943142, + 0.010488693602383137, + 0.005195366684347391, + 0.006138926837593317, + 0.006393037270754576 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.0021962826140224934 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera lens, held at eye level, slowly pans forward through a dreamy forest scene bathed in a warm, ethereal glow. In the background, a soft golden light pulses rhythmically, illuminating lazy white smoke drifting across a soft-focus backdrop of ancient trees. As the camera draws closer, a delightful group of small animals gathers on a patch of lush green moss. A fluffy white bunny with twitching pink ears, a ginger kitten with bright green eyes, and a fluffy yellow chick, all lost in pure joy. In a captivatingly out-of-sync rhythm, they begin to shake their bodies; the bunny vigorously exhales a fine spray of dew from its fur, the kitten playfully wiggles from head to tail, and the chick adorably fluffs its feathers in a rapid, quivering throb. The magical golden light catches the edges of their soft fur and down, creating a gentle halo effect that casts long, soft shadows and creates a deeply serene and warm atmosphere of pure, innocent magic.", + "seed": 9324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9324_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1963499908684753, + "vlm_logprob": -0.1963499908684753, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.1963499908684753, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7068797945976257 + ], + "opposite_logprob_tokens": [ + -0.7068797945976257 + ], + "logprob_reward": -0.7068797945976257, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7068797945976257, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06790966540575027, + 0.03989516571164131, + 0.03850366547703743, + 0.11191925406455994, + 0.07988794147968292, + 0.08744248747825623, + 0.06397131830453873, + 0.10353376716375351, + 0.14539892971515656, + 0.2615378499031067 + ], + "credit_map": [ + 0.0046351137571036816, + 0.012366271577775478, + 0.005814067088067532, + 0.008191040717065334, + 0.008545765653252602, + 0.013171929866075516, + 0.009575716219842434, + 0.005609759595245123, + 0.003040463663637638, + 0.006477838847786188, + 0.004129608161747456, + 0.004060911014676094, + 0.004693071823567152, + 0.00810784287750721, + 0.005548608489334583, + 0.0038368182722479105, + 0.0033379129599779844, + 0.009154265746474266, + 0.0034628731664270163, + 0.0051863486878573895, + 0.004500859417021275, + 0.005718073341995478, + 0.004576509352773428, + 0.0025668214075267315, + 0.01013505831360817, + 0.01670539379119873, + 0.013525540009140968, + 0.011995719745755196, + 0.012630987912416458, + 0.02018948271870613, + 0.01575932465493679, + 0.010977738536894321, + 0.009537632577121258, + 0.011329751461744308, + 0.017367342486977577, + 0.015152917243540287, + 0.0061875334940850735, + 0.008510603569447994, + 0.007319662254303694, + 0.0044824955984950066, + 0.009771629236638546, + 0.00984949804842472, + 0.009643360041081905, + 0.013647741638123989, + 0.011919112876057625, + 0.015504627488553524, + 0.009581298567354679, + 0.007525212597101927, + 0.0076637184247374535, + 0.0074660819955170155, + 0.010599769651889801, + 0.007984631694853306, + 0.005822130478918552, + 0.009844498708844185, + 0.008116387762129307, + 0.006474099587649107, + 0.004479382187128067, + 0.0060041784308850765, + 0.003575690556317568, + 0.002698667114600539, + 0.03987983986735344, + 0.040810342878103256, + 0.003526266198605299, + 0.002559386193752289, + 0.02812609076499939, + 0.017552107572555542, + 0.03397631272673607, + 0.011692152358591557, + 0.014865033328533173, + 0.01653343252837658, + 0.012950686737895012, + 0.00970310065895319, + 0.00551202055066824, + 0.006926465779542923, + 0.0035807101521641016, + 0.004581824876368046, + 0.11215826123952866, + 0.11525394022464752, + 0.008310549892485142, + 0.005214071832597256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.7068797945976257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521214783191681 + ], + "opposite_logprob_tokens": [ + -1.5021214485168457 + ], + "logprob_reward": -0.2521214783191681, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521214783191681, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23625034093856812, + 0.04825641214847565, + 0.020506272092461586, + 0.024035532027482986, + 0.03404800966382027, + 0.16327211260795593, + 0.07653592526912689, + 0.298500120639801, + 0.028927626088261604, + 0.06966757029294968 + ], + "credit_map": [ + 0.02123315818607807, + 0.036867283284664154, + 0.02765893004834652, + 0.027128851041197777, + 0.026269059628248215, + 0.041821617633104324, + 0.031516704708337784, + 0.023754775524139404, + 0.004276154097169638, + 0.010115141049027443, + 0.004602506291121244, + 0.004432199988514185, + 0.005316876340657473, + 0.007353161461651325, + 0.006652200594544411, + 0.00550817558541894, + 0.001306575140915811, + 0.004759905859827995, + 0.001689956639893353, + 0.0018924684263765812, + 0.0025315829552710056, + 0.004249188583344221, + 0.002015316393226385, + 0.0020612794905900955, + 0.002226670039817691, + 0.003511489136144519, + 0.00253406073898077, + 0.003557987278327346, + 0.0033724824897944927, + 0.00369689567014575, + 0.0028340108692646027, + 0.002301938598975539, + 0.0038775948341935873, + 0.0058843945153057575, + 0.002802216913551092, + 0.003814256750047207, + 0.003939686808735132, + 0.00794371496886015, + 0.0029054787009954453, + 0.002880669431760907, + 0.01419845037162304, + 0.014529683627188206, + 0.019133031368255615, + 0.014423871412873268, + 0.03701437637209892, + 0.032903578132390976, + 0.022356338798999786, + 0.008712801150977612, + 0.008633680641651154, + 0.009465024806559086, + 0.006720591336488724, + 0.007966897450387478, + 0.010232308879494667, + 0.014507538639008999, + 0.011498472653329372, + 0.0075114136561751366, + 0.01095829252153635, + 0.015579159371554852, + 0.008461047895252705, + 0.008534676395356655, + 0.11671555042266846, + 0.11961351335048676, + 0.012545297853648663, + 0.0060925837606191635, + 0.002217361470684409, + 0.0015259634237736464, + 0.002319542458280921, + 0.0017668848158791661, + 0.008256484754383564, + 0.009004605934023857, + 0.0021551132667809725, + 0.0016816722927615047, + 0.0023382031358778477, + 0.0021349445451050997, + 0.002495512133464217, + 0.003017779905349016, + 0.02634071744978428, + 0.026825128123164177, + 0.004337071906775236, + 0.002178218914195895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.2521214783191681 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008204193436540663 + ], + "opposite_logprob_tokens": [ + -7.125820636749268 + ], + "logprob_reward": -0.0008204193436540663, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008204193436540663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10998506844043732, + 0.02379128336906433, + 0.07195605337619781, + 0.08041241019964218, + 0.09267350286245346, + 0.06987713277339935, + 0.10442706197500229, + 0.18224160373210907, + 0.05118389055132866, + 0.21345195174217224 + ], + "credit_map": [ + 0.008240323513746262, + 0.015030909329652786, + 0.009122476913034916, + 0.009933234192430973, + 0.0148809514939785, + 0.027062028646469116, + 0.015458259731531143, + 0.01025688648223877, + 0.0016914482694119215, + 0.0029861589428037405, + 0.0023242617025971413, + 0.0034331742208451033, + 0.0029287640936672688, + 0.004145072307437658, + 0.004140425939112902, + 0.0021419809199869633, + 0.00555867375805974, + 0.012304585427045822, + 0.0072451988235116005, + 0.008980581536889076, + 0.007514720782637596, + 0.01434027124196291, + 0.00923134759068489, + 0.0067806849256157875, + 0.006613654550164938, + 0.010478006675839424, + 0.009117808192968369, + 0.009788691066205502, + 0.009175810031592846, + 0.015834202989935875, + 0.011244114488363266, + 0.008160131052136421, + 0.009909669868648052, + 0.012732653878629208, + 0.01827055960893631, + 0.0153889125213027, + 0.008162056095898151, + 0.010237278416752815, + 0.010680020786821842, + 0.00729236239567399, + 0.00738842086866498, + 0.010608658194541931, + 0.006608137860894203, + 0.006947483401745558, + 0.013539042323827744, + 0.012667205184698105, + 0.007738799322396517, + 0.0043793972581624985, + 0.013983052223920822, + 0.0117555633187294, + 0.01301357802003622, + 0.013709207996726036, + 0.011028716340661049, + 0.013484599068760872, + 0.01880907453596592, + 0.008643289096653461, + 0.0069887880235910416, + 0.007497868034988642, + 0.00693547073751688, + 0.006540903355926275, + 0.06735188513994217, + 0.06883080303668976, + 0.01360970176756382, + 0.004486187361180782, + 0.009220748208463192, + 0.005635411478579044, + 0.011384891346096992, + 0.004361662548035383, + 0.00625350559130311, + 0.007103842217475176, + 0.003651566803455353, + 0.003572260495275259, + 0.008442490361630917, + 0.008657360449433327, + 0.008512590080499649, + 0.012192158959805965, + 0.05711342766880989, + 0.0646328255534172, + 0.04334365576505661, + 0.010557468980550766 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0008204193436540663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13221018016338348, + 0.060857292264699936, + 0.07565464824438095, + 0.041566137224435806, + 0.06261100620031357, + 0.089585080742836, + 0.05676460638642311, + 0.06564731895923615, + 0.2390911877155304, + 0.17601248621940613 + ], + "credit_map": [ + 0.009084303863346577, + 0.019622022286057472, + 0.013603425584733486, + 0.01637016050517559, + 0.016627850010991096, + 0.02145574800670147, + 0.023256385698914528, + 0.012190301902592182, + 0.004048496019095182, + 0.008463810198009014, + 0.00669515784829855, + 0.006484886631369591, + 0.006920232903212309, + 0.011343788355588913, + 0.009096789173781872, + 0.007804131601005793, + 0.0053522950038313866, + 0.013362443074584007, + 0.007217261474579573, + 0.0103297783061862, + 0.009091239422559738, + 0.014736692421138287, + 0.009202836081385612, + 0.00636210385710001, + 0.004003632813692093, + 0.009633935987949371, + 0.00389310740865767, + 0.00407533859834075, + 0.005637119058519602, + 0.007008794695138931, + 0.004191590938717127, + 0.0031226191204041243, + 0.008976468816399574, + 0.008976392447948456, + 0.016390418633818626, + 0.007026494946330786, + 0.0067445519380271435, + 0.006468807812780142, + 0.004173749592155218, + 0.0038541252724826336, + 0.01051761582493782, + 0.009044794365763664, + 0.0071719251573085785, + 0.008067835122346878, + 0.016718385741114616, + 0.019592994824051857, + 0.011509940028190613, + 0.006961597595363855, + 0.007966799661517143, + 0.007200227119028568, + 0.008071945048868656, + 0.00737948901951313, + 0.005294691771268845, + 0.0075031486339867115, + 0.009738016873598099, + 0.0036102919839322567, + 0.004671146161854267, + 0.005707674194127321, + 0.004159136675298214, + 0.005859730299562216, + 0.014772986061871052, + 0.014886111952364445, + 0.011427033692598343, + 0.004163503646850586, + 0.04268456622958183, + 0.02973247319459915, + 0.05103954300284386, + 0.017941780388355255, + 0.029044384136795998, + 0.030591653659939766, + 0.022197477519512177, + 0.01585930958390236, + 0.005962833762168884, + 0.0056863101199269295, + 0.0037946519441902637, + 0.005781229119747877, + 0.0710015743970871, + 0.07256536185741425, + 0.007484119851142168, + 0.0037364026065915823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001335444045253098 + ], + "opposite_logprob_tokens": [ + -6.626335620880127 + ], + "logprob_reward": -0.001335444045253098, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.001335444045253098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16931118071079254, + 0.04741334915161133, + 0.06657444685697556, + 0.05151001363992691, + 0.13165222108364105, + 0.0921216607093811, + 0.07936329394578934, + 0.042170580476522446, + 0.13330861926078796, + 0.18657465279102325 + ], + "credit_map": [ + 0.011946414597332478, + 0.027121812105178833, + 0.017303405329585075, + 0.02045157551765442, + 0.020832562819123268, + 0.03345056623220444, + 0.023147964850068092, + 0.015056862495839596, + 0.00226981402374804, + 0.008252314291894436, + 0.004560641944408417, + 0.008708443492650986, + 0.0044556837528944016, + 0.006736093200743198, + 0.008406487293541431, + 0.0040238723158836365, + 0.005080313887447119, + 0.012212383560836315, + 0.006674287375062704, + 0.009515141136944294, + 0.007245287299156189, + 0.012776725925505161, + 0.007554388605058193, + 0.005515928380191326, + 0.0058702994138002396, + 0.006851318757981062, + 0.00612689508125186, + 0.004864232614636421, + 0.007821595296263695, + 0.008727507665753365, + 0.006280818022787571, + 0.0049673463217914104, + 0.019695693626999855, + 0.013075511902570724, + 0.02032659761607647, + 0.00934567116200924, + 0.01669851318001747, + 0.03254654258489609, + 0.00965395849198103, + 0.010309724137187004, + 0.009805790148675442, + 0.011072066612541676, + 0.012249412946403027, + 0.007436962332576513, + 0.017219070345163345, + 0.02043987810611725, + 0.007634466513991356, + 0.0062640076503157616, + 0.011732198297977448, + 0.010843484662473202, + 0.007929270155727863, + 0.00936959870159626, + 0.006654045544564724, + 0.01239662617444992, + 0.010848978534340858, + 0.00958909559994936, + 0.0019684366416186094, + 0.0028490356635302305, + 0.002295664744451642, + 0.0018794034840539098, + 0.014695737510919571, + 0.015116753987967968, + 0.0018480313010513783, + 0.0015175212174654007, + 0.024145785719156265, + 0.015399559400975704, + 0.030719442293047905, + 0.010930136777460575, + 0.01578926295042038, + 0.01675395667552948, + 0.010515235364437103, + 0.00905522983521223, + 0.0075574349611997604, + 0.00872486736625433, + 0.006474665366113186, + 0.007035250309854746, + 0.06641534715890884, + 0.07000462710857391, + 0.01380536425858736, + 0.006557097192853689 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.001335444045253098 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "camera lens, held at eye level, slowly pans forward through a dreamy forest scene bathed in a warm, ethereal glow. In the background, a soft golden light pulses rhythmically, illuminating lazy white smoke drifting across a soft-focus backdrop of ancient trees. As the camera draws closer, a delightful group of small animals gathers on a patch of lush green moss. A fluffy white bunny with twitching pink ears, a ginger kitten with bright green eyes, and a fluffy yellow chick, all lost in pure joy. In a captivatingly out-of-sync rhythm, they begin to shake their bodies; the bunny vigorously exhales a fine spray of dew from its fur, the kitten playfully wiggles from head to tail, and the chick adorably fluffs its feathers in a rapid, quivering throb. The magical golden light catches the edges of their soft fur and down, creating a gentle halo effect that casts long, soft shadows and creates a deeply serene and warm atmosphere of pure, innocent magic.", + "seed": 9325, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_13/index_989_seed_9325_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11283859576797113, + "vlm_logprob": -0.11283859576797113, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.11283859576797113, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.484413743019104 + ], + "opposite_logprob_tokens": [ + -0.984413743019104 + ], + "logprob_reward": -0.484413743019104, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.484413743019104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1598413735628128, + 0.12497896701097488, + 0.23236525058746338, + 0.11149641871452332, + 0.03775053843855858, + 0.04063444957137108, + 0.09527769684791565, + 0.05143999308347702, + 0.10921285301446915, + 0.03700244426727295 + ], + "credit_map": [ + 0.013212516903877258, + 0.017237750813364983, + 0.013132438994944096, + 0.01367223635315895, + 0.02864808402955532, + 0.03293003514409065, + 0.019693648442626, + 0.02131466753780842, + 0.009958594106137753, + 0.024920396506786346, + 0.010912948288023472, + 0.0147643331438303, + 0.025616200640797615, + 0.01807536371052265, + 0.010756053030490875, + 0.0099750692024827, + 0.023335814476013184, + 0.02793431282043457, + 0.040529899299144745, + 0.023610977455973625, + 0.036451034247875214, + 0.0430617593228817, + 0.017625339329242706, + 0.01981610246002674, + 0.01592911034822464, + 0.012485045939683914, + 0.008548758924007416, + 0.011404073797166348, + 0.019355563446879387, + 0.014625512063503265, + 0.015581626445055008, + 0.013566727750003338, + 0.0032367375679314137, + 0.003973618149757385, + 0.0033527619671076536, + 0.0050813364796340466, + 0.00651306239888072, + 0.005503159016370773, + 0.004372058901935816, + 0.005717802327126265, + 0.00670798122882843, + 0.008955910801887512, + 0.003944293595850468, + 0.0035975694190710783, + 0.004826107062399387, + 0.004987460095435381, + 0.0033388989977538586, + 0.004276226740330458, + 0.007314293645322323, + 0.010222761891782284, + 0.00847044587135315, + 0.008647166192531586, + 0.02646814100444317, + 0.018371842801570892, + 0.009226946160197258, + 0.006556095089763403, + 0.00808787066489458, + 0.005943691823631525, + 0.007745977491140366, + 0.005259225610643625, + 0.006921527441591024, + 0.006155924405902624, + 0.00537334056571126, + 0.0059524341486394405, + 0.006543589755892754, + 0.04234038293361664, + 0.0076243518851697445, + 0.00743387034162879, + 0.014813700690865517, + 0.005525460001081228, + 0.005323647987097502, + 0.01960785500705242, + 0.005001100245863199, + 0.0064023942686617374, + 0.0032295440323650837, + 0.0034536852035671473, + 0.004368974361568689, + 0.00427114637568593, + 0.0038084255065768957, + 0.006467175204306841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.484413743019104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497325211763382 + ], + "opposite_logprob_tokens": [ + -2.9299733638763428 + ], + "logprob_reward": -0.05497325211763382, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497325211763382, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2579577565193176, + 0.10874824225902557, + 0.03404476121068001, + 0.11186757683753967, + 0.028859209269285202, + 0.03652041405439377, + 0.06201380491256714, + 0.03716953471302986, + 0.29861733317375183, + 0.02420133724808693 + ], + "credit_map": [ + 0.024992739781737328, + 0.019545163959264755, + 0.03173879161477089, + 0.01690317504107952, + 0.03842772915959358, + 0.04684200882911682, + 0.030524831265211105, + 0.04898332431912422, + 0.011356192640960217, + 0.01472598034888506, + 0.010355250909924507, + 0.010948949493467808, + 0.014614136889576912, + 0.01257743127644062, + 0.015910211950540543, + 0.01826009713113308, + 0.003715187544003129, + 0.004093310330063105, + 0.003075480693951249, + 0.0030688343103975058, + 0.0034182108938694, + 0.004064793232828379, + 0.004547921475023031, + 0.008061024360358715, + 0.019631339237093925, + 0.011468078941106796, + 0.009648391976952553, + 0.010262276977300644, + 0.01431330107152462, + 0.014078580774366856, + 0.01780313439667225, + 0.014662479981780052, + 0.002180386334657669, + 0.0029226846527308226, + 0.002253845566883683, + 0.003123586531728506, + 0.005697799380868673, + 0.004049048293381929, + 0.0037689097225666046, + 0.004862951580435038, + 0.004844102542847395, + 0.005105362739413977, + 0.004507832694798708, + 0.0030206802766770124, + 0.006238928064703941, + 0.0045935530215501785, + 0.0032914914190769196, + 0.004918466787785292, + 0.005455495323985815, + 0.005500665865838528, + 0.007209098897874355, + 0.004071536939591169, + 0.01699652522802353, + 0.01147522870451212, + 0.005926866549998522, + 0.005378394387662411, + 0.004961326718330383, + 0.004095913376659155, + 0.005709596909582615, + 0.003424607217311859, + 0.006467627827078104, + 0.00441804388538003, + 0.0035050446167588234, + 0.004587376024574041, + 0.010890944860875607, + 0.12748320400714874, + 0.0070468648336827755, + 0.014031377620995045, + 0.033490631729364395, + 0.01434649433940649, + 0.01154811680316925, + 0.0797797366976738, + 0.002198385540395975, + 0.0037166401743888855, + 0.0039019701071083546, + 0.002140222815796733, + 0.0032844373490661383, + 0.005686490796506405, + 0.0017403924139216542, + 0.0015328001463785768 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.05497325211763382 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00012063252506777644 + ], + "opposite_logprob_tokens": [ + -9.125120162963867 + ], + "logprob_reward": -0.00012063252506777644, + "yes_no_margin_reward": 9.125, + "token_credit_reward": -0.00012063252506777644, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1963941603899002, + 0.1110035628080368, + 0.12832039594650269, + 0.04488803446292877, + 0.02948838472366333, + 0.11109708249568939, + 0.09108544141054153, + 0.03470411151647568, + 0.20936886966228485, + 0.043649934232234955 + ], + "credit_map": [ + 0.015444567427039146, + 0.018020469695329666, + 0.018646331503987312, + 0.011968035250902176, + 0.03263331949710846, + 0.03563753142952919, + 0.0218861922621727, + 0.04215770587325096, + 0.008700812235474586, + 0.013365508988499641, + 0.010507927276194096, + 0.010588400065898895, + 0.028349358588457108, + 0.01653466187417507, + 0.013888767920434475, + 0.009068124927580357, + 0.012854394502937794, + 0.015687575563788414, + 0.01957845129072666, + 0.01270765159279108, + 0.016439788043498993, + 0.020167455077171326, + 0.011446448974311352, + 0.01943862810730934, + 0.0037612607702612877, + 0.0063605583272874355, + 0.005410917103290558, + 0.005584409460425377, + 0.007829069159924984, + 0.005603800527751446, + 0.005183680448681116, + 0.005154337268322706, + 0.0024978185538202524, + 0.004337918944656849, + 0.003333787899464369, + 0.002875483827665448, + 0.004673073999583721, + 0.0038382739294320345, + 0.003296848852187395, + 0.004635178949683905, + 0.02894027903676033, + 0.016542447730898857, + 0.012103050015866756, + 0.009590818546712399, + 0.0128935556858778, + 0.014883694238960743, + 0.00788114219903946, + 0.008262095041573048, + 0.007218279875814915, + 0.015159402042627335, + 0.012484168633818626, + 0.00919424556195736, + 0.02118806540966034, + 0.00997097697108984, + 0.009103409014642239, + 0.006766891572624445, + 0.005491266958415508, + 0.0037290435284376144, + 0.004626115784049034, + 0.003689019475132227, + 0.005090813152492046, + 0.004214930813759565, + 0.0033655394800007343, + 0.004497383255511522, + 0.009529775008559227, + 0.08007373660802841, + 0.01158198993653059, + 0.012206322513520718, + 0.02210499718785286, + 0.012484516017138958, + 0.010142939165234566, + 0.0512445829808712, + 0.005004054866731167, + 0.0076064360328018665, + 0.00648641400039196, + 0.00402517756447196, + 0.0060354662127792835, + 0.00595493521541357, + 0.0036813910119235516, + 0.00485606025904417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.00012063252506777644 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020592818036675453 + ], + "opposite_logprob_tokens": [ + -3.8955929279327393 + ], + "logprob_reward": -0.020592818036675453, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020592818036675453, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14062204957008362, + 0.07617901265621185, + 0.1310950517654419, + 0.10250293463468552, + 0.04684239625930786, + 0.09171479940414429, + 0.23046797513961792, + 0.03800292685627937, + 0.1057194247841835, + 0.036853380501270294 + ], + "credit_map": [ + 0.010943993926048279, + 0.015666242688894272, + 0.011573785915970802, + 0.011534339748322964, + 0.021016402170062065, + 0.022393593564629555, + 0.02877936139702797, + 0.018714332953095436, + 0.0063676354475319386, + 0.013100706972181797, + 0.006531625520437956, + 0.009294836781919003, + 0.017748093232512474, + 0.011031181551516056, + 0.007385174743831158, + 0.0047197588719427586, + 0.01250419020652771, + 0.01907574012875557, + 0.018462445586919785, + 0.014369459822773933, + 0.018954936414957047, + 0.023779336363077164, + 0.011756650172173977, + 0.01219228096306324, + 0.008801385760307312, + 0.01288569439202547, + 0.009230691008269787, + 0.011144953779876232, + 0.024089714512228966, + 0.014432746917009354, + 0.012911081314086914, + 0.009006669744849205, + 0.0044984095729887486, + 0.0063299937173724174, + 0.005324205383658409, + 0.00547382514923811, + 0.007235857658088207, + 0.0066002532839775085, + 0.005817296449095011, + 0.005562555510550737, + 0.021563047543168068, + 0.013032305985689163, + 0.00931557361036539, + 0.007241035811603069, + 0.010864774696528912, + 0.013682982884347439, + 0.008129294030368328, + 0.007885782979428768, + 0.02127205580472946, + 0.02593577653169632, + 0.028179189190268517, + 0.029804058372974396, + 0.05200357735157013, + 0.035196784883737564, + 0.02120913937687874, + 0.016867388039827347, + 0.007265916559845209, + 0.004031078424304724, + 0.007564873434603214, + 0.0032237509731203318, + 0.004780907183885574, + 0.0038095167838037014, + 0.0038946326822042465, + 0.003432252909988165, + 0.005303933285176754, + 0.04208339750766754, + 0.0037748615723103285, + 0.0055648330599069595, + 0.01643231138586998, + 0.006825304590165615, + 0.004709742963314056, + 0.021025046706199646, + 0.004109826870262623, + 0.005205500405281782, + 0.005554285831749439, + 0.004713051021099091, + 0.004626182373613119, + 0.006078333593904972, + 0.0035128742456436157, + 0.0030533261597156525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.020592818036675453 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the camera slowly pan forward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene a forest with ancient trees and lush green moss?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are a fluffy white bunny, a ginger kitten, and a fluffy yellow chick visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the animals perform specific actions: the bunny exhaling dew, the kitten wiggling, and the chick fluffing its feathers?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a soft golden light creating a halo effect on the animals' fur?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17757759988307953, + 0.10017415881156921, + 0.09948885440826416, + 0.10726538300514221, + 0.021831730380654335, + 0.11553177982568741, + 0.037685006856918335, + 0.03607379272580147, + 0.22216473519802094, + 0.08220694959163666 + ], + "credit_map": [ + 0.013061098754405975, + 0.015470383688807487, + 0.014631162397563457, + 0.015225527808070183, + 0.03545562922954559, + 0.02518681064248085, + 0.028147755190730095, + 0.03039923869073391, + 0.010572408325970173, + 0.014708762988448143, + 0.009920824319124222, + 0.011227966286242008, + 0.013639470562338829, + 0.013808246701955795, + 0.013001568615436554, + 0.013294919393956661, + 0.010294409468770027, + 0.011707666330039501, + 0.019035015255212784, + 0.00976291298866272, + 0.013668757863342762, + 0.017095496878027916, + 0.007888520136475563, + 0.01003606803715229, + 0.013619305565953255, + 0.012087532319128513, + 0.008095547556877136, + 0.0100076450034976, + 0.020800935104489326, + 0.019923841580748558, + 0.013592804782092571, + 0.009137768298387527, + 0.00223210034891963, + 0.002731706714257598, + 0.0024115389678627253, + 0.0030741398222744465, + 0.0023690371308475733, + 0.0025865500792860985, + 0.0029141930863261223, + 0.0035124635323882103, + 0.01571749523282051, + 0.01466525997966528, + 0.008787372149527073, + 0.013861026614904404, + 0.018855635076761246, + 0.01720966026186943, + 0.013078803196549416, + 0.013356529176235199, + 0.003937636502087116, + 0.006096672732383013, + 0.00573333352804184, + 0.004588649608194828, + 0.00426348764449358, + 0.0047769672237336636, + 0.003451181575655937, + 0.004837078507989645, + 0.005940363742411137, + 0.004378600977361202, + 0.005822670180350542, + 0.004431786946952343, + 0.00426137913018465, + 0.00336557999253273, + 0.0035232482478022575, + 0.004350163973867893, + 0.007193347439169884, + 0.09732623398303986, + 0.004809285048395395, + 0.010706041008234024, + 0.029496679082512856, + 0.010181237012147903, + 0.008234295062720776, + 0.05421760305762291, + 0.014138116501271725, + 0.011528472416102886, + 0.010058101266622543, + 0.0066095865331590176, + 0.011106192134320736, + 0.01809569261968136, + 0.005274822004139423, + 0.005395969375967979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.004092533141374588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_tensor([576, 576])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_tensor([576, 576])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..92f621bf4bb7c8516ffaa3e07affa1790933dd84 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_tensor([576, 576])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, low-angle tracking shot glides smoothly above the ground, keeping pace with a lone skateboarder as they expertly navigate the vast, cavernous spaces of a dimly lit underground garage. The camera remains close to the subject, emphasizing their speed and skill. The skateboarder is a focused figure, dressed in a loose-fitting, dark gray hoodie with the hood pulled up, obscuring their features in shadow, and baggy black overalls. They maintain a low, balanced stance on their skateboard, with its visibly worn wooden deck and vibrant orange wheels, providing a striking pop of color in the quiet surroundings. With fluid precision, they carve a winding path between massive, rough-hewn concrete pillars that recede into the deep, shadowy background. The action is a seamless flow; they lean deeply, their bodies almost parallel to the ground, then effortlessly shift their weight to navigate around the next obstacle. The only sound is the consistent, hypnotic hum of polyurethane wheels rolling across the smooth, polished concrete floor, which is stained with light oil stains and faded yellow parking lines. The sparse, cool white light from flickering fluorescent tubes overhead illuminates the scene, casting long, dancing shadows that stretch and twist with the skateboarders’ movements, creating a stark, high-contrast visual palette and an atmosphere of solitary, focused freedom.", + "seed": 9611, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_576_seed_9611_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10869782292284072, + "vlm_logprob": -0.10869782292284072, + "vlm_yes_no_margin": 3.375000011920929, + "vlm_token_credit": -0.10869782292284072, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09133970737457275 + ], + "opposite_logprob_tokens": [ + -3.591339588165283 + ], + "logprob_reward": -0.09133970737457275, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.09133970737457275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19900090992450714, + 0.07270561158657074, + 0.049330923706293106, + 0.05490676686167717, + 0.05118434503674507, + 0.05015762895345688, + 0.1866951435804367, + 0.049641989171504974, + 0.10216812789440155, + 0.18420855700969696 + ], + "credit_map": [ + 0.028332984074950218, + 0.018955091014504433, + 0.03760303556919098, + 0.020686335861682892, + 0.031195715069770813, + 0.027425646781921387, + 0.015342793427407742, + 0.019459286704659462, + 0.007072200533002615, + 0.008618426509201527, + 0.016885068267583847, + 0.00965250376611948, + 0.007887860760092735, + 0.008562165312469006, + 0.007640837226063013, + 0.0063865454867482185, + 0.004727454390376806, + 0.005498587619513273, + 0.006374905351549387, + 0.009136387147009373, + 0.00531410239636898, + 0.0058214180171489716, + 0.005660625174641609, + 0.006797440350055695, + 0.006698234938085079, + 0.004839618224650621, + 0.006241058465093374, + 0.005878940224647522, + 0.010524418205022812, + 0.007060774136334658, + 0.00802721455693245, + 0.005636504385620356, + 0.007107993122190237, + 0.008113875053822994, + 0.006913181859999895, + 0.005176841747015715, + 0.008265797980129719, + 0.006254423409700394, + 0.004631165415048599, + 0.004721060395240784, + 0.006009602453559637, + 0.008617410436272621, + 0.007120028138160706, + 0.0065754251554608345, + 0.005683984607458115, + 0.006607614923268557, + 0.006075750570744276, + 0.0034678038209676743, + 0.016317442059516907, + 0.026753999292850494, + 0.032793596386909485, + 0.020925471559166908, + 0.02770083025097847, + 0.029542330652475357, + 0.01877274550497532, + 0.013888717629015446, + 0.005634633358567953, + 0.007095120847225189, + 0.005767793860286474, + 0.005151795223355293, + 0.005609140731394291, + 0.006443405523896217, + 0.0057023558765649796, + 0.008237737230956554, + 0.003420980414375663, + 0.008275335654616356, + 0.005487442947924137, + 0.012047315947711468, + 0.031558528542518616, + 0.028381222859025, + 0.008524023927748203, + 0.004473268054425716, + 0.012283369898796082, + 0.013622388243675232, + 0.010918015614151955, + 0.010644685477018356, + 0.016899913549423218, + 0.03766924887895584, + 0.031456876546144485, + 0.05071403831243515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.09133970737457275 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023308804258704185 + ], + "opposite_logprob_tokens": [ + -3.773308753967285 + ], + "logprob_reward": -0.023308804258704185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023308804258704185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.279289573431015, + 0.04230983555316925, + 0.03492717817425728, + 0.08103711158037186, + 0.043583113700151443, + 0.10899248719215393, + 0.11654633283615112, + 0.08947142958641052, + 0.14062076807022095, + 0.06322219222784042 + ], + "credit_map": [ + 0.03347644582390785, + 0.016520081087946892, + 0.07444382458925247, + 0.02464376576244831, + 0.02228812873363495, + 0.028863979503512383, + 0.0189084280282259, + 0.06014491617679596, + 0.004519065376371145, + 0.00470760790631175, + 0.0058829812332987785, + 0.006407246459275484, + 0.00693730590865016, + 0.0037430208176374435, + 0.0065783290192484856, + 0.003534275572746992, + 0.0035827315878123045, + 0.0022616658825427294, + 0.008001759648323059, + 0.007542987819761038, + 0.0034996718168258667, + 0.002761764219030738, + 0.003943100105971098, + 0.003333496628329158, + 0.010960112325847149, + 0.00682915048673749, + 0.012023056857287884, + 0.0064159599132835865, + 0.020480774343013763, + 0.009760753251612186, + 0.00903358869254589, + 0.005533715710043907, + 0.007445231545716524, + 0.006863675080239773, + 0.0037603897508233786, + 0.004528450779616833, + 0.008718503639101982, + 0.004042671527713537, + 0.005258956458419561, + 0.0029652337543666363, + 0.01019322406500578, + 0.02100576087832451, + 0.011480872519314289, + 0.010784052312374115, + 0.021691011264920235, + 0.013479750603437424, + 0.013519320636987686, + 0.006838483735918999, + 0.006446686573326588, + 0.013443768955767155, + 0.01173577643930912, + 0.011488465592265129, + 0.023513833060860634, + 0.018344663083553314, + 0.024230174720287323, + 0.007342962082475424, + 0.008433103561401367, + 0.013451593928039074, + 0.005380312446504831, + 0.011439227499067783, + 0.01313859410583973, + 0.011077061295509338, + 0.014308590441942215, + 0.012242943048477173, + 0.0034782656002789736, + 0.010126546956598759, + 0.004368479363620281, + 0.007332698442041874, + 0.05456603318452835, + 0.049374472349882126, + 0.006675221491605043, + 0.0046990495175123215, + 0.005511829163879156, + 0.008326455950737, + 0.004942078609019518, + 0.007550109177827835, + 0.011348772794008255, + 0.009720705449581146, + 0.007422338705509901, + 0.008399899117648602 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.023308804258704185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38711947202682495 + ], + "opposite_logprob_tokens": [ + -1.1371195316314697 + ], + "logprob_reward": -0.38711947202682495, + "yes_no_margin_reward": 0.7500000596046448, + "token_credit_reward": -0.38711947202682495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1544925570487976, + 0.24813710153102875, + 0.03541964292526245, + 0.03661368414759636, + 0.055655669420957565, + 0.08087055385112762, + 0.13364794850349426, + 0.10493361949920654, + 0.10089874267578125, + 0.04933049529790878 + ], + "credit_map": [ + 0.015901679173111916, + 0.010842213407158852, + 0.025244589895009995, + 0.013441805727779865, + 0.018206559121608734, + 0.037856679409742355, + 0.011950044892728329, + 0.021048961207270622, + 0.02642623335123062, + 0.028086477890610695, + 0.03263777494430542, + 0.025894103571772575, + 0.0462023988366127, + 0.019568344578146935, + 0.05214384198188782, + 0.017177924513816833, + 0.0029164371080696583, + 0.002565978327766061, + 0.003772668307647109, + 0.004280777648091316, + 0.0055190627463161945, + 0.005138331558555365, + 0.006354689598083496, + 0.00487169437110424, + 0.004523574374616146, + 0.003063047770410776, + 0.0034214998595416546, + 0.003774787997826934, + 0.007829428650438786, + 0.004516514949500561, + 0.006367812864482403, + 0.003117015352472663, + 0.007422343362122774, + 0.007896725088357925, + 0.0066839708015322685, + 0.00794257503002882, + 0.005554728209972382, + 0.007544028107076883, + 0.00746896443888545, + 0.005142333451658487, + 0.009100417606532574, + 0.016707438975572586, + 0.010994520969688892, + 0.007032510358840227, + 0.013101656921207905, + 0.01128979679197073, + 0.006868530996143818, + 0.005775675177574158, + 0.008859886787831783, + 0.016424288973212242, + 0.012471008114516735, + 0.013010682538151741, + 0.031440798193216324, + 0.02108851633965969, + 0.022343289107084274, + 0.008009470999240875, + 0.01183225680142641, + 0.014831287786364555, + 0.010618193075060844, + 0.015084635466337204, + 0.014811290428042412, + 0.014320479705929756, + 0.013370372354984283, + 0.010065094567835331, + 0.0037382093723863363, + 0.011168399825692177, + 0.003598907496780157, + 0.0053760637529194355, + 0.035640619695186615, + 0.030069313943386078, + 0.006748187821358442, + 0.004559028893709183, + 0.004424532875418663, + 0.004029758740216494, + 0.003381139598786831, + 0.006823420990258455, + 0.006943879183381796, + 0.011127372272312641, + 0.006095157004892826, + 0.006505231838673353 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.38711947202682495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036171742249280214 + ], + "opposite_logprob_tokens": [ + -5.628617286682129 + ], + "logprob_reward": -0.0036171742249280214, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036171742249280214, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3677906095981598, + 0.09082531183958054, + 0.03844461217522621, + 0.05121511593461037, + 0.05764492228627205, + 0.07287100702524185, + 0.05409299582242966, + 0.07007595896720886, + 0.09845270961523056, + 0.09858684241771698 + ], + "credit_map": [ + 0.040268827229738235, + 0.024910159409046173, + 0.11021614819765091, + 0.03268053010106087, + 0.027298638597130775, + 0.06026579067111015, + 0.020664799958467484, + 0.051485661417245865, + 0.010778190568089485, + 0.010684119537472725, + 0.012745941989123821, + 0.010094957426190376, + 0.017915774136781693, + 0.0069502596743404865, + 0.015933319926261902, + 0.005722736474126577, + 0.004469018895179033, + 0.0032574820797890425, + 0.006015799008309841, + 0.0051748184487223625, + 0.004586788360029459, + 0.004873289261013269, + 0.004754985217005014, + 0.0053124274127185345, + 0.00692725507542491, + 0.007017199881374836, + 0.006248872261494398, + 0.007794243283569813, + 0.00617242930456996, + 0.006981474347412586, + 0.0048527889885008335, + 0.005220853723585606, + 0.007340027950704098, + 0.01072404719889164, + 0.005884710233658552, + 0.0066027273423969746, + 0.00872818287461996, + 0.0073341140523552895, + 0.005946231074631214, + 0.005084876902401447, + 0.008780532516539097, + 0.015985213220119476, + 0.007523516193032265, + 0.010139702819287777, + 0.008941292762756348, + 0.009804931469261646, + 0.007708828430622816, + 0.003986984491348267, + 0.003942236769944429, + 0.006627746857702732, + 0.006662154104560614, + 0.0058862208388745785, + 0.01076341699808836, + 0.008916068822145462, + 0.006925071123987436, + 0.004370075184851885, + 0.007386062294244766, + 0.009773475117981434, + 0.007584040984511375, + 0.008919897489249706, + 0.012775847688317299, + 0.0076215630397200584, + 0.007213900797069073, + 0.008801170624792576, + 0.006419360172003508, + 0.015359954908490181, + 0.01094729546457529, + 0.009894002228975296, + 0.017162218689918518, + 0.01688172109425068, + 0.008537865243852139, + 0.013250294141471386, + 0.012574688531458378, + 0.010900430381298065, + 0.009208455681800842, + 0.006953367497771978, + 0.013725461438298225, + 0.01653546467423439, + 0.012016681022942066, + 0.01667228527367115 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.0036171742249280214 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810395672917366 + ], + "opposite_logprob_tokens": [ + -3.2881040573120117 + ], + "logprob_reward": -0.03810395672917366, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810395672917366, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09877476841211319, + 0.04967968165874481, + 0.03548908606171608, + 0.03919040039181709, + 0.04202313721179962, + 0.07989956438541412, + 0.06947939097881317, + 0.057294197380542755, + 0.2012988030910492, + 0.32687100768089294 + ], + "credit_map": [ + 0.010848275385797024, + 0.005736174993216991, + 0.025700034573674202, + 0.007732548285275698, + 0.007000401150435209, + 0.02714957296848297, + 0.006466608960181475, + 0.00814114697277546, + 0.005662239622324705, + 0.008416025899350643, + 0.007354080211371183, + 0.007189550902694464, + 0.006472188513725996, + 0.004388648085296154, + 0.007066470105201006, + 0.003130478784441948, + 0.0037879429291933775, + 0.0033583673648536205, + 0.004428592510521412, + 0.005731415469199419, + 0.003908001817762852, + 0.003821475664153695, + 0.005986062344163656, + 0.0044672260992228985, + 0.005120603367686272, + 0.004785387311130762, + 0.0059714606031775475, + 0.005517532583326101, + 0.004492767155170441, + 0.006042194552719593, + 0.0032166868913918734, + 0.004043764900416136, + 0.00590715603902936, + 0.005935750901699066, + 0.006052051670849323, + 0.005233694799244404, + 0.004526496399194002, + 0.006080613937228918, + 0.004708097316324711, + 0.003579276381060481, + 0.009658134542405605, + 0.018540574237704277, + 0.010344481095671654, + 0.007104614283889532, + 0.008173382841050625, + 0.01225235778838396, + 0.0064375135116279125, + 0.00738850561901927, + 0.006623898632824421, + 0.009784436784684658, + 0.006536848843097687, + 0.007932866923511028, + 0.014132651500403881, + 0.011613370850682259, + 0.008206081576645374, + 0.004649235401302576, + 0.007128002122044563, + 0.005609879270195961, + 0.007276441901922226, + 0.006296894978731871, + 0.00709945522248745, + 0.006393556483089924, + 0.006480950862169266, + 0.011009016074240208, + 0.006882449612021446, + 0.01147390715777874, + 0.01127992570400238, + 0.009554218500852585, + 0.06702803075313568, + 0.06538952887058258, + 0.011505457572638988, + 0.018185291439294815, + 0.023312591016292572, + 0.017379583790898323, + 0.02278410643339157, + 0.013112076558172703, + 0.03012816049158573, + 0.06715379655361176, + 0.05352623015642166, + 0.09947443753480911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.03810395672917366 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides smoothly above the ground, keeping pace with a lone skateboarder as they expertly navigate the vast, cavernous spaces of a dimly lit underground garage. The camera remains close to the subject, emphasizing their speed and skill. The skateboarder is a focused figure, dressed in a loose-fitting, dark gray hoodie with the hood pulled up, obscuring their features in shadow, and baggy black overalls. They maintain a low, balanced stance on their skateboard, with its visibly worn wooden deck and vibrant orange wheels, providing a striking pop of color in the quiet surroundings. With fluid precision, they carve a winding path between massive, rough-hewn concrete pillars that recede into the deep, shadowy background. The action is a seamless flow; they lean deeply, their bodies almost parallel to the ground, then effortlessly shift their weight to navigate around the next obstacle. The only sound is the consistent, hypnotic hum of polyurethane wheels rolling across the smooth, polished concrete floor, which is stained with light oil stains and faded yellow parking lines. The sparse, cool white light from flickering fluorescent tubes overhead illuminates the scene, casting long, dancing shadows that stretch and twist with the skateboarders’ movements, creating a stark, high-contrast visual palette and an atmosphere of solitary, focused freedom.", + "seed": 9612, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_576_seed_9612_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10066517665982247, + "vlm_logprob": -0.10066517665982247, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.10066517665982247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1045852079987526 + ], + "opposite_logprob_tokens": [ + -4.22958517074585 + ], + "logprob_reward": -0.1045852079987526, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.1045852079987526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16874440014362335, + 0.10725905001163483, + 0.20654846727848053, + 0.08706177771091461, + 0.13273823261260986, + 0.07473213225603104, + 0.05726794898509979, + 0.05367572605609894, + 0.03938542678952217, + 0.07258693128824234 + ], + "credit_map": [ + 0.017411403357982635, + 0.02258366532623768, + 0.0247691310942173, + 0.019933898001909256, + 0.02669261023402214, + 0.026208240538835526, + 0.012100105173885822, + 0.019045328721404076, + 0.011967208236455917, + 0.012997145764529705, + 0.010724710300564766, + 0.02264128252863884, + 0.015445448458194733, + 0.011116175912320614, + 0.013866178691387177, + 0.008500893600285053, + 0.02070361003279686, + 0.041556574404239655, + 0.055623818188905716, + 0.023444591090083122, + 0.017966095358133316, + 0.016591206192970276, + 0.014713259413838387, + 0.015949292108416557, + 0.010761972516775131, + 0.009272066876292229, + 0.017029426991939545, + 0.010105556808412075, + 0.008331967517733574, + 0.01747841015458107, + 0.006076015532016754, + 0.008006352931261063, + 0.012241069227457047, + 0.01911163702607155, + 0.024279121309518814, + 0.018809624016284943, + 0.016526050865650177, + 0.016562048345804214, + 0.01016149576753378, + 0.015047174878418446, + 0.008374286815524101, + 0.0075754206627607346, + 0.011356278322637081, + 0.009055579081177711, + 0.008377687074244022, + 0.01244663167744875, + 0.006161669734865427, + 0.01138457003980875, + 0.00809275358915329, + 0.0066282195039093494, + 0.007081006653606892, + 0.010622560977935791, + 0.003528759116306901, + 0.0032595449592918158, + 0.0066490196622908115, + 0.011406080797314644, + 0.00524386391043663, + 0.009579470381140709, + 0.006976882927119732, + 0.011288759298622608, + 0.00504936883226037, + 0.004148316569626331, + 0.00708055030554533, + 0.004308509640395641, + 0.0069495914503932, + 0.0026550909969955683, + 0.007859631441533566, + 0.004867514129728079, + 0.004915118217468262, + 0.003621306736022234, + 0.004177401307970285, + 0.0043397690169513226, + 0.010182268917560577, + 0.011512748897075653, + 0.01257841382175684, + 0.008279208093881607, + 0.00920804776251316, + 0.006189044564962387, + 0.009607604704797268, + 0.005029590334743261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.1045852079987526 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048689182847738266 + ], + "opposite_logprob_tokens": [ + -3.048689126968384 + ], + "logprob_reward": -0.048689182847738266, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048689182847738266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13342127203941345, + 0.11715800315141678, + 0.2821922302246094, + 0.12770359218120575, + 0.06435222923755646, + 0.07087080180644989, + 0.03803970292210579, + 0.05785440281033516, + 0.039977483451366425, + 0.06843026727437973 + ], + "credit_map": [ + 0.013192392885684967, + 0.01695910468697548, + 0.017028357833623886, + 0.020151864737272263, + 0.029003743082284927, + 0.014232402667403221, + 0.009419175796210766, + 0.01343423780053854, + 0.013285970315337181, + 0.0114750936627388, + 0.012958112172782421, + 0.024009402841329575, + 0.023013176396489143, + 0.010118488222360611, + 0.015919826924800873, + 0.00637793866917491, + 0.015752526000142097, + 0.04609047621488571, + 0.038978252559900284, + 0.031127087771892548, + 0.06451541930437088, + 0.028768355026841164, + 0.03636225685477257, + 0.02059783972799778, + 0.014287377707660198, + 0.010202077217400074, + 0.019740475341677666, + 0.010294180363416672, + 0.019057773053646088, + 0.024580230936408043, + 0.014153500087559223, + 0.015387976542115211, + 0.006256752647459507, + 0.008550981990993023, + 0.008734253235161304, + 0.009069550782442093, + 0.00879642553627491, + 0.007045223377645016, + 0.00817274209111929, + 0.007726301904767752, + 0.004998120944947004, + 0.007143151015043259, + 0.017520416527986526, + 0.007667139638215303, + 0.008805167861282825, + 0.01217874139547348, + 0.007587955798953772, + 0.004970106296241283, + 0.0048140957951545715, + 0.0041895522736012936, + 0.0044014896266162395, + 0.005311841610819101, + 0.005941662006080151, + 0.005122481379657984, + 0.0037917825393378735, + 0.00446679862216115, + 0.004173985682427883, + 0.00876594241708517, + 0.007302821613848209, + 0.012323703616857529, + 0.006488109938800335, + 0.004801873117685318, + 0.00949025433510542, + 0.004507709760218859, + 0.006865794770419598, + 0.003569605527445674, + 0.006040169857442379, + 0.004650686867535114, + 0.005915324203670025, + 0.0036811567842960358, + 0.0040267580188810825, + 0.005227991379797459, + 0.007792425807565451, + 0.00989355705678463, + 0.008492667227983475, + 0.008174571208655834, + 0.010600702837109566, + 0.009023518301546574, + 0.009515957906842232, + 0.004936872515827417 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.048689182847738266 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3135129511356354 + ], + "opposite_logprob_tokens": [ + -1.313512921333313 + ], + "logprob_reward": -0.3135129511356354, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3135129511356354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10358727723360062, + 0.08069399744272232, + 0.16865573823451996, + 0.1386824995279312, + 0.15001574158668518, + 0.23100024461746216, + 0.02732817828655243, + 0.028388477861881256, + 0.016959192231297493, + 0.054688695818185806 + ], + "credit_map": [ + 0.008766258135437965, + 0.019737184047698975, + 0.012909254990518093, + 0.018606139346957207, + 0.013741990551352501, + 0.012522134929895401, + 0.008610210381448269, + 0.008694101125001907, + 0.008927593939006329, + 0.009556788951158524, + 0.008541464805603027, + 0.01426803320646286, + 0.01283399946987629, + 0.008598394691944122, + 0.01134954672306776, + 0.006618180777877569, + 0.01150897890329361, + 0.03208378329873085, + 0.03623092547059059, + 0.017245210707187653, + 0.025726990774273872, + 0.011065744794905186, + 0.020835421979427338, + 0.013958669267594814, + 0.014621957205235958, + 0.010552081279456615, + 0.024878382682800293, + 0.011830216273665428, + 0.026928799226880074, + 0.02537623979151249, + 0.012587231583893299, + 0.011907593347132206, + 0.011930665001273155, + 0.019620444625616074, + 0.02197566069662571, + 0.018391355872154236, + 0.025393472984433174, + 0.013177218846976757, + 0.020969564095139503, + 0.018557339906692505, + 0.025869281962513924, + 0.023946426808834076, + 0.036205701529979706, + 0.03087923303246498, + 0.032049983739852905, + 0.030772391706705093, + 0.021855728700757027, + 0.029421484097838402, + 0.004498389083892107, + 0.003493008902296424, + 0.003884549718350172, + 0.004553766921162605, + 0.002510254504159093, + 0.0026552067138254642, + 0.0021932809613645077, + 0.0035397193860262632, + 0.0032197153195738792, + 0.006589026656001806, + 0.003061420749872923, + 0.00509387394413352, + 0.0031059652101248503, + 0.0032866757828742266, + 0.0021191046107560396, + 0.001912693027406931, + 0.0027304922696202993, + 0.0014767562970519066, + 0.003247719956561923, + 0.0016012771520763636, + 0.0022836849093437195, + 0.001954555045813322, + 0.001695880200713873, + 0.0019688280299305916, + 0.007349634077399969, + 0.00745792081579566, + 0.008565320633351803, + 0.0065730237402021885, + 0.006718114018440247, + 0.006914741825312376, + 0.007576281204819679, + 0.0035336592700332403 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.3135129511356354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735360249876976 + ], + "opposite_logprob_tokens": [ + -5.006735324859619 + ], + "logprob_reward": -0.006735360249876976, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735360249876976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21858182549476624, + 0.11222951859235764, + 0.11408720165491104, + 0.16540004312992096, + 0.07361991703510284, + 0.10817613452672958, + 0.05555494502186775, + 0.03987709805369377, + 0.0257786326110363, + 0.08669468760490417 + ], + "credit_map": [ + 0.020594550296664238, + 0.024723123759031296, + 0.030197028070688248, + 0.03876425698399544, + 0.03012900799512863, + 0.026484109461307526, + 0.018489239737391472, + 0.029200514778494835, + 0.011527699418365955, + 0.011592879891395569, + 0.010765805840492249, + 0.027231188490986824, + 0.015112915076315403, + 0.010707150213420391, + 0.0166834257543087, + 0.008608444593846798, + 0.012956935912370682, + 0.017276346683502197, + 0.013499066233634949, + 0.01613065041601658, + 0.014551635831594467, + 0.010476227849721909, + 0.01803448610007763, + 0.011161847971379757, + 0.023695437237620354, + 0.021334074437618256, + 0.026965990662574768, + 0.016735831275582314, + 0.020547879859805107, + 0.03198382258415222, + 0.009412751533091068, + 0.014724250882863998, + 0.00795790832489729, + 0.01088008750230074, + 0.011742113158106804, + 0.009698932059109211, + 0.008981947787106037, + 0.007989084348082542, + 0.007596900686621666, + 0.008772946894168854, + 0.012240997515618801, + 0.009732968173921108, + 0.030336113646626472, + 0.01137320976704359, + 0.009894776158034801, + 0.014458729885518551, + 0.009549614042043686, + 0.010589727200567722, + 0.007016122341156006, + 0.006857777014374733, + 0.008908100426197052, + 0.014008235186338425, + 0.0039293477311730385, + 0.0042286040261387825, + 0.004848258104175329, + 0.005758500192314386, + 0.005189628805965185, + 0.00814052950590849, + 0.005170510150492191, + 0.00890255905687809, + 0.0035184489097446203, + 0.0029843368101865053, + 0.003366017946973443, + 0.002605065004900098, + 0.0030702915973961353, + 0.0031612850725650787, + 0.002779806964099407, + 0.002898439532145858, + 0.0031264026183634996, + 0.0031844365876168013, + 0.0033566486090421677, + 0.004201321862637997, + 0.011351126246154308, + 0.01411238219588995, + 0.012758634984493256, + 0.010227849707007408, + 0.011152892373502254, + 0.01032997202128172, + 0.009545373730361462, + 0.0072164516896009445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006735360249876976 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803181067109108 + ], + "opposite_logprob_tokens": [ + -3.5298032760620117 + ], + "logprob_reward": -0.029803181067109108, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803181067109108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19068390130996704, + 0.1291167438030243, + 0.0842156857252121, + 0.08941299468278885, + 0.093702033162117, + 0.06634636968374252, + 0.17485299706459045, + 0.07506225258111954, + 0.02697017230093479, + 0.06963682174682617 + ], + "credit_map": [ + 0.023088455200195312, + 0.01781037449836731, + 0.022296270355582237, + 0.018966231495141983, + 0.02625393308699131, + 0.03576064854860306, + 0.019902881234884262, + 0.02660512924194336, + 0.014045697636902332, + 0.016345189884305, + 0.01256266888231039, + 0.0265911053866148, + 0.017154710367321968, + 0.01405488420277834, + 0.01787056401371956, + 0.010491922497749329, + 0.00983746163547039, + 0.01146094873547554, + 0.011735357344150543, + 0.011468356475234032, + 0.010454755276441574, + 0.00758599815890193, + 0.014799416065216064, + 0.006873397156596184, + 0.010507158003747463, + 0.007460957393050194, + 0.014800118282437325, + 0.007321479264646769, + 0.012620718218386173, + 0.015529884025454521, + 0.008928162045776844, + 0.012244527228176594, + 0.009305205196142197, + 0.014367908239364624, + 0.017468033358454704, + 0.012781084515154362, + 0.011691723950207233, + 0.010699267499148846, + 0.007452412508428097, + 0.009936406277120113, + 0.008776369504630566, + 0.006959112826734781, + 0.01550136785954237, + 0.0063130478374660015, + 0.00818206463009119, + 0.006847131531685591, + 0.00757680693641305, + 0.006190477404743433, + 0.027248192578554153, + 0.02854502759873867, + 0.024124789983034134, + 0.029733162373304367, + 0.013064108788967133, + 0.01386871375143528, + 0.013930491171777248, + 0.024338509887456894, + 0.005785398185253143, + 0.012166974134743214, + 0.010371606796979904, + 0.019524510949850082, + 0.0057298047468066216, + 0.009871809743344784, + 0.0065250881016254425, + 0.0050870636478066444, + 0.0038190283812582493, + 0.002996427472680807, + 0.003610430285334587, + 0.002886021975427866, + 0.002699383767321706, + 0.0038989889435470104, + 0.0030491608195006847, + 0.0040107304230332375, + 0.009413771331310272, + 0.009587142616510391, + 0.012974568642675877, + 0.007772373966872692, + 0.006164115853607655, + 0.010712040588259697, + 0.006771272048354149, + 0.006241540424525738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.029803181067109108 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides smoothly above the ground, keeping pace with a lone skateboarder as they expertly navigate the vast, cavernous spaces of a dimly lit underground garage. The camera remains close to the subject, emphasizing their speed and skill. The skateboarder is a focused figure, dressed in a loose-fitting, dark gray hoodie with the hood pulled up, obscuring their features in shadow, and baggy black overalls. They maintain a low, balanced stance on their skateboard, with its visibly worn wooden deck and vibrant orange wheels, providing a striking pop of color in the quiet surroundings. With fluid precision, they carve a winding path between massive, rough-hewn concrete pillars that recede into the deep, shadowy background. The action is a seamless flow; they lean deeply, their bodies almost parallel to the ground, then effortlessly shift their weight to navigate around the next obstacle. The only sound is the consistent, hypnotic hum of polyurethane wheels rolling across the smooth, polished concrete floor, which is stained with light oil stains and faded yellow parking lines. The sparse, cool white light from flickering fluorescent tubes overhead illuminates the scene, casting long, dancing shadows that stretch and twist with the skateboarders’ movements, creating a stark, high-contrast visual palette and an atmosphere of solitary, focused freedom.", + "seed": 9613, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_576_seed_9613_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07825994389131666, + "vlm_logprob": -0.07825994389131666, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.07825994389131666, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05546237528324127 + ], + "opposite_logprob_tokens": [ + -4.43046236038208 + ], + "logprob_reward": -0.05546237528324127, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.05546237528324127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3960094749927521, + 0.09588290005922318, + 0.10937566310167313, + 0.05826699733734131, + 0.03920936957001686, + 0.06134567782282829, + 0.05803126469254494, + 0.03849940001964569, + 0.05373818427324295, + 0.08964104950428009 + ], + "credit_map": [ + 0.02257482334971428, + 0.029508911073207855, + 0.023903738707304, + 0.09181976318359375, + 0.042343977838754654, + 0.1126389130949974, + 0.04119536280632019, + 0.032023992389440536, + 0.010635320097208023, + 0.005631784442812204, + 0.02373087964951992, + 0.005468487739562988, + 0.009943167679011822, + 0.009878451004624367, + 0.013264493085443974, + 0.017330320551991463, + 0.011157593689858913, + 0.016224196180701256, + 0.012458965182304382, + 0.010625950060784817, + 0.023553967475891113, + 0.016005858778953552, + 0.011546803638339043, + 0.00780233321711421, + 0.0052079204469919205, + 0.009470295161008835, + 0.008692336268723011, + 0.006180890370160341, + 0.009822217747569084, + 0.005372520536184311, + 0.008745028637349606, + 0.004775789100676775, + 0.002876878948882222, + 0.0049907853826880455, + 0.004120701923966408, + 0.004863079637289047, + 0.008427257649600506, + 0.006101455073803663, + 0.004591363947838545, + 0.0032378467731177807, + 0.004851111676543951, + 0.005761526525020599, + 0.006826743949204683, + 0.01138733234256506, + 0.008223545737564564, + 0.008633758872747421, + 0.008384870365262032, + 0.007276786491274834, + 0.00578879751265049, + 0.006919885519891977, + 0.005878683179616928, + 0.010254893451929092, + 0.00954425148665905, + 0.0035650753416121006, + 0.0107937166467309, + 0.005285961553454399, + 0.004765721037983894, + 0.003284488571807742, + 0.008962095715105534, + 0.003340711584314704, + 0.0029693262185901403, + 0.0030964810866862535, + 0.003974316641688347, + 0.008106255903840065, + 0.004792532417923212, + 0.006095249671489, + 0.008148714900016785, + 0.008235739544034004, + 0.0072407955303788185, + 0.007796528283506632, + 0.00828389823436737, + 0.0031447275541722775, + 0.007888223975896835, + 0.008020578883588314, + 0.011444021947681904, + 0.010719669982790947, + 0.010381151922047138, + 0.012359987013041973, + 0.02063344419002533, + 0.008193972520530224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.05546237528324127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048666927963495255 + ], + "opposite_logprob_tokens": [ + -3.0486669540405273 + ], + "logprob_reward": -0.048666927963495255, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048666927963495255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32973068952560425, + 0.08367107808589935, + 0.05290762335062027, + 0.04495418816804886, + 0.05254244804382324, + 0.07990888506174088, + 0.0808006003499031, + 0.06242092698812485, + 0.07552003860473633, + 0.13754349946975708 + ], + "credit_map": [ + 0.019395187497138977, + 0.017697773873806, + 0.0945136770606041, + 0.10284139215946198, + 0.014060311019420624, + 0.034777529537677765, + 0.01967683993279934, + 0.02676800638437271, + 0.008951925672590733, + 0.005513409152626991, + 0.014374750666320324, + 0.005827535409480333, + 0.010295961983501911, + 0.009190227836370468, + 0.01233278401196003, + 0.017184490337967873, + 0.005293555557727814, + 0.0066735344007611275, + 0.00493485014885664, + 0.005034847185015678, + 0.0071503398939967155, + 0.007083177100867033, + 0.010777306742966175, + 0.005960014648735523, + 0.00392807275056839, + 0.007573132403194904, + 0.00533171696588397, + 0.004612666089087725, + 0.008633055724203587, + 0.0034539918415248394, + 0.008279547095298767, + 0.003142012981697917, + 0.004159348551183939, + 0.0053917765617370605, + 0.006420462857931852, + 0.0048714736476540565, + 0.012431996874511242, + 0.007878663018345833, + 0.006840416695922613, + 0.0045483168214559555, + 0.004137752577662468, + 0.006094349082559347, + 0.005846651736646891, + 0.014996163547039032, + 0.017838390544056892, + 0.013112680986523628, + 0.010740969330072403, + 0.007141936104744673, + 0.00868249498307705, + 0.009531931951642036, + 0.010353906080126762, + 0.0108251441270113, + 0.013855001889169216, + 0.0047072614543139935, + 0.014836982823908329, + 0.008007887750864029, + 0.009254812262952328, + 0.006835737265646458, + 0.010512817651033401, + 0.006330551579594612, + 0.007841871120035648, + 0.006384899374097586, + 0.007063634227961302, + 0.008196615613996983, + 0.007721649948507547, + 0.008349304087460041, + 0.009951192885637283, + 0.012627344578504562, + 0.011261481791734695, + 0.009307444095611572, + 0.011297207325696945, + 0.005004428327083588, + 0.008631613105535507, + 0.010087910108268261, + 0.01093179453164339, + 0.014676540158689022, + 0.03104228526353836, + 0.023383889347314835, + 0.027750341221690178, + 0.011039132252335548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.048666927963495255 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212082266807556 + ], + "opposite_logprob_tokens": [ + -1.502120852470398 + ], + "logprob_reward": -0.25212082266807556, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212082266807556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3192135989665985, + 0.04771117866039276, + 0.12839621305465698, + 0.12223254144191742, + 0.03491656109690666, + 0.0548434816300869, + 0.06320678442716599, + 0.053346071392297745, + 0.04483330622315407, + 0.13130025565624237 + ], + "credit_map": [ + 0.01574750617146492, + 0.012459784746170044, + 0.0957939401268959, + 0.1232282966375351, + 0.014661530964076519, + 0.030203018337488174, + 0.012221723794937134, + 0.014897814020514488, + 0.006165166851133108, + 0.004696968477219343, + 0.00880784634500742, + 0.0060125007294118404, + 0.005255930125713348, + 0.006099809892475605, + 0.005856875795871019, + 0.004816078580915928, + 0.01352809090167284, + 0.019809983670711517, + 0.011461806483566761, + 0.013946869410574436, + 0.025848032906651497, + 0.01704210415482521, + 0.014742502011358738, + 0.01201681885868311, + 0.01264591421931982, + 0.01636473461985588, + 0.03067895397543907, + 0.022003892809152603, + 0.013577580451965332, + 0.008148483000695705, + 0.011873312294483185, + 0.006939659360796213, + 0.0026154201477766037, + 0.005728280637413263, + 0.004203558433800936, + 0.003223182400688529, + 0.0061427694745361805, + 0.005423451773822308, + 0.003950432874262333, + 0.003629464190453291, + 0.0033066740725189447, + 0.0045796651393175125, + 0.0033006404992192984, + 0.008740386925637722, + 0.013684524223208427, + 0.008971599861979485, + 0.007532721851021051, + 0.0047272625379264355, + 0.007044036407023668, + 0.0084106819704175, + 0.006486111786216497, + 0.009048918262124062, + 0.00891671422868967, + 0.004414759110659361, + 0.013228680938482285, + 0.005656878929585218, + 0.006840417627245188, + 0.003600504482164979, + 0.015134545043110847, + 0.005396922118961811, + 0.003357469569891691, + 0.0026564092840999365, + 0.005762905348092318, + 0.010596894659101963, + 0.005310014821588993, + 0.005191659554839134, + 0.005911341868340969, + 0.007599599193781614, + 0.005636322312057018, + 0.00548016931861639, + 0.006751541048288345, + 0.002952651586383581, + 0.01289115846157074, + 0.009893380105495453, + 0.016006778925657272, + 0.020574547350406647, + 0.018051687628030777, + 0.019895005971193314, + 0.023052267730236053, + 0.010935422964394093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.25212082266807556 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249999929219484 + ], + "opposite_logprob_tokens": [ + -5.255249977111816 + ], + "logprob_reward": -0.005249999929219484, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249999929219484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7373583316802979, + 0.06319225579500198, + 0.025223003700375557, + 0.022168083116412163, + 0.016124706715345383, + 0.025375425815582275, + 0.014360956847667694, + 0.018215373158454895, + 0.019879398867487907, + 0.05810249224305153 + ], + "credit_map": [ + 0.03583504632115364, + 0.031019432470202446, + 0.253421425819397, + 0.2933969497680664, + 0.015145319513976574, + 0.04537102207541466, + 0.02791665680706501, + 0.035252511501312256, + 0.01150849275290966, + 0.005608606617897749, + 0.01003737561404705, + 0.0073887077160179615, + 0.004170004744082689, + 0.008272250182926655, + 0.005663793068379164, + 0.01054302603006363, + 0.0025183430407196283, + 0.0033922719303518534, + 0.0024769441224634647, + 0.0032150098122656345, + 0.004961478523910046, + 0.0027065379545092583, + 0.003616488538682461, + 0.0023359295446425676, + 0.0018362327245995402, + 0.00431950856000185, + 0.0022369527723640203, + 0.002097946824505925, + 0.005210763309150934, + 0.0018317182548344135, + 0.003190777264535427, + 0.0014441842213273048, + 0.0013232131022959948, + 0.002510137390345335, + 0.002057209610939026, + 0.0018978029256686568, + 0.0027348909061402082, + 0.0026003390084952116, + 0.0016192892799153924, + 0.0013818239094689488, + 0.0021055657416582108, + 0.002421720651909709, + 0.0024576231371611357, + 0.005003681872040033, + 0.003796712961047888, + 0.004060786683112383, + 0.0031311213970184326, + 0.0023982131388038397, + 0.001813192735426128, + 0.001877497066743672, + 0.00175102602224797, + 0.0018564293859526515, + 0.002118948847055435, + 0.000982699217274785, + 0.0022678563836961985, + 0.0016933074221014977, + 0.003048143582418561, + 0.0015459980349987745, + 0.00251537817530334, + 0.0024135238490998745, + 0.0014463153202086687, + 0.0020374737214297056, + 0.0026458355132490396, + 0.0025627061259001493, + 0.0024025822058320045, + 0.0023527578450739384, + 0.002166224177926779, + 0.003095854539424181, + 0.0030656110029667616, + 0.002545327879488468, + 0.002427321160212159, + 0.0018237204058095813, + 0.005747838877141476, + 0.0047558206133544445, + 0.010230355896055698, + 0.00777583010494709, + 0.0069304839707911015, + 0.0094658387824893, + 0.008388265036046505, + 0.004808060359209776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.005249999929219484 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979959361255169 + ], + "opposite_logprob_tokens": [ + -3.529799699783325 + ], + "logprob_reward": -0.02979959361255169, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979959361255169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5743487477302551, + 0.06875024735927582, + 0.020533911883831024, + 0.046120453625917435, + 0.021997064352035522, + 0.08876402676105499, + 0.008864044211804867, + 0.009351002052426338, + 0.02115263231098652, + 0.1401178389787674 + ], + "credit_map": [ + 0.02593417838215828, + 0.022809674963355064, + 0.10552575439214706, + 0.21649357676506042, + 0.04205353930592537, + 0.10772629082202911, + 0.022186974063515663, + 0.03161875531077385, + 0.006176755763590336, + 0.0028959575574845076, + 0.02085700072348118, + 0.005679885856807232, + 0.004192201420664787, + 0.0052538649179041386, + 0.0037852460518479347, + 0.019909339025616646, + 0.001822320045903325, + 0.002940220758318901, + 0.0022948873229324818, + 0.0021578555461019278, + 0.004404966253787279, + 0.0029476105701178312, + 0.002019851002842188, + 0.001946200616657734, + 0.004747353959828615, + 0.006301395129412413, + 0.011588464491069317, + 0.008722683414816856, + 0.005285318940877914, + 0.003323265118524432, + 0.003949243575334549, + 0.002202728996053338, + 0.0018852301873266697, + 0.0032288478687405586, + 0.002372300485149026, + 0.003587460145354271, + 0.0031502866186201572, + 0.0033425975125283003, + 0.001979759894311428, + 0.0024505823384970427, + 0.006348487455397844, + 0.009952830150723457, + 0.008420034311711788, + 0.017920829355716705, + 0.013139924965798855, + 0.013290833681821823, + 0.010968863032758236, + 0.008722230792045593, + 0.0014441575622186065, + 0.0012210747227072716, + 0.0009157628519460559, + 0.0011521773412823677, + 0.0012335897190496325, + 0.0006604528171010315, + 0.001483038067817688, + 0.000753791187889874, + 0.0012622467475011945, + 0.0007406965014524758, + 0.0018416474340483546, + 0.000988178071565926, + 0.0007682928699068725, + 0.0010506876278668642, + 0.001101852860301733, + 0.0015974000561982393, + 0.002310965908691287, + 0.0022573135793209076, + 0.002822493202984333, + 0.0026751887053251266, + 0.0036893505603075027, + 0.003056662855669856, + 0.002540722256526351, + 0.0017999347764998674, + 0.012484808452427387, + 0.011035741306841373, + 0.023623401299118996, + 0.02302851900458336, + 0.017029892653226852, + 0.022654704749584198, + 0.018094506114721298, + 0.012166270054876804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.02979959361255169 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, low-angle tracking shot glides smoothly above the ground, keeping pace with a lone skateboarder as they expertly navigate the vast, cavernous spaces of a dimly lit underground garage. The camera remains close to the subject, emphasizing their speed and skill. The skateboarder is a focused figure, dressed in a loose-fitting, dark gray hoodie with the hood pulled up, obscuring their features in shadow, and baggy black overalls. They maintain a low, balanced stance on their skateboard, with its visibly worn wooden deck and vibrant orange wheels, providing a striking pop of color in the quiet surroundings. With fluid precision, they carve a winding path between massive, rough-hewn concrete pillars that recede into the deep, shadowy background. The action is a seamless flow; they lean deeply, their bodies almost parallel to the ground, then effortlessly shift their weight to navigate around the next obstacle. The only sound is the consistent, hypnotic hum of polyurethane wheels rolling across the smooth, polished concrete floor, which is stained with light oil stains and faded yellow parking lines. The sparse, cool white light from flickering fluorescent tubes overhead illuminates the scene, casting long, dancing shadows that stretch and twist with the skateboarders’ movements, creating a stark, high-contrast visual palette and an atmosphere of solitary, focused freedom.", + "seed": 9614, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_14/index_576_seed_9614_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06928610736504197, + "vlm_logprob": -0.06928610736504197, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06928610736504197, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03652061149477959 + ], + "opposite_logprob_tokens": [ + -5.036520481109619 + ], + "logprob_reward": -0.03652061149477959, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.03652061149477959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07577143609523773, + 0.1230822205543518, + 0.07045523822307587, + 0.044721007347106934, + 0.10297054797410965, + 0.049302857369184494, + 0.11396634578704834, + 0.06094690039753914, + 0.26346758008003235, + 0.09531588107347488 + ], + "credit_map": [ + 0.010361313819885254, + 0.009794721379876137, + 0.01506136730313301, + 0.01504897978156805, + 0.006745437625795603, + 0.006991363130509853, + 0.0067254058085381985, + 0.005042848642915487, + 0.00805653166025877, + 0.015626857057213783, + 0.007653464563190937, + 0.009539313614368439, + 0.011348633095622063, + 0.03853422403335571, + 0.01133191678673029, + 0.02099127508699894, + 0.01247800700366497, + 0.006476682145148516, + 0.007335017900913954, + 0.006745604798197746, + 0.00905745942145586, + 0.016510793939232826, + 0.006340111140161753, + 0.005511563736945391, + 0.0033871429041028023, + 0.007437011227011681, + 0.0049798027612268925, + 0.0062925200909376144, + 0.0055938903242349625, + 0.0052392189390957355, + 0.007241697050631046, + 0.004549723118543625, + 0.008568836376070976, + 0.008748026564717293, + 0.02661173790693283, + 0.008048980496823788, + 0.005537007469683886, + 0.006301363930106163, + 0.014055304229259491, + 0.025099286809563637, + 0.006887125317007303, + 0.011440825648605824, + 0.005739876534789801, + 0.010554320178925991, + 0.0030751321464776993, + 0.003128478303551674, + 0.0052406457252800465, + 0.0032364563085138798, + 0.023725291714072227, + 0.0119113614782691, + 0.014226791448891163, + 0.013967910781502724, + 0.010786443017423153, + 0.01407846063375473, + 0.013024210929870605, + 0.012245872989296913, + 0.006976642180234194, + 0.004427509848028421, + 0.012641945853829384, + 0.005903025157749653, + 0.0076964073814451694, + 0.008198611438274384, + 0.004992421716451645, + 0.010110337287187576, + 0.01825961284339428, + 0.01308836042881012, + 0.0854104608297348, + 0.015252123586833477, + 0.010294434614479542, + 0.017237873747944832, + 0.01848599500954151, + 0.08543872088193893, + 0.01306269969791174, + 0.011670461855828762, + 0.018558336421847343, + 0.014937730506062508, + 0.007448806893080473, + 0.012227500788867474, + 0.009255072101950645, + 0.008155271410942078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.03652061149477959 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811577707529068 + ], + "opposite_logprob_tokens": [ + -3.2881157398223877 + ], + "logprob_reward": -0.03811577707529068, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811577707529068, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1441972851753235, + 0.25759437680244446, + 0.07280060648918152, + 0.10216718912124634, + 0.10465359687805176, + 0.07280364632606506, + 0.07493908703327179, + 0.04953950271010399, + 0.07537851482629776, + 0.045926228165626526 + ], + "credit_map": [ + 0.02225547283887863, + 0.017577091231942177, + 0.01317015290260315, + 0.012368058785796165, + 0.015578539110720158, + 0.02245114929974079, + 0.025481965392827988, + 0.015314849093556404, + 0.010161944665014744, + 0.03856953606009483, + 0.021559730172157288, + 0.01202354859560728, + 0.024110782891511917, + 0.0827198326587677, + 0.014351775869727135, + 0.054097194224596024, + 0.012698912061750889, + 0.00927757564932108, + 0.009821012616157532, + 0.007475053425878286, + 0.005002468358725309, + 0.017064902931451797, + 0.005401351023465395, + 0.006059329025447369, + 0.007850132882595062, + 0.018910745158791542, + 0.02164272964000702, + 0.005976927932351828, + 0.02143060229718685, + 0.012095378711819649, + 0.00483496580272913, + 0.009425700642168522, + 0.010844373144209385, + 0.013414229266345501, + 0.011194207705557346, + 0.013690556399524212, + 0.008071556687355042, + 0.0072932373732328415, + 0.029782792553305626, + 0.010362641885876656, + 0.010342879220843315, + 0.014953468926250935, + 0.010295618325471878, + 0.012426755391061306, + 0.004610637668520212, + 0.00661618635058403, + 0.006626257672905922, + 0.0069318413734436035, + 0.011725811287760735, + 0.008889532648026943, + 0.009584366343915462, + 0.014677821658551693, + 0.005811117123812437, + 0.0077522555366158485, + 0.008517800830304623, + 0.007980376482009888, + 0.004629315342754126, + 0.003659465117380023, + 0.010279415175318718, + 0.007908034138381481, + 0.003715739818289876, + 0.004827316850423813, + 0.004250980447977781, + 0.010269230231642723, + 0.006909836549311876, + 0.004681956022977829, + 0.01591147668659687, + 0.006615650840103626, + 0.007739434950053692, + 0.0074645462445914745, + 0.008425869047641754, + 0.01762974262237549, + 0.00715135270729661, + 0.003910100553184748, + 0.009351714514195919, + 0.005377684719860554, + 0.004158060532063246, + 0.004455328453332186, + 0.005082879681140184, + 0.006439106073230505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.03811577707529068 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25211629271507263 + ], + "opposite_logprob_tokens": [ + -1.502116322517395 + ], + "logprob_reward": -0.25211629271507263, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25211629271507263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09411367774009705, + 0.06562527269124985, + 0.10842148959636688, + 0.15140995383262634, + 0.17980660498142242, + 0.1441882848739624, + 0.07079067081212997, + 0.09467136114835739, + 0.039108097553253174, + 0.05186459794640541 + ], + "credit_map": [ + 0.012563485652208328, + 0.009763662703335285, + 0.011925172060728073, + 0.00790848582983017, + 0.0088152214884758, + 0.009512986987829208, + 0.020964624360203743, + 0.012660032138228416, + 0.004321408458054066, + 0.005984313786029816, + 0.01697918027639389, + 0.005409824196249247, + 0.007210732903331518, + 0.005726150702685118, + 0.005664757452905178, + 0.014328899793326855, + 0.015164797194302082, + 0.014433206990361214, + 0.021392548456788063, + 0.010229554027318954, + 0.011740170419216156, + 0.017814353108406067, + 0.010018661618232727, + 0.007628192659467459, + 0.010030470788478851, + 0.017062723636627197, + 0.039186056703329086, + 0.01368893776088953, + 0.02608371153473854, + 0.011465157382190228, + 0.013240029104053974, + 0.020652851089835167, + 0.01816120371222496, + 0.01601523533463478, + 0.02032425068318844, + 0.01676441729068756, + 0.016567407175898552, + 0.03388025611639023, + 0.04022131860256195, + 0.017872506752610207, + 0.0241249892860651, + 0.009988104924559593, + 0.034606046974658966, + 0.011931891553103924, + 0.015032299794256687, + 0.02088121511042118, + 0.01406684797257185, + 0.013556871563196182, + 0.017212146893143654, + 0.00862786453217268, + 0.009223302826285362, + 0.008687811903655529, + 0.006325477734208107, + 0.0064458176493644714, + 0.006602129898965359, + 0.00766611285507679, + 0.010145574808120728, + 0.00783835630863905, + 0.011714600957930088, + 0.021724753081798553, + 0.010430112481117249, + 0.010979157872498035, + 0.01044574286788702, + 0.011393057182431221, + 0.003266621381044388, + 0.00274688214994967, + 0.007106529548764229, + 0.004359536338597536, + 0.004340755753219128, + 0.004328425507992506, + 0.007272326387465, + 0.0056870183907449245, + 0.006043219938874245, + 0.003929976373910904, + 0.007601686753332615, + 0.007907385937869549, + 0.0052323308773338795, + 0.007965571247041225, + 0.0075325630605220795, + 0.005651857238262892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.25211629271507263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361218536272645 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.00361218536272645, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361218536272645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10005657374858856, + 0.058995332568883896, + 0.05877859517931938, + 0.06737779825925827, + 0.15253892540931702, + 0.1057606041431427, + 0.0898047685623169, + 0.06127815693616867, + 0.19161711633205414, + 0.11379208415746689 + ], + "credit_map": [ + 0.01154463179409504, + 0.02643292024731636, + 0.008743279613554478, + 0.024336067959666252, + 0.007590785622596741, + 0.00752596789970994, + 0.008011439815163612, + 0.005871487781405449, + 0.004480152390897274, + 0.006321793422102928, + 0.015197809785604477, + 0.004788621328771114, + 0.006519741378724575, + 0.004346964880824089, + 0.005959215573966503, + 0.011381035670638084, + 0.01049460843205452, + 0.0063760061748325825, + 0.007961510680615902, + 0.004687978886067867, + 0.005102648865431547, + 0.013919123448431492, + 0.005643809214234352, + 0.004592906218022108, + 0.005407674703747034, + 0.0069311680272221565, + 0.017148328945040703, + 0.00525751244276762, + 0.012419596314430237, + 0.007901689037680626, + 0.004818705376237631, + 0.007493131328374147, + 0.0131832929328084, + 0.01155410148203373, + 0.03174262121319771, + 0.012924251146614552, + 0.009336806833744049, + 0.019076067954301834, + 0.026245001703500748, + 0.0284767784178257, + 0.01850687712430954, + 0.008450426161289215, + 0.022869622334837914, + 0.00828205794095993, + 0.01162263285368681, + 0.017592662945389748, + 0.008632862009108067, + 0.009803461842238903, + 0.01764940842986107, + 0.0098973223939538, + 0.010450449772179127, + 0.00833209604024887, + 0.011916717514395714, + 0.013333515264093876, + 0.00866096094250679, + 0.009564300999045372, + 0.006667998619377613, + 0.0066846683621406555, + 0.007276460994035006, + 0.00953125674277544, + 0.009216323494911194, + 0.008822997100651264, + 0.006708371918648481, + 0.006370081566274166, + 0.011126728728413582, + 0.009487450122833252, + 0.06990466266870499, + 0.011653156951069832, + 0.0055001769214868546, + 0.01257274765521288, + 0.011028691194951534, + 0.06034349277615547, + 0.012736116535961628, + 0.012232822366058826, + 0.023021802306175232, + 0.020678916946053505, + 0.009153029881417751, + 0.012965130619704723, + 0.011139164678752422, + 0.011865099892020226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.00361218536272645 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera positioned at a low angle and tracking smoothly above the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the skateboarder wearing a dark gray hoodie with the hood pulled up and baggy black overalls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the skateboard's deck visibly worn wood and its wheels vibrant orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene take place in a dimly lit underground garage with massive rough-hewn concrete pillars?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the floor smooth polished concrete with light oil stains and faded yellow parking lines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065670177340508 + ], + "opposite_logprob_tokens": [ + -4.14106559753418 + ], + "logprob_reward": -0.016065670177340508, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065670177340508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06707505136728287, + 0.2691341042518616, + 0.04632314294576645, + 0.05260050669312477, + 0.23977433145046234, + 0.0549132414162159, + 0.05267234146595001, + 0.08819883316755295, + 0.06940466910600662, + 0.059903834015131 + ], + "credit_map": [ + 0.007494746241718531, + 0.015254674479365349, + 0.01669459044933319, + 0.0069442749954760075, + 0.004799650516360998, + 0.004502515774220228, + 0.006661335937678814, + 0.004723262507468462, + 0.013565289787948132, + 0.0335414819419384, + 0.06939515471458435, + 0.01630203053355217, + 0.015488997101783752, + 0.047419410198926926, + 0.01006222516298294, + 0.06335952877998352, + 0.008105766959488392, + 0.005365316290408373, + 0.009156721644103527, + 0.0033405148424208164, + 0.003190325340256095, + 0.010150313377380371, + 0.003741751192137599, + 0.003272430971264839, + 0.0034355977550148964, + 0.006121619138866663, + 0.013169221580028534, + 0.0039527821354568005, + 0.006323961075395346, + 0.005636320915073156, + 0.00417511910200119, + 0.009785888716578484, + 0.024097222834825516, + 0.021028904244303703, + 0.03449295088648796, + 0.02403832972049713, + 0.014860540628433228, + 0.028723815456032753, + 0.05984853580594063, + 0.032684020698070526, + 0.008876307867467403, + 0.004449212457984686, + 0.009916027076542377, + 0.004670563619583845, + 0.004951010458171368, + 0.009143666364252567, + 0.006618266459554434, + 0.006288188509643078, + 0.008547690697014332, + 0.004204360768198967, + 0.006490370258688927, + 0.0042595816776156425, + 0.006645631976425648, + 0.007260291371494532, + 0.006097281351685524, + 0.009167129173874855, + 0.009127413854002953, + 0.009723678231239319, + 0.016711663454771042, + 0.01037287525832653, + 0.007336008828133345, + 0.007367631420493126, + 0.009554724209010601, + 0.01800483837723732, + 0.007567048072814941, + 0.008632670156657696, + 0.01793164759874344, + 0.005403122399002314, + 0.003544256091117859, + 0.005396750755608082, + 0.00651538698002696, + 0.01441378053277731, + 0.008053455501794815, + 0.005212742369621992, + 0.010535471141338348, + 0.010842486284673214, + 0.003956330940127373, + 0.006712808273732662, + 0.005281647201627493, + 0.009308887645602226 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 7 + ], + "selected_reward": -0.016065670177340508 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_tensor([3144, 3144])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_tensor([3144, 3144])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1e114a27046313697d4f25e0591a1ee0f6c4cdd6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_tensor([3144, 3144])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, wide shot captures a vast circular room, rendered in a high-contrast, noir-style anime, its curved walls constructed of massive, dark-gray stone blocks. In the center, a massive, multi-tiered bookshelf, constructed of dark, aged wood and wire, rotates with a silent, relentless slowness. Its countless ancient tomes, bound in crumbling, shadowy leather, bear not titles but mysterious, cryptic symbols that glow with an ethereal blue light, pulsing faintly in the darkness. The room's sole source of light comes from a magical sphere of crackling white energy floating near the high, domed ceiling; its erratic flickering causes the entire scene to flicker violently, casting long, twisting, fiberglass-like shadows that wriggle and dance across the polished, cold stone floor. In the foreground, a solitary, cloaked figure stands motionless, their body completely concealed by the heavy, dark fabric, their head tilted upward, watching the celestial rotation, a silent observer lost in the heart of an ancient, shimmering mystery.", + "seed": 24316, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_3144_seed_24316_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.16580862253904344, + "vlm_logprob": -0.16580862253904344, + "vlm_yes_no_margin": 2.225, + "vlm_token_credit": -0.16580862253904344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18767720460891724 + ], + "opposite_logprob_tokens": [ + -2.6876771450042725 + ], + "logprob_reward": -0.18767720460891724, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.18767720460891724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1309012621641159, + 0.05523784086108208, + 0.32284146547317505, + 0.03384977579116821, + 0.14453211426734924, + 0.03693915531039238, + 0.07863236218690872, + 0.016338033601641655, + 0.025413693860173225, + 0.15531428158283234 + ], + "credit_map": [ + 0.018734492361545563, + 0.020122995600104332, + 0.01969287171959877, + 0.01863914355635643, + 0.009320748969912529, + 0.010854732245206833, + 0.019827011972665787, + 0.013709262944757938, + 0.008200674317777157, + 0.00867679063230753, + 0.0061640203930437565, + 0.01093385275453329, + 0.005614276975393295, + 0.00628077844157815, + 0.0055182743817567825, + 0.003849169472232461, + 0.0496877059340477, + 0.045234281569719315, + 0.04376678541302681, + 0.04737783223390579, + 0.022672317922115326, + 0.026426570490002632, + 0.04281511902809143, + 0.0448608472943306, + 0.004945516586303711, + 0.00659272912889719, + 0.006039671134203672, + 0.005258128512650728, + 0.002027845708653331, + 0.0019581515807658434, + 0.004071366507560015, + 0.0029563659336417913, + 0.018004631623625755, + 0.018088659271597862, + 0.017006346955895424, + 0.02003089152276516, + 0.012046254239976406, + 0.012415922246873379, + 0.017652995884418488, + 0.02928641065955162, + 0.006546790711581707, + 0.004854502622038126, + 0.0061339703388512135, + 0.005082623567432165, + 0.001884898985736072, + 0.0028758710250258446, + 0.004226576071232557, + 0.005333922803401947, + 0.0130933141335845, + 0.010252735577523708, + 0.011857285164296627, + 0.012686041183769703, + 0.0046293470077216625, + 0.0065436167642474174, + 0.00989834126085043, + 0.009671676903963089, + 0.002240452216938138, + 0.0015684021636843681, + 0.0022439041640609503, + 0.002374839037656784, + 0.001006225822493434, + 0.001338869333267212, + 0.0036957920528948307, + 0.0018695492763072252, + 0.004991891793906689, + 0.004808229394257069, + 0.0035115829668939114, + 0.0038224661257117987, + 0.001586070517078042, + 0.00131248461548239, + 0.0029248809441924095, + 0.0024560876190662384, + 0.019664853811264038, + 0.0202268548309803, + 0.018917258828878403, + 0.02265462651848793, + 0.010297790169715881, + 0.011640176177024841, + 0.020977728068828583, + 0.030934980139136314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.18767720460891724 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31348446011543274 + ], + "opposite_logprob_tokens": [ + -1.3134844303131104 + ], + "logprob_reward": -0.31348446011543274, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31348446011543274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35935118794441223, + 0.04746986925601959, + 0.2683257758617401, + 0.023782934993505478, + 0.12408410757780075, + 0.016455350443720818, + 0.1128045991063118, + 0.00877404771745205, + 0.009588651359081268, + 0.029363475739955902 + ], + "credit_map": [ + 0.04894265532493591, + 0.03695233538746834, + 0.06238793209195137, + 0.04954373463988304, + 0.03317728266119957, + 0.0367036797106266, + 0.05783039331436157, + 0.03381318971514702, + 0.006396203767508268, + 0.006464108359068632, + 0.0069121625274419785, + 0.006510223727673292, + 0.006212189793586731, + 0.005835193209350109, + 0.004990818910300732, + 0.004148965235799551, + 0.041090965270996094, + 0.037602826952934265, + 0.03641223534941673, + 0.03932248428463936, + 0.019566792994737625, + 0.02240884117782116, + 0.034520912915468216, + 0.03740071505308151, + 0.003346404293552041, + 0.003113165497779846, + 0.003982061520218849, + 0.003436120692640543, + 0.0020155860111117363, + 0.0026206099428236485, + 0.002861043205484748, + 0.0024079398717731237, + 0.0158417746424675, + 0.013232368975877762, + 0.017145296558737755, + 0.018555117771029472, + 0.012628750875592232, + 0.010359757579863071, + 0.012757215648889542, + 0.023563826456665993, + 0.0023833252489566803, + 0.0023643551394343376, + 0.002721603959798813, + 0.0022524071391671896, + 0.0011345609091222286, + 0.0011723265051841736, + 0.0018546492792665958, + 0.0025721220299601555, + 0.017766322940587997, + 0.014022568240761757, + 0.016280775889754295, + 0.019052976742386818, + 0.0070512001402676105, + 0.008929915726184845, + 0.015015829354524612, + 0.014684999361634254, + 0.0011108379112556577, + 0.0014134431257843971, + 0.0011523516150191426, + 0.0013299124548211694, + 0.0006204484961926937, + 0.0009220686042681336, + 0.0014224647311493754, + 0.000802520546130836, + 0.001773109077475965, + 0.0009286857675760984, + 0.001445143367163837, + 0.0017270235111936927, + 0.0007880267221480608, + 0.0006902295281179249, + 0.0012372690252959728, + 0.000999164069071412, + 0.003925015218555927, + 0.004454373847693205, + 0.0041020638309419155, + 0.004558858461678028, + 0.0026573992799967527, + 0.002595888450741768, + 0.0035519874654710293, + 0.00351788941770792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.31348446011543274 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07900167256593704 + ], + "opposite_logprob_tokens": [ + -2.5790016651153564 + ], + "logprob_reward": -0.07900167256593704, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07900167256593704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2834750711917877, + 0.08393491804599762, + 0.22588683664798737, + 0.07122336328029633, + 0.19569362699985504, + 0.04030010476708412, + 0.026580296456813812, + 0.025859879329800606, + 0.02038075216114521, + 0.026665184646844864 + ], + "credit_map": [ + 0.038132548332214355, + 0.024146558716893196, + 0.033423930406570435, + 0.05322740599513054, + 0.023324230685830116, + 0.021880511194467545, + 0.05800479277968407, + 0.03133507817983627, + 0.01110763568431139, + 0.010859772562980652, + 0.01154042687267065, + 0.015400493517518044, + 0.0059442888014018536, + 0.006744407583028078, + 0.011552953161299229, + 0.010784929618239403, + 0.033084284514188766, + 0.03152699023485184, + 0.03461303189396858, + 0.03301919996738434, + 0.016198070719838142, + 0.017837584018707275, + 0.026171360164880753, + 0.03343629837036133, + 0.008625248447060585, + 0.012031139805912971, + 0.010341397486627102, + 0.011802692897617817, + 0.005573499482125044, + 0.006054646335542202, + 0.010749036446213722, + 0.006045698653906584, + 0.024328211322426796, + 0.024448344483971596, + 0.023120203986763954, + 0.025885188952088356, + 0.020547164604067802, + 0.017306631430983543, + 0.025073299184441566, + 0.03498456999659538, + 0.00571161275729537, + 0.005277807358652353, + 0.007331092841923237, + 0.005269392393529415, + 0.0029653755482286215, + 0.004684437066316605, + 0.005023385398089886, + 0.004036996979266405, + 0.004440912511199713, + 0.0035414285957813263, + 0.0031405738554894924, + 0.0048455605283379555, + 0.0019053687574341893, + 0.003109171986579895, + 0.003108011791482568, + 0.002489263890311122, + 0.0035208133049309254, + 0.003994198516011238, + 0.0050263297744095325, + 0.004217264708131552, + 0.0012748058652505279, + 0.0017522363923490047, + 0.002816123655065894, + 0.0032581023406237364, + 0.0035058483481407166, + 0.0036899393890053034, + 0.0026037676725536585, + 0.004098663106560707, + 0.0011574930977076292, + 0.0012383407447487116, + 0.0023139547556638718, + 0.0017727427184581757, + 0.0031214465852826834, + 0.0035560508258640766, + 0.0031219404190778732, + 0.003765333443880081, + 0.0030035413801670074, + 0.003373153507709503, + 0.0031413757242262363, + 0.003582339733839035 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.07900167256593704 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232793390750885 + ], + "opposite_logprob_tokens": [ + -3.7732794284820557 + ], + "logprob_reward": -0.0232793390750885, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232793390750885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14846357703208923, + 0.056512270122766495, + 0.4781658351421356, + 0.027901137247681618, + 0.16107890009880066, + 0.042641814798116684, + 0.024513261392712593, + 0.013099588453769684, + 0.011104238219559193, + 0.036519430577754974 + ], + "credit_map": [ + 0.021024372428655624, + 0.020220596343278885, + 0.026207977905869484, + 0.02086016908288002, + 0.010219289921224117, + 0.010630154050886631, + 0.02244081161916256, + 0.01686021499335766, + 0.01001053024083376, + 0.00944962166249752, + 0.007850534282624722, + 0.008188770152628422, + 0.002863801782950759, + 0.004690077155828476, + 0.006217968184500933, + 0.0072409613057971, + 0.07366012036800385, + 0.06684882193803787, + 0.06485330313444138, + 0.07089632749557495, + 0.033556561917066574, + 0.037899624556303024, + 0.06321630626916885, + 0.06723474711179733, + 0.003212900599464774, + 0.003962417598813772, + 0.003962838556617498, + 0.004203826654702425, + 0.0018891342915594578, + 0.002306634560227394, + 0.004126161336898804, + 0.004237224813550711, + 0.01875915750861168, + 0.02996370568871498, + 0.01898150146007538, + 0.02269640378654003, + 0.0114965308457613, + 0.012205365113914013, + 0.019847797229886055, + 0.0271284356713295, + 0.007356069982051849, + 0.005047814454883337, + 0.007197946775704622, + 0.007264158688485622, + 0.0025192962493747473, + 0.003551134606823325, + 0.005145465023815632, + 0.004559929016977549, + 0.0034441978204995394, + 0.0038588482420891523, + 0.0035492628812789917, + 0.0037292472552508116, + 0.001594105502590537, + 0.0017787222750484943, + 0.0031007423531264067, + 0.0034581338986754417, + 0.0015962207689881325, + 0.0017165170283988118, + 0.0025111401919275522, + 0.001917906804010272, + 0.0010537664638832211, + 0.0011184944305568933, + 0.0015375192742794752, + 0.0016480243066325784, + 0.0018384542781859636, + 0.001903464668430388, + 0.0017841266235336661, + 0.0015541346510872245, + 0.0007869530818425119, + 0.0007303841994144022, + 0.0015685891266912222, + 0.0009381319396197796, + 0.005377355497330427, + 0.007933303713798523, + 0.004924429580569267, + 0.005031614564359188, + 0.0022086871322244406, + 0.0016714051598683, + 0.004542771726846695, + 0.0048298644833266735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0232793390750885 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256004363298416 + ], + "opposite_logprob_tokens": [ + -1.6006004810333252 + ], + "logprob_reward": -0.2256004363298416, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256004363298416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1073458269238472, + 0.047592971473932266, + 0.6275946497917175, + 0.01417250744998455, + 0.09483113139867783, + 0.0587250217795372, + 0.01510546077042818, + 0.007722639478743076, + 0.010843452997505665, + 0.016066351905465126 + ], + "credit_map": [ + 0.015187566168606281, + 0.014377517625689507, + 0.013326553627848625, + 0.017472820356488228, + 0.007160730194300413, + 0.004708475898951292, + 0.02072504162788391, + 0.014387121424078941, + 0.007472256664186716, + 0.005991294514387846, + 0.006801456678658724, + 0.006984748411923647, + 0.004583802539855242, + 0.003537532174959779, + 0.005084105301648378, + 0.007137778215110302, + 0.09857996553182602, + 0.08943353593349457, + 0.08587663620710373, + 0.09157248586416245, + 0.04396417737007141, + 0.04880469664931297, + 0.07789675891399384, + 0.09146637469530106, + 0.001770937000401318, + 0.002479131566360593, + 0.0016166751738637686, + 0.00209250720217824, + 0.0014333314029499888, + 0.001492440002039075, + 0.0019401154713705182, + 0.00134736904874444, + 0.010841085575520992, + 0.011565649881958961, + 0.012064171023666859, + 0.012971932999789715, + 0.009309117682278156, + 0.008164765313267708, + 0.01281709410250187, + 0.017097316682338715, + 0.009686661884188652, + 0.0078091928735375404, + 0.009569834917783737, + 0.007383212447166443, + 0.003445782233029604, + 0.0053615388460457325, + 0.006930923089385033, + 0.008537876419723034, + 0.0024283435195684433, + 0.0018627304816618562, + 0.0024310315493494272, + 0.002369142835959792, + 0.0012150785187259316, + 0.0013724738964810967, + 0.001963694579899311, + 0.001462966320104897, + 0.001164810499176383, + 0.0007454445003531873, + 0.0017275194404646754, + 0.0012305488344281912, + 0.0004539258952718228, + 0.0005350065184757113, + 0.0010626513976603746, + 0.0008027328876778483, + 0.0018505246844142675, + 0.001647808589041233, + 0.001333998516201973, + 0.0021088195499032736, + 0.0007053744047880173, + 0.0007852504495531321, + 0.0014494576025754213, + 0.0009622194338589907, + 0.0024387543089687824, + 0.0026642531156539917, + 0.002575865713879466, + 0.0023356634192168713, + 0.0009386593592353165, + 0.0009711412712931633, + 0.0022934407461434603, + 0.0018485741456970572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.2256004363298416 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide shot captures a vast circular room, rendered in a high-contrast, noir-style anime, its curved walls constructed of massive, dark-gray stone blocks. In the center, a massive, multi-tiered bookshelf, constructed of dark, aged wood and wire, rotates with a silent, relentless slowness. Its countless ancient tomes, bound in crumbling, shadowy leather, bear not titles but mysterious, cryptic symbols that glow with an ethereal blue light, pulsing faintly in the darkness. The room's sole source of light comes from a magical sphere of crackling white energy floating near the high, domed ceiling; its erratic flickering causes the entire scene to flicker violently, casting long, twisting, fiberglass-like shadows that wriggle and dance across the polished, cold stone floor. In the foreground, a solitary, cloaked figure stands motionless, their body completely concealed by the heavy, dark fabric, their head tilted upward, watching the celestial rotation, a silent observer lost in the heart of an ancient, shimmering mystery.", + "seed": 24317, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_3144_seed_24317_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04468120755627751, + "vlm_logprob": -0.04468120755627751, + "vlm_yes_no_margin": 3.775, + "vlm_token_credit": -0.04468120755627751, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08413596451282501 + ], + "opposite_logprob_tokens": [ + -3.5841360092163086 + ], + "logprob_reward": -0.08413596451282501, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.08413596451282501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1214958056807518, + 0.056057099252939224, + 0.04506794363260269, + 0.03695663437247276, + 0.057494960725307465, + 0.2729886770248413, + 0.06129347160458565, + 0.21779341995716095, + 0.10347867012023926, + 0.027373237535357475 + ], + "credit_map": [ + 0.01709487847983837, + 0.012119930237531662, + 0.014976398088037968, + 0.020199328660964966, + 0.015289979055523872, + 0.014767262153327465, + 0.015604395419359207, + 0.011443635448813438, + 0.006005249917507172, + 0.006735042203217745, + 0.005646784324198961, + 0.009667263366281986, + 0.010946580208837986, + 0.00512350257486105, + 0.00829535536468029, + 0.003637326415628195, + 0.005192541051656008, + 0.005035842768847942, + 0.006038407329469919, + 0.007731887977570295, + 0.00301101035438478, + 0.004303994122892618, + 0.007571803871542215, + 0.006182457320392132, + 0.0053236172534525394, + 0.0037449912633746862, + 0.0039003107231110334, + 0.009904395788908005, + 0.0024284974206238985, + 0.005066802259534597, + 0.0029263333417475224, + 0.0036616921424865723, + 0.005982814822345972, + 0.006122184451669455, + 0.00513276644051075, + 0.00978191290050745, + 0.007019038777798414, + 0.004830396268516779, + 0.012868689373135567, + 0.0057571628130972385, + 0.0369715690612793, + 0.023556482046842575, + 0.031120453029870987, + 0.028966756537556648, + 0.01727532036602497, + 0.07042605429887772, + 0.026793241500854492, + 0.037878818809986115, + 0.007260552607476711, + 0.006752700079232454, + 0.007752525620162487, + 0.009489193558692932, + 0.004417175427079201, + 0.007654367946088314, + 0.00985002238303423, + 0.00811693537980318, + 0.03143042325973511, + 0.02496316283941269, + 0.04314064979553223, + 0.026832563802599907, + 0.013004874810576439, + 0.028180712834000587, + 0.028163988143205643, + 0.022077064961194992, + 0.019213950261473656, + 0.009641516953706741, + 0.020762339234352112, + 0.009190313518047333, + 0.008421201258897781, + 0.01746736839413643, + 0.00861434731632471, + 0.01016764622181654, + 0.0037422606255859137, + 0.003521260805428028, + 0.004829047713428736, + 0.003822011174634099, + 0.0019725956954061985, + 0.0030606461223214865, + 0.003857631469145417, + 0.002567786956205964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.08413596451282501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009805607609450817 + ], + "opposite_logprob_tokens": [ + -4.634805679321289 + ], + "logprob_reward": -0.009805607609450817, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009805607609450817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22853270173072815, + 0.12171008437871933, + 0.1460069715976715, + 0.07001333683729172, + 0.05120817944407463, + 0.19677869975566864, + 0.09706585854291916, + 0.06348902732133865, + 0.012540708296000957, + 0.012654431164264679 + ], + "credit_map": [ + 0.039566200226545334, + 0.01961555890738964, + 0.05629546567797661, + 0.03140722215175629, + 0.013347749598324299, + 0.02172071300446987, + 0.02799721248447895, + 0.018582558259367943, + 0.014876725152134895, + 0.01896657794713974, + 0.017755992710590363, + 0.014989045448601246, + 0.01729084737598896, + 0.012786039151251316, + 0.017458397895097733, + 0.007586457766592503, + 0.04230755940079689, + 0.00563313253223896, + 0.056946270167827606, + 0.011574633419513702, + 0.004166088532656431, + 0.006146608851850033, + 0.008990032598376274, + 0.010242648422718048, + 0.012752940878272057, + 0.0039077820256352425, + 0.008052022196352482, + 0.026763098314404488, + 0.0027856093365699053, + 0.0053510647267103195, + 0.0058956146240234375, + 0.004505198914557695, + 0.005586098879575729, + 0.00497801136225462, + 0.005606795661151409, + 0.00648918142542243, + 0.004801521543413401, + 0.007326564751565456, + 0.010353494435548782, + 0.006066513247787952, + 0.025244032964110374, + 0.02117813378572464, + 0.02355257049202919, + 0.021511387079954147, + 0.015377020463347435, + 0.04740669205784798, + 0.017283417284488678, + 0.025225456804037094, + 0.013150036334991455, + 0.01305888220667839, + 0.00726042315363884, + 0.01781173050403595, + 0.007235272787511349, + 0.01101864967495203, + 0.01810324750840664, + 0.009427623823285103, + 0.00907061342149973, + 0.006680995691567659, + 0.01258884184062481, + 0.007842948660254478, + 0.008939094841480255, + 0.007286960259079933, + 0.0057670921087265015, + 0.005312483292073011, + 0.0015572033589705825, + 0.0018554666312411427, + 0.0013778379652649164, + 0.001325914403423667, + 0.0024220605846494436, + 0.0012366059236228466, + 0.0017914666095748544, + 0.0009741532849147916, + 0.0016439604805782437, + 0.0014577078400179744, + 0.001635727589018643, + 0.00211617280729115, + 0.0015672381268814206, + 0.0010491340653970838, + 0.002074877033010125, + 0.0011096133384853601 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.009805607609450817 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674506975337863 + ], + "opposite_logprob_tokens": [ + -5.006744861602783 + ], + "logprob_reward": -0.00674506975337863, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674506975337863, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14257901906967163, + 0.1981458067893982, + 0.10899901390075684, + 0.07092346251010895, + 0.05557771027088165, + 0.1220000684261322, + 0.043212149292230606, + 0.1471935212612152, + 0.07524605095386505, + 0.03612319752573967 + ], + "credit_map": [ + 0.02039577253162861, + 0.017010582610964775, + 0.025038303807377815, + 0.02314058691263199, + 0.01594003662467003, + 0.013910656794905663, + 0.01635373942553997, + 0.010789342224597931, + 0.024322329089045525, + 0.023920945823192596, + 0.03204029053449631, + 0.02115335874259472, + 0.035509709268808365, + 0.025498367846012115, + 0.02161111868917942, + 0.014089686796069145, + 0.023795248940587044, + 0.00964049156755209, + 0.033848751336336136, + 0.0133449025452137, + 0.0045554121024906635, + 0.0061733960174024105, + 0.010248844511806965, + 0.007391971070319414, + 0.014610139653086662, + 0.005631553940474987, + 0.007168566808104515, + 0.02639157325029373, + 0.0030641185585409403, + 0.004357092082500458, + 0.0051483334973454475, + 0.004552093800157309, + 0.00711542833596468, + 0.005491918418556452, + 0.008086947724223137, + 0.009823204018175602, + 0.0047292690724134445, + 0.005923098884522915, + 0.008812139742076397, + 0.005595705937594175, + 0.01619395986199379, + 0.01347243133932352, + 0.015886792913079262, + 0.013786151073873043, + 0.00668903486803174, + 0.030369561165571213, + 0.010808383114635944, + 0.014793765731155872, + 0.006093562114983797, + 0.0049512614496052265, + 0.005385763477534056, + 0.006549985613673925, + 0.0031421035528182983, + 0.0051029399037361145, + 0.007452534511685371, + 0.004534002859145403, + 0.01797870732843876, + 0.017350850626826286, + 0.014332505874335766, + 0.014881324954330921, + 0.024869367480278015, + 0.02223432995378971, + 0.021264486014842987, + 0.014281953684985638, + 0.011471029371023178, + 0.00731972511857748, + 0.011900625191628933, + 0.007396213244646788, + 0.01031638402491808, + 0.01214808039367199, + 0.007675688713788986, + 0.007018308620899916, + 0.0049357605166733265, + 0.0049109868705272675, + 0.004026670008897781, + 0.006223863456398249, + 0.0036147984210401773, + 0.00393632659688592, + 0.00561508210375905, + 0.0028597123455256224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.00674506975337863 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890204906463623 + ], + "opposite_logprob_tokens": [ + -2.4640204906463623 + ], + "logprob_reward": -0.0890204906463623, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890204906463623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06493399292230606, + 0.12789513170719147, + 0.2270752638578415, + 0.030451277270913124, + 0.10563187301158905, + 0.11313877999782562, + 0.1095341145992279, + 0.056146759539842606, + 0.1408342719078064, + 0.024358589202165604 + ], + "credit_map": [ + 0.009260330349206924, + 0.008424809202551842, + 0.009403376840054989, + 0.009585310705006123, + 0.006264341063797474, + 0.008911185897886753, + 0.006482010707259178, + 0.006602628622204065, + 0.01412068959325552, + 0.019268378615379333, + 0.011751946061849594, + 0.012582402676343918, + 0.030141528695821762, + 0.016408676281571388, + 0.014702285639941692, + 0.008919238112866879, + 0.06511446833610535, + 0.008584833703935146, + 0.09014081209897995, + 0.016752125695347786, + 0.006028937641531229, + 0.010473497211933136, + 0.014911913312971592, + 0.01506868377327919, + 0.0044572255574166775, + 0.002185758436098695, + 0.003542571095749736, + 0.00940336287021637, + 0.0018619421171024442, + 0.0032663228921592236, + 0.0023715433198958635, + 0.0033625485375523567, + 0.011476337909698486, + 0.010480500757694244, + 0.008802435360848904, + 0.016280250623822212, + 0.009482630528509617, + 0.015087787993252277, + 0.021609701216220856, + 0.012412226758897305, + 0.016650328412652016, + 0.011027473025023937, + 0.01820077933371067, + 0.010734009556472301, + 0.0064315651543438435, + 0.027660200372338295, + 0.010753833688795567, + 0.011680599302053452, + 0.017831608653068542, + 0.01089794933795929, + 0.0178492683917284, + 0.013075584545731544, + 0.005899105221033096, + 0.018739955499768257, + 0.009936868213117123, + 0.015303761698305607, + 0.008197486400604248, + 0.006348273251205683, + 0.010932795703411102, + 0.006373149808496237, + 0.005320006515830755, + 0.007198323495686054, + 0.0055335210636258125, + 0.006243200972676277, + 0.024823913350701332, + 0.013243229128420353, + 0.02467883750796318, + 0.013659649528563023, + 0.009891991503536701, + 0.02600458264350891, + 0.012501065619289875, + 0.016030987724661827, + 0.003071543062105775, + 0.0029298020526766777, + 0.0037884421180933714, + 0.004046895075589418, + 0.00259082461707294, + 0.0025660768151283264, + 0.0035049591679126024, + 0.0018600457115098834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 1 + ], + "selected_reward": -0.0890204906463623 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033698905259370804 + ], + "opposite_logprob_tokens": [ + -3.408698797225952 + ], + "logprob_reward": -0.033698905259370804, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033698905259370804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14420771598815918, + 0.08975325524806976, + 0.13490240275859833, + 0.10187895596027374, + 0.04916653782129288, + 0.11948815733194351, + 0.036770109087228775, + 0.1128581091761589, + 0.13561661541461945, + 0.07535819709300995 + ], + "credit_map": [ + 0.019995512440800667, + 0.012378247454762459, + 0.02078952081501484, + 0.023159530013799667, + 0.01763579063117504, + 0.01498771458864212, + 0.0185385849326849, + 0.01672281324863434, + 0.010540167801082134, + 0.01351664587855339, + 0.010451534762978554, + 0.01388562098145485, + 0.009414108470082283, + 0.010186958126723766, + 0.013521553948521614, + 0.00823665875941515, + 0.038453154265880585, + 0.00710704829543829, + 0.0467817597091198, + 0.013450956903398037, + 0.005307558923959732, + 0.0059147365391254425, + 0.008786564692854881, + 0.009100616909563541, + 0.019984716549515724, + 0.006517332047224045, + 0.011655689217150211, + 0.04058676213026047, + 0.0035784209612756968, + 0.00635916693136096, + 0.006844228599220514, + 0.00635263929143548, + 0.008510636165738106, + 0.004963638260960579, + 0.00543974107131362, + 0.0048874192871153355, + 0.004781639203429222, + 0.009068027138710022, + 0.006151287816464901, + 0.005364148411899805, + 0.017311319708824158, + 0.011122205294668674, + 0.015875274315476418, + 0.011195671744644642, + 0.009488570503890514, + 0.028330914676189423, + 0.013551279902458191, + 0.01261292677372694, + 0.004914348945021629, + 0.004425378516316414, + 0.004729590378701687, + 0.004223308060318232, + 0.0027006675954908133, + 0.006618271116167307, + 0.0047651613131165504, + 0.004393383860588074, + 0.015215091407299042, + 0.015610488131642342, + 0.01171157043427229, + 0.01612555980682373, + 0.01851181872189045, + 0.012641306966543198, + 0.010901750065386295, + 0.01214052364230156, + 0.01741851307451725, + 0.015218188986182213, + 0.015333965420722961, + 0.017468653619289398, + 0.0194026418030262, + 0.01826469972729683, + 0.018154945224523544, + 0.014355001971125603, + 0.01051514595746994, + 0.008509498089551926, + 0.010086997412145138, + 0.0064711603336036205, + 0.01378015149384737, + 0.01102969329804182, + 0.008700991049408913, + 0.006264558993279934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 2 + ], + "selected_reward": -0.033698905259370804 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide shot captures a vast circular room, rendered in a high-contrast, noir-style anime, its curved walls constructed of massive, dark-gray stone blocks. In the center, a massive, multi-tiered bookshelf, constructed of dark, aged wood and wire, rotates with a silent, relentless slowness. Its countless ancient tomes, bound in crumbling, shadowy leather, bear not titles but mysterious, cryptic symbols that glow with an ethereal blue light, pulsing faintly in the darkness. The room's sole source of light comes from a magical sphere of crackling white energy floating near the high, domed ceiling; its erratic flickering causes the entire scene to flicker violently, casting long, twisting, fiberglass-like shadows that wriggle and dance across the polished, cold stone floor. In the foreground, a solitary, cloaked figure stands motionless, their body completely concealed by the heavy, dark fabric, their head tilted upward, watching the celestial rotation, a silent observer lost in the heart of an ancient, shimmering mystery.", + "seed": 24318, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_3144_seed_24318_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04518303349614143, + "vlm_logprob": -0.04518303349614143, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.04518303349614143, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10146959125995636 + ], + "opposite_logprob_tokens": [ + -3.8514695167541504 + ], + "logprob_reward": -0.10146959125995636, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.10146959125995636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22344492375850677, + 0.17161603271961212, + 0.08794745802879333, + 0.0741920918226242, + 0.0650135725736618, + 0.07691074162721634, + 0.07265764474868774, + 0.1147167980670929, + 0.06782463192939758, + 0.0456760935485363 + ], + "credit_map": [ + 0.02207900770008564, + 0.020162515342235565, + 0.02794710360467434, + 0.04731136932969093, + 0.021449217572808266, + 0.035595815628767014, + 0.020525997504591942, + 0.028373895213007927, + 0.010594512335956097, + 0.007890183478593826, + 0.021219493821263313, + 0.026353614404797554, + 0.014545147307217121, + 0.030110038816928864, + 0.03665750101208687, + 0.02424555830657482, + 0.0077482727356255054, + 0.006393822375684977, + 0.011232860386371613, + 0.008873526938259602, + 0.008930878713726997, + 0.016392365097999573, + 0.019132908433675766, + 0.009242825210094452, + 0.007945556193590164, + 0.005652009509503841, + 0.00958799384534359, + 0.00978289358317852, + 0.005457951221615076, + 0.012416780926287174, + 0.013255628757178783, + 0.010093272663652897, + 0.01002699788659811, + 0.00579442223533988, + 0.00850437581539154, + 0.011497939936816692, + 0.004892223980277777, + 0.01053284015506506, + 0.006550784222781658, + 0.007213990204036236, + 0.007575110998004675, + 0.00841633789241314, + 0.009036497212946415, + 0.016729183495044708, + 0.006471030879765749, + 0.012435189448297024, + 0.009871930815279484, + 0.006375463213771582, + 0.005017637275159359, + 0.0049750711768865585, + 0.007846861146390438, + 0.019642436876893044, + 0.003496413119137287, + 0.012209108099341393, + 0.013847908936440945, + 0.005622209049761295, + 0.014225435443222523, + 0.012575658969581127, + 0.01465196069329977, + 0.016944346949458122, + 0.009305709972977638, + 0.020245229825377464, + 0.016032230108976364, + 0.010736227966845036, + 0.006685513537377119, + 0.005429521668702364, + 0.008250538259744644, + 0.016488967463374138, + 0.00442079920321703, + 0.011133292689919472, + 0.010464342311024666, + 0.004951658193022013, + 0.005604248959571123, + 0.003538434160873294, + 0.0060822623781859875, + 0.006982885301113129, + 0.0035403696820139885, + 0.006369163282215595, + 0.007745717186480761, + 0.005813016556203365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.10146959125995636 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01256337109953165 + ], + "opposite_logprob_tokens": [ + -4.387563228607178 + ], + "logprob_reward": -0.01256337109953165, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01256337109953165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29427286982536316, + 0.21220120787620544, + 0.08056133985519409, + 0.14484389126300812, + 0.056146685034036636, + 0.03362166881561279, + 0.027955668047070503, + 0.07582668215036392, + 0.04986520856618881, + 0.024704791605472565 + ], + "credit_map": [ + 0.028101392090320587, + 0.03830209746956825, + 0.039400242269039154, + 0.04954957216978073, + 0.03806572034955025, + 0.046860478818416595, + 0.028450025245547295, + 0.025543315336108208, + 0.014573568478226662, + 0.012883543968200684, + 0.02190565876662731, + 0.02386111579835415, + 0.02103332057595253, + 0.03663891926407814, + 0.0549645759165287, + 0.026340482756495476, + 0.006501632742583752, + 0.00985343474894762, + 0.010659709572792053, + 0.016403593122959137, + 0.008934149518609047, + 0.012919927947223186, + 0.007595698349177837, + 0.007693185470998287, + 0.01699059084057808, + 0.01958037167787552, + 0.01745789311826229, + 0.017503801733255386, + 0.01792091876268387, + 0.02393925003707409, + 0.015979429706931114, + 0.015471626073122025, + 0.003936737775802612, + 0.003370411694049835, + 0.006200455129146576, + 0.012716900557279587, + 0.00504810456186533, + 0.009742884896695614, + 0.008134138770401478, + 0.006997049320489168, + 0.0036516329273581505, + 0.002536223502829671, + 0.005079243797808886, + 0.007282552774995565, + 0.0023815901950001717, + 0.005386045668274164, + 0.004157546907663345, + 0.003146831877529621, + 0.0022619045339524746, + 0.002481117844581604, + 0.0029859805945307016, + 0.005581458564847708, + 0.002523579401895404, + 0.004781702999025583, + 0.005109997466206551, + 0.002229922218248248, + 0.0077256495133042336, + 0.010860093869268894, + 0.0074188136495649815, + 0.011358710005879402, + 0.007903954945504665, + 0.011487698182463646, + 0.01278697233647108, + 0.006284779869019985, + 0.005758966784924269, + 0.005164404399693012, + 0.0062883635982871056, + 0.010887743905186653, + 0.004195664077997208, + 0.00810289941728115, + 0.0065824901685118675, + 0.002884668530896306, + 0.002278592437505722, + 0.0026816162280738354, + 0.0025766799226403236, + 0.004280515015125275, + 0.002621977822855115, + 0.00418511126190424, + 0.004098150413483381, + 0.001982145244255662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.01256337109953165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014205054379999638 + ], + "opposite_logprob_tokens": [ + -4.264204978942871 + ], + "logprob_reward": -0.014205054379999638, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014205054379999638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27924129366874695, + 0.20559383928775787, + 0.14114946126937866, + 0.06520168483257294, + 0.06798065453767776, + 0.06949324160814285, + 0.043302521109580994, + 0.05340486392378807, + 0.05005756393074989, + 0.0245748832821846 + ], + "credit_map": [ + 0.024106184020638466, + 0.029986754059791565, + 0.03224467486143112, + 0.0477895587682724, + 0.02608851157128811, + 0.04223622381687164, + 0.03544750064611435, + 0.041341885924339294, + 0.014170458540320396, + 0.013279536738991737, + 0.03455796092748642, + 0.0353478267788887, + 0.01661843992769718, + 0.035624947398900986, + 0.02641964517533779, + 0.0295750442892313, + 0.013737614266574383, + 0.022431373596191406, + 0.018602102994918823, + 0.018052982166409492, + 0.01666909269988537, + 0.022526627406477928, + 0.01764926314353943, + 0.011480411514639854, + 0.008781853131949902, + 0.005841620732098818, + 0.007888678461313248, + 0.006806975230574608, + 0.005320245400071144, + 0.011727656237781048, + 0.009965709410607815, + 0.008868947625160217, + 0.008559056557714939, + 0.006769488099962473, + 0.008168740198016167, + 0.008110462687909603, + 0.00798085331916809, + 0.010806197300553322, + 0.008684861473739147, + 0.008900999091565609, + 0.005543755833059549, + 0.005655760411173105, + 0.009987013414502144, + 0.014500634744763374, + 0.004770003259181976, + 0.011355862952768803, + 0.009591471403837204, + 0.008088750764727592, + 0.005379119422286749, + 0.002592463744804263, + 0.0044129760935902596, + 0.004462572745978832, + 0.0036270786076784134, + 0.0071084024384617805, + 0.009305322542786598, + 0.006414588540792465, + 0.006689727306365967, + 0.005784212611615658, + 0.005933180917054415, + 0.008621062152087688, + 0.004849940072745085, + 0.00947373453527689, + 0.007473524659872055, + 0.004579488653689623, + 0.004898851737380028, + 0.004100185818970203, + 0.006153461989015341, + 0.012312769889831543, + 0.0035082621034234762, + 0.008009939454495907, + 0.007151228841394186, + 0.003922868054360151, + 0.0027737279888242483, + 0.002274584723636508, + 0.0023461359087377787, + 0.0041643488220870495, + 0.0025200447998940945, + 0.004176310263574123, + 0.004876183345913887, + 0.0014435483608394861 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.014205054379999638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636267855763435 + ], + "opposite_logprob_tokens": [ + -4.758636474609375 + ], + "logprob_reward": -0.008636267855763435, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636267855763435, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17371460795402527, + 0.12413769960403442, + 0.15623100101947784, + 0.1632186621427536, + 0.11377479135990143, + 0.0384780690073967, + 0.028816798701882362, + 0.07228779047727585, + 0.06502436101436615, + 0.06431619077920914 + ], + "credit_map": [ + 0.015101075172424316, + 0.014753506518900394, + 0.018113646656274796, + 0.018568389117717743, + 0.021161068230867386, + 0.028207672759890556, + 0.020602887496352196, + 0.03720635920763016, + 0.010690332390367985, + 0.010200470685958862, + 0.010452277958393097, + 0.018847212195396423, + 0.010543071664869785, + 0.02135336771607399, + 0.024800432845950127, + 0.017250532284379005, + 0.013917102478444576, + 0.013878176920115948, + 0.019430505111813545, + 0.01743467152118683, + 0.01601090282201767, + 0.025647902861237526, + 0.029184453189373016, + 0.020727289840579033, + 0.01662679761648178, + 0.007621935103088617, + 0.015867117792367935, + 0.01704627089202404, + 0.013684865087270737, + 0.027605045586824417, + 0.03965277969837189, + 0.02511386200785637, + 0.010971049778163433, + 0.00882792565971613, + 0.010432098060846329, + 0.016054892912507057, + 0.019972847774624825, + 0.018743624910712242, + 0.016449322924017906, + 0.012323034927248955, + 0.003232539864256978, + 0.00402564462274313, + 0.004767003934830427, + 0.006920650601387024, + 0.00333751505240798, + 0.006288966629654169, + 0.006128858309239149, + 0.0037768911570310593, + 0.0030087395571172237, + 0.001788764027878642, + 0.0030112385284155607, + 0.005792846437543631, + 0.001752100302837789, + 0.004735208582133055, + 0.004678803030401468, + 0.004049096722155809, + 0.008480562828481197, + 0.0051930369809269905, + 0.007877213880419731, + 0.007629925850778818, + 0.007152656093239784, + 0.012218156829476357, + 0.017298124730587006, + 0.006438107695430517, + 0.00919404998421669, + 0.006721621844917536, + 0.006812208332121372, + 0.012078702449798584, + 0.005170161835849285, + 0.010419323109090328, + 0.009543641470372677, + 0.00508465338498354, + 0.00629441486671567, + 0.0028473427519202232, + 0.004895100835710764, + 0.00632678484544158, + 0.005070116836577654, + 0.012437903322279453, + 0.020992932841181755, + 0.005451603326946497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.008636267855763435 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904088288545609 + ], + "opposite_logprob_tokens": [ + -2.464040994644165 + ], + "logprob_reward": -0.08904088288545609, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904088288545609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16037710011005402, + 0.21259668469429016, + 0.055206213146448135, + 0.054610725492239, + 0.10934395343065262, + 0.11128398030996323, + 0.036988917738199234, + 0.19189947843551636, + 0.036557380110025406, + 0.03113558143377304 + ], + "credit_map": [ + 0.014703498221933842, + 0.014131131581962109, + 0.020341238006949425, + 0.033483367413282394, + 0.011746973730623722, + 0.021852171048521996, + 0.021923774853348732, + 0.022194935008883476, + 0.017435651272535324, + 0.01597425900399685, + 0.03480418398976326, + 0.021090742200613022, + 0.022081829607486725, + 0.03526528924703598, + 0.03753393888473511, + 0.028410756960511208, + 0.004466526675969362, + 0.00477322144433856, + 0.0077504548244178295, + 0.010135037824511528, + 0.0058317300863564014, + 0.009073792025446892, + 0.007798794191330671, + 0.005376654677093029, + 0.005537303630262613, + 0.006244469899684191, + 0.0059274109080433846, + 0.005475183948874474, + 0.0055650160647928715, + 0.008742650970816612, + 0.009189681150019169, + 0.007929006591439247, + 0.01090980414301157, + 0.007717561442404985, + 0.013700107112526894, + 0.0292331725358963, + 0.00804441049695015, + 0.019508175551891327, + 0.009912625886499882, + 0.010318086482584476, + 0.008252718485891819, + 0.005986528005450964, + 0.017379552125930786, + 0.028040027245879173, + 0.005788508802652359, + 0.019473372027277946, + 0.01647179014980793, + 0.009891485795378685, + 0.0022707893513143063, + 0.0035343049094080925, + 0.0032235817052423954, + 0.00586842093616724, + 0.0031129277776926756, + 0.006440724246203899, + 0.009116923436522484, + 0.003421243978664279, + 0.020473917946219444, + 0.02336280420422554, + 0.030732518061995506, + 0.04084828495979309, + 0.012381821870803833, + 0.028207791969180107, + 0.02105037122964859, + 0.014841940253973007, + 0.0036665762308984995, + 0.0029939087107777596, + 0.004569496028125286, + 0.007929563522338867, + 0.002805038820952177, + 0.005929569713771343, + 0.005922835785895586, + 0.002740388736128807, + 0.002937827492132783, + 0.003271441673859954, + 0.0034952927380800247, + 0.005964512471109629, + 0.003017696551978588, + 0.004808065015822649, + 0.004637464880943298, + 0.003003279911354184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.08904088288545609 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, wide shot captures a vast circular room, rendered in a high-contrast, noir-style anime, its curved walls constructed of massive, dark-gray stone blocks. In the center, a massive, multi-tiered bookshelf, constructed of dark, aged wood and wire, rotates with a silent, relentless slowness. Its countless ancient tomes, bound in crumbling, shadowy leather, bear not titles but mysterious, cryptic symbols that glow with an ethereal blue light, pulsing faintly in the darkness. The room's sole source of light comes from a magical sphere of crackling white energy floating near the high, domed ceiling; its erratic flickering causes the entire scene to flicker violently, casting long, twisting, fiberglass-like shadows that wriggle and dance across the polished, cold stone floor. In the foreground, a solitary, cloaked figure stands motionless, their body completely concealed by the heavy, dark fabric, their head tilted upward, watching the celestial rotation, a silent observer lost in the heart of an ancient, shimmering mystery.", + "seed": 24319, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_15/index_3144_seed_24319_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05251057967543602, + "vlm_logprob": -0.05251057967543602, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05251057967543602, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0632171630859375 + ], + "opposite_logprob_tokens": [ + -3.4382171630859375 + ], + "logprob_reward": -0.0632171630859375, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0632171630859375, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1643487960100174, + 0.06171233206987381, + 0.053477007895708084, + 0.04605668783187866, + 0.05948427692055702, + 0.05629447102546692, + 0.051129113882780075, + 0.11099457740783691, + 0.3550230860710144, + 0.041479647159576416 + ], + "credit_map": [ + 0.026551734656095505, + 0.01620478183031082, + 0.017528237774968147, + 0.01622546650469303, + 0.011490902863442898, + 0.03381100669503212, + 0.015931738540530205, + 0.026604915037751198, + 0.009108529426157475, + 0.0058427611365914345, + 0.004687880165874958, + 0.007363337092101574, + 0.0039048781618475914, + 0.010448704473674297, + 0.014933948405086994, + 0.005422290880233049, + 0.007359404116868973, + 0.003248520428314805, + 0.0034349458292126656, + 0.014092834666371346, + 0.0027749433647841215, + 0.009331353940069675, + 0.01011644583195448, + 0.0031185552943497896, + 0.0059342836029827595, + 0.006948727648705244, + 0.0037783475127071142, + 0.0062921177595853806, + 0.006541167385876179, + 0.006416399497538805, + 0.006527372635900974, + 0.0036182685289531946, + 0.0070051224902272224, + 0.004541333764791489, + 0.005006089806556702, + 0.006648173090070486, + 0.004039933905005455, + 0.009152008220553398, + 0.013568468391895294, + 0.00952314306050539, + 0.006903586443513632, + 0.007989414036273956, + 0.0055251759476959705, + 0.006844870746135712, + 0.004685491789132357, + 0.007841397076845169, + 0.006023262161761522, + 0.010481266304850578, + 0.0064352755434811115, + 0.004426086321473122, + 0.004808127414435148, + 0.00805133581161499, + 0.005125786177814007, + 0.008596953004598618, + 0.009270012378692627, + 0.004415533971041441, + 0.01400165818631649, + 0.008092833682894707, + 0.015557745471596718, + 0.019019801169633865, + 0.005896393675357103, + 0.014718863181769848, + 0.01608567126095295, + 0.017621610313653946, + 0.04429573565721512, + 0.06724943220615387, + 0.02625357359647751, + 0.049073487520217896, + 0.03131238743662834, + 0.047162942588329315, + 0.047458548098802567, + 0.04221697896718979, + 0.005686952266842127, + 0.0053564500994980335, + 0.004335269331932068, + 0.005555045325309038, + 0.003234216244891286, + 0.006265539675951004, + 0.007910439744591713, + 0.0031357337720692158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0632171630859375 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011096080765128136 + ], + "opposite_logprob_tokens": [ + -4.511096000671387 + ], + "logprob_reward": -0.011096080765128136, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011096080765128136, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2544794976711273, + 0.04967121034860611, + 0.031054411083459854, + 0.024908224120736122, + 0.07126068323850632, + 0.04241439700126648, + 0.07057639956474304, + 0.03362405672669411, + 0.370686411857605, + 0.05132479593157768 + ], + "credit_map": [ + 0.03242434933781624, + 0.027239911258220673, + 0.029857732355594635, + 0.0260267723351717, + 0.023578202351927757, + 0.029461557045578957, + 0.02710619568824768, + 0.058784764260053635, + 0.007853214628994465, + 0.0035434220917522907, + 0.005268700420856476, + 0.0074321129359304905, + 0.0029715734999626875, + 0.009619668126106262, + 0.0062050293199718, + 0.006777486763894558, + 0.005273445043712854, + 0.0023080785758793354, + 0.002896746387705207, + 0.00529314624145627, + 0.002149696461856365, + 0.00653406698256731, + 0.004595547914505005, + 0.0020036804489791393, + 0.004137116484344006, + 0.0023723579943180084, + 0.0020526840817183256, + 0.0037518462631851435, + 0.0019527726108208299, + 0.0035841972567141056, + 0.004439841024577618, + 0.0026174045633524656, + 0.010373790748417377, + 0.007242708932608366, + 0.005631015170365572, + 0.0070775882340967655, + 0.0042499671690166, + 0.012642703019082546, + 0.011516068130731583, + 0.01252683438360691, + 0.0054481192491948605, + 0.005582667421549559, + 0.004341007210314274, + 0.005047800950706005, + 0.002995183924213052, + 0.006964321248233318, + 0.004215934779495001, + 0.00781935639679432, + 0.009101425297558308, + 0.007446741685271263, + 0.007083766628056765, + 0.011079693213105202, + 0.005931484512984753, + 0.01181832142174244, + 0.012475250288844109, + 0.005639705341309309, + 0.004095441661775112, + 0.0022898700553923845, + 0.004048389382660389, + 0.006562694441527128, + 0.002039827173575759, + 0.00490319449454546, + 0.004519525915384293, + 0.005165106616914272, + 0.0487784780561924, + 0.0675124004483223, + 0.02954976074397564, + 0.048146914690732956, + 0.030349774286150932, + 0.05154189094901085, + 0.0473695769906044, + 0.04743758216500282, + 0.008568614721298218, + 0.0038545397110283375, + 0.005372714716941118, + 0.008137864992022514, + 0.003048859303817153, + 0.009640620090067387, + 0.006860222201794386, + 0.00584135577082634 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.011096080765128136 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544417753815651 + ], + "opposite_logprob_tokens": [ + -4.387544631958008 + ], + "logprob_reward": -0.012544417753815651, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544417753815651, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1640395224094391, + 0.06700035184621811, + 0.08926880359649658, + 0.053160879760980606, + 0.16981364786624908, + 0.081988126039505, + 0.18219050765037537, + 0.06265904754400253, + 0.07523791491985321, + 0.054641228169202805 + ], + "credit_map": [ + 0.020676571875810623, + 0.03072292171418667, + 0.015036014840006828, + 0.024328185245394707, + 0.015787480399012566, + 0.024417610839009285, + 0.014244570396840572, + 0.018826168030500412, + 0.010873125866055489, + 0.0081272441893816, + 0.006804157979786396, + 0.006850257050246, + 0.004070102237164974, + 0.011948113329708576, + 0.010243368335068226, + 0.008083983324468136, + 0.011602873913943768, + 0.005485735833644867, + 0.005899589508771896, + 0.016238078474998474, + 0.006900659296661615, + 0.013553322292864323, + 0.01845022849738598, + 0.011138313449919224, + 0.006712250877171755, + 0.007805097848176956, + 0.004196731373667717, + 0.0070507884956896305, + 0.005205847322940826, + 0.006926191039383411, + 0.010903993621468544, + 0.004359979648143053, + 0.021620312705636024, + 0.009478208608925343, + 0.013648158870637417, + 0.017126547172665596, + 0.012034890241920948, + 0.026164226233959198, + 0.038648657500743866, + 0.031092653051018715, + 0.010040540248155594, + 0.007968836463987827, + 0.013233885169029236, + 0.007563078310340643, + 0.007256681565195322, + 0.010711119510233402, + 0.008138079196214676, + 0.017075907438993454, + 0.01753593422472477, + 0.02196122519671917, + 0.015885286033153534, + 0.035164110362529755, + 0.02279093861579895, + 0.02077505551278591, + 0.03645610436797142, + 0.011621852405369282, + 0.008648416958749294, + 0.003921191208064556, + 0.007470841985195875, + 0.009363963268697262, + 0.004079196602106094, + 0.009463271126151085, + 0.008817289024591446, + 0.010894880630075932, + 0.009691179729998112, + 0.013046045787632465, + 0.008249466307461262, + 0.007595603819936514, + 0.006696344818919897, + 0.010112249292433262, + 0.009776228107511997, + 0.010070792399346828, + 0.008996865712106228, + 0.003846857463940978, + 0.004706642124801874, + 0.009080554358661175, + 0.0040230597369372845, + 0.010324424132704735, + 0.009281924925744534, + 0.00438090180978179 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.012544417753815651 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12703682482242584 + ], + "opposite_logprob_tokens": [ + -2.1270368099212646 + ], + "logprob_reward": -0.12703682482242584, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12703682482242584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1407039761543274, + 0.035162314772605896, + 0.015945160761475563, + 0.023425454273819923, + 0.5569717884063721, + 0.02846454083919525, + 0.029518477618694305, + 0.04237785190343857, + 0.10438951104879379, + 0.023040924221277237 + ], + "credit_map": [ + 0.017074786126613617, + 0.018208811059594154, + 0.015558400191366673, + 0.015803437680006027, + 0.013331198133528233, + 0.018515098839998245, + 0.014983599074184895, + 0.027228645980358124, + 0.003760806517675519, + 0.0038139375392347574, + 0.0035651589278131723, + 0.0063531952910125256, + 0.002657008357346058, + 0.004618827719241381, + 0.007475084625184536, + 0.0029182955622673035, + 0.001821367535740137, + 0.0013670194894075394, + 0.0015221553621813655, + 0.002888710703700781, + 0.001665520598180592, + 0.0021722414530813694, + 0.0029584565199911594, + 0.0015496892156079412, + 0.002676805481314659, + 0.0032449904829263687, + 0.0020995172671973705, + 0.003757930127903819, + 0.001813410548493266, + 0.002860110718756914, + 0.0034627430140972137, + 0.0035099468659609556, + 0.07124151289463043, + 0.060179971158504486, + 0.045620452612638474, + 0.05354691296815872, + 0.03197464346885681, + 0.08037081360816956, + 0.08950208127498627, + 0.12453541159629822, + 0.003482065862044692, + 0.0022864246275275946, + 0.006158505100756884, + 0.002847913419827819, + 0.0017292400589212775, + 0.004055489785969257, + 0.002454797737300396, + 0.005450104828923941, + 0.003608092200011015, + 0.0034308666363358498, + 0.002938894322142005, + 0.003940494731068611, + 0.0028238447848707438, + 0.003737084800377488, + 0.00553599838167429, + 0.0035032012965530157, + 0.005144530441612005, + 0.0036538683343678713, + 0.0031153010204434395, + 0.010181967169046402, + 0.00347843486815691, + 0.006801907904446125, + 0.006231239996850491, + 0.0037705996073782444, + 0.01386326551437378, + 0.019992023706436157, + 0.008442469872534275, + 0.012226530350744724, + 0.009338281117379665, + 0.015542466193437576, + 0.013135351240634918, + 0.011849119327962399, + 0.0028474440332502127, + 0.0020568440668284893, + 0.0022150620352476835, + 0.004714237526059151, + 0.0017265004571527243, + 0.0034378201235085726, + 0.004106583073735237, + 0.001936432789079845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.12703682482242584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera angle low and the shot static?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a vast circular room with curved walls made of massive dark-gray stone blocks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a massive multi-tiered bookshelf in the center constructed of dark aged wood and wire that rotates slowly?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a solitary cloaked figure standing motionless in the foreground with their head tilted upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting provided by a magical sphere of crackling white energy near the domed ceiling causing violent flickering and long shadows?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865841194987297 + ], + "opposite_logprob_tokens": [ + -3.0486583709716797 + ], + "logprob_reward": -0.04865841194987297, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865841194987297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10158140957355499, + 0.0755094438791275, + 0.04561643674969673, + 0.08473959565162659, + 0.08856524527072906, + 0.06880491226911545, + 0.061982639133930206, + 0.05219339579343796, + 0.3943311870098114, + 0.026675758883357048 + ], + "credit_map": [ + 0.01285783015191555, + 0.009001125581562519, + 0.012294787913560867, + 0.011545763351023197, + 0.00863613560795784, + 0.013868252746760845, + 0.014605077914893627, + 0.018772438168525696, + 0.008182554505765438, + 0.006659033242613077, + 0.007603097707033157, + 0.01188142504543066, + 0.00485304556787014, + 0.010193641297519207, + 0.01639983430504799, + 0.00973681453615427, + 0.007337275426834822, + 0.004737101495265961, + 0.003899981500580907, + 0.0060391537845134735, + 0.0034341372083872557, + 0.007580336183309555, + 0.006562601309269667, + 0.006025854032486677, + 0.01207395363599062, + 0.00856063049286604, + 0.006590174976736307, + 0.009343961253762245, + 0.014297827146947384, + 0.014680445194244385, + 0.010592227801680565, + 0.008600378409028053, + 0.010932723060250282, + 0.011689682491123676, + 0.007847408764064312, + 0.00904015637934208, + 0.005396273452788591, + 0.012070118449628353, + 0.013373201712965965, + 0.01821567863225937, + 0.00748086255043745, + 0.011425122618675232, + 0.009326211176812649, + 0.00923570804297924, + 0.0057986765168607235, + 0.008648605085909367, + 0.007921815849840641, + 0.008967908099293709, + 0.008656032383441925, + 0.0067687863484025, + 0.004613903351128101, + 0.00848023034632206, + 0.0068612112663686275, + 0.009222269058227539, + 0.010303677059710026, + 0.007076524198055267, + 0.005862298421561718, + 0.004286606330424547, + 0.005186166614294052, + 0.011953342705965042, + 0.004425886087119579, + 0.007276664022356272, + 0.00790606439113617, + 0.0052963667549192905, + 0.04481752589344978, + 0.08775205165147781, + 0.03047359175980091, + 0.056546956300735474, + 0.03919325768947601, + 0.05003027990460396, + 0.05794288590550423, + 0.027574613690376282, + 0.0032999319955706596, + 0.002545934868976474, + 0.0037357043474912643, + 0.0037455062847584486, + 0.002108487533405423, + 0.004015173297375441, + 0.0034905271604657173, + 0.003734494559466839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.04865841194987297 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_tensor([1514, 1514])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_tensor([1514, 1514])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3ba911572669aea4d8d3ad755f609c589a74a0a4 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_tensor([1514, 1514])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins from a slight upward angle, the camera slowly and steadily pulls back to reveal a sacred golden-red phoenix hovering in the deep night sky, a backdrop dotted with a tranquil sea of stars, punctuated by countless tiny stars. The phoenix is draped in shimmering golden feathers, their edges and underbelly gleaming with a fiery crimson glow. Its slender neck gracefully arches, its eyes like molten gold gaze into the distance, its powerful wings fully extended, maintaining a posture of tranquility and power. As its body unfurls, its magnificent tail feathers—composed of countless feathers that resemble the condensation of pure flame—unfold layer by layer in rhythmic slow motion. Each feather has a sharp edge, its center ablaze with golden light, gradually fading to a transparent orange-red. The entire unfolding process resembles a vast cosmic flower of light and fire blooming in the void. The phoenix itself is the only light source, and its blazing light forms a warm halo around it, giving the entire picture an atmosphere of solemnity, awe and fantasy.", + "seed": 25683, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25683_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1181848106905818, + "vlm_logprob": -0.1181848106905818, + "vlm_yes_no_margin": 2.55, + "vlm_token_credit": -0.1181848106905818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19335207343101501 + ], + "opposite_logprob_tokens": [ + -1.8183521032333374 + ], + "logprob_reward": -0.19335207343101501, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19335207343101501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26858630776405334, + 0.08216297626495361, + 0.09373478591442108, + 0.14037658274173737, + 0.042219214141368866, + 0.021733136847615242, + 0.036911383271217346, + 0.2107153683900833, + 0.044154174625873566, + 0.0594061017036438 + ], + "credit_map": [ + 0.0787491425871849, + 0.11528252810239792, + 0.018448302522301674, + 0.00903155654668808, + 0.02008916810154915, + 0.01165292039513588, + 0.008498135022819042, + 0.006834549363702536, + 0.013611448928713799, + 0.016254432499408722, + 0.009583857841789722, + 0.0076364800333976746, + 0.006149648688733578, + 0.00877496600151062, + 0.007172341458499432, + 0.012979802675545216, + 0.01571541838347912, + 0.014223971404135227, + 0.018706411123275757, + 0.007685247343033552, + 0.012293110601603985, + 0.007505265064537525, + 0.010255939327180386, + 0.007349415682256222, + 0.038048747926950455, + 0.012576420791447163, + 0.0508820116519928, + 0.009332859888672829, + 0.008910370990633965, + 0.0073778447695076466, + 0.005064512137323618, + 0.008183806203305721, + 0.005557833705097437, + 0.0030013713985681534, + 0.0051673296838998795, + 0.004738561809062958, + 0.0044151912443339825, + 0.004710290115326643, + 0.00948688667267561, + 0.005141743458807468, + 0.002144559985026717, + 0.0014670680975541472, + 0.002309119561687112, + 0.002241360954940319, + 0.00218777172267437, + 0.002028899034485221, + 0.007209856063127518, + 0.002144500147551298, + 0.0038848232943564653, + 0.004299431573599577, + 0.005685259122401476, + 0.002685732673853636, + 0.004382261540740728, + 0.006096804514527321, + 0.003483254462480545, + 0.00639381306245923, + 0.009825984947383404, + 0.013449796475470066, + 0.040878210216760635, + 0.0795200765132904, + 0.022318262606859207, + 0.01267259381711483, + 0.0188905056566, + 0.013159919530153275, + 0.0055294823832809925, + 0.004152204841375351, + 0.011401667259633541, + 0.004812685772776604, + 0.005655596032738686, + 0.003852792549878359, + 0.004301491659134626, + 0.004448253195732832, + 0.006946349982172251, + 0.0077905477955937386, + 0.007928350940346718, + 0.008090244606137276, + 0.006122511811554432, + 0.005448541138321161, + 0.007550605572760105, + 0.009528943337500095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.19335207343101501 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16036830842494965 + ], + "opposite_logprob_tokens": [ + -1.9103683233261108 + ], + "logprob_reward": -0.16036830842494965, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16036830842494965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.036403339356184006, + 0.033061686903238297, + 0.04442545026540756, + 0.0692608505487442, + 0.04648585245013237, + 0.019703544676303864, + 0.04906192049384117, + 0.6057705283164978, + 0.06081219017505646, + 0.03501461446285248 + ], + "credit_map": [ + 0.006600214168429375, + 0.010970162227749825, + 0.0051393709145486355, + 0.00300841243006289, + 0.002780755516141653, + 0.00263209524564445, + 0.003016564529389143, + 0.0022557650227099657, + 0.005219125188887119, + 0.004580567125231028, + 0.004277445841580629, + 0.0031866610515862703, + 0.0029169966001063585, + 0.004908833187073469, + 0.003150208154693246, + 0.004821851383894682, + 0.00954778864979744, + 0.004513631109148264, + 0.015038356184959412, + 0.0030373006593436003, + 0.0033073073718696833, + 0.002718785312026739, + 0.0032567083835601807, + 0.003005573758855462, + 0.013403456658124924, + 0.011996501125395298, + 0.015722457319498062, + 0.005970320198684931, + 0.005925155710428953, + 0.0051362160593271255, + 0.006271066144108772, + 0.004835677333176136, + 0.009011692367494106, + 0.003446723334491253, + 0.0069456035271286964, + 0.00532556651160121, + 0.0034220460802316666, + 0.0050463867373764515, + 0.007790753617882729, + 0.0054970793426036835, + 0.00199105660431087, + 0.002337055979296565, + 0.0022090086713433266, + 0.0019737014081329107, + 0.0023262142203748226, + 0.003513377159833908, + 0.003299183677881956, + 0.0020539474207907915, + 0.004935014992952347, + 0.005353226792067289, + 0.007688898127526045, + 0.00429962994530797, + 0.007223495282232761, + 0.005223757121711969, + 0.00524513702839613, + 0.00909275934100151, + 0.029529770836234093, + 0.03940042480826378, + 0.1205686703324318, + 0.22134673595428467, + 0.0710858479142189, + 0.036873314529657364, + 0.051315899938344955, + 0.035649921745061874, + 0.009677411057054996, + 0.005694011691957712, + 0.010274598374962807, + 0.004785728175193071, + 0.005961803253740072, + 0.005402897950261831, + 0.008361976593732834, + 0.010653759352862835, + 0.004755668807774782, + 0.0049296896904706955, + 0.005578991957008839, + 0.004316363949328661, + 0.0034554211888462305, + 0.0029168485198169947, + 0.004271724261343479, + 0.004789907950907946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.16036830842494965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35821688175201416, + 0.05134229734539986, + 0.2578839957714081, + 0.09574796259403229, + 0.01883763074874878, + 0.012914543971419334, + 0.020875347778201103, + 0.13279739022254944, + 0.03788965567946434, + 0.01349437702447176 + ], + "credit_map": [ + 0.08742865175008774, + 0.15985672175884247, + 0.029824240133166313, + 0.01232216227799654, + 0.029745526611804962, + 0.018721742555499077, + 0.010552718304097652, + 0.00976509228348732, + 0.007317572832107544, + 0.005102318245917559, + 0.006085123401135206, + 0.0044095516204833984, + 0.0057274410501122475, + 0.006493467837572098, + 0.008199100382626057, + 0.008007720112800598, + 0.056348539888858795, + 0.021041277796030045, + 0.07533573359251022, + 0.02216785028576851, + 0.028736092150211334, + 0.01703954115509987, + 0.02230643853545189, + 0.014908505603671074, + 0.02437652461230755, + 0.007258400321006775, + 0.035576850175857544, + 0.006680490914732218, + 0.0066402386873960495, + 0.004274291917681694, + 0.006045665591955185, + 0.004895488265901804, + 0.0036005142610520124, + 0.0021128307562321424, + 0.003037736052647233, + 0.001993961166590452, + 0.0014368599513545632, + 0.002299700165167451, + 0.0021368719171732664, + 0.0022191540338099003, + 0.0009090852690860629, + 0.0010084799723699689, + 0.001278679701499641, + 0.0013119272189214826, + 0.001189369591884315, + 0.00152514164801687, + 0.004117490258067846, + 0.0015743689145892859, + 0.0020595903042703867, + 0.0030913420487195253, + 0.003276206785812974, + 0.00252539967186749, + 0.002247265074402094, + 0.002657366683706641, + 0.0023463580291718245, + 0.002671818481758237, + 0.006216839421540499, + 0.007929573766887188, + 0.0280627328902483, + 0.05066882073879242, + 0.013712423853576183, + 0.00799198355525732, + 0.01110625360161066, + 0.007108758669346571, + 0.005768067203462124, + 0.0032216571271419525, + 0.009145783260464668, + 0.0038000226486474276, + 0.0031827304046601057, + 0.003471450647339225, + 0.0046021598391234875, + 0.0046977815218269825, + 0.0014774882001802325, + 0.0019434201531112194, + 0.0016783089376986027, + 0.0019688571337610483, + 0.0012406606692820787, + 0.0012153609422966838, + 0.001737723359838128, + 0.00223255786113441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865216836333275 + ], + "opposite_logprob_tokens": [ + -3.048652172088623 + ], + "logprob_reward": -0.04865216836333275, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865216836333275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07726017385721207, + 0.10927305370569229, + 0.15968932211399078, + 0.15290246903896332, + 0.050121285021305084, + 0.08852583914995193, + 0.12382453680038452, + 0.16858822107315063, + 0.0345628522336483, + 0.03525226563215256 + ], + "credit_map": [ + 0.01523733139038086, + 0.02081744745373726, + 0.008259743452072144, + 0.008050891570746899, + 0.005934680346399546, + 0.005832070484757423, + 0.007194141857326031, + 0.005933861248195171, + 0.014306232333183289, + 0.014688961207866669, + 0.015598091296851635, + 0.010663270950317383, + 0.0132748419418931, + 0.012276450172066689, + 0.015336807817220688, + 0.013128400780260563, + 0.03132542967796326, + 0.014728674665093422, + 0.035212572664022446, + 0.011164550669491291, + 0.019675370305776596, + 0.01432647556066513, + 0.018670760095119476, + 0.014585490338504314, + 0.03040209785103798, + 0.022121986374258995, + 0.03550872951745987, + 0.008840986527502537, + 0.012747022323310375, + 0.016187766566872597, + 0.012283924967050552, + 0.014809958636760712, + 0.006672692485153675, + 0.005007543601095676, + 0.00667483638972044, + 0.0054222834296524525, + 0.007323705591261387, + 0.0052134389989078045, + 0.006024369969964027, + 0.007782415021210909, + 0.006462363991886377, + 0.00813439954072237, + 0.010510559193789959, + 0.009021858684718609, + 0.00991506315767765, + 0.007239175960421562, + 0.028852304443717003, + 0.008390119299292564, + 0.009789862670004368, + 0.011685105971992016, + 0.018708204850554466, + 0.012681719847023487, + 0.0159800723195076, + 0.019229883328080177, + 0.01652630604803562, + 0.019223373383283615, + 0.009572775103151798, + 0.013977507129311562, + 0.026829035952687263, + 0.060762569308280945, + 0.01087773684412241, + 0.011307589709758759, + 0.020828302949666977, + 0.014432705007493496, + 0.005892538465559483, + 0.001967053161934018, + 0.010309016332030296, + 0.0032527842558920383, + 0.002196531742811203, + 0.0028839153237640858, + 0.005152047146111727, + 0.0029089655727148056, + 0.004182131960988045, + 0.0035733100958168507, + 0.004811575636267662, + 0.004325645510107279, + 0.002981791039928794, + 0.002542330650612712, + 0.0051206182688474655, + 0.0077148620039224625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.04865216836333275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799032986164093 + ], + "opposite_logprob_tokens": [ + -1.804903268814087 + ], + "logprob_reward": -0.1799032986164093, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799032986164093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10483333468437195, + 0.09134555608034134, + 0.07540441304445267, + 0.22825230658054352, + 0.04166029766201973, + 0.04179159924387932, + 0.057694002985954285, + 0.21127787232398987, + 0.10334267467260361, + 0.044397857040166855 + ], + "credit_map": [ + 0.020133236423134804, + 0.02708694525063038, + 0.018626239150762558, + 0.008835171349346638, + 0.0089040482416749, + 0.008137017488479614, + 0.006564562674611807, + 0.0065461271442472935, + 0.01279439590871334, + 0.012961684726178646, + 0.01332802139222622, + 0.009546281769871712, + 0.008851802907884121, + 0.009923746809363365, + 0.011906499043107033, + 0.012033137492835522, + 0.01514830719679594, + 0.01178436167538166, + 0.017332345247268677, + 0.006374229211360216, + 0.0073014432564377785, + 0.004898622166365385, + 0.006623181514441967, + 0.005941940471529961, + 0.0570186972618103, + 0.024086931720376015, + 0.078887440264225, + 0.015014341101050377, + 0.015000933781266212, + 0.013068970292806625, + 0.011347656138241291, + 0.013827376067638397, + 0.008664238266646862, + 0.004025330301374197, + 0.006583062931895256, + 0.00472394609823823, + 0.0043221465311944485, + 0.003635087516158819, + 0.005166796036064625, + 0.004539697431027889, + 0.003593156347051263, + 0.003722683759406209, + 0.004572248086333275, + 0.006391953211277723, + 0.0036939363926649094, + 0.003716233652085066, + 0.011026307009160519, + 0.005075086839497089, + 0.004903897177428007, + 0.007424852345138788, + 0.008943827822804451, + 0.008500426076352596, + 0.006215956527739763, + 0.008460599929094315, + 0.005750264972448349, + 0.007494182325899601, + 0.015496586449444294, + 0.017909666523337364, + 0.035167887806892395, + 0.07009761035442352, + 0.01355611439794302, + 0.016165373846888542, + 0.023341700434684753, + 0.019542977213859558, + 0.014908970333635807, + 0.01040028128772974, + 0.022108091041445732, + 0.010428651235997677, + 0.008881776593625546, + 0.010434959083795547, + 0.008589615114033222, + 0.01759033463895321, + 0.004941427614539862, + 0.005290400702506304, + 0.005566638894379139, + 0.00542978523299098, + 0.00541550712659955, + 0.003455270081758499, + 0.00543978251516819, + 0.008859055116772652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.1799032986164093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from a slight upward angle, the camera slowly and steadily pulls back to reveal a sacred golden-red phoenix hovering in the deep night sky, a backdrop dotted with a tranquil sea of stars, punctuated by countless tiny stars. The phoenix is draped in shimmering golden feathers, their edges and underbelly gleaming with a fiery crimson glow. Its slender neck gracefully arches, its eyes like molten gold gaze into the distance, its powerful wings fully extended, maintaining a posture of tranquility and power. As its body unfurls, its magnificent tail feathers—composed of countless feathers that resemble the condensation of pure flame—unfold layer by layer in rhythmic slow motion. Each feather has a sharp edge, its center ablaze with golden light, gradually fading to a transparent orange-red. The entire unfolding process resembles a vast cosmic flower of light and fire blooming in the void. The phoenix itself is the only light source, and its blazing light forms a warm halo around it, giving the entire picture an atmosphere of solemnity, awe and fantasy.", + "seed": 25684, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25684_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.14698127778247, + "vlm_logprob": -0.14698127778247, + "vlm_yes_no_margin": 2.374999976158142, + "vlm_token_credit": -0.14698127778247, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2956112027168274 + ], + "opposite_logprob_tokens": [ + -1.4206111431121826 + ], + "logprob_reward": -0.2956112027168274, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2956112027168274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10306461900472641, + 0.131235733628273, + 0.1838206797838211, + 0.04874271899461746, + 0.08980687707662582, + 0.06611470133066177, + 0.20096932351589203, + 0.09035618603229523, + 0.05465277284383774, + 0.031236402690410614 + ], + "credit_map": [ + 0.020834609866142273, + 0.009535923600196838, + 0.012715247459709644, + 0.0069748410023748875, + 0.012495389208197594, + 0.022627586498856544, + 0.004869882948696613, + 0.013011137954890728, + 0.020568443462252617, + 0.01466389000415802, + 0.017645783722400665, + 0.013533252291381359, + 0.012502559460699558, + 0.019996631890535355, + 0.011770885437726974, + 0.020554279908537865, + 0.019129855558276176, + 0.010338437743484974, + 0.009965373203158379, + 0.011070024222135544, + 0.008768410421907902, + 0.012279560789465904, + 0.06893406063318253, + 0.0433349646627903, + 0.007735353894531727, + 0.0035910068545490503, + 0.007376905996352434, + 0.00472834799438715, + 0.005976776126772165, + 0.0045557525008916855, + 0.006954205688089132, + 0.007824370637536049, + 0.012956003658473492, + 0.008923707529902458, + 0.00994415208697319, + 0.009145674295723438, + 0.014193476177752018, + 0.01350653637200594, + 0.0077894097194075584, + 0.013347918167710304, + 0.008377522230148315, + 0.010439455509185791, + 0.005592934787273407, + 0.004420692101120949, + 0.009761976078152657, + 0.010630563832819462, + 0.007295582443475723, + 0.009595973417162895, + 0.04666537418961525, + 0.013812633231282234, + 0.007298135664314032, + 0.010436891578137875, + 0.01909569464623928, + 0.05606871843338013, + 0.01237539853900671, + 0.035216476768255234, + 0.016558686271309853, + 0.008789300918579102, + 0.004709801636636257, + 0.005758917424827814, + 0.01078488864004612, + 0.022689172998070717, + 0.007378521375358105, + 0.013686894439160824, + 0.0077234795317053795, + 0.010250935330986977, + 0.005021050106734037, + 0.0050389813259243965, + 0.010142730548977852, + 0.004989015404134989, + 0.0050292168743908405, + 0.006457364186644554, + 0.0024688062258064747, + 0.002212653635069728, + 0.0036799742374569178, + 0.003160718595609069, + 0.0022883687634021044, + 0.005422514863312244, + 0.0030945991165935993, + 0.0089087700471282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 1 + ], + "selected_reward": -0.2956112027168274 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270802915096283 + ], + "opposite_logprob_tokens": [ + -2.127080202102661 + ], + "logprob_reward": -0.1270802915096283, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270802915096283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08921338617801666, + 0.07377232611179352, + 0.1139691099524498, + 0.06628569215536118, + 0.05727299302816391, + 0.1277744323015213, + 0.25450682640075684, + 0.10523644089698792, + 0.08829504996538162, + 0.023673683404922485 + ], + "credit_map": [ + 0.017442261800169945, + 0.0077216546051204205, + 0.011726194992661476, + 0.00681223114952445, + 0.01118405070155859, + 0.01836339198052883, + 0.004405761603266001, + 0.011557837948203087, + 0.007835269905626774, + 0.007560238242149353, + 0.010312661528587341, + 0.011939536780118942, + 0.005666142795234919, + 0.008879711851477623, + 0.008159298449754715, + 0.013419468887150288, + 0.02854723110795021, + 0.004439563024789095, + 0.006603207904845476, + 0.007829210720956326, + 0.004808196797966957, + 0.007466474547982216, + 0.03684151545166969, + 0.017433714121580124, + 0.006547355093061924, + 0.00422330666333437, + 0.013216073624789715, + 0.009609430097043514, + 0.007681590970605612, + 0.005677725654095411, + 0.011208230629563332, + 0.008121981285512447, + 0.006383771076798439, + 0.009603460319340229, + 0.006236107554286718, + 0.006997524295002222, + 0.009035860188305378, + 0.009046653285622597, + 0.003753896337002516, + 0.006215721368789673, + 0.015408352948725224, + 0.018976153805851936, + 0.011094415560364723, + 0.01014785934239626, + 0.018024541437625885, + 0.02111300267279148, + 0.01142974104732275, + 0.021580366417765617, + 0.053087156265974045, + 0.018177660182118416, + 0.013987588696181774, + 0.013191642239689827, + 0.030275540426373482, + 0.04583563655614853, + 0.01797298900783062, + 0.06197861582040787, + 0.018424017354846, + 0.008165488950908184, + 0.00593024305999279, + 0.00779105955734849, + 0.014782346785068512, + 0.022972317412495613, + 0.00898714829236269, + 0.018183831125497818, + 0.012799263000488281, + 0.014249109663069248, + 0.007339126896113157, + 0.007761725690215826, + 0.016692884266376495, + 0.014354896731674671, + 0.006121219135820866, + 0.008976821787655354, + 0.0028386719059199095, + 0.0032829889096319675, + 0.0027359991800040007, + 0.0035498167853802443, + 0.0036326253321021795, + 0.0025861812755465508, + 0.0022489360999315977, + 0.002798462985083461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 2 + ], + "selected_reward": -0.1270802915096283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639123592525721 + ], + "opposite_logprob_tokens": [ + -4.882638931274414 + ], + "logprob_reward": -0.007639123592525721, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639123592525721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0449344664812088, + 0.04269229248166084, + 0.09037533402442932, + 0.03817606717348099, + 0.029031775891780853, + 0.06512757390737534, + 0.49445781111717224, + 0.16117389500141144, + 0.02424371801316738, + 0.009787056595087051 + ], + "credit_map": [ + 0.006246218457818031, + 0.0035470423754304647, + 0.005800454877316952, + 0.004352778196334839, + 0.0066449022851884365, + 0.009111043997108936, + 0.0037570304702967405, + 0.005474995356053114, + 0.005237941164523363, + 0.004841095767915249, + 0.004607778508216143, + 0.0049652899615466595, + 0.004333636723458767, + 0.006677052937448025, + 0.004404234234243631, + 0.007625260856002569, + 0.02418942004442215, + 0.0027656794991344213, + 0.004424674436450005, + 0.004920138977468014, + 0.004407432861626148, + 0.007644536439329386, + 0.03168141469359398, + 0.010342036373913288, + 0.0042562903836369514, + 0.003473301650956273, + 0.0038252617232501507, + 0.005043040495365858, + 0.007823272608220577, + 0.0033202299382537603, + 0.006704411935061216, + 0.0037302591372281313, + 0.0033944069873541594, + 0.0023344899527728558, + 0.0037551678251475096, + 0.0033420627005398273, + 0.00551494350656867, + 0.004995943512767553, + 0.002885277383029461, + 0.0028094854205846786, + 0.00680320942774415, + 0.006106739863753319, + 0.004888228606432676, + 0.004099973011761904, + 0.00595810916274786, + 0.011341854929924011, + 0.006775361020117998, + 0.019154099747538567, + 0.09992410987615585, + 0.03448057919740677, + 0.0208031814545393, + 0.025662977248430252, + 0.05123279243707657, + 0.1344795525074005, + 0.031338077038526535, + 0.0965365469455719, + 0.03826042637228966, + 0.010285688564181328, + 0.007927009835839272, + 0.0076111434027552605, + 0.019169528037309647, + 0.05989206209778786, + 0.005752189084887505, + 0.01227584108710289, + 0.0035625521559268236, + 0.004431465175002813, + 0.0020478577353060246, + 0.0028185024857521057, + 0.0027313618920743465, + 0.003774615004658699, + 0.002762510674074292, + 0.00211485312320292, + 0.001332489075139165, + 0.0012549615930765867, + 0.0012789188185706735, + 0.0012595888692885637, + 0.0010393086122348905, + 0.001569550484418869, + 0.0009340629912912846, + 0.0011181759182363749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.007639123592525721 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898503541946411 + ], + "opposite_logprob_tokens": [ + -2.5789849758148193 + ], + "logprob_reward": -0.07898503541946411, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898503541946411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11017510294914246, + 0.14633196592330933, + 0.24431239068508148, + 0.07772721350193024, + 0.06800395995378494, + 0.039463527500629425, + 0.09051217138767242, + 0.06496105343103409, + 0.12151233851909637, + 0.037000227719545364 + ], + "credit_map": [ + 0.02373405173420906, + 0.007399373687803745, + 0.014189068228006363, + 0.007642567157745361, + 0.013947722502052784, + 0.023646589368581772, + 0.0064046066254377365, + 0.01321112085133791, + 0.022197866812348366, + 0.019789472222328186, + 0.016388464719057083, + 0.015742510557174683, + 0.013714063912630081, + 0.017494114115834236, + 0.01609954982995987, + 0.024905923753976822, + 0.040763065218925476, + 0.005615622270852327, + 0.006651397794485092, + 0.01471019722521305, + 0.00856306217610836, + 0.01444865483790636, + 0.11105970293283463, + 0.04250069335103035, + 0.008776319213211536, + 0.0045397477224469185, + 0.01421299297362566, + 0.01019349880516529, + 0.011193091981112957, + 0.006708088796585798, + 0.012637450359761715, + 0.009466024115681648, + 0.007956796325743198, + 0.008280500769615173, + 0.0065834494307637215, + 0.008112150244414806, + 0.012346132658421993, + 0.010094456374645233, + 0.0065510887652635574, + 0.00807938352227211, + 0.004919310100376606, + 0.006528505589812994, + 0.0035041377414017916, + 0.003538416465744376, + 0.006368589587509632, + 0.004730599001049995, + 0.004684815648943186, + 0.005189154297113419, + 0.01666182465851307, + 0.007330772932618856, + 0.005074101034551859, + 0.006098743993788958, + 0.012655648402869701, + 0.01627175323665142, + 0.009017730131745338, + 0.01740158535540104, + 0.010918494313955307, + 0.006933816708624363, + 0.005141549278050661, + 0.009489068761467934, + 0.005647098179906607, + 0.011553149670362473, + 0.006237298250198364, + 0.009040581062436104, + 0.0176459439098835, + 0.01226380281150341, + 0.007859420031309128, + 0.012185033410787582, + 0.015165604650974274, + 0.030051598325371742, + 0.012415051460266113, + 0.013925892300903797, + 0.0028746933676302433, + 0.004091642331331968, + 0.0041582961566746235, + 0.003946852870285511, + 0.0037846744526177645, + 0.0057944441214203835, + 0.0036088149063289165, + 0.008740810677409172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 8 + ], + "selected_reward": -0.07898503541946411 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559073567390442 + ], + "opposite_logprob_tokens": [ + -1.600590705871582 + ], + "logprob_reward": -0.22559073567390442, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559073567390442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05277131870388985, + 0.08148477971553802, + 0.31158724427223206, + 0.03905156999826431, + 0.05889394134283066, + 0.11211651563644409, + 0.13857893645763397, + 0.13023661077022552, + 0.060120634734630585, + 0.015158465132117271 + ], + "credit_map": [ + 0.006833210121840239, + 0.006719599012285471, + 0.006509300321340561, + 0.004813557956367731, + 0.006779659539461136, + 0.009482771158218384, + 0.004843730479478836, + 0.006789491977542639, + 0.01129445992410183, + 0.010090486146509647, + 0.01028222031891346, + 0.008085466921329498, + 0.007059004157781601, + 0.009440504014492035, + 0.011646055616438389, + 0.013586577028036118, + 0.068842813372612, + 0.009254132397472858, + 0.011645900085568428, + 0.01664317026734352, + 0.013487983494997025, + 0.02056574448943138, + 0.12484195083379745, + 0.046305540949106216, + 0.004854225087910891, + 0.0030334717594087124, + 0.004674465395510197, + 0.005319130606949329, + 0.007471560034900904, + 0.0035747624933719635, + 0.006827341392636299, + 0.0032966129947453737, + 0.008014344610273838, + 0.006321347318589687, + 0.005744204856455326, + 0.007290254347026348, + 0.008443252183496952, + 0.008546531200408936, + 0.00530698336660862, + 0.009227026253938675, + 0.01917526312172413, + 0.015317970886826515, + 0.007950810715556145, + 0.007368703372776508, + 0.016128413379192352, + 0.01786196045577526, + 0.008223866112530231, + 0.020089536905288696, + 0.02396058663725853, + 0.01066883746534586, + 0.00861744862049818, + 0.008330922573804855, + 0.016632873564958572, + 0.033444713801145554, + 0.011762688867747784, + 0.02516087330877781, + 0.023294944316148758, + 0.012963292188942432, + 0.008912140503525734, + 0.011170821264386177, + 0.01799366995692253, + 0.034216366708278656, + 0.009982732124626637, + 0.011702636256814003, + 0.009208169765770435, + 0.010506097227334976, + 0.0054319933988153934, + 0.009502965956926346, + 0.008714442141354084, + 0.005796121433377266, + 0.006506488658487797, + 0.004454358015209436, + 0.0017321493942290545, + 0.0019120752112939954, + 0.001847239676862955, + 0.00157175085041672, + 0.0019031907431781292, + 0.0020148104522377253, + 0.0017084261635318398, + 0.002468822291120887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 7 + ], + "selected_reward": -0.22559073567390442 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from a slight upward angle, the camera slowly and steadily pulls back to reveal a sacred golden-red phoenix hovering in the deep night sky, a backdrop dotted with a tranquil sea of stars, punctuated by countless tiny stars. The phoenix is draped in shimmering golden feathers, their edges and underbelly gleaming with a fiery crimson glow. Its slender neck gracefully arches, its eyes like molten gold gaze into the distance, its powerful wings fully extended, maintaining a posture of tranquility and power. As its body unfurls, its magnificent tail feathers—composed of countless feathers that resemble the condensation of pure flame—unfold layer by layer in rhythmic slow motion. Each feather has a sharp edge, its center ablaze with golden light, gradually fading to a transparent orange-red. The entire unfolding process resembles a vast cosmic flower of light and fire blooming in the void. The phoenix itself is the only light source, and its blazing light forms a warm halo around it, giving the entire picture an atmosphere of solemnity, awe and fantasy.", + "seed": 25685, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25685_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10085918679833412, + "vlm_logprob": -0.10085918679833412, + "vlm_yes_no_margin": 2.725, + "vlm_token_credit": -0.10085918679833412, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.19315898418426514 + ], + "opposite_logprob_tokens": [ + -1.8181589841842651 + ], + "logprob_reward": -0.19315898418426514, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.19315898418426514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1342373490333557, + 0.05956384539604187, + 0.03483898192644119, + 0.025018183514475822, + 0.0360758937895298, + 0.043462324887514114, + 0.04183751717209816, + 0.43726685643196106, + 0.05164341256022453, + 0.1360556185245514 + ], + "credit_map": [ + 0.03189227357506752, + 0.013271775096654892, + 0.012542299926280975, + 0.015014194883406162, + 0.016930365934967995, + 0.01922089420258999, + 0.010233377106487751, + 0.015132172033190727, + 0.014335175976157188, + 0.00606298353523016, + 0.00518898107111454, + 0.008787508122622967, + 0.0051641808822751045, + 0.010183785110712051, + 0.00490987254306674, + 0.0049313572235405445, + 0.004894831217825413, + 0.004085528198629618, + 0.004328704439103603, + 0.005715701263397932, + 0.002886145142838359, + 0.004968233406543732, + 0.0033377401996403933, + 0.004622098058462143, + 0.003328985534608364, + 0.0027728828135877848, + 0.0020948457531630993, + 0.004376869183033705, + 0.0038850349374115467, + 0.002806359902024269, + 0.0032632227521389723, + 0.0024899833370000124, + 0.00807048100978136, + 0.0034664953127503395, + 0.0036658316385000944, + 0.00394394202157855, + 0.0030041292775422335, + 0.00558274332433939, + 0.0031995195895433426, + 0.00514275161549449, + 0.007209566421806812, + 0.0060731167905032635, + 0.004298618994653225, + 0.00436741765588522, + 0.004431131761521101, + 0.0070473854430019855, + 0.005884737707674503, + 0.004150354769080877, + 0.005914853885769844, + 0.00458063930273056, + 0.00526717584580183, + 0.0034880724269896746, + 0.004080037586390972, + 0.006006218027323484, + 0.003900151001289487, + 0.008600368164479733, + 0.030709149315953255, + 0.010768133215606213, + 0.14909085631370544, + 0.022326217964291573, + 0.015030134469270706, + 0.1584860384464264, + 0.016206394881010056, + 0.034649934619665146, + 0.009334838949143887, + 0.006257370579987764, + 0.007981255650520325, + 0.004864071495831013, + 0.004753676243126392, + 0.006693820934742689, + 0.0039010324981063604, + 0.007857351563870907, + 0.009446723386645317, + 0.003535271156579256, + 0.04842623695731163, + 0.006321192719042301, + 0.0033512611407786608, + 0.007742505986243486, + 0.00610252283513546, + 0.05112990736961365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.19315898418426514 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16038455069065094 + ], + "opposite_logprob_tokens": [ + -1.9103845357894897 + ], + "logprob_reward": -0.16038455069065094, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16038455069065094, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023159176111221313, + 0.008266517892479897, + 0.008153965696692467, + 0.007603773847222328, + 0.009895194321870804, + 0.01054011657834053, + 0.015669940039515495, + 0.8963073492050171, + 0.0071801249869167805, + 0.01322387345135212 + ], + "credit_map": [ + 0.0058585889637470245, + 0.002411765744909644, + 0.00207880767993629, + 0.002677038311958313, + 0.001987819094210863, + 0.002858829451724887, + 0.00189884752035141, + 0.0033874798100441694, + 0.0012523437617346644, + 0.0009751193574629724, + 0.0005644215852953494, + 0.000909438997041434, + 0.0014233671827241778, + 0.001628983300179243, + 0.0006513542030006647, + 0.0008614897960796952, + 0.0010724446037784219, + 0.0010489211417734623, + 0.000695509952493012, + 0.0009608296677470207, + 0.0009404143784195185, + 0.0013676751405000687, + 0.0009343806887045503, + 0.0011337901232764125, + 0.0012682884698733687, + 0.0008074470097199082, + 0.0006589899421669543, + 0.0012335199862718582, + 0.0008824234246276319, + 0.000963909609708935, + 0.0010896803578361869, + 0.0006995150470174849, + 0.0015080742305144668, + 0.00120640080422163, + 0.0011158213019371033, + 0.0009379324619658291, + 0.0011581346625462174, + 0.001236684387549758, + 0.001026776502840221, + 0.0017053695628419518, + 0.0017770598642528057, + 0.0011095853988081217, + 0.0011145112803205848, + 0.0010605745483189821, + 0.0010308751370757818, + 0.0016858231974765658, + 0.0013375331182032824, + 0.0014241539174690843, + 0.0021457436960190535, + 0.001610853592865169, + 0.0015373973874375224, + 0.001536677242256701, + 0.0017684552585706115, + 0.0022912495769560337, + 0.0014586339239031076, + 0.0033209293615072966, + 0.06524413079023361, + 0.01052149012684822, + 0.3033686578273773, + 0.03132752329111099, + 0.020983384922146797, + 0.37870293855667114, + 0.026632113382220268, + 0.05952707305550575, + 0.001175502547994256, + 0.0005644246703013778, + 0.0012317737564444542, + 0.0011067395098507404, + 0.0004955482436344028, + 0.000981016200967133, + 0.000594224373344332, + 0.0010308959754183888, + 0.0009911261731758714, + 0.0007935495814308524, + 0.003692805767059326, + 0.0011700585018843412, + 0.0006840542773716152, + 0.0007087161648087204, + 0.0010071502765640616, + 0.004176412709057331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.16038455069065094 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006745543330907822 + ], + "opposite_logprob_tokens": [ + -5.006745338439941 + ], + "logprob_reward": -0.006745543330907822, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006745543330907822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24505220353603363, + 0.16112664341926575, + 0.03747212514281273, + 0.04912324249744415, + 0.11576276272535324, + 0.08498650789260864, + 0.026683136820793152, + 0.25221961736679077, + 0.007018555887043476, + 0.02055520750582218 + ], + "credit_map": [ + 0.048529963940382004, + 0.030215086415410042, + 0.014490129426121712, + 0.02339683100581169, + 0.02628868818283081, + 0.06479567289352417, + 0.013445458374917507, + 0.023890376091003418, + 0.05095497891306877, + 0.01787709817290306, + 0.011336005292832851, + 0.012888340279459953, + 0.01821202039718628, + 0.026713727042078972, + 0.01174676138907671, + 0.01139771193265915, + 0.0051985252648591995, + 0.0033381772227585316, + 0.0031604839023202658, + 0.007711909711360931, + 0.003414259059354663, + 0.004328056704252958, + 0.0060740504413843155, + 0.004246666096150875, + 0.009202208369970322, + 0.00501260906457901, + 0.003950503189116716, + 0.00769263319671154, + 0.006505940109491348, + 0.006059662438929081, + 0.006022769957780838, + 0.004676923155784607, + 0.03244292736053467, + 0.011008057743310928, + 0.01005749125033617, + 0.012234739027917385, + 0.01044042780995369, + 0.017995715141296387, + 0.010640477761626244, + 0.010942935012280941, + 0.023677514865994453, + 0.007232763338834047, + 0.0065697599202394485, + 0.007413691375404596, + 0.005760094616562128, + 0.017904840409755707, + 0.00687059573829174, + 0.00955725647509098, + 0.004221661016345024, + 0.003041589632630348, + 0.0019122605444863439, + 0.0022707399912178516, + 0.0027771673630923033, + 0.003758820937946439, + 0.002455615671351552, + 0.006245285272598267, + 0.017115140333771706, + 0.006915113423019648, + 0.08506374061107635, + 0.01167413778603077, + 0.008743668906390667, + 0.09640170633792877, + 0.00944044254720211, + 0.016865693032741547, + 0.0008821286028251052, + 0.0007066645193845034, + 0.001153173390775919, + 0.0009572193375788629, + 0.0007767262868583202, + 0.0009066137718036771, + 0.0005683332565240562, + 0.001067697536200285, + 0.0019199136877432466, + 0.0008069365285336971, + 0.005832068156450987, + 0.0011257905280217528, + 0.0007690935744903982, + 0.0016445898218080401, + 0.001394630060531199, + 0.007062187418341637 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.006745543330907822 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494809150695801 + ], + "opposite_logprob_tokens": [ + -2.929948091506958 + ], + "logprob_reward": -0.05494809150695801, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494809150695801, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15180513262748718, + 0.04028969258069992, + 0.03817691653966904, + 0.07025205343961716, + 0.02475588023662567, + 0.014276809990406036, + 0.025753645226359367, + 0.522278368473053, + 0.033071424812078476, + 0.07934007793664932 + ], + "credit_map": [ + 0.040040019899606705, + 0.018237853422760963, + 0.009449156001210213, + 0.00956027489155531, + 0.017761066555976868, + 0.03718091547489166, + 0.008262582123279572, + 0.011313270777463913, + 0.008026154711842537, + 0.0034964878577739, + 0.0037499412428587675, + 0.006189694162458181, + 0.003668000688776374, + 0.0055694738402962685, + 0.005033770110458136, + 0.0045561701990664005, + 0.007616016082465649, + 0.004112580791115761, + 0.003604672849178314, + 0.0054927365854382515, + 0.0033739067148417234, + 0.005230403970927, + 0.00430862745270133, + 0.00443797092884779, + 0.007977979257702827, + 0.0071117570623755455, + 0.00915270484983921, + 0.012392884120345116, + 0.008998896926641464, + 0.008694496937096119, + 0.009231853298842907, + 0.006691477261483669, + 0.0029832010623067617, + 0.002610276686027646, + 0.003835272742435336, + 0.0023450960870832205, + 0.002289694268256426, + 0.0030146557837724686, + 0.0027051870711147785, + 0.004972496069967747, + 0.002076538745313883, + 0.0015252305893227458, + 0.002046379027888179, + 0.0011683139018714428, + 0.0014040247770026326, + 0.00244142091833055, + 0.001375723397359252, + 0.0022391777019947767, + 0.0041473121382296085, + 0.0026090743485838175, + 0.0019008747767657042, + 0.0024789629969745874, + 0.0022793221287429333, + 0.00420116912573576, + 0.0018004333833232522, + 0.006336494814604521, + 0.03839615732431412, + 0.01061246357858181, + 0.16439832746982574, + 0.018601786345243454, + 0.013256973586976528, + 0.21974581480026245, + 0.016279151663184166, + 0.040987737476825714, + 0.005165232811123133, + 0.005087778903543949, + 0.004070736467838287, + 0.002966625150293112, + 0.00363734969869256, + 0.004179914947599173, + 0.003980694804340601, + 0.003983092959970236, + 0.006526473443955183, + 0.0038029218558222055, + 0.024306725710630417, + 0.0065435124561190605, + 0.0031576501205563545, + 0.004681993275880814, + 0.004815611056983471, + 0.02550519071519375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.05494809150695801 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890587642788887 + ], + "opposite_logprob_tokens": [ + -2.4640588760375977 + ], + "logprob_reward": -0.0890587642788887, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890587642788887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2314128279685974, + 0.07844321429729462, + 0.0756542757153511, + 0.07165950536727905, + 0.08689746260643005, + 0.06795769184827805, + 0.12751919031143188, + 0.054421499371528625, + 0.03299005329608917, + 0.17304429411888123 + ], + "credit_map": [ + 0.053672805428504944, + 0.026969389989972115, + 0.01485690101981163, + 0.020222937688231468, + 0.033317163586616516, + 0.05033411830663681, + 0.013656475581228733, + 0.018383042886853218, + 0.02105090022087097, + 0.007872244343161583, + 0.0058257924392819405, + 0.010579310357570648, + 0.00629767682403326, + 0.011848102323710918, + 0.0067661297507584095, + 0.0082030538469553, + 0.007847013883292675, + 0.0075423698872327805, + 0.0071371691301465034, + 0.014419945888221264, + 0.008525770157575607, + 0.008950593881309032, + 0.01224875170737505, + 0.008982663974165916, + 0.008354811929166317, + 0.006460292264819145, + 0.00886460766196251, + 0.00908698420971632, + 0.011025062762200832, + 0.009812753647565842, + 0.010027842596173286, + 0.008027147501707077, + 0.013579166494309902, + 0.009736365638673306, + 0.00964033417403698, + 0.009879023768007755, + 0.008536064997315407, + 0.013182459399104118, + 0.0077963522635400295, + 0.014547700993716717, + 0.018391136080026627, + 0.006887948606163263, + 0.0044747269712388515, + 0.00412770127877593, + 0.006255964282900095, + 0.013003868982195854, + 0.005573175847530365, + 0.009243171662092209, + 0.014285468496382236, + 0.017293306067585945, + 0.01101746503263712, + 0.01510928850620985, + 0.01625210978090763, + 0.016218002885580063, + 0.013324510306119919, + 0.024019038304686546, + 0.005372957326471806, + 0.005715787876397371, + 0.01134833786636591, + 0.005192664917558432, + 0.004758404102176428, + 0.013491191901266575, + 0.004081599414348602, + 0.004460552241653204, + 0.0034185273107141256, + 0.0025162003003060818, + 0.005742393434047699, + 0.005018217954784632, + 0.002557929838076234, + 0.0042497809045016766, + 0.00312518491409719, + 0.006361817475408316, + 0.011079263873398304, + 0.004236223176121712, + 0.06170692667365074, + 0.007858614437282085, + 0.005167079623788595, + 0.010495070368051529, + 0.008773507550358772, + 0.06372761726379395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.0890587642788887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins from a slight upward angle, the camera slowly and steadily pulls back to reveal a sacred golden-red phoenix hovering in the deep night sky, a backdrop dotted with a tranquil sea of stars, punctuated by countless tiny stars. The phoenix is draped in shimmering golden feathers, their edges and underbelly gleaming with a fiery crimson glow. Its slender neck gracefully arches, its eyes like molten gold gaze into the distance, its powerful wings fully extended, maintaining a posture of tranquility and power. As its body unfurls, its magnificent tail feathers—composed of countless feathers that resemble the condensation of pure flame—unfold layer by layer in rhythmic slow motion. Each feather has a sharp edge, its center ablaze with golden light, gradually fading to a transparent orange-red. The entire unfolding process resembles a vast cosmic flower of light and fire blooming in the void. The phoenix itself is the only light source, and its blazing light forms a warm halo around it, giving the entire picture an atmosphere of solemnity, awe and fantasy.", + "seed": 25686, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_2/index_1514_seed_25686_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11898372024297714, + "vlm_logprob": -0.11898372024297714, + "vlm_yes_no_margin": 2.4, + "vlm_token_credit": -0.11898372024297714, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2610129415988922 + ], + "opposite_logprob_tokens": [ + -1.5110129117965698 + ], + "logprob_reward": -0.2610129415988922, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2610129415988922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36702385544776917, + 0.11518103629350662, + 0.04980865493416786, + 0.07381444424390793, + 0.06892475485801697, + 0.05950777232646942, + 0.07795844227075577, + 0.03255576267838478, + 0.06202580779790878, + 0.0931994840502739 + ], + "credit_map": [ + 0.04465284198522568, + 0.06712434440851212, + 0.05165698379278183, + 0.05008730664849281, + 0.026690464466810226, + 0.06130973622202873, + 0.027968507260084152, + 0.03753368556499481, + 0.008927577175199986, + 0.007654600311070681, + 0.021104617044329643, + 0.02087562531232834, + 0.0066556064411997795, + 0.00940390769392252, + 0.019012432545423508, + 0.021546663716435432, + 0.004538300912827253, + 0.0042996397241950035, + 0.012157113291323185, + 0.0058813439682126045, + 0.0035675414837896824, + 0.006571151781827211, + 0.005575050134211779, + 0.00721851596608758, + 0.007585304789245129, + 0.007943454198539257, + 0.014377201907336712, + 0.00840793363749981, + 0.006978142540901899, + 0.007195573300123215, + 0.009398441761732101, + 0.011928393505513668, + 0.010236717760562897, + 0.0063819256611168385, + 0.01006384938955307, + 0.01018919050693512, + 0.00892032589763403, + 0.005088604521006346, + 0.009798838756978512, + 0.008245306089520454, + 0.011794590391218662, + 0.005660529714077711, + 0.00978470966219902, + 0.0055776238441467285, + 0.004928795155137777, + 0.00687154708430171, + 0.005415773019194603, + 0.009474202990531921, + 0.010829801671206951, + 0.01031579915434122, + 0.012196166440844536, + 0.008874619379639626, + 0.007307346910238266, + 0.01002151146531105, + 0.00934672262519598, + 0.009066475555300713, + 0.006410875357687473, + 0.00390978017821908, + 0.004324283450841904, + 0.0024353244807571173, + 0.003724205307662487, + 0.0038807038217782974, + 0.0026362843345850706, + 0.00523430434986949, + 0.011359261348843575, + 0.005476095713675022, + 0.00880967453122139, + 0.00722733186557889, + 0.0048812380991876125, + 0.006091565825045109, + 0.006061683874577284, + 0.012118957005441189, + 0.014464621432125568, + 0.013523103669285774, + 0.016391785815358162, + 0.005329377017915249, + 0.00674293702468276, + 0.009588446468114853, + 0.006061889231204987, + 0.021097328513860703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.2610129415988922 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991405725479126 + ], + "opposite_logprob_tokens": [ + -1.8049139976501465 + ], + "logprob_reward": -0.17991405725479126, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991405725479126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13413117825984955, + 0.22076494991779327, + 0.04043613374233246, + 0.061327237635850906, + 0.06482242047786713, + 0.09386668354272842, + 0.16666346788406372, + 0.06337831914424896, + 0.09275823831558228, + 0.06185135990381241 + ], + "credit_map": [ + 0.012712676078081131, + 0.014170067384839058, + 0.0196104496717453, + 0.029827330261468887, + 0.0069158668629825115, + 0.013199934735894203, + 0.01574883610010147, + 0.021946022287011147, + 0.017422938719391823, + 0.016021504998207092, + 0.041411492973566055, + 0.03330392390489578, + 0.012716093100607395, + 0.017797421663999557, + 0.03275170177221298, + 0.04933988302946091, + 0.004161265213042498, + 0.004718543495982885, + 0.006364169996231794, + 0.0049508134834468365, + 0.004596162587404251, + 0.0038522735703736544, + 0.006769703235477209, + 0.005023202393203974, + 0.007023538928478956, + 0.0072894832119345665, + 0.008588134311139584, + 0.007190454285591841, + 0.006108857225626707, + 0.004589239601045847, + 0.00789229292422533, + 0.012645237147808075, + 0.009803046472370625, + 0.006585422437638044, + 0.010128859430551529, + 0.00534581346437335, + 0.00803561881184578, + 0.00767156295478344, + 0.007215764373540878, + 0.010036329738795757, + 0.025309747084975243, + 0.009011968038976192, + 0.02049006149172783, + 0.0078690554946661, + 0.005930005107074976, + 0.009782737120985985, + 0.007000457961112261, + 0.008472651243209839, + 0.022512760013341904, + 0.02166767604649067, + 0.028664758428931236, + 0.020374132320284843, + 0.018071390688419342, + 0.017344720661640167, + 0.020959164947271347, + 0.017068859189748764, + 0.01283142901957035, + 0.006471899803727865, + 0.00852489098906517, + 0.006662397179752588, + 0.0071309590712189674, + 0.005546429194509983, + 0.006090216338634491, + 0.010120092891156673, + 0.020394550636410713, + 0.007382159121334553, + 0.014084999449551105, + 0.008435788564383984, + 0.005248541943728924, + 0.008923997171223164, + 0.0079019321128726, + 0.02038627117872238, + 0.009685665369033813, + 0.009823204018175602, + 0.010812118649482727, + 0.004278311971575022, + 0.0050449431873857975, + 0.006397934630513191, + 0.004809417296200991, + 0.010999767109751701 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.17991405725479126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026351071894168854 + ], + "opposite_logprob_tokens": [ + -3.651350975036621 + ], + "logprob_reward": -0.026351071894168854, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026351071894168854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2069157063961029, + 0.044572632759809494, + 0.05891614407300949, + 0.03700551763176918, + 0.03319800645112991, + 0.05454401671886444, + 0.07461302727460861, + 0.04562702029943466, + 0.144426628947258, + 0.3001812994480133 + ], + "credit_map": [ + 0.022972160950303078, + 0.04508613422513008, + 0.026886843144893646, + 0.01685395836830139, + 0.013885901309549809, + 0.04724765196442604, + 0.015786536037921906, + 0.01819651946425438, + 0.0038562784902751446, + 0.0035266121849417686, + 0.0060553220100700855, + 0.008732166141271591, + 0.0027076546102762222, + 0.004114019684493542, + 0.006882233079522848, + 0.008698347955942154, + 0.004655386786907911, + 0.005006192717701197, + 0.01593163050711155, + 0.00618856493383646, + 0.003252205206081271, + 0.007800104096531868, + 0.004654129035770893, + 0.011427929624915123, + 0.004349140450358391, + 0.004741523414850235, + 0.005641933064907789, + 0.0036310101859271526, + 0.003628211561590433, + 0.005020815413445234, + 0.004384543281048536, + 0.005608338862657547, + 0.00516519183292985, + 0.004141667392104864, + 0.004764745477586985, + 0.004528652876615524, + 0.0028785092290490866, + 0.0029344370122998953, + 0.0038764600176364183, + 0.004908340983092785, + 0.013497692532837391, + 0.004278628155589104, + 0.012405379675328732, + 0.004313845187425613, + 0.0027289048302918673, + 0.0063950056210160255, + 0.003766231006011367, + 0.00715832831338048, + 0.014407523907721043, + 0.010544080287218094, + 0.008399678394198418, + 0.006681438535451889, + 0.006617904640734196, + 0.007927329279482365, + 0.007244298700243235, + 0.012790773995220661, + 0.010551567189395428, + 0.004669832065701485, + 0.006972620729357004, + 0.003632476320490241, + 0.004382594954222441, + 0.0049072327092289925, + 0.0034326352179050446, + 0.007078058551996946, + 0.03582865372300148, + 0.010339276865124702, + 0.024487212300300598, + 0.008002732880413532, + 0.006549264770001173, + 0.012407888658344746, + 0.010714472271502018, + 0.03609712794423103, + 0.058712106198072433, + 0.03951545059680939, + 0.048587534576654434, + 0.01542908139526844, + 0.024083660915493965, + 0.027686290442943573, + 0.017284788191318512, + 0.06888236850500107 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.026351071894168854 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048653073608875275 + ], + "opposite_logprob_tokens": [ + -3.0486531257629395 + ], + "logprob_reward": -0.048653073608875275, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048653073608875275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25738322734832764, + 0.1074436604976654, + 0.07103623449802399, + 0.06519120186567307, + 0.06641574949026108, + 0.08598804473876953, + 0.1193460002541542, + 0.057607974857091904, + 0.09576313197612762, + 0.07382477074861526 + ], + "credit_map": [ + 0.020193330943584442, + 0.03380127251148224, + 0.04279755428433418, + 0.03532285988330841, + 0.012752250768244267, + 0.037215568125247955, + 0.03436661884188652, + 0.040933772921562195, + 0.012321169488132, + 0.009444542229175568, + 0.018075253814458847, + 0.01341775618493557, + 0.00980681274086237, + 0.009686551988124847, + 0.014103945344686508, + 0.020587630569934845, + 0.008674291893839836, + 0.005845131352543831, + 0.017187202349305153, + 0.007147624157369137, + 0.005564406979829073, + 0.00866263173520565, + 0.006410838104784489, + 0.011544102802872658, + 0.005371949635446072, + 0.0060277339071035385, + 0.01734165847301483, + 0.00655656773597002, + 0.00484823202714324, + 0.007568863220512867, + 0.006234422326087952, + 0.011241779662668705, + 0.00817992351949215, + 0.00859057903289795, + 0.009932010434567928, + 0.007651317398995161, + 0.005796162411570549, + 0.008685690350830555, + 0.006538711488246918, + 0.011041351594030857, + 0.01518629863858223, + 0.009431322105228901, + 0.015427528880536556, + 0.006164010148495436, + 0.0067731416784226894, + 0.011782368645071983, + 0.0064299451187253, + 0.014793426729738712, + 0.01707475818693638, + 0.012504528276622295, + 0.023035669699311256, + 0.015916725620627403, + 0.012779546901583672, + 0.010321881622076035, + 0.013616238720715046, + 0.014096655882894993, + 0.010285766795277596, + 0.005197875201702118, + 0.008861029520630836, + 0.007282985839992762, + 0.004755996633321047, + 0.00594761036336422, + 0.006584769114851952, + 0.008691943250596523, + 0.021781956776976585, + 0.007145834155380726, + 0.015305940993130207, + 0.0074139791540801525, + 0.005522516090422869, + 0.008807370439171791, + 0.006949335802346468, + 0.022836199030280113, + 0.017189515754580498, + 0.006717383395880461, + 0.010829218663275242, + 0.003664284013211727, + 0.005288832355290651, + 0.006440144032239914, + 0.0043706209398806095, + 0.01932477205991745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.048653073608875275 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot beginning from a slight upward angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly and steadily pull back to reveal the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sacred golden-red phoenix visible hovering in a deep night sky with a starry backdrop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the phoenix's tail feathers unfold layer by layer in rhythmic slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the phoenix the only light source in the scene, creating a warm halo around it?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898745685815811 + ], + "opposite_logprob_tokens": [ + -2.5789873600006104 + ], + "logprob_reward": -0.07898745685815811, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898745685815811, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21722941100597382, + 0.13525786995887756, + 0.06675603985786438, + 0.07919575273990631, + 0.03179062530398369, + 0.09978798031806946, + 0.11407189071178436, + 0.07506196200847626, + 0.06795573979616165, + 0.11289273202419281 + ], + "credit_map": [ + 0.02952018938958645, + 0.03333933278918266, + 0.029826536774635315, + 0.02510208822786808, + 0.013059419579803944, + 0.026785491034388542, + 0.024210412055253983, + 0.03538595139980316, + 0.011860291473567486, + 0.011198682710528374, + 0.022883007302880287, + 0.018759865313768387, + 0.007038681767880917, + 0.011953864246606827, + 0.019787458702921867, + 0.03177601099014282, + 0.006708451546728611, + 0.006506309378892183, + 0.011060447432100773, + 0.0072361184284091, + 0.0050248499028384686, + 0.006212495267391205, + 0.007327877450734377, + 0.016679489985108376, + 0.007455391809344292, + 0.00514810299500823, + 0.02051446959376335, + 0.008736486546695232, + 0.0056643118150532246, + 0.011271568015217781, + 0.008246091194450855, + 0.012159334495663643, + 0.004623714834451675, + 0.0034375153481960297, + 0.004736150614917278, + 0.003273849142715335, + 0.0035612229257822037, + 0.003933405503630638, + 0.003513303818181157, + 0.0047114635817706585, + 0.03448048233985901, + 0.008400935679674149, + 0.030017584562301636, + 0.006034692749381065, + 0.003931560553610325, + 0.006671446841210127, + 0.004686777945607901, + 0.005564498249441385, + 0.021023893728852272, + 0.014314896427094936, + 0.016962965950369835, + 0.011842839419841766, + 0.009424705989658833, + 0.010227191261947155, + 0.01145960297435522, + 0.018815794959664345, + 0.015089735388755798, + 0.007176576182246208, + 0.011960766278207302, + 0.007288867142051458, + 0.00663924403488636, + 0.007865329273045063, + 0.005817309021949768, + 0.01322413980960846, + 0.013811647891998291, + 0.006558647844940424, + 0.008719969540834427, + 0.006655664183199406, + 0.006086157169193029, + 0.007008479908108711, + 0.006997887510806322, + 0.012117283418774605, + 0.018250394612550735, + 0.014576800167560577, + 0.01808888092637062, + 0.006146125961095095, + 0.007626786362379789, + 0.010561246424913406, + 0.008439813740551472, + 0.029202675446867943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.07898745685815811 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_tensor([4941, 4941])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_tensor([4941, 4941])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e1c19ca41278b1dc6343fe3e2efdf05bc4be450a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_tensor([4941, 4941])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "sweeping, cinematic shot begins with a low-angle tracking shot, moving parallel to a young Southeast Asian man as he rides his bicycle at a steady, relaxed pace. He has short, windswept black hair, slightly tousled, and a subtle, contented smile on his lips. He wears a lightweight, dark teal windbreaker, zipped up over a simple white T-shirt, slim-fitting black sweatpants, and dark sneakers. He rides a sleek, matte black fixed-gear bicycle, its minimalist frame cutting cleanly through the city air. The camera then smoothly pulls back and ascends, transitioning to a high-angle wide shot, revealing the breathtaking scale of this vibrant city at dusk. He is cycling on a dedicated bike lane on a bridge or elevated roadway, surrounded by a trail of red and white lights, reflecting the busy traffic below. The sky is a deep indigo canvas, punctuated by the last fiery orange glow of sunset. Towering skyscrapers, their windows forming a mosaic of warm and cool lights, create a dazzling, futuristic skyline. Recent rainfall left the asphalt shimmering, reflecting the city's neon signs and streetlights in a shimmering, impressionistic blur of colors. The evening's predominantly cool tones were interrupted by a vibrant palette of pinks, cyans, and warm yellows in the digital billboards and architectural lighting, creating a scene both vibrant and beautiful. The overall atmosphere was one of exhilarating freedom and tranquil contemplation, capturing a single, perfect moment of journey and connection to this vast, vibrant metropolis.", + "seed": 28064, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28064_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.052959496481344105, + "vlm_logprob": -0.052959496481344105, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.052959496481344105, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14838677644729614 + ], + "opposite_logprob_tokens": [ + -2.0233867168426514 + ], + "logprob_reward": -0.14838677644729614, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14838677644729614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17419078946113586, + 0.05307868868112564, + 0.03270633518695831, + 0.052981190383434296, + 0.14505815505981445, + 0.1337997168302536, + 0.2840452790260315, + 0.048168301582336426, + 0.027888214215636253, + 0.04808337241411209 + ], + "credit_map": [ + 0.02903894893825054, + 0.043092239648103714, + 0.018895629793405533, + 0.014953630045056343, + 0.015903817489743233, + 0.021045340225100517, + 0.015908222645521164, + 0.015352953225374222, + 0.006089349277317524, + 0.0059642493724823, + 0.006604632828384638, + 0.006857964210212231, + 0.006370083428919315, + 0.006646655034273863, + 0.003881573909893632, + 0.010664183646440506, + 0.00516901770606637, + 0.005523777101188898, + 0.004397791810333729, + 0.0038782821502536535, + 0.0030310123693197966, + 0.005219864659011364, + 0.0028415312990546227, + 0.002645057626068592, + 0.009159480221569538, + 0.007039827760308981, + 0.01074914075434208, + 0.007259033620357513, + 0.0038709675427526236, + 0.0038362357299774885, + 0.0054100509732961655, + 0.005656452849507332, + 0.0075750271789729595, + 0.008173555135726929, + 0.027737930417060852, + 0.015792593359947205, + 0.008761950768530369, + 0.012483961880207062, + 0.015929225832223892, + 0.04860391095280647, + 0.011794591322541237, + 0.011604409664869308, + 0.04516811668872833, + 0.008229546248912811, + 0.00896640494465828, + 0.02043929137289524, + 0.0126718794927001, + 0.014925479888916016, + 0.10217782855033875, + 0.01553014200180769, + 0.08369727432727814, + 0.022174933925271034, + 0.009452978149056435, + 0.0089231813326478, + 0.017204690724611282, + 0.024884240701794624, + 0.005414757877588272, + 0.009232167154550552, + 0.004594313446432352, + 0.006077260244637728, + 0.004934592172503471, + 0.005607308819890022, + 0.00695496890693903, + 0.005352931562811136, + 0.0035609197802841663, + 0.0029541885014623404, + 0.0037234658375382423, + 0.005243341438472271, + 0.002248163102194667, + 0.004007081035524607, + 0.0029935496859252453, + 0.0031575073953717947, + 0.0041076792404055595, + 0.005109691061079502, + 0.010060234926640987, + 0.007544040679931641, + 0.004735168069601059, + 0.0047483304515480995, + 0.006815960165113211, + 0.0049622682854533195 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 4 + ], + "selected_reward": -0.14838677644729614 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016079628840088844 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.016079628840088844, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016079628840088844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07854639738798141, + 0.027923230081796646, + 0.04752713441848755, + 0.056553006172180176, + 0.11966200172901154, + 0.08018186688423157, + 0.41873061656951904, + 0.05745141953229904, + 0.030692577362060547, + 0.08273177593946457 + ], + "credit_map": [ + 0.008477378636598587, + 0.00997763592749834, + 0.011660557240247726, + 0.010574309155344963, + 0.011173324659466743, + 0.007404300384223461, + 0.010736493393778801, + 0.00854239147156477, + 0.002205297350883484, + 0.002560890279710293, + 0.0034100674092769623, + 0.004076942801475525, + 0.002558730775490403, + 0.0024277768097817898, + 0.0030473635997623205, + 0.0076361605897545815, + 0.00397254154086113, + 0.00624274043366313, + 0.0074336715042591095, + 0.005476079415529966, + 0.006098269950598478, + 0.008816621266305447, + 0.004709811881184578, + 0.004777394700795412, + 0.007684529758989811, + 0.006881524343043566, + 0.012401360087096691, + 0.007750917226076126, + 0.007261165417730808, + 0.003871844382956624, + 0.006332247518002987, + 0.004369410220533609, + 0.007309645414352417, + 0.01092500053346157, + 0.020849013701081276, + 0.012619172222912312, + 0.01040663756430149, + 0.009413918480277061, + 0.013536747545003891, + 0.034601856023073196, + 0.0075576514936983585, + 0.010124673135578632, + 0.010465547442436218, + 0.013140653260052204, + 0.01040463987737894, + 0.00953938439488411, + 0.008849186822772026, + 0.010100130923092365, + 0.13660719990730286, + 0.022370733320713043, + 0.14470481872558594, + 0.03149988129734993, + 0.013501754961907864, + 0.013152151368558407, + 0.021188784390687943, + 0.03570523113012314, + 0.007981376722455025, + 0.0054509458132088184, + 0.006951414048671722, + 0.00748196616768837, + 0.006070329342037439, + 0.00871313363313675, + 0.0065290783531963825, + 0.008273174054920673, + 0.004379346966743469, + 0.0035458593629300594, + 0.004753640852868557, + 0.0047961995005607605, + 0.0026816262397915125, + 0.003481500782072544, + 0.0022263815626502037, + 0.004828018136322498, + 0.010380183346569538, + 0.006472391542047262, + 0.019631417468190193, + 0.011235389858484268, + 0.008634364232420921, + 0.009370727464556694, + 0.008931268937885761, + 0.00807602796703577 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.016079628840088844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186602741479874 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186602741479874, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186602741479874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10538945347070694, + 0.06047888472676277, + 0.056988976895809174, + 0.025956889614462852, + 0.04738645628094673, + 0.4903982877731323, + 0.05853773280978203, + 0.0450144000351429, + 0.040180519223213196, + 0.06966838240623474 + ], + "credit_map": [ + 0.015003630891442299, + 0.02356337569653988, + 0.01058753952383995, + 0.010884230956435204, + 0.011988039128482342, + 0.010961905121803284, + 0.013643264770507812, + 0.008757475763559341, + 0.007414167746901512, + 0.007455587387084961, + 0.00820037629455328, + 0.00697932718321681, + 0.006383408792316914, + 0.008840815164148808, + 0.00542050413787365, + 0.009784702211618423, + 0.006567684002220631, + 0.008478708565235138, + 0.007231833413243294, + 0.006466277409344912, + 0.007774793077260256, + 0.008636768907308578, + 0.005125148221850395, + 0.006707769352942705, + 0.004266221541911364, + 0.003778428304940462, + 0.004195421002805233, + 0.003268050728365779, + 0.0034542062785476446, + 0.0018137304577976465, + 0.0028231891337782145, + 0.002357643097639084, + 0.005192250479012728, + 0.007553935516625643, + 0.005891631823033094, + 0.004151161760091782, + 0.006914753001183271, + 0.004118320997804403, + 0.005772117991000414, + 0.00779228750616312, + 0.043377429246902466, + 0.035126954317092896, + 0.1835300773382187, + 0.02911072038114071, + 0.025856319814920425, + 0.06782490760087967, + 0.05543079599738121, + 0.050141092389822006, + 0.0133630046620965, + 0.008681042119860649, + 0.007265739142894745, + 0.0061956909485161304, + 0.004894469864666462, + 0.003708863165229559, + 0.005572689231485128, + 0.008856236934661865, + 0.005188301671296358, + 0.006353694945573807, + 0.005546113010495901, + 0.0072767590172588825, + 0.002870923141017556, + 0.004546016920357943, + 0.006551309954375029, + 0.0066812834702432156, + 0.003576531307771802, + 0.00507946265861392, + 0.005097576417028904, + 0.008828239515423775, + 0.003548265667632222, + 0.003900925163179636, + 0.004026102367788553, + 0.006123418919742107, + 0.008725768886506557, + 0.004864878486841917, + 0.020660312846302986, + 0.009334731847047806, + 0.006584026850759983, + 0.005341143347322941, + 0.008313581347465515, + 0.005843945778906345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.018186602741479874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07895760983228683 + ], + "opposite_logprob_tokens": [ + -2.5789575576782227 + ], + "logprob_reward": -0.07895760983228683, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07895760983228683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03791625052690506, + 0.02761254459619522, + 0.018715139478445053, + 0.03164146840572357, + 0.017959244549274445, + 0.4843295216560364, + 0.13953860104084015, + 0.03471793606877327, + 0.028629813343286514, + 0.17893949151039124 + ], + "credit_map": [ + 0.006089302245527506, + 0.008799370378255844, + 0.003967795055359602, + 0.002740375464782119, + 0.003907080739736557, + 0.0037606079131364822, + 0.003934219945222139, + 0.004717493895441294, + 0.0030485468450933695, + 0.002821119735017419, + 0.005636543035507202, + 0.003876717062667012, + 0.002745965262874961, + 0.002647839020937681, + 0.0016534168971702456, + 0.005182396154850721, + 0.0018957221182063222, + 0.002708166604861617, + 0.0022893217392265797, + 0.002861220855265856, + 0.0018612106796354055, + 0.0024217863101512194, + 0.0018657903419807553, + 0.0028119191993027925, + 0.003791022812947631, + 0.003968001808971167, + 0.00701950816437602, + 0.004575871396809816, + 0.003584787482395768, + 0.0027027775067836046, + 0.0030383721459656954, + 0.002961125923320651, + 0.002112331800162792, + 0.0017429533181712031, + 0.0049193864688277245, + 0.00192293431609869, + 0.0022308826446533203, + 0.0013970856089144945, + 0.001634522806853056, + 0.0019991458393633366, + 0.04431957006454468, + 0.03425416350364685, + 0.18737126886844635, + 0.025172607973217964, + 0.02221774496138096, + 0.06807851791381836, + 0.048145513981580734, + 0.05477011203765869, + 0.03286757320165634, + 0.005371844861656427, + 0.04958018288016319, + 0.009144737385213375, + 0.004482767544686794, + 0.005810135509818792, + 0.006976509466767311, + 0.02530483901500702, + 0.0046111238189041615, + 0.004555351100862026, + 0.004937831312417984, + 0.00559031218290329, + 0.0030178315937519073, + 0.0037493249401450157, + 0.003605755977332592, + 0.004650404676795006, + 0.0030945888720452785, + 0.004013714846223593, + 0.0048706987872719765, + 0.006284319795668125, + 0.0022167323622852564, + 0.002151348628103733, + 0.002427133033052087, + 0.003571274457499385, + 0.020654039457440376, + 0.011208528652787209, + 0.05671780928969383, + 0.02479386515915394, + 0.013520199805498123, + 0.01157368440181017, + 0.0253574438393116, + 0.015113907866179943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 6 + ], + "selected_reward": -0.07895760983228683 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003186864545568824 + ], + "opposite_logprob_tokens": [ + -5.7531867027282715 + ], + "logprob_reward": -0.003186864545568824, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003186864545568824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08306681364774704, + 0.06773964315652847, + 0.027247775346040726, + 0.0627644807100296, + 0.03677789866924286, + 0.5278266668319702, + 0.04278785362839699, + 0.04341994225978851, + 0.031242137774825096, + 0.07712683826684952 + ], + "credit_map": [ + 0.0122224697843194, + 0.0158652663230896, + 0.013346794992685318, + 0.009268776513636112, + 0.004961545579135418, + 0.010239838622510433, + 0.007495745085179806, + 0.009666364639997482, + 0.00698568020015955, + 0.0068903290666639805, + 0.017930705100297928, + 0.009349260479211807, + 0.005528842099010944, + 0.006499629467725754, + 0.004104245454072952, + 0.010450944304466248, + 0.0049286410212516785, + 0.005487171467393637, + 0.0032779511529952288, + 0.0037558774929493666, + 0.002170209540054202, + 0.0032319563906639814, + 0.0017530808690935373, + 0.002642882987856865, + 0.01093866303563118, + 0.007767346687614918, + 0.015868879854679108, + 0.007245399057865143, + 0.005701996386051178, + 0.005126336123794317, + 0.006662831176072359, + 0.0034530269913375378, + 0.004803009796887636, + 0.0036372332833707333, + 0.00718676159158349, + 0.0043009244836866856, + 0.004735295660793781, + 0.002590246731415391, + 0.0031338511034846306, + 0.006390572991222143, + 0.04874899610877037, + 0.03799339011311531, + 0.20186491310596466, + 0.02759833261370659, + 0.02469627745449543, + 0.07529179006814957, + 0.05142788589000702, + 0.060205020010471344, + 0.010384813882410526, + 0.0027509229257702827, + 0.011230106465518475, + 0.0035633211955428123, + 0.00401234719902277, + 0.0025279573164880276, + 0.00410178117454052, + 0.004216598346829414, + 0.006947451736778021, + 0.003495562355965376, + 0.005415262188762426, + 0.004669557325541973, + 0.003968523815274239, + 0.005222159903496504, + 0.005442269612103701, + 0.008259149268269539, + 0.004225396551191807, + 0.0034924016799777746, + 0.0031605467665940523, + 0.005284361075609922, + 0.003097918350249529, + 0.005069824866950512, + 0.0034219282679259777, + 0.0034897581208497286, + 0.011846009641885757, + 0.004957367666065693, + 0.019104555249214172, + 0.006315871607512236, + 0.011379595845937729, + 0.0055230590514838696, + 0.008637516759335995, + 0.009362860582768917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.003186864545568824 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, cinematic shot begins with a low-angle tracking shot, moving parallel to a young Southeast Asian man as he rides his bicycle at a steady, relaxed pace. He has short, windswept black hair, slightly tousled, and a subtle, contented smile on his lips. He wears a lightweight, dark teal windbreaker, zipped up over a simple white T-shirt, slim-fitting black sweatpants, and dark sneakers. He rides a sleek, matte black fixed-gear bicycle, its minimalist frame cutting cleanly through the city air. The camera then smoothly pulls back and ascends, transitioning to a high-angle wide shot, revealing the breathtaking scale of this vibrant city at dusk. He is cycling on a dedicated bike lane on a bridge or elevated roadway, surrounded by a trail of red and white lights, reflecting the busy traffic below. The sky is a deep indigo canvas, punctuated by the last fiery orange glow of sunset. Towering skyscrapers, their windows forming a mosaic of warm and cool lights, create a dazzling, futuristic skyline. Recent rainfall left the asphalt shimmering, reflecting the city's neon signs and streetlights in a shimmering, impressionistic blur of colors. The evening's predominantly cool tones were interrupted by a vibrant palette of pinks, cyans, and warm yellows in the digital billboards and architectural lighting, creating a scene both vibrant and beautiful. The overall atmosphere was one of exhilarating freedom and tranquil contemplation, capturing a single, perfect moment of journey and connection to this vast, vibrant metropolis.", + "seed": 28065, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28065_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05614277571439743, + "vlm_logprob": -0.05614277571439743, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.05614277571439743, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20574453473091125 + ], + "opposite_logprob_tokens": [ + -1.7057445049285889 + ], + "logprob_reward": -0.20574453473091125, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20574453473091125, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1571209877729416, + 0.09198027104139328, + 0.06088262423872948, + 0.3435651361942291, + 0.09755197912454605, + 0.04665525257587433, + 0.06063144654035568, + 0.03363039717078209, + 0.01787398010492325, + 0.09010791033506393 + ], + "credit_map": [ + 0.02242748998105526, + 0.01249780785292387, + 0.021433725953102112, + 0.022627759724855423, + 0.01902882568538189, + 0.018148148432374, + 0.019546497613191605, + 0.021410733461380005, + 0.012208850122988224, + 0.008748341351747513, + 0.0173175185918808, + 0.008256255649030209, + 0.012779944576323032, + 0.010228984989225864, + 0.008300907909870148, + 0.014139467850327492, + 0.006845699157565832, + 0.007022128906100988, + 0.008295508101582527, + 0.008196310140192509, + 0.01130813267081976, + 0.007607133127748966, + 0.006774744484573603, + 0.00483296811580658, + 0.07943956553936005, + 0.08287286758422852, + 0.061644040048122406, + 0.02657601609826088, + 0.02459847927093506, + 0.04106314852833748, + 0.013416429050266743, + 0.013954592868685722, + 0.014262422919273376, + 0.025567861273884773, + 0.012324913404881954, + 0.010953221470117569, + 0.00828884169459343, + 0.01238496508449316, + 0.0051563759334385395, + 0.00861337874084711, + 0.008249680511653423, + 0.0042734649032354355, + 0.0029359390027821064, + 0.004370810929685831, + 0.00791264045983553, + 0.007153043989092112, + 0.004996448755264282, + 0.0067632258869707584, + 0.015541107393801212, + 0.0059945955872535706, + 0.0024337563663721085, + 0.00290956674143672, + 0.00678744213655591, + 0.00533277727663517, + 0.003915843553841114, + 0.017716357484459877, + 0.005066922400146723, + 0.004824793431907892, + 0.0037592479493469, + 0.003102953778579831, + 0.004717839416116476, + 0.005054634064435959, + 0.0032603247091174126, + 0.0038436821196228266, + 0.0018402996938675642, + 0.0021490762010216713, + 0.0015134030254557729, + 0.0014053378254175186, + 0.0033105260226875544, + 0.002917511621490121, + 0.0014940298860892653, + 0.0032437951304018497, + 0.010325923562049866, + 0.010423400439321995, + 0.005584691185504198, + 0.005234266631305218, + 0.010479348711669445, + 0.01335176918655634, + 0.019387805834412575, + 0.015320707112550735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.20574453473091125 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214574359357357 + ], + "opposite_logprob_tokens": [ + -4.264214515686035 + ], + "logprob_reward": -0.014214574359357357, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214574359357357, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07340038567781448, + 0.03542031720280647, + 0.03702855110168457, + 0.6785743832588196, + 0.019378170371055603, + 0.01959441788494587, + 0.05675012245774269, + 0.021465009078383446, + 0.012388041242957115, + 0.046000637114048004 + ], + "credit_map": [ + 0.010454811155796051, + 0.00587951997295022, + 0.01167116966098547, + 0.009920510463416576, + 0.007586187217384577, + 0.009060660377144814, + 0.008768469095230103, + 0.0100590530782938, + 0.0038721137680113316, + 0.003154104808345437, + 0.0036413855850696564, + 0.0036251551937311888, + 0.006737235467880964, + 0.005454184953123331, + 0.0039850566536188126, + 0.0049510826356709, + 0.002855036873370409, + 0.004796229302883148, + 0.00484957592561841, + 0.005056985188275576, + 0.007094383239746094, + 0.00331432675011456, + 0.005183025728911161, + 0.0038789885584264994, + 0.15387868881225586, + 0.15370583534240723, + 0.12726658582687378, + 0.05388693884015083, + 0.046314362436532974, + 0.08284606784582138, + 0.031701188534498215, + 0.02897470071911812, + 0.002637038938701153, + 0.004696080926805735, + 0.0021464598830789328, + 0.0024168845266103745, + 0.0017099257092922926, + 0.00302515784278512, + 0.0011548073962330818, + 0.0015918146818876266, + 0.0031779196579009295, + 0.0018693902529776096, + 0.0022561366204172373, + 0.0012492244131863117, + 0.002106057945638895, + 0.0034509156830608845, + 0.0016965147806331515, + 0.0037882577162235975, + 0.015403239987790585, + 0.005066805984824896, + 0.0025198212824761868, + 0.002585859037935734, + 0.0034944091457873583, + 0.005999621469527483, + 0.0031709272880107164, + 0.018509438261389732, + 0.003138996195048094, + 0.003610888496041298, + 0.0016914961161091924, + 0.0019495016895234585, + 0.0028514962177723646, + 0.0036321061197668314, + 0.0018864002777263522, + 0.002704123966395855, + 0.0015477922279387712, + 0.0015340258833020926, + 0.0007651241030544043, + 0.0016124822432175279, + 0.001978615764528513, + 0.002308294642716646, + 0.0012916240375488997, + 0.0013500818749889731, + 0.00942334346473217, + 0.004600454121828079, + 0.00258520245552063, + 0.0030826327856630087, + 0.004236655309796333, + 0.007547774817794561, + 0.005992894526571035, + 0.008531678467988968 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.014214574359357357 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192597009241581 + ], + "opposite_logprob_tokens": [ + -4.264192581176758 + ], + "logprob_reward": -0.014192597009241581, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192597009241581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14808541536331177, + 0.22294411063194275, + 0.08737380802631378, + 0.24746164679527283, + 0.05358255282044411, + 0.033513158559799194, + 0.07766561210155487, + 0.040607940405607224, + 0.018782958388328552, + 0.06998282670974731 + ], + "credit_map": [ + 0.017947740852832794, + 0.01584196463227272, + 0.0200241319835186, + 0.018572455272078514, + 0.021123051643371582, + 0.016855811700224876, + 0.023769337683916092, + 0.013950926251709461, + 0.026423685252666473, + 0.01544548012316227, + 0.018480198457837105, + 0.025517094880342484, + 0.024595309048891068, + 0.059366919100284576, + 0.02084495685994625, + 0.03227047622203827, + 0.017213178798556328, + 0.009218180552124977, + 0.0080717159435153, + 0.01227977778762579, + 0.007641335017979145, + 0.019840875640511513, + 0.006820097099989653, + 0.006288650445640087, + 0.05646546557545662, + 0.042404718697071075, + 0.0504484660923481, + 0.018840458244085312, + 0.019200023263692856, + 0.033398158848285675, + 0.01349656842648983, + 0.013207790441811085, + 0.006866182666271925, + 0.005795423407107592, + 0.01228470727801323, + 0.005861859302967787, + 0.0055985115468502045, + 0.007566419430077076, + 0.0039863549172878265, + 0.0056230914779007435, + 0.005645603407174349, + 0.004101512487977743, + 0.002596373902633786, + 0.0020304585341364145, + 0.004536170046776533, + 0.005845147185027599, + 0.0029324465431272984, + 0.005825444590300322, + 0.01766677014529705, + 0.007780618965625763, + 0.0050651743076741695, + 0.006165697705000639, + 0.0053165932185947895, + 0.00818968378007412, + 0.005408080760389566, + 0.022072991356253624, + 0.006428199354559183, + 0.0043334621004760265, + 0.0045191035605967045, + 0.0030703633092343807, + 0.004234544467180967, + 0.008834109641611576, + 0.0038562878035008907, + 0.005331868305802345, + 0.0021166731603443623, + 0.0018545427592471242, + 0.0016150240553542972, + 0.0011046389117836952, + 0.003839803161099553, + 0.004043865483254194, + 0.0012097982689738274, + 0.0029986125882714987, + 0.011049519293010235, + 0.010187434032559395, + 0.005120850633829832, + 0.004336607176810503, + 0.008628182113170624, + 0.012442953884601593, + 0.006138273980468512, + 0.012078999541699886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.014192597009241581 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328236773610115 + ], + "opposite_logprob_tokens": [ + -3.773282289505005 + ], + "logprob_reward": -0.02328236773610115, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328236773610115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0645952820777893, + 0.06377245485782623, + 0.12597188353538513, + 0.33156338334083557, + 0.08097636699676514, + 0.04464951530098915, + 0.08739536255598068, + 0.07853865623474121, + 0.057222191244363785, + 0.06531483680009842 + ], + "credit_map": [ + 0.00886806845664978, + 0.005522383842617273, + 0.009467885829508305, + 0.008326075039803982, + 0.006340762600302696, + 0.009976361878216267, + 0.00894229393452406, + 0.007151457946747541, + 0.006629160605370998, + 0.006613388657569885, + 0.007055429741740227, + 0.00474513228982687, + 0.00822010263800621, + 0.012303054332733154, + 0.0058192783035337925, + 0.012386913411319256, + 0.02308136411011219, + 0.012533677741885185, + 0.012620469555258751, + 0.014415975660085678, + 0.01278765220195055, + 0.030793773010373116, + 0.01023658737540245, + 0.009502401575446129, + 0.07223895192146301, + 0.07594691962003708, + 0.060882870107889175, + 0.02934255078434944, + 0.023708755150437355, + 0.04107895493507385, + 0.013984020799398422, + 0.01438039168715477, + 0.007636763621121645, + 0.016497675329446793, + 0.016277309507131577, + 0.006606996525079012, + 0.008551990613341331, + 0.012025387957692146, + 0.005658777430653572, + 0.007721473928540945, + 0.0058660442009568214, + 0.006386537570506334, + 0.006629290524870157, + 0.004040812607854605, + 0.004734313581138849, + 0.007560592610388994, + 0.005090971942991018, + 0.004340957384556532, + 0.020308872684836388, + 0.009961935691535473, + 0.0039701275527477264, + 0.00393298314884305, + 0.007603538688272238, + 0.011610572226345539, + 0.006347962189465761, + 0.023659376427531242, + 0.008627469651401043, + 0.01179936621338129, + 0.007367509417235851, + 0.009877339005470276, + 0.01071435771882534, + 0.009977723471820354, + 0.008736196905374527, + 0.011438703164458275, + 0.005565963685512543, + 0.00796493049710989, + 0.006252257619053125, + 0.007255997974425554, + 0.009597345255315304, + 0.00773482583463192, + 0.0069278329610824585, + 0.005923042539507151, + 0.008795978501439095, + 0.007091626524925232, + 0.0028885710053145885, + 0.0031258147209882736, + 0.007349072955548763, + 0.011408334597945213, + 0.01566365733742714, + 0.008991790004074574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.02328236773610115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327980473637581 + ], + "opposite_logprob_tokens": [ + -3.773279905319214 + ], + "logprob_reward": -0.02327980473637581, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327980473637581, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1890821009874344, + 0.17824453115463257, + 0.02710423246026039, + 0.18397347629070282, + 0.03918996453285217, + 0.09089351445436478, + 0.08387935161590576, + 0.03940756618976593, + 0.03569094464182854, + 0.13253434002399445 + ], + "credit_map": [ + 0.03151743486523628, + 0.014326201751828194, + 0.03574749454855919, + 0.023925332352519035, + 0.020802341401576996, + 0.021649040281772614, + 0.019885797053575516, + 0.02122846618294716, + 0.02932804636657238, + 0.01087024062871933, + 0.028246363624930382, + 0.008965132758021355, + 0.01852271519601345, + 0.056271255016326904, + 0.007870230823755264, + 0.0181705504655838, + 0.0025656165089458227, + 0.0030774257611483335, + 0.004170611500740051, + 0.004902963526546955, + 0.0031481699552387, + 0.0035952217876911163, + 0.002790410304442048, + 0.002853814512491226, + 0.04166237637400627, + 0.04285164549946785, + 0.03255360201001167, + 0.013680209405720234, + 0.01465508434921503, + 0.0184409711509943, + 0.010857411660254002, + 0.00927218422293663, + 0.0050813923589885235, + 0.008379651233553886, + 0.005442052148282528, + 0.0039239306934177876, + 0.0035515849012881517, + 0.005473737604916096, + 0.0031111089047044516, + 0.004226505756378174, + 0.015084884129464626, + 0.011456976644694805, + 0.010178107768297195, + 0.00466515589505434, + 0.004606510512530804, + 0.019873708486557007, + 0.007295157294720411, + 0.01773301512002945, + 0.019374128431081772, + 0.008951844647526741, + 0.0071535781025886536, + 0.005668869707733393, + 0.005360123235732317, + 0.011429768986999989, + 0.005528941284865141, + 0.02041209675371647, + 0.006200634408742189, + 0.004144111182540655, + 0.003798705991357565, + 0.0020129105541855097, + 0.004314672201871872, + 0.011046789586544037, + 0.0030827668961137533, + 0.004806975834071636, + 0.004661599174141884, + 0.003862581215798855, + 0.006199321709573269, + 0.00241715251468122, + 0.0059556239284574986, + 0.006573264952749014, + 0.0023244167678058147, + 0.003696982515975833, + 0.02617456391453743, + 0.013151162303984165, + 0.007032151799649, + 0.005473438650369644, + 0.020435014739632607, + 0.026942046359181404, + 0.012858781032264233, + 0.020467178896069527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.02327980473637581 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, cinematic shot begins with a low-angle tracking shot, moving parallel to a young Southeast Asian man as he rides his bicycle at a steady, relaxed pace. He has short, windswept black hair, slightly tousled, and a subtle, contented smile on his lips. He wears a lightweight, dark teal windbreaker, zipped up over a simple white T-shirt, slim-fitting black sweatpants, and dark sneakers. He rides a sleek, matte black fixed-gear bicycle, its minimalist frame cutting cleanly through the city air. The camera then smoothly pulls back and ascends, transitioning to a high-angle wide shot, revealing the breathtaking scale of this vibrant city at dusk. He is cycling on a dedicated bike lane on a bridge or elevated roadway, surrounded by a trail of red and white lights, reflecting the busy traffic below. The sky is a deep indigo canvas, punctuated by the last fiery orange glow of sunset. Towering skyscrapers, their windows forming a mosaic of warm and cool lights, create a dazzling, futuristic skyline. Recent rainfall left the asphalt shimmering, reflecting the city's neon signs and streetlights in a shimmering, impressionistic blur of colors. The evening's predominantly cool tones were interrupted by a vibrant palette of pinks, cyans, and warm yellows in the digital billboards and architectural lighting, creating a scene both vibrant and beautiful. The overall atmosphere was one of exhilarating freedom and tranquil contemplation, capturing a single, perfect moment of journey and connection to this vast, vibrant metropolis.", + "seed": 28066, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28066_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.051605035457760094, + "vlm_logprob": -0.051605035457760094, + "vlm_yes_no_margin": 3.675, + "vlm_token_credit": -0.051605035457760094, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.118588887155056 + ], + "opposite_logprob_tokens": [ + -2.243588924407959 + ], + "logprob_reward": -0.118588887155056, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.118588887155056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1246304139494896, + 0.12643061578273773, + 0.08241868019104004, + 0.13738568127155304, + 0.1275743544101715, + 0.10291295498609543, + 0.044336557388305664, + 0.08559370040893555, + 0.05401548743247986, + 0.1147015392780304 + ], + "credit_map": [ + 0.016690295189619064, + 0.010314815677702427, + 0.012445912696421146, + 0.013651901856064796, + 0.009770004078745842, + 0.03217935189604759, + 0.010332725942134857, + 0.01924540475010872, + 0.022345269098877907, + 0.021412555128335953, + 0.01442362554371357, + 0.011701352894306183, + 0.018121013417840004, + 0.02013692632317543, + 0.00965862162411213, + 0.008631240576505661, + 0.009504366666078568, + 0.010429014451801777, + 0.008332821540534496, + 0.008787098340690136, + 0.011861843056976795, + 0.010102531872689724, + 0.00996093638241291, + 0.013440060429275036, + 0.014254896901547909, + 0.01211564801633358, + 0.013251272961497307, + 0.01816570572555065, + 0.025730108842253685, + 0.015089906752109528, + 0.014733166433870792, + 0.024044983088970184, + 0.010555701330304146, + 0.016121316701173782, + 0.008875786326825619, + 0.019441571086645126, + 0.026624590158462524, + 0.013542146421968937, + 0.016711391508579254, + 0.01570184901356697, + 0.017230739817023277, + 0.01416073739528656, + 0.013086117804050446, + 0.009671284817159176, + 0.01460113562643528, + 0.021611616015434265, + 0.00724394153803587, + 0.005307384300976992, + 0.005685877054929733, + 0.005175974685698748, + 0.005619911476969719, + 0.006879759952425957, + 0.00630742684006691, + 0.006838510278612375, + 0.004673928488045931, + 0.003155170939862728, + 0.01572779007256031, + 0.009935218840837479, + 0.008066387847065926, + 0.00701107457280159, + 0.01368582621216774, + 0.0067126075737178326, + 0.00924739334732294, + 0.01520740706473589, + 0.00895758718252182, + 0.005112692713737488, + 0.007718371693044901, + 0.005777612328529358, + 0.006538575980812311, + 0.005308777093887329, + 0.007314110640436411, + 0.007287757005542517, + 0.017141103744506836, + 0.01049275416880846, + 0.01423751749098301, + 0.011064608581364155, + 0.011940347030758858, + 0.011599580757319927, + 0.03166511282324791, + 0.006560507230460644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 1 + ], + "selected_reward": -0.118588887155056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020594101399183273 + ], + "opposite_logprob_tokens": [ + -3.8955941200256348 + ], + "logprob_reward": -0.020594101399183273, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020594101399183273, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10870460420846939, + 0.0721503347158432, + 0.11195680499076843, + 0.17951589822769165, + 0.2271537035703659, + 0.057691790163517, + 0.053268857300281525, + 0.06616725772619247, + 0.0730060562491417, + 0.05038471519947052 + ], + "credit_map": [ + 0.019922656938433647, + 0.009994973428547382, + 0.01724402606487274, + 0.009304349310696125, + 0.007277889642864466, + 0.01959044486284256, + 0.011909258551895618, + 0.013461008667945862, + 0.00929878931492567, + 0.01756594516336918, + 0.00797249935567379, + 0.008492656983435154, + 0.008202111348509789, + 0.006330549251288176, + 0.008456991985440254, + 0.005830800160765648, + 0.01578870601952076, + 0.01321487594395876, + 0.01138058491051197, + 0.011783464811742306, + 0.01507074385881424, + 0.022142956033349037, + 0.010683447122573853, + 0.011892021633684635, + 0.025243017822504044, + 0.013002855703234673, + 0.022476980462670326, + 0.023053040727972984, + 0.023687109351158142, + 0.017294973134994507, + 0.017992334440350533, + 0.036765582859516144, + 0.017012538388371468, + 0.026394812390208244, + 0.014988820999860764, + 0.03876255452632904, + 0.044451188296079636, + 0.02307947725057602, + 0.02289055474102497, + 0.039573755115270615, + 0.00799858570098877, + 0.007328715641051531, + 0.005607933271676302, + 0.008128335699439049, + 0.010697425343096256, + 0.009156057611107826, + 0.0054255141876637936, + 0.0033492213115096092, + 0.009063716046512127, + 0.007174085360020399, + 0.0046830070205032825, + 0.004301893524825573, + 0.0075068301521241665, + 0.012378494255244732, + 0.004671657923609018, + 0.0034891737159341574, + 0.014631133526563644, + 0.009666495025157928, + 0.010526965372264385, + 0.0058539798483252525, + 0.012345788069069386, + 0.004970632493495941, + 0.003929733764380217, + 0.004242532886564732, + 0.008719531819224358, + 0.00959114357829094, + 0.01020027045160532, + 0.008779407478868961, + 0.0078094410710036755, + 0.00794899184256792, + 0.01026440691202879, + 0.009692860767245293, + 0.006981094367802143, + 0.005079356487840414, + 0.0074423812329769135, + 0.004133265465497971, + 0.006012990605086088, + 0.005286314524710178, + 0.011889071203768253, + 0.0035602429416030645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.020594101399183273 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606672629714012 + ], + "opposite_logprob_tokens": [ + -4.141066551208496 + ], + "logprob_reward": -0.01606672629714012, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606672629714012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.196298286318779, + 0.03803228586912155, + 0.09552950412034988, + 0.2107405662536621, + 0.07264840602874756, + 0.12122680991888046, + 0.05641593411564827, + 0.04984000325202942, + 0.08829311281442642, + 0.07097508013248444 + ], + "credit_map": [ + 0.041579682379961014, + 0.010392328724265099, + 0.021056141704320908, + 0.007996953092515469, + 0.012317655608057976, + 0.0720956102013588, + 0.011774946004152298, + 0.01908496581017971, + 0.005998165346682072, + 0.004874454811215401, + 0.005929721053689718, + 0.002931889146566391, + 0.004834899213165045, + 0.005585587117820978, + 0.0028433615807443857, + 0.005034210160374641, + 0.013266180641949177, + 0.014563208445906639, + 0.009159923531115055, + 0.010058267042040825, + 0.013174992054700851, + 0.012579151429235935, + 0.0098893316462636, + 0.012838451191782951, + 0.02335749752819538, + 0.02573455311357975, + 0.01978447660803795, + 0.029545357450842857, + 0.03944431617856026, + 0.027406007051467896, + 0.02319902926683426, + 0.022269319742918015, + 0.00494863698258996, + 0.0055251955054700375, + 0.00796053558588028, + 0.019655130803585052, + 0.008718347176909447, + 0.008112586103379726, + 0.010582123883068562, + 0.007145851384848356, + 0.016556890681385994, + 0.01726582646369934, + 0.014433220028877258, + 0.014885998331010342, + 0.020026380196213722, + 0.01953997276723385, + 0.010524235665798187, + 0.007994286715984344, + 0.009942794218659401, + 0.007058887276798487, + 0.006970860995352268, + 0.0074889929965138435, + 0.005783585365861654, + 0.010709499940276146, + 0.004371204413473606, + 0.004090110771358013, + 0.007818099111318588, + 0.00534049142152071, + 0.008622723631560802, + 0.003598950570449233, + 0.007053655572235584, + 0.004572280216962099, + 0.00486276438459754, + 0.007971038110554218, + 0.007507490459829569, + 0.007979086600244045, + 0.010180356912314892, + 0.012611023150384426, + 0.010428240522742271, + 0.007173500489443541, + 0.015871085226535797, + 0.016542335972189903, + 0.008556855842471123, + 0.009086428210139275, + 0.008770656771957874, + 0.006192452274262905, + 0.007776098325848579, + 0.007935917936265469, + 0.013781910762190819, + 0.008874758146703243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.01606672629714012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028994083404541 + ], + "opposite_logprob_tokens": [ + -2.350289821624756 + ], + "logprob_reward": -0.10028994083404541, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028994083404541, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08708833903074265, + 0.05459579825401306, + 0.05470111593604088, + 0.08963604271411896, + 0.04530489072203636, + 0.04575653001666069, + 0.09362021088600159, + 0.06153411418199539, + 0.20260444283485413, + 0.2651585340499878 + ], + "credit_map": [ + 0.012346207164227962, + 0.009469614364206791, + 0.009454398415982723, + 0.01895478367805481, + 0.00804055668413639, + 0.007508891634643078, + 0.012247364968061447, + 0.009066508151590824, + 0.010027852840721607, + 0.005611406173557043, + 0.007706188131123781, + 0.004635433200746775, + 0.005626019090414047, + 0.008949794806540012, + 0.004272212274372578, + 0.007766887079924345, + 0.00813826359808445, + 0.005571529734879732, + 0.005367971491068602, + 0.004588272422552109, + 0.007123022340238094, + 0.009675499983131886, + 0.005908104125410318, + 0.008328452706336975, + 0.011899967677891254, + 0.009867334738373756, + 0.008194486610591412, + 0.011225024238228798, + 0.015123575925827026, + 0.010678941383957863, + 0.009592418558895588, + 0.01305429171770811, + 0.0052079916931688786, + 0.005769939627498388, + 0.005967059172689915, + 0.00795498862862587, + 0.00724684027954936, + 0.004292530007660389, + 0.003878571791574359, + 0.004986965097486973, + 0.00548752024769783, + 0.005688582547008991, + 0.004712024703621864, + 0.005822020582854748, + 0.008408825844526291, + 0.007827156223356724, + 0.00413756025955081, + 0.003672841237857938, + 0.018399816006422043, + 0.01298829447478056, + 0.009107622317969799, + 0.008730529807507992, + 0.010926934890449047, + 0.01972304657101631, + 0.007144929375499487, + 0.006599029991775751, + 0.01180575042963028, + 0.007574624847620726, + 0.007351239211857319, + 0.004074889700859785, + 0.008633222430944443, + 0.006266310345381498, + 0.008509489707648754, + 0.007318580057471991, + 0.020579978823661804, + 0.025685181841254234, + 0.017086710780858994, + 0.026484299451112747, + 0.026791458949446678, + 0.016388090327382088, + 0.03664376586675644, + 0.03294495493173599, + 0.03589720278978348, + 0.022716490551829338, + 0.017115220427513123, + 0.0225412305444479, + 0.02957700379192829, + 0.02100701443850994, + 0.09270801395177841, + 0.023596355691552162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.10028994083404541 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024855216033756733 + ], + "opposite_logprob_tokens": [ + -6.002485752105713 + ], + "logprob_reward": -0.0024855216033756733, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024855216033756733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19378620386123657, + 0.056622464209795, + 0.05692136287689209, + 0.14139197766780853, + 0.09501281380653381, + 0.11624575406312943, + 0.051340360194444656, + 0.05974330008029938, + 0.10469407588243484, + 0.12424174696207047 + ], + "credit_map": [ + 0.03613927960395813, + 0.012955882586538792, + 0.020473849028348923, + 0.020054208114743233, + 0.019405802711844444, + 0.019446948543190956, + 0.027801940217614174, + 0.03750826418399811, + 0.010309776291251183, + 0.006337431259453297, + 0.008124214597046375, + 0.004768356680870056, + 0.00613057054579258, + 0.008106482215225697, + 0.0056220474652945995, + 0.007223580963909626, + 0.008491436950862408, + 0.005856105126440525, + 0.005884242244064808, + 0.004826562944799662, + 0.007209514267742634, + 0.008897068910300732, + 0.005914878100156784, + 0.009841547347605228, + 0.020280081778764725, + 0.008924695663154125, + 0.01691203936934471, + 0.018783999606966972, + 0.018458856269717216, + 0.014710393734276295, + 0.014188718050718307, + 0.02913316898047924, + 0.010462584905326366, + 0.019216123968362808, + 0.010505967773497105, + 0.015615235082805157, + 0.010482536628842354, + 0.012183803133666515, + 0.00654686288908124, + 0.009999689646065235, + 0.021975906565785408, + 0.017483502626419067, + 0.016699573025107384, + 0.011062153615057468, + 0.012314927764236927, + 0.022146854549646378, + 0.008016959764063358, + 0.006545869167894125, + 0.008642276749014854, + 0.00512090465053916, + 0.007985414937138557, + 0.008510932326316833, + 0.005230404902249575, + 0.009747800417244434, + 0.0033620260655879974, + 0.0027405947912484407, + 0.01130303181707859, + 0.007271880749613047, + 0.011447363533079624, + 0.007742350455373526, + 0.007809935603290796, + 0.0051147048361599445, + 0.004789615981280804, + 0.004264410585165024, + 0.011642396450042725, + 0.010834850370883942, + 0.016405008733272552, + 0.009942338801920414, + 0.012447352521121502, + 0.0123415132984519, + 0.015171242877840996, + 0.015909358859062195, + 0.019316954538226128, + 0.01007892657071352, + 0.008543120697140694, + 0.01132316142320633, + 0.014632067643105984, + 0.00933905690908432, + 0.04434988647699356, + 0.006658564787358046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0024855216033756733 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "sweeping, cinematic shot begins with a low-angle tracking shot, moving parallel to a young Southeast Asian man as he rides his bicycle at a steady, relaxed pace. He has short, windswept black hair, slightly tousled, and a subtle, contented smile on his lips. He wears a lightweight, dark teal windbreaker, zipped up over a simple white T-shirt, slim-fitting black sweatpants, and dark sneakers. He rides a sleek, matte black fixed-gear bicycle, its minimalist frame cutting cleanly through the city air. The camera then smoothly pulls back and ascends, transitioning to a high-angle wide shot, revealing the breathtaking scale of this vibrant city at dusk. He is cycling on a dedicated bike lane on a bridge or elevated roadway, surrounded by a trail of red and white lights, reflecting the busy traffic below. The sky is a deep indigo canvas, punctuated by the last fiery orange glow of sunset. Towering skyscrapers, their windows forming a mosaic of warm and cool lights, create a dazzling, futuristic skyline. Recent rainfall left the asphalt shimmering, reflecting the city's neon signs and streetlights in a shimmering, impressionistic blur of colors. The evening's predominantly cool tones were interrupted by a vibrant palette of pinks, cyans, and warm yellows in the digital billboards and architectural lighting, creating a scene both vibrant and beautiful. The overall atmosphere was one of exhilarating freedom and tranquil contemplation, capturing a single, perfect moment of journey and connection to this vast, vibrant metropolis.", + "seed": 28067, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_3/index_4941_seed_28067_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.039089585607871416, + "vlm_logprob": -0.039089585607871416, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.039089585607871416, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10478372126817703 + ], + "opposite_logprob_tokens": [ + -2.354783773422241 + ], + "logprob_reward": -0.10478372126817703, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10478372126817703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09539035707712173, + 0.06573062390089035, + 0.07766436785459518, + 0.07541482895612717, + 0.08352475613355637, + 0.16855718195438385, + 0.09242710471153259, + 0.13660266995429993, + 0.12450716644525528, + 0.08018095791339874 + ], + "credit_map": [ + 0.00764781329780817, + 0.013772985897958279, + 0.01145592425018549, + 0.016136467456817627, + 0.008846685290336609, + 0.010935266502201557, + 0.011591402813792229, + 0.015003811568021774, + 0.008847810328006744, + 0.006896066013723612, + 0.009322662837803364, + 0.007616420742124319, + 0.0037703218404203653, + 0.008753894828259945, + 0.006551417056471109, + 0.013972029089927673, + 0.009144800715148449, + 0.008242241106927395, + 0.015515429899096489, + 0.008646205998957157, + 0.009490828029811382, + 0.007333219517022371, + 0.007545611821115017, + 0.011746032163500786, + 0.01964920200407505, + 0.011532185599207878, + 0.012183144688606262, + 0.008601074106991291, + 0.0056516812182962894, + 0.005660362541675568, + 0.005588599015027285, + 0.006548580247908831, + 0.007538177073001862, + 0.009301111102104187, + 0.015022503212094307, + 0.010203298181295395, + 0.014532791450619698, + 0.010259959846735, + 0.008085631765425205, + 0.008581278845667839, + 0.018816420808434486, + 0.04196867719292641, + 0.031756822019815445, + 0.029330985620617867, + 0.015318784862756729, + 0.005960032343864441, + 0.008483887650072575, + 0.01692158542573452, + 0.010458921082317829, + 0.020995404571294785, + 0.012487172149121761, + 0.009224348701536655, + 0.00997158233076334, + 0.011763657443225384, + 0.007929314859211445, + 0.009596697986125946, + 0.042155083268880844, + 0.0092313252389431, + 0.006673506461083889, + 0.005624484270811081, + 0.012417114339768887, + 0.040049564093351364, + 0.014152600429952145, + 0.00629899837076664, + 0.03233207389712334, + 0.0068980623036623, + 0.05288096144795418, + 0.014108814299106598, + 0.004991958849132061, + 0.0030676666647195816, + 0.004582827445119619, + 0.0056448024697601795, + 0.010392265394330025, + 0.008954334072768688, + 0.028505712747573853, + 0.01607515849173069, + 0.003072300460189581, + 0.0030212225392460823, + 0.004092119168490171, + 0.006067842710763216 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 8 + ], + "selected_reward": -0.10478372126817703 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029813941568136215 + ], + "opposite_logprob_tokens": [ + -3.5298140048980713 + ], + "logprob_reward": -0.029813941568136215, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029813941568136215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06718483567237854, + 0.06735497713088989, + 0.044831544160842896, + 0.03029467537999153, + 0.05277758464217186, + 0.41745615005493164, + 0.016856634989380836, + 0.0626155436038971, + 0.22258323431015015, + 0.01804484613239765 + ], + "credit_map": [ + 0.00516748521476984, + 0.011151663027703762, + 0.00831746868789196, + 0.011944620870053768, + 0.004646233282983303, + 0.0077523537911474705, + 0.007860110141336918, + 0.010344905778765678, + 0.008678996935486794, + 0.0086958147585392, + 0.008716413751244545, + 0.008267452009022236, + 0.0049457065761089325, + 0.00873610470443964, + 0.008616163395345211, + 0.010698322206735611, + 0.0036039024125784636, + 0.005310835316777229, + 0.008173776790499687, + 0.007263406179845333, + 0.004603300243616104, + 0.0045912195928394794, + 0.004481641110032797, + 0.006803461816161871, + 0.00859852321445942, + 0.004288642667233944, + 0.004384201485663652, + 0.0030076850671321154, + 0.0024733925238251686, + 0.002684255363419652, + 0.0023739642929285765, + 0.002484011696651578, + 0.00455549405887723, + 0.004603485111147165, + 0.009174109436571598, + 0.0063855694606900215, + 0.010163399390876293, + 0.006100881844758987, + 0.005530862137675285, + 0.006263785529881716, + 0.05177083984017372, + 0.12289926409721375, + 0.09756248444318771, + 0.03631591796875, + 0.04041502624750137, + 0.01587594486773014, + 0.016645299270749092, + 0.03597140312194824, + 0.00204722979106009, + 0.0025300667621195316, + 0.002531657926738262, + 0.0024575728457421064, + 0.0015720123192295432, + 0.0018281142693012953, + 0.0016613666666671634, + 0.002228614641353488, + 0.009241641499102116, + 0.005319865886121988, + 0.006983987055718899, + 0.00463741272687912, + 0.012233301065862179, + 0.010486647486686707, + 0.007817096076905727, + 0.005895588546991348, + 0.0446622259914875, + 0.03957407549023628, + 0.059436336159706116, + 0.019567713141441345, + 0.02332307957112789, + 0.01379997469484806, + 0.012464032508432865, + 0.009755780920386314, + 0.0017655764240771532, + 0.0017456116620451212, + 0.004458116367459297, + 0.002807999262586236, + 0.001381746493279934, + 0.0012915530242025852, + 0.002045579254627228, + 0.0025486627127975225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 1 + ], + "selected_reward": -0.029813941568136215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419412437826395 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.01419412437826395, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419412437826395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0668923482298851, + 0.040092114359140396, + 0.05113838613033295, + 0.07251229137182236, + 0.06392037123441696, + 0.03984680771827698, + 0.03860991820693016, + 0.25957047939300537, + 0.2601107954978943, + 0.10730649530887604 + ], + "credit_map": [ + 0.007155056577175856, + 0.005194553639739752, + 0.007058761082589626, + 0.014505688101053238, + 0.005060853436589241, + 0.006616389378905296, + 0.009145783260464668, + 0.012155265547335148, + 0.005713187623769045, + 0.004348957911133766, + 0.005569051951169968, + 0.004319976549595594, + 0.0027852305211126804, + 0.005544692277908325, + 0.005780736915767193, + 0.006030283402651548, + 0.006150944624096155, + 0.007799079176038504, + 0.00807426031678915, + 0.006111156661063433, + 0.004306718707084656, + 0.006435480434447527, + 0.0052495994605124, + 0.007011145353317261, + 0.01812063716351986, + 0.010475707240402699, + 0.014609720557928085, + 0.010419435799121857, + 0.004395296331495047, + 0.005023981910198927, + 0.004507363308221102, + 0.004960149060934782, + 0.005527921952307224, + 0.008541532792150974, + 0.00901706051081419, + 0.008674068376421928, + 0.008912437595427036, + 0.005410958081483841, + 0.008520140312612057, + 0.009316259995102882, + 0.00454396428540349, + 0.0037706377916038036, + 0.004555590450763702, + 0.007382864598184824, + 0.003978640306740999, + 0.006255568470805883, + 0.0034452846739441156, + 0.00591425783932209, + 0.004430554807186127, + 0.007867718115448952, + 0.004606521222740412, + 0.004076458513736725, + 0.005798619240522385, + 0.004328939598053694, + 0.004095721524208784, + 0.003405383788049221, + 0.07863812893629074, + 0.014841081574559212, + 0.011378447525203228, + 0.006546981632709503, + 0.02285904996097088, + 0.08510623127222061, + 0.01852128468453884, + 0.02167927287518978, + 0.05199968442320824, + 0.04723517969250679, + 0.06768004596233368, + 0.02270752191543579, + 0.02869703248143196, + 0.01689867489039898, + 0.014422034844756126, + 0.010470598004758358, + 0.01253245398402214, + 0.011778220534324646, + 0.03327234461903572, + 0.01706620119512081, + 0.007195115089416504, + 0.005987199954688549, + 0.00642592366784811, + 0.013049031607806683 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.01419412437826395 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043044786900281906 + ], + "opposite_logprob_tokens": [ + -3.1680448055267334 + ], + "logprob_reward": -0.043044786900281906, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043044786900281906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03596189245581627, + 0.012924117967486382, + 0.01787923090159893, + 0.06024618074297905, + 0.03163891285657883, + 0.20437905192375183, + 0.027106966823339462, + 0.20557449758052826, + 0.37370267510414124, + 0.030586490407586098 + ], + "credit_map": [ + 0.005311121698468924, + 0.005432494450360537, + 0.0034713728819042444, + 0.00679316371679306, + 0.002116125775501132, + 0.00392582081258297, + 0.003503997577354312, + 0.005407798569649458, + 0.0020432143937796354, + 0.0014328802935779095, + 0.001858735573478043, + 0.002250795951113105, + 0.0009013881790451705, + 0.0009197033359669149, + 0.001204388914629817, + 0.0023130092304199934, + 0.0016464263899251819, + 0.003075634827837348, + 0.0026916672941297293, + 0.0024934494867920876, + 0.0012430093483999372, + 0.0018986336654052138, + 0.001795498887076974, + 0.0030349106527864933, + 0.015131382271647453, + 0.010184228420257568, + 0.009832904674112797, + 0.007353899069130421, + 0.004230298567563295, + 0.004416999407112598, + 0.0039932685904204845, + 0.005103196483105421, + 0.003530454356223345, + 0.004614535253494978, + 0.005502115469425917, + 0.004242032300680876, + 0.0037623087409883738, + 0.0036875521764159203, + 0.0031752220820635557, + 0.0031246906146407127, + 0.025407979264855385, + 0.0615244135260582, + 0.0488419346511364, + 0.016244811937212944, + 0.02103608101606369, + 0.006469985470175743, + 0.007965623401105404, + 0.016888221725821495, + 0.004009594209492207, + 0.004420646466314793, + 0.0023658964782953262, + 0.0020882096141576767, + 0.004573590122163296, + 0.0025721504352986813, + 0.0022926011588424444, + 0.004784275311976671, + 0.05459234118461609, + 0.011364220641553402, + 0.00969309639185667, + 0.007457927335053682, + 0.016772108152508736, + 0.06890049576759338, + 0.015281632542610168, + 0.02151266485452652, + 0.07140286266803741, + 0.07250723242759705, + 0.09037848562002182, + 0.03259032964706421, + 0.04352991655468941, + 0.025726595893502235, + 0.02175609953701496, + 0.015811102464795113, + 0.004224950447678566, + 0.0037131391000002623, + 0.010391517542302608, + 0.00567466439679265, + 0.0013612277107313275, + 0.0012481707381084561, + 0.0015890210634097457, + 0.0023837971966713667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 5 + ], + "selected_reward": -0.043044786900281906 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young Southeast Asian man riding a bicycle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a dark teal windbreaker over a white T-shirt and black sweatpants?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the bicycle a sleek, matte black fixed-gear model?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video begin with a low-angle tracking shot and transition to a high-angle wide shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the setting a city at dusk with a deep indigo sky and vibrant neon lights reflecting on wet asphalt?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003611353924497962 + ], + "opposite_logprob_tokens": [ + -5.6286115646362305 + ], + "logprob_reward": -0.003611353924497962, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003611353924497962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043230168521404266, + 0.11388108879327774, + 0.020940128713846207, + 0.0942544937133789, + 0.0242617130279541, + 0.28478965163230896, + 0.03804858773946762, + 0.13115359842777252, + 0.13323725759983063, + 0.11620331555604935 + ], + "credit_map": [ + 0.005345022305846214, + 0.0070670051500201225, + 0.00504430802538991, + 0.007209856994450092, + 0.0027859702240675688, + 0.004901417065411806, + 0.0038406183011829853, + 0.007035968359559774, + 0.019614091143012047, + 0.012887859717011452, + 0.016176369041204453, + 0.013585682958364487, + 0.0076352800242602825, + 0.01326330378651619, + 0.011438068933784962, + 0.01928042620420456, + 0.001709811738692224, + 0.002268858253955841, + 0.002946489956229925, + 0.0050863721407949924, + 0.0016145693371072412, + 0.0016066735843196511, + 0.001976185478270054, + 0.003731165546923876, + 0.02405388467013836, + 0.019528549164533615, + 0.010966326110064983, + 0.009413708001375198, + 0.0073281098157167435, + 0.00889265164732933, + 0.005576264578849077, + 0.008494997397065163, + 0.002720785792917013, + 0.004340077750384808, + 0.004914165940135717, + 0.003218651982024312, + 0.00257471832446754, + 0.0026332035195082426, + 0.001956075429916382, + 0.0019040339393541217, + 0.03573139011859894, + 0.08623874187469482, + 0.0698775127530098, + 0.01934773102402687, + 0.02804512158036232, + 0.010985404253005981, + 0.010671642608940601, + 0.023892123252153397, + 0.004957288969308138, + 0.005457752384245396, + 0.005434796214103699, + 0.0045426408760249615, + 0.0034630149602890015, + 0.005894479341804981, + 0.004035714082419872, + 0.00426289951428771, + 0.036513518542051315, + 0.009388746693730354, + 0.00797130074352026, + 0.00519793014973402, + 0.010705854743719101, + 0.04287073016166687, + 0.01018540095537901, + 0.008320106193423271, + 0.028140783309936523, + 0.014873393811285496, + 0.0467880442738533, + 0.01570184901356697, + 0.008139334619045258, + 0.006312647834420204, + 0.005825172644108534, + 0.007456028368324041, + 0.015413441695272923, + 0.013703382574021816, + 0.04164406284689903, + 0.02089855819940567, + 0.004408559761941433, + 0.0041061448864638805, + 0.0059069618582725525, + 0.010122193954885006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.003611353924497962 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_tensor([2649, 2649])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_tensor([2649, 2649])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..0ac5d41d6fb0c08987b4f1443b90048812e56459 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_tensor([2649, 2649])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot at eye level, rendered in a bright fantasy art style, begins with a slow, deliberate zoom-in to an enchanted forest clearing. The scene is bathed in a cool, magical glow, with silvery moonlight filtering through the dense canopy of ancient, gnarled trees and dotting the mossy ground with spots. In the center of the clearing, a sleek black cat with fur the color of polished obsidian and large, expressive amethyst eyes, its body curled protectively into a crescent, surrounded by a sleeping kitten with fuzzy gray fur. Clusters of bioluminescent mushrooms, pulsating shades of sapphire blue and ethereal violet, sprout from the fertile earth, casting a soft, otherworldly glow that illuminates the pair and nearby delicate, shimmering flowers, their petals seemingly spun from the moonlight. As the camera zooms in, tightening the frame to focus on the tender moment, the mother cat cautiously lowers her head, her whiskers twitching slightly, and gently nuzzles the kitten's cheek. Her long, graceful tail sways in a slow, hypnotic S-curve, her fingertips twitching contentedly, creating a serene and moving atmosphere of pure, protective love.", + "seed": 11113, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_2649_seed_11113_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.29316684864461423, + "vlm_logprob": -0.29316684864461423, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.29316684864461423, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3142259120941162 + ], + "opposite_logprob_tokens": [ + -0.3142259120941162 + ], + "logprob_reward": -1.3142259120941162, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3142259120941162, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.044800955802202225, + 0.019261285662651062, + 0.014650515280663967, + 0.022846639156341553, + 0.011395017616450787, + 0.012323292903602123, + 0.03726176172494888, + 0.5704530477523804, + 0.2528334856033325, + 0.01417402271181345 + ], + "credit_map": [ + 0.004789610393345356, + 0.005170220043510199, + 0.005592628847807646, + 0.005704257637262344, + 0.00731618981808424, + 0.005829400382936001, + 0.006711529567837715, + 0.0036871181800961494, + 0.0025655862409621477, + 0.0029373373836278915, + 0.0021033240482211113, + 0.002127629006281495, + 0.003051775274798274, + 0.002152550732716918, + 0.002492993837222457, + 0.0018300883239135146, + 0.0017716537695378065, + 0.0018623123178258538, + 0.0012959875166416168, + 0.0011297313030809164, + 0.0010713398223742843, + 0.004092210438102484, + 0.0010007928358390927, + 0.002426487859338522, + 0.003039936302229762, + 0.0013092472217977047, + 0.0019021086627617478, + 0.004710568580776453, + 0.0019609699957072735, + 0.0038869313430041075, + 0.0026389765553176403, + 0.0033978987485170364, + 0.0015780609101057053, + 0.00100901466794312, + 0.00190561986528337, + 0.0013788728974759579, + 0.000693580717779696, + 0.0019600174855440855, + 0.0014692950062453747, + 0.0014005553675815463, + 0.0012675630860030651, + 0.0008991313516162336, + 0.0019095733296126127, + 0.0018992384430021048, + 0.001186506124213338, + 0.0017027519643306732, + 0.0014852469321340322, + 0.0019732823129743338, + 0.008324428461492062, + 0.003171479795128107, + 0.00321765523403883, + 0.004029395990073681, + 0.003835609182715416, + 0.007476907689124346, + 0.0029971483163535595, + 0.004209136124700308, + 0.03648294880986214, + 0.032975371927022934, + 0.181565523147583, + 0.21812704205513, + 0.010465255007147789, + 0.03310219570994377, + 0.034924305975437164, + 0.022810403257608414, + 0.020162902772426605, + 0.00885874591767788, + 0.10414136946201324, + 0.06624270975589752, + 0.0059579540975391865, + 0.02613971009850502, + 0.011513844132423401, + 0.009816261939704418, + 0.001134667545557022, + 0.0009690063307061791, + 0.003423952031880617, + 0.0027084050234407187, + 0.0010869044344872236, + 0.0018590566469356418, + 0.001697103725746274, + 0.0012949276715517044 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -1.3142259120941162 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009809385053813457 + ], + "opposite_logprob_tokens": [ + -4.634809494018555 + ], + "logprob_reward": -0.009809385053813457, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009809385053813457, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10246537625789642, + 0.04583857208490372, + 0.03508101403713226, + 0.058791663497686386, + 0.018034636974334717, + 0.06397461146116257, + 0.08301705121994019, + 0.3259718716144562, + 0.23431700468063354, + 0.032508280128240585 + ], + "credit_map": [ + 0.010931327007710934, + 0.012536395341157913, + 0.012856666930019855, + 0.012739254161715508, + 0.01684771105647087, + 0.012510789558291435, + 0.01613091491162777, + 0.007912307977676392, + 0.005916272755712271, + 0.006726887077093124, + 0.005193077493458986, + 0.005267431493848562, + 0.007615947164595127, + 0.004947581794112921, + 0.006134164519608021, + 0.004037205595523119, + 0.004287585616111755, + 0.004709691274911165, + 0.0030537100974470377, + 0.0030251468997448683, + 0.002781192772090435, + 0.00959005020558834, + 0.0026251315139234066, + 0.005008503329008818, + 0.008453063666820526, + 0.003263165708631277, + 0.004990516696125269, + 0.011200403794646263, + 0.00525343744084239, + 0.00998162105679512, + 0.006986893713474274, + 0.008662556298077106, + 0.0023201070725917816, + 0.0017909288872033358, + 0.0033311781007796526, + 0.002346252789720893, + 0.0015254694735631347, + 0.002418105024844408, + 0.0023228470236063004, + 0.001979745924472809, + 0.004983397200703621, + 0.0052057113498449326, + 0.012934999540448189, + 0.011143247596919537, + 0.008531162515282631, + 0.006982848513871431, + 0.00753584923222661, + 0.006657386664301157, + 0.01919396035373211, + 0.006645535584539175, + 0.006471728906035423, + 0.009148307144641876, + 0.00795948039740324, + 0.017673233523964882, + 0.007145251613110304, + 0.008779539726674557, + 0.020344190299510956, + 0.01922936737537384, + 0.1032053753733635, + 0.12345730513334274, + 0.006854906678199768, + 0.019157424569129944, + 0.019817106425762177, + 0.013906157575547695, + 0.018428387120366096, + 0.008321051485836506, + 0.09604840725660324, + 0.06131860986351967, + 0.0056894319131970406, + 0.02450861781835556, + 0.010636329650878906, + 0.00936614815145731, + 0.0028186794370412827, + 0.002404456492513418, + 0.007398013956844807, + 0.006095222197473049, + 0.0024322406388819218, + 0.004502883180975914, + 0.003825058229267597, + 0.003031725063920021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.009809385053813457 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865625500679016 + ], + "opposite_logprob_tokens": [ + -3.0486562252044678 + ], + "logprob_reward": -0.04865625500679016, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865625500679016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051354166120290756, + 0.014612521976232529, + 0.02150578796863556, + 0.011677342467010021, + 0.11824226379394531, + 0.03886932507157326, + 0.05492893606424332, + 0.6482905149459839, + 0.022993523627519608, + 0.017525583505630493 + ], + "credit_map": [ + 0.005598020274192095, + 0.0066938153468072414, + 0.005774957127869129, + 0.006772243417799473, + 0.00854676216840744, + 0.00598222017288208, + 0.007442370522767305, + 0.004543778486549854, + 0.0016134741017594934, + 0.0019951241556555033, + 0.0017120616976171732, + 0.0020278487354516983, + 0.0015446654288098216, + 0.002029591705650091, + 0.00196481472812593, + 0.0017249403754249215, + 0.0027751275338232517, + 0.0029863757081329823, + 0.0017599387792870402, + 0.002087893197312951, + 0.001772291143424809, + 0.005953632760792971, + 0.0017801758367568254, + 0.0023903511464595795, + 0.0016620251117274165, + 0.0009536822908557951, + 0.0010030876146629453, + 0.0012984349159523845, + 0.0018005408346652985, + 0.0018159194150939584, + 0.0017292080447077751, + 0.0014144449960440397, + 0.015641575679183006, + 0.007204336114227772, + 0.013085980899631977, + 0.021641956642270088, + 0.007572122849524021, + 0.022594328969717026, + 0.013748855330049992, + 0.01675311103463173, + 0.003006291575729847, + 0.0029501491226255894, + 0.008158218115568161, + 0.007227364461869001, + 0.005188888870179653, + 0.004190229810774326, + 0.004292280413210392, + 0.0038559038657695055, + 0.008918619714677334, + 0.004937854595482349, + 0.006324521731585264, + 0.005502739455550909, + 0.007286894600838423, + 0.007647802587598562, + 0.006687969900667667, + 0.007622532080858946, + 0.04270974174141884, + 0.037453606724739075, + 0.2034645825624466, + 0.2454434037208557, + 0.013685078360140324, + 0.03873157501220703, + 0.04007376357913017, + 0.026728728786110878, + 0.0018889295170083642, + 0.0018131384858861566, + 0.004101743921637535, + 0.0033808525186032057, + 0.0018992829136550426, + 0.004070273134857416, + 0.002525780815631151, + 0.0033135239500552416, + 0.0012776571093127131, + 0.0014125347370281816, + 0.004022177774459124, + 0.003047540318220854, + 0.0017970424378290772, + 0.002300909487530589, + 0.0022900826297700405, + 0.001377637730911374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.04865625500679016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194007031619549 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194007031619549, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194007031619549, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025391723960638046, + 0.016618672758340836, + 0.013101915828883648, + 0.0084584541618824, + 0.017990589141845703, + 0.011552252806723118, + 0.034212227910757065, + 0.6770424246788025, + 0.18344475328922272, + 0.012186991982161999 + ], + "credit_map": [ + 0.0027159862220287323, + 0.003302260534837842, + 0.0031754677183926105, + 0.003021207405254245, + 0.004179373849183321, + 0.003524996340274811, + 0.00350305438041687, + 0.0019693763460963964, + 0.0022762881126254797, + 0.0025897701270878315, + 0.0019942300859838724, + 0.0015664442908018827, + 0.002913541393354535, + 0.0017583138542249799, + 0.0021095857955515385, + 0.0014104973524808884, + 0.0015488560311496258, + 0.0016478101024404168, + 0.001115836901590228, + 0.00110817386303097, + 0.0011635152623057365, + 0.003376416862010956, + 0.001005662837997079, + 0.0021356423385441303, + 0.000993036781437695, + 0.0005784336244687438, + 0.0007649548933841288, + 0.001488835085183382, + 0.000634031486697495, + 0.001798861543647945, + 0.0008469392196275294, + 0.0013533603632822633, + 0.0026247412897646427, + 0.001287525868974626, + 0.0027033775113523006, + 0.0024940890725702047, + 0.001113929320126772, + 0.0033249715343117714, + 0.0021773173939436674, + 0.0022646356374025345, + 0.0010476289317011833, + 0.0008511788328178227, + 0.0017194325337186456, + 0.0017895260825753212, + 0.0012710103765130043, + 0.0014281567418947816, + 0.0015200716443359852, + 0.0019252465572208166, + 0.007893644273281097, + 0.0030548584181815386, + 0.003289043204858899, + 0.0034527487587183714, + 0.0031689470633864403, + 0.006613840349018574, + 0.0027655812446027994, + 0.003973560407757759, + 0.04339487850666046, + 0.039121486246585846, + 0.21561193466186523, + 0.25886276364326477, + 0.012406161054968834, + 0.03910817578434944, + 0.04152575135231018, + 0.027011217549443245, + 0.014588980004191399, + 0.006436247378587723, + 0.07564894109964371, + 0.04799656569957733, + 0.004341904539614916, + 0.01899377629160881, + 0.008384277112782001, + 0.007054040674120188, + 0.001000518910586834, + 0.0008383207605220377, + 0.002795559586957097, + 0.002345861168578267, + 0.0008425673004239798, + 0.0016233031637966633, + 0.001447490998543799, + 0.001293369336053729 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.014194007031619549 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894868403673172 + ], + "opposite_logprob_tokens": [ + -2.578948736190796 + ], + "logprob_reward": -0.07894868403673172, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894868403673172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.022886011749505997, + 0.020412815734744072, + 0.00837291032075882, + 0.00974232703447342, + 0.0705963671207428, + 0.008456766605377197, + 0.014028279110789299, + 0.7702754139900208, + 0.06100306659936905, + 0.014226092025637627 + ], + "credit_map": [ + 0.0021641904022544622, + 0.0019863927736878395, + 0.002191868843510747, + 0.004671128001064062, + 0.0025112375151365995, + 0.0035457536578178406, + 0.003892158390954137, + 0.0019232805352658033, + 0.0022432394325733185, + 0.002118076663464308, + 0.003915346693247557, + 0.002260324079543352, + 0.0028800833970308304, + 0.0019576800987124443, + 0.0029391974676400423, + 0.0020988688338547945, + 0.0013057285686954856, + 0.0012520153541117907, + 0.00070047756889835, + 0.0009503576438874006, + 0.000991717097349465, + 0.0012678108178079128, + 0.0009072912507690489, + 0.0009975117864087224, + 0.0009928862564265728, + 0.0009503249893896282, + 0.0009166168747469783, + 0.0010742407757788897, + 0.0016679371474310756, + 0.0010304820025339723, + 0.0018894713139161468, + 0.0012203673832118511, + 0.009763333015143871, + 0.004414145834743977, + 0.008632190525531769, + 0.01267135702073574, + 0.004353654105216265, + 0.013158518821001053, + 0.008227542974054813, + 0.009375626221299171, + 0.0007564689731225371, + 0.0006196689209900796, + 0.0008115344098769128, + 0.0008508199243806303, + 0.0005071433843113482, + 0.0012702796375378966, + 0.0013610424939543009, + 0.002279809210449457, + 0.002792968414723873, + 0.0014294972643256187, + 0.001596317975781858, + 0.001721946056932211, + 0.0013873777352273464, + 0.002383571583777666, + 0.001153739751316607, + 0.0015628592809662223, + 0.04940565302968025, + 0.04448694363236427, + 0.24485836923122406, + 0.2943033277988434, + 0.014187073335051537, + 0.0450545959174633, + 0.047256939113140106, + 0.030722441151738167, + 0.004928234964609146, + 0.002619927516207099, + 0.02346694841980934, + 0.014896582812070847, + 0.002341552171856165, + 0.005891561508178711, + 0.003259652527049184, + 0.003598603419959545, + 0.0010922724613919854, + 0.0014992402866482735, + 0.0030456525273621082, + 0.0026371520943939686, + 0.001336465822532773, + 0.0016435791039839387, + 0.0015159872127696872, + 0.0014557434478774667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 8 + ], + "selected_reward": -0.07894868403673172 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot at eye level, rendered in a bright fantasy art style, begins with a slow, deliberate zoom-in to an enchanted forest clearing. The scene is bathed in a cool, magical glow, with silvery moonlight filtering through the dense canopy of ancient, gnarled trees and dotting the mossy ground with spots. In the center of the clearing, a sleek black cat with fur the color of polished obsidian and large, expressive amethyst eyes, its body curled protectively into a crescent, surrounded by a sleeping kitten with fuzzy gray fur. Clusters of bioluminescent mushrooms, pulsating shades of sapphire blue and ethereal violet, sprout from the fertile earth, casting a soft, otherworldly glow that illuminates the pair and nearby delicate, shimmering flowers, their petals seemingly spun from the moonlight. As the camera zooms in, tightening the frame to focus on the tender moment, the mother cat cautiously lowers her head, her whiskers twitching slightly, and gently nuzzles the kitten's cheek. Her long, graceful tail sways in a slow, hypnotic S-curve, her fingertips twitching contentedly, creating a serene and moving atmosphere of pure, protective love.", + "seed": 11114, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_2649_seed_11114_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.8216025285422802, + "vlm_logprob": -0.8216025285422802, + "vlm_yes_no_margin": 1.1, + "vlm_token_credit": -0.8216025285422802, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -3.6520233154296875 + ], + "opposite_logprob_tokens": [ + -0.027023233473300934 + ], + "logprob_reward": -3.6520233154296875, + "yes_no_margin_reward": -3.625, + "token_credit_reward": -3.6520233154296875, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.11082879453897476, + 0.05738678574562073, + 0.029317451640963554, + 0.009390648454427719, + 0.014640727080404758, + 0.007353673223406076, + 0.013174707069993019, + 0.00498151034116745, + 0.7254424691200256, + 0.027483193203806877 + ], + "credit_map": [ + 0.010693933814764023, + 0.011053057387471199, + 0.017078332602977753, + 0.011313383467495441, + 0.011723114177584648, + 0.011661946773529053, + 0.011500355787575245, + 0.025804677978157997, + 0.005898396018892527, + 0.00814791303128004, + 0.007414879743009806, + 0.006094163283705711, + 0.004673006013035774, + 0.008932889439165592, + 0.005204492248594761, + 0.011021045036613941, + 0.0028960523195564747, + 0.004147508181631565, + 0.004884160589426756, + 0.0032091534230858088, + 0.0026471049059182405, + 0.004444057121872902, + 0.0024779378436505795, + 0.004611476324498653, + 0.001003603800199926, + 0.0011148924240842462, + 0.0012095660204067826, + 0.0011717049637809396, + 0.0009019450517371297, + 0.0014384647365659475, + 0.0011012020986527205, + 0.0014492691261693835, + 0.0015255550388246775, + 0.001244836370460689, + 0.0019558658823370934, + 0.0023890191223472357, + 0.002936406759545207, + 0.0013708597980439663, + 0.001811571535654366, + 0.0014066133880987763, + 0.0010342529276385903, + 0.0006230107392184436, + 0.001460313331335783, + 0.0006393113872036338, + 0.0007211551419459283, + 0.0010821057949215174, + 0.0008061923435889184, + 0.0009873317321762443, + 0.0018155297730118036, + 0.0009632317232899368, + 0.002470343606546521, + 0.0016758784186095, + 0.0013867056695744395, + 0.001525780069641769, + 0.001987817697227001, + 0.0013494208687916398, + 0.0005455534555949271, + 0.0004296527476981282, + 0.0006642907974310219, + 0.000693612964823842, + 0.0006638053455390036, + 0.0005734325968660414, + 0.0005300675984472036, + 0.0008810944855213165, + 0.07284849882125854, + 0.038495954126119614, + 0.039972346276044846, + 0.2070164829492569, + 0.045046377927064896, + 0.21652692556381226, + 0.055038291960954666, + 0.0504976250231266, + 0.003279836382716894, + 0.0038656594697386026, + 0.0026953204069286585, + 0.007764892186969519, + 0.0025023494381457567, + 0.0023270421661436558, + 0.002542585600167513, + 0.0025055084843188524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -3.6520233154296875 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06232789158821106 + ], + "opposite_logprob_tokens": [ + -2.8123278617858887 + ], + "logprob_reward": -0.06232789158821106, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06232789158821106, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.113568015396595, + 0.04393387213349342, + 0.02672131545841694, + 0.011458653956651688, + 0.012378696352243423, + 0.009885248728096485, + 0.01474297046661377, + 0.012833473272621632, + 0.735893964767456, + 0.018583836033940315 + ], + "credit_map": [ + 0.011114812456071377, + 0.012444455176591873, + 0.02009093388915062, + 0.010268842801451683, + 0.008412704803049564, + 0.011247160844504833, + 0.010402821004390717, + 0.029586277902126312, + 0.004481842275708914, + 0.006445905659347773, + 0.005377526860684156, + 0.00488591194152832, + 0.0034358613193035126, + 0.006879725493490696, + 0.0041243163868784904, + 0.008302779868245125, + 0.0024592664558440447, + 0.0037921268958598375, + 0.004276231862604618, + 0.0032618672121316195, + 0.002305980073288083, + 0.003896815236657858, + 0.002359789563342929, + 0.0043692379258573055, + 0.0011352028232067823, + 0.0014047777513042092, + 0.0013575420016422868, + 0.0014846600824967027, + 0.0014459890080615878, + 0.0015582098858430982, + 0.0013466669479385018, + 0.001725605339743197, + 0.001279932097531855, + 0.0011931644985452294, + 0.0016512476140633225, + 0.0017307086382061243, + 0.002400939119979739, + 0.0013713384978473186, + 0.00147859170101583, + 0.001272774301469326, + 0.0010397000005468726, + 0.0008451915346086025, + 0.0019348773639649153, + 0.0011654353002086282, + 0.001471440540626645, + 0.0011710090329870582, + 0.0012269417056813836, + 0.0010306538315489888, + 0.001971434336155653, + 0.001025036326609552, + 0.002831656252965331, + 0.0018143433844670653, + 0.0016638422384858131, + 0.0018666853429749608, + 0.0020126516465097666, + 0.0015573218697682023, + 0.0013903991784900427, + 0.001305494923144579, + 0.001144868554547429, + 0.0026118026580661535, + 0.0016521555371582508, + 0.0010877865133807063, + 0.0016923047369346023, + 0.0019486609380692244, + 0.07436595857143402, + 0.038508594036102295, + 0.04074075073003769, + 0.2086879014968872, + 0.04534535855054855, + 0.2204275131225586, + 0.056382860988378525, + 0.0514349602162838, + 0.002716470742598176, + 0.001649711630307138, + 0.0019868321251124144, + 0.003217174671590328, + 0.001551143592223525, + 0.00254974071867764, + 0.00278115039691329, + 0.0021316108759492636 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.06232789158821106 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11304840445518494 + ], + "opposite_logprob_tokens": [ + -2.2380483150482178 + ], + "logprob_reward": -0.11304840445518494, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11304840445518494, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08978945016860962, + 0.03332860767841339, + 0.043541278690099716, + 0.009359682910144329, + 0.010313164442777634, + 0.011583670973777771, + 0.014558072201907635, + 0.024367477744817734, + 0.7527610063552856, + 0.010397599078714848 + ], + "credit_map": [ + 0.008622846566140652, + 0.010585072450339794, + 0.02050023153424263, + 0.007649072911590338, + 0.005772152449935675, + 0.009084033779799938, + 0.007795108016580343, + 0.019780941307544708, + 0.0036550506483763456, + 0.005213530734181404, + 0.004532586317509413, + 0.003549617948010564, + 0.0029462084639817476, + 0.0048581380397081375, + 0.002711100038141012, + 0.005862378049641848, + 0.0031138979829847813, + 0.0051915026269853115, + 0.00648646242916584, + 0.007437560241669416, + 0.0061850217171013355, + 0.004640435799956322, + 0.006335425656288862, + 0.0041509754955768585, + 0.0008530417690053582, + 0.0013004988431930542, + 0.0011166766053065658, + 0.00167566176969558, + 0.0015915617113932967, + 0.0010733932722359896, + 0.0009168903343379498, + 0.0008319596527144313, + 0.0010142327519133687, + 0.001021409872919321, + 0.0014932892518118024, + 0.0016654400387778878, + 0.0017504901625216007, + 0.0011296756565570831, + 0.0013182148104533553, + 0.0009204125381074846, + 0.0014616711996495724, + 0.001199970138259232, + 0.0020073596388101578, + 0.0011759760091081262, + 0.0014829448191449046, + 0.0011891401372849941, + 0.001357864704914391, + 0.001708745607174933, + 0.0016981104854494333, + 0.0010466675739735365, + 0.0026215522084385157, + 0.0017652156529948115, + 0.00175633339677006, + 0.0016020485199987888, + 0.002215363783761859, + 0.0018527822103351355, + 0.0031769948545843363, + 0.001814192859455943, + 0.0019080648198723793, + 0.007143697235733271, + 0.001843011355958879, + 0.0017538091633468866, + 0.0020623805467039347, + 0.0046653300523757935, + 0.07565196603536606, + 0.03975725919008255, + 0.04157569259405136, + 0.21438749134540558, + 0.046820227056741714, + 0.2243458479642868, + 0.05733737722039223, + 0.052885204553604126, + 0.0014422237873077393, + 0.0012885495088994503, + 0.0011572432704269886, + 0.0021908876951783895, + 0.0011894591152668, + 0.001176804187707603, + 0.000865681329742074, + 0.0010867505334317684 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.11304840445518494 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05500326678156853 + ], + "opposite_logprob_tokens": [ + -2.9300031661987305 + ], + "logprob_reward": -0.05500326678156853, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05500326678156853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05893060564994812, + 0.022928329184651375, + 0.0170392245054245, + 0.007809614297002554, + 0.011865674518048763, + 0.010674390010535717, + 0.009259001351892948, + 0.007688478566706181, + 0.8376670479774475, + 0.016137713566422462 + ], + "credit_map": [ + 0.0056485943496227264, + 0.006204874720424414, + 0.007672221399843693, + 0.006423004902899265, + 0.006080055609345436, + 0.006471272092312574, + 0.00579213397577405, + 0.014638442546129227, + 0.0017541649285703897, + 0.0026979048270732164, + 0.0035617377143353224, + 0.002306526293978095, + 0.002259893575683236, + 0.0034698352683335543, + 0.0021304842084646225, + 0.004747779108583927, + 0.001970010343939066, + 0.0017702901968732476, + 0.002240299480035901, + 0.0023166837636381388, + 0.0015321369282901287, + 0.0021590597461909056, + 0.001886627054773271, + 0.003164115361869335, + 0.000791441067121923, + 0.0008150919456966221, + 0.0010073946323245764, + 0.0010110152652487159, + 0.0012645572423934937, + 0.0009538194863125682, + 0.0008815600886009634, + 0.0010847331723198295, + 0.0012155186850577593, + 0.001126574003137648, + 0.0016750185750424862, + 0.001226103282533586, + 0.0009019272984005511, + 0.002017176477238536, + 0.0013827031943947077, + 0.002320651663467288, + 0.0011150556383654475, + 0.0010479396441951394, + 0.0018898055423051119, + 0.0012438867706805468, + 0.0014475374482572079, + 0.0016691350610926747, + 0.0009102949989028275, + 0.001350733800791204, + 0.001274453243240714, + 0.0005979631678201258, + 0.0017869975417852402, + 0.0011545905144885182, + 0.0010690772905945778, + 0.0011655999114736915, + 0.0012649648124352098, + 0.0009453542879782617, + 0.0010335968108847737, + 0.0007321869488805532, + 0.0009906202321872115, + 0.0011584381572902203, + 0.0006636265898123384, + 0.0012542534386739135, + 0.0007565930718556046, + 0.0010991620365530252, + 0.08454044163227081, + 0.043791987001895905, + 0.046065837144851685, + 0.23773875832557678, + 0.05125480145215988, + 0.2525159418582916, + 0.06327126175165176, + 0.05848793312907219, + 0.0023183664306998253, + 0.0019801966845989227, + 0.0017544451402500272, + 0.0031822463497519493, + 0.0013413671404123306, + 0.0016732978401705623, + 0.0026047739665955305, + 0.0012830179184675217 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.05500326678156853 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560976445674896 + ], + "opposite_logprob_tokens": [ + -1.6006097793579102 + ], + "logprob_reward": -0.22560976445674896, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560976445674896, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03262726217508316, + 0.018003258854150772, + 0.012252208776772022, + 0.00672845309600234, + 0.00904654897749424, + 0.01015481911599636, + 0.005810399539768696, + 0.010193952359259129, + 0.8798522353172302, + 0.015330837108194828 + ], + "credit_map": [ + 0.0031664015259593725, + 0.004151766654103994, + 0.006907619535923004, + 0.002631769049912691, + 0.0024716586340218782, + 0.0032550503965467215, + 0.002418037038296461, + 0.007624958176165819, + 0.0017898684600368142, + 0.0028065363876521587, + 0.0025871568359434605, + 0.0020765371154993773, + 0.0016139984363690019, + 0.002861954504624009, + 0.0015874147647991776, + 0.002679790835827589, + 0.0012199855409562588, + 0.0014615723630413413, + 0.0020810673013329506, + 0.001458970713429153, + 0.0009838946862146258, + 0.0016305932076647878, + 0.0015421495772898197, + 0.0018739751540124416, + 0.0005688603851012886, + 0.0006222527590580285, + 0.0010712542571127415, + 0.001056823180988431, + 0.001054240739904344, + 0.0005668806843459606, + 0.0010066470131278038, + 0.000781493610702455, + 0.0007878839387558401, + 0.0007816657307557762, + 0.0012386260787025094, + 0.0012434072559699416, + 0.0020935898646712303, + 0.0007918178453110158, + 0.001267569838091731, + 0.000841987959574908, + 0.0014191429363563657, + 0.0013416502624750137, + 0.0014965016162022948, + 0.0011358880437910557, + 0.0015352960908785462, + 0.0008456816431134939, + 0.0011662916513159871, + 0.0012143671046942472, + 0.0005737753235734999, + 0.0006812709616497159, + 0.0007101609953679144, + 0.0007881142082624137, + 0.0009582281345501542, + 0.0004580638196785003, + 0.0010511999716982245, + 0.0005895855138078332, + 0.0008468865999020636, + 0.0012040473520755768, + 0.0008775662281550467, + 0.0015936372801661491, + 0.001343120587989688, + 0.001213312498293817, + 0.0009412401122972369, + 0.002174140652641654, + 0.0894898846745491, + 0.04547252133488655, + 0.04787036404013634, + 0.2474399209022522, + 0.05335473641753197, + 0.26780131459236145, + 0.06659494340419769, + 0.06182856857776642, + 0.0019437439041212201, + 0.002530498895794153, + 0.002742868149653077, + 0.0016893105348572135, + 0.00187874271068722, + 0.0011615909170359373, + 0.0020590329077094793, + 0.0013250475749373436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.22560976445674896 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot at eye level, rendered in a bright fantasy art style, begins with a slow, deliberate zoom-in to an enchanted forest clearing. The scene is bathed in a cool, magical glow, with silvery moonlight filtering through the dense canopy of ancient, gnarled trees and dotting the mossy ground with spots. In the center of the clearing, a sleek black cat with fur the color of polished obsidian and large, expressive amethyst eyes, its body curled protectively into a crescent, surrounded by a sleeping kitten with fuzzy gray fur. Clusters of bioluminescent mushrooms, pulsating shades of sapphire blue and ethereal violet, sprout from the fertile earth, casting a soft, otherworldly glow that illuminates the pair and nearby delicate, shimmering flowers, their petals seemingly spun from the moonlight. As the camera zooms in, tightening the frame to focus on the tender moment, the mother cat cautiously lowers her head, her whiskers twitching slightly, and gently nuzzles the kitten's cheek. Her long, graceful tail sways in a slow, hypnotic S-curve, her fingertips twitching contentedly, creating a serene and moving atmosphere of pure, protective love.", + "seed": 11115, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_2649_seed_11115_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017527350690215826, + "vlm_logprob": -0.017527350690215826, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.017527350690215826, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006875073071569204 + ], + "opposite_logprob_tokens": [ + -5.131875038146973 + ], + "logprob_reward": -0.006875073071569204, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.006875073071569204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08052507787942886, + 0.04723995178937912, + 0.04399843513965607, + 0.5022903084754944, + 0.11412336677312851, + 0.06765329092741013, + 0.05632719025015831, + 0.03090331144630909, + 0.01999899558722973, + 0.0369400791823864 + ], + "credit_map": [ + 0.015158232301473618, + 0.007914728485047817, + 0.017619097605347633, + 0.006819447968155146, + 0.007791211828589439, + 0.008564259856939316, + 0.006719754543155432, + 0.009938341565430164, + 0.004912123084068298, + 0.004252508282661438, + 0.00871005468070507, + 0.0052345627918839455, + 0.005694039165973663, + 0.007011068519204855, + 0.006160613149404526, + 0.005264974199235439, + 0.007733489852398634, + 0.0043436200357973576, + 0.005045411176979542, + 0.007804510183632374, + 0.003926794044673443, + 0.005877388175576925, + 0.004959092941135168, + 0.0043081240728497505, + 0.1213030219078064, + 0.028728395700454712, + 0.036731742322444916, + 0.039479486644268036, + 0.03194345161318779, + 0.19341561198234558, + 0.016540678218007088, + 0.03414793685078621, + 0.026411017403006554, + 0.012407206930220127, + 0.02014858089387417, + 0.007865817286074162, + 0.010881817899644375, + 0.016194768249988556, + 0.007169580087065697, + 0.013044562190771103, + 0.00785181112587452, + 0.006834452506154776, + 0.013551215641200542, + 0.005845045205205679, + 0.007986596785485744, + 0.00988007616251707, + 0.006043230649083853, + 0.009660857729613781, + 0.008018012158572674, + 0.005073137581348419, + 0.012491686269640923, + 0.003906505648046732, + 0.005324817728251219, + 0.006930721458047628, + 0.004774932283908129, + 0.009807370603084564, + 0.009411961771547794, + 0.010409751906991005, + 0.0021168270613998175, + 0.0014935459475964308, + 0.0023041691165417433, + 0.002592136850580573, + 0.0013326292391866446, + 0.001242287689819932, + 0.003282711375504732, + 0.001970185898244381, + 0.00454867584630847, + 0.0012966213980689645, + 0.0014821036020293832, + 0.003853047266602516, + 0.0015497503336519003, + 0.002015896840021014, + 0.00956185907125473, + 0.002338164485991001, + 0.004722002428025007, + 0.0019479147158563137, + 0.0018880865536630154, + 0.008934779092669487, + 0.005373358726501465, + 0.0021739136427640915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.006875073071569204 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864903163164854 + ], + "opposite_logprob_tokens": [ + -4.758648872375488 + ], + "logprob_reward": -0.00864903163164854, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864903163164854, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14975014328956604, + 0.08364617824554443, + 0.11861176788806915, + 0.07779047638177872, + 0.15475954115390778, + 0.0714188888669014, + 0.11524925380945206, + 0.08130523562431335, + 0.09715098142623901, + 0.05031752586364746 + ], + "credit_map": [ + 0.019358614459633827, + 0.015508043579757214, + 0.020406058058142662, + 0.015161246992647648, + 0.01413897518068552, + 0.022276179865002632, + 0.016456153243780136, + 0.026444876566529274, + 0.010655767284333706, + 0.00925176590681076, + 0.008583311922848225, + 0.006233501713722944, + 0.00790454726666212, + 0.024971207603812218, + 0.009769614785909653, + 0.006276464555412531, + 0.019027017056941986, + 0.00976163987070322, + 0.014067109674215317, + 0.022566786035895348, + 0.010843377560377121, + 0.015038562007248402, + 0.014912799932062626, + 0.012394470162689686, + 0.009394073858857155, + 0.008753732778131962, + 0.012704570777714252, + 0.006335295736789703, + 0.007469083648175001, + 0.015230078250169754, + 0.005110488273203373, + 0.01279315073043108, + 0.02531830221414566, + 0.018694087862968445, + 0.01925053261220455, + 0.02178122289478779, + 0.017184888944029808, + 0.01839275285601616, + 0.02001386694610119, + 0.014123894274234772, + 0.009754045866429806, + 0.007130427751690149, + 0.011719759553670883, + 0.005774897523224354, + 0.01066403929144144, + 0.012246090918779373, + 0.007068556267768145, + 0.0070610749535262585, + 0.016332658007740974, + 0.014571447856724262, + 0.027067597955465317, + 0.009068521670997143, + 0.01404508575797081, + 0.013301088474690914, + 0.009485184215009212, + 0.011377671733498573, + 0.015192670747637749, + 0.02318236418068409, + 0.0076949140056967735, + 0.006573698949068785, + 0.009640823118388653, + 0.008096275851130486, + 0.006094141397625208, + 0.0048303487710654736, + 0.018400857225060463, + 0.009492731653153896, + 0.012415452860295773, + 0.00841407012194395, + 0.010143478401005268, + 0.013113859109580517, + 0.01457594707608223, + 0.010594586841762066, + 0.010815036483108997, + 0.005218188744038343, + 0.006604338996112347, + 0.0028497532475739717, + 0.0037787032779306173, + 0.012323323637247086, + 0.004287390038371086, + 0.00444078492000699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.00864903163164854 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369740769267082 + ], + "opposite_logprob_tokens": [ + -3.4086973667144775 + ], + "logprob_reward": -0.03369740769267082, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369740769267082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10668065398931503, + 0.08237507194280624, + 0.052716534584760666, + 0.25351881980895996, + 0.18063007295131683, + 0.049496352672576904, + 0.04842553287744522, + 0.11868692934513092, + 0.07287462055683136, + 0.03459544479846954 + ], + "credit_map": [ + 0.022512024268507957, + 0.010976847261190414, + 0.024629740044474602, + 0.01083110086619854, + 0.007687931880354881, + 0.010808766819536686, + 0.008739721029996872, + 0.010494520887732506, + 0.012584385462105274, + 0.006535794120281935, + 0.011971660889685154, + 0.00707289157435298, + 0.009543301537632942, + 0.01912892423570156, + 0.008012487553060055, + 0.007525627966970205, + 0.007684266194701195, + 0.004057057201862335, + 0.011391052976250648, + 0.007611501030623913, + 0.004739466588944197, + 0.005758718121796846, + 0.005030693486332893, + 0.006443776655942202, + 0.06098927929997444, + 0.014553967863321304, + 0.01980193518102169, + 0.019844576716423035, + 0.016347937285900116, + 0.09311710298061371, + 0.010086877271533012, + 0.018777139484882355, + 0.028052954003214836, + 0.018316349014639854, + 0.023645896464586258, + 0.027434242889285088, + 0.0175933875143528, + 0.018953269347548485, + 0.027667468413710594, + 0.018966494128108025, + 0.006516723893582821, + 0.005675408523529768, + 0.007962321862578392, + 0.004938209429383278, + 0.006537713576108217, + 0.008697070181369781, + 0.004568743519484997, + 0.004600159358233213, + 0.005849191918969154, + 0.007379337213933468, + 0.011814289726316929, + 0.003617004258558154, + 0.005457962863147259, + 0.005715781822800636, + 0.0037381164729595184, + 0.004853845573961735, + 0.033052124083042145, + 0.034529753029346466, + 0.009619876742362976, + 0.00804599467664957, + 0.008605871349573135, + 0.010016153566539288, + 0.009092611260712147, + 0.005724532064050436, + 0.009290610440075397, + 0.007997088134288788, + 0.011471421457827091, + 0.004568070638924837, + 0.005404012277722359, + 0.014394535683095455, + 0.006881411653012037, + 0.012867467477917671, + 0.005331877153366804, + 0.002900591818615794, + 0.004492043983191252, + 0.0029105881694704294, + 0.002987136133015156, + 0.0075952401384711266, + 0.005461385007947683, + 0.0029165814630687237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.03369740769267082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635086007416248 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008635086007416248, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635086007416248, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08655014634132385, + 0.12077852338552475, + 0.0876353308558464, + 0.1587994247674942, + 0.19587574899196625, + 0.07024955749511719, + 0.0470135323703289, + 0.07805399596691132, + 0.05380748584866524, + 0.10123622417449951 + ], + "credit_map": [ + 0.016396092250943184, + 0.00913525652140379, + 0.01808476448059082, + 0.009400194510817528, + 0.005561936646699905, + 0.009212362580001354, + 0.008299951441586018, + 0.010459587909281254, + 0.01345848198980093, + 0.011685888282954693, + 0.02047673799097538, + 0.012756990268826485, + 0.014600328169763088, + 0.012877359054982662, + 0.015757378190755844, + 0.019165361300110817, + 0.01194408256560564, + 0.006637818645685911, + 0.011059435084462166, + 0.019620222970843315, + 0.007606318220496178, + 0.010396808385848999, + 0.01212247833609581, + 0.00824817456305027, + 0.03445262089371681, + 0.009887787513434887, + 0.012758223339915276, + 0.013719301670789719, + 0.010078269988298416, + 0.05991160124540329, + 0.00578662334010005, + 0.012205015867948532, + 0.026018451899290085, + 0.018830951303243637, + 0.04413721337914467, + 0.0156411025673151, + 0.01576535776257515, + 0.030475273728370667, + 0.013478864915668964, + 0.031528543680906296, + 0.010500219650566578, + 0.007411045953631401, + 0.011699605733156204, + 0.005968879442662001, + 0.007043993100523949, + 0.012782586738467216, + 0.005843242164701223, + 0.008999988436698914, + 0.0067053744569420815, + 0.004546457435935736, + 0.006464123725891113, + 0.0033782985992729664, + 0.004167977254837751, + 0.007445537019520998, + 0.0045100306160748005, + 0.009795737452805042, + 0.020643513649702072, + 0.02552316151559353, + 0.005893259774893522, + 0.004122959915548563, + 0.005962735507637262, + 0.00631377100944519, + 0.00441545806825161, + 0.005179134663194418, + 0.008114808239042759, + 0.004483995959162712, + 0.008306585252285004, + 0.004349098075181246, + 0.004582727327942848, + 0.011317398399114609, + 0.004285164177417755, + 0.00836771260946989, + 0.027978578582406044, + 0.006630091927945614, + 0.009956547059118748, + 0.0051917871460318565, + 0.004548602271825075, + 0.027876228094100952, + 0.013303354382514954, + 0.005751032382249832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 1 + ], + "selected_reward": -0.008635086007416248 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029780155047774315 + ], + "opposite_logprob_tokens": [ + -3.529780149459839 + ], + "logprob_reward": -0.029780155047774315, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029780155047774315, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06870467960834503, + 0.1246747076511383, + 0.08329979330301285, + 0.19683991372585297, + 0.08130884170532227, + 0.11603191494941711, + 0.10861998796463013, + 0.05054774507880211, + 0.047148749232292175, + 0.12282370030879974 + ], + "credit_map": [ + 0.007574241142719984, + 0.00855542253702879, + 0.009192117489874363, + 0.008329825475811958, + 0.006683498155325651, + 0.00968770682811737, + 0.0073029883205890656, + 0.0113788777962327, + 0.013704045675694942, + 0.011605266481637955, + 0.01994025893509388, + 0.01023741066455841, + 0.019090035930275917, + 0.02229173295199871, + 0.014470349065959454, + 0.013335604220628738, + 0.008551420643925667, + 0.009166279807686806, + 0.013141190633177757, + 0.015264956280589104, + 0.010492091998457909, + 0.007781120017170906, + 0.010135318152606487, + 0.008767412975430489, + 0.05432937666773796, + 0.010566215962171555, + 0.018951648846268654, + 0.012948596850037575, + 0.014298886992037296, + 0.06277710199356079, + 0.007740796077996492, + 0.015227271243929863, + 0.009039172902703285, + 0.009271213784813881, + 0.008615388534963131, + 0.00906485877931118, + 0.011082720942795277, + 0.015213475562632084, + 0.009788493625819683, + 0.009233520366251469, + 0.01759655587375164, + 0.01820174790918827, + 0.013090921565890312, + 0.00921512208878994, + 0.01854671910405159, + 0.013191488571465015, + 0.010703466832637787, + 0.01548589114099741, + 0.01678074151277542, + 0.011573848314583302, + 0.017409225925803185, + 0.009394358843564987, + 0.019292118027806282, + 0.011767814867198467, + 0.01132366806268692, + 0.011078204028308392, + 0.012804516591131687, + 0.015777425840497017, + 0.003993476741015911, + 0.0035401470959186554, + 0.0044752927497029305, + 0.00422329269349575, + 0.0028382267337292433, + 0.0028953631408512592, + 0.007490397430956364, + 0.008128711953759193, + 0.006162711419165134, + 0.0039270296692848206, + 0.0071257962845265865, + 0.007227577734738588, + 0.003799332305788994, + 0.003287192666903138, + 0.02363506704568863, + 0.010073366574943066, + 0.020705273374915123, + 0.013934305869042873, + 0.00709492526948452, + 0.030522050336003304, + 0.009107680059969425, + 0.007751033175736666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 9 + ], + "selected_reward": -0.029780155047774315 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot at eye level, rendered in a bright fantasy art style, begins with a slow, deliberate zoom-in to an enchanted forest clearing. The scene is bathed in a cool, magical glow, with silvery moonlight filtering through the dense canopy of ancient, gnarled trees and dotting the mossy ground with spots. In the center of the clearing, a sleek black cat with fur the color of polished obsidian and large, expressive amethyst eyes, its body curled protectively into a crescent, surrounded by a sleeping kitten with fuzzy gray fur. Clusters of bioluminescent mushrooms, pulsating shades of sapphire blue and ethereal violet, sprout from the fertile earth, casting a soft, otherworldly glow that illuminates the pair and nearby delicate, shimmering flowers, their petals seemingly spun from the moonlight. As the camera zooms in, tightening the frame to focus on the tender moment, the mother cat cautiously lowers her head, her whiskers twitching slightly, and gently nuzzles the kitten's cheek. Her long, graceful tail sways in a slow, hypnotic S-curve, her fingertips twitching contentedly, creating a serene and moving atmosphere of pure, protective love.", + "seed": 11116, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_4/index_2649_seed_11116_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04728490319103003, + "vlm_logprob": -0.04728490319103003, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.04728490319103003, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0285510066896677 + ], + "opposite_logprob_tokens": [ + -3.6535511016845703 + ], + "logprob_reward": -0.0285510066896677, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.0285510066896677, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4168674051761627, + 0.16744795441627502, + 0.03982778266072273, + 0.15257596969604492, + 0.03803092613816261, + 0.05638852342963219, + 0.040288373827934265, + 0.019468259066343307, + 0.05683653801679611, + 0.01226817723363638 + ], + "credit_map": [ + 0.036035407334566116, + 0.060725074261426926, + 0.04807000234723091, + 0.038498517125844955, + 0.07173268496990204, + 0.0661897286772728, + 0.03188847005367279, + 0.06372757256031036, + 0.009327271021902561, + 0.010888158343732357, + 0.014248923398554325, + 0.06526143848896027, + 0.006619511637836695, + 0.010558784008026123, + 0.009636694565415382, + 0.04090718924999237, + 0.004393602721393108, + 0.0041361921466887, + 0.0036367452703416348, + 0.004828390199691057, + 0.004545426927506924, + 0.004560269881039858, + 0.003927997779101133, + 0.009799163788557053, + 0.018855687230825424, + 0.01071900688111782, + 0.01975269243121147, + 0.009295313619077206, + 0.020953739061951637, + 0.020809663459658623, + 0.025406284257769585, + 0.02678358554840088, + 0.004597342573106289, + 0.003531441558152437, + 0.0047688172198832035, + 0.002009584568440914, + 0.008735815063118935, + 0.005583405494689941, + 0.004955779295414686, + 0.0038487426936626434, + 0.005893047899007797, + 0.005522452294826508, + 0.004444321151822805, + 0.004041444510221481, + 0.011478432454168797, + 0.009458858519792557, + 0.005582220386713743, + 0.009967755526304245, + 0.004174290224909782, + 0.003680754452943802, + 0.004216471221297979, + 0.005897446069866419, + 0.006898574996739626, + 0.006478039547801018, + 0.004542413633316755, + 0.004400390665978193, + 0.002520886017009616, + 0.0022383045870810747, + 0.002043546177446842, + 0.0025206501595675945, + 0.0028759283013641834, + 0.002408072119578719, + 0.0020520531106740236, + 0.0028088223189115524, + 0.004556081257760525, + 0.009413395076990128, + 0.004424008075147867, + 0.008016999810934067, + 0.009200725704431534, + 0.005125278141349554, + 0.007182717323303223, + 0.008917339146137238, + 0.001324429758824408, + 0.0020085496362298727, + 0.001288481755182147, + 0.0015198623295873404, + 0.0016682787099853158, + 0.0016671039629727602, + 0.0012413566000759602, + 0.0015501161105930805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.0285510066896677 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081975772976875 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081975772976875, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081975772976875, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1667029708623886, + 0.12024838477373123, + 0.09159457683563232, + 0.12299729883670807, + 0.06529685109853745, + 0.04889577999711037, + 0.06085903197526932, + 0.04579741507768631, + 0.2440488487482071, + 0.03355881944298744 + ], + "credit_map": [ + 0.016167158260941505, + 0.03338036686182022, + 0.017640573903918266, + 0.029514454305171967, + 0.02363496460020542, + 0.016642319038510323, + 0.013263877481222153, + 0.0164592657238245, + 0.008621994405984879, + 0.012877591885626316, + 0.014323563314974308, + 0.03550993278622627, + 0.008129999972879887, + 0.011119942180812359, + 0.008473260328173637, + 0.021192098036408424, + 0.010088787414133549, + 0.011533107608556747, + 0.010365431196987629, + 0.01290189940482378, + 0.011965586803853512, + 0.010650335811078548, + 0.010168978944420815, + 0.013920443132519722, + 0.020445197820663452, + 0.006263704504817724, + 0.014789026230573654, + 0.006011242978274822, + 0.00731813907623291, + 0.010931077413260937, + 0.024253519251942635, + 0.0329853855073452, + 0.005623713601380587, + 0.003763670101761818, + 0.020033761858940125, + 0.004418487660586834, + 0.0042640394531190395, + 0.006798333488404751, + 0.01536654680967331, + 0.0050283027812838554, + 0.004815395921468735, + 0.0042414977215230465, + 0.006674112286418676, + 0.003810312831774354, + 0.011015462689101696, + 0.007737215142697096, + 0.004216533154249191, + 0.006385249551385641, + 0.006555645260959864, + 0.007889232598245144, + 0.006296204403042793, + 0.005932734813541174, + 0.008739857003092766, + 0.006403591949492693, + 0.0070723420940339565, + 0.011969425715506077, + 0.004068189766258001, + 0.003935716580599546, + 0.00479468610137701, + 0.0058820173144340515, + 0.009438907727599144, + 0.00562490988522768, + 0.005577282980084419, + 0.006475701462477446, + 0.01620474085211754, + 0.029997292906045914, + 0.021434037014842033, + 0.04426649585366249, + 0.03555869311094284, + 0.024552257731556892, + 0.03703158348798752, + 0.03500373288989067, + 0.003139672102406621, + 0.005943338852375746, + 0.0038508460856974125, + 0.004635966382920742, + 0.005165461916476488, + 0.004003336653113365, + 0.0027281197253614664, + 0.00409207446500659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.016081975772976875 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898581027984619 + ], + "opposite_logprob_tokens": [ + -2.5789856910705566 + ], + "logprob_reward": -0.07898581027984619, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898581027984619, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10232804715633392, + 0.08198224008083344, + 0.043707650154829025, + 0.14580614864826202, + 0.10100901871919632, + 0.21878156065940857, + 0.09450043737888336, + 0.03271125257015228, + 0.13295534253120422, + 0.04621834307909012 + ], + "credit_map": [ + 0.00990012101829052, + 0.020084720104932785, + 0.010197850875556469, + 0.015787625685334206, + 0.01739276759326458, + 0.009550631046295166, + 0.007837029173970222, + 0.011577295139431953, + 0.006328738760203123, + 0.012791824527084827, + 0.011604136787354946, + 0.01882823556661606, + 0.006678321864455938, + 0.009742890484631062, + 0.006887421943247318, + 0.009120662696659565, + 0.004623653367161751, + 0.005113100633025169, + 0.006730337161570787, + 0.0043978579342365265, + 0.0051845465786755085, + 0.0058457436971366405, + 0.004466267302632332, + 0.007346141152083874, + 0.018143825232982635, + 0.009278631769120693, + 0.0182937104254961, + 0.008860844187438488, + 0.009474003687500954, + 0.014082133769989014, + 0.030305612832307816, + 0.037367381155490875, + 0.010070139542222023, + 0.005033445078879595, + 0.03226293995976448, + 0.0056342170573771, + 0.009216250851750374, + 0.010081750340759754, + 0.021268093958497047, + 0.007442182395607233, + 0.029314719140529633, + 0.02292194589972496, + 0.01261263806372881, + 0.012297607026994228, + 0.060731083154678345, + 0.025871174409985542, + 0.021211447194218636, + 0.03382093831896782, + 0.009405797347426414, + 0.011399153620004654, + 0.008201755583286285, + 0.012323234230279922, + 0.013300000689923763, + 0.00852430984377861, + 0.013267568312585354, + 0.018078617751598358, + 0.0035992746707051992, + 0.003337060334160924, + 0.003289314918220043, + 0.004994898568838835, + 0.005211591254919767, + 0.003462033113464713, + 0.003778849495574832, + 0.005038228817284107, + 0.010038618929684162, + 0.02516600303351879, + 0.011281386017799377, + 0.01659606210887432, + 0.023023812100291252, + 0.012484378181397915, + 0.014760485850274563, + 0.01960458792746067, + 0.003973342012614012, + 0.008720777928829193, + 0.004758405964821577, + 0.007172972429543734, + 0.006079448387026787, + 0.005230160430073738, + 0.0046625882387161255, + 0.005620643496513367 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 8 + ], + "selected_reward": -0.07898581027984619 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012533470056951046 + ], + "opposite_logprob_tokens": [ + -4.387533664703369 + ], + "logprob_reward": -0.012533470056951046, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012533470056951046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2863578200340271, + 0.164768248796463, + 0.12175673246383667, + 0.22157713770866394, + 0.06171339377760887, + 0.030988426879048347, + 0.030085407197475433, + 0.02977636642754078, + 0.041232287883758545, + 0.011744103394448757 + ], + "credit_map": [ + 0.032050713896751404, + 0.047787170857191086, + 0.034663666039705276, + 0.027790620923042297, + 0.0374397374689579, + 0.041502684354782104, + 0.018031487241387367, + 0.04709175229072571, + 0.010457598604261875, + 0.00986393727362156, + 0.013759161345660686, + 0.06251011043787003, + 0.007245281711220741, + 0.010843870230019093, + 0.009645001031458378, + 0.04044327884912491, + 0.014082522131502628, + 0.014724130742251873, + 0.014673389494419098, + 0.01821853034198284, + 0.015060258097946644, + 0.016087787225842476, + 0.01453521940857172, + 0.014374894089996815, + 0.022860147058963776, + 0.01713499426841736, + 0.02105545438826084, + 0.012381723150610924, + 0.0333394818007946, + 0.02972617745399475, + 0.03947538882493973, + 0.045603781938552856, + 0.005213133525103331, + 0.0032439930364489555, + 0.019174810498952866, + 0.0035750900860875845, + 0.005107380915433168, + 0.00624769926071167, + 0.01395006850361824, + 0.005201222375035286, + 0.003108773846179247, + 0.0032013063319027424, + 0.004916244652122259, + 0.0022138601634651423, + 0.005920742638409138, + 0.004704764112830162, + 0.0031148656271398067, + 0.0038078706711530685, + 0.0028742081485688686, + 0.002548071090131998, + 0.003301538061350584, + 0.00580243207514286, + 0.002701470861211419, + 0.0037943124771118164, + 0.00397467939183116, + 0.00508869718760252, + 0.0034831769298762083, + 0.0034568393602967262, + 0.003470346098765731, + 0.004099453799426556, + 0.004739934578537941, + 0.0036881768610328436, + 0.003216812154278159, + 0.0036216273438185453, + 0.0033843321725726128, + 0.006926145404577255, + 0.003511535469442606, + 0.006062038708478212, + 0.0057582007721066475, + 0.003926001023501158, + 0.005337064620107412, + 0.006326969712972641, + 0.0012149014510214329, + 0.001349812955595553, + 0.0015377853997051716, + 0.0010810724925249815, + 0.0012566689401865005, + 0.0022866465151309967, + 0.0012944465270265937, + 0.0017227685311809182 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.012533470056951046 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in an enchanted forest clearing with ancient, gnarled trees and mossy ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is there a sleek black cat with fur the color of polished obsidian and large amethyst eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a sleeping kitten with fuzzy gray fur curled protectively near the mother cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are bioluminescent mushrooms in shades of sapphire blue and ethereal violet visible on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin with a slow zoom-in that tightens the frame to focus on the cats?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027225315570831 + ], + "opposite_logprob_tokens": [ + -2.3502721786499023 + ], + "logprob_reward": -0.10027225315570831, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027225315570831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3228083848953247, + 0.07394696772098541, + 0.10968627780675888, + 0.12201574444770813, + 0.08426908403635025, + 0.06250753253698349, + 0.0732106864452362, + 0.03439037874341011, + 0.10775711387395859, + 0.009407809004187584 + ], + "credit_map": [ + 0.02929612621665001, + 0.05523719638586044, + 0.032841138541698456, + 0.059785518795251846, + 0.04340190812945366, + 0.039531949907541275, + 0.02941611036658287, + 0.03329845145344734, + 0.007278247270733118, + 0.011192516423761845, + 0.011193668469786644, + 0.011018100194633007, + 0.007690091617405415, + 0.006959124933928251, + 0.009515277110040188, + 0.009099943563342094, + 0.010824139229953289, + 0.01806071028113365, + 0.012182220816612244, + 0.01442046370357275, + 0.01644931547343731, + 0.011459712870419025, + 0.012143888510763645, + 0.014145826920866966, + 0.011404833756387234, + 0.012195213697850704, + 0.024890845641493797, + 0.011717470362782478, + 0.020548976957798004, + 0.01456481497734785, + 0.013684043660759926, + 0.013009539805352688, + 0.007556285243481398, + 0.0048546153120696545, + 0.01641257293522358, + 0.005229892209172249, + 0.009994237683713436, + 0.015516040846705437, + 0.015070612542331219, + 0.009634827263653278, + 0.006029533687978983, + 0.01154356636106968, + 0.0070163775235414505, + 0.007307504769414663, + 0.007239691913127899, + 0.0074814697727561, + 0.0069969031028449535, + 0.008892487734556198, + 0.005528939422219992, + 0.00971186626702547, + 0.006953391712158918, + 0.015395919792354107, + 0.006900254171341658, + 0.007920893840491772, + 0.007887298241257668, + 0.012912124395370483, + 0.0031078089959919453, + 0.004195594694465399, + 0.0032528317533433437, + 0.004138004034757614, + 0.0054952166974544525, + 0.004339388106018305, + 0.004227485973387957, + 0.0056340512819588184, + 0.00797220878303051, + 0.017843367531895638, + 0.008004622533917427, + 0.017620448023080826, + 0.014482320286333561, + 0.01020759716629982, + 0.01389975007623434, + 0.017726797610521317, + 0.0007854768773540854, + 0.0017264563357457519, + 0.0010015069274231791, + 0.0010035247541964054, + 0.0012930603697896004, + 0.001029000268317759, + 0.0009678800706751645, + 0.0016009033424779773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.10027225315570831 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_tensor([6362, 6362])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_tensor([6362, 6362])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..495b81ed32c134a6add661632b16467f3d92f39b --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_tensor([6362, 6362])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle captures a nearly perfect cube of ice resting on the dark, polished surface of a black slate countertop. A single, cool, overhead light illuminates the scene, creating sharp, shimmering highlights on the ice's damp surface, reflected in the slowly forming pool of meltwater beneath. The translucent ice reveals a network of tiny internal bubbles, frozen in its crystalline structure. Suddenly, a fine crack appears, originating at the core and moving swiftly and decisively across the cube's surface in a jagged, lightning-fast path. With a silent, violent pop, internal thermal stresses reach a peak, fracturing the cube into two separate, unequal fragments. The force of the fracture abruptly separates the two pieces; they slide briefly across the smooth, damp countertop, their newly exposed, rough-textured fractured surfaces catching the light, before falling completely still and separating. The entire event unfolds with stark, dramatic clarity, creating an atmosphere of intense focus, isolating this small, yet devastating moment against the dark, minimalist backdrop.", + "seed": 31588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_6362_seed_31588_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03459073938429356, + "vlm_logprob": -0.03459073938429356, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.03459073938429356, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.022057093679904938 + ], + "opposite_logprob_tokens": [ + -5.147057056427002 + ], + "logprob_reward": -0.022057093679904938, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.022057093679904938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21431894600391388, + 0.045954011380672455, + 0.06263551861047745, + 0.04306058958172798, + 0.02326079085469246, + 0.19826626777648926, + 0.04698530212044716, + 0.24648910760879517, + 0.0838087797164917, + 0.0352206788957119 + ], + "credit_map": [ + 0.06103174388408661, + 0.013745526783168316, + 0.022746000438928604, + 0.012819535098969936, + 0.014480998739600182, + 0.06417979300022125, + 0.008144591003656387, + 0.017170747742056847, + 0.006378151476383209, + 0.002925560576841235, + 0.009541278705000877, + 0.0032204806338995695, + 0.004311594180762768, + 0.007171009201556444, + 0.003446279326453805, + 0.00895965937525034, + 0.010206126607954502, + 0.005295048467814922, + 0.010534761473536491, + 0.005685423035174608, + 0.00484270928427577, + 0.010858885012567043, + 0.004646453075110912, + 0.010566112585365772, + 0.007908659987151623, + 0.005357337184250355, + 0.008021509274840355, + 0.003973652608692646, + 0.0039048301987349987, + 0.004275010898709297, + 0.003485135966911912, + 0.00613445695489645, + 0.0032979780808091164, + 0.00296954158693552, + 0.0028819392900913954, + 0.00253149657510221, + 0.003571505192667246, + 0.0027604217175394297, + 0.0024201984051615, + 0.002827713033184409, + 0.07086917757987976, + 0.013386821374297142, + 0.031796861439943314, + 0.006096820812672377, + 0.051471613347530365, + 0.010545982047915459, + 0.008673490956425667, + 0.005425502546131611, + 0.005826118402183056, + 0.0058839148841798306, + 0.00602743960916996, + 0.004288730211555958, + 0.007942886091768742, + 0.007022614125162363, + 0.00487168412655592, + 0.0051219165325164795, + 0.016435395926237106, + 0.08817926049232483, + 0.012862361036241055, + 0.00882510095834732, + 0.02121419832110405, + 0.08311045914888382, + 0.007554089650511742, + 0.008308228105306625, + 0.006879106163978577, + 0.013348469510674477, + 0.014004320837557316, + 0.010044815950095654, + 0.011301379650831223, + 0.010392608121037483, + 0.011060803197324276, + 0.006777276284992695, + 0.005729250609874725, + 0.004463476128876209, + 0.006142947357147932, + 0.003258896991610527, + 0.005915095563977957, + 0.0035211793147027493, + 0.0030277392361313105, + 0.003162094159051776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.022057093679904938 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016093941405415535 + ], + "opposite_logprob_tokens": [ + -4.141093730926514 + ], + "logprob_reward": -0.016093941405415535, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016093941405415535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09071145951747894, + 0.12633225321769714, + 0.1336226910352707, + 0.04331555217504501, + 0.03281714394688606, + 0.038168612867593765, + 0.02033872902393341, + 0.3975820243358612, + 0.07956565171480179, + 0.03754587471485138 + ], + "credit_map": [ + 0.025990281254053116, + 0.0054810442961752415, + 0.008190173655748367, + 0.004451229237020016, + 0.005943777039647102, + 0.02772519737482071, + 0.005411017220467329, + 0.007518742699176073, + 0.035069603472948074, + 0.005783032160252333, + 0.017079442739486694, + 0.005393346771597862, + 0.0053076231852173805, + 0.040800657123327255, + 0.004127759952098131, + 0.012770785950124264, + 0.03828806430101395, + 0.005436806473881006, + 0.010380683466792107, + 0.005574494134634733, + 0.005086416844278574, + 0.043746571987867355, + 0.0047408295795321465, + 0.020368827506899834, + 0.008189613930881023, + 0.005523103289306164, + 0.008500823751091957, + 0.003979208879172802, + 0.0038429750129580498, + 0.004181305877864361, + 0.00347375450655818, + 0.00562477158382535, + 0.00505861034616828, + 0.00516517274081707, + 0.004965654574334621, + 0.0029536690562963486, + 0.005006501916795969, + 0.003014825051650405, + 0.002826408017426729, + 0.0038263038732111454, + 0.014908815734088421, + 0.002246302319690585, + 0.009285901673138142, + 0.001653673592954874, + 0.004571845754981041, + 0.002309004543349147, + 0.0015908342320472002, + 0.001602238160558045, + 0.0025317880790680647, + 0.003295030677691102, + 0.0027586189098656178, + 0.001398454885929823, + 0.003290057647973299, + 0.0031271290499716997, + 0.0020192654337733984, + 0.0019183863187208772, + 0.02900487743318081, + 0.14301089942455292, + 0.021394122391939163, + 0.015473867766559124, + 0.033844951540231705, + 0.13119058310985565, + 0.011534133926033974, + 0.012128639966249466, + 0.005853326991200447, + 0.009201906621456146, + 0.01004093885421753, + 0.009951972402632236, + 0.01226477324962616, + 0.014567377977073193, + 0.010107672773301601, + 0.007577689364552498, + 0.0034755757078528404, + 0.005000157747417688, + 0.004674296826124191, + 0.004569664131850004, + 0.006232109852135181, + 0.005672204773873091, + 0.004690179601311684, + 0.0032316893339157104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 1 + ], + "selected_reward": -0.016093941405415535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906083554029465 + ], + "opposite_logprob_tokens": [ + -2.4640607833862305 + ], + "logprob_reward": -0.08906083554029465, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906083554029465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20634056627750397, + 0.053171295672655106, + 0.048746488988399506, + 0.03527548164129257, + 0.03924913331866264, + 0.3804328739643097, + 0.033912308514118195, + 0.0545479916036129, + 0.11464052647352219, + 0.03368344530463219 + ], + "credit_map": [ + 0.059274837374687195, + 0.01324627734720707, + 0.01936393603682518, + 0.009019425138831139, + 0.010191108100116253, + 0.06416655331850052, + 0.0108224181458354, + 0.020255979150533676, + 0.013437354937195778, + 0.002785792574286461, + 0.00550088519230485, + 0.0032526522409170866, + 0.004086047876626253, + 0.013430947437882423, + 0.003508526599034667, + 0.007169080898165703, + 0.008977592922747135, + 0.00417353818193078, + 0.005537114106118679, + 0.004134646151214838, + 0.0060952408239245415, + 0.009643463417887688, + 0.004492714535444975, + 0.005692168138921261, + 0.005943541415035725, + 0.005728548858314753, + 0.004186347126960754, + 0.0033429914619773626, + 0.00419944804161787, + 0.004395351745188236, + 0.0035787613596767187, + 0.0039004841819405556, + 0.0056624095886945724, + 0.004633251577615738, + 0.00549020990729332, + 0.003842139383777976, + 0.0074483975768089294, + 0.0037838793359696865, + 0.004311195574700832, + 0.004077643156051636, + 0.11301521211862564, + 0.03326893225312233, + 0.04001500457525253, + 0.013217948377132416, + 0.1279374212026596, + 0.021236497908830643, + 0.02048434130847454, + 0.011257452890276909, + 0.0039162966422736645, + 0.0039054404478520155, + 0.004118939395993948, + 0.002441191580146551, + 0.005427854601293802, + 0.006571851670742035, + 0.0030005830340087414, + 0.0045301467180252075, + 0.004433267284184694, + 0.014516560360789299, + 0.003860217286273837, + 0.0025800252333283424, + 0.0065038613975048065, + 0.017153745517134666, + 0.003223867155611515, + 0.0022764336317777634, + 0.013120325282216072, + 0.014339434914290905, + 0.017383666709065437, + 0.013599538244307041, + 0.01689562201499939, + 0.013415957801043987, + 0.01442809123545885, + 0.011457856744527817, + 0.0031017547007650137, + 0.004416125826537609, + 0.004671134985983372, + 0.002934713615104556, + 0.005001997109502554, + 0.006289759650826454, + 0.003448452102020383, + 0.0038194984663277864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.08906083554029465 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007643263787031174 + ], + "opposite_logprob_tokens": [ + -4.882643222808838 + ], + "logprob_reward": -0.007643263787031174, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007643263787031174, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11980999261140823, + 0.04315126687288284, + 0.03667447343468666, + 0.020459096878767014, + 0.019330235198140144, + 0.23705555498600006, + 0.08191905170679092, + 0.3340948224067688, + 0.08106277137994766, + 0.026442736387252808 + ], + "credit_map": [ + 0.03816532716155052, + 0.007223628461360931, + 0.0076147569343447685, + 0.005548844579607248, + 0.008948754519224167, + 0.04045848175883293, + 0.005127027630805969, + 0.006723162718117237, + 0.012727877125144005, + 0.002286784118041396, + 0.004368038848042488, + 0.001990520628169179, + 0.002390161855146289, + 0.013912186026573181, + 0.0016367110656574368, + 0.0038389889523386955, + 0.006961885839700699, + 0.0028402903117239475, + 0.003637002781033516, + 0.002896493300795555, + 0.003942757844924927, + 0.007237582467496395, + 0.002981226658448577, + 0.006177234463393688, + 0.004053774289786816, + 0.0022078785113990307, + 0.0033411637414246798, + 0.001751623465679586, + 0.001680614659562707, + 0.0023784686345607042, + 0.0013801633613184094, + 0.003665411612018943, + 0.002447588136419654, + 0.002587730996310711, + 0.0025750286877155304, + 0.0020200686994940042, + 0.0031037582084536552, + 0.0017065684078261256, + 0.0021579584572464228, + 0.0027315348852425814, + 0.06968823820352554, + 0.020067185163497925, + 0.02326376736164093, + 0.008347395807504654, + 0.08100759983062744, + 0.013727115467190742, + 0.013332167640328407, + 0.007622084580361843, + 0.009229011833667755, + 0.00922642182558775, + 0.011583476327359676, + 0.00597213814035058, + 0.016858961433172226, + 0.01052525918930769, + 0.008074731566011906, + 0.010449046269059181, + 0.022912275046110153, + 0.11891797930002213, + 0.017787618562579155, + 0.012388455681502819, + 0.02999446541070938, + 0.11074435710906982, + 0.010507212020456791, + 0.010842456482350826, + 0.0071034119464457035, + 0.01157352328300476, + 0.013453537598252296, + 0.009750866331160069, + 0.011179988272488117, + 0.011133638210594654, + 0.010481199249625206, + 0.0063866060227155685, + 0.003120260778814554, + 0.0029534550849348307, + 0.003779339836910367, + 0.0036237328313291073, + 0.003884947160258889, + 0.0030235438607633114, + 0.003333603497594595, + 0.0027238535694777966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 0 + ], + "selected_reward": -0.007643263787031174 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809856250882149 + ], + "opposite_logprob_tokens": [ + -3.2880985736846924 + ], + "logprob_reward": -0.03809856250882149, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809856250882149, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19930356740951538, + 0.05421051010489464, + 0.11351339519023895, + 0.021142931655049324, + 0.018510395660996437, + 0.4139237701892853, + 0.06087799742817879, + 0.02193594165146351, + 0.057474419474601746, + 0.039107028394937515 + ], + "credit_map": [ + 0.056069549173116684, + 0.011621258221566677, + 0.03228197991847992, + 0.008913600817322731, + 0.009857691824436188, + 0.055634308606386185, + 0.008956442587077618, + 0.015968753024935722, + 0.011898466385900974, + 0.00294350809417665, + 0.010671300813555717, + 0.002466876059770584, + 0.0033367942087352276, + 0.014140387065708637, + 0.0024420116096735, + 0.0063111670315265656, + 0.03062760829925537, + 0.0043654427863657475, + 0.009856217540800571, + 0.006505122408270836, + 0.004398433957248926, + 0.034886736422777176, + 0.005132324527949095, + 0.01774151809513569, + 0.0034424366895109415, + 0.002937609562650323, + 0.003917367197573185, + 0.0015413021901622415, + 0.002564782975241542, + 0.002972619840875268, + 0.0013618809171020985, + 0.002404934959486127, + 0.0025867451913654804, + 0.0018562666373327374, + 0.0028240440879017115, + 0.0019527202239260077, + 0.0027849215548485518, + 0.001935814623720944, + 0.002151837572455406, + 0.0024180468171834946, + 0.12896095216274261, + 0.031569745391607285, + 0.04755663126707077, + 0.015452701598405838, + 0.1312721222639084, + 0.023700788617134094, + 0.02225942723453045, + 0.013151442632079124, + 0.009375556372106075, + 0.00920738372951746, + 0.0070548285730183125, + 0.004528391174972057, + 0.009500270709395409, + 0.007524057291448116, + 0.0071618277579545975, + 0.006525687873363495, + 0.0026688033249229193, + 0.003625277429819107, + 0.002499500522390008, + 0.00144517223816365, + 0.0031889949459582567, + 0.005359070375561714, + 0.001814494258724153, + 0.0013346290215849876, + 0.005760573782026768, + 0.006278064101934433, + 0.007511948235332966, + 0.006174834445118904, + 0.00679012993350625, + 0.012021554633975029, + 0.004543873947113752, + 0.008393445052206516, + 0.00427156500518322, + 0.004591826815158129, + 0.005393297411501408, + 0.004435035400092602, + 0.006267128046602011, + 0.005786655005067587, + 0.0042587462812662125, + 0.004102778155356646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 2 + ], + "selected_reward": -0.03809856250882149 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle captures a nearly perfect cube of ice resting on the dark, polished surface of a black slate countertop. A single, cool, overhead light illuminates the scene, creating sharp, shimmering highlights on the ice's damp surface, reflected in the slowly forming pool of meltwater beneath. The translucent ice reveals a network of tiny internal bubbles, frozen in its crystalline structure. Suddenly, a fine crack appears, originating at the core and moving swiftly and decisively across the cube's surface in a jagged, lightning-fast path. With a silent, violent pop, internal thermal stresses reach a peak, fracturing the cube into two separate, unequal fragments. The force of the fracture abruptly separates the two pieces; they slide briefly across the smooth, damp countertop, their newly exposed, rough-textured fractured surfaces catching the light, before falling completely still and separating. The entire event unfolds with stark, dramatic clarity, creating an atmosphere of intense focus, isolating this small, yet devastating moment against the dark, minimalist backdrop.", + "seed": 31589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_6362_seed_31589_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.037665099184960124, + "vlm_logprob": -0.037665099184960124, + "vlm_yes_no_margin": 3.95, + "vlm_token_credit": -0.037665099184960124, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030451403930783272 + ], + "opposite_logprob_tokens": [ + -3.9054512977600098 + ], + "logprob_reward": -0.030451403930783272, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030451403930783272, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.117862269282341, + 0.041439320892095566, + 0.29510167241096497, + 0.05290662497282028, + 0.03335132822394371, + 0.039735130965709686, + 0.35732969641685486, + 0.017412109300494194, + 0.030485868453979492, + 0.014375921338796616 + ], + "credit_map": [ + 0.018420016393065453, + 0.009395119734108448, + 0.023207494989037514, + 0.011057700030505657, + 0.008375310339033604, + 0.01793765090405941, + 0.012583023868501186, + 0.016885947436094284, + 0.006170616485178471, + 0.0031119592022150755, + 0.009064983576536179, + 0.00436311773955822, + 0.003340360941365361, + 0.005959331523627043, + 0.00286740530282259, + 0.006561544723808765, + 0.04581057280302048, + 0.006693935487419367, + 0.10810652375221252, + 0.008753888309001923, + 0.006520335096865892, + 0.08855048567056656, + 0.0033086531329900026, + 0.027357278391718864, + 0.014432840049266815, + 0.0033676265738904476, + 0.014631431549787521, + 0.0036227675154805183, + 0.005718558095395565, + 0.0030503009911626577, + 0.0034128595143556595, + 0.00467023765668273, + 0.009223364293575287, + 0.0020488461013883352, + 0.010522532276809216, + 0.0023611390497535467, + 0.0027073752135038376, + 0.0016704895533621311, + 0.0025137353222817183, + 0.0023038466461002827, + 0.006144025828689337, + 0.0031086718663573265, + 0.004641435574740171, + 0.0032422984950244427, + 0.004217943176627159, + 0.0037713043857365847, + 0.0072050211019814014, + 0.0074044279754161835, + 0.1472596526145935, + 0.015054365620017052, + 0.1506107598543167, + 0.010247386991977692, + 0.00917551014572382, + 0.010723024606704712, + 0.005835918243974447, + 0.008423065766692162, + 0.0041121793910861015, + 0.002143880818039179, + 0.00445813313126564, + 0.001391654834151268, + 0.0013934342423453927, + 0.0012718960642814636, + 0.0012440934078767896, + 0.0013968385756015778, + 0.006499451585114002, + 0.007792499847710133, + 0.0033363522961735725, + 0.0031154623720794916, + 0.0035337498411536217, + 0.0021706102415919304, + 0.0018347603036090732, + 0.0022029816173017025, + 0.001901065232232213, + 0.0016234719660133123, + 0.0017324541695415974, + 0.00166771758813411, + 0.003336218185722828, + 0.0012848185142502189, + 0.0015349152963608503, + 0.0012952600372955203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.030451403930783272 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268380511552095 + ], + "opposite_logprob_tokens": [ + -5.255268573760986 + ], + "logprob_reward": -0.005268380511552095, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268380511552095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08436845988035202, + 0.030494919046759605, + 0.12339354306459427, + 0.05864409729838371, + 0.058618586510419846, + 0.04447023198008537, + 0.4874168038368225, + 0.0938824713230133, + 0.011173955164849758, + 0.007536949124187231 + ], + "credit_map": [ + 0.010142670013010502, + 0.005017153453081846, + 0.020944751799106598, + 0.006815228145569563, + 0.0033600220922380686, + 0.015714505687355995, + 0.004693695344030857, + 0.017680438235402107, + 0.0032743741758167744, + 0.002374825533479452, + 0.007048890460282564, + 0.003032078966498375, + 0.003140741493552923, + 0.004699425306171179, + 0.0021044970490038395, + 0.004820086993277073, + 0.018825314939022064, + 0.0032451138831675053, + 0.04385359212756157, + 0.0039936453104019165, + 0.0031193927861750126, + 0.036811463534832, + 0.0019695525988936424, + 0.01157546229660511, + 0.00905359536409378, + 0.004513595253229141, + 0.01161538902670145, + 0.009171578101813793, + 0.006367126479744911, + 0.005435307044535875, + 0.006738139316439629, + 0.005749364849179983, + 0.019707433879375458, + 0.0027483946178108454, + 0.022447405382990837, + 0.0026215084362775087, + 0.0025297868996858597, + 0.0031165212858468294, + 0.0023106331937015057, + 0.0031368988566100597, + 0.010033825412392616, + 0.004667569883167744, + 0.008816158398985863, + 0.002711895387619734, + 0.005033982917666435, + 0.003306994214653969, + 0.005718419328331947, + 0.004181388299912214, + 0.1941818743944168, + 0.021220428869128227, + 0.19625364243984222, + 0.014166554436087608, + 0.013402065262198448, + 0.015336564742028713, + 0.011259987950325012, + 0.02159568853676319, + 0.034288935363292694, + 0.005950741469860077, + 0.03422369062900543, + 0.004083577077835798, + 0.003347117453813553, + 0.004463689401745796, + 0.002841423498466611, + 0.004683293402194977, + 0.0018667051335796714, + 0.0019160413648933172, + 0.0015254153404384851, + 0.0012585787335410714, + 0.0012474247487261891, + 0.0009941603057086468, + 0.0009324053535237908, + 0.0014332233695313334, + 0.0010419267928227782, + 0.0008137857657857239, + 0.0012435876997187734, + 0.0009293444454669952, + 0.0012027519987896085, + 0.0007041682256385684, + 0.0007065744139254093, + 0.0008948097820393741 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 7 + ], + "selected_reward": -0.005268380511552095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294735968112946 + ], + "opposite_logprob_tokens": [ + -2.237947463989258 + ], + "logprob_reward": -0.11294735968112946, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294735968112946, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0336143933236599, + 0.03643827512860298, + 0.4326675832271576, + 0.06994566321372986, + 0.02413717843592167, + 0.018956009298563004, + 0.32012856006622314, + 0.012209760956466198, + 0.027695847675204277, + 0.02420678362250328 + ], + "credit_map": [ + 0.005111090373247862, + 0.0020432560704648495, + 0.008182134479284286, + 0.00245135510340333, + 0.0019150215666741133, + 0.005541892256587744, + 0.0032765045762062073, + 0.0050931344740092754, + 0.004869110882282257, + 0.002827262505888939, + 0.006103814113885164, + 0.0036162398755550385, + 0.0032188582699745893, + 0.005995078012347221, + 0.002066883957013488, + 0.007741025649011135, + 0.06912680715322495, + 0.009441383183002472, + 0.16076335310935974, + 0.012438146397471428, + 0.009758224710822105, + 0.12748442590236664, + 0.00528437877073884, + 0.03837083280086517, + 0.016711313277482986, + 0.005062193144112825, + 0.018378904089331627, + 0.006117200944572687, + 0.007671766448765993, + 0.004288827534765005, + 0.0060781012289226055, + 0.005637355148792267, + 0.004942188039422035, + 0.0024531609378755093, + 0.004480527713894844, + 0.0025146566331386566, + 0.0028699967078864574, + 0.002191276289522648, + 0.0023291874676942825, + 0.0023561841808259487, + 0.002874098252505064, + 0.0024718486238271, + 0.0028364176396280527, + 0.0017399381613358855, + 0.0020204035099595785, + 0.0016518073389306664, + 0.0022329676430672407, + 0.0031285262666642666, + 0.13285069167613983, + 0.013503259979188442, + 0.13516202569007874, + 0.00890771672129631, + 0.008133074268698692, + 0.009271281771361828, + 0.004919414408504963, + 0.007381082512438297, + 0.0017929886234924197, + 0.0018561935285106301, + 0.0014992246869951487, + 0.001667239353992045, + 0.0012725484557449818, + 0.0012299978407099843, + 0.0011521268170326948, + 0.0017394404858350754, + 0.002878208179026842, + 0.004485938232392073, + 0.0037566882092505693, + 0.003522885497659445, + 0.0031132977455854416, + 0.0026574188377708197, + 0.0028483185451477766, + 0.004433089401572943, + 0.003848422085866332, + 0.0033434638753533363, + 0.0016911675920709968, + 0.0021348614245653152, + 0.006931247655302286, + 0.002045489614829421, + 0.0020135000813752413, + 0.002198627684265375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 3 + ], + "selected_reward": -0.11294735968112946 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958176217973232 + ], + "opposite_logprob_tokens": [ + -5.130958080291748 + ], + "logprob_reward": -0.005958176217973232, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958176217973232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05611090734601021, + 0.04285118728876114, + 0.22987748682498932, + 0.06892916560173035, + 0.04328494891524315, + 0.05635853111743927, + 0.3844763934612274, + 0.049917083233594894, + 0.03323569893836975, + 0.034958600997924805 + ], + "credit_map": [ + 0.006665015127509832, + 0.0038987132720649242, + 0.012014714069664478, + 0.004608233459293842, + 0.004276410676538944, + 0.008576377294957638, + 0.005559660494327545, + 0.010511784814298153, + 0.004204932134598494, + 0.0038267269264906645, + 0.00989428162574768, + 0.004747019149363041, + 0.004180902615189552, + 0.006305285729467869, + 0.0033751872833818197, + 0.006316855549812317, + 0.03659839928150177, + 0.005744086112827063, + 0.0844632238149643, + 0.006915396545082331, + 0.0066309659741818905, + 0.06602492928504944, + 0.00407120818272233, + 0.019429299980401993, + 0.014538314193487167, + 0.005096311680972576, + 0.018641022965312004, + 0.006701913196593523, + 0.0073011270724236965, + 0.004646570421755314, + 0.006083380896598101, + 0.005920530762523413, + 0.012428691610693932, + 0.0022749449126422405, + 0.01323615200817585, + 0.0036874311044812202, + 0.0030476560350507498, + 0.0023537366650998592, + 0.0032733618281781673, + 0.0029829801060259342, + 0.008500811643898487, + 0.003908173181116581, + 0.008296447806060314, + 0.005257779266685247, + 0.00618589436635375, + 0.004576194100081921, + 0.01102039311081171, + 0.008612842299044132, + 0.15273447334766388, + 0.015938246622681618, + 0.1538962423801422, + 0.010634971782565117, + 0.010617272928357124, + 0.012602968141436577, + 0.009422985836863518, + 0.01862923800945282, + 0.013328982517123222, + 0.004817793145775795, + 0.014035383239388466, + 0.0032670979853719473, + 0.0026009823195636272, + 0.003845693776383996, + 0.0021638181060552597, + 0.005857336800545454, + 0.0051746065728366375, + 0.0035033810418099165, + 0.004794579930603504, + 0.004644515924155712, + 0.004558556247502565, + 0.0029614733066409826, + 0.003060293383896351, + 0.004538296721875668, + 0.005103330127894878, + 0.0030987521167844534, + 0.006252448074519634, + 0.0043546343222260475, + 0.004286160692572594, + 0.0037316654343158007, + 0.0035577453672885895, + 0.004573867190629244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 3 + ], + "selected_reward": -0.005958176217973232 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370017558336258 + ], + "opposite_logprob_tokens": [ + -3.4087002277374268 + ], + "logprob_reward": -0.03370017558336258, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370017558336258, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0666651651263237, + 0.061638232320547104, + 0.05628746375441551, + 0.09625852108001709, + 0.047971587628126144, + 0.09134051203727722, + 0.4049546718597412, + 0.03493363782763481, + 0.03171491250395775, + 0.10823528468608856 + ], + "credit_map": [ + 0.009409990161657333, + 0.004102344624698162, + 0.015464412979781628, + 0.006810377351939678, + 0.004379876423627138, + 0.01037796400487423, + 0.005615445785224438, + 0.010504755191504955, + 0.005521922837942839, + 0.00549395103007555, + 0.013605033047497272, + 0.007017928641289473, + 0.005671813618391752, + 0.009783958084881306, + 0.004049278330057859, + 0.010494343936443329, + 0.00979960709810257, + 0.0028534389566630125, + 0.014751207083463669, + 0.0046676467172801495, + 0.0037353618536144495, + 0.012128230184316635, + 0.0033128655049949884, + 0.005039109382778406, + 0.02401866763830185, + 0.005275372415781021, + 0.028047971427440643, + 0.007741840090602636, + 0.009840385057032108, + 0.005954221822321415, + 0.007717551663517952, + 0.007662508171051741, + 0.01656191423535347, + 0.002465310972183943, + 0.018612587824463844, + 0.0022491507697850466, + 0.0018143793568015099, + 0.0022895359434187412, + 0.0016207396984100342, + 0.0023579727858304977, + 0.02549125626683235, + 0.007196251302957535, + 0.02434379793703556, + 0.007477891631424427, + 0.006018046289682388, + 0.006921145599335432, + 0.006957929581403732, + 0.0069341943599283695, + 0.16086338460445404, + 0.02447490021586418, + 0.16170430183410645, + 0.012323352508246899, + 0.010896905325353146, + 0.011765411123633385, + 0.009006621316075325, + 0.01391979493200779, + 0.0061171697452664375, + 0.0037742957938462496, + 0.006836744956672192, + 0.0032085285056382418, + 0.0024945586919784546, + 0.0030311867594718933, + 0.0032565402798354626, + 0.006214614026248455, + 0.005129075609147549, + 0.005588306114077568, + 0.003752492368221283, + 0.0044831461273133755, + 0.002499018795788288, + 0.0027273150626569986, + 0.0031365356408059597, + 0.0043990216217935085, + 0.017928684130311012, + 0.013104681856930256, + 0.007845154963433743, + 0.008679184131324291, + 0.03286127373576164, + 0.00914676208049059, + 0.009821219369769096, + 0.008848323486745358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 3 + ], + "selected_reward": -0.03370017558336258 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle captures a nearly perfect cube of ice resting on the dark, polished surface of a black slate countertop. A single, cool, overhead light illuminates the scene, creating sharp, shimmering highlights on the ice's damp surface, reflected in the slowly forming pool of meltwater beneath. The translucent ice reveals a network of tiny internal bubbles, frozen in its crystalline structure. Suddenly, a fine crack appears, originating at the core and moving swiftly and decisively across the cube's surface in a jagged, lightning-fast path. With a silent, violent pop, internal thermal stresses reach a peak, fracturing the cube into two separate, unequal fragments. The force of the fracture abruptly separates the two pieces; they slide briefly across the smooth, damp countertop, their newly exposed, rough-textured fractured surfaces catching the light, before falling completely still and separating. The entire event unfolds with stark, dramatic clarity, creating an atmosphere of intense focus, isolating this small, yet devastating moment against the dark, minimalist backdrop.", + "seed": 31590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_6362_seed_31590_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02968821693211794, + "vlm_logprob": -0.02968821693211794, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02968821693211794, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06918541342020035 + ], + "opposite_logprob_tokens": [ + -2.944185495376587 + ], + "logprob_reward": -0.06918541342020035, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.06918541342020035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08034563064575195, + 0.022356972098350525, + 0.2957131266593933, + 0.036639731377363205, + 0.0353737398982048, + 0.023483751341700554, + 0.41845715045928955, + 0.05714070424437523, + 0.015612611547112465, + 0.014876563102006912 + ], + "credit_map": [ + 0.00834584329277277, + 0.008131095208227634, + 0.011405246332287788, + 0.006779210176318884, + 0.008658001199364662, + 0.011028038337826729, + 0.01133915688842535, + 0.014659038744866848, + 0.0019170331070199609, + 0.0018577033188194036, + 0.003085675183683634, + 0.00260546593926847, + 0.003961686976253986, + 0.0029326530639082193, + 0.0027935649268329144, + 0.0032031885348260403, + 0.017843974754214287, + 0.00708589656278491, + 0.007704002782702446, + 0.004562895279377699, + 0.00892276968806982, + 0.12454293668270111, + 0.009853390045464039, + 0.11519724875688553, + 0.006348360795527697, + 0.005819989833980799, + 0.003921217750757933, + 0.004748405888676643, + 0.004090844187885523, + 0.0033509274944663048, + 0.004213923122733831, + 0.004146060440689325, + 0.003255467163398862, + 0.0027462062425911427, + 0.0075013292953372, + 0.003124064998701215, + 0.011089028790593147, + 0.0035224698949605227, + 0.0018433763179928064, + 0.0022917964961379766, + 0.0034442474134266376, + 0.0016277377726510167, + 0.003552344162017107, + 0.0025399804580956697, + 0.0028008557856082916, + 0.002555032493546605, + 0.005052493419498205, + 0.0019110587891191244, + 0.029152167961001396, + 0.03022320568561554, + 0.004149056971073151, + 0.006159691605716944, + 0.185152068734169, + 0.13670478761196136, + 0.016577821224927902, + 0.010338331572711468, + 0.007137557957321405, + 0.0047043683007359505, + 0.006559696048498154, + 0.006998517084866762, + 0.006240846123546362, + 0.009378650225698948, + 0.008103406988084316, + 0.00801766011863947, + 0.002186936093494296, + 0.0012004681630060077, + 0.0030337318312376738, + 0.0014366372488439083, + 0.0016860052710399032, + 0.0022024991922080517, + 0.002055626828223467, + 0.0018107077339664102, + 0.0019990825094282627, + 0.0010049007833003998, + 0.0015226390678435564, + 0.0013238991377875209, + 0.002063262974843383, + 0.0026770946569740772, + 0.0017400442156940699, + 0.0025456398725509644 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.06918541342020035 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004120432306081057 + ], + "opposite_logprob_tokens": [ + -5.504120349884033 + ], + "logprob_reward": -0.004120432306081057, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004120432306081057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30176886916160583, + 0.04936617612838745, + 0.16091811656951904, + 0.02979864366352558, + 0.12936227023601532, + 0.02130221202969551, + 0.15334925055503845, + 0.07292632013559341, + 0.017757272347807884, + 0.0634509027004242 + ], + "credit_map": [ + 0.0312698557972908, + 0.020476488396525383, + 0.05042504519224167, + 0.02629539929330349, + 0.02551392838358879, + 0.050452347844839096, + 0.028171174228191376, + 0.06916462630033493, + 0.005643866490572691, + 0.0030611574184149504, + 0.010563341900706291, + 0.003544155741110444, + 0.004739571362733841, + 0.008380896411836147, + 0.0035264401230961084, + 0.009906744584441185, + 0.01035421621054411, + 0.005032092798501253, + 0.006046726368367672, + 0.0026227072812616825, + 0.004180673509836197, + 0.06819254904985428, + 0.0035242652520537376, + 0.06096486747264862, + 0.004276415798813105, + 0.0032490382436662912, + 0.004363480489701033, + 0.003616957226768136, + 0.003798222867771983, + 0.00391176575794816, + 0.003181342501193285, + 0.0034014189150184393, + 0.02154206670820713, + 0.005523940082639456, + 0.04382365196943283, + 0.006827611476182938, + 0.008927418850362301, + 0.0086530651897192, + 0.004048945382237434, + 0.030015548691153526, + 0.0024940550792962313, + 0.001677239197306335, + 0.0030335444025695324, + 0.0020158204715698957, + 0.0024855411611497402, + 0.0031035118736326694, + 0.0032567642629146576, + 0.0032357354648411274, + 0.010478860698640347, + 0.011236518621444702, + 0.0019918959587812424, + 0.002537997905164957, + 0.06707320362329483, + 0.04933625087141991, + 0.00620286725461483, + 0.004491637460887432, + 0.008687692694365978, + 0.006879766006022692, + 0.006879989989101887, + 0.00849295873194933, + 0.011216200888156891, + 0.011432670056819916, + 0.0086378064006567, + 0.010699235834181309, + 0.001974129816517234, + 0.0014435313642024994, + 0.0018800543621182442, + 0.0012482110178098083, + 0.0033029820770025253, + 0.003029515966773033, + 0.0020112486090511084, + 0.0028675966896116734, + 0.007380227092653513, + 0.004723252262920141, + 0.008441408164799213, + 0.0041174148209393024, + 0.00882506463676691, + 0.009189935401082039, + 0.0093832490965724, + 0.011390343308448792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 6 + ], + "selected_reward": -0.004120432306081057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497054383158684 + ], + "opposite_logprob_tokens": [ + -2.9299705028533936 + ], + "logprob_reward": -0.05497054383158684, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497054383158684, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0679296925663948, + 0.045806676149368286, + 0.5424888730049133, + 0.04619254171848297, + 0.03897864744067192, + 0.07736456394195557, + 0.09045535326004028, + 0.03325437009334564, + 0.04424123466014862, + 0.013288027606904507 + ], + "credit_map": [ + 0.009693735279142857, + 0.00529427221044898, + 0.00917046982795, + 0.003997458145022392, + 0.006477024871855974, + 0.011849630624055862, + 0.006359259132295847, + 0.015087842009961605, + 0.004142823629081249, + 0.004122607875615358, + 0.008312448859214783, + 0.0038006941322237253, + 0.006234558764845133, + 0.0055461544543504715, + 0.005824762396514416, + 0.007822629995644093, + 0.028021803125739098, + 0.011688020080327988, + 0.010421312414109707, + 0.00501612713560462, + 0.010785616934299469, + 0.24032406508922577, + 0.006960871163755655, + 0.2292710393667221, + 0.006773778237402439, + 0.006348409689962864, + 0.0060882060788571835, + 0.005375273991376162, + 0.004146010614931583, + 0.004915196914225817, + 0.0050535909831523895, + 0.0074920752085745335, + 0.0038514770567417145, + 0.0055692559108138084, + 0.006629427894949913, + 0.0029418454505503178, + 0.007395930588245392, + 0.0045689987018704414, + 0.0033802876714617014, + 0.004641421604901552, + 0.010632445104420185, + 0.005944705568253994, + 0.01130981557071209, + 0.007738224696367979, + 0.008355729281902313, + 0.01174794603139162, + 0.011638374999165535, + 0.009997325949370861, + 0.007678482681512833, + 0.006655428092926741, + 0.005947743076831102, + 0.004827211145311594, + 0.03322147578001022, + 0.022298332303762436, + 0.005930375307798386, + 0.0038963044062256813, + 0.004876443184912205, + 0.00295992917381227, + 0.0031180684454739094, + 0.003338898764923215, + 0.0040795523673295975, + 0.005034408997744322, + 0.003782624378800392, + 0.006064441986382008, + 0.004593753255903721, + 0.003170607378706336, + 0.005859045777469873, + 0.003908348735421896, + 0.007550911512225866, + 0.006245523225516081, + 0.005025468301028013, + 0.007887579500675201, + 0.0016801415476948023, + 0.0008760697091929615, + 0.0012310340534895658, + 0.0010253202635794878, + 0.001687976997345686, + 0.00243241130374372, + 0.001229622052051127, + 0.0031254517380148172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 5 + ], + "selected_reward": -0.05497054383158684 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005964694079011679 + ], + "opposite_logprob_tokens": [ + -5.130964756011963 + ], + "logprob_reward": -0.005964694079011679, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005964694079011679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.016269436106085777, + 0.023297665640711784, + 0.33193516731262207, + 0.020318012684583664, + 0.05117601156234741, + 0.004439756274223328, + 0.533636212348938, + 0.006063046399503946, + 0.00337421428412199, + 0.009490440599620342 + ], + "credit_map": [ + 0.002451021922752261, + 0.001351374201476574, + 0.0022887003142386675, + 0.001605506637133658, + 0.0017283125780522823, + 0.0025775216054171324, + 0.0015119998715817928, + 0.0027549993246793747, + 0.0029886101838201284, + 0.0019299823325127363, + 0.002640752587467432, + 0.001969735836610198, + 0.0038969048764556646, + 0.003555628005415201, + 0.0027354233898222446, + 0.0035806309897452593, + 0.019271738827228546, + 0.006950159091502428, + 0.005963996518403292, + 0.0030615886207669973, + 0.0062911780551075935, + 0.1485234498977661, + 0.0034550391137599945, + 0.1384180337190628, + 0.0026912360917776823, + 0.002760660368949175, + 0.00297592137940228, + 0.002641307422891259, + 0.0023102222476154566, + 0.001995118334889412, + 0.0018623941577970982, + 0.0030811545439064503, + 0.005106806755065918, + 0.002172286855056882, + 0.018743565306067467, + 0.0031027861405164003, + 0.011092610657215118, + 0.004140887409448624, + 0.0014113361248746514, + 0.005405739415436983, + 0.0007026309031061828, + 0.00039075917447917163, + 0.0005243739578872919, + 0.0005149035714566708, + 0.00045964133460074663, + 0.0007286701584234834, + 0.0004809233942069113, + 0.000637854274827987, + 0.037378940731287, + 0.03875980153679848, + 0.004891455173492432, + 0.007537629455327988, + 0.23678019642829895, + 0.17399421334266663, + 0.02096588723361492, + 0.01332810427993536, + 0.0008031798643060029, + 0.000536922481842339, + 0.0006675535114482045, + 0.000649137538857758, + 0.0007321301964111626, + 0.000991357141174376, + 0.0008142265724018216, + 0.0008685399079695344, + 0.00038453645538538694, + 0.0002474781358614564, + 0.0004149754240643233, + 0.0002860466775018722, + 0.00040260140667669475, + 0.0005259756580926478, + 0.00043043302139267325, + 0.0006821679417043924, + 0.000892516749445349, + 0.000741005118470639, + 0.0011565756285563111, + 0.0010206917067989707, + 0.0009328955784440041, + 0.0013847706140950322, + 0.0014819859061390162, + 0.0018800002289935946 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.005964694079011679 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200001023709774 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200001023709774, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200001023709774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04452639818191528, + 0.06368236243724823, + 0.30894696712493896, + 0.08175013214349747, + 0.08390791714191437, + 0.025511043146252632, + 0.3101577162742615, + 0.026752110570669174, + 0.020454997196793556, + 0.03431034833192825 + ], + "credit_map": [ + 0.0058935945853590965, + 0.0033327299170196056, + 0.006123697850853205, + 0.003201575018465519, + 0.00467791547998786, + 0.008180609904229641, + 0.0030979954171925783, + 0.010018284432590008, + 0.006011222954839468, + 0.006639434490352869, + 0.005339526571333408, + 0.005925833247601986, + 0.01166505180299282, + 0.006825885269790888, + 0.011803867295384407, + 0.009471545927226543, + 0.018758434802293777, + 0.007234275806695223, + 0.007296839263290167, + 0.004300912842154503, + 0.0064904470928013325, + 0.1350332498550415, + 0.005435623228549957, + 0.1243971511721611, + 0.010468041524291039, + 0.013846123591065407, + 0.007937749847769737, + 0.010061531327664852, + 0.011178474873304367, + 0.008021625690162182, + 0.010900932364165783, + 0.009335645474493504, + 0.010533120483160019, + 0.004984999541193247, + 0.03037147782742977, + 0.005299470853060484, + 0.009851029142737389, + 0.00666783144697547, + 0.003628769423812628, + 0.012571223080158234, + 0.0035993459168821573, + 0.001612963154911995, + 0.003663684008643031, + 0.0026931825559586287, + 0.003460376989096403, + 0.002403286285698414, + 0.004994295537471771, + 0.003083908697590232, + 0.021190078929066658, + 0.022471150383353233, + 0.003487365087494254, + 0.004745692014694214, + 0.13728967308998108, + 0.10108032077550888, + 0.012305892072618008, + 0.007587562780827284, + 0.0033024330623447895, + 0.0019377891439944506, + 0.0027227045502513647, + 0.003289905609562993, + 0.0032534922938793898, + 0.0038936054334044456, + 0.004311794880777597, + 0.004040384199470282, + 0.0027025246527045965, + 0.0018665720708668232, + 0.002228666329756379, + 0.0018354548374190927, + 0.003823889186605811, + 0.0030763261020183563, + 0.0024477127008140087, + 0.0024738514330238104, + 0.0031266026198863983, + 0.0026720245368778706, + 0.0033836557995527983, + 0.0031525574158877134, + 0.004391867201775312, + 0.005066760815680027, + 0.0055959695018827915, + 0.006920906715095043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 4 + ], + "selected_reward": -0.014200001023709774 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "An extreme close-up, static shot from a slightly elevated angle captures a nearly perfect cube of ice resting on the dark, polished surface of a black slate countertop. A single, cool, overhead light illuminates the scene, creating sharp, shimmering highlights on the ice's damp surface, reflected in the slowly forming pool of meltwater beneath. The translucent ice reveals a network of tiny internal bubbles, frozen in its crystalline structure. Suddenly, a fine crack appears, originating at the core and moving swiftly and decisively across the cube's surface in a jagged, lightning-fast path. With a silent, violent pop, internal thermal stresses reach a peak, fracturing the cube into two separate, unequal fragments. The force of the fracture abruptly separates the two pieces; they slide briefly across the smooth, damp countertop, their newly exposed, rough-textured fractured surfaces catching the light, before falling completely still and separating. The entire event unfolds with stark, dramatic clarity, creating an atmosphere of intense focus, isolating this small, yet devastating moment against the dark, minimalist backdrop.", + "seed": 31591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_5/index_6362_seed_31591_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01699745934456587, + "vlm_logprob": -0.01699745934456587, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.01699745934456587, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02501748315989971 + ], + "opposite_logprob_tokens": [ + -6.400017261505127 + ], + "logprob_reward": -0.02501748315989971, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.02501748315989971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042316555976867676, + 0.07903432101011276, + 0.05063965171575546, + 0.1315319985151291, + 0.042872555553913116, + 0.08546659350395203, + 0.1636909544467926, + 0.14616481959819794, + 0.07349281758069992, + 0.18478979170322418 + ], + "credit_map": [ + 0.006796412169933319, + 0.00397875951603055, + 0.00660751573741436, + 0.004939446225762367, + 0.004048270173370838, + 0.0063239759765565395, + 0.005776042118668556, + 0.0038461333606392145, + 0.014302331954240799, + 0.008976013399660587, + 0.015140634030103683, + 0.0084577901288867, + 0.006709067616611719, + 0.012358525767922401, + 0.004982161335647106, + 0.008107795380055904, + 0.008343823254108429, + 0.004591930191963911, + 0.00777416629716754, + 0.007258806377649307, + 0.005966986529529095, + 0.009970189072191715, + 0.003991861827671528, + 0.0027418911922723055, + 0.0311183650046587, + 0.0073117162100970745, + 0.03232264891266823, + 0.013027061708271503, + 0.008376061916351318, + 0.021434025838971138, + 0.010027316398918629, + 0.00791480578482151, + 0.008637765422463417, + 0.004032015334814787, + 0.010870014317333698, + 0.0042925989255309105, + 0.002844801638275385, + 0.003468862036243081, + 0.004118958953768015, + 0.004607535898685455, + 0.034713707864284515, + 0.0021274168975651264, + 0.033766791224479675, + 0.003331890096887946, + 0.003086170880123973, + 0.002962252823635936, + 0.0029225628823041916, + 0.00255581084638834, + 0.05238890275359154, + 0.0068446253426373005, + 0.05562426149845123, + 0.007476218976080418, + 0.008184741251170635, + 0.008322153240442276, + 0.00985739566385746, + 0.014992665499448776, + 0.05886032059788704, + 0.004578702617436647, + 0.05818101763725281, + 0.003845548490062356, + 0.00752295320853591, + 0.003869663691148162, + 0.003756930585950613, + 0.0055496771819889545, + 0.011424115858972073, + 0.009575488045811653, + 0.015454057604074478, + 0.007895613089203835, + 0.007866079919040203, + 0.0069901649840176105, + 0.00801512785255909, + 0.0062721665017306805, + 0.0525909885764122, + 0.007594120688736439, + 0.011275350116193295, + 0.008019263856112957, + 0.020710406824946404, + 0.006720689591020346, + 0.06973115354776382, + 0.008147834800183773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.02501748315989971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008672667667269707 + ], + "opposite_logprob_tokens": [ + -4.758672714233398 + ], + "logprob_reward": -0.008672667667269707, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008672667667269707, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12196799367666245, + 0.13971321284770966, + 0.11601632088422775, + 0.12632650136947632, + 0.04500166326761246, + 0.05149666219949722, + 0.14893825352191925, + 0.15132029354572296, + 0.021774685010313988, + 0.07744444906711578 + ], + "credit_map": [ + 0.01975785195827484, + 0.011707405559718609, + 0.023569650948047638, + 0.017392072826623917, + 0.007785116787999868, + 0.020009182393550873, + 0.00883547868579626, + 0.012911234050989151, + 0.025586960837244987, + 0.01587938889861107, + 0.02559884451329708, + 0.014525458216667175, + 0.010252532549202442, + 0.026828385889530182, + 0.008165640756487846, + 0.012875990942120552, + 0.022908248007297516, + 0.010020569898188114, + 0.02711503393948078, + 0.015340222977101803, + 0.009200640954077244, + 0.011499151587486267, + 0.008650093339383602, + 0.011282358318567276, + 0.03184179216623306, + 0.006420247256755829, + 0.030262531712651253, + 0.012215983122587204, + 0.007990431971848011, + 0.021757105365395546, + 0.008514368906617165, + 0.0073240455240011215, + 0.007920993492007256, + 0.005103277508169413, + 0.009136728942394257, + 0.005871717818081379, + 0.003985434304922819, + 0.0032855698373168707, + 0.005031873006373644, + 0.00466606579720974, + 0.021755492314696312, + 0.0012461264850571752, + 0.021123753860592842, + 0.0017673090333119035, + 0.0014924760907888412, + 0.001455856952816248, + 0.001537802629172802, + 0.001117844134569168, + 0.048627499490976334, + 0.004068916197866201, + 0.054855067282915115, + 0.005250037647783756, + 0.003827506909146905, + 0.006884438917040825, + 0.005733386147767305, + 0.01969139091670513, + 0.06153295934200287, + 0.004469818435609341, + 0.06175042688846588, + 0.003984837327152491, + 0.005555633455514908, + 0.004620446357876062, + 0.004365278407931328, + 0.0050408923998475075, + 0.003210090333595872, + 0.0031448551453649998, + 0.004038037266582251, + 0.0022815882693976164, + 0.0019144233083352447, + 0.0021922439336776733, + 0.003165925620123744, + 0.0018275214824825525, + 0.022126607596874237, + 0.00350763532333076, + 0.0049264635890722275, + 0.0036279461346566677, + 0.00841752253472805, + 0.0032797164749354124, + 0.027222974225878716, + 0.004335584584623575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.008672667667269707 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201351165771484 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201351165771484, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201351165771484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052853189408779144, + 0.1140422597527504, + 0.07551724463701248, + 0.1904362589120865, + 0.0495094433426857, + 0.14150914549827576, + 0.14119063317775726, + 0.1095660850405693, + 0.06727045029401779, + 0.058105289936065674 + ], + "credit_map": [ + 0.0108293192461133, + 0.004458120558410883, + 0.012216201983392239, + 0.0055120219476521015, + 0.003511647926643491, + 0.006273193284869194, + 0.0034803461749106646, + 0.0065723299048841, + 0.020609695464372635, + 0.011988715268671513, + 0.02199230156838894, + 0.012150316499173641, + 0.009031390771269798, + 0.017596997320652008, + 0.007868209853768349, + 0.012804622761905193, + 0.012188632041215897, + 0.0067982664331793785, + 0.014040609821677208, + 0.012469572015106678, + 0.009562362916767597, + 0.008326411247253418, + 0.0066340952180325985, + 0.00549728749319911, + 0.045883677899837494, + 0.0156019926071167, + 0.04490792751312256, + 0.030189480632543564, + 0.013647526502609253, + 0.014250040054321289, + 0.013323516584932804, + 0.012632076628506184, + 0.008316649124026299, + 0.004752693232148886, + 0.008758322335779667, + 0.008589593693614006, + 0.004346907138824463, + 0.0037217261269688606, + 0.006382029037922621, + 0.00464151706546545, + 0.05508244410157204, + 0.004547553136944771, + 0.05415070429444313, + 0.007041610777378082, + 0.005918524693697691, + 0.004600245505571365, + 0.006457220762968063, + 0.00371082522906363, + 0.04986324906349182, + 0.00344089069403708, + 0.0546385794878006, + 0.004546345677226782, + 0.003322190372273326, + 0.005532912444323301, + 0.004807412624359131, + 0.015039040707051754, + 0.030829327180981636, + 0.007274710573256016, + 0.03225325047969818, + 0.00570943346247077, + 0.01276933029294014, + 0.0060851857997477055, + 0.007250258699059486, + 0.007394576445221901, + 0.009511074982583523, + 0.0076210107654333115, + 0.015140342526137829, + 0.00786482822149992, + 0.0056156194768846035, + 0.007870105095207691, + 0.006699139717966318, + 0.006948323454707861, + 0.012229613959789276, + 0.00217385939322412, + 0.006445177365094423, + 0.0025569337885826826, + 0.0069832876324653625, + 0.0026954251807183027, + 0.02202232927083969, + 0.0029986537992954254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 6 + ], + "selected_reward": -0.018201351165771484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009798170067369938 + ], + "opposite_logprob_tokens": [ + -4.634798049926758 + ], + "logprob_reward": -0.009798170067369938, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009798170067369938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04065637290477753, + 0.025463787838816643, + 0.05855393409729004, + 0.2006201595067978, + 0.028689580038189888, + 0.07007985562086105, + 0.20213325321674347, + 0.33138027787208557, + 0.02596236765384674, + 0.01646038144826889 + ], + "credit_map": [ + 0.007655487395823002, + 0.003964702133089304, + 0.007888942956924438, + 0.004569251090288162, + 0.0035597134847193956, + 0.005664590746164322, + 0.003188468748703599, + 0.00416521867737174, + 0.0043703895062208176, + 0.00284305727109313, + 0.005390369798988104, + 0.003464075271040201, + 0.002024484798312187, + 0.003053401131182909, + 0.002189998049288988, + 0.0021280134096741676, + 0.010931418277323246, + 0.005917280446738005, + 0.010392406955361366, + 0.007930515334010124, + 0.005418376065790653, + 0.0078886142000556, + 0.00585517194122076, + 0.004220149014145136, + 0.05336284264922142, + 0.013917917385697365, + 0.05071204900741577, + 0.030536456033587456, + 0.013138250447809696, + 0.013805757276713848, + 0.012030811980366707, + 0.013116094283759594, + 0.005448196083307266, + 0.0031675202772021294, + 0.0053589907474815845, + 0.0036320413928478956, + 0.0034267783630639315, + 0.0022150417789816856, + 0.003113750834017992, + 0.0023272638209164143, + 0.025079395622015, + 0.0026234930846840143, + 0.023596471175551414, + 0.0028810531366616488, + 0.005090075079351664, + 0.003459300845861435, + 0.0037610798608511686, + 0.0035889926366508007, + 0.06432175636291504, + 0.011207984760403633, + 0.06987248361110687, + 0.011160165071487427, + 0.009007394313812256, + 0.01166163757443428, + 0.01148980762809515, + 0.01341202761977911, + 0.10976029187440872, + 0.01822728104889393, + 0.10773172229528427, + 0.015077329240739346, + 0.020025301724672318, + 0.016029447317123413, + 0.024486342445015907, + 0.02004256844520569, + 0.004100874997675419, + 0.004237794782966375, + 0.004221381153911352, + 0.0027912193909287453, + 0.0024729224387556314, + 0.002661103615537286, + 0.002698182128369808, + 0.0027788898441940546, + 0.0016654249047860503, + 0.0009847043547779322, + 0.002402693498879671, + 0.0012579289032146335, + 0.0035953763872385025, + 0.0016019509639590979, + 0.0038933863397687674, + 0.0010589183075353503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.009798170067369938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an extreme close-up of a nearly perfect cube of ice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the ice resting on a dark, polished black slate countertop?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera remain static from a slightly elevated angle throughout the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the ice cube fracture into two separate, unequal fragments?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the two fragments slide briefly across the countertop before falling completely still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232976246625185 + ], + "opposite_logprob_tokens": [ + -3.7732975482940674 + ], + "logprob_reward": -0.0232976246625185, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232976246625185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04410393163561821, + 0.090462327003479, + 0.0708051323890686, + 0.19958876073360443, + 0.06865768879652023, + 0.024454766884446144, + 0.14619916677474976, + 0.2284592092037201, + 0.027511974796652794, + 0.09975703805685043 + ], + "credit_map": [ + 0.008097942918539047, + 0.004117401782423258, + 0.00883590616285801, + 0.004960429854691029, + 0.003312202636152506, + 0.006330462172627449, + 0.00351965450681746, + 0.0049299271777272224, + 0.01660757139325142, + 0.00901939906179905, + 0.01810622774064541, + 0.008983825333416462, + 0.007145164534449577, + 0.01490156352519989, + 0.0057757203467190266, + 0.009922856464982033, + 0.014141219668090343, + 0.007274629082530737, + 0.014701337553560734, + 0.00957384705543518, + 0.003622364951297641, + 0.010476583614945412, + 0.006113023497164249, + 0.004902123939245939, + 0.05429232120513916, + 0.013591808266937733, + 0.05055782571434975, + 0.030498607084155083, + 0.012180388905107975, + 0.015237879008054733, + 0.011238250881433487, + 0.011991684325039387, + 0.014577770605683327, + 0.006302869413048029, + 0.014487264677882195, + 0.00897546298801899, + 0.004866074770689011, + 0.005581342615187168, + 0.00741586135700345, + 0.00645104655995965, + 0.004934072494506836, + 0.0016045343363657594, + 0.005727480165660381, + 0.003050450002774596, + 0.0028268720488995314, + 0.0024667540565133095, + 0.0021406172309070826, + 0.001703985850326717, + 0.0542990043759346, + 0.006259433459490538, + 0.056017905473709106, + 0.00581539748236537, + 0.005263341590762138, + 0.005357092712074518, + 0.0066280788742005825, + 0.006558921188116074, + 0.07651810348033905, + 0.01139605138450861, + 0.0801425650715828, + 0.009950241073966026, + 0.012386258691549301, + 0.01019467692822218, + 0.015584638342261314, + 0.012286693789064884, + 0.0038374720606952906, + 0.003947137389332056, + 0.004106825217604637, + 0.0024172309786081314, + 0.0033058219123631716, + 0.003973602782934904, + 0.003378399880602956, + 0.0025454869028180838, + 0.02845011278986931, + 0.0039000925607979298, + 0.0055524082854390144, + 0.004040620289742947, + 0.010527696460485458, + 0.004303301684558392, + 0.03790631890296936, + 0.005076481960713863 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.0232976246625185 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_tensor([5899, 5899])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_tensor([5899, 5899])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e2f873c37e32740e58cdc9999ac9ac7cf37f31d3 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_tensor([5899, 5899])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "long drone shot, taken from high above, slowly zooms out, capturing a tranquil and poetic scene. Beneath the morning mist, the vast river's surface is as smooth as a mirror, a deep, dark green. A quaint, dark brown wooden boat, its hull etched with age, steadily and leisurely paddles from one side of the frame toward the center. A boatman, wearing a conical hat and a gray linen gown, slowly and steadily paddles his oars. Each silent stroke of the oars creates a widening V-shaped wave at the stern, stirring up a series of perfectly concentric ripples. These ripples spread outward smoothly and gracefully, gradually growing larger and shallower from a distinct ripple, ultimately fading silently into the vast river. The soft morning light penetrates the mist, casting mottled golden spots on the water, which together with the vague silhouettes of mountains in the distance form a flowing ink painting. The whole scene is filled with meditative tranquility and lonely beauty.", + "seed": 17582, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17582_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.24312400966882705, + "vlm_logprob": -0.24312400966882705, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.24312400966882705, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.37139272689819336 + ], + "opposite_logprob_tokens": [ + -1.2463927268981934 + ], + "logprob_reward": -0.37139272689819336, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.37139272689819336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5598140358924866, + 0.045204855501651764, + 0.05157437548041344, + 0.0433928444981575, + 0.026160361245274544, + 0.045499108731746674, + 0.040711596608161926, + 0.10555854439735413, + 0.06085091829299927, + 0.02123337797820568 + ], + "credit_map": [ + 0.06075747311115265, + 0.08110900968313217, + 0.031441185623407364, + 0.04142060875892639, + 0.10778693109750748, + 0.08715007454156876, + 0.05145932734012604, + 0.09868939965963364, + 0.00971677340567112, + 0.005906622391194105, + 0.005881014280021191, + 0.004592341370880604, + 0.004297947511076927, + 0.006000938359647989, + 0.0035355640575289726, + 0.005273650400340557, + 0.005887149833142757, + 0.006392683833837509, + 0.0040495735593140125, + 0.002620452083647251, + 0.01117991004139185, + 0.010173904709517956, + 0.0031318562105298042, + 0.00813884288072586, + 0.005395420361310244, + 0.0029809968546032906, + 0.008944221772253513, + 0.002152060391381383, + 0.004209194332361221, + 0.013721964322030544, + 0.001602092757821083, + 0.0043868920765817165, + 0.0029128198511898518, + 0.002769047161564231, + 0.0043007247149944305, + 0.003056693123653531, + 0.0036505323369055986, + 0.002928872825577855, + 0.0035865369718521833, + 0.0029551349580287933, + 0.006427247542887926, + 0.005252373870462179, + 0.007725692354142666, + 0.004800901748239994, + 0.004483767785131931, + 0.005343593657016754, + 0.003853162517771125, + 0.007612368557602167, + 0.004318809602409601, + 0.003666155505925417, + 0.007474994752556086, + 0.003218288067728281, + 0.0033861359115689993, + 0.005090022459626198, + 0.0025335338432341814, + 0.011023657396435738, + 0.012095796875655651, + 0.010548330843448639, + 0.015288012102246284, + 0.010935728438198566, + 0.014465762302279472, + 0.014192626811563969, + 0.012245851568877697, + 0.0157864261418581, + 0.006208449136465788, + 0.004870091564953327, + 0.018350815400481224, + 0.0036534094251692295, + 0.004482550080865622, + 0.0175128486007452, + 0.002358213532716036, + 0.003414542879909277, + 0.002011901931837201, + 0.0019873178098350763, + 0.0026639814022928476, + 0.0029411183204501867, + 0.0036755744367837906, + 0.003945269156247377, + 0.002074648393318057, + 0.0019335668766871095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.37139272689819336 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578065991401672 + ], + "opposite_logprob_tokens": [ + -0.6328065991401672 + ], + "logprob_reward": -0.7578065991401672, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578065991401672, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.274579793214798, + 0.046915289014577866, + 0.05157412961125374, + 0.01754370518028736, + 0.04847622662782669, + 0.0831419825553894, + 0.04010150954127312, + 0.34327825903892517, + 0.062168821692466736, + 0.032220225781202316 + ], + "credit_map": [ + 0.0311961118131876, + 0.03944111242890358, + 0.014616799540817738, + 0.019817549735307693, + 0.0544569231569767, + 0.04223025590181351, + 0.025594398379325867, + 0.047226667404174805, + 0.004008464980870485, + 0.007781957741826773, + 0.002867290982976556, + 0.0030906882602721453, + 0.010590879246592522, + 0.007592774927616119, + 0.0035861111246049404, + 0.007397127337753773, + 0.006290108896791935, + 0.008295418694615364, + 0.00456662755459547, + 0.004549241159111261, + 0.0059019397012889385, + 0.007220158353447914, + 0.0032674255780875683, + 0.011483212932944298, + 0.0012815550435334444, + 0.0020853031892329454, + 0.0015920954756438732, + 0.0023409423884004354, + 0.002274105791002512, + 0.0033277752809226513, + 0.0020648047793656588, + 0.0025771239306777716, + 0.00453515350818634, + 0.005268468055874109, + 0.007121392525732517, + 0.006487112492322922, + 0.006894783116877079, + 0.0052527268417179585, + 0.007095368579030037, + 0.0058212243020534515, + 0.010373823344707489, + 0.006867676507681608, + 0.021205725148320198, + 0.004650446120649576, + 0.007842677645385265, + 0.005463650915771723, + 0.004770235158503056, + 0.021967755630612373, + 0.0029664398171007633, + 0.00555152865126729, + 0.0036699920892715454, + 0.004304512869566679, + 0.006246465723961592, + 0.006390017922967672, + 0.003924883436411619, + 0.007047670893371105, + 0.030233843252062798, + 0.027235284447669983, + 0.08007108420133591, + 0.021969228982925415, + 0.03015179932117462, + 0.05074748396873474, + 0.01739467866718769, + 0.0854749009013176, + 0.006252691615372896, + 0.004760183393955231, + 0.01951872929930687, + 0.003127412172034383, + 0.0044258516281843185, + 0.017903534695506096, + 0.0023072606418281794, + 0.0038731605745851994, + 0.0028122286312282085, + 0.004176421090960503, + 0.00320123927667737, + 0.0050797476433217525, + 0.005622758995741606, + 0.004283026326447725, + 0.0032336274161934853, + 0.003811178496107459 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.7578065991401672 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708129078149796 + ], + "opposite_logprob_tokens": [ + -3.408708095550537 + ], + "logprob_reward": -0.033708129078149796, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708129078149796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3109428584575653, + 0.050139378756284714, + 0.15071965754032135, + 0.09440971165895462, + 0.017939381301403046, + 0.013872845098376274, + 0.03239000588655472, + 0.2867630422115326, + 0.020763851702213287, + 0.022059261798858643 + ], + "credit_map": [ + 0.033072154968976974, + 0.04584337770938873, + 0.019929010421037674, + 0.02253665216267109, + 0.06117520481348038, + 0.04676403850317001, + 0.02904239483177662, + 0.052580058574676514, + 0.0068864780478179455, + 0.007568712346255779, + 0.0037291632033884525, + 0.006007384043186903, + 0.0061724563129246235, + 0.006964782252907753, + 0.003506395500153303, + 0.009304012171924114, + 0.012824584729969501, + 0.018820015713572502, + 0.013397020287811756, + 0.008898074738681316, + 0.0262931939214468, + 0.02755512110888958, + 0.008554666303098202, + 0.03437698632478714, + 0.007177543360739946, + 0.006077007856220007, + 0.00995644647628069, + 0.0028536138124763966, + 0.006961814593523741, + 0.032854970544576645, + 0.0022923059295862913, + 0.02623601444065571, + 0.0016937959007918835, + 0.0022383464965969324, + 0.0018186186207458377, + 0.001860475866124034, + 0.0018626556266099215, + 0.0025097050238400698, + 0.0017539415275678039, + 0.004201844800263643, + 0.0014059593668207526, + 0.0015645053936168551, + 0.0010250986088067293, + 0.0017291516996920109, + 0.0019666345324367285, + 0.001956365304067731, + 0.0020393908489495516, + 0.002185741439461708, + 0.003407467622309923, + 0.0041182441636919975, + 0.00339769059792161, + 0.0029707460198551416, + 0.004436388146132231, + 0.00710717961192131, + 0.0024973333347588778, + 0.004454959183931351, + 0.026583772152662277, + 0.021847544237971306, + 0.0628090426325798, + 0.017932353541254997, + 0.024872342124581337, + 0.037059564143419266, + 0.019826248288154602, + 0.07583217322826385, + 0.002069039735943079, + 0.0023052352480590343, + 0.002277493244037032, + 0.0026977898087352514, + 0.0021039501298218966, + 0.003694905200973153, + 0.002166878432035446, + 0.003448561765253544, + 0.0022707905154675245, + 0.002580825472250581, + 0.0018027700716629624, + 0.0029669031500816345, + 0.0031479382887482643, + 0.003119139699265361, + 0.0024886794853955507, + 0.003682216862216592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.033708129078149796 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636430785059929 + ], + "opposite_logprob_tokens": [ + -3.651364326477051 + ], + "logprob_reward": -0.02636430785059929, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636430785059929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20479388535022736, + 0.028019152581691742, + 0.11660532653331757, + 0.0751643031835556, + 0.023558132350444794, + 0.03612932562828064, + 0.021798785775899887, + 0.4583364427089691, + 0.019928837195038795, + 0.015665777027606964 + ], + "credit_map": [ + 0.022318730130791664, + 0.029481535777449608, + 0.012038243003189564, + 0.014437135308980942, + 0.04017481207847595, + 0.03133624047040939, + 0.01871659606695175, + 0.03629060089588165, + 0.0032581922132521868, + 0.004113533068448305, + 0.002986035542562604, + 0.0026932384353131056, + 0.004497545771300793, + 0.003789677983149886, + 0.0020715934224426746, + 0.004609336145222187, + 0.012564144097268581, + 0.014398659579455853, + 0.010761460289359093, + 0.0060363351367414, + 0.021687719970941544, + 0.021850667893886566, + 0.006742204073816538, + 0.022564135491847992, + 0.009720596484839916, + 0.00733181182295084, + 0.01842944696545601, + 0.0032382041681557894, + 0.004963308107107878, + 0.005067368037998676, + 0.0024020541459321976, + 0.02401151694357395, + 0.002824655268341303, + 0.002575048478320241, + 0.004238265100866556, + 0.0028763748705387115, + 0.0026821040082722902, + 0.00256172614172101, + 0.0026026254054158926, + 0.0031973335426300764, + 0.00392542127519846, + 0.0037747249007225037, + 0.006580682937055826, + 0.0038053393363952637, + 0.003458375111222267, + 0.0040754894725978374, + 0.0030405246652662754, + 0.007468765135854483, + 0.0018457522382959723, + 0.0025462093763053417, + 0.0030603150371462107, + 0.0013382353354245424, + 0.0023949274327605963, + 0.003212951822206378, + 0.0012132979463785887, + 0.0061870962381362915, + 0.04232625290751457, + 0.035645946860313416, + 0.1010928750038147, + 0.029169322922825813, + 0.03975522518157959, + 0.061167165637016296, + 0.03161536529660225, + 0.11756432056427002, + 0.0017155916430056095, + 0.0019035905133932829, + 0.00428389897570014, + 0.0017913095653057098, + 0.0017668359214439988, + 0.005487547721713781, + 0.0013031018897891045, + 0.0016769608482718468, + 0.0017757819732651114, + 0.0015815779333934188, + 0.0015243323286995292, + 0.0021110696252435446, + 0.0020214158575981855, + 0.0023026939015835524, + 0.0018315495690330863, + 0.0025173560716211796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.02636430785059929 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026348285377025604 + ], + "opposite_logprob_tokens": [ + -3.651348352432251 + ], + "logprob_reward": -0.026348285377025604, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026348285377025604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10220801830291748, + 0.06547728180885315, + 0.09388656169176102, + 0.15925884246826172, + 0.01262657344341278, + 0.12149704247713089, + 0.07343171536922455, + 0.18482105433940887, + 0.1539974808692932, + 0.032795362174510956 + ], + "credit_map": [ + 0.010677472688257694, + 0.013210361823439598, + 0.013128268532454967, + 0.010910598561167717, + 0.019034679979085922, + 0.012813721783459187, + 0.010087281465530396, + 0.012345627881586552, + 0.01396158616989851, + 0.008935006335377693, + 0.008169857785105705, + 0.006605576258152723, + 0.006444008555263281, + 0.006894183345139027, + 0.004767071455717087, + 0.009699990972876549, + 0.012982734479010105, + 0.013803242705762386, + 0.008721845224499702, + 0.008290397003293037, + 0.013030449859797955, + 0.01321676466614008, + 0.006945417728275061, + 0.016895705834031105, + 0.013123305514454842, + 0.01376822218298912, + 0.027391944080591202, + 0.005236372817307711, + 0.008648883551359177, + 0.024494653567671776, + 0.005602849181741476, + 0.060992613434791565, + 0.0013603483093902469, + 0.0016990723088383675, + 0.001014718203805387, + 0.0011880005476996303, + 0.001525436993688345, + 0.001835149829275906, + 0.00122248032130301, + 0.0027813671622425318, + 0.01446557231247425, + 0.01015462726354599, + 0.03245816379785538, + 0.004587990231812, + 0.010687514208257198, + 0.00807665754109621, + 0.004361707717180252, + 0.036704812198877335, + 0.007453935220837593, + 0.007939121685922146, + 0.011602560058236122, + 0.004311679862439632, + 0.00661635585129261, + 0.010241958312690258, + 0.00414844648912549, + 0.02111765928566456, + 0.018826650455594063, + 0.017751138657331467, + 0.038955576717853546, + 0.012353898026049137, + 0.017068209126591682, + 0.028023384511470795, + 0.013195769861340523, + 0.038646433502435684, + 0.016910936683416367, + 0.010446750558912754, + 0.05123224854469299, + 0.007715578190982342, + 0.010207454673945904, + 0.04025709629058838, + 0.0053316946141421795, + 0.01189572736620903, + 0.0029171197675168514, + 0.004188134800642729, + 0.002540871035307646, + 0.00414065457880497, + 0.0057376776821911335, + 0.005067003425210714, + 0.003464581910520792, + 0.004739318508654833 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 8 + ], + "selected_reward": -0.026348285377025604 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long drone shot, taken from high above, slowly zooms out, capturing a tranquil and poetic scene. Beneath the morning mist, the vast river's surface is as smooth as a mirror, a deep, dark green. A quaint, dark brown wooden boat, its hull etched with age, steadily and leisurely paddles from one side of the frame toward the center. A boatman, wearing a conical hat and a gray linen gown, slowly and steadily paddles his oars. Each silent stroke of the oars creates a widening V-shaped wave at the stern, stirring up a series of perfectly concentric ripples. These ripples spread outward smoothly and gracefully, gradually growing larger and shallower from a distinct ripple, ultimately fading silently into the vast river. The soft morning light penetrates the mist, casting mottled golden spots on the water, which together with the vague silhouettes of mountains in the distance form a flowing ink painting. The whole scene is filled with meditative tranquility and lonely beauty.", + "seed": 17583, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17583_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2286882797256112, + "vlm_logprob": -0.2286882797256112, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.2286882797256112, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4678078591823578 + ], + "opposite_logprob_tokens": [ + -1.0928078889846802 + ], + "logprob_reward": -0.4678078591823578, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4678078591823578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12918348610401154, + 0.04982222616672516, + 0.04038640856742859, + 0.045815929770469666, + 0.23038573563098907, + 0.09476108103990555, + 0.1415647715330124, + 0.10850582271814346, + 0.11715042591094971, + 0.042424075305461884 + ], + "credit_map": [ + 0.013782406225800514, + 0.01791824959218502, + 0.013889197260141373, + 0.021099790930747986, + 0.012618273496627808, + 0.016669662669301033, + 0.015681272372603416, + 0.01752464286983013, + 0.006925555877387524, + 0.0033762056846171618, + 0.0065849544480443, + 0.007448568008840084, + 0.003931785002350807, + 0.007156258448958397, + 0.005818026605993509, + 0.008580878376960754, + 0.003921593073755503, + 0.003850071458145976, + 0.005825518164783716, + 0.004767928272485733, + 0.0047759078443050385, + 0.007205708883702755, + 0.0043360586278140545, + 0.0057036252692341805, + 0.006559228058904409, + 0.00502661382779479, + 0.0055107432417571545, + 0.005340467672795057, + 0.003705705516040325, + 0.007242633029818535, + 0.004775128792971373, + 0.00765541335567832, + 0.025818707421422005, + 0.022365480661392212, + 0.031151963397860527, + 0.026332493871450424, + 0.019655145704746246, + 0.03721117600798607, + 0.01784379407763481, + 0.05000699684023857, + 0.015788767486810684, + 0.008792745880782604, + 0.010567459277808666, + 0.00822879932820797, + 0.010957987047731876, + 0.013202299363911152, + 0.011633466929197311, + 0.015589556656777859, + 0.02449280023574829, + 0.016889112070202827, + 0.01599741540849209, + 0.01456219982355833, + 0.015549524687230587, + 0.02029336802661419, + 0.020303167402744293, + 0.013477201573550701, + 0.016007404774427414, + 0.012721575796604156, + 0.012835431843996048, + 0.012701096013188362, + 0.007918432354927063, + 0.013674735091626644, + 0.008330565877258778, + 0.024316592141985893, + 0.015548087656497955, + 0.011594781652092934, + 0.01455171499401331, + 0.012486177496612072, + 0.012391328811645508, + 0.016819654032588005, + 0.017736144363880157, + 0.01602254807949066, + 0.006500099319964647, + 0.005306466482579708, + 0.005845530889928341, + 0.003723274450749159, + 0.004898986779153347, + 0.006223408505320549, + 0.004574775695800781, + 0.005351537372916937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.4678078591823578 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328195929527283 + ], + "opposite_logprob_tokens": [ + -0.7578195929527283 + ], + "logprob_reward": -0.6328195929527283, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328195929527283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028225691989064217, + 0.04432966560125351, + 0.028011519461870193, + 0.03684752807021141, + 0.3679462671279907, + 0.047272007912397385, + 0.1855936199426651, + 0.20520582795143127, + 0.021714311093091965, + 0.03485354408621788 + ], + "credit_map": [ + 0.003547322005033493, + 0.002196487272158265, + 0.003726835362613201, + 0.004719860851764679, + 0.0035592992790043354, + 0.00465917494148016, + 0.0031971305143088102, + 0.0026195812970399857, + 0.006596683524549007, + 0.0038524495903402567, + 0.005376975052058697, + 0.008859332650899887, + 0.0041246572509408, + 0.006403641775250435, + 0.005164585541933775, + 0.0039513432420790195, + 0.005443121772259474, + 0.0023890610318630934, + 0.0035036758054047823, + 0.002855262253433466, + 0.0025040421169251204, + 0.004044945351779461, + 0.002746125217527151, + 0.004525286145508289, + 0.005361094605177641, + 0.0038301595486700535, + 0.004585911985486746, + 0.004796057473868132, + 0.0032444552052766085, + 0.005345732904970646, + 0.004080132115632296, + 0.0056039863266050816, + 0.046507660299539566, + 0.03155609220266342, + 0.04894494637846947, + 0.039861734956502914, + 0.028650710359215736, + 0.058109451085329056, + 0.029687980189919472, + 0.0846276581287384, + 0.007855264469981194, + 0.005095818545669317, + 0.006007737014442682, + 0.00477477815002203, + 0.003801206359639764, + 0.007067827507853508, + 0.005790692754089832, + 0.0068786852061748505, + 0.03361448645591736, + 0.01994863711297512, + 0.019685395061969757, + 0.017839422449469566, + 0.020519398152828217, + 0.024998091161251068, + 0.019138846546411514, + 0.0298493392765522, + 0.02393503114581108, + 0.025277849286794662, + 0.02495255507528782, + 0.025540903210639954, + 0.013144600205123425, + 0.02552873268723488, + 0.017673108726739883, + 0.04915305972099304, + 0.004829375073313713, + 0.0033879380207508802, + 0.002295030513778329, + 0.002001196378841996, + 0.002156413858756423, + 0.0030186001677066088, + 0.0016833420377224684, + 0.0023424127139151096, + 0.00647730054333806, + 0.0033146198838949203, + 0.004943175707012415, + 0.002870551310479641, + 0.0033709132112562656, + 0.006245047319680452, + 0.003992775920778513, + 0.0036391604226082563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 6 + ], + "selected_reward": -0.6328195929527283 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012544064782559872 + ], + "opposite_logprob_tokens": [ + -4.38754415512085 + ], + "logprob_reward": -0.012544064782559872, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012544064782559872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15387648344039917, + 0.14729946851730347, + 0.03370889648795128, + 0.02364383637905121, + 0.15547451376914978, + 0.02976173721253872, + 0.24026179313659668, + 0.07858449965715408, + 0.0607086643576622, + 0.07668006420135498 + ], + "credit_map": [ + 0.020669642835855484, + 0.01366551872342825, + 0.01942138560116291, + 0.021450035274028778, + 0.014253347180783749, + 0.024664558470249176, + 0.018048303201794624, + 0.021703699603676796, + 0.018853407353162766, + 0.011733693070709705, + 0.013354863971471786, + 0.03608649596571922, + 0.01093860249966383, + 0.0241312924772501, + 0.017919300124049187, + 0.014281818643212318, + 0.005750434938818216, + 0.0034214260522276163, + 0.0033743150997906923, + 0.003008246188983321, + 0.0069369603879749775, + 0.005396281369030476, + 0.003185170702636242, + 0.0026360624469816685, + 0.004167551174759865, + 0.002282715868204832, + 0.0036083210725337267, + 0.0028744437731802464, + 0.0019888482056558132, + 0.0036758114583790302, + 0.0023073814809322357, + 0.002738764975219965, + 0.017655469477176666, + 0.013966606929898262, + 0.020255019888281822, + 0.01675768941640854, + 0.011776312254369259, + 0.026780374348163605, + 0.012698656879365444, + 0.03558439761400223, + 0.0035104111302644014, + 0.004017445724457502, + 0.0031206414569169283, + 0.004074264783412218, + 0.0036819030065089464, + 0.004158735740929842, + 0.00320228049531579, + 0.003996053710579872, + 0.037650998681783676, + 0.02760138362646103, + 0.0280928798019886, + 0.026896143332123756, + 0.02329646423459053, + 0.03289337083697319, + 0.03555718809366226, + 0.02827334962785244, + 0.007838058285415173, + 0.009456021711230278, + 0.009978659451007843, + 0.00950407050549984, + 0.006225041579455137, + 0.010429231449961662, + 0.007011816371232271, + 0.018141601234674454, + 0.007046231999993324, + 0.005532748997211456, + 0.007566106040030718, + 0.007155636325478554, + 0.0067372629418969154, + 0.008922817185521126, + 0.009818801656365395, + 0.007929058745503426, + 0.01545031275600195, + 0.009945990517735481, + 0.008188619278371334, + 0.008564090356230736, + 0.009559431113302708, + 0.010646235197782516, + 0.007745944894850254, + 0.006579441484063864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.012544064782559872 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607646234333515 + ], + "opposite_logprob_tokens": [ + -4.141076564788818 + ], + "logprob_reward": -0.01607646234333515, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607646234333515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12102195620536804, + 0.1552371084690094, + 0.034087326377630234, + 0.022166334092617035, + 0.09746815264225006, + 0.03779624029994011, + 0.1371070295572281, + 0.07424908131361008, + 0.05120145529508591, + 0.26966536045074463 + ], + "credit_map": [ + 0.01496024988591671, + 0.014995682053267956, + 0.015693161636590958, + 0.018084926530718803, + 0.012842499651014805, + 0.018368905410170555, + 0.01130435150116682, + 0.014772170223295689, + 0.024345779791474342, + 0.013021595776081085, + 0.01674284227192402, + 0.03057723119854927, + 0.01219645794481039, + 0.02683275379240513, + 0.01663965731859207, + 0.014880793169140816, + 0.0035023600794374943, + 0.0032217518892139196, + 0.003491183975711465, + 0.003589659696444869, + 0.006880471482872963, + 0.0053853136487305164, + 0.003915654961019754, + 0.004100930877029896, + 0.003773436415940523, + 0.0023242111783474684, + 0.0026853266172111034, + 0.0025039264000952244, + 0.001802294049412012, + 0.0033825498539954424, + 0.0016509437700733542, + 0.0040436433628201485, + 0.01265797670930624, + 0.007898256182670593, + 0.01179787702858448, + 0.008024720475077629, + 0.007691462058573961, + 0.01607806794345379, + 0.007643082644790411, + 0.025676708668470383, + 0.004648946225643158, + 0.004095281008630991, + 0.004855123348534107, + 0.0030150734819471836, + 0.003281734883785248, + 0.006058546714484692, + 0.0030763994436711073, + 0.00876513123512268, + 0.03242460638284683, + 0.020799526944756508, + 0.01405041292309761, + 0.010874243453145027, + 0.010493472218513489, + 0.02038857340812683, + 0.010147360153496265, + 0.017928825691342354, + 0.007460766471922398, + 0.008750137872993946, + 0.008087091147899628, + 0.00903153046965599, + 0.0071089137345552444, + 0.008515560068190098, + 0.010208736173808575, + 0.015086336992681026, + 0.00649589067324996, + 0.006101729813963175, + 0.006951464340090752, + 0.0044235545210540295, + 0.004865733440965414, + 0.007948617450892925, + 0.005638107191771269, + 0.008776355534791946, + 0.06453211605548859, + 0.034454409033060074, + 0.03088402934372425, + 0.01858137734234333, + 0.028097184374928474, + 0.04187021031975746, + 0.023922443389892578, + 0.027323586866259575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.01607646234333515 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193419367074966 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193419367074966, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193419367074966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07378067821264267, + 0.04058196023106575, + 0.02097906544804573, + 0.05819189175963402, + 0.5697990655899048, + 0.05121180787682533, + 0.07388333976268768, + 0.03816523030400276, + 0.053160712122917175, + 0.020246341824531555 + ], + "credit_map": [ + 0.011005343869328499, + 0.009398975409567356, + 0.00860627368092537, + 0.007787114474922419, + 0.01232105866074562, + 0.009755280800163746, + 0.007154142949730158, + 0.007752492558211088, + 0.00563904969021678, + 0.003430688753724098, + 0.005599708762019873, + 0.004814455285668373, + 0.0040903957560658455, + 0.007614172529429197, + 0.0037908204831182957, + 0.005602668505162001, + 0.0016286424361169338, + 0.001726852497085929, + 0.003524367930367589, + 0.0018683848902583122, + 0.0034245226997882128, + 0.004173989873379469, + 0.0018834868678823113, + 0.002748817903921008, + 0.0069307005032896996, + 0.005575213115662336, + 0.007495546247810125, + 0.0065248156897723675, + 0.006289640907198191, + 0.008349062874913216, + 0.005666503217071295, + 0.011360405944287777, + 0.07164870202541351, + 0.049957938492298126, + 0.07573890686035156, + 0.0474344827234745, + 0.04027760028839111, + 0.09864497184753418, + 0.040251415222883224, + 0.1458449810743332, + 0.008854844607412815, + 0.00631700037047267, + 0.00532903615385294, + 0.004590964410454035, + 0.0042518507689237595, + 0.00638610078021884, + 0.005012141074985266, + 0.01046986598521471, + 0.011721502989530563, + 0.008697724901139736, + 0.007780376821756363, + 0.008329970762133598, + 0.006682754494249821, + 0.010015562176704407, + 0.008958130143582821, + 0.011697319336235523, + 0.006771839689463377, + 0.005011359695345163, + 0.004511843901127577, + 0.003334997920319438, + 0.00446974067017436, + 0.004881794098764658, + 0.005180348176509142, + 0.004003308713436127, + 0.009233890101313591, + 0.006106268148869276, + 0.005660437047481537, + 0.007231655064970255, + 0.005249991547316313, + 0.007957610301673412, + 0.0069110500626266, + 0.004809808451682329, + 0.00296996277756989, + 0.0022317825350910425, + 0.003124731592833996, + 0.0021090847440063953, + 0.002297009574249387, + 0.0031327996402978897, + 0.0020512710325419903, + 0.0023296992294490337 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.014193419367074966 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long drone shot, taken from high above, slowly zooms out, capturing a tranquil and poetic scene. Beneath the morning mist, the vast river's surface is as smooth as a mirror, a deep, dark green. A quaint, dark brown wooden boat, its hull etched with age, steadily and leisurely paddles from one side of the frame toward the center. A boatman, wearing a conical hat and a gray linen gown, slowly and steadily paddles his oars. Each silent stroke of the oars creates a widening V-shaped wave at the stern, stirring up a series of perfectly concentric ripples. These ripples spread outward smoothly and gracefully, gradually growing larger and shallower from a distinct ripple, ultimately fading silently into the vast river. The soft morning light penetrates the mist, casting mottled golden spots on the water, which together with the vague silhouettes of mountains in the distance form a flowing ink painting. The whole scene is filled with meditative tranquility and lonely beauty.", + "seed": 17584, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17584_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1565138442441821, + "vlm_logprob": -0.1565138442441821, + "vlm_yes_no_margin": 3.425, + "vlm_token_credit": -0.1565138442441821, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05890776962041855 + ], + "opposite_logprob_tokens": [ + -3.1839077472686768 + ], + "logprob_reward": -0.05890776962041855, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05890776962041855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12498622387647629, + 0.046722106635570526, + 0.0702390968799591, + 0.05955742299556732, + 0.3728770315647125, + 0.045245856046676636, + 0.07135194540023804, + 0.04473283141851425, + 0.06472029536962509, + 0.09956712275743484 + ], + "credit_map": [ + 0.019801290705800056, + 0.027222344651818275, + 0.017511310055851936, + 0.00903412513434887, + 0.010887949727475643, + 0.015472576022148132, + 0.011732319369912148, + 0.013324311003088951, + 0.006676294840872288, + 0.010331997647881508, + 0.005547752603888512, + 0.003815773408859968, + 0.005240348633378744, + 0.006556652951985598, + 0.00434225145727396, + 0.004211034160107374, + 0.00618140259757638, + 0.011609380133450031, + 0.009906415827572346, + 0.004474143963307142, + 0.007543359883129597, + 0.01866716705262661, + 0.005005931481719017, + 0.006851290352642536, + 0.007378204260021448, + 0.01818828657269478, + 0.0031732888892292976, + 0.00381641392596066, + 0.008486890234053135, + 0.004039901774376631, + 0.005091547034680843, + 0.009382893331348896, + 0.04085356369614601, + 0.12035345286130905, + 0.016052592545747757, + 0.016354994848370552, + 0.06019968539476395, + 0.019331974908709526, + 0.031090067699551582, + 0.06864068657159805, + 0.004442117642611265, + 0.005311568267643452, + 0.005307495594024658, + 0.007815171964466572, + 0.005613618530333042, + 0.005161950830370188, + 0.006199374794960022, + 0.005394557490944862, + 0.004573649261146784, + 0.010398534126579762, + 0.007774501107633114, + 0.005095901433378458, + 0.011293663643300533, + 0.017770415171980858, + 0.00637012766674161, + 0.00807515624910593, + 0.0031677186489105225, + 0.004428334068506956, + 0.005410394631326199, + 0.011086995713412762, + 0.004463633522391319, + 0.006535850465297699, + 0.004950906150043011, + 0.004689001012593508, + 0.004569092765450478, + 0.008989213034510612, + 0.007067108526825905, + 0.011288847774267197, + 0.007792504504323006, + 0.007570796646177769, + 0.010073156096041203, + 0.007369573228061199, + 0.008572006598114967, + 0.012480109930038452, + 0.009611791931092739, + 0.012763281352818012, + 0.014094912447035313, + 0.012180333025753498, + 0.018406618386507034, + 0.01145806536078453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.05890776962041855 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.693326473236084 + ], + "opposite_logprob_tokens": [ + -0.693326473236084 + ], + "logprob_reward": -0.693326473236084, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.693326473236084, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10865345597267151, + 0.045706089586019516, + 0.0981021523475647, + 0.020430391654372215, + 0.3193971514701843, + 0.014280462637543678, + 0.06505480408668518, + 0.03179969638586044, + 0.18200582265853882, + 0.11457002907991409 + ], + "credit_map": [ + 0.02120944857597351, + 0.033457107841968536, + 0.006216687150299549, + 0.0068678660318255424, + 0.01058114878833294, + 0.00899434182792902, + 0.007255212403833866, + 0.01407164242118597, + 0.004757720045745373, + 0.007883494719862938, + 0.005867445841431618, + 0.003993974532932043, + 0.005493301898241043, + 0.009528541006147861, + 0.003879146184772253, + 0.0043024662882089615, + 0.013319827616214752, + 0.03096575289964676, + 0.00640771584585309, + 0.006385329645127058, + 0.012060100212693214, + 0.006706054322421551, + 0.008573278784751892, + 0.01368409488350153, + 0.002478731330484152, + 0.0042579700239002705, + 0.0021846021991223097, + 0.0017380270874127746, + 0.002499597379937768, + 0.0021167437080293894, + 0.0019595979247242212, + 0.00319512072019279, + 0.034095942974090576, + 0.10161142796278, + 0.013337965123355389, + 0.015126707032322884, + 0.05257214605808258, + 0.01594710350036621, + 0.026083316653966904, + 0.060622524470090866, + 0.00144335743971169, + 0.0011998119298368692, + 0.001907396363094449, + 0.0015806397423148155, + 0.0017217658460140228, + 0.0024587379302829504, + 0.002215434331446886, + 0.0017533197533339262, + 0.004132981877774, + 0.009614990092813969, + 0.006297504995018244, + 0.0073563819751143456, + 0.009598356671631336, + 0.014048612676560879, + 0.006699822843074799, + 0.007306150160729885, + 0.0019097407348453999, + 0.003061780473217368, + 0.003746551228687167, + 0.005636800080537796, + 0.003996663726866245, + 0.00538264773786068, + 0.004588654264807701, + 0.003476856742054224, + 0.013958181254565716, + 0.026588594540953636, + 0.021750446408987045, + 0.024349305778741837, + 0.03295664116740227, + 0.03189396113157272, + 0.01699179783463478, + 0.013516891747713089, + 0.011757687665522099, + 0.010776747018098831, + 0.012139824219048023, + 0.01605585590004921, + 0.014667954295873642, + 0.015355082228779793, + 0.019726520404219627, + 0.01409036386758089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.693326473236084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253794603049755 + ], + "opposite_logprob_tokens": [ + -5.255253791809082 + ], + "logprob_reward": -0.005253794603049755, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253794603049755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1385030597448349, + 0.28880664706230164, + 0.03639579191803932, + 0.08861251175403595, + 0.1438002586364746, + 0.04418215528130531, + 0.03608724847435951, + 0.0514477863907814, + 0.10723119229078293, + 0.06493332237005234 + ], + "credit_map": [ + 0.013622635044157505, + 0.015277862548828125, + 0.024605821818113327, + 0.015559851191937923, + 0.009673554450273514, + 0.018948465585708618, + 0.018229009583592415, + 0.02258586697280407, + 0.0508827269077301, + 0.09959881752729416, + 0.01645885407924652, + 0.016984671354293823, + 0.02868698164820671, + 0.015166057273745537, + 0.021591229364275932, + 0.03943730518221855, + 0.0036712803412228823, + 0.005761777050793171, + 0.0037262358237057924, + 0.0038403556682169437, + 0.0035516449715942144, + 0.006793714128434658, + 0.003741527907550335, + 0.0053092543967068195, + 0.011341373436152935, + 0.02863333933055401, + 0.004426596220582724, + 0.003940093331038952, + 0.013378314673900604, + 0.0058594681322574615, + 0.006818963680416346, + 0.014214362017810345, + 0.014190162532031536, + 0.04072503373026848, + 0.009198375046253204, + 0.009113138541579247, + 0.021285517141222954, + 0.014304903335869312, + 0.01246589794754982, + 0.02251722849905491, + 0.0026706571225076914, + 0.006028245203197002, + 0.00509018637239933, + 0.006978559773415327, + 0.0050566899590194225, + 0.006627402734011412, + 0.006902377121150494, + 0.004828033968806267, + 0.0026000633370131254, + 0.005975381471216679, + 0.003964344039559364, + 0.0026826094835996628, + 0.00554371252655983, + 0.007830816321074963, + 0.003014683024957776, + 0.0044756378047168255, + 0.0031704450957477093, + 0.0047632381319999695, + 0.006536426488310099, + 0.011133197695016861, + 0.006259100046008825, + 0.009509248659014702, + 0.005650283303111792, + 0.00442584790289402, + 0.007217461243271828, + 0.010775682516396046, + 0.013919051736593246, + 0.021549660712480545, + 0.010404025204479694, + 0.013392381370067596, + 0.01753774657845497, + 0.01243518479168415, + 0.006215700879693031, + 0.009137485176324844, + 0.008886460214853287, + 0.008252184838056564, + 0.008062462322413921, + 0.00851250346750021, + 0.009418664500117302, + 0.006447862833738327 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.005253794603049755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543593533337116 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543593533337116, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543593533337116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15309561789035797, + 0.08022929728031158, + 0.06670788675546646, + 0.060148414224386215, + 0.3280624747276306, + 0.014676649123430252, + 0.019115034490823746, + 0.07563333958387375, + 0.16157829761505127, + 0.04075298830866814 + ], + "credit_map": [ + 0.019261514768004417, + 0.027303004637360573, + 0.015739165246486664, + 0.017665604129433632, + 0.013831193558871746, + 0.013658300042152405, + 0.01818515732884407, + 0.02745167165994644, + 0.01215244922786951, + 0.023223986849188805, + 0.008144492283463478, + 0.004813477862626314, + 0.00796485599130392, + 0.010541186667978764, + 0.005814505275338888, + 0.0075743431225419044, + 0.00777426129207015, + 0.01754244975745678, + 0.006060033105313778, + 0.0055594234727323055, + 0.007933237589895725, + 0.00704586785286665, + 0.006924229208379984, + 0.007868378423154354, + 0.007694311905652285, + 0.01886069029569626, + 0.00369058339856565, + 0.0034587124828249216, + 0.00894967745989561, + 0.004343621898442507, + 0.004849480930715799, + 0.008301332592964172, + 0.035878125578165054, + 0.10599339753389359, + 0.014043034985661507, + 0.0140389958396554, + 0.05379883199930191, + 0.016947412863373756, + 0.0268698763102293, + 0.06049276515841484, + 0.001710834912955761, + 0.0018093774560838938, + 0.0018146425718441606, + 0.0018419313710182905, + 0.001607044250704348, + 0.0025127932894974947, + 0.001980146858841181, + 0.0013998785289004445, + 0.0016251466004177928, + 0.0021536690182983875, + 0.0029352426063269377, + 0.0030796348582953215, + 0.0020450332667678595, + 0.0016769117210060358, + 0.003147789975628257, + 0.002451605862006545, + 0.0050421711057424545, + 0.006837671622633934, + 0.01040731742978096, + 0.01621536910533905, + 0.00923998560756445, + 0.011420764029026031, + 0.009982059709727764, + 0.006487993523478508, + 0.012449546717107296, + 0.019017735496163368, + 0.017611432820558548, + 0.03271595388650894, + 0.015648694708943367, + 0.015144388191401958, + 0.029149290174245834, + 0.019841257482767105, + 0.002969682915136218, + 0.0041564893908798695, + 0.004880646709352732, + 0.0044283075258135796, + 0.006560659501701593, + 0.00803295336663723, + 0.005846348591148853, + 0.0038778968155384064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.012543593533337116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23538583517074585, + 0.23768581449985504, + 0.062334805727005005, + 0.09326691925525665, + 0.058575354516506195, + 0.025910083204507828, + 0.03678714483976364, + 0.05696006864309311, + 0.08607038110494614, + 0.10702360421419144 + ], + "credit_map": [ + 0.026573359966278076, + 0.0412910059094429, + 0.03161943331360817, + 0.011640735901892185, + 0.015469495207071304, + 0.025271417573094368, + 0.02179204300045967, + 0.061728350818157196, + 0.02954472042620182, + 0.037046197801828384, + 0.037944499403238297, + 0.012173568829894066, + 0.022581379860639572, + 0.05306567624211311, + 0.016541864722967148, + 0.02878793142735958, + 0.005880493205040693, + 0.0059469412080943584, + 0.009257320314645767, + 0.0044109937734901905, + 0.006858255714178085, + 0.018564734607934952, + 0.004926016554236412, + 0.006490051746368408, + 0.011905607767403126, + 0.03052276186645031, + 0.005412769038230181, + 0.0038894694298505783, + 0.013693980872631073, + 0.006691301241517067, + 0.006629487965255976, + 0.014521540142595768, + 0.005428621079772711, + 0.007320762611925602, + 0.008650892414152622, + 0.0073472014628350735, + 0.007406826131045818, + 0.007949449121952057, + 0.007906842045485973, + 0.00656475592404604, + 0.002665409352630377, + 0.0027070275973528624, + 0.0029579177498817444, + 0.00360320508480072, + 0.0036221209447830915, + 0.003110631601884961, + 0.003753447672352195, + 0.003490324132144451, + 0.0034751566126942635, + 0.0049300892278552055, + 0.004713168367743492, + 0.0029392484575510025, + 0.005772706121206284, + 0.007139549590647221, + 0.003841439029201865, + 0.00397578626871109, + 0.004804270341992378, + 0.005707968957722187, + 0.007927075028419495, + 0.011847629211843014, + 0.006280240137130022, + 0.007135544903576374, + 0.007469812408089638, + 0.0057875229977071285, + 0.0064743030816316605, + 0.006248004734516144, + 0.011604255065321922, + 0.019476469606161118, + 0.008520129136741161, + 0.00783495232462883, + 0.015011542476713657, + 0.010900726541876793, + 0.009044860489666462, + 0.014282063581049442, + 0.010261982679367065, + 0.01151277031749487, + 0.01621718890964985, + 0.018500052392482758, + 0.015893055126070976, + 0.01131162978708744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.012537590228021145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "long drone shot, taken from high above, slowly zooms out, capturing a tranquil and poetic scene. Beneath the morning mist, the vast river's surface is as smooth as a mirror, a deep, dark green. A quaint, dark brown wooden boat, its hull etched with age, steadily and leisurely paddles from one side of the frame toward the center. A boatman, wearing a conical hat and a gray linen gown, slowly and steadily paddles his oars. Each silent stroke of the oars creates a widening V-shaped wave at the stern, stirring up a series of perfectly concentric ripples. These ripples spread outward smoothly and gracefully, gradually growing larger and shallower from a distinct ripple, ultimately fading silently into the vast river. The soft morning light penetrates the mist, casting mottled golden spots on the water, which together with the vague silhouettes of mountains in the distance form a flowing ink painting. The whole scene is filled with meditative tranquility and lonely beauty.", + "seed": 17585, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_6/index_5899_seed_17585_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.076423195656389, + "vlm_logprob": -0.076423195656389, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.076423195656389, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04371803626418114 + ], + "opposite_logprob_tokens": [ + -3.793718099594116 + ], + "logprob_reward": -0.04371803626418114, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04371803626418114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21729333698749542, + 0.11110630631446838, + 0.024358797818422318, + 0.052970632910728455, + 0.026661869138479233, + 0.10786257684230804, + 0.05308638513088226, + 0.02298167161643505, + 0.2634838819503784, + 0.12019447982311249 + ], + "credit_map": [ + 0.023587824776768684, + 0.011413516476750374, + 0.060472749173641205, + 0.008747397921979427, + 0.008606472983956337, + 0.0193009115755558, + 0.01045303326100111, + 0.07471144944429398, + 0.007150923367589712, + 0.0036853402853012085, + 0.03817341476678848, + 0.00799258891493082, + 0.004426080733537674, + 0.008572954684495926, + 0.0058035170659422874, + 0.03530149161815643, + 0.002041471190750599, + 0.0026085858698934317, + 0.003779481165111065, + 0.003333575092256069, + 0.0018195579759776592, + 0.0042343963868916035, + 0.002249393379315734, + 0.004292337689548731, + 0.005844785366207361, + 0.006226003635674715, + 0.010999818332493305, + 0.004863781854510307, + 0.0031825483310967684, + 0.007497584912925959, + 0.004067550878971815, + 0.010288563556969166, + 0.004168768413364887, + 0.003493962110951543, + 0.0032584560103714466, + 0.0029991043265908957, + 0.001989404670894146, + 0.004454046953469515, + 0.0026043830439448357, + 0.003693743608891964, + 0.010892234742641449, + 0.013840685598552227, + 0.01764569990336895, + 0.007061083801090717, + 0.009138645604252815, + 0.02707175351679325, + 0.005208163522183895, + 0.01700432039797306, + 0.0026289040688425303, + 0.017660047858953476, + 0.006905887275934219, + 0.003479917300865054, + 0.008181419223546982, + 0.007991713471710682, + 0.0019522320944815874, + 0.0042862663976848125, + 0.001609933446161449, + 0.004139013588428497, + 0.005128256045281887, + 0.002504815114662051, + 0.0027208311948925257, + 0.0019991269800812006, + 0.0018321797251701355, + 0.0030475170351564884, + 0.01112019456923008, + 0.03413962200284004, + 0.057675570249557495, + 0.011707634665071964, + 0.03277205303311348, + 0.09088615328073502, + 0.004260167479515076, + 0.02092248946428299, + 0.008112736977636814, + 0.00936001818627119, + 0.009461943991482258, + 0.00696554547175765, + 0.036962639540433884, + 0.030923141166567802, + 0.005717855878174305, + 0.012690612114965916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.04371803626418114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31342923641204834 + ], + "opposite_logprob_tokens": [ + -1.3134292364120483 + ], + "logprob_reward": -0.31342923641204834, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31342923641204834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14209286868572235, + 0.03751732036471367, + 0.03572136536240578, + 0.05656096339225769, + 0.023801680654287338, + 0.14287447929382324, + 0.23944582045078278, + 0.08510580658912659, + 0.07314295321702957, + 0.1637367457151413 + ], + "credit_map": [ + 0.008815779350697994, + 0.009428368881344795, + 0.029034819453954697, + 0.009249359369277954, + 0.011003226041793823, + 0.03425179049372673, + 0.009394781664013863, + 0.030914749950170517, + 0.0022392466198652983, + 0.0017930467147380114, + 0.011969713494181633, + 0.0031541846692562103, + 0.001786593347787857, + 0.0036732442677021027, + 0.0026835775934159756, + 0.010217715054750443, + 0.0037134913727641106, + 0.0026324372738599777, + 0.0065910955891013145, + 0.003253090428188443, + 0.0018308808794245124, + 0.008851950988173485, + 0.0026833817828446627, + 0.006165036465972662, + 0.006158972159028053, + 0.005377489607781172, + 0.009837732650339603, + 0.005563502199947834, + 0.003514850977808237, + 0.010932974517345428, + 0.0048756045289337635, + 0.010299832560122013, + 0.003996448591351509, + 0.0027602873742580414, + 0.003658836940303445, + 0.0020271686371415854, + 0.0017737924354150891, + 0.004121304024010897, + 0.001702547655440867, + 0.0037612956948578358, + 0.011406229808926582, + 0.020214509218931198, + 0.02206510864198208, + 0.01008750218898058, + 0.01071134302765131, + 0.03217214718461037, + 0.008752331137657166, + 0.027465293183922768, + 0.0129292793571949, + 0.08296969532966614, + 0.03823043033480644, + 0.008996078744530678, + 0.03637557476758957, + 0.035638581961393356, + 0.006053046323359013, + 0.018253130838274956, + 0.006055073346942663, + 0.01357156690210104, + 0.012381418608129025, + 0.01074038166552782, + 0.0082350242882967, + 0.017041753977537155, + 0.0074681430123746395, + 0.009612446650862694, + 0.003978626802563667, + 0.0097184544429183, + 0.011233811266720295, + 0.00640774704515934, + 0.008830158039927483, + 0.022281276062130928, + 0.005376866087317467, + 0.005316013004630804, + 0.009274792857468128, + 0.014602994546294212, + 0.01512330211699009, + 0.020278487354516983, + 0.035181496292352676, + 0.036451783031225204, + 0.01797758787870407, + 0.01484630536288023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 5 + ], + "selected_reward": -0.31342923641204834 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255098920315504 + ], + "opposite_logprob_tokens": [ + -5.255255222320557 + ], + "logprob_reward": -0.005255098920315504, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255098920315504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3661314845085144, + 0.017623750492930412, + 0.012845157645642757, + 0.02425217814743519, + 0.03315415233373642, + 0.05608586221933365, + 0.052006810903549194, + 0.09960214793682098, + 0.25837892293930054, + 0.07991958409547806 + ], + "credit_map": [ + 0.050275903195142746, + 0.017143765464425087, + 0.10045076161623001, + 0.010652217082679272, + 0.01379383448511362, + 0.025198079645633698, + 0.016756577417254448, + 0.13186031579971313, + 0.0013791375095024705, + 0.001783984713256359, + 0.003157887374982238, + 0.0024602310732007027, + 0.0013258927501738071, + 0.0023781289346516132, + 0.0020595104433596134, + 0.003078975947573781, + 0.0009649862768128514, + 0.0010138442739844322, + 0.0024044038727879524, + 0.0019906230736523867, + 0.0008162054000422359, + 0.0016928723780438304, + 0.0014332360588014126, + 0.002528984099626541, + 0.002364115556702018, + 0.0032077599316835403, + 0.002723176032304764, + 0.002311763586476445, + 0.0014962442219257355, + 0.004994719754904509, + 0.002210599137470126, + 0.0049437968991696835, + 0.002950790338218212, + 0.007429399993270636, + 0.006176545284688473, + 0.002396579599007964, + 0.0033056908287107944, + 0.00551410298794508, + 0.0016763813327997923, + 0.0037046552170068026, + 0.005702298134565353, + 0.005956949200481176, + 0.00968726072460413, + 0.0036942462902516127, + 0.004293761216104031, + 0.013525795191526413, + 0.002938668243587017, + 0.010286872275173664, + 0.004592880606651306, + 0.00966410618275404, + 0.013407211750745773, + 0.004031119402498007, + 0.004927055444568396, + 0.005445262882858515, + 0.0030397505033761263, + 0.006899415049701929, + 0.004704682622104883, + 0.03016914613544941, + 0.01116801705211401, + 0.008701459504663944, + 0.015751052647829056, + 0.014384696260094643, + 0.0074611385352909565, + 0.00726194167509675, + 0.01093475241214037, + 0.033307529985904694, + 0.05809410661458969, + 0.01232435368001461, + 0.03040793538093567, + 0.08728592842817307, + 0.0051733339205384254, + 0.02085094340145588, + 0.006288716569542885, + 0.00786521378904581, + 0.007287752348929644, + 0.007845006883144379, + 0.020510323345661163, + 0.017188653349876404, + 0.006093532312661409, + 0.006840369664132595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.005255098920315504 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075802147388458 + ], + "opposite_logprob_tokens": [ + -4.511075973510742 + ], + "logprob_reward": -0.011075802147388458, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075802147388458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15163813531398773, + 0.025473367422819138, + 0.0068891230039298534, + 0.011486525647342205, + 0.010406003333628178, + 0.0474676713347435, + 0.5852576494216919, + 0.09801871329545975, + 0.009007197804749012, + 0.05435552820563316 + ], + "credit_map": [ + 0.016475873067975044, + 0.0072396728210151196, + 0.04230358079075813, + 0.006017104256898165, + 0.006358502898365259, + 0.01055603101849556, + 0.00871527660638094, + 0.053972095251083374, + 0.003375762840732932, + 0.0022652032785117626, + 0.0035165564622730017, + 0.0027382224798202515, + 0.0023301485925912857, + 0.004175453446805477, + 0.0020315642468631268, + 0.005040458403527737, + 0.0005355667090043426, + 0.0005605335463769734, + 0.0012607218232005835, + 0.0008615927072241902, + 0.000494906329549849, + 0.0010825135977938771, + 0.0007703225710429251, + 0.0013229663018137217, + 0.0011744871735572815, + 0.0016827912768349051, + 0.0012954039266332984, + 0.0010942775988951325, + 0.0008209916413761675, + 0.0019961169455200434, + 0.0011584501480683684, + 0.002264008391648531, + 0.0010860504116863012, + 0.0016341968439519405, + 0.0012959735468029976, + 0.0011059946846216917, + 0.0009614620357751846, + 0.002122674137353897, + 0.0007938658236525953, + 0.0014057856751605868, + 0.004778928589075804, + 0.006735621020197868, + 0.007059101946651936, + 0.002974933013319969, + 0.0035747389774769545, + 0.01201997883617878, + 0.002397837582975626, + 0.007926533930003643, + 0.01539669930934906, + 0.1868327260017395, + 0.1361374706029892, + 0.025905774906277657, + 0.08162029832601547, + 0.0679023265838623, + 0.01081777922809124, + 0.06064465269446373, + 0.004513769410550594, + 0.029046250507235527, + 0.012425899505615234, + 0.008746462874114513, + 0.01456077117472887, + 0.014502027072012424, + 0.007089382037520409, + 0.007134157232940197, + 0.0006405683816410601, + 0.0020139727275818586, + 0.0010560473892837763, + 0.0009569857502356172, + 0.0009894751710817218, + 0.001524674124084413, + 0.0010309283388778567, + 0.0007945467950776219, + 0.004283508751541376, + 0.005577996838837862, + 0.006686246022582054, + 0.007046309765428305, + 0.010451790876686573, + 0.009076318703591824, + 0.004975341726094484, + 0.006258019711822271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.011075802147388458 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a high angle looking down on the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera slowly zoom out during the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a dark brown wooden boat with an aged hull visible paddling from one side of the frame toward the center?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the boatman wearing a conical hat and a gray linen gown while paddling?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the oar strokes create widening V-shaped waves and concentric ripples that spread outward on the water?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637804538011551 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637804538011551, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637804538011551, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.039392102509737015, + 0.016144651919603348, + 0.006603937596082687, + 0.022565340623259544, + 0.020558100193738937, + 0.02060813643038273, + 0.16424120962619781, + 0.014282825402915478, + 0.6643997430801392, + 0.0312039777636528 + ], + "credit_map": [ + 0.0022816259879618883, + 0.0017954062204807997, + 0.011087119579315186, + 0.0027224153745919466, + 0.001942350878380239, + 0.005373928230255842, + 0.002842563670128584, + 0.011346687562763691, + 0.0018285741098225117, + 0.0010639819083735347, + 0.0029276334680616856, + 0.0017178113339468837, + 0.0013353113317862153, + 0.0032876383047550917, + 0.0014724580105394125, + 0.0025112417060881853, + 0.0005701882764697075, + 0.00043853287934325635, + 0.0012892743106931448, + 0.000850708456709981, + 0.00042185228085145354, + 0.001143783563748002, + 0.0006711098831146955, + 0.001218487392179668, + 0.002709699561819434, + 0.0035257351119071245, + 0.0029304197523742914, + 0.0015852084616199136, + 0.001538677141070366, + 0.005659958813339472, + 0.0013287034817039967, + 0.0032869363203644753, + 0.0018456155667081475, + 0.0034854889381676912, + 0.0030867145396769047, + 0.0014324975199997425, + 0.001468527945689857, + 0.005250983405858278, + 0.0011154187377542257, + 0.002872851677238941, + 0.0021237109322100878, + 0.0038778509479016066, + 0.0028968313708901405, + 0.0021178426686674356, + 0.0015902518061921, + 0.00324210780672729, + 0.0016460687620565295, + 0.0031134700402617455, + 0.004769276827573776, + 0.04915471747517586, + 0.04094390198588371, + 0.0081822220236063, + 0.021753549575805664, + 0.017222536727786064, + 0.003755047917366028, + 0.01845994032919407, + 0.001239315839484334, + 0.001835864270105958, + 0.0026847843546420336, + 0.0018195417942479253, + 0.0013913626316934824, + 0.0025148987770080566, + 0.0011704275384545326, + 0.0016266288002952933, + 0.028069699183106422, + 0.0861116573214531, + 0.14875298738479614, + 0.029883313924074173, + 0.08047811686992645, + 0.22744624316692352, + 0.01021361444145441, + 0.053444065153598785, + 0.00176506326533854, + 0.0026625257451087236, + 0.0019203796982765198, + 0.001407759147696197, + 0.010958842001855373, + 0.009219986386597157, + 0.0016358596039935946, + 0.0016335607506334782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.008637804538011551 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_tensor([1640, 1640])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_tensor([1640, 1640])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c8a6a0bab55acc6eab3048255ff389ada51c80af --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_tensor([1640, 1640])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, medium shot, positioned just above the ground, captures a soccer player striding across a lush green field under the bright midday sun. The camera then executes a rapid, powerful zoom, darting forward to frame an extreme close-up focused entirely on their feet and the ball. The player's muscular legs, clad in royal blue athletic shorts and matching knee-high socks, power across the field. Their feet, a blur of movement within the intricately designed, laceless studs gleaming in white material, execute a dazzling display of complex footwork. The classic black-and-white paneled ball appears tethered to their feet, being relentlessly yet deftly nudged, tapped, and rolled with lightning-fast precision. Each explosive stride and sharp change of direction drives the ball into the pristine turf, sending visible dark soil and blades of vibrant green grass blasting into the air, suspended for a moment in the sunlight. High-contrast lighting casts sharp, rapidly moving shadows across the neatly manicured grass, creating a dynamic and high-energy atmosphere imbued with intense focus and raw kinetic energy.", + "seed": 921, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_921_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0747716114623472, + "vlm_logprob": -0.0747716114623472, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.0747716114623472, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04134664312005043 + ], + "opposite_logprob_tokens": [ + -3.416346549987793 + ], + "logprob_reward": -0.04134664312005043, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04134664312005043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.082004614174366, + 0.04818527400493622, + 0.028471553698182106, + 0.0640396922826767, + 0.2681852877140045, + 0.043714117258787155, + 0.29401737451553345, + 0.016715865582227707, + 0.01766296476125717, + 0.13700328767299652 + ], + "credit_map": [ + 0.012793357484042645, + 0.010320262052118778, + 0.013228971511125565, + 0.012201882898807526, + 0.007850999012589455, + 0.009022202342748642, + 0.0068612247705459595, + 0.009725712239742279, + 0.00686295609921217, + 0.004492013715207577, + 0.011124265380203724, + 0.00558624230325222, + 0.003205864457413554, + 0.004441544413566589, + 0.005714148283004761, + 0.006758239585906267, + 0.003676628228276968, + 0.003487899899482727, + 0.004949639551341534, + 0.0031632205937057734, + 0.002003675326704979, + 0.003995276056230068, + 0.0038314734119921923, + 0.003363742958754301, + 0.008481052704155445, + 0.008251904509961605, + 0.01010682713240385, + 0.00754729425534606, + 0.008042710833251476, + 0.00817340612411499, + 0.006779383402317762, + 0.006657111924141645, + 0.06534935534000397, + 0.01607581600546837, + 0.021678803488612175, + 0.012651551514863968, + 0.024817895144224167, + 0.07825696468353271, + 0.013719572685658932, + 0.035635337233543396, + 0.004069426562637091, + 0.005145564675331116, + 0.006391525734215975, + 0.00529048778116703, + 0.004344608169049025, + 0.005699215456843376, + 0.007385158445686102, + 0.0053881327621638775, + 0.024249626323580742, + 0.022696346044540405, + 0.010310199111700058, + 0.012908313423395157, + 0.013216962106525898, + 0.10198390483856201, + 0.018827158957719803, + 0.0898248553276062, + 0.001992640784010291, + 0.0014287703670561314, + 0.0021634178701788187, + 0.001961729722097516, + 0.0019496980821713805, + 0.00221726531162858, + 0.002316004829481244, + 0.0026863368693739176, + 0.0032694151159375906, + 0.001963858725503087, + 0.002796388929709792, + 0.0018849280895665288, + 0.0012708631111308932, + 0.0017968307947739959, + 0.0024994274135679007, + 0.0021812529303133488, + 0.022283241152763367, + 0.013976279646158218, + 0.020052725449204445, + 0.01131709385663271, + 0.014460362493991852, + 0.02214890904724598, + 0.011973532848060131, + 0.02079114504158497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 9 + ], + "selected_reward": -0.04134664312005043 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981521561741829 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.02981521561741829, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981521561741829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11074628680944443, + 0.13107503950595856, + 0.14126056432724, + 0.0401209220290184, + 0.054137539118528366, + 0.021462082862854004, + 0.08680711686611176, + 0.06335937976837158, + 0.09313272684812546, + 0.25789836049079895 + ], + "credit_map": [ + 0.011751560494303703, + 0.012799285352230072, + 0.027563825249671936, + 0.01743587851524353, + 0.006945667788386345, + 0.00971500389277935, + 0.011660877615213394, + 0.012874191626906395, + 0.0107842106372118, + 0.018714889883995056, + 0.027135349810123444, + 0.01830846630036831, + 0.01286755409091711, + 0.0128810815513134, + 0.012798517011106014, + 0.017584970220923424, + 0.018142372369766235, + 0.021764930337667465, + 0.018593279644846916, + 0.018038121983408928, + 0.010545855388045311, + 0.02127755433320999, + 0.01752440817654133, + 0.015374038368463516, + 0.004105089232325554, + 0.004387966822832823, + 0.006335278041660786, + 0.005898458417505026, + 0.005942163057625294, + 0.0041458928026258945, + 0.004553912207484245, + 0.004752159118652344, + 0.007581094279885292, + 0.0065588089637458324, + 0.009352444671094418, + 0.004947172477841377, + 0.005663138348609209, + 0.009976264089345932, + 0.004576124716550112, + 0.005482492037117481, + 0.002943873405456543, + 0.001691716373898089, + 0.003804924665018916, + 0.002500882837921381, + 0.001971449935808778, + 0.0024607116356492043, + 0.003686661133542657, + 0.0024018618278205395, + 0.0145870978012681, + 0.006976442877203226, + 0.012121988460421562, + 0.004609823226928711, + 0.005675963591784239, + 0.020144814625382423, + 0.006874788086861372, + 0.01581619493663311, + 0.006952047813683748, + 0.00795262586325407, + 0.004908109549432993, + 0.003597947536036372, + 0.004773054737597704, + 0.012825330719351768, + 0.005386972334235907, + 0.0169632937759161, + 0.013462092727422714, + 0.010725707747042179, + 0.012326452881097794, + 0.009103613905608654, + 0.010465490631759167, + 0.014722245745360851, + 0.008775076828897, + 0.01355204451829195, + 0.04174336791038513, + 0.026038991287350655, + 0.03904571011662483, + 0.02019524574279785, + 0.0270500760525465, + 0.038186170160770416, + 0.02493358589708805, + 0.04070519283413887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 1 + ], + "selected_reward": -0.02981521561741829 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521135210990906 + ], + "opposite_logprob_tokens": [ + -1.5021135807037354 + ], + "logprob_reward": -0.2521135210990906, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521135210990906, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11103701591491699, + 0.09906857460737228, + 0.04891175031661987, + 0.03317452222108841, + 0.09731480479240417, + 0.03537306189537048, + 0.12053783982992172, + 0.2588144540786743, + 0.12071805447340012, + 0.07504991441965103 + ], + "credit_map": [ + 0.013301406055688858, + 0.01677190139889717, + 0.012422214262187481, + 0.016391314566135406, + 0.01096141617745161, + 0.015107415616512299, + 0.011797869578003883, + 0.014283482916653156, + 0.011556447483599186, + 0.010388301685452461, + 0.012224163860082626, + 0.010624106042087078, + 0.017333898693323135, + 0.015097374096512794, + 0.010160794481635094, + 0.011683491989970207, + 0.006362448446452618, + 0.0069054472260177135, + 0.005869782995432615, + 0.005472134333103895, + 0.004106293432414532, + 0.007086518220603466, + 0.0070381262339651585, + 0.0060709998942911625, + 0.003919532056897879, + 0.0027687486726790667, + 0.004433231893926859, + 0.005731071345508099, + 0.00418183533474803, + 0.0035188477486371994, + 0.00473122438415885, + 0.0038900328800082207, + 0.016754211857914925, + 0.007968029007315636, + 0.013484127819538116, + 0.010161668062210083, + 0.0078114112839102745, + 0.021950168535113335, + 0.008774894289672375, + 0.010410292074084282, + 0.0031997275073081255, + 0.003083599731326103, + 0.005016162525862455, + 0.005285186227411032, + 0.0044899024069309235, + 0.003518827725201845, + 0.006447168532758951, + 0.004332485608756542, + 0.013634846545755863, + 0.01221583690494299, + 0.0072332420386374, + 0.008117271587252617, + 0.00988621823489666, + 0.02535191923379898, + 0.01309065893292427, + 0.03100784868001938, + 0.03096921741962433, + 0.018993061035871506, + 0.019605441018939018, + 0.03309940546751022, + 0.019936375319957733, + 0.05459142476320267, + 0.032354116439819336, + 0.04926541820168495, + 0.017974667251110077, + 0.01371736265718937, + 0.013190223835408688, + 0.011735635809600353, + 0.01283287163823843, + 0.015091600827872753, + 0.018459266051650047, + 0.0177164264023304, + 0.012349081225693226, + 0.008191670291125774, + 0.011839864775538445, + 0.006289070472121239, + 0.008134671486914158, + 0.012276748195290565, + 0.005886125843971968, + 0.010082680732011795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.2521135210990906 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019427727675065398 + ], + "opposite_logprob_tokens": [ + -6.2519426345825195 + ], + "logprob_reward": -0.0019427727675065398, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019427727675065398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08861044049263, + 0.07107432931661606, + 0.03168649971485138, + 0.05443125218153, + 0.06926272064447403, + 0.0896923765540123, + 0.3502520024776459, + 0.17892979085445404, + 0.04170248284935951, + 0.024358080700039864 + ], + "credit_map": [ + 0.010621801018714905, + 0.012201123870909214, + 0.012427533976733685, + 0.01134863868355751, + 0.0077397762797772884, + 0.010303058661520481, + 0.013162232004106045, + 0.010806280188262463, + 0.00766059011220932, + 0.006885900162160397, + 0.013134312815964222, + 0.008503819815814495, + 0.006890202406793833, + 0.008191728964447975, + 0.010251091793179512, + 0.009556690230965614, + 0.004123631399124861, + 0.0034502388443797827, + 0.004046802408993244, + 0.0038888175040483475, + 0.00271215895190835, + 0.004445062484592199, + 0.005142012611031532, + 0.003877779236063361, + 0.00584582332521677, + 0.006750695873051882, + 0.00805568415671587, + 0.007170879747718573, + 0.0067536295391619205, + 0.006945571396499872, + 0.006990675814449787, + 0.005918292794376612, + 0.012137508019804955, + 0.004692974034696817, + 0.01007409580051899, + 0.006534318905323744, + 0.0055641671642661095, + 0.01686784252524376, + 0.0053094783797860146, + 0.00808233767747879, + 0.0075072324834764, + 0.008121653459966183, + 0.014288397505879402, + 0.010789372026920319, + 0.01041574776172638, + 0.009303146973252296, + 0.01841801404953003, + 0.010848816484212875, + 0.03378383070230484, + 0.0459890179336071, + 0.028539283201098442, + 0.03160443156957626, + 0.03389200568199158, + 0.06398438662290573, + 0.0518343560397625, + 0.06062468886375427, + 0.019901681691408157, + 0.020474720746278763, + 0.01925111748278141, + 0.011481602676212788, + 0.01646801456809044, + 0.031511012464761734, + 0.017519444227218628, + 0.042322203516960144, + 0.005437859334051609, + 0.0032154780346900225, + 0.0070608388632535934, + 0.0047418479807674885, + 0.003346722573041916, + 0.004778645932674408, + 0.007256744895130396, + 0.005864351522177458, + 0.004020319785922766, + 0.0020795704331249, + 0.006995942909270525, + 0.0030720122158527374, + 0.0015251269796863198, + 0.0014766036765649915, + 0.0022908970713615417, + 0.0028976069297641516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.0019427727675065398 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863990470767021 + ], + "opposite_logprob_tokens": [ + -3.048640012741089 + ], + "logprob_reward": -0.04863990470767021, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863990470767021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07892563194036484, + 0.03741177171468735, + 0.03469130024313927, + 0.06148424372076988, + 0.05758223682641983, + 0.17893606424331665, + 0.41311028599739075, + 0.03153320774435997, + 0.08863665908575058, + 0.01768861897289753 + ], + "credit_map": [ + 0.00772685743868351, + 0.009445805102586746, + 0.009880853816866875, + 0.013310699723660946, + 0.009305811487138271, + 0.01043027639389038, + 0.009821371175348759, + 0.009003961458802223, + 0.0029005995020270348, + 0.004446410574018955, + 0.004502356052398682, + 0.005619962699711323, + 0.0063672978430986404, + 0.0049476465210318565, + 0.004132784903049469, + 0.004494715481996536, + 0.0046942769549787045, + 0.004287069663405418, + 0.0043470850214362144, + 0.003996409010142088, + 0.0028112500440329313, + 0.00475488044321537, + 0.005216646008193493, + 0.0045836856588721275, + 0.006567677948623896, + 0.006678663194179535, + 0.010655352845788002, + 0.0071277497336268425, + 0.006528125144541264, + 0.0112313162535429, + 0.006181488744914532, + 0.006513876840472221, + 0.008078488521277905, + 0.007541248574852943, + 0.00823790580034256, + 0.00789989810436964, + 0.0057841818779706955, + 0.006418929435312748, + 0.0080555509775877, + 0.00556603679433465, + 0.016143063083291054, + 0.0159654188901186, + 0.02966918610036373, + 0.020885135978460312, + 0.019542448222637177, + 0.018969666212797165, + 0.0361519493162632, + 0.02160920947790146, + 0.04715753719210625, + 0.05307212099432945, + 0.035267557948827744, + 0.04110167548060417, + 0.043867308646440506, + 0.05391179025173187, + 0.06576734781265259, + 0.07296495139598846, + 0.00371028040535748, + 0.00570537569001317, + 0.003419640939682722, + 0.003861779347062111, + 0.0022149754222482443, + 0.004565967712551355, + 0.0028643906116485596, + 0.005190799944102764, + 0.015226075425744057, + 0.007897633127868176, + 0.013091927394270897, + 0.007172360550612211, + 0.010366514325141907, + 0.013946613296866417, + 0.007877985015511513, + 0.013057554140686989, + 0.0027185101062059402, + 0.0012017086846753955, + 0.003446288639679551, + 0.0024092234671115875, + 0.001481155282817781, + 0.002137678675353527, + 0.0015043191378936172, + 0.0027897367253899574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.04863990470767021 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, positioned just above the ground, captures a soccer player striding across a lush green field under the bright midday sun. The camera then executes a rapid, powerful zoom, darting forward to frame an extreme close-up focused entirely on their feet and the ball. The player's muscular legs, clad in royal blue athletic shorts and matching knee-high socks, power across the field. Their feet, a blur of movement within the intricately designed, laceless studs gleaming in white material, execute a dazzling display of complex footwork. The classic black-and-white paneled ball appears tethered to their feet, being relentlessly yet deftly nudged, tapped, and rolled with lightning-fast precision. Each explosive stride and sharp change of direction drives the ball into the pristine turf, sending visible dark soil and blades of vibrant green grass blasting into the air, suspended for a moment in the sunlight. High-contrast lighting casts sharp, rapidly moving shadows across the neatly manicured grass, creating a dynamic and high-energy atmosphere imbued with intense focus and raw kinetic energy.", + "seed": 922, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_922_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04492860424797982, + "vlm_logprob": -0.04492860424797982, + "vlm_yes_no_margin": 3.8749999523162844, + "vlm_token_credit": -0.04492860424797982, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026829907670617104 + ], + "opposite_logprob_tokens": [ + -4.026829719543457 + ], + "logprob_reward": -0.026829907670617104, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.026829907670617104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1684945821762085, + 0.1259937286376953, + 0.018317027017474174, + 0.061301544308662415, + 0.0525597482919693, + 0.05104172229766846, + 0.06970633566379547, + 0.05426826700568199, + 0.3745304346084595, + 0.023786647245287895 + ], + "credit_map": [ + 0.022350165992975235, + 0.016781693324446678, + 0.03184366226196289, + 0.01975291036069393, + 0.020918674767017365, + 0.0175620224326849, + 0.01785174570977688, + 0.021433696150779724, + 0.010122003965079784, + 0.01663544587790966, + 0.015364020131528378, + 0.0238884799182415, + 0.012623719871044159, + 0.010632996447384357, + 0.022600362077355385, + 0.014126695692539215, + 0.0026016433257609606, + 0.00191132549662143, + 0.003685892792418599, + 0.0017235849518328905, + 0.0017895377241075039, + 0.0022545685060322285, + 0.0017120424890890718, + 0.002638431964442134, + 0.010614401660859585, + 0.006685451604425907, + 0.005434585735201836, + 0.005817319266498089, + 0.0096048004925251, + 0.007505787070840597, + 0.00893472135066986, + 0.006704478524625301, + 0.007145013194531202, + 0.007146783638745546, + 0.007242375984787941, + 0.006647899746894836, + 0.004290463402867317, + 0.00647050840780139, + 0.005767707712948322, + 0.007848997600376606, + 0.008402080275118351, + 0.011260688304901123, + 0.006552623584866524, + 0.0056957160122692585, + 0.005244726315140724, + 0.005255540832877159, + 0.0038318331353366375, + 0.004798515234142542, + 0.010705240070819855, + 0.013040639460086823, + 0.009437177330255508, + 0.010286439210176468, + 0.006952534429728985, + 0.006903544999659061, + 0.0073753586038947105, + 0.005005392711609602, + 0.007664669305086136, + 0.006745634134858847, + 0.0075917793437838554, + 0.011467981152236462, + 0.00482144532725215, + 0.0042722406797111034, + 0.004922199994325638, + 0.006782314740121365, + 0.03602529317140579, + 0.010630344972014427, + 0.13073661923408508, + 0.016257600858807564, + 0.007556538563221693, + 0.0216342993080616, + 0.013090845197439194, + 0.13859888911247253, + 0.0028231414034962654, + 0.0026506951544433832, + 0.003821049816906452, + 0.0021287112031131983, + 0.002333437092602253, + 0.0027213890571147203, + 0.0043899500742554665, + 0.002918275073170662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.026829907670617104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867339879274368 + ], + "opposite_logprob_tokens": [ + -3.048673391342163 + ], + "logprob_reward": -0.04867339879274368, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867339879274368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10292471945285797, + 0.27482420206069946, + 0.02013205550611019, + 0.0432845763862133, + 0.04290714114904404, + 0.023466845974326134, + 0.02373155578970909, + 0.02154991403222084, + 0.4154910743236542, + 0.03168793395161629 + ], + "credit_map": [ + 0.010204490274190903, + 0.014396507292985916, + 0.012260643765330315, + 0.018942710012197495, + 0.011254868470132351, + 0.00632693013176322, + 0.01610243320465088, + 0.013436133041977882, + 0.040988773107528687, + 0.051481980830430984, + 0.03257761895656586, + 0.029904639348387718, + 0.03861471265554428, + 0.02749549224972725, + 0.0240035243332386, + 0.029757419601082802, + 0.0025859482120722532, + 0.0020172041840851307, + 0.003259422490373254, + 0.0023147722240537405, + 0.0036001012194901705, + 0.001987116876989603, + 0.002368002664297819, + 0.0019994883332401514, + 0.005062995012849569, + 0.005222945474088192, + 0.004575125407427549, + 0.0059527927078306675, + 0.008162271231412888, + 0.004634194541722536, + 0.005688794422894716, + 0.003985452465713024, + 0.003648308804258704, + 0.00722170015797019, + 0.003988929092884064, + 0.006053001619875431, + 0.007054361514747143, + 0.003979567438364029, + 0.00560906995087862, + 0.005352201405912638, + 0.002350467722862959, + 0.0030462502036243677, + 0.002812256570905447, + 0.0027663065120577812, + 0.0023230325896292925, + 0.0037523461505770683, + 0.0024360462557524443, + 0.003980138339102268, + 0.002456734422594309, + 0.0027832086198031902, + 0.003391398349776864, + 0.003509088885039091, + 0.0030764644034206867, + 0.0030305087566375732, + 0.0029812774155288935, + 0.00250287470407784, + 0.0019813552498817444, + 0.002352125011384487, + 0.0024723433889448643, + 0.004642090294510126, + 0.002954754512757063, + 0.0022439316380769014, + 0.0020843243692070246, + 0.002818988636136055, + 0.02931930497288704, + 0.008392644114792347, + 0.1540948897600174, + 0.015833595767617226, + 0.009420706890523434, + 0.023622678592801094, + 0.016332749277353287, + 0.15847450494766235, + 0.006462838966399431, + 0.003115586005151272, + 0.004286615177989006, + 0.003279551165178418, + 0.0032559391111135483, + 0.003925248049199581, + 0.0033086237963289022, + 0.004053529351949692 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.04867339879274368 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12705016136169434 + ], + "opposite_logprob_tokens": [ + -2.1270501613616943 + ], + "logprob_reward": -0.12705016136169434, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12705016136169434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14810335636138916, + 0.3394656479358673, + 0.033281195908784866, + 0.06896685808897018, + 0.07904224842786789, + 0.052027810364961624, + 0.033050816506147385, + 0.03334707394242287, + 0.1379348784685135, + 0.07478012144565582 + ], + "credit_map": [ + 0.02059364505112171, + 0.020151875913143158, + 0.017478929832577705, + 0.020474079996347427, + 0.0238901749253273, + 0.01209107507020235, + 0.016212159767746925, + 0.017211416736245155, + 0.022898871451616287, + 0.04321970045566559, + 0.045525748282670975, + 0.06962329149246216, + 0.03445242717862129, + 0.02663424052298069, + 0.058625299483537674, + 0.0384860523045063, + 0.0031013379339128733, + 0.003212276380509138, + 0.003309244057163596, + 0.004176871385425329, + 0.007573246490210295, + 0.004063545726239681, + 0.004271442070603371, + 0.0035732316318899393, + 0.009967430494725704, + 0.009651399217545986, + 0.0062892683781683445, + 0.009615917690098286, + 0.011203508824110031, + 0.006977957207709551, + 0.008292904123663902, + 0.006968472618609667, + 0.008215448819100857, + 0.012432330287992954, + 0.009022950194776058, + 0.012340499088168144, + 0.009668923914432526, + 0.010708833113312721, + 0.008512238040566444, + 0.00814102403819561, + 0.00794996228069067, + 0.008888335898518562, + 0.0049758669920265675, + 0.0063002705574035645, + 0.0069769201800227165, + 0.006719752214848995, + 0.005818519741296768, + 0.0043981801718473434, + 0.0041307746432721615, + 0.004686649888753891, + 0.004683363251388073, + 0.005524552892893553, + 0.0033593298867344856, + 0.0038479964714497328, + 0.003962274175137281, + 0.0028558748308569193, + 0.003291629021987319, + 0.004377839621156454, + 0.0047623091377317905, + 0.007611447479575872, + 0.00334542035125196, + 0.0028574722819030285, + 0.0034638261422514915, + 0.0036371296737343073, + 0.009641356766223907, + 0.0028273307252675295, + 0.05032646656036377, + 0.006635183002799749, + 0.0033778271172195673, + 0.008594609797000885, + 0.0051576425321400166, + 0.05137445777654648, + 0.012144641950726509, + 0.009982991963624954, + 0.007601116318255663, + 0.007106972858309746, + 0.007935749366879463, + 0.012327891774475574, + 0.0106247803196311, + 0.007055976428091526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.12705016136169434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015148365637287498 + ], + "opposite_logprob_tokens": [ + -6.501514911651611 + ], + "logprob_reward": -0.0015148365637287498, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015148365637287498, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09611393511295319, + 0.05304136872291565, + 0.04882942512631416, + 0.09659287333488464, + 0.07290203869342804, + 0.07200771570205688, + 0.12145617604255676, + 0.0667782723903656, + 0.26106128096580505, + 0.11121691763401031 + ], + "credit_map": [ + 0.008824898861348629, + 0.013301407918334007, + 0.012249956838786602, + 0.01801787130534649, + 0.013106499798595905, + 0.0074113220907747746, + 0.01185469701886177, + 0.01134728267788887, + 0.0057132956571877, + 0.008917934261262417, + 0.0067858463153243065, + 0.0066800969652831554, + 0.008025315590202808, + 0.006099449470639229, + 0.005878383293747902, + 0.004941045772284269, + 0.00477950181812048, + 0.0037406820338219404, + 0.009758943691849709, + 0.007251422386616468, + 0.006640402600169182, + 0.004267239477485418, + 0.0067619518376886845, + 0.005629281047731638, + 0.012437717989087105, + 0.012577367946505547, + 0.008038925938308239, + 0.009625104255974293, + 0.016861144453287125, + 0.011777471750974655, + 0.014761141501367092, + 0.010513996705412865, + 0.008372708223760128, + 0.007570832502096891, + 0.009430723264813423, + 0.009485906921327114, + 0.009999727830290794, + 0.009578081779181957, + 0.00938548892736435, + 0.009078559465706348, + 0.011747445911169052, + 0.017725426703691483, + 0.0078691141679883, + 0.006344510708004236, + 0.007526922971010208, + 0.008442402817308903, + 0.006618511863052845, + 0.005733382888138294, + 0.009146024473011494, + 0.013427317142486572, + 0.018634699285030365, + 0.01883290894329548, + 0.01504682470113039, + 0.01485253032296896, + 0.018827203661203384, + 0.012688666582107544, + 0.007627114187926054, + 0.007060552481561899, + 0.007962934672832489, + 0.013080856762826443, + 0.007173172198235989, + 0.007431407924741507, + 0.0068344115279614925, + 0.009607828222215176, + 0.02458791993558407, + 0.011809449642896652, + 0.07858866453170776, + 0.01259317435324192, + 0.014699661172926426, + 0.023372601717710495, + 0.016776252537965775, + 0.07863353937864304, + 0.02553134597837925, + 0.011479290202260017, + 0.010471874848008156, + 0.009893814101815224, + 0.016448471695184708, + 0.010253054089844227, + 0.01182238943874836, + 0.015316670760512352 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.0015148365637287498 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020574716851115227 + ], + "opposite_logprob_tokens": [ + -3.8955748081207275 + ], + "logprob_reward": -0.020574716851115227, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020574716851115227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12613065540790558, + 0.09813503175973892, + 0.03455480933189392, + 0.11773829907178879, + 0.026280051097273827, + 0.04485276713967323, + 0.05494190379977226, + 0.017547011375427246, + 0.4605955183506012, + 0.019223924726247787 + ], + "credit_map": [ + 0.011565334163606167, + 0.01707765832543373, + 0.012824708595871925, + 0.024011405184864998, + 0.016204524785280228, + 0.009741539135575294, + 0.019562022760510445, + 0.01514346431940794, + 0.006752367131412029, + 0.011265372857451439, + 0.01249814871698618, + 0.02102656662464142, + 0.009265516884624958, + 0.0067950342781841755, + 0.019079821184277534, + 0.0114522073417902, + 0.0038176116067916155, + 0.0047422898933291435, + 0.005402814131230116, + 0.003936564549803734, + 0.005971010774374008, + 0.0030693390872329473, + 0.003968741279095411, + 0.0036464363802224398, + 0.013569096103310585, + 0.016347141936421394, + 0.008746963925659657, + 0.011639845557510853, + 0.026183893904089928, + 0.012471802532672882, + 0.01612509787082672, + 0.012654458172619343, + 0.0033745975233614445, + 0.003071531653404236, + 0.003271389054134488, + 0.0027807438746094704, + 0.003007841994985938, + 0.0037762033753097057, + 0.0034316759556531906, + 0.003566067200154066, + 0.005522994324564934, + 0.007480685133486986, + 0.004355963319540024, + 0.006507692858576775, + 0.00639765989035368, + 0.005198174621909857, + 0.006354328244924545, + 0.0030352685134857893, + 0.005831531714648008, + 0.006061109248548746, + 0.0079758046194911, + 0.009976050816476345, + 0.006247719284147024, + 0.007370797451585531, + 0.0060827345587313175, + 0.005396154709160328, + 0.001529088942334056, + 0.0017114889342337847, + 0.002152134431526065, + 0.0037281052209436893, + 0.0023690841626375914, + 0.0019778027199208736, + 0.001621910952962935, + 0.002457395661622286, + 0.032074350863695145, + 0.012360792607069016, + 0.1652083694934845, + 0.017570113763213158, + 0.014054502360522747, + 0.02905452251434326, + 0.019196484237909317, + 0.17107635736465454, + 0.004727921448647976, + 0.0019633236806839705, + 0.0018940242007374763, + 0.0014229476219043136, + 0.0029575987718999386, + 0.002129779662936926, + 0.0019291085191071033, + 0.0021992202382534742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.020574716851115227 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, positioned just above the ground, captures a soccer player striding across a lush green field under the bright midday sun. The camera then executes a rapid, powerful zoom, darting forward to frame an extreme close-up focused entirely on their feet and the ball. The player's muscular legs, clad in royal blue athletic shorts and matching knee-high socks, power across the field. Their feet, a blur of movement within the intricately designed, laceless studs gleaming in white material, execute a dazzling display of complex footwork. The classic black-and-white paneled ball appears tethered to their feet, being relentlessly yet deftly nudged, tapped, and rolled with lightning-fast precision. Each explosive stride and sharp change of direction drives the ball into the pristine turf, sending visible dark soil and blades of vibrant green grass blasting into the air, suspended for a moment in the sunlight. High-contrast lighting casts sharp, rapidly moving shadows across the neatly manicured grass, creating a dynamic and high-energy atmosphere imbued with intense focus and raw kinetic energy.", + "seed": 923, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_923_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05397463664412498, + "vlm_logprob": -0.05397463664412498, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.05397463664412498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011912751011550426 + ], + "opposite_logprob_tokens": [ + -5.136912822723389 + ], + "logprob_reward": -0.011912751011550426, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.011912751011550426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1376822143793106, + 0.02835388481616974, + 0.05820168927311897, + 0.051682788878679276, + 0.0325578935444355, + 0.0389275886118412, + 0.03430483862757683, + 0.03355235606431961, + 0.5343175530433655, + 0.050419166684150696 + ], + "credit_map": [ + 0.016475530341267586, + 0.013444606214761734, + 0.027709079906344414, + 0.015178975649178028, + 0.012118411250412464, + 0.02105916105210781, + 0.017495710402727127, + 0.014200725592672825, + 0.0036448678001761436, + 0.0036936572287231684, + 0.003917321562767029, + 0.004442301578819752, + 0.002632837975397706, + 0.0032080651726573706, + 0.003684115828946233, + 0.0031307172030210495, + 0.006184068508446217, + 0.00630252156406641, + 0.006625958252698183, + 0.003743305103853345, + 0.015380071476101875, + 0.008521076291799545, + 0.005990682635456324, + 0.0054540070705115795, + 0.00895107351243496, + 0.006215910892933607, + 0.010838196612894535, + 0.006172118242830038, + 0.004165979102253914, + 0.003476809710264206, + 0.004771430045366287, + 0.007091272622346878, + 0.0031412041280418634, + 0.003360890317708254, + 0.0059147183783352375, + 0.003470079507678747, + 0.005535013973712921, + 0.002252854872494936, + 0.004644262604415417, + 0.004238868597894907, + 0.0035377175081521273, + 0.0038262035232037306, + 0.010043435730040073, + 0.012008925899863243, + 0.0017839791253209114, + 0.001984129659831524, + 0.0030166450887918472, + 0.0027265518438071012, + 0.006706105545163155, + 0.005068700760602951, + 0.006120843347162008, + 0.004925772547721863, + 0.0024608608800917864, + 0.0025134903844445944, + 0.0032294506672769785, + 0.0032796121668070555, + 0.004171444568783045, + 0.004959665238857269, + 0.00665265740826726, + 0.004413051530718803, + 0.0029658342245966196, + 0.0048621539026498795, + 0.002437830902636051, + 0.0030897180549800396, + 0.0539834126830101, + 0.07632379978895187, + 0.1714007407426834, + 0.12156757712364197, + 0.030591633170843124, + 0.020678451284766197, + 0.03494749590754509, + 0.024824468418955803, + 0.003974521066993475, + 0.015149999409914017, + 0.01849176548421383, + 0.005201296415179968, + 0.0019216937944293022, + 0.001794314244762063, + 0.0017467669676989317, + 0.002138808136805892 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.011912751011550426 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301001638174057 + ], + "opposite_logprob_tokens": [ + -3.773300886154175 + ], + "logprob_reward": -0.023301001638174057, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301001638174057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18037734925746918, + 0.051406972110271454, + 0.13359381258487701, + 0.024153508245944977, + 0.05347130820155144, + 0.09728222340345383, + 0.05306391417980194, + 0.0894295945763588, + 0.283507764339447, + 0.03371356055140495 + ], + "credit_map": [ + 0.01596006192266941, + 0.03336586058139801, + 0.023290127515792847, + 0.022761356085538864, + 0.016929445788264275, + 0.01201536413282156, + 0.03585561737418175, + 0.020199518650770187, + 0.006577952764928341, + 0.0054875947535037994, + 0.010580625385046005, + 0.0067706359550356865, + 0.0048047020100057125, + 0.004796231165528297, + 0.006898395251482725, + 0.005490834359079599, + 0.013989884406328201, + 0.013333534821867943, + 0.02014855481684208, + 0.013538709841668606, + 0.033670950680971146, + 0.010327484458684921, + 0.017434358596801758, + 0.011150321923196316, + 0.0032954677008092403, + 0.0032614129595458508, + 0.0037452750839293003, + 0.002549429191276431, + 0.0023008405696600676, + 0.002481929026544094, + 0.0030245399102568626, + 0.0034946161322295666, + 0.00454908050596714, + 0.004243174102157354, + 0.009787427261471748, + 0.006201942451298237, + 0.0064339591190218925, + 0.004522272851318121, + 0.00919739156961441, + 0.008536061272025108, + 0.006196658592671156, + 0.004886281210929155, + 0.03132133558392525, + 0.034588173031806946, + 0.00604153610765934, + 0.004590496886521578, + 0.004888781346380711, + 0.004768973216414452, + 0.00692998431622982, + 0.006474516354501247, + 0.007115528918802738, + 0.005658280570060015, + 0.006844364572316408, + 0.008053230121731758, + 0.006501643918454647, + 0.005486363545060158, + 0.0113499965518713, + 0.011895977891981602, + 0.011025398038327694, + 0.009359823539853096, + 0.008763104677200317, + 0.015734413638710976, + 0.008446029387414455, + 0.012854850850999355, + 0.028558745980262756, + 0.039760056883096695, + 0.09471862763166428, + 0.0616755485534668, + 0.016945134848356247, + 0.011158152483403683, + 0.018413517624139786, + 0.012277969159185886, + 0.0038797687739133835, + 0.003895909059792757, + 0.004955472890287638, + 0.006923007313162088, + 0.003869525622576475, + 0.003516131779178977, + 0.0035333922132849693, + 0.003140352200716734 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 2 + ], + "selected_reward": -0.023301001638174057 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015928477048874 + ], + "opposite_logprob_tokens": [ + -1.7015928030014038 + ], + "logprob_reward": -0.2015928477048874, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015928477048874, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09417219460010529, + 0.09820454567670822, + 0.3127397298812866, + 0.022073974832892418, + 0.05601900443434715, + 0.04648160934448242, + 0.0804142951965332, + 0.11567451804876328, + 0.15528210997581482, + 0.018937984481453896 + ], + "credit_map": [ + 0.008162826299667358, + 0.009409231133759022, + 0.019952796399593353, + 0.014281845651566982, + 0.008867781609296799, + 0.0056420802138745785, + 0.017337065190076828, + 0.010518576018512249, + 0.012764734216034412, + 0.009505163878202438, + 0.014701667241752148, + 0.015492814593017101, + 0.011742088943719864, + 0.010401836596429348, + 0.01309977937489748, + 0.010496476665139198, + 0.03265748172998428, + 0.03275584429502487, + 0.03889237344264984, + 0.024614613503217697, + 0.0835288017988205, + 0.032777994871139526, + 0.04106884449720383, + 0.02644381672143936, + 0.001917095622047782, + 0.0038268023636192083, + 0.0031991053838282824, + 0.0029532634653151035, + 0.0019509262638166547, + 0.0019169794395565987, + 0.0035120851825922728, + 0.0027977190911769867, + 0.004835641011595726, + 0.006699286866933107, + 0.00753109622746706, + 0.0072321817278862, + 0.008156077936291695, + 0.004522583447396755, + 0.008636992424726486, + 0.00840514712035656, + 0.004330000374466181, + 0.0036448317114263773, + 0.013810526579618454, + 0.012551662512123585, + 0.0032297042198479176, + 0.0028546631801873446, + 0.0035903984680771828, + 0.0024698248598724604, + 0.010824648663401604, + 0.010698900558054447, + 0.011585576459765434, + 0.008799265138804913, + 0.008922909386456013, + 0.010908976197242737, + 0.009055557660758495, + 0.009618470445275307, + 0.012214694172143936, + 0.015122516080737114, + 0.01726498268544674, + 0.014290654100477695, + 0.012948024086654186, + 0.01664760522544384, + 0.012878053821623325, + 0.01430799812078476, + 0.016001014038920403, + 0.02176343835890293, + 0.050612471997737885, + 0.03373152017593384, + 0.008940530009567738, + 0.006895286962389946, + 0.010179846547544003, + 0.007158020045608282, + 0.002382055390626192, + 0.003657195484265685, + 0.0036386139690876007, + 0.003944705706089735, + 0.001083049806766212, + 0.0015011419309303164, + 0.0014987951144576073, + 0.001232429058291018 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.2015928477048874 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782352484762669 + ], + "opposite_logprob_tokens": [ + -4.634782314300537 + ], + "logprob_reward": -0.009782352484762669, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782352484762669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1122414767742157, + 0.047335125505924225, + 0.1275157481431961, + 0.023516319692134857, + 0.12477954477071762, + 0.07209296524524689, + 0.10659930855035782, + 0.24564430117607117, + 0.04804115369915962, + 0.0922340378165245 + ], + "credit_map": [ + 0.013924123719334602, + 0.023155469447374344, + 0.015191207639873028, + 0.013885637745261192, + 0.012092596851289272, + 0.008331140503287315, + 0.014437748119235039, + 0.01122355181723833, + 0.006439894903451204, + 0.004298937041312456, + 0.007951254956424236, + 0.005547639913856983, + 0.006792041473090649, + 0.004746759310364723, + 0.006613901816308498, + 0.004944700747728348, + 0.016190458089113235, + 0.013598849065601826, + 0.017849138006567955, + 0.014169110916554928, + 0.027004849165678024, + 0.013325552456080914, + 0.014490647241473198, + 0.010887152515351772, + 0.0016475303564220667, + 0.0031971405260264874, + 0.0026948151644319296, + 0.002936264034360647, + 0.002863915404304862, + 0.002562971320003271, + 0.004068148788064718, + 0.003545532701537013, + 0.015583324246108532, + 0.01087605208158493, + 0.01438699010759592, + 0.012650934979319572, + 0.017196157947182655, + 0.012250933796167374, + 0.02451704442501068, + 0.017318107187747955, + 0.003435988910496235, + 0.003231785958632827, + 0.022300558164715767, + 0.025760481134057045, + 0.0033816362265497446, + 0.003581096651032567, + 0.005849388428032398, + 0.004552031867206097, + 0.01245209388434887, + 0.014454432763159275, + 0.012211943045258522, + 0.008910703472793102, + 0.015517094172537327, + 0.017081983387470245, + 0.013153046369552612, + 0.01281801238656044, + 0.0319347158074379, + 0.03516879677772522, + 0.03189946711063385, + 0.024449091404676437, + 0.02431933395564556, + 0.04138128459453583, + 0.023575229570269585, + 0.03291639685630798, + 0.004928550682961941, + 0.006560066249221563, + 0.013776013627648354, + 0.009075511246919632, + 0.0036808347795158625, + 0.003343313466757536, + 0.0036412219051271677, + 0.003035640576854348, + 0.011259627528488636, + 0.01975748874247074, + 0.025164326652884483, + 0.011387040838599205, + 0.006161052733659744, + 0.007140524219721556, + 0.0047028218396008015, + 0.006661164574325085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 4 + ], + "selected_reward": -0.009782352484762669 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328423038125038 + ], + "opposite_logprob_tokens": [ + -3.7732841968536377 + ], + "logprob_reward": -0.02328423038125038, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328423038125038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15881133079528809, + 0.07606007903814316, + 0.06183675304055214, + 0.12150315195322037, + 0.18477922677993774, + 0.17505192756652832, + 0.07747422158718109, + 0.03507512807846069, + 0.09132203459739685, + 0.018086176365613937 + ], + "credit_map": [ + 0.015131359919905663, + 0.037339914590120316, + 0.02634158544242382, + 0.024214550852775574, + 0.014301218092441559, + 0.010985184460878372, + 0.01662938855588436, + 0.013868124224245548, + 0.009841462597250938, + 0.008099676109850407, + 0.010515572503209114, + 0.011422554962337017, + 0.008492817170917988, + 0.007035304792225361, + 0.013720951043069363, + 0.006931731011718512, + 0.008941174484789371, + 0.005024902056902647, + 0.009719962254166603, + 0.006140443962067366, + 0.006668214686214924, + 0.009670743718743324, + 0.008617430925369263, + 0.00705387769266963, + 0.019602984189987183, + 0.02401764877140522, + 0.0242720115929842, + 0.015613324008882046, + 0.009414232335984707, + 0.008182714693248272, + 0.008641394786536694, + 0.011758831329643726, + 0.020457226783037186, + 0.014516097493469715, + 0.028619512915611267, + 0.027878934517502785, + 0.019173678010702133, + 0.015212906524538994, + 0.02994983270764351, + 0.02897101640701294, + 0.021503854542970657, + 0.01570025272667408, + 0.03677326813340187, + 0.03059856779873371, + 0.01727241650223732, + 0.01458925660699606, + 0.023234065622091293, + 0.015380238182842731, + 0.00983683206140995, + 0.00910174660384655, + 0.00712643563747406, + 0.006151633802801371, + 0.010756012052297592, + 0.012275593355298042, + 0.011210654862225056, + 0.011015301570296288, + 0.003957212436944246, + 0.004988010972738266, + 0.005072543863207102, + 0.0045106844045221806, + 0.0036648474633693695, + 0.004993493668735027, + 0.0027502826415002346, + 0.005138049367815256, + 0.008418470621109009, + 0.011286317370831966, + 0.019724326208233833, + 0.0115469079464674, + 0.011995824053883553, + 0.009915749542415142, + 0.00970133114606142, + 0.008733110502362251, + 0.0017015928169712424, + 0.0023094117641448975, + 0.00333667011000216, + 0.004628741182386875, + 0.0013264077715575695, + 0.0019738664850592613, + 0.001369654550217092, + 0.0014398309867829084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.02328423038125038 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, medium shot, positioned just above the ground, captures a soccer player striding across a lush green field under the bright midday sun. The camera then executes a rapid, powerful zoom, darting forward to frame an extreme close-up focused entirely on their feet and the ball. The player's muscular legs, clad in royal blue athletic shorts and matching knee-high socks, power across the field. Their feet, a blur of movement within the intricately designed, laceless studs gleaming in white material, execute a dazzling display of complex footwork. The classic black-and-white paneled ball appears tethered to their feet, being relentlessly yet deftly nudged, tapped, and rolled with lightning-fast precision. Each explosive stride and sharp change of direction drives the ball into the pristine turf, sending visible dark soil and blades of vibrant green grass blasting into the air, suspended for a moment in the sunlight. High-contrast lighting casts sharp, rapidly moving shadows across the neatly manicured grass, creating a dynamic and high-energy atmosphere imbued with intense focus and raw kinetic energy.", + "seed": 924, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_7/index_1640_seed_924_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.029457408282905818, + "vlm_logprob": -0.029457408282905818, + "vlm_yes_no_margin": 4.475, + "vlm_token_credit": -0.029457408282905818, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004397958982735872 + ], + "opposite_logprob_tokens": [ + -6.129397869110107 + ], + "logprob_reward": -0.004397958982735872, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.004397958982735872, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17106835544109344, + 0.03423507139086723, + 0.034819211810827255, + 0.016503192484378815, + 0.025412550196051598, + 0.03408217057585716, + 0.18499477207660675, + 0.056199006736278534, + 0.06072698533535004, + 0.3819587230682373 + ], + "credit_map": [ + 0.015032312832772732, + 0.010020554065704346, + 0.047417644411325455, + 0.038263291120529175, + 0.012875872664153576, + 0.011252548545598984, + 0.015435263514518738, + 0.020770860835909843, + 0.003728809067979455, + 0.003908826503902674, + 0.004497410263866186, + 0.004186592530459166, + 0.005487176589667797, + 0.0032985114958137274, + 0.00453288946300745, + 0.004594852216541767, + 0.003564827609807253, + 0.0031817813869565725, + 0.005993423517793417, + 0.0045123170129954815, + 0.005179254803806543, + 0.0037743421271443367, + 0.004522524308413267, + 0.004090744536370039, + 0.0015786591684445739, + 0.0014979930128902197, + 0.0024919298011809587, + 0.0033770715817809105, + 0.0019920652266591787, + 0.001605071360245347, + 0.001953753875568509, + 0.002006648341193795, + 0.0028381997253745794, + 0.0028165311086922884, + 0.003494078991934657, + 0.0022550837602466345, + 0.0031450949609279633, + 0.002599206054583192, + 0.003251455957069993, + 0.005012901034206152, + 0.004036658443510532, + 0.0055600847117602825, + 0.004991834983229637, + 0.005702091380953789, + 0.003971886821091175, + 0.0026217265985906124, + 0.0031369670759886503, + 0.004060918930917978, + 0.015149553306400776, + 0.023898065090179443, + 0.030649622902274132, + 0.021292010322213173, + 0.018760962411761284, + 0.018724894151091576, + 0.02519029565155506, + 0.031329378485679626, + 0.010020343586802483, + 0.009513967670500278, + 0.008727631531655788, + 0.005629759281873703, + 0.0062966556288301945, + 0.004006068222224712, + 0.004744262434542179, + 0.007260319776833057, + 0.01124411541968584, + 0.005076099187135696, + 0.007643396500498056, + 0.004616135731339455, + 0.007507314905524254, + 0.006569352466613054, + 0.00968257151544094, + 0.008387998677790165, + 0.07325680553913116, + 0.1611802726984024, + 0.016079753637313843, + 0.024069769307971, + 0.029238667339086533, + 0.0389321893453598, + 0.026206422597169876, + 0.012994840741157532 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.004397958982735872 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016067663207650185 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.016067663207650185, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016067663207650185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11476081609725952, + 0.0404655896127224, + 0.03105560503900051, + 0.03617461025714874, + 0.013845889829099178, + 0.0472228080034256, + 0.0939575806260109, + 0.13250797986984253, + 0.04149799793958664, + 0.4485112130641937 + ], + "credit_map": [ + 0.007202348671853542, + 0.006842286325991154, + 0.028430238366127014, + 0.025890784338116646, + 0.011988830752670765, + 0.006865173112601042, + 0.01786397583782673, + 0.009677182883024216, + 0.0034782150760293007, + 0.0035507516004145145, + 0.004714785143733025, + 0.005655848886817694, + 0.007685461081564426, + 0.0035457478370517492, + 0.006813121493905783, + 0.005021660588681698, + 0.002709819469600916, + 0.0031262682750821114, + 0.005275784991681576, + 0.0047097597271203995, + 0.004762991331517696, + 0.0030000610277056694, + 0.003723681904375553, + 0.0037472404073923826, + 0.002937681507319212, + 0.0035616459790617228, + 0.003736402839422226, + 0.0036621401086449623, + 0.004788780584931374, + 0.004908755421638489, + 0.008403555490076542, + 0.00417564669623971, + 0.0017269025556743145, + 0.0016604065895080566, + 0.0019429074600338936, + 0.0019383792532607913, + 0.0019526422256603837, + 0.001112784375436604, + 0.0018455951940268278, + 0.0016662721754983068, + 0.004740497563034296, + 0.005265471525490284, + 0.007216480560600758, + 0.006626353599131107, + 0.006370618473738432, + 0.003684057155624032, + 0.005997086875140667, + 0.007322240620851517, + 0.008070997893810272, + 0.011649602092802525, + 0.02054056152701378, + 0.010293704457581043, + 0.012557536363601685, + 0.007452887482941151, + 0.015383555553853512, + 0.008008738048374653, + 0.021152185276150703, + 0.017841482535004616, + 0.04741787537932396, + 0.012166706845164299, + 0.01066784281283617, + 0.006074868142604828, + 0.007490183226764202, + 0.00969682913273573, + 0.005454080644994974, + 0.006578095257282257, + 0.006026808638125658, + 0.004014826845377684, + 0.004396054428070784, + 0.0041337101720273495, + 0.005453494377434254, + 0.005440928973257542, + 0.07972396910190582, + 0.1942279040813446, + 0.017341237515211105, + 0.027952203527092934, + 0.03407309949398041, + 0.04753894731402397, + 0.02848641946911812, + 0.019167395308613777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.016067663207650185 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293266713619232 + ], + "opposite_logprob_tokens": [ + -2.2379326820373535 + ], + "logprob_reward": -0.11293266713619232, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293266713619232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07858899980783463, + 0.019227273762226105, + 0.028812376782298088, + 0.01911514811217785, + 0.027997473254799843, + 0.017391793429851532, + 0.1063109040260315, + 0.1128455102443695, + 0.05160391330718994, + 0.5381066203117371 + ], + "credit_map": [ + 0.006137568037956953, + 0.005852171219885349, + 0.017893271520733833, + 0.015140830539166927, + 0.009567366912961006, + 0.007323452737182379, + 0.009233226999640465, + 0.007441099267452955, + 0.002341431798413396, + 0.0018481607548892498, + 0.002347031608223915, + 0.0016679378459230065, + 0.00284002348780632, + 0.0028739033732563257, + 0.0021437134128063917, + 0.0031650704331696033, + 0.003124210052192211, + 0.0028717273380607367, + 0.004291045945137739, + 0.0033202164340764284, + 0.0045044561848044395, + 0.003224158426746726, + 0.00355141912586987, + 0.003925136756151915, + 0.0018489371286705136, + 0.0019615169148892164, + 0.002866951050236821, + 0.0018481194274500012, + 0.002901029773056507, + 0.002780740149319172, + 0.0025398810394108295, + 0.002367972396314144, + 0.0028242391999810934, + 0.002628399059176445, + 0.004557213746011257, + 0.0023575532250106335, + 0.002592334756627679, + 0.002700975863263011, + 0.003037708578631282, + 0.007299046032130718, + 0.0023393097799271345, + 0.002194681204855442, + 0.0029598705004900694, + 0.0021993357222527266, + 0.0021881922148168087, + 0.0015130346873775125, + 0.0021287614945322275, + 0.0018686061957851052, + 0.008928850293159485, + 0.011000040918588638, + 0.021789349615573883, + 0.010112976655364037, + 0.01106909941881895, + 0.011740007437765598, + 0.01478596217930317, + 0.01688460074365139, + 0.017914626747369766, + 0.014530437998473644, + 0.034973278641700745, + 0.012990458868443966, + 0.011783736757934093, + 0.006285183597356081, + 0.007541918195784092, + 0.006825849413871765, + 0.0058307768777012825, + 0.006928980816155672, + 0.008167749270796776, + 0.004890557378530502, + 0.0062878490425646305, + 0.0051376610063016415, + 0.00790441781282425, + 0.006455916911363602, + 0.10007733106613159, + 0.23282529413700104, + 0.021526144817471504, + 0.03288442641496658, + 0.04061971604824066, + 0.057153645902872086, + 0.03189639747142792, + 0.021123630926012993 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.11293266713619232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637921884655952 + ], + "opposite_logprob_tokens": [ + -4.75863790512085 + ], + "logprob_reward": -0.008637921884655952, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637921884655952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02418474107980728, + 0.01592007838189602, + 0.012726173736155033, + 0.016828930005431175, + 0.01494141947478056, + 0.010136944241821766, + 0.027959316968917847, + 0.08140318095684052, + 0.06374476104974747, + 0.7321544289588928 + ], + "credit_map": [ + 0.0026560500264167786, + 0.0029165996238589287, + 0.0033134282566607, + 0.002649367321282625, + 0.002671055728569627, + 0.003337414003908634, + 0.0036503851879388094, + 0.002990440232679248, + 0.0013068553525954485, + 0.001322322990745306, + 0.002484011696651578, + 0.0020454253535717726, + 0.001987763913348317, + 0.0016239427495747805, + 0.003425820730626583, + 0.0017239346634596586, + 0.0014411204028874636, + 0.0014445912092924118, + 0.0017450215527787805, + 0.0012099567102268338, + 0.002626113360747695, + 0.0012629472184926271, + 0.0017102017300203443, + 0.0012862208532169461, + 0.0010514401365071535, + 0.0011303479550406337, + 0.0018824057187885046, + 0.001673471531830728, + 0.002073200885206461, + 0.001071370206773281, + 0.005481387488543987, + 0.0024653065484017134, + 0.0014647995121777058, + 0.0011592056835070252, + 0.0033352645114064217, + 0.001344032702036202, + 0.00158434989862144, + 0.0012705455301329494, + 0.001549679902382195, + 0.0032335412688553333, + 0.0015568342059850693, + 0.001901109702885151, + 0.001627075020223856, + 0.0013562787789851427, + 0.0012232117587700486, + 0.0006681185332126915, + 0.0010148995788767934, + 0.0007894161390140653, + 0.003311742562800646, + 0.005442061461508274, + 0.005797888617962599, + 0.003199646482244134, + 0.002403095830231905, + 0.00203294912353158, + 0.002588429721072316, + 0.003183505032211542, + 0.01094836089760065, + 0.011860961094498634, + 0.016966987401247025, + 0.006217720918357372, + 0.006583198439329863, + 0.006203759461641312, + 0.009580738842487335, + 0.013041448779404163, + 0.012256162241101265, + 0.005888377781957388, + 0.00710891280323267, + 0.004313788376748562, + 0.00971654336899519, + 0.004821054171770811, + 0.01252097450196743, + 0.007118952926248312, + 0.13269150257110596, + 0.3146817088127136, + 0.02904924750328064, + 0.04476833716034889, + 0.055982012301683426, + 0.076487235724926, + 0.04783221334218979, + 0.030662216246128082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.008637921884655952 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set on a lush green field under bright midday sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the player wearing royal blue athletic shorts and matching knee-high socks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the player's feet wearing white laceless studs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show a classic black-and-white paneled ball being nudged, tapped, and rolled?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start with a low-angle medium shot and then execute a rapid zoom to an extreme close-up of the feet and ball?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250830203294754 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005250830203294754, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250830203294754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24252672493457794, + 0.017350707203149796, + 0.025746436789631844, + 0.017363158985972404, + 0.034948576241731644, + 0.044725313782691956, + 0.208114355802536, + 0.04223394766449928, + 0.09704198688268661, + 0.2699487805366516 + ], + "credit_map": [ + 0.012822562828660011, + 0.011552286334335804, + 0.07186906784772873, + 0.06553901731967926, + 0.022707819938659668, + 0.010850328952074051, + 0.030292045325040817, + 0.01689358986914158, + 0.001751835341565311, + 0.00220215180888772, + 0.0020461883395910263, + 0.0017248508520424366, + 0.0029072128236293793, + 0.0020912105683237314, + 0.002427836647257209, + 0.0021994190756231546, + 0.0019356460543349385, + 0.0018103983020409942, + 0.003671585815027356, + 0.0027383272536098957, + 0.0025883717462420464, + 0.0017827097326517105, + 0.0024779080413281918, + 0.008741489611566067, + 0.0015496478881686926, + 0.0015704130055382848, + 0.0020147585310041904, + 0.002728069666773081, + 0.002756016794592142, + 0.002193262567743659, + 0.0025545810349285603, + 0.001996408449485898, + 0.003324158489704132, + 0.002774176187813282, + 0.0068067992106080055, + 0.003172867698594928, + 0.004119112156331539, + 0.003342011710628867, + 0.004317357204854488, + 0.0070920949801802635, + 0.004848080221563578, + 0.005608647596091032, + 0.006990739144384861, + 0.007485643960535526, + 0.006305258255451918, + 0.0030596957076340914, + 0.005944835487753153, + 0.004482411779463291, + 0.017432359978556633, + 0.024445531889796257, + 0.03476069122552872, + 0.021406622603535652, + 0.0202433243393898, + 0.023011991754174232, + 0.02667343243956566, + 0.04014041647315025, + 0.00535751273855567, + 0.005073219072073698, + 0.009541057050228119, + 0.0037458431906998158, + 0.003975974395871162, + 0.0036574294790625572, + 0.003970503807067871, + 0.006912413984537125, + 0.011298708617687225, + 0.012970229610800743, + 0.013886875472962856, + 0.009208930656313896, + 0.010786238126456738, + 0.016122976318001747, + 0.011736487038433552, + 0.011031536385416985, + 0.051701098680496216, + 0.11833760887384415, + 0.010437729768455029, + 0.016618812456727028, + 0.02034420147538185, + 0.02908048965036869, + 0.013464942574501038, + 0.00996390637010336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.005250830203294754 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_tensor([7373, 7373])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_tensor([7373, 7373])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d117c8b8aafcfdd03b447846d8a9d9da6138a714 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_tensor([7373, 7373])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, tautly focused close-up shot, beautifully captured at a high shutter speed, focuses on a blond soldier at the peak of his fury. His short, flaming blond hair trembles slightly in the chaotic air currents, his face contorted with utter fury, his eyes wide, and veins bulging in his forehead and neck from exertion. He wears a dark green, scuffed tactical vest over a tight-fitting black combat uniform, clearly outlining the muscular mass of his arms, shoulders, and chest bulging with exertion. Against a dim, rubble-strewn backdrop of battlefield ruins, he leans forward, his legs firmly planted in a lunge. A dazzling, translucent, bluish-white crescent-shaped wave of energy erupts from between his thrusting palms. This wave itself becomes the primary source of light in the scene, casting sharp highlights and deep shadows across his enraged face. Under the extremely clear lens, every bead of sweat splashing from his forehead and the tiny stones shaken up by the energy shock wave were perfectly solidified in the air, presenting a crystal-clear still state. The whole picture was full of explosive power and a tense atmosphere of life and death.", + "seed": 11018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_7373_seed_11018_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.031791404169052837, + "vlm_logprob": -0.031791404169052837, + "vlm_yes_no_margin": 4.449999976158142, + "vlm_token_credit": -0.031791404169052837, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12806758284568787 + ], + "opposite_logprob_tokens": [ + -2.1280674934387207 + ], + "logprob_reward": -0.12806758284568787, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12806758284568787, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1650390475988388, + 0.06311789155006409, + 0.028783654794096947, + 0.12828461825847626, + 0.009218545630574226, + 0.013679489493370056, + 0.0737144947052002, + 0.043257009238004684, + 0.42272713780403137, + 0.052178140729665756 + ], + "credit_map": [ + 0.015617120079696178, + 0.009795501828193665, + 0.0077885654754936695, + 0.028861450031399727, + 0.017548752948641777, + 0.010237662121653557, + 0.0587611086666584, + 0.016428882256150246, + 0.012715466320514679, + 0.0044672926887869835, + 0.009163969196379185, + 0.0076940348371863365, + 0.0065749757923185825, + 0.005636645946651697, + 0.008425060659646988, + 0.008440439589321613, + 0.0019967423286288977, + 0.0018936576088890433, + 0.0021614280994981527, + 0.006001467816531658, + 0.0030762336682528257, + 0.00353584554977715, + 0.004085610155016184, + 0.006032668519765139, + 0.007658704649657011, + 0.009454723447561264, + 0.005811139941215515, + 0.011381566524505615, + 0.045614093542099, + 0.007658915128558874, + 0.010035177692770958, + 0.030670303851366043, + 0.0011120917042717338, + 0.0016087143449112773, + 0.0008351456490345299, + 0.0009539628517813981, + 0.0015124949859455228, + 0.001282362500205636, + 0.0013098771451041102, + 0.0006038962164893746, + 0.0016164390835911036, + 0.0018619467737153172, + 0.0018970133969560266, + 0.001823753584176302, + 0.001583100762218237, + 0.0014540149131789804, + 0.00228927587158978, + 0.0011539459228515625, + 0.0044797491282224655, + 0.00982135720551014, + 0.00986531376838684, + 0.00898495502769947, + 0.013080799952149391, + 0.006388650741428137, + 0.015236707404255867, + 0.005856954492628574, + 0.0021948509383946657, + 0.0034388417843729258, + 0.004280574154108763, + 0.008824584074318409, + 0.006156417075544596, + 0.009392950683832169, + 0.005442902445793152, + 0.0035258904099464417, + 0.021018136292696, + 0.05147496238350868, + 0.03286430984735489, + 0.044862985610961914, + 0.10269474983215332, + 0.06539260596036911, + 0.06740592420101166, + 0.0370134636759758, + 0.0023193645756691694, + 0.0036234152503311634, + 0.0028333812952041626, + 0.012430517934262753, + 0.004724052269011736, + 0.0030863897409290075, + 0.007143957074731588, + 0.016017060726881027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.12806758284568787 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006746253930032253 + ], + "opposite_logprob_tokens": [ + -5.006746292114258 + ], + "logprob_reward": -0.006746253930032253, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006746253930032253, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14088836312294006, + 0.0645161122083664, + 0.1249222531914711, + 0.03948567062616348, + 0.019618717953562737, + 0.03676645830273628, + 0.16743262112140656, + 0.11350825428962708, + 0.2424076348543167, + 0.05045396462082863 + ], + "credit_map": [ + 0.009885584935545921, + 0.0076655554585158825, + 0.007012506481260061, + 0.03263644129037857, + 0.01406286470592022, + 0.010277298279106617, + 0.04696931689977646, + 0.01237879041582346, + 0.010954223573207855, + 0.00423907907679677, + 0.009696346707642078, + 0.007186304312199354, + 0.006638164632022381, + 0.0064290533773601055, + 0.0077836280688643456, + 0.01158931851387024, + 0.007993302308022976, + 0.006879129447042942, + 0.005516348406672478, + 0.03237500786781311, + 0.012965402565896511, + 0.00724082812666893, + 0.009868212975561619, + 0.04208401218056679, + 0.003914894536137581, + 0.0050031570717692375, + 0.003871324472129345, + 0.0059510781429708, + 0.007873747497797012, + 0.0034614631440490484, + 0.003723894013091922, + 0.0056861094199121, + 0.001563906786032021, + 0.003117164596915245, + 0.0021074768155813217, + 0.00155986740719527, + 0.0037872230168431997, + 0.0018374136416241527, + 0.0032883465755730867, + 0.0023573192302137613, + 0.004798657726496458, + 0.00498422235250473, + 0.0056536379270255566, + 0.005067117512226105, + 0.0035504403058439493, + 0.004251894075423479, + 0.0035256282426416874, + 0.004934858065098524, + 0.01367164682596922, + 0.01966719888150692, + 0.029632503166794777, + 0.022665029391646385, + 0.0254782997071743, + 0.013335776515305042, + 0.022954650223255157, + 0.020027514547109604, + 0.009228931739926338, + 0.009996950626373291, + 0.012768607586622238, + 0.017263052985072136, + 0.0174087081104517, + 0.020939901471138, + 0.014604425057768822, + 0.011297674849629402, + 0.01387124415487051, + 0.01965673826634884, + 0.02752133458852768, + 0.02742592617869377, + 0.051584985107183456, + 0.036615584045648575, + 0.03463074192404747, + 0.031101107597351074, + 0.004310277756303549, + 0.008418832905590534, + 0.003968802746385336, + 0.008089850656688213, + 0.004488661885261536, + 0.0060446420684456825, + 0.005618867930024862, + 0.009514027275145054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.006746253930032253 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253201816231012 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253201816231012, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253201816231012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23124872148036957, + 0.04585901275277138, + 0.3210007846355438, + 0.1238289326429367, + 0.019497817382216454, + 0.014400099404156208, + 0.07446257025003433, + 0.07380034029483795, + 0.08403573930263519, + 0.011865944601595402 + ], + "credit_map": [ + 0.020691506564617157, + 0.012885555624961853, + 0.01001785509288311, + 0.04515141248703003, + 0.02192741446197033, + 0.012846341356635094, + 0.08525818586349487, + 0.022470448166131973, + 0.007911310531198978, + 0.0028875567950308323, + 0.0064244200475513935, + 0.006170119158923626, + 0.004331124946475029, + 0.00454356474801898, + 0.00718011474236846, + 0.006410799454897642, + 0.019810602068901062, + 0.017982838675379753, + 0.011177914217114449, + 0.0748765841126442, + 0.03029578924179077, + 0.029398947954177856, + 0.021863220259547234, + 0.11559489369392395, + 0.007111092563718557, + 0.005461565684527159, + 0.005341012496501207, + 0.007225716020911932, + 0.04734280705451965, + 0.00864835549145937, + 0.011618525721132755, + 0.031079856678843498, + 0.0017758436733856797, + 0.0026881040539592505, + 0.002160601783543825, + 0.0025991073343902826, + 0.0031471012625843287, + 0.0023290058597922325, + 0.0029963653068989515, + 0.0018016868270933628, + 0.0018182909116148949, + 0.0017982356948778033, + 0.0016963790403679013, + 0.001900090486742556, + 0.0014448075089603662, + 0.0014706365764141083, + 0.0017859118524938822, + 0.002485746517777443, + 0.004974688868969679, + 0.007407152093946934, + 0.007659771014004946, + 0.009699572809040546, + 0.01342273224145174, + 0.006314089987426996, + 0.018534807488322258, + 0.006449758540838957, + 0.005101380404084921, + 0.004360905382782221, + 0.005154382903128862, + 0.007983800023794174, + 0.02173185721039772, + 0.01819436624646187, + 0.007136401254683733, + 0.004137251526117325, + 0.007836087606847286, + 0.009009742178022861, + 0.008071037009358406, + 0.00984600093215704, + 0.013631954789161682, + 0.009703490883111954, + 0.01234058104455471, + 0.013596837408840656, + 0.0009762943955138326, + 0.0016646500444039702, + 0.0010476455790922046, + 0.0014668252551928163, + 0.0016097218031063676, + 0.001554853399284184, + 0.0023541508708149195, + 0.001191803952679038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.005253201816231012 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069775447249413 + ], + "opposite_logprob_tokens": [ + -4.1410698890686035 + ], + "logprob_reward": -0.016069775447249413, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069775447249413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2920190989971161, + 0.06858517229557037, + 0.2797178030014038, + 0.06852087378501892, + 0.04826682433485985, + 0.02420797199010849, + 0.05484374985098839, + 0.09581945836544037, + 0.045041993260383606, + 0.02297699823975563 + ], + "credit_map": [ + 0.01808767020702362, + 0.013359966687858105, + 0.010416220873594284, + 0.08039911836385727, + 0.022845081984996796, + 0.01638534478843212, + 0.10598437488079071, + 0.02454133704304695, + 0.011342219077050686, + 0.004438857547938824, + 0.009292465634644032, + 0.01036700326949358, + 0.006510025355964899, + 0.006038198247551918, + 0.010656832717359066, + 0.009939566254615784, + 0.018986770883202553, + 0.01644435152411461, + 0.011263887397944927, + 0.07615689933300018, + 0.029325295239686966, + 0.015116343274712563, + 0.021824948489665985, + 0.09059928357601166, + 0.004674278665333986, + 0.004054444842040539, + 0.0030385551508516073, + 0.004999727942049503, + 0.024802938103675842, + 0.005227336194366217, + 0.006369263399392366, + 0.01535432506352663, + 0.004756180569529533, + 0.006851572543382645, + 0.005167317111045122, + 0.006091105751693249, + 0.008369088172912598, + 0.006357067730277777, + 0.006728861015290022, + 0.003945631440728903, + 0.0033255978487432003, + 0.002359426114708185, + 0.0027379735838621855, + 0.0032319712918251753, + 0.0023094341158866882, + 0.0027353086043149233, + 0.002697388641536236, + 0.004810872487723827, + 0.003288437845185399, + 0.005598614923655987, + 0.009023520164191723, + 0.00820996891707182, + 0.00944178830832243, + 0.005185291171073914, + 0.009852075949311256, + 0.004244053736329079, + 0.005628177430480719, + 0.004353412892669439, + 0.0062360274605453014, + 0.010371271520853043, + 0.027503598481416702, + 0.02738035097718239, + 0.007795589976012707, + 0.006551025900989771, + 0.004393189214169979, + 0.00423789955675602, + 0.0048583997413516045, + 0.006265521049499512, + 0.006101303733885288, + 0.0047276876866817474, + 0.008216091431677341, + 0.006241899449378252, + 0.0015862943837419152, + 0.001765385502949357, + 0.002182700904086232, + 0.00539345433935523, + 0.0016948749544098973, + 0.0033793027978390455, + 0.0027358816005289555, + 0.004239105153828859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.016069775447249413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002820206806063652 + ], + "opposite_logprob_tokens": [ + -5.877820014953613 + ], + "logprob_reward": -0.002820206806063652, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002820206806063652, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4024898409843445, + 0.02027062326669693, + 0.09842820465564728, + 0.03465767949819565, + 0.028908586129546165, + 0.006328764371573925, + 0.09475576132535934, + 0.11113973706960678, + 0.14568635821342468, + 0.05733450502157211 + ], + "credit_map": [ + 0.03389209136366844, + 0.019328515976667404, + 0.015743328258395195, + 0.08807945251464844, + 0.03582495450973511, + 0.020561836659908295, + 0.1514306217432022, + 0.03762901946902275, + 0.002666264772415161, + 0.0017938455566763878, + 0.0027455289382487535, + 0.0024177159648388624, + 0.002718189964070916, + 0.002304578199982643, + 0.002751938533037901, + 0.002872559241950512, + 0.006622178014367819, + 0.0059275198727846146, + 0.004511592909693718, + 0.02641087956726551, + 0.010073843412101269, + 0.005485761910676956, + 0.007443102542310953, + 0.031953323632478714, + 0.0020919854287058115, + 0.0035442402586340904, + 0.002634594216942787, + 0.00419938750565052, + 0.009080502204596996, + 0.002170529682189226, + 0.0031462032347917557, + 0.007790234871208668, + 0.002966903615742922, + 0.004382291808724403, + 0.003325406927615404, + 0.0031652001198381186, + 0.005160341039299965, + 0.00320809637196362, + 0.004726653452962637, + 0.0019736902322620153, + 0.0006117263692431152, + 0.0008888982702046633, + 0.0007722056470811367, + 0.0009703866671770811, + 0.0006989884423092008, + 0.0008129565976560116, + 0.0007566315471194685, + 0.0008169698412530124, + 0.006640959531068802, + 0.007725600153207779, + 0.009672781452536583, + 0.014751683920621872, + 0.014291474595665932, + 0.008382735773921013, + 0.025516707450151443, + 0.00777381332591176, + 0.0063627613708376884, + 0.005727024748921394, + 0.0057571157813072205, + 0.011942924931645393, + 0.035019032657146454, + 0.029918324202299118, + 0.009554951451718807, + 0.006857585161924362, + 0.007782501168549061, + 0.016975445672869682, + 0.014030351303517818, + 0.02364593744277954, + 0.027232319116592407, + 0.01798875257372856, + 0.02302374877035618, + 0.015007277950644493, + 0.0027738974895328283, + 0.0030877080280333757, + 0.0034396974369883537, + 0.015962054952979088, + 0.004590626806020737, + 0.0034760828129947186, + 0.006012811325490475, + 0.017991622909903526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.002820206806063652 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tautly focused close-up shot, beautifully captured at a high shutter speed, focuses on a blond soldier at the peak of his fury. His short, flaming blond hair trembles slightly in the chaotic air currents, his face contorted with utter fury, his eyes wide, and veins bulging in his forehead and neck from exertion. He wears a dark green, scuffed tactical vest over a tight-fitting black combat uniform, clearly outlining the muscular mass of his arms, shoulders, and chest bulging with exertion. Against a dim, rubble-strewn backdrop of battlefield ruins, he leans forward, his legs firmly planted in a lunge. A dazzling, translucent, bluish-white crescent-shaped wave of energy erupts from between his thrusting palms. This wave itself becomes the primary source of light in the scene, casting sharp highlights and deep shadows across his enraged face. Under the extremely clear lens, every bead of sweat splashing from his forehead and the tiny stones shaken up by the energy shock wave were perfectly solidified in the air, presenting a crystal-clear still state. The whole picture was full of explosive power and a tense atmosphere of life and death.", + "seed": 11019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_7373_seed_11019_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016153807379305364, + "vlm_logprob": -0.016153807379305364, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.016153807379305364, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04426629841327667 + ], + "opposite_logprob_tokens": [ + -3.1692662239074707 + ], + "logprob_reward": -0.04426629841327667, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04426629841327667, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04759930074214935, + 0.0696147084236145, + 0.04466640204191208, + 0.0853879451751709, + 0.11554962396621704, + 0.24108092486858368, + 0.21135982871055603, + 0.0761723667383194, + 0.03298328071832657, + 0.07558566331863403 + ], + "credit_map": [ + 0.0051925769075751305, + 0.006211708765476942, + 0.006057078950107098, + 0.005875672213733196, + 0.00652615400031209, + 0.004915333352982998, + 0.006847168784588575, + 0.005973612889647484, + 0.00642292108386755, + 0.007001897320151329, + 0.005326573736965656, + 0.006364033091813326, + 0.0088424663990736, + 0.02087455242872238, + 0.006201402749866247, + 0.00858086347579956, + 0.007083761040121317, + 0.003958074375987053, + 0.003981972578912973, + 0.007140664383769035, + 0.005542324390262365, + 0.005384993739426136, + 0.005940170958638191, + 0.0056344387121498585, + 0.011027620173990726, + 0.010762494057416916, + 0.015025961212813854, + 0.008348105475306511, + 0.010537288151681423, + 0.00977073423564434, + 0.011326330713927746, + 0.008589413948357105, + 0.009845159016549587, + 0.022450504824519157, + 0.008846639655530453, + 0.016089532524347305, + 0.01284499280154705, + 0.026919910684227943, + 0.011041882447898388, + 0.007510996423661709, + 0.01661265641450882, + 0.0487423837184906, + 0.022364221513271332, + 0.06657624989748001, + 0.027878036722540855, + 0.017008772119879723, + 0.029734812676906586, + 0.01216379925608635, + 0.01878511905670166, + 0.042545415461063385, + 0.026941485702991486, + 0.04164038971066475, + 0.026946019381284714, + 0.012660765089094639, + 0.023719605058431625, + 0.018121035769581795, + 0.0033207733649760485, + 0.007421908900141716, + 0.0046808612532913685, + 0.015306013636291027, + 0.0253630168735981, + 0.004891545977443457, + 0.003983022645115852, + 0.011205221526324749, + 0.003090176498517394, + 0.004269102588295937, + 0.003718859516084194, + 0.004109813831746578, + 0.006013413425534964, + 0.0031311202328652143, + 0.005064521916210651, + 0.0035862720105797052, + 0.005533040035516024, + 0.01347185019403696, + 0.006669745314866304, + 0.00863383524119854, + 0.014685765840113163, + 0.011859177611768246, + 0.0072712767869234085, + 0.007460966240614653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.04426629841327667 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011086177080869675 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011086177080869675, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011086177080869675, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05079371854662895, + 0.10331787168979645, + 0.2379683256149292, + 0.054777923971414566, + 0.049615684896707535, + 0.10327564179897308, + 0.09153056144714355, + 0.09931088984012604, + 0.05037900060415268, + 0.15903036296367645 + ], + "credit_map": [ + 0.006401210557669401, + 0.0065008229576051235, + 0.008092238567769527, + 0.00514601357281208, + 0.006814668886363506, + 0.005578662268817425, + 0.00581839494407177, + 0.006441706325858831, + 0.008081379346549511, + 0.007879686541855335, + 0.005999195389449596, + 0.008987465873360634, + 0.013658595271408558, + 0.03667229786515236, + 0.008846943266689777, + 0.013192306272685528, + 0.030398482456803322, + 0.016811635345220566, + 0.02397245168685913, + 0.042017485946416855, + 0.03421055153012276, + 0.02919670380651951, + 0.03150516375899315, + 0.029855836182832718, + 0.00807805173099041, + 0.008292311802506447, + 0.007283750921487808, + 0.006529918871819973, + 0.007337995804846287, + 0.007864165119826794, + 0.005864516366273165, + 0.003527210094034672, + 0.00561842555180192, + 0.008416448719799519, + 0.008119134232401848, + 0.0067270006984472275, + 0.005398715380579233, + 0.004643430933356285, + 0.005518656689673662, + 0.005173868034034967, + 0.00705266697332263, + 0.013498381711542606, + 0.009339887648820877, + 0.021028101444244385, + 0.011702755466103554, + 0.015485486947000027, + 0.015116952359676361, + 0.010051406919956207, + 0.009335001930594444, + 0.01845923624932766, + 0.011977246031165123, + 0.017267020419239998, + 0.010704964399337769, + 0.005743253510445356, + 0.011358299292623997, + 0.006685539614409208, + 0.005833946634083986, + 0.01571284420788288, + 0.009623083285987377, + 0.017467455938458443, + 0.027423830702900887, + 0.005871541332453489, + 0.006371491122990847, + 0.011006690561771393, + 0.0038583322893828154, + 0.005953225307166576, + 0.004227757919579744, + 0.007194461766630411, + 0.01328600849956274, + 0.004713920876383781, + 0.007603212725371122, + 0.003542076563462615, + 0.008316760882735252, + 0.034587930887937546, + 0.012481029145419598, + 0.013384598307311535, + 0.043404143303632736, + 0.017271190881729126, + 0.01911637745797634, + 0.01046832650899887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.011086177080869675 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004094788804650307 + ], + "opposite_logprob_tokens": [ + -5.50409460067749 + ], + "logprob_reward": -0.004094788804650307, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004094788804650307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22038128972053528, + 0.08724281191825867, + 0.07655805349349976, + 0.13565979897975922, + 0.05719318985939026, + 0.059988267719745636, + 0.092183418571949, + 0.08171618729829788, + 0.12559308111667633, + 0.06348393112421036 + ], + "credit_map": [ + 0.029957963153719902, + 0.028432859107851982, + 0.0255887471139431, + 0.022965867072343826, + 0.026989717036485672, + 0.019092250615358353, + 0.02661110833287239, + 0.040742773562669754, + 0.008183656260371208, + 0.007283014245331287, + 0.007313396781682968, + 0.013156477361917496, + 0.009832759387791157, + 0.010714818723499775, + 0.010544363409280777, + 0.020214319229125977, + 0.0065521011129021645, + 0.007149364799261093, + 0.0066722990013659, + 0.008658030070364475, + 0.014801356010138988, + 0.01676073856651783, + 0.009640549309551716, + 0.006323599256575108, + 0.021615924313664436, + 0.010738126933574677, + 0.021759910508990288, + 0.012639171443879604, + 0.015416220761835575, + 0.015133765526115894, + 0.02074538543820381, + 0.017611274495720863, + 0.0056051830761134624, + 0.006633194163441658, + 0.007461173459887505, + 0.00713270902633667, + 0.006060357671231031, + 0.007728101685643196, + 0.010780149139463902, + 0.005792318377643824, + 0.004640123341232538, + 0.005843705032020807, + 0.006287139840424061, + 0.013181631453335285, + 0.006923082750290632, + 0.009666881524026394, + 0.007992534898221493, + 0.005453162826597691, + 0.009479214437305927, + 0.01579897291958332, + 0.01357396598905325, + 0.01607002131640911, + 0.011733156628906727, + 0.007515068165957928, + 0.010954489931464195, + 0.007058522664010525, + 0.00606800802052021, + 0.011338765732944012, + 0.009679430164396763, + 0.012721451930701733, + 0.014508864842355251, + 0.007316854316741228, + 0.011658918112516403, + 0.008423884399235249, + 0.013445163145661354, + 0.015786459669470787, + 0.015296335332095623, + 0.01571081578731537, + 0.018707329407334328, + 0.016378073021769524, + 0.014336120337247849, + 0.015932779759168625, + 0.005341395735740662, + 0.008761066943407059, + 0.006652750540524721, + 0.007178810890763998, + 0.013350000604987144, + 0.007062508258968592, + 0.008948422968387604, + 0.006188971921801567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.004094788804650307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016071535646915436 + ], + "opposite_logprob_tokens": [ + -4.141071319580078 + ], + "logprob_reward": -0.016071535646915436, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016071535646915436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07574883103370667, + 0.055368442088365555, + 0.08503714203834534, + 0.2433934360742569, + 0.11703374236822128, + 0.10056155920028687, + 0.06428656727075577, + 0.18690231442451477, + 0.04188230261206627, + 0.02978568524122238 + ], + "credit_map": [ + 0.009851476177573204, + 0.009862992912530899, + 0.010061218403279781, + 0.007651679217815399, + 0.011302564293146133, + 0.0068937037140131, + 0.011290321126580238, + 0.00883487705141306, + 0.0046701012179255486, + 0.006396015640348196, + 0.0048252553679049015, + 0.0061011603102087975, + 0.009388361126184464, + 0.00914786383509636, + 0.006128141190856695, + 0.00871154386550188, + 0.007462562993168831, + 0.008752713911235332, + 0.006070394068956375, + 0.009945780970156193, + 0.015913672745227814, + 0.009232956916093826, + 0.015338534489274025, + 0.012320528738200665, + 0.034417569637298584, + 0.03376920893788338, + 0.030521659180521965, + 0.027182670310139656, + 0.0351562574505806, + 0.026490043848752975, + 0.029634695500135422, + 0.026221347972750664, + 0.009027710184454918, + 0.014607889577746391, + 0.00628128694370389, + 0.014537502080202103, + 0.016030237078666687, + 0.023247644305229187, + 0.017646243795752525, + 0.01565522886812687, + 0.006417797412723303, + 0.007729355711489916, + 0.0098357442766428, + 0.01578742079436779, + 0.011976946145296097, + 0.02574564702808857, + 0.014359716325998306, + 0.008708931505680084, + 0.00555369583889842, + 0.010574664920568466, + 0.005689918529242277, + 0.012209877371788025, + 0.009055418893694878, + 0.0047302572056651115, + 0.01221269927918911, + 0.004260037560015917, + 0.008220173418521881, + 0.01462829764932394, + 0.007217333186417818, + 0.03868730366230011, + 0.06845930218696594, + 0.011022968217730522, + 0.00920773670077324, + 0.029459189623594284, + 0.00288126477971673, + 0.003362990915775299, + 0.0034801322035491467, + 0.005162845831364393, + 0.013947865925729275, + 0.004605032037943602, + 0.0053683179430663586, + 0.0030738513451069593, + 0.002425573067739606, + 0.0031358867418020964, + 0.003540942445397377, + 0.003079015761613846, + 0.004218257497996092, + 0.006129308138042688, + 0.004268481861799955, + 0.0029882185626775026 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 4 + ], + "selected_reward": -0.016071535646915436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005250236950814724 + ], + "opposite_logprob_tokens": [ + -5.255250453948975 + ], + "logprob_reward": -0.005250236950814724, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005250236950814724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054428212344646454, + 0.06386101990938187, + 0.05422027036547661, + 0.10003755241632462, + 0.09731481224298477, + 0.0901234969496727, + 0.17513272166252136, + 0.29233840107917786, + 0.023274773731827736, + 0.049268726259469986 + ], + "credit_map": [ + 0.006493815220892429, + 0.008630223572254181, + 0.005895461421459913, + 0.006038788706064224, + 0.006360779516398907, + 0.004756877664476633, + 0.008054756559431553, + 0.008197514340281487, + 0.005758891347795725, + 0.005688842851668596, + 0.004529206547886133, + 0.005654104985296726, + 0.007964760065078735, + 0.016746077686548233, + 0.008149601519107819, + 0.00936953816562891, + 0.007248023524880409, + 0.004356340505182743, + 0.004606421571224928, + 0.008225511759519577, + 0.007836573757231236, + 0.007244089152663946, + 0.007836885750293732, + 0.006866426207125187, + 0.013632550835609436, + 0.010920598171651363, + 0.016921089962124825, + 0.009176097810268402, + 0.012616747990250587, + 0.009684127755463123, + 0.014528242871165276, + 0.012558097951114178, + 0.01032002829015255, + 0.02091449499130249, + 0.011188273318111897, + 0.011411979794502258, + 0.010346239432692528, + 0.018657412379980087, + 0.009162535890936852, + 0.005313849076628685, + 0.010521737858653069, + 0.009532999247312546, + 0.0129977623000741, + 0.019036995247006416, + 0.008476822637021542, + 0.01077823992818594, + 0.011297306977212429, + 0.007481630425900221, + 0.015507927164435387, + 0.02978016808629036, + 0.02000352367758751, + 0.029570920392870903, + 0.0208074189722538, + 0.028022494167089462, + 0.01687299646437168, + 0.014567282982170582, + 0.01745603047311306, + 0.03273029252886772, + 0.016309766098856926, + 0.055682603269815445, + 0.09377975016832352, + 0.01725294068455696, + 0.014957273378968239, + 0.04416978731751442, + 0.0025959336198866367, + 0.0036801069509238005, + 0.0027345125563442707, + 0.003795493394136429, + 0.0027089447248727083, + 0.0028246406000107527, + 0.0023244100157171488, + 0.0026107346639037132, + 0.004907655529677868, + 0.008773315697908401, + 0.004029514268040657, + 0.003906647674739361, + 0.011134966276586056, + 0.005742900539189577, + 0.005942760035395622, + 0.0048309676349163055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.005250236950814724 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tautly focused close-up shot, beautifully captured at a high shutter speed, focuses on a blond soldier at the peak of his fury. His short, flaming blond hair trembles slightly in the chaotic air currents, his face contorted with utter fury, his eyes wide, and veins bulging in his forehead and neck from exertion. He wears a dark green, scuffed tactical vest over a tight-fitting black combat uniform, clearly outlining the muscular mass of his arms, shoulders, and chest bulging with exertion. Against a dim, rubble-strewn backdrop of battlefield ruins, he leans forward, his legs firmly planted in a lunge. A dazzling, translucent, bluish-white crescent-shaped wave of energy erupts from between his thrusting palms. This wave itself becomes the primary source of light in the scene, casting sharp highlights and deep shadows across his enraged face. Under the extremely clear lens, every bead of sweat splashing from his forehead and the tiny stones shaken up by the energy shock wave were perfectly solidified in the air, presenting a crystal-clear still state. The whole picture was full of explosive power and a tense atmosphere of life and death.", + "seed": 11020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_7373_seed_11020_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.058264664467424156, + "vlm_logprob": -0.058264664467424156, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.058264664467424156, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25286170840263367 + ], + "opposite_logprob_tokens": [ + -1.502861738204956 + ], + "logprob_reward": -0.25286170840263367, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25286170840263367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10541848093271255, + 0.18413473665714264, + 0.08834392577409744, + 0.08477349579334259, + 0.012109511531889439, + 0.026397552341222763, + 0.04482964426279068, + 0.042147196829319, + 0.181039497256279, + 0.23080596327781677 + ], + "credit_map": [ + 0.007166809868067503, + 0.009243178181350231, + 0.006063372828066349, + 0.03283317759633064, + 0.006195887923240662, + 0.010661601088941097, + 0.009138424880802631, + 0.02411603182554245, + 0.011575614102184772, + 0.034399233758449554, + 0.020417461171746254, + 0.037755753844976425, + 0.01977461762726307, + 0.019758110865950584, + 0.02615942247211933, + 0.014294536784291267, + 0.005854772869497538, + 0.01072821207344532, + 0.014913379214704037, + 0.010817247442901134, + 0.017088262364268303, + 0.007336531300097704, + 0.015374702401459217, + 0.00623082322999835, + 0.0074821044690907, + 0.013705922290682793, + 0.013399792835116386, + 0.011793422512710094, + 0.01313263364136219, + 0.006175487767904997, + 0.009913931600749493, + 0.009170208126306534, + 0.002006255090236664, + 0.0011722701601684093, + 0.0014812714653089643, + 0.0015281725209206343, + 0.001349200145341456, + 0.001532499329186976, + 0.0016261935234069824, + 0.0014136508107185364, + 0.0044590868055820465, + 0.00408348860219121, + 0.0039942278526723385, + 0.003986869473010302, + 0.0024296664632856846, + 0.0033909897319972515, + 0.0024203432258218527, + 0.0016328823985531926, + 0.005592058412730694, + 0.005995277315378189, + 0.005250053480267525, + 0.0053719040006399155, + 0.008606219664216042, + 0.005467374809086323, + 0.004434334579855204, + 0.004112426191568375, + 0.005448358599096537, + 0.004725094418972731, + 0.005745717324316502, + 0.005039778538048267, + 0.005201167427003384, + 0.003373063402250409, + 0.009340980090200901, + 0.0032730421517044306, + 0.006694288458675146, + 0.007543726824223995, + 0.006869544740766287, + 0.06872164458036423, + 0.009213718585669994, + 0.005284040234982967, + 0.07038017362356186, + 0.006332372315227985, + 0.024540459737181664, + 0.021716095507144928, + 0.023440930992364883, + 0.05190887302160263, + 0.03536147251725197, + 0.02002382092177868, + 0.03338291123509407, + 0.020431427285075188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 8 + ], + "selected_reward": -0.25286170840263367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007638176903128624 + ], + "opposite_logprob_tokens": [ + -4.882637977600098 + ], + "logprob_reward": -0.007638176903128624, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007638176903128624, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.038329530507326126, + 0.059962403029203415, + 0.09753230959177017, + 0.043109338730573654, + 0.05431738123297691, + 0.13357074558734894, + 0.05664025992155075, + 0.04341049864888191, + 0.23326213657855988, + 0.2398654669523239 + ], + "credit_map": [ + 0.002215906511992216, + 0.003427928313612938, + 0.003579221898689866, + 0.009986229240894318, + 0.0021442577708512545, + 0.004627784248441458, + 0.004101807717233896, + 0.008246387355029583, + 0.003603126620873809, + 0.009381810203194618, + 0.0086147990077734, + 0.011975102126598358, + 0.006862649694085121, + 0.006536117289215326, + 0.008345276117324829, + 0.0046435086987912655, + 0.005958955734968185, + 0.007638381328433752, + 0.019038857892155647, + 0.011239471845328808, + 0.01381253357976675, + 0.014239630661904812, + 0.018652481958270073, + 0.006951982621103525, + 0.005367424804717302, + 0.004234474152326584, + 0.006420797668397427, + 0.0033176164142787457, + 0.004950776696205139, + 0.004936163779348135, + 0.008362279273569584, + 0.0055198026821017265, + 0.004853354301303625, + 0.003607367631047964, + 0.007566039450466633, + 0.01263698935508728, + 0.004149638582020998, + 0.005377584602683783, + 0.011279354803264141, + 0.004847043193876743, + 0.028390252962708473, + 0.01105364877730608, + 0.039345018565654755, + 0.017088187858462334, + 0.008352212607860565, + 0.013783521950244904, + 0.007901218719780445, + 0.007656663656234741, + 0.0043154386803507805, + 0.008436892181634903, + 0.007506109774112701, + 0.00723280431702733, + 0.008150147274136543, + 0.01058797538280487, + 0.006029027514159679, + 0.00438185827806592, + 0.004091824870556593, + 0.003182097338140011, + 0.005829072091728449, + 0.0041450150310993195, + 0.011104388162493706, + 0.003140331245958805, + 0.009007959626615047, + 0.0029098023660480976, + 0.008982150815427303, + 0.008577783592045307, + 0.007234702352434397, + 0.0930599644780159, + 0.008781577460467815, + 0.0041311257518827915, + 0.09392023086547852, + 0.00857458170503378, + 0.02931429073214531, + 0.018882984295487404, + 0.03544283285737038, + 0.026297198608517647, + 0.033972304314374924, + 0.03764287382364273, + 0.024762600660324097, + 0.033550336956977844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 5 + ], + "selected_reward": -0.007638176903128624 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012536530382931232 + ], + "opposite_logprob_tokens": [ + -4.387536525726318 + ], + "logprob_reward": -0.012536530382931232, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012536530382931232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10767849534749985, + 0.02987685054540634, + 0.045390110462903976, + 0.014260916970670223, + 0.4937070310115814, + 0.035900842398405075, + 0.023518046364188194, + 0.011198652908205986, + 0.20719929039478302, + 0.03126969188451767 + ], + "credit_map": [ + 0.007606204133480787, + 0.009084965102374554, + 0.006126357242465019, + 0.03264975920319557, + 0.006421941332519054, + 0.012546095065772533, + 0.009760964661836624, + 0.023482201620936394, + 0.0025501290801912546, + 0.0029918397776782513, + 0.0035085955169051886, + 0.0042093428783118725, + 0.0032750340178608894, + 0.0036612211260944605, + 0.004243231378495693, + 0.0054374560713768005, + 0.004192030057311058, + 0.0032293577678501606, + 0.008972130715847015, + 0.004244182724505663, + 0.004233256448060274, + 0.004826030693948269, + 0.009654401801526546, + 0.006038723047822714, + 0.001849376829341054, + 0.0014454692136496305, + 0.0016795515548437834, + 0.0012415562523528934, + 0.002315906109288335, + 0.001332723069936037, + 0.00205708434805274, + 0.002339249476790428, + 0.05122947320342064, + 0.03174998611211777, + 0.080013707280159, + 0.11974355578422546, + 0.03213413432240486, + 0.04513219743967056, + 0.09210618585348129, + 0.041597817093133926, + 0.004574749618768692, + 0.00307616894133389, + 0.003467773785814643, + 0.0050298129208385944, + 0.0034460534807294607, + 0.005019059870392084, + 0.007319736760109663, + 0.003967485390603542, + 0.0015348566230386496, + 0.0038258368149399757, + 0.0023226046469062567, + 0.0029125686269253492, + 0.006325228605419397, + 0.0022803505416959524, + 0.0027665046509355307, + 0.0015500939916819334, + 0.0017510215984657407, + 0.0009091707179322839, + 0.0020693663973361254, + 0.0010632849298417568, + 0.0012650477001443505, + 0.0009661923977546394, + 0.0017568275798112154, + 0.0014177414122968912, + 0.007569458335638046, + 0.007207483984529972, + 0.005099112633615732, + 0.08275432884693146, + 0.008127407170832157, + 0.002786888973787427, + 0.08648073673248291, + 0.00717386556789279, + 0.003206852125003934, + 0.0033232492860406637, + 0.004469576291739941, + 0.004234201740473509, + 0.00429023802280426, + 0.004586221184581518, + 0.0038213881198316813, + 0.0033379674423485994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.012536530382931232 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014192949049174786 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014192949049174786, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014192949049174786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09831781685352325, + 0.03126616030931473, + 0.0488467738032341, + 0.02461337111890316, + 0.06217209994792938, + 0.02333925850689411, + 0.029326187446713448, + 0.02340390533208847, + 0.4548986256122589, + 0.20381589233875275 + ], + "credit_map": [ + 0.0071306535974144936, + 0.008175911381840706, + 0.009740820154547691, + 0.02452617697417736, + 0.007848224602639675, + 0.012175406329333782, + 0.009586920961737633, + 0.019133688881993294, + 0.0018836419330909848, + 0.006242962088435888, + 0.004944831132888794, + 0.003960566595196724, + 0.004581979941576719, + 0.0024872105568647385, + 0.004602538421750069, + 0.0025624253321439028, + 0.002536757616326213, + 0.0022964985109865665, + 0.011428183875977993, + 0.004717459436506033, + 0.0042500258423388, + 0.0037838921416550875, + 0.012750961817800999, + 0.007082990370690823, + 0.0032829679548740387, + 0.002463739598169923, + 0.0038752437103539705, + 0.002394084818661213, + 0.002620079554617405, + 0.0029768897220492363, + 0.0029074682388454676, + 0.00409289263188839, + 0.005437745712697506, + 0.004841706715524197, + 0.011112719774246216, + 0.010950912721455097, + 0.005566478706896305, + 0.0061192987486720085, + 0.010308602824807167, + 0.007834624499082565, + 0.004880569409579039, + 0.001887061633169651, + 0.004319861065596342, + 0.002777885878458619, + 0.0015830089105293155, + 0.0025628223083913326, + 0.0033250905107706785, + 0.0020029570441693068, + 0.0026164986193180084, + 0.0031237935181707144, + 0.003657425520941615, + 0.0034928652457892895, + 0.0064431666396558285, + 0.003650517435744405, + 0.003428200026974082, + 0.00291371438652277, + 0.0028231150936335325, + 0.0019630221650004387, + 0.0029149616602808237, + 0.0018852388020604849, + 0.0050285267643630505, + 0.0022755740210413933, + 0.004828281234949827, + 0.0016851802356541157, + 0.017973948270082474, + 0.015415964648127556, + 0.01252524834126234, + 0.1782420575618744, + 0.019227486103773117, + 0.006982059217989445, + 0.18866777420043945, + 0.015863995999097824, + 0.025732791051268578, + 0.015880681574344635, + 0.02804568037390709, + 0.02280406467616558, + 0.03417903557419777, + 0.02662595920264721, + 0.02086711674928665, + 0.029680542647838593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.014192949049174786 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093957599252462 + ], + "opposite_logprob_tokens": [ + -5.504094123840332 + ], + "logprob_reward": -0.004093957599252462, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093957599252462, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029795793816447258, + 0.05426523834466934, + 0.05095808953046799, + 0.03491964936256409, + 0.06891214102506638, + 0.01713922992348671, + 0.12780073285102844, + 0.05490465834736824, + 0.44114333391189575, + 0.12016106396913528 + ], + "credit_map": [ + 0.002893614349886775, + 0.003165858332067728, + 0.005930395796895027, + 0.004085118882358074, + 0.0028316867537796497, + 0.0033259098418056965, + 0.0037315564695745707, + 0.0038316543214023113, + 0.005861739162355661, + 0.007982982322573662, + 0.007078306749463081, + 0.009791205637156963, + 0.0057355621829628944, + 0.004332696087658405, + 0.007092082407325506, + 0.006390662863850594, + 0.003308730199933052, + 0.00606849230825901, + 0.010088508017361164, + 0.004935143981128931, + 0.008780132047832012, + 0.0033170573879033327, + 0.01086218561977148, + 0.0035978369414806366, + 0.0033764224499464035, + 0.007722294889390469, + 0.004888995550572872, + 0.003892771666869521, + 0.005632154177874327, + 0.0034614813048392534, + 0.003207216504961252, + 0.002738313050940633, + 0.007393541280180216, + 0.005537348799407482, + 0.009513122960925102, + 0.012650999240577221, + 0.006583233829587698, + 0.008291811682283878, + 0.01181200984865427, + 0.007130071520805359, + 0.0022919350303709507, + 0.0017516607185825706, + 0.0029344779904931784, + 0.0022253633942455053, + 0.002148575382307172, + 0.001901006791740656, + 0.0025470952969044447, + 0.0013391153188422322, + 0.00790233165025711, + 0.021528666839003563, + 0.011894249357283115, + 0.016754860058426857, + 0.029432276263833046, + 0.01610446721315384, + 0.015993239358067513, + 0.00819065235555172, + 0.008653361350297928, + 0.006096167024224997, + 0.007652204483747482, + 0.007855565287172794, + 0.005537844263017178, + 0.006399349309504032, + 0.008174320682883263, + 0.004535848740488291, + 0.017261208966374397, + 0.01833377592265606, + 0.020840685814619064, + 0.16655437648296356, + 0.018821794539690018, + 0.01213766634464264, + 0.17009097337722778, + 0.01710284687578678, + 0.01332393754273653, + 0.013083184137940407, + 0.013915405608713627, + 0.015377831645309925, + 0.02250177226960659, + 0.014018996618688107, + 0.017547503113746643, + 0.010392441414296627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.004093957599252462 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, tautly focused close-up shot, beautifully captured at a high shutter speed, focuses on a blond soldier at the peak of his fury. His short, flaming blond hair trembles slightly in the chaotic air currents, his face contorted with utter fury, his eyes wide, and veins bulging in his forehead and neck from exertion. He wears a dark green, scuffed tactical vest over a tight-fitting black combat uniform, clearly outlining the muscular mass of his arms, shoulders, and chest bulging with exertion. Against a dim, rubble-strewn backdrop of battlefield ruins, he leans forward, his legs firmly planted in a lunge. A dazzling, translucent, bluish-white crescent-shaped wave of energy erupts from between his thrusting palms. This wave itself becomes the primary source of light in the scene, casting sharp highlights and deep shadows across his enraged face. Under the extremely clear lens, every bead of sweat splashing from his forehead and the tiny stones shaken up by the energy shock wave were perfectly solidified in the air, presenting a crystal-clear still state. The whole picture was full of explosive power and a tense atmosphere of life and death.", + "seed": 11021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_8/index_7373_seed_11021_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11011099051684141, + "vlm_logprob": -0.11011099051684141, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.11011099051684141, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4751584529876709 + ], + "opposite_logprob_tokens": [ + -0.9751584529876709 + ], + "logprob_reward": -0.4751584529876709, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4751584529876709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10830328613519669, + 0.07037215679883957, + 0.0564032606780529, + 0.02986275590956211, + 0.04183448106050491, + 0.018862148746848106, + 0.010099471546709538, + 0.10569242388010025, + 0.2242414504289627, + 0.3343285322189331 + ], + "credit_map": [ + 0.009550617076456547, + 0.010864492505788803, + 0.016254490241408348, + 0.014946423470973969, + 0.014895608648657799, + 0.01544902939349413, + 0.013818242587149143, + 0.01252437848597765, + 0.006592773832380772, + 0.009434563107788563, + 0.007040560245513916, + 0.004242901690304279, + 0.02156290039420128, + 0.006732459645718336, + 0.0056609222665429115, + 0.0091050760820508, + 0.005214549135416746, + 0.00449470104649663, + 0.007687213830649853, + 0.008327143266797066, + 0.005674860440194607, + 0.007208110764622688, + 0.012080617249011993, + 0.005716066341847181, + 0.0038303944747895002, + 0.002619568957015872, + 0.0034686410799622536, + 0.0032833374571055174, + 0.0031946543604135513, + 0.004103393759578466, + 0.005465666763484478, + 0.0038970964960753918, + 0.0029384670779109, + 0.0023754993453621864, + 0.0022238604724407196, + 0.0018900209106504917, + 0.013658024370670319, + 0.013660707511007786, + 0.0024743801914155483, + 0.0026135204825550318, + 0.002579173306003213, + 0.002576927188783884, + 0.0026108059100806713, + 0.0020090509206056595, + 0.0026764736976474524, + 0.0015528140356764197, + 0.0022699006367474794, + 0.002587003167718649, + 0.001401696354150772, + 0.0014193546958267689, + 0.0016433466225862503, + 0.0021225891541689634, + 0.0007830475806258619, + 0.0010118253994733095, + 0.0008433176553808153, + 0.0008742944337427616, + 0.00896001048386097, + 0.010816408321261406, + 0.013979888521134853, + 0.01508578471839428, + 0.014172472059726715, + 0.00703493831679225, + 0.022523649036884308, + 0.01311927568167448, + 0.03653763607144356, + 0.012435807846486568, + 0.08963222801685333, + 0.053555928170681, + 0.007462930865585804, + 0.008927139453589916, + 0.007974574342370033, + 0.007715214975178242, + 0.026900241151452065, + 0.04484403878450394, + 0.03603631258010864, + 0.037973664700984955, + 0.07795385271310806, + 0.029399853199720383, + 0.03836383298039436, + 0.04285672679543495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.4751584529876709 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018211886286735535 + ], + "opposite_logprob_tokens": [ + -4.018211841583252 + ], + "logprob_reward": -0.018211886286735535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018211886286735535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12356194108724594, + 0.13274356722831726, + 0.05905841663479805, + 0.044919777661561966, + 0.03622167184948921, + 0.018948709592223167, + 0.010042276233434677, + 0.042458102107048035, + 0.3863740861415863, + 0.14567141234874725 + ], + "credit_map": [ + 0.01340749766677618, + 0.017175370827317238, + 0.021430760622024536, + 0.02289034053683281, + 0.01174574252218008, + 0.011837365105748177, + 0.015804052352905273, + 0.00927081611007452, + 0.014288865961134434, + 0.019863426685333252, + 0.01809065230190754, + 0.011459565721452236, + 0.02903446927666664, + 0.010599474422633648, + 0.011433563195168972, + 0.017973553389310837, + 0.0066641634330153465, + 0.006247047800570726, + 0.011226306669414043, + 0.007638475392013788, + 0.0068555232137441635, + 0.007308019790798426, + 0.0069050975143909454, + 0.006213782355189323, + 0.006452586036175489, + 0.0045138029381632805, + 0.006214909255504608, + 0.006319679785519838, + 0.005187760107219219, + 0.005268683657050133, + 0.0065424940548837185, + 0.0044198608957231045, + 0.0036101010628044605, + 0.0022177465725690126, + 0.010962202213704586, + 0.0025322039145976305, + 0.0028554927557706833, + 0.009598824195563793, + 0.0021979922894388437, + 0.0022471053525805473, + 0.0032906378619372845, + 0.0024020366836339235, + 0.00246760924346745, + 0.002548412885516882, + 0.0018527940846979618, + 0.0016652608755975962, + 0.002297044964507222, + 0.0024249153211712837, + 0.0014613880775868893, + 0.001519385608844459, + 0.0013839463936164975, + 0.0011072979541495442, + 0.001155017875134945, + 0.0010406639194115996, + 0.0013315532123669982, + 0.0010430238908156753, + 0.007097864523530006, + 0.0058914534747600555, + 0.005830378737300634, + 0.006435689050704241, + 0.004753148648887873, + 0.003241227939724922, + 0.005460723768919706, + 0.0037476159632205963, + 0.06458504498004913, + 0.026746196672320366, + 0.15109175443649292, + 0.08298047631978989, + 0.017413778230547905, + 0.016752460971474648, + 0.012977566570043564, + 0.01382682379335165, + 0.012289549224078655, + 0.021175287663936615, + 0.01668519154191017, + 0.017927538603544235, + 0.03183237835764885, + 0.011503738351166248, + 0.01675337739288807, + 0.017504343762993813 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 1 + ], + "selected_reward": -0.018211886286735535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004097281955182552 + ], + "opposite_logprob_tokens": [ + -5.5040974617004395 + ], + "logprob_reward": -0.004097281955182552, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004097281955182552, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3247077167034149, + 0.07765133678913116, + 0.09652890264987946, + 0.055441468954086304, + 0.03931276127696037, + 0.0406859815120697, + 0.03492553159594536, + 0.14241209626197815, + 0.09628172963857651, + 0.0920524150133133 + ], + "credit_map": [ + 0.02490730956196785, + 0.021089406684041023, + 0.05836249142885208, + 0.03307044878602028, + 0.06249050796031952, + 0.027391280978918076, + 0.06568655371665955, + 0.03170975670218468, + 0.00849907286465168, + 0.010042688809335232, + 0.008616955950856209, + 0.007145662792026997, + 0.01627005636692047, + 0.005122343078255653, + 0.010258128866553307, + 0.011696434579789639, + 0.007793471682816744, + 0.008116299286484718, + 0.011368141509592533, + 0.011269962415099144, + 0.007750951685011387, + 0.013194864615797997, + 0.02611548826098442, + 0.01091972179710865, + 0.007121761795133352, + 0.004074450582265854, + 0.008070109412074089, + 0.006723407190293074, + 0.006176423747092485, + 0.008927400223910809, + 0.008594413287937641, + 0.005753503646701574, + 0.004007812589406967, + 0.005114940460771322, + 0.00394041370600462, + 0.005234090145677328, + 0.004192540422081947, + 0.00566956726834178, + 0.006740838289260864, + 0.00441256258636713, + 0.00371527299284935, + 0.00589523371309042, + 0.003599021350964904, + 0.004132022615522146, + 0.008324961178004742, + 0.006456400733441114, + 0.00508986646309495, + 0.0034732059575617313, + 0.0037618721835315228, + 0.004919444676488638, + 0.004441394004970789, + 0.003969191107898951, + 0.004452493507415056, + 0.0032329061068594456, + 0.0044641317799687386, + 0.005684100091457367, + 0.013674394227564335, + 0.015929535031318665, + 0.018250171095132828, + 0.01751432940363884, + 0.016349835321307182, + 0.009018602780997753, + 0.030325261875987053, + 0.02134997583925724, + 0.02108232118189335, + 0.008133490569889545, + 0.03006657585501671, + 0.010792783461511135, + 0.0065656076185405254, + 0.006822169292718172, + 0.005263843573629856, + 0.007554950658231974, + 0.009851890616118908, + 0.01138173695653677, + 0.012746668420732021, + 0.011446175165474415, + 0.018129197880625725, + 0.00906068179756403, + 0.009637566283345222, + 0.009798502549529076 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.004097281955182552 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328958734869957 + ], + "opposite_logprob_tokens": [ + -3.773289680480957 + ], + "logprob_reward": -0.02328958734869957, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328958734869957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14013567566871643, + 0.15205150842666626, + 0.053914882242679596, + 0.05970150604844093, + 0.11300325393676758, + 0.03946848213672638, + 0.01463343296200037, + 0.05113082751631737, + 0.30241188406944275, + 0.07354854047298431 + ], + "credit_map": [ + 0.015377134084701538, + 0.01439129002392292, + 0.023984212428331375, + 0.021211739629507065, + 0.02305934578180313, + 0.015194585546851158, + 0.016290394589304924, + 0.010626958683133125, + 0.01713383011519909, + 0.02147897332906723, + 0.023005375638604164, + 0.015519584529101849, + 0.030404020100831985, + 0.010805885307490826, + 0.01445192750543356, + 0.019251901656389236, + 0.007513066753745079, + 0.005591737572103739, + 0.009501281194388866, + 0.006906238850206137, + 0.006972269620746374, + 0.006850471254438162, + 0.004959655459970236, + 0.005620163399726152, + 0.007677615620195866, + 0.007564395200461149, + 0.009283029474318027, + 0.0069555859081447124, + 0.006736855488270521, + 0.007593058981001377, + 0.007544250693172216, + 0.006346710026264191, + 0.008509484119713306, + 0.005087858531624079, + 0.027282394468784332, + 0.006396036595106125, + 0.01610252447426319, + 0.04037851095199585, + 0.004225463140755892, + 0.005020973738282919, + 0.00631678756326437, + 0.0056794374249875546, + 0.005712742451578379, + 0.004376641940325499, + 0.004231744445860386, + 0.003992386627942324, + 0.00461182463914156, + 0.004546915180981159, + 0.0024210656993091106, + 0.0022909489925950766, + 0.0023112448398023844, + 0.002219005022197962, + 0.0013844884233549237, + 0.001401087618432939, + 0.0013698858674615622, + 0.0012357053346931934, + 0.008390925824642181, + 0.006994101218879223, + 0.006713502109050751, + 0.007167511619627476, + 0.006056691519916058, + 0.004173659719526768, + 0.005117357708513737, + 0.006517076399177313, + 0.05756776034832001, + 0.01864679716527462, + 0.12478505820035934, + 0.06392084807157516, + 0.009396887384355068, + 0.011228951625525951, + 0.008236085996031761, + 0.008629501797258854, + 0.0072283376939594746, + 0.00920869316905737, + 0.011629633605480194, + 0.009556008502840996, + 0.014230991713702679, + 0.006609524600207806, + 0.007871691137552261, + 0.00721365213394165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.02328958734869957 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a blond soldier with short, flaming blond hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the soldier wearing a dark green, scuffed tactical vest over a black combat uniform?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a bluish-white crescent-shaped wave of energy erupting from the soldier's palms?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a dim, rubble-strewn battlefield with visible beads of sweat and tiny stones frozen in mid-air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a close-up where the soldier's face is contorted with fury, showing wide eyes and bulging veins?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029797744005918503 + ], + "opposite_logprob_tokens": [ + -3.5297977924346924 + ], + "logprob_reward": -0.029797744005918503, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029797744005918503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11248888075351715, + 0.09577915072441101, + 0.0770520344376564, + 0.038371678441762924, + 0.0629529282450676, + 0.027127670124173164, + 0.016923967748880386, + 0.1278100460767746, + 0.19083647429943085, + 0.25065717101097107 + ], + "credit_map": [ + 0.009287341497838497, + 0.008829492144286633, + 0.021905196830630302, + 0.015495127998292446, + 0.01451666746288538, + 0.015560434199869633, + 0.013964778743684292, + 0.012929849326610565, + 0.008512146770954132, + 0.01384278479963541, + 0.011865707114338875, + 0.007077834103256464, + 0.023675568401813507, + 0.008178004994988441, + 0.008433709852397442, + 0.014193403534591198, + 0.007359794341027737, + 0.0075391060672700405, + 0.01383108738809824, + 0.009327187202870846, + 0.00856349803507328, + 0.012097363360226154, + 0.010670600458979607, + 0.007663403172045946, + 0.004481355659663677, + 0.004258761182427406, + 0.005641665309667587, + 0.005756472237408161, + 0.004349557217210531, + 0.005624806508421898, + 0.004916083067655563, + 0.0033429760951548815, + 0.005644467659294605, + 0.003207785077393055, + 0.017944324761629105, + 0.0044265566393733025, + 0.0051932306960225105, + 0.02037762850522995, + 0.0030559066217392683, + 0.003103026421740651, + 0.0050233276560902596, + 0.003382441820576787, + 0.003911491017788649, + 0.003763866610825062, + 0.002769825281575322, + 0.003169367788359523, + 0.0030012172646820545, + 0.002106134779751301, + 0.001977312844246626, + 0.0018539066659286618, + 0.002389964647591114, + 0.0025351503863930702, + 0.0021037026308476925, + 0.0022172625176608562, + 0.0022294705267995596, + 0.0016171978786587715, + 0.012101237662136555, + 0.012345556169748306, + 0.01760266348719597, + 0.02101537398993969, + 0.016760746017098427, + 0.008328245021402836, + 0.02441723644733429, + 0.01523899007588625, + 0.023626942187547684, + 0.023110870271921158, + 0.04918385669589043, + 0.033117372542619705, + 0.022017745301127434, + 0.01352175883948803, + 0.011357473209500313, + 0.014900445938110352, + 0.02165394462645054, + 0.03420739620923996, + 0.027400238439440727, + 0.029351964592933655, + 0.05255423113703728, + 0.026541506871581078, + 0.025479162111878395, + 0.033468734472990036 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.029797744005918503 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_tensor([554, 554])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_tensor([554, 554])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..a56a25776cc86a482627c28af4b4b69dc98cebf2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_tensor([554, 554])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "close-up, static shot from an extremely low angle, draws the viewer's gaze squarely to the weathered face of an elderly woman. A soft, golden beam of afternoon sunlight slants in from the side of the frame, meticulously sculpting every wrinkle in her features. Deep crow's feet and nasolabial folds resemble a dry riverbed, and faint age spots dot her skin. A few strands of her thinning silver-white hair, ruffled by the breeze, lay plastered to her temple. Her cloudy, yet luminous eyes gaze upward, their rims reddened, her lips pursed, her chin trembling imperceptibly with suppressed emotion. In this silence, a full, glistening tear slowly escapes from the corner of her right eye. It lingers briefly for half a second before steadily and leisurely sliding down a deep wrinkle on her cheek, like a river flowing in a predetermined path. The tear reflects the warm sunlight as it glides, leaving a wet, glistening trail before disappearing from the camera's field of view. The entire picture is dominated by warm gold and deep shadows under the strong contrast of light and shadow, creating a quiet, deep and melancholy atmosphere full of the sense of time. The background is completely blurred into a blurred warm light spot.", + "seed": 16068, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_554_seed_16068_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.15471226591616868, + "vlm_logprob": -0.15471226591616868, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.15471226591616868, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018472423776984215 + ], + "opposite_logprob_tokens": [ + -4.768472194671631 + ], + "logprob_reward": -0.018472423776984215, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.018472423776984215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14808803796768188, + 0.10378068685531616, + 0.08016280084848404, + 0.040499843657016754, + 0.0367923304438591, + 0.02764143794775009, + 0.14605548977851868, + 0.1707945019006729, + 0.0756308063864708, + 0.170554056763649 + ], + "credit_map": [ + 0.012949868105351925, + 0.02664157748222351, + 0.016117876395583153, + 0.025363760069012642, + 0.0191356148570776, + 0.014531094580888748, + 0.02104816399514675, + 0.012300089932978153, + 0.011090157553553581, + 0.013264539651572704, + 0.009549045003950596, + 0.013539143837988377, + 0.01593911275267601, + 0.01534905657172203, + 0.014530694112181664, + 0.010518940165638924, + 0.01652337796986103, + 0.009213426150381565, + 0.006424889434129, + 0.011366704478859901, + 0.009496107697486877, + 0.010094515047967434, + 0.009659670293331146, + 0.007384105119854212, + 0.004688742104917765, + 0.005347299389541149, + 0.002764313481748104, + 0.006279415916651487, + 0.0036631880793720484, + 0.006966070272028446, + 0.005679713562130928, + 0.005111100152134895, + 0.009250793606042862, + 0.004481330048292875, + 0.002595816971734166, + 0.004406555090099573, + 0.003275627037510276, + 0.004884494934231043, + 0.004019642248749733, + 0.003878069343045354, + 0.006305290851742029, + 0.002589255804196, + 0.002738828305155039, + 0.0035362180788069963, + 0.0025479907635599375, + 0.004156702198088169, + 0.0028208286967128515, + 0.0029463227838277817, + 0.03927495703101158, + 0.012686589732766151, + 0.013505778275430202, + 0.020058268681168556, + 0.008793974295258522, + 0.024275369942188263, + 0.009099164977669716, + 0.018361395224928856, + 0.046207308769226074, + 0.017678378149867058, + 0.015110500156879425, + 0.02034713327884674, + 0.0119051868095994, + 0.027830783277750015, + 0.010257814079523087, + 0.02145739085972309, + 0.015981024131178856, + 0.005274498835206032, + 0.007054279558360577, + 0.008570021018385887, + 0.004879009444266558, + 0.015919845551252365, + 0.004667716100811958, + 0.013284411281347275, + 0.04150453954935074, + 0.012464504688978195, + 0.02058415114879608, + 0.01977742649614811, + 0.011165004223585129, + 0.02942756935954094, + 0.01205009687691927, + 0.023580752313137054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.018472423776984215 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16042426228523254 + ], + "opposite_logprob_tokens": [ + -1.9104242324829102 + ], + "logprob_reward": -0.16042426228523254, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16042426228523254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15947073698043823, + 0.05104701220989227, + 0.10666321218013763, + 0.1303207129240036, + 0.04277477413415909, + 0.11572082340717316, + 0.17844906449317932, + 0.1146593913435936, + 0.023305952548980713, + 0.0775882825255394 + ], + "credit_map": [ + 0.01743469201028347, + 0.013886294327676296, + 0.009915720671415329, + 0.016933634877204895, + 0.012669733725488186, + 0.04304148629307747, + 0.010111921466886997, + 0.03547725826501846, + 0.006610284559428692, + 0.006154386792331934, + 0.00777042843401432, + 0.008552225306630135, + 0.005198645405471325, + 0.006972710136324167, + 0.005225134082138538, + 0.004563198424875736, + 0.028226802125573158, + 0.00881896447390318, + 0.014821181073784828, + 0.012495285831391811, + 0.006721709854900837, + 0.016558440402150154, + 0.006683951243758202, + 0.012336879037320614, + 0.02961675263941288, + 0.010764164850115776, + 0.012631515972316265, + 0.013620886020362377, + 0.008785849437117577, + 0.025365693494677544, + 0.008944044820964336, + 0.020591814070940018, + 0.010789426043629646, + 0.005354073829948902, + 0.00399816082790494, + 0.005408193450421095, + 0.0033258874900639057, + 0.005396953783929348, + 0.0035404805094003677, + 0.004961599130183458, + 0.0272932481020689, + 0.007783855311572552, + 0.008981066755950451, + 0.01211409829556942, + 0.005225663539022207, + 0.026785988360643387, + 0.005371786188334227, + 0.022165119647979736, + 0.04278945177793503, + 0.01640583574771881, + 0.020516887307167053, + 0.023865139111876488, + 0.012728105299174786, + 0.0289210993796587, + 0.013674727641046047, + 0.019547805190086365, + 0.03342606872320175, + 0.012844396755099297, + 0.009850403293967247, + 0.016039203852415085, + 0.008056114427745342, + 0.015183431096374989, + 0.00807082001119852, + 0.011188961565494537, + 0.004781513940542936, + 0.0034199641086161137, + 0.001844012294895947, + 0.0030931199435144663, + 0.0022721022833138704, + 0.0026059304364025593, + 0.0027933097444474697, + 0.002496002009138465, + 0.023202206939458847, + 0.007226990070194006, + 0.00758477533236146, + 0.009485677815973759, + 0.005026544909924269, + 0.011012104339897633, + 0.00515930587425828, + 0.008890670724213123 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.16042426228523254 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865340419113636 + ], + "opposite_logprob_tokens": [ + -4.75865364074707 + ], + "logprob_reward": -0.00865340419113636, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865340419113636, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26005539298057556, + 0.09204374253749847, + 0.11889824271202087, + 0.06251727789640427, + 0.11659662425518036, + 0.08128642290830612, + 0.10092542320489883, + 0.10851220786571503, + 0.0210783202201128, + 0.038086358457803726 + ], + "credit_map": [ + 0.0175621397793293, + 0.01422948483377695, + 0.01714906096458435, + 0.021731050685048103, + 0.0119608910754323, + 0.09032242000102997, + 0.01336559560149908, + 0.07373477518558502, + 0.010406311601400375, + 0.008767781779170036, + 0.008468663319945335, + 0.009369906969368458, + 0.00942049827426672, + 0.020277122035622597, + 0.009020599536597729, + 0.0163128599524498, + 0.016058998182415962, + 0.013341653160750866, + 0.0067429617047309875, + 0.013540142215788364, + 0.01565907895565033, + 0.01972053386271, + 0.012912190519273281, + 0.02092268131673336, + 0.013601870276033878, + 0.006849145982414484, + 0.005567785818129778, + 0.008121933788061142, + 0.005668897647410631, + 0.007733532227575779, + 0.008298593573272228, + 0.006675524637103081, + 0.022600851953029633, + 0.012323252856731415, + 0.01213020645081997, + 0.013799735344946384, + 0.009493805468082428, + 0.01915721222758293, + 0.012657762505114079, + 0.014433801174163818, + 0.017445959150791168, + 0.00840089377015829, + 0.004774067550897598, + 0.007976588793098927, + 0.0054955086670815945, + 0.017149753868579865, + 0.0051484135910868645, + 0.014895230531692505, + 0.019101018086075783, + 0.0041906945407390594, + 0.007955538108944893, + 0.009163642302155495, + 0.004071501083672047, + 0.029639823362231255, + 0.0034835864789783955, + 0.023319624364376068, + 0.026978712528944016, + 0.009966305457055569, + 0.009366475977003574, + 0.01279783807694912, + 0.005982123780995607, + 0.01926322653889656, + 0.006896443199366331, + 0.01726108230650425, + 0.004887224640697241, + 0.0018411166965961456, + 0.002678181044757366, + 0.002748281229287386, + 0.0014356086030602455, + 0.003287391271442175, + 0.001638215035200119, + 0.0025622998364269733, + 0.012442529201507568, + 0.004234218969941139, + 0.003942539915442467, + 0.004676368087530136, + 0.0023223937023431063, + 0.00451635429635644, + 0.0024729527067393064, + 0.003479000646620989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.00865340419113636 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762259364128113 + ], + "opposite_logprob_tokens": [ + -0.8262259364128113 + ], + "logprob_reward": -0.5762259364128113, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762259364128113, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23100827634334564, + 0.11465851962566376, + 0.07873602211475372, + 0.10160167515277863, + 0.05494260415434837, + 0.11520690470933914, + 0.1410450041294098, + 0.07649245858192444, + 0.022809455171227455, + 0.06349913030862808 + ], + "credit_map": [ + 0.03683698922395706, + 0.02580629102885723, + 0.030553504824638367, + 0.03538799285888672, + 0.018380766734480858, + 0.03405878692865372, + 0.02750043384730816, + 0.022483481094241142, + 0.018763870000839233, + 0.013204729184508324, + 0.013278558850288391, + 0.01988465152680874, + 0.009343558922410011, + 0.017767641693353653, + 0.011646668426692486, + 0.010768829844892025, + 0.01599428988993168, + 0.007398910354822874, + 0.006007473450154066, + 0.01158867683261633, + 0.006020606029778719, + 0.01333844568580389, + 0.00814803782850504, + 0.01023957785218954, + 0.027979359030723572, + 0.009267961606383324, + 0.010076508857309818, + 0.013906092382967472, + 0.006963854655623436, + 0.013737821951508522, + 0.009209812618792057, + 0.010460253804922104, + 0.01169012300670147, + 0.005241946782916784, + 0.005074567161500454, + 0.006842156872153282, + 0.0042604804039001465, + 0.008866483345627785, + 0.006703305058181286, + 0.006263533607125282, + 0.014976647682487965, + 0.0053336708806455135, + 0.005967971868813038, + 0.009407684206962585, + 0.004592066165059805, + 0.03905382379889488, + 0.005005030892789364, + 0.030870001763105392, + 0.026284178718924522, + 0.008725335821509361, + 0.010255096480250359, + 0.015132215805351734, + 0.007576017640531063, + 0.035928595811128616, + 0.008004121482372284, + 0.029139438644051552, + 0.009286587126553059, + 0.007119691930711269, + 0.008844657801091671, + 0.009391076862812042, + 0.004678402561694384, + 0.015827344730496407, + 0.006062245462089777, + 0.01528244186192751, + 0.004369197878986597, + 0.0029973082710057497, + 0.002007384318858385, + 0.003989282064139843, + 0.0019975819159299135, + 0.0026977157685905695, + 0.0025653860066086054, + 0.0021855970844626427, + 0.006753253750503063, + 0.00553932273760438, + 0.003954470623284578, + 0.006371012423187494, + 0.003994351718574762, + 0.01883695460855961, + 0.00511637469753623, + 0.01293338742107153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.5762259364128113 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978530291467905 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.00978530291467905, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978530291467905, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295177936553955, + 0.07366090267896652, + 0.13692013919353485, + 0.16211152076721191, + 0.10038945078849792, + 0.15185807645320892, + 0.041215620934963226, + 0.10735712945461273, + 0.0589824877679348, + 0.03798693045973778 + ], + "credit_map": [ + 0.015180488117039204, + 0.012220523320138454, + 0.014948965981602669, + 0.015367165207862854, + 0.009387326426804066, + 0.026852687820792198, + 0.015569863840937614, + 0.01999078132212162, + 0.008364553563296795, + 0.007788518443703651, + 0.008174311369657516, + 0.007625631056725979, + 0.005363292992115021, + 0.016506386920809746, + 0.005381436087191105, + 0.01445677224546671, + 0.031124651432037354, + 0.012163584120571613, + 0.01463660690933466, + 0.016689853742718697, + 0.006928361486643553, + 0.025156205520033836, + 0.012256589718163013, + 0.017964286729693413, + 0.04255214333534241, + 0.01633036881685257, + 0.013496587984263897, + 0.01842101477086544, + 0.00886142160743475, + 0.027225980535149574, + 0.011757737025618553, + 0.023466268554329872, + 0.018849315121769905, + 0.013967294245958328, + 0.00783390924334526, + 0.011630438268184662, + 0.010133173316717148, + 0.014057372696697712, + 0.013047663494944572, + 0.010870290920138359, + 0.013399912044405937, + 0.007707636803388596, + 0.007314911112189293, + 0.01367438305169344, + 0.006932563614100218, + 0.05335754528641701, + 0.0079598193988204, + 0.041511304676532745, + 0.0056998045183718204, + 0.0044935038313269615, + 0.006058238912373781, + 0.006587184499949217, + 0.003264346159994602, + 0.005471840035170317, + 0.006137968506664038, + 0.003502737497910857, + 0.021241307258605957, + 0.013238250277936459, + 0.008350258693099022, + 0.01114575657993555, + 0.01207305770367384, + 0.01720094308257103, + 0.008961469866335392, + 0.01514608133584261, + 0.0037716457154601812, + 0.004785760771483183, + 0.004195492248982191, + 0.00586566561833024, + 0.00373642984777689, + 0.01740117371082306, + 0.0051381695084273815, + 0.014088149182498455, + 0.010638111270964146, + 0.00378889637067914, + 0.0038243357557803392, + 0.0048956070095300674, + 0.0029023075476288795, + 0.00520779425278306, + 0.003031217260286212, + 0.003698658663779497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 2 + ], + "selected_reward": -0.00978530291467905 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot from an extremely low angle, draws the viewer's gaze squarely to the weathered face of an elderly woman. A soft, golden beam of afternoon sunlight slants in from the side of the frame, meticulously sculpting every wrinkle in her features. Deep crow's feet and nasolabial folds resemble a dry riverbed, and faint age spots dot her skin. A few strands of her thinning silver-white hair, ruffled by the breeze, lay plastered to her temple. Her cloudy, yet luminous eyes gaze upward, their rims reddened, her lips pursed, her chin trembling imperceptibly with suppressed emotion. In this silence, a full, glistening tear slowly escapes from the corner of her right eye. It lingers briefly for half a second before steadily and leisurely sliding down a deep wrinkle on her cheek, like a river flowing in a predetermined path. The tear reflects the warm sunlight as it glides, leaving a wet, glistening trail before disappearing from the camera's field of view. The entire picture is dominated by warm gold and deep shadows under the strong contrast of light and shadow, creating a quiet, deep and melancholy atmosphere full of the sense of time. The background is completely blurred into a blurred warm light spot.", + "seed": 16069, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_554_seed_16069_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.12281903512775898, + "vlm_logprob": -0.12281903512775898, + "vlm_yes_no_margin": 3.375000023841858, + "vlm_token_credit": -0.12281903512775898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.013006022199988365 + ], + "opposite_logprob_tokens": [ + -5.763006210327148 + ], + "logprob_reward": -0.013006022199988365, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.013006022199988365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15865878760814667, + 0.08295802026987076, + 0.054761141538619995, + 0.05208129808306694, + 0.06953227519989014, + 0.05877477303147316, + 0.08560486137866974, + 0.09784265607595444, + 0.1189463660120964, + 0.22083982825279236 + ], + "credit_map": [ + 0.02807006798684597, + 0.026739509776234627, + 0.020006073638796806, + 0.02983666956424713, + 0.014551130123436451, + 0.013969415798783302, + 0.015636665746569633, + 0.009849253110587597, + 0.021270183846354485, + 0.010551082901656628, + 0.012761980295181274, + 0.011476341634988785, + 0.008149594068527222, + 0.005613307468593121, + 0.007951914332807064, + 0.005183614324778318, + 0.008867387659847736, + 0.014410960488021374, + 0.006413342896848917, + 0.007947113364934921, + 0.004799052141606808, + 0.0030220469925552607, + 0.005947158671915531, + 0.00335408141836524, + 0.01350180059671402, + 0.0063477628864347935, + 0.008605336770415306, + 0.009670470841228962, + 0.0036550138611346483, + 0.0023891516029834747, + 0.004913580138236284, + 0.0029981806874275208, + 0.01891395077109337, + 0.008447026833891869, + 0.0098104327917099, + 0.009955068118870258, + 0.005356454756110907, + 0.004941074177622795, + 0.008006254211068153, + 0.004102017264813185, + 0.01397054921835661, + 0.009447915479540825, + 0.009684902615845203, + 0.008607360534369946, + 0.004016180522739887, + 0.004610874690115452, + 0.0053077032789587975, + 0.0031292845960706472, + 0.02206076867878437, + 0.010755267925560474, + 0.012400425970554352, + 0.00955742783844471, + 0.009104707278311253, + 0.005895824637264013, + 0.010352274402976036, + 0.005478166043758392, + 0.016061654314398766, + 0.008364680223166943, + 0.02788829430937767, + 0.024836676195263863, + 0.004673061892390251, + 0.004622872918844223, + 0.007344089914113283, + 0.0040513318963348866, + 0.00942311529070139, + 0.007534591946750879, + 0.040735483169555664, + 0.04412113502621651, + 0.004172193817794323, + 0.003120886394754052, + 0.006035848055034876, + 0.0038031095173209906, + 0.016875185072422028, + 0.012478874064981937, + 0.08263659477233887, + 0.08664068579673767, + 0.004335720092058182, + 0.003853969043120742, + 0.00869965460151434, + 0.005319146439433098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.013006022199988365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.142875075340271 + ], + "opposite_logprob_tokens": [ + -2.0178751945495605 + ], + "logprob_reward": -0.142875075340271, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.142875075340271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2550067603588104, + 0.05181495100259781, + 0.052077293395996094, + 0.05965372547507286, + 0.041124578565359116, + 0.06618917733430862, + 0.05283404141664505, + 0.10813993960618973, + 0.09691917151212692, + 0.21624040603637695 + ], + "credit_map": [ + 0.08628261834383011, + 0.030149590224027634, + 0.04095916450023651, + 0.035683225840330124, + 0.016500040888786316, + 0.013774791732430458, + 0.017873700708150864, + 0.01378361601382494, + 0.009326686151325703, + 0.007541544735431671, + 0.00788355153053999, + 0.008555627427995205, + 0.004867060109972954, + 0.0028246366418898106, + 0.006730927620083094, + 0.004084908403456211, + 0.009906451217830181, + 0.010332653298974037, + 0.00594492070376873, + 0.006066285073757172, + 0.005341949872672558, + 0.004113026894629002, + 0.007318708114326, + 0.003053290769457817, + 0.012296155095100403, + 0.008854960091412067, + 0.010879691690206528, + 0.00827892404049635, + 0.005082298535853624, + 0.004331540782004595, + 0.006739232689142227, + 0.0031909223180264235, + 0.01197985839098692, + 0.006623501889407635, + 0.006410585716366768, + 0.004537458531558514, + 0.003428362077102065, + 0.0017862439854070544, + 0.004424504470080137, + 0.0019340615253895521, + 0.0142593402415514, + 0.009690387174487114, + 0.007822460494935513, + 0.009639706462621689, + 0.005589482374489307, + 0.00610815966501832, + 0.007805543020367622, + 0.005274095572531223, + 0.016420893371105194, + 0.006881930399686098, + 0.008343740366399288, + 0.005494569428265095, + 0.003915510606020689, + 0.0038793161511421204, + 0.005008659325540066, + 0.002889416180551052, + 0.012481841258704662, + 0.013316545635461807, + 0.022859087213873863, + 0.023656217381358147, + 0.009761717170476913, + 0.006283544935286045, + 0.013152796775102615, + 0.006628178991377354, + 0.014189288951456547, + 0.007790058385580778, + 0.02541985921561718, + 0.026032665744423866, + 0.006102562882006168, + 0.00437539117410779, + 0.00813099555671215, + 0.00487833796069026, + 0.014679020270705223, + 0.011499408632516861, + 0.08243558555841446, + 0.08546958118677139, + 0.0037060966715216637, + 0.003714493243023753, + 0.008992918767035007, + 0.005743283312767744 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.142875075340271 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020603327080607414 + ], + "opposite_logprob_tokens": [ + -3.8956034183502197 + ], + "logprob_reward": -0.020603327080607414, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020603327080607414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20482082664966583, + 0.06805998086929321, + 0.09071174263954163, + 0.0693529024720192, + 0.0902831107378006, + 0.07865594327449799, + 0.036654215306043625, + 0.1158994510769844, + 0.054614122956991196, + 0.19094765186309814 + ], + "credit_map": [ + 0.028791021555662155, + 0.03056730329990387, + 0.035299137234687805, + 0.04301619157195091, + 0.019969264045357704, + 0.015287062153220177, + 0.017644798383116722, + 0.014246061444282532, + 0.014199241995811462, + 0.009372825734317303, + 0.008828066289424896, + 0.007863947190344334, + 0.006928715389221907, + 0.007596466690301895, + 0.008102827705442905, + 0.005167888943105936, + 0.0146826496347785, + 0.01654990203678608, + 0.01101690623909235, + 0.011328545399010181, + 0.01177055574953556, + 0.00577556574717164, + 0.013534748926758766, + 0.006052866578102112, + 0.013015457428991795, + 0.012524772435426712, + 0.01286251749843359, + 0.010394980199635029, + 0.005633232183754444, + 0.0035967847798019648, + 0.007589965593069792, + 0.0037351902574300766, + 0.018897168338298798, + 0.0128758754581213, + 0.01305660791695118, + 0.01479114405810833, + 0.007570179179310799, + 0.007550508715212345, + 0.009703592397272587, + 0.0058380309492349625, + 0.018600136041641235, + 0.010676565580070019, + 0.014407617039978504, + 0.010893778875470161, + 0.007209484465420246, + 0.005367792211472988, + 0.0065005142241716385, + 0.005000054836273193, + 0.006307083647698164, + 0.005106073804199696, + 0.007232611533254385, + 0.005294617265462875, + 0.002476948080584407, + 0.0031632876489311457, + 0.0037238537333905697, + 0.0033497402910143137, + 0.02470279298722744, + 0.007794046308845282, + 0.03524995222687721, + 0.026551077142357826, + 0.005371464416384697, + 0.005293369293212891, + 0.005846826825290918, + 0.005089926067739725, + 0.009642078541219234, + 0.006143689621239901, + 0.010595915839076042, + 0.012991945259273052, + 0.0037922863848507404, + 0.0030212802812457085, + 0.004721018020063639, + 0.0037059090100228786, + 0.016708901152014732, + 0.00903286598622799, + 0.0730639323592186, + 0.07285186648368835, + 0.003382601775228977, + 0.003682464826852083, + 0.00711468979716301, + 0.005110329482704401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.020603327080607414 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289681017398834 + ], + "opposite_logprob_tokens": [ + -1.053968071937561 + ], + "logprob_reward": -0.4289681017398834, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289681017398834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.288968026638031, + 0.10650142282247543, + 0.1257144957780838, + 0.10505399107933044, + 0.11474735289812088, + 0.04425612837076187, + 0.029461298137903214, + 0.07725568115711212, + 0.05436081066727638, + 0.053680822253227234 + ], + "credit_map": [ + 0.031152693554759026, + 0.021654069423675537, + 0.08255378156900406, + 0.09496178478002548, + 0.016602879390120506, + 0.01077018491923809, + 0.019384436309337616, + 0.011888205073773861, + 0.027040433138608932, + 0.014290675520896912, + 0.01483679935336113, + 0.017689518630504608, + 0.00849626399576664, + 0.007509555201977491, + 0.009105940349400043, + 0.007532231975346804, + 0.012053067795932293, + 0.02692343108355999, + 0.012790980748832226, + 0.01724114455282688, + 0.01556460838764906, + 0.008006392978131771, + 0.021482368931174278, + 0.011652510613203049, + 0.013595459051430225, + 0.017690811306238174, + 0.02248590998351574, + 0.01915789395570755, + 0.008005445823073387, + 0.006328616291284561, + 0.011254887096583843, + 0.006534966640174389, + 0.019552310928702354, + 0.0187800582498312, + 0.013915792107582092, + 0.01667809672653675, + 0.010820060037076473, + 0.01138953398913145, + 0.015017515979707241, + 0.00859397929161787, + 0.007394088432192802, + 0.007297234609723091, + 0.00548404548317194, + 0.00647173123434186, + 0.004880597814917564, + 0.003287362866103649, + 0.005579301156103611, + 0.003861767239868641, + 0.007357303984463215, + 0.0037998005282133818, + 0.0038261879235506058, + 0.003967931494116783, + 0.0027332629542797804, + 0.0021046046167612076, + 0.003905707970261574, + 0.001766497502103448, + 0.013275266624987125, + 0.006743502803146839, + 0.02179531194269657, + 0.019369816407561302, + 0.0042811050079762936, + 0.004008198156952858, + 0.004727480001747608, + 0.003055004868656397, + 0.011895135045051575, + 0.006736859679222107, + 0.007472801487892866, + 0.0073829093016684055, + 0.005191451404243708, + 0.004063460510224104, + 0.007557989098131657, + 0.004060203209519386, + 0.0036606749054044485, + 0.0031563113443553448, + 0.019405057653784752, + 0.02055381052196026, + 0.0013121289666742086, + 0.0010726782493293285, + 0.0028234210330992937, + 0.0016967441188171506 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.4289681017398834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0086426492780447 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.0086426492780447, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.0086426492780447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2901886999607086, + 0.09238890558481216, + 0.0469827838242054, + 0.04553854838013649, + 0.05257129669189453, + 0.05576738715171814, + 0.0654117539525032, + 0.10297657549381256, + 0.0835379809141159, + 0.1646360158920288 + ], + "credit_map": [ + 0.03910787031054497, + 0.029268868267536163, + 0.07795296609401703, + 0.082072414457798, + 0.017109135165810585, + 0.013712962158024311, + 0.01825401745736599, + 0.012710470706224442, + 0.026955094188451767, + 0.011145148426294327, + 0.01249224878847599, + 0.011505800299346447, + 0.007172999903559685, + 0.0064319646917283535, + 0.009135711006820202, + 0.007549935486167669, + 0.01307798083871603, + 0.008211102336645126, + 0.007279993500560522, + 0.007230621296912432, + 0.0035573348868638277, + 0.0021597971208393574, + 0.003273514099419117, + 0.0021924369502812624, + 0.01290606427937746, + 0.007764236070215702, + 0.007376980036497116, + 0.005608039442449808, + 0.003162299282848835, + 0.0030249059200286865, + 0.0033674698788672686, + 0.002328556263819337, + 0.008413196541368961, + 0.007182312663644552, + 0.011351500637829304, + 0.008034905418753624, + 0.004222443792968988, + 0.0044143470004200935, + 0.005416749976575375, + 0.003535839729011059, + 0.01123800314962864, + 0.00809108093380928, + 0.0061402758583426476, + 0.009621835313737392, + 0.007414942141622305, + 0.0036645301152020693, + 0.007219457998871803, + 0.0023772623389959335, + 0.01868002861738205, + 0.008688018657267094, + 0.008506838232278824, + 0.008107098750770092, + 0.006075861863791943, + 0.003929069731384516, + 0.0074089569970965385, + 0.004015884827822447, + 0.013162401504814625, + 0.008563760668039322, + 0.0313645638525486, + 0.027854878455400467, + 0.004776907153427601, + 0.004321650601923466, + 0.007716658990830183, + 0.005215756595134735, + 0.0093239676207304, + 0.007941707037389278, + 0.02358812652528286, + 0.024833835661411285, + 0.004276707302778959, + 0.0031113503500819206, + 0.006430812645703554, + 0.004031474236398935, + 0.011462806724011898, + 0.008759137243032455, + 0.06247175484895706, + 0.06610482931137085, + 0.0028045126236975193, + 0.002581824781373143, + 0.00646543363109231, + 0.003985727205872536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.0086426492780447 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot from an extremely low angle, draws the viewer's gaze squarely to the weathered face of an elderly woman. A soft, golden beam of afternoon sunlight slants in from the side of the frame, meticulously sculpting every wrinkle in her features. Deep crow's feet and nasolabial folds resemble a dry riverbed, and faint age spots dot her skin. A few strands of her thinning silver-white hair, ruffled by the breeze, lay plastered to her temple. Her cloudy, yet luminous eyes gaze upward, their rims reddened, her lips pursed, her chin trembling imperceptibly with suppressed emotion. In this silence, a full, glistening tear slowly escapes from the corner of her right eye. It lingers briefly for half a second before steadily and leisurely sliding down a deep wrinkle on her cheek, like a river flowing in a predetermined path. The tear reflects the warm sunlight as it glides, leaving a wet, glistening trail before disappearing from the camera's field of view. The entire picture is dominated by warm gold and deep shadows under the strong contrast of light and shadow, creating a quiet, deep and melancholy atmosphere full of the sense of time. The background is completely blurred into a blurred warm light spot.", + "seed": 16070, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_554_seed_16070_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07741630119271577, + "vlm_logprob": -0.07741630119271577, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.07741630119271577, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02590162865817547 + ], + "opposite_logprob_tokens": [ + -4.150901794433594 + ], + "logprob_reward": -0.02590162865817547, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02590162865817547, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07924647629261017, + 0.03372867777943611, + 0.16944418847560883, + 0.026980074122548103, + 0.008738919161260128, + 0.611797034740448, + 0.019482256844639778, + 0.014182528480887413, + 0.023435764014720917, + 0.012964156456291676 + ], + "credit_map": [ + 0.006261168979108334, + 0.0022375464905053377, + 0.004034338053315878, + 0.002246757736429572, + 0.006761922035366297, + 0.01169289369136095, + 0.03227221220731735, + 0.013739638961851597, + 0.0022721870336681604, + 0.001526699517853558, + 0.003206889610737562, + 0.001614708686247468, + 0.0037301196716725826, + 0.004301983863115311, + 0.012330650351941586, + 0.0047454372979700565, + 0.01342343632131815, + 0.00429700268432498, + 0.013072469271719456, + 0.004919537343084812, + 0.017873235046863556, + 0.018433358520269394, + 0.06900780647993088, + 0.02841733582317829, + 0.0035050332080572844, + 0.0012110702227801085, + 0.002023873617872596, + 0.000841572240460664, + 0.0021533933468163013, + 0.0038231562357395887, + 0.004073628690093756, + 0.009348345920443535, + 0.0006709094741381705, + 0.0004928355920128524, + 0.0006927915383130312, + 0.0006705019623041153, + 0.0012512803077697754, + 0.0011886563152074814, + 0.0018537357682362199, + 0.0019182084361091256, + 0.09239864349365234, + 0.01117353979498148, + 0.22311857342720032, + 0.019296210259199142, + 0.03245186060667038, + 0.03249998763203621, + 0.16735433042049408, + 0.03350384905934334, + 0.0023156413808465004, + 0.0011994277592748404, + 0.0030917066615074873, + 0.0014765687519684434, + 0.0020771727431565523, + 0.00283026322722435, + 0.0036654481664299965, + 0.0028260278049856424, + 0.0012556472793221474, + 0.0010486348764970899, + 0.0010221124393865466, + 0.0012118015438318253, + 0.002048828173428774, + 0.0020629141945391893, + 0.0028321840800344944, + 0.0027004058938473463, + 0.0019507668912410736, + 0.001754201133735478, + 0.002810370409861207, + 0.0019861168693751097, + 0.0028981559444218874, + 0.0033599059097468853, + 0.0035753825213760138, + 0.0051008653827011585, + 0.0009068017243407667, + 0.0011980477720499039, + 0.001390016172081232, + 0.001040751812979579, + 0.0017376623582094908, + 0.0023414529860019684, + 0.0019358852878212929, + 0.0024135380517691374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.02590162865817547 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22562859952449799 + ], + "opposite_logprob_tokens": [ + -1.6006286144256592 + ], + "logprob_reward": -0.22562859952449799, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22562859952449799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.47926098108291626, + 0.022044271230697632, + 0.14868199825286865, + 0.05546757951378822, + 0.030459415167570114, + 0.1824444979429245, + 0.01350944209843874, + 0.03187183663249016, + 0.018360069021582603, + 0.0178998913615942 + ], + "credit_map": [ + 0.0604204423725605, + 0.013061740435659885, + 0.019764676690101624, + 0.011878805235028267, + 0.04467612877488136, + 0.06611625850200653, + 0.1914442628622055, + 0.07189865410327911, + 0.001811168622225523, + 0.0019478866597637534, + 0.002664936939254403, + 0.002088072244077921, + 0.0038971714675426483, + 0.004265946801751852, + 0.003092428669333458, + 0.002276659943163395, + 0.010761793702840805, + 0.0038685197941958904, + 0.012944099493324757, + 0.004377706907689571, + 0.01579899713397026, + 0.01665056124329567, + 0.059707220643758774, + 0.024573111906647682, + 0.008487408049404621, + 0.0021496342960745096, + 0.007102522067725658, + 0.0014594384701922536, + 0.003848203457891941, + 0.007585823070257902, + 0.00831368938088417, + 0.01652085781097412, + 0.0024812917690724134, + 0.0017261753091588616, + 0.0042248317040503025, + 0.0017814619932323694, + 0.0040953150019049644, + 0.003752951743081212, + 0.005534212104976177, + 0.00686317402869463, + 0.030032603070139885, + 0.003777948673814535, + 0.0672537162899971, + 0.00619502505287528, + 0.010188831947743893, + 0.009616567753255367, + 0.044949036091566086, + 0.010430771857500076, + 0.0008577496046200395, + 0.0005221896572038531, + 0.0009091873653233051, + 0.0005924172583036125, + 0.0019338647834956646, + 0.003949517384171486, + 0.0022338740527629852, + 0.0025106424000114202, + 0.003482551546767354, + 0.0012926225317642093, + 0.0026144932489842176, + 0.0017796348547562957, + 0.0030173175036907196, + 0.005246773362159729, + 0.0045120976865291595, + 0.009926345199346542, + 0.0015218191547319293, + 0.0009185480303131044, + 0.0017643616301938891, + 0.0009288859437219799, + 0.001578273600898683, + 0.0032621321734040976, + 0.0030375595670193434, + 0.005348487291485071, + 0.0009710724116303027, + 0.0018550000386312604, + 0.0012125802459195256, + 0.002667526016011834, + 0.003215054515749216, + 0.0023730318062007427, + 0.003180513624101877, + 0.002425112994387746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.22562859952449799 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032051641028374434 + ], + "opposite_logprob_tokens": [ + -5.753205299377441 + ], + "logprob_reward": -0.0032051641028374434, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032051641028374434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2240062952041626, + 0.11078524589538574, + 0.07390061765909195, + 0.17227742075920105, + 0.10763412714004517, + 0.04879164323210716, + 0.16349470615386963, + 0.03672104328870773, + 0.03959861025214195, + 0.022790303453803062 + ], + "credit_map": [ + 0.02223532274365425, + 0.005779901519417763, + 0.00976505596190691, + 0.006378571502864361, + 0.020326388999819756, + 0.030523711815476418, + 0.09096561372280121, + 0.03803173825144768, + 0.006961748003959656, + 0.004394353833049536, + 0.011757735162973404, + 0.004537216387689114, + 0.012058008462190628, + 0.012859292328357697, + 0.04182837903499603, + 0.016388509422540665, + 0.00500105693936348, + 0.003659370820969343, + 0.008938516490161419, + 0.0036672267597168684, + 0.009272413328289986, + 0.009054861031472683, + 0.024778170511126518, + 0.009529003873467445, + 0.020063480362296104, + 0.006483701057732105, + 0.020229248329997063, + 0.007852836512029171, + 0.014192258939146996, + 0.025856632739305496, + 0.020318463444709778, + 0.057280801236629486, + 0.013102399185299873, + 0.007705552037805319, + 0.01893410086631775, + 0.008671891875565052, + 0.013250348158180714, + 0.015756212174892426, + 0.015627874061465263, + 0.014585744589567184, + 0.004280486609786749, + 0.0015782348345965147, + 0.008022981695830822, + 0.0026857545599341393, + 0.0044739446602761745, + 0.0069494061172008514, + 0.009181269444525242, + 0.011619566939771175, + 0.007876354269683361, + 0.003106188727542758, + 0.00904210563749075, + 0.003990560304373503, + 0.01200435683131218, + 0.015249636955559254, + 0.06241277605295181, + 0.049812719225883484, + 0.003725050948560238, + 0.003007692750543356, + 0.004992901347577572, + 0.003156043589115143, + 0.004422531928867102, + 0.005926745943725109, + 0.005070825107395649, + 0.006419254932552576, + 0.0046104793436825275, + 0.0029367057140916586, + 0.005563083570450544, + 0.003109017387032509, + 0.004600316286087036, + 0.004861124325543642, + 0.005888252053409815, + 0.008029629476368427, + 0.002239454537630081, + 0.0013038592878729105, + 0.0015278368955478072, + 0.0014409360010176897, + 0.003241152036935091, + 0.003947275225073099, + 0.003636701498180628, + 0.005453087855130434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.0032051641028374434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270899474620819 + ], + "opposite_logprob_tokens": [ + -2.1270899772644043 + ], + "logprob_reward": -0.1270899474620819, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270899474620819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34499606490135193, + 0.12428317219018936, + 0.06582939624786377, + 0.07063446938991547, + 0.17914675176143646, + 0.057550910860300064, + 0.05753997340798378, + 0.06755769997835159, + 0.016193812713027, + 0.016267774626612663 + ], + "credit_map": [ + 0.04016685113310814, + 0.008938848040997982, + 0.014092504046857357, + 0.008765860460698605, + 0.03143096715211868, + 0.0521077997982502, + 0.1385214626789093, + 0.05097179114818573, + 0.008667084388434887, + 0.012919369153678417, + 0.009617863222956657, + 0.020149169489741325, + 0.02540687844157219, + 0.015601681545376778, + 0.018278220668435097, + 0.013642898760735989, + 0.003447421593591571, + 0.0019458052702248096, + 0.00742591405287385, + 0.002594345947727561, + 0.006489086430519819, + 0.006469005718827248, + 0.026064977049827576, + 0.011392836458981037, + 0.009065620601177216, + 0.0034534293226897717, + 0.004246517084538937, + 0.00325249251909554, + 0.008634834550321102, + 0.010367387905716896, + 0.009778667241334915, + 0.02183551900088787, + 0.01620304584503174, + 0.01260353997349739, + 0.025361258536577225, + 0.021153565496206284, + 0.02492995373904705, + 0.025047700852155685, + 0.031153930351138115, + 0.022693762555718422, + 0.01062520407140255, + 0.0017863019602373242, + 0.02153996378183365, + 0.0025611864402890205, + 0.003926110453903675, + 0.0033686591777950525, + 0.0066148992627859116, + 0.007128587923943996, + 0.003065217286348343, + 0.0014645438641309738, + 0.002512704348191619, + 0.001821103855036199, + 0.004582882393151522, + 0.005444465670734644, + 0.020915936678647995, + 0.017733117565512657, + 0.007126411888748407, + 0.00449153920635581, + 0.011619951575994492, + 0.007553535047918558, + 0.007729083765298128, + 0.010075048543512821, + 0.00944834016263485, + 0.009513793513178825, + 0.001157754217274487, + 0.001481935614719987, + 0.0013645295985043049, + 0.001804026192985475, + 0.002467429731041193, + 0.00200583110563457, + 0.0027833867352455854, + 0.0031289183534681797, + 0.0009170947014354169, + 0.0011604019673541188, + 0.0011540462728589773, + 0.001719710067845881, + 0.002422392601147294, + 0.0028747827745974064, + 0.0032814873848110437, + 0.002737858798354864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.1270899474620819 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005256166215986013 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.005256166215986013, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005256166215986013, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18136458098888397, + 0.061687249690294266, + 0.0792425200343132, + 0.15228156745433807, + 0.0486215241253376, + 0.15530557930469513, + 0.13518300652503967, + 0.04384204372763634, + 0.04862336069345474, + 0.09384854882955551 + ], + "credit_map": [ + 0.01458145771175623, + 0.005082322750240564, + 0.008672446012496948, + 0.0049363235011696815, + 0.016068316996097565, + 0.02523648738861084, + 0.07510929554700851, + 0.031677935272455215, + 0.005226295907050371, + 0.0038037130143493414, + 0.004789537284523249, + 0.004688188433647156, + 0.007390730548650026, + 0.007524826563894749, + 0.021492864936590195, + 0.006771090906113386, + 0.004730698652565479, + 0.0046331319026649, + 0.007768018636852503, + 0.004606739617884159, + 0.010845641605556011, + 0.010455647483468056, + 0.02437455579638481, + 0.011828083544969559, + 0.024227969348430634, + 0.007118095178157091, + 0.027917394414544106, + 0.007118024863302708, + 0.010219994001090527, + 0.019316747784614563, + 0.021655358374118805, + 0.034707993268966675, + 0.0035359219182282686, + 0.0036387864965945482, + 0.006393320392817259, + 0.004210669081658125, + 0.006863253656774759, + 0.00982146430760622, + 0.00787960086017847, + 0.006278510205447674, + 0.0143484091386199, + 0.003495947690680623, + 0.03161707893013954, + 0.004845497664064169, + 0.010476702824234962, + 0.01802176423370838, + 0.050485581159591675, + 0.022014595568180084, + 0.005463385488837957, + 0.002333100186660886, + 0.007092183921486139, + 0.003276500850915909, + 0.01114024966955185, + 0.013449889607727528, + 0.05509113520383835, + 0.0373365692794323, + 0.003026992315426469, + 0.003999945241957903, + 0.003326398553326726, + 0.004669107962399721, + 0.007899591699242592, + 0.006289039272814989, + 0.00793304294347763, + 0.006697928532958031, + 0.0030781649984419346, + 0.0034802237059921026, + 0.0043017384596168995, + 0.0055163269862532616, + 0.008820637129247189, + 0.007932125590741634, + 0.009099348448216915, + 0.0063947951421141624, + 0.007379659917205572, + 0.0035717010032385588, + 0.004782500211149454, + 0.0037382368464022875, + 0.008712606504559517, + 0.026002468541264534, + 0.016099611297249794, + 0.023561770096421242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.005256166215986013 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "close-up, static shot from an extremely low angle, draws the viewer's gaze squarely to the weathered face of an elderly woman. A soft, golden beam of afternoon sunlight slants in from the side of the frame, meticulously sculpting every wrinkle in her features. Deep crow's feet and nasolabial folds resemble a dry riverbed, and faint age spots dot her skin. A few strands of her thinning silver-white hair, ruffled by the breeze, lay plastered to her temple. Her cloudy, yet luminous eyes gaze upward, their rims reddened, her lips pursed, her chin trembling imperceptibly with suppressed emotion. In this silence, a full, glistening tear slowly escapes from the corner of her right eye. It lingers briefly for half a second before steadily and leisurely sliding down a deep wrinkle on her cheek, like a river flowing in a predetermined path. The tear reflects the warm sunlight as it glides, leaving a wet, glistening trail before disappearing from the camera's field of view. The entire picture is dominated by warm gold and deep shadows under the strong contrast of light and shadow, creating a quiet, deep and melancholy atmosphere full of the sense of time. The background is completely blurred into a blurred warm light spot.", + "seed": 16071, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000070/group_9/index_554_seed_16071_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1410024369135499, + "vlm_logprob": -0.1410024369135499, + "vlm_yes_no_margin": 3.399999976158142, + "vlm_token_credit": -0.1410024369135499, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018052909523248672 + ], + "opposite_logprob_tokens": [ + -6.26805305480957 + ], + "logprob_reward": -0.018052909523248672, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.018052909523248672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22292718291282654, + 0.08482727408409119, + 0.08434005081653595, + 0.2174384444952011, + 0.08862321078777313, + 0.06245024874806404, + 0.05530652776360512, + 0.0570494644343853, + 0.07886042445898056, + 0.048177193850278854 + ], + "credit_map": [ + 0.024477703496813774, + 0.02220812626183033, + 0.039016302675008774, + 0.021259034052491188, + 0.028782907873392105, + 0.02772071585059166, + 0.028571277856826782, + 0.030891109257936478, + 0.00931042805314064, + 0.008249841630458832, + 0.010085702873766422, + 0.007782422471791506, + 0.009883022867143154, + 0.011055405251681805, + 0.008561885915696621, + 0.019898558035492897, + 0.009636437520384789, + 0.007372928783297539, + 0.008097708225250244, + 0.00821791309863329, + 0.008326158858835697, + 0.012280422262847424, + 0.010039830580353737, + 0.02036864124238491, + 0.025731738656759262, + 0.017443208023905754, + 0.048587437719106674, + 0.021190792322158813, + 0.021739041432738304, + 0.03165529668331146, + 0.03021611087024212, + 0.02087482437491417, + 0.010868558660149574, + 0.009889662265777588, + 0.019233813509345055, + 0.007692395243793726, + 0.009904852136969566, + 0.01124438177794218, + 0.011504636146128178, + 0.008284910582005978, + 0.008259744383394718, + 0.006990457884967327, + 0.006677708122879267, + 0.005616430193185806, + 0.009714662097394466, + 0.008890974335372448, + 0.007995153777301311, + 0.008305121213197708, + 0.0071234614588320255, + 0.005796574521809816, + 0.009704070165753365, + 0.006906843278557062, + 0.005550432950258255, + 0.007038788869976997, + 0.007347647566348314, + 0.005838704761117697, + 0.008185936138033867, + 0.007656875066459179, + 0.006730343215167522, + 0.005812385585159063, + 0.007391962222754955, + 0.007685461081564426, + 0.007351874373853207, + 0.006234625820070505, + 0.008606175892055035, + 0.007952087558805943, + 0.015159840695559978, + 0.007127255201339722, + 0.008641289547085762, + 0.01139706652611494, + 0.009606670588254929, + 0.010370034724473953, + 0.007998599670827389, + 0.005122253205627203, + 0.006944986991584301, + 0.005550142843276262, + 0.005499913822859526, + 0.005676885135471821, + 0.00484788604080677, + 0.006536528468132019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.018052909523248672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351906061172485 + ], + "opposite_logprob_tokens": [ + -1.31351900100708 + ], + "logprob_reward": -0.31351906061172485, + "yes_no_margin_reward": 0.9999999403953552, + "token_credit_reward": -0.31351906061172485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12593460083007812, + 0.23677538335323334, + 0.09141476452350616, + 0.24895182251930237, + 0.07966209203004837, + 0.056640978902578354, + 0.045873358845710754, + 0.04494587704539299, + 0.05043323338031769, + 0.019367925822734833 + ], + "credit_map": [ + 0.011209794320166111, + 0.015684625133872032, + 0.018772197887301445, + 0.016358353197574615, + 0.016532888635993004, + 0.018664151430130005, + 0.016776299104094505, + 0.011936268769204617, + 0.029108168557286263, + 0.017092334106564522, + 0.03393945097923279, + 0.014733590185642242, + 0.022156748920679092, + 0.031807832419872284, + 0.024057848379015923, + 0.06387940794229507, + 0.008164534345269203, + 0.011158555746078491, + 0.008450738154351711, + 0.008024985902011395, + 0.011848147958517075, + 0.012313824146986008, + 0.014490654692053795, + 0.01696331985294819, + 0.026691725477576256, + 0.03061714582145214, + 0.02488706260919571, + 0.02390531450510025, + 0.033435337245464325, + 0.033370498567819595, + 0.041894521564245224, + 0.03415020555257797, + 0.007995957508683205, + 0.00846631359308958, + 0.013716085813939571, + 0.00884744618088007, + 0.010599801316857338, + 0.010189578868448734, + 0.01177363470196724, + 0.00807326938956976, + 0.005651560612022877, + 0.006463563535362482, + 0.011593451723456383, + 0.0051346090622246265, + 0.008475558832287788, + 0.0072102793492376804, + 0.0071744415909051895, + 0.00493751373142004, + 0.005622928496450186, + 0.004254074767231941, + 0.012355577200651169, + 0.003867085324600339, + 0.004729380831122398, + 0.0057705058716237545, + 0.004374300595372915, + 0.004899503663182259, + 0.005005258601158857, + 0.004414015915244818, + 0.006323568522930145, + 0.0053275092504918575, + 0.005758604500442743, + 0.006007189862430096, + 0.007458436768501997, + 0.004651292692869902, + 0.0038729712832719088, + 0.004105365835130215, + 0.005752078723162413, + 0.007177396211773157, + 0.005205713678151369, + 0.007746715098619461, + 0.010347696952521801, + 0.006225291173905134, + 0.0021057352423667908, + 0.0021219050977379084, + 0.002747577615082264, + 0.002014141995459795, + 0.002789846621453762, + 0.0025849223602563143, + 0.0028945147059857845, + 0.002109281253069639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.31351906061172485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865092221647501 + ], + "opposite_logprob_tokens": [ + -4.758650779724121 + ], + "logprob_reward": -0.00865092221647501, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865092221647501, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21870110929012299, + 0.08049493283033371, + 0.31605884432792664, + 0.06957142800092697, + 0.07979744672775269, + 0.09929218888282776, + 0.03423833101987839, + 0.030110960826277733, + 0.03134047985076904, + 0.040394339710474014 + ], + "credit_map": [ + 0.01789212040603161, + 0.027480388060212135, + 0.02374962903559208, + 0.02687358297407627, + 0.02293907105922699, + 0.03308885917067528, + 0.027420317754149437, + 0.039257150143384933, + 0.007138031534850597, + 0.008943343535065651, + 0.006462571211159229, + 0.006933676078915596, + 0.011541325598955154, + 0.0110721904784441, + 0.00898872409015894, + 0.01941506750881672, + 0.0264535304158926, + 0.03337141498923302, + 0.04812777414917946, + 0.02891349606215954, + 0.03358951956033707, + 0.04216981679201126, + 0.05067884549498558, + 0.05275445058941841, + 0.0070173852145671844, + 0.008217213675379753, + 0.010753294453024864, + 0.008405391126871109, + 0.007483778987079859, + 0.010397759266197681, + 0.009079298004508018, + 0.008217304944992065, + 0.008347639814019203, + 0.010484136641025543, + 0.01641378179192543, + 0.007173985708504915, + 0.009267956018447876, + 0.009353654459118843, + 0.009364819154143333, + 0.009391466155648232, + 0.012743346393108368, + 0.008878265507519245, + 0.025505278259515762, + 0.008305724710226059, + 0.011946478858590126, + 0.012781490571796894, + 0.009715363383293152, + 0.009416243992745876, + 0.004194095730781555, + 0.004493545275181532, + 0.005803173873573542, + 0.0034834023099392653, + 0.004562500398606062, + 0.004232239909470081, + 0.004020450171083212, + 0.003448923584073782, + 0.0025986928958445787, + 0.003414321457967162, + 0.003301250282675028, + 0.0043476237915456295, + 0.004053672309964895, + 0.004006864503026009, + 0.005485927686095238, + 0.0029026081319898367, + 0.003411229932680726, + 0.004340737592428923, + 0.005705078598111868, + 0.0036729974672198296, + 0.0027769545558840036, + 0.003608413739129901, + 0.004397425800561905, + 0.0034276407677680254, + 0.003489559981971979, + 0.005285425111651421, + 0.003097331849858165, + 0.004710123408585787, + 0.005175760015845299, + 0.005372599232941866, + 0.005146712996065617, + 0.008116827346384525 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.00865092221647501 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.34871822595596313 + ], + "opposite_logprob_tokens": [ + -1.2237181663513184 + ], + "logprob_reward": -0.34871822595596313, + "yes_no_margin_reward": 0.8749999403953552, + "token_credit_reward": -0.34871822595596313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20335647463798523, + 0.2773038446903229, + 0.13589753210544586, + 0.12598268687725067, + 0.0859457477927208, + 0.03377385810017586, + 0.024831319227814674, + 0.05216243118047714, + 0.019431503489613533, + 0.041314613074064255 + ], + "credit_map": [ + 0.029405955225229263, + 0.019444022327661514, + 0.026118138805031776, + 0.018333489075303078, + 0.025865525007247925, + 0.02706444449722767, + 0.02708122693002224, + 0.030043667182326317, + 0.030838381499052048, + 0.019486520439386368, + 0.035497333854436874, + 0.02213551662862301, + 0.02684534713625908, + 0.03991797938942909, + 0.02841844968497753, + 0.07416430115699768, + 0.014096795581281185, + 0.013810631819069386, + 0.016819266602396965, + 0.012875538319349289, + 0.013897138647735119, + 0.019233185797929764, + 0.0163341723382473, + 0.02883080393075943, + 0.013368017971515656, + 0.0122475391253829, + 0.01812712848186493, + 0.012705880217254162, + 0.01662496291100979, + 0.016838235780596733, + 0.019024066627025604, + 0.01704685389995575, + 0.011121791787445545, + 0.008497660048305988, + 0.019522691145539284, + 0.006473828107118607, + 0.010194345377385616, + 0.01040935143828392, + 0.01075651403516531, + 0.008969573304057121, + 0.003303881036117673, + 0.005261955317109823, + 0.003922447562217712, + 0.004588725045323372, + 0.005559259094297886, + 0.0039954278618097305, + 0.004337001591920853, + 0.0028051622211933136, + 0.0030313029419630766, + 0.002844420960173011, + 0.00303547247312963, + 0.0033347555436193943, + 0.00235366215929389, + 0.0034848456270992756, + 0.003719908185303211, + 0.0030269508715718985, + 0.006011328659951687, + 0.0062936595641076565, + 0.009770565666258335, + 0.00613141804933548, + 0.006268507335335016, + 0.0074650682508945465, + 0.0064087677747011185, + 0.0038131196051836014, + 0.0019756117835640907, + 0.0018920854199677706, + 0.0020484086126089096, + 0.002318157348781824, + 0.0022706424351781607, + 0.0026284835767000914, + 0.003513375297188759, + 0.002784739714115858, + 0.004762907046824694, + 0.004208693280816078, + 0.008130214177072048, + 0.003375793108716607, + 0.004817545413970947, + 0.005572016816586256, + 0.005218883510679007, + 0.005228559486567974 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.34871822595596313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a close-up of an elderly woman's face taken from an extremely low angle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a soft, golden beam of sunlight slanting in from the side, illuminating wrinkles and age spots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does a full, glistening tear slowly escape from the corner of the woman's right eye and slide down her cheek?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred into a warm light spot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607106626033783 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.01607106626033783, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607106626033783, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4002881944179535, + 0.21311822533607483, + 0.10270289331674576, + 0.07507415860891342, + 0.030179312452673912, + 0.04286721348762512, + 0.03364153951406479, + 0.05472823977470398, + 0.016025947406888008, + 0.03137422725558281 + ], + "credit_map": [ + 0.03699600696563721, + 0.041647449135780334, + 0.072463758289814, + 0.038526877760887146, + 0.047088757157325745, + 0.06498009711503983, + 0.044536735862493515, + 0.05404851958155632, + 0.025815440341830254, + 0.01431118231266737, + 0.03156145289540291, + 0.015613384544849396, + 0.020287247374653816, + 0.029787631705403328, + 0.018702495843172073, + 0.05703938379883766, + 0.010783880949020386, + 0.012616514228284359, + 0.013457829132676125, + 0.010199184529483318, + 0.012273812666535378, + 0.015445109456777573, + 0.010380927473306656, + 0.017545640468597412, + 0.006548341829329729, + 0.007720853667706251, + 0.006357272155582905, + 0.009746287018060684, + 0.008995995856821537, + 0.010742267593741417, + 0.015097521245479584, + 0.009865614585578442, + 0.003203872824087739, + 0.0041458019986748695, + 0.004225846379995346, + 0.0033284283708781004, + 0.004780138377100229, + 0.003950811922550201, + 0.004156102892011404, + 0.002388309920206666, + 0.004544926341623068, + 0.00523929949849844, + 0.007187848910689354, + 0.005087284371256828, + 0.006612060591578484, + 0.00513135502114892, + 0.00423468928784132, + 0.004829749930649996, + 0.004807291552424431, + 0.003905495395883918, + 0.006704910192638636, + 0.0027020613197237253, + 0.003700737375766039, + 0.004320394713431597, + 0.0034946552477777004, + 0.0040059941820800304, + 0.005720441695302725, + 0.007682110648602247, + 0.00941358134150505, + 0.0059729162603616714, + 0.005327138584107161, + 0.006536807399243116, + 0.007311802823096514, + 0.006763442885130644, + 0.0018398996908217669, + 0.0019399526063352823, + 0.002374762436375022, + 0.0019205576973035932, + 0.0016981355147436261, + 0.0021937380079180002, + 0.0023688944056630135, + 0.0016900072805583477, + 0.0040217177011072636, + 0.003379817819222808, + 0.005981074180454016, + 0.0031863460317254066, + 0.0038384359795600176, + 0.003729606745764613, + 0.0037953616119921207, + 0.003441866720095277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.01607106626033783 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_tensor([2342, 2342])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_tensor([2342, 2342])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..57bce6dc61f58ae3565639cd382f2b1b045518a9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_tensor([2342, 2342])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot captures the tense scene, rendered in a clean, graphic animation style against a deep blue void. Thin, sharp white lines radiate outward from an invisible point off-screen, creating a stark web-like pattern that defines the abstract space. In the center, several pale figures gather. The leftmost figure, frozen in mid-shouting, their mouths agape, their faces contorted in pain in silent despair, clad in a heavy, dark purple robe. Just behind them, a second figure in a somber gray robe places a steady hand on the shouting figure's shoulder, a gesture of comfort or restraint, their own expression serious and focused. The relative stillness of the central group contrasts sharply with the surrounding chaos, as blurred, shadowy outlines race through the foreground and background, their forms blurred and ghostly. The lighting is harsh and cool, casting sharp shadows that highlight the pale tones of their skin against the saturated colors of their purple, pink, and gray robes, creating a dark and tense atmosphere that freezes a moment of high-stakes drama and emotional crisis.", + "seed": 31533, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_2342_seed_31533_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.016617601388134062, + "vlm_logprob": -0.016617601388134062, + "vlm_yes_no_margin": 5.375, + "vlm_token_credit": -0.016617601388134062, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032567339949309826 + ], + "opposite_logprob_tokens": [ + -7.753256797790527 + ], + "logprob_reward": -0.0032567339949309826, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.0032567339949309826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07614131271839142, + 0.027465933933854103, + 0.014044961892068386, + 0.09311974048614502, + 0.01587137021124363, + 0.019703010097146034, + 0.027613984420895576, + 0.03999032825231552, + 0.049625836312770844, + 0.6364235281944275 + ], + "credit_map": [ + 0.012370046228170395, + 0.00932428427040577, + 0.007525351829826832, + 0.008390003815293312, + 0.007134571671485901, + 0.007731336168944836, + 0.011194327846169472, + 0.012471390888094902, + 0.0037642954848706722, + 0.0044189696200191975, + 0.001877261558547616, + 0.0027129247318953276, + 0.004610599018633366, + 0.0038720467127859592, + 0.003365684999153018, + 0.002844153670594096, + 0.0012703234096989036, + 0.0019541846122592688, + 0.0013549969298765063, + 0.0014202609891071916, + 0.0014765433734282851, + 0.0029255750123411417, + 0.0018852349603548646, + 0.00175784295424819, + 0.0030254425946623087, + 0.0035344217903912067, + 0.0022736673709005117, + 0.0029124501161277294, + 0.004438858013600111, + 0.03724338486790657, + 0.004303854890167713, + 0.035387661308050156, + 0.0022068535909056664, + 0.0011678652372211218, + 0.001579070114530623, + 0.0012340210378170013, + 0.0014970683259889483, + 0.004574581515043974, + 0.0018792462069541216, + 0.0017326633678749204, + 0.0019522246439009905, + 0.0016757099656388164, + 0.0009969205129891634, + 0.0017102161655202508, + 0.003019969444721937, + 0.004377993289381266, + 0.0027315455954521894, + 0.0032384302467107773, + 0.002846428891643882, + 0.002897299360483885, + 0.0031646452844142914, + 0.003521210979670286, + 0.0030095584224909544, + 0.003145939437672496, + 0.0036047021858394146, + 0.005424200091511011, + 0.011685536243021488, + 0.0020861136727035046, + 0.0020118786487728357, + 0.0033509947825223207, + 0.0028704367578029633, + 0.003406435251235962, + 0.0030949036590754986, + 0.011484027840197086, + 0.004420962184667587, + 0.005573093891143799, + 0.0074293045327067375, + 0.005884991027414799, + 0.004733389709144831, + 0.008410804904997349, + 0.006536560133099556, + 0.0066367327235639095, + 0.23231880366802216, + 0.0322108156979084, + 0.02733648382127285, + 0.01158691942691803, + 0.02444593608379364, + 0.254695326089859, + 0.01195240207016468, + 0.041876863688230515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.0032567339949309826 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00195229088421911 + ], + "opposite_logprob_tokens": [ + -6.251952171325684 + ], + "logprob_reward": -0.00195229088421911, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.00195229088421911, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03382014110684395, + 0.05021710693836212, + 0.025434931740164757, + 0.19502657651901245, + 0.06047520786523819, + 0.021115785464644432, + 0.030900025740265846, + 0.06873692572116852, + 0.009694679640233517, + 0.5045786499977112 + ], + "credit_map": [ + 0.0058630406856536865, + 0.003413270926102996, + 0.0035502666141837835, + 0.004288697615265846, + 0.004298230167478323, + 0.005008403677493334, + 0.0044950577430427074, + 0.0029031711164861917, + 0.006801238749176264, + 0.005080289673060179, + 0.007154807914048433, + 0.006323720794171095, + 0.00404464453458786, + 0.01015112828463316, + 0.005582077894359827, + 0.005079198163002729, + 0.002478615613654256, + 0.0027801389805972576, + 0.0018072096863761544, + 0.0017801853828132153, + 0.0021414279472082853, + 0.0067407796159386635, + 0.0034664722625166178, + 0.004240101203322411, + 0.0075246719643473625, + 0.006178903393447399, + 0.003014804096892476, + 0.0027828433085232973, + 0.009374082088470459, + 0.08459824323654175, + 0.004796895664185286, + 0.07675613462924957, + 0.008414115756750107, + 0.009727834723889828, + 0.005592252127826214, + 0.00652189739048481, + 0.00552366254851222, + 0.011390388943254948, + 0.0041100988164544106, + 0.009194958955049515, + 0.002168317325413227, + 0.002434954745694995, + 0.0011048776796087623, + 0.0017089885659515858, + 0.0025752768851816654, + 0.0051151118241250515, + 0.0025490126572549343, + 0.0034592465963214636, + 0.0032449443824589252, + 0.005658145993947983, + 0.0030507519841194153, + 0.0041760727763175964, + 0.0035299721639603376, + 0.003871859982609749, + 0.00399587070569396, + 0.003372407052665949, + 0.009664311073720455, + 0.005459851585328579, + 0.007324348669499159, + 0.010502012446522713, + 0.006120398174971342, + 0.008227529004216194, + 0.011443579569458961, + 0.009994895197451115, + 0.0011213149409741163, + 0.0011365762911736965, + 0.001160152954980731, + 0.0012613440630957484, + 0.0012644424568861723, + 0.0013838371960446239, + 0.0011924150167033076, + 0.0011745963711291552, + 0.1798066794872284, + 0.025205671787261963, + 0.021785007789731026, + 0.010502743534743786, + 0.020023275166749954, + 0.20205923914909363, + 0.00979445781558752, + 0.03540156036615372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.00195229088421911 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017149040941148996 + ], + "opposite_logprob_tokens": [ + -6.376714706420898 + ], + "logprob_reward": -0.0017149040941148996, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017149040941148996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10152973234653473, + 0.054559376090765, + 0.01913587562739849, + 0.4187101721763611, + 0.019384996965527534, + 0.01696711964905262, + 0.030199691653251648, + 0.015520750544965267, + 0.008168455213308334, + 0.31582385301589966 + ], + "credit_map": [ + 0.020124360918998718, + 0.010320227593183517, + 0.010397575795650482, + 0.008298368193209171, + 0.006403358653187752, + 0.014036878012120724, + 0.011242158710956573, + 0.020706800743937492, + 0.005204350221902132, + 0.004652454983443022, + 0.0055010621435940266, + 0.006009332835674286, + 0.003793268697336316, + 0.014610071666538715, + 0.0066242641769349575, + 0.008164568804204464, + 0.0022189111914485693, + 0.002287546405568719, + 0.002161326352506876, + 0.0025451299734413624, + 0.0026375853922218084, + 0.002093174261972308, + 0.0028716151136904955, + 0.002320585772395134, + 0.016278184950351715, + 0.012455293908715248, + 0.00602194806560874, + 0.00546058127656579, + 0.01976925879716873, + 0.1823832392692566, + 0.010942927561700344, + 0.1653987020254135, + 0.002902005333453417, + 0.0015240528155118227, + 0.0017694062553346157, + 0.0020799196790903807, + 0.0018796335207298398, + 0.004616508260369301, + 0.0021494287066161633, + 0.002464042045176029, + 0.0018400303088128567, + 0.0014293554704636335, + 0.0009026351617649198, + 0.0012985898647457361, + 0.0023432904854416847, + 0.004597307648509741, + 0.0016622960101813078, + 0.0028936131857335567, + 0.003601107746362686, + 0.002819700399413705, + 0.003184722503647208, + 0.004162340424954891, + 0.0033511738292872906, + 0.00396003620699048, + 0.0040437001734972, + 0.005076907109469175, + 0.0023638613056391478, + 0.0019363544415682554, + 0.0013267198810353875, + 0.0017192792147397995, + 0.001500684767961502, + 0.0020860699005424976, + 0.002114917617291212, + 0.0024728626012802124, + 0.0010850632097572088, + 0.0008498269598931074, + 0.0011505443835631013, + 0.00096669647609815, + 0.0012700422666966915, + 0.0010206843726336956, + 0.0010408188682049513, + 0.0007847783854231238, + 0.11494819819927216, + 0.015887128189206123, + 0.013790319673717022, + 0.005615245550870895, + 0.012186071835458279, + 0.12676624953746796, + 0.006033680867403746, + 0.020596938207745552 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.0017149040941148996 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03807767853140831 + ], + "opposite_logprob_tokens": [ + -3.2880775928497314 + ], + "logprob_reward": -0.03807767853140831, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03807767853140831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13738131523132324, + 0.05586908757686615, + 0.02192506194114685, + 0.3903588652610779, + 0.02632828801870346, + 0.03033933974802494, + 0.035431262105703354, + 0.02367871254682541, + 0.019316865131258965, + 0.25937116146087646 + ], + "credit_map": [ + 0.022822333499789238, + 0.015235747210681438, + 0.014239128679037094, + 0.011294671334326267, + 0.01720828004181385, + 0.01789937913417816, + 0.014992062933743, + 0.023689720779657364, + 0.0038619600236415863, + 0.006296373438090086, + 0.006550290621817112, + 0.006170738488435745, + 0.005667453166097403, + 0.010703854262828827, + 0.009828267619013786, + 0.006790152285248041, + 0.0023942498955875635, + 0.0017403231468051672, + 0.002961129415780306, + 0.0026030410081148148, + 0.0016293843509629369, + 0.004395608324557543, + 0.004010312724858522, + 0.0021910122595727444, + 0.014592444524168968, + 0.01194856595247984, + 0.006458907388150692, + 0.006374251563102007, + 0.0185879934579134, + 0.1671907752752304, + 0.011699541471898556, + 0.15350639820098877, + 0.004043067805469036, + 0.0022247578017413616, + 0.0019648598972707987, + 0.0020033784676343203, + 0.0026548339519649744, + 0.00872854981571436, + 0.0020773112773895264, + 0.0026315273717045784, + 0.0031739254482090473, + 0.0028315603267401457, + 0.0016468856483697891, + 0.002202878938987851, + 0.0044041466899216175, + 0.008272632956504822, + 0.0025303030852228403, + 0.005277007818222046, + 0.004162989556789398, + 0.003212059149518609, + 0.003935148008167744, + 0.005212271586060524, + 0.003393821883946657, + 0.004408644512295723, + 0.005786324851214886, + 0.005320001859217882, + 0.003880831180140376, + 0.002823382616043091, + 0.002200416289269924, + 0.0026850332506000996, + 0.001860103220678866, + 0.0029982186388224363, + 0.003304783022031188, + 0.0039259446784853935, + 0.0023015267215669155, + 0.0025448433589190245, + 0.0019913981668651104, + 0.002707430627197027, + 0.0018961348105221987, + 0.0033195645082741976, + 0.002789201680570841, + 0.0017667640931904316, + 0.09463781118392944, + 0.013677546754479408, + 0.01132956612855196, + 0.005652531981468201, + 0.010193676687777042, + 0.10186450928449631, + 0.00579972006380558, + 0.016215801239013672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.03807767853140831 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808639943599701 + ], + "opposite_logprob_tokens": [ + -3.288086414337158 + ], + "logprob_reward": -0.03808639943599701, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808639943599701, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06341009587049484, + 0.03275962546467781, + 0.017697181552648544, + 0.6178182363510132, + 0.015007422305643559, + 0.024429189041256905, + 0.018310043960809708, + 0.035114485770463943, + 0.023932771757245064, + 0.1515209972858429 + ], + "credit_map": [ + 0.013650854118168354, + 0.0077271913178265095, + 0.006041988730430603, + 0.004695860203355551, + 0.005080519709736109, + 0.007562853395938873, + 0.00767888966947794, + 0.010971933603286743, + 0.0035694034304469824, + 0.003501655999571085, + 0.0030579743906855583, + 0.0036915161181241274, + 0.004366781562566757, + 0.006135414354503155, + 0.0039534554816782475, + 0.004483423661440611, + 0.0016483329236507416, + 0.0017059865640476346, + 0.0017222532769665122, + 0.001969295786693692, + 0.002484510885551572, + 0.002164091682061553, + 0.0038152874913066626, + 0.0021874206140637398, + 0.024722086265683174, + 0.017782049253582954, + 0.009072101674973965, + 0.007475729566067457, + 0.028850922361016273, + 0.270321786403656, + 0.0157619696110487, + 0.2438315749168396, + 0.0029119038954377174, + 0.0013427551602944732, + 0.0010836737928912044, + 0.0010346209164708853, + 0.002253413200378418, + 0.003435574471950531, + 0.0014420778024941683, + 0.001503401785157621, + 0.0025554411113262177, + 0.0020784863736480474, + 0.0011970854829996824, + 0.001644120435230434, + 0.0035601642448455095, + 0.006674571894109249, + 0.002359449863433838, + 0.004359868820756674, + 0.002238909713923931, + 0.0020753026474267244, + 0.0019857711158692837, + 0.0025683133862912655, + 0.0015192202990874648, + 0.0017984810983762145, + 0.0030149808153510094, + 0.0031090632546693087, + 0.004488887265324593, + 0.004079185426235199, + 0.003215107833966613, + 0.00259279552847147, + 0.0028626269195228815, + 0.0060170311480760574, + 0.0028049314860254526, + 0.009053919464349747, + 0.0019372073002159595, + 0.0030448376201093197, + 0.0025237775407731533, + 0.0034553275909274817, + 0.0032194985542446375, + 0.0032394067384302616, + 0.00405051838606596, + 0.002462194999679923, + 0.05528402328491211, + 0.008486972190439701, + 0.006585803814232349, + 0.004074261523783207, + 0.0064201317727565765, + 0.05800700932741165, + 0.004165807738900185, + 0.00849697832018137 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.03808639943599701 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the tense scene, rendered in a clean, graphic animation style against a deep blue void. Thin, sharp white lines radiate outward from an invisible point off-screen, creating a stark web-like pattern that defines the abstract space. In the center, several pale figures gather. The leftmost figure, frozen in mid-shouting, their mouths agape, their faces contorted in pain in silent despair, clad in a heavy, dark purple robe. Just behind them, a second figure in a somber gray robe places a steady hand on the shouting figure's shoulder, a gesture of comfort or restraint, their own expression serious and focused. The relative stillness of the central group contrasts sharply with the surrounding chaos, as blurred, shadowy outlines race through the foreground and background, their forms blurred and ghostly. The lighting is harsh and cool, casting sharp shadows that highlight the pale tones of their skin against the saturated colors of their purple, pink, and gray robes, creating a dark and tense atmosphere that freezes a moment of high-stakes drama and emotional crisis.", + "seed": 31534, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_2342_seed_31534_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008717246539890765, + "vlm_logprob": -0.008717246539890765, + "vlm_yes_no_margin": 6.1, + "vlm_token_credit": -0.008717246539890765, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0038790472317487 + ], + "opposite_logprob_tokens": [ + -8.253878593444824 + ], + "logprob_reward": -0.0038790472317487, + "yes_no_margin_reward": 8.25, + "token_credit_reward": -0.0038790472317487, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11621497571468353, + 0.11006946116685867, + 0.0588463731110096, + 0.07672964036464691, + 0.09280030429363251, + 0.0628366470336914, + 0.10702764987945557, + 0.3016842305660248, + 0.03732297569513321, + 0.03646774962544441 + ], + "credit_map": [ + 0.03049680031836033, + 0.008868357166647911, + 0.00872600544244051, + 0.010054268874228, + 0.007145864889025688, + 0.019208675250411034, + 0.01726210117340088, + 0.01445289608091116, + 0.017034024000167847, + 0.00690183462575078, + 0.012483879923820496, + 0.010136914439499378, + 0.008376427926123142, + 0.022481758147478104, + 0.011778326705098152, + 0.02087630145251751, + 0.018105970695614815, + 0.002389003988355398, + 0.0032584951259195805, + 0.0033100543078035116, + 0.004076821729540825, + 0.015791859477758408, + 0.004649748094379902, + 0.007264418993145227, + 0.0053707147017121315, + 0.0036658085882663727, + 0.0041937194764614105, + 0.005224150139838457, + 0.006252774968743324, + 0.012982736341655254, + 0.017073802649974823, + 0.021965928375720978, + 0.007983606308698654, + 0.003611107124015689, + 0.0062757451087236404, + 0.006946254521608353, + 0.006251298822462559, + 0.008999771438539028, + 0.03336668759584427, + 0.019365835934877396, + 0.007203646004199982, + 0.003770257579162717, + 0.005047937855124474, + 0.005437787156552076, + 0.005473110359162092, + 0.011427956633269787, + 0.01166156679391861, + 0.01281438022851944, + 0.01457157637923956, + 0.006322448141872883, + 0.0084375636652112, + 0.008694145828485489, + 0.010583938099443913, + 0.01506130676716566, + 0.013968347571790218, + 0.02938832901418209, + 0.09218912571668625, + 0.007989060133695602, + 0.009957739152014256, + 0.009153935126960278, + 0.007831520400941372, + 0.04258377477526665, + 0.013798625208437443, + 0.11818043887615204, + 0.006042638793587685, + 0.0049270386807620525, + 0.004335090052336454, + 0.005101464223116636, + 0.0033123348839581013, + 0.005795502569526434, + 0.0042289323173463345, + 0.00357997533865273, + 0.006434463895857334, + 0.003548399079591036, + 0.003540908684954047, + 0.003901067888364196, + 0.0033068594057112932, + 0.004557645879685879, + 0.002913873875513673, + 0.008264533244073391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.0038790472317487 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025022882036864758 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.0025022882036864758, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025022882036864758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13950316607952118, + 0.09791658073663712, + 0.057313285768032074, + 0.2384471595287323, + 0.11547493189573288, + 0.0718904435634613, + 0.047073639929294586, + 0.174093559384346, + 0.020915992558002472, + 0.03737126663327217 + ], + "credit_map": [ + 0.04370519518852234, + 0.006228002719581127, + 0.006996132433414459, + 0.009412122890353203, + 0.00855580810457468, + 0.025812305510044098, + 0.024831200018525124, + 0.013962404802441597, + 0.014848566614091396, + 0.00796908512711525, + 0.01067289337515831, + 0.010340186767280102, + 0.009826412424445152, + 0.012905503623187542, + 0.01548896823078394, + 0.015864968299865723, + 0.008786232210695744, + 0.00368630257435143, + 0.00534848403185606, + 0.005265137646347284, + 0.005067904945462942, + 0.012156901881098747, + 0.008757366798818111, + 0.008244955912232399, + 0.02967800572514534, + 0.011436490342020988, + 0.013745181262493134, + 0.015325462445616722, + 0.018372859805822372, + 0.03708792105317116, + 0.048983827233314514, + 0.06381741166114807, + 0.02310161665081978, + 0.0063887061551213264, + 0.009474757127463818, + 0.008487668819725513, + 0.008033804595470428, + 0.0195353701710701, + 0.0170661099255085, + 0.02338690124452114, + 0.012517516501247883, + 0.003175540128722787, + 0.0045351292937994, + 0.006028228905051947, + 0.004378853365778923, + 0.011140686459839344, + 0.017766479402780533, + 0.012348007410764694, + 0.005682196468114853, + 0.002751656575128436, + 0.0038901674561202526, + 0.0042665572836995125, + 0.004141961690038443, + 0.00885061826556921, + 0.004828143399208784, + 0.012662336230278015, + 0.05261903628706932, + 0.0049719070084393024, + 0.005540837068110704, + 0.0045607006177306175, + 0.004869361873716116, + 0.026627251878380775, + 0.005560948513448238, + 0.06934351474046707, + 0.004544718191027641, + 0.001677890308201313, + 0.002074332209303975, + 0.0017001921078190207, + 0.0020543006248772144, + 0.0036911966744810343, + 0.0032307482324540615, + 0.0019426146754994988, + 0.009715585969388485, + 0.0021121581085026264, + 0.0022111486177891493, + 0.002050666371360421, + 0.0017075061332434416, + 0.0051248990930616856, + 0.002032579854130745, + 0.012416723184287548 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.0025022882036864758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00104301399551332 + ], + "opposite_logprob_tokens": [ + -6.87604284286499 + ], + "logprob_reward": -0.00104301399551332, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00104301399551332, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21280479431152344, + 0.1997000277042389, + 0.10142843425273895, + 0.04540890455245972, + 0.09137112647294998, + 0.05820288136601448, + 0.07301981747150421, + 0.10166563093662262, + 0.04605865851044655, + 0.07033982127904892 + ], + "credit_map": [ + 0.058847930282354355, + 0.019355757161974907, + 0.01454393845051527, + 0.01684108003973961, + 0.018605606630444527, + 0.030282868072390556, + 0.03668372333049774, + 0.017643863335251808, + 0.036344967782497406, + 0.012521575205028057, + 0.027184076607227325, + 0.021395491436123848, + 0.018021833151578903, + 0.02581210434436798, + 0.024251841008663177, + 0.034168120473623276, + 0.022870445623993874, + 0.011676265858113766, + 0.006001616828143597, + 0.008078139275312424, + 0.009867900982499123, + 0.018327416852116585, + 0.01065229345113039, + 0.013954345136880875, + 0.007422182243317366, + 0.0026940391398966312, + 0.004627654794603586, + 0.004715881310403347, + 0.003547665663063526, + 0.009536229074001312, + 0.00641068909317255, + 0.00645455764606595, + 0.00826989021152258, + 0.003754905890673399, + 0.007742931134998798, + 0.007771920412778854, + 0.005873079877346754, + 0.011415800079703331, + 0.025583993643522263, + 0.020958591252565384, + 0.01048977766185999, + 0.0026522690895944834, + 0.0029123080894351006, + 0.006514678243547678, + 0.0041264733299613, + 0.007977615110576153, + 0.01908883824944496, + 0.004440920893102884, + 0.008726958185434341, + 0.0056845564395189285, + 0.005807928275316954, + 0.005765735171735287, + 0.010924692265689373, + 0.010155199095606804, + 0.010626757517457008, + 0.015327981673181057, + 0.0158157330006361, + 0.00956970825791359, + 0.00763857364654541, + 0.009006904438138008, + 0.008072126656770706, + 0.018313048407435417, + 0.010266672819852829, + 0.02298286184668541, + 0.007947341538965702, + 0.009917995892465115, + 0.004884083289653063, + 0.005329894833266735, + 0.004447728395462036, + 0.006232365965843201, + 0.0037614735774695873, + 0.003537773620337248, + 0.01737617328763008, + 0.00188166880980134, + 0.002970166737213731, + 0.0028913463465869427, + 0.002340511418879032, + 0.012088149785995483, + 0.0031601411756128073, + 0.027631662786006927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.00104301399551332 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002483262214809656 + ], + "opposite_logprob_tokens": [ + -6.002483367919922 + ], + "logprob_reward": -0.002483262214809656, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002483262214809656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2104029804468155, + 0.12844449281692505, + 0.07907196134328842, + 0.021305739879608154, + 0.057295795530080795, + 0.02982526645064354, + 0.12757901847362518, + 0.21666067838668823, + 0.025619545951485634, + 0.10379447042942047 + ], + "credit_map": [ + 0.05674955993890762, + 0.014678449369966984, + 0.012511539272964, + 0.015892453491687775, + 0.01658741757273674, + 0.03367800638079643, + 0.030781470239162445, + 0.029524091631174088, + 0.021523011848330498, + 0.009976307861506939, + 0.01659693941473961, + 0.015369505621492863, + 0.01297144964337349, + 0.01652953028678894, + 0.014293637126684189, + 0.021184133365750313, + 0.023828228935599327, + 0.006060806103050709, + 0.00523221492767334, + 0.007865307852625847, + 0.006211903877556324, + 0.01674089953303337, + 0.007817042991518974, + 0.005315566901117563, + 0.0031150823924690485, + 0.0017256776336580515, + 0.002055657096207142, + 0.00236784596927464, + 0.0016802652971819043, + 0.004149550572037697, + 0.0030777640640735626, + 0.0031338981352746487, + 0.00459542078897357, + 0.0029120452236384153, + 0.004924138076603413, + 0.00444850604981184, + 0.004092360381036997, + 0.007722104899585247, + 0.014668646268546581, + 0.013932580128312111, + 0.005756538826972246, + 0.002310737268999219, + 0.002209070371463895, + 0.0035496645141392946, + 0.0024322685785591602, + 0.004504137206822634, + 0.0060778227634727955, + 0.0029850297141820192, + 0.02017720229923725, + 0.00962095893919468, + 0.008705566637217999, + 0.011337969452142715, + 0.012458937242627144, + 0.0190496314316988, + 0.013690771535038948, + 0.03253799304366112, + 0.05076179280877113, + 0.006321536842733622, + 0.00734803918749094, + 0.008389392867684364, + 0.009627973660826683, + 0.03873473033308983, + 0.01434707548469305, + 0.08113015443086624, + 0.00449552945792675, + 0.004932698793709278, + 0.002432155655696988, + 0.0026991949416697025, + 0.0022698561660945415, + 0.0039051519706845284, + 0.0024873611982911825, + 0.0023975989315658808, + 0.024557994678616524, + 0.003923341631889343, + 0.006836386863142252, + 0.007826427929103374, + 0.005101310554891825, + 0.012887183576822281, + 0.008748614229261875, + 0.03391322121024132 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 1 + ], + "selected_reward": -0.002483262214809656 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367862105369568 + ], + "opposite_logprob_tokens": [ + -3.4086785316467285 + ], + "logprob_reward": -0.03367862105369568, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367862105369568, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24032942950725555, + 0.11650042235851288, + 0.08167080581188202, + 0.06285165995359421, + 0.05007370561361313, + 0.08136767148971558, + 0.07290361821651459, + 0.22880002856254578, + 0.03960822522640228, + 0.025894463062286377 + ], + "credit_map": [ + 0.04813581705093384, + 0.012167876586318016, + 0.017919575795531273, + 0.01486917119473219, + 0.022729290649294853, + 0.045779015868902206, + 0.02957846224308014, + 0.049150221049785614, + 0.020630547776818275, + 0.005781444255262613, + 0.014238646253943443, + 0.00954029243439436, + 0.009726615622639656, + 0.01961640641093254, + 0.01099316868931055, + 0.02597329579293728, + 0.019482700154185295, + 0.004567693453282118, + 0.005474672187119722, + 0.005697106942534447, + 0.005446588154882193, + 0.019951600581407547, + 0.007056776434183121, + 0.013993663713335991, + 0.005985239055007696, + 0.004897930193692446, + 0.0057806032709777355, + 0.006356921512633562, + 0.0064098467119038105, + 0.008521472103893757, + 0.012268287129700184, + 0.012631355784833431, + 0.005810179747641087, + 0.0029661247972398996, + 0.004005428869277239, + 0.004752494860440493, + 0.004096039105206728, + 0.008346554823219776, + 0.013495398685336113, + 0.006601481698453426, + 0.014229738153517246, + 0.004991880618035793, + 0.00582514563575387, + 0.009715339168906212, + 0.006375543773174286, + 0.014003190211951733, + 0.017984429374337196, + 0.008242398500442505, + 0.009950588457286358, + 0.006240431219339371, + 0.006383518222719431, + 0.005740546155720949, + 0.007523697335273027, + 0.01388796977698803, + 0.006536041386425495, + 0.01664082519710064, + 0.0658494159579277, + 0.006574955768883228, + 0.006604071240872145, + 0.00702527305111289, + 0.0066694035194814205, + 0.033998310565948486, + 0.01093417126685381, + 0.09114442765712738, + 0.007215069606900215, + 0.002679607830941677, + 0.0031423228792846203, + 0.0027211501728743315, + 0.0039873444475233555, + 0.008483299054205418, + 0.005136256571859121, + 0.0062431711703538895, + 0.005830186419188976, + 0.001555425813421607, + 0.0014290125109255314, + 0.001845783437602222, + 0.0016309797065332532, + 0.003911771345883608, + 0.0021222440991550684, + 0.007569059729576111 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.03367862105369568 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the tense scene, rendered in a clean, graphic animation style against a deep blue void. Thin, sharp white lines radiate outward from an invisible point off-screen, creating a stark web-like pattern that defines the abstract space. In the center, several pale figures gather. The leftmost figure, frozen in mid-shouting, their mouths agape, their faces contorted in pain in silent despair, clad in a heavy, dark purple robe. Just behind them, a second figure in a somber gray robe places a steady hand on the shouting figure's shoulder, a gesture of comfort or restraint, their own expression serious and focused. The relative stillness of the central group contrasts sharply with the surrounding chaos, as blurred, shadowy outlines race through the foreground and background, their forms blurred and ghostly. The lighting is harsh and cool, casting sharp shadows that highlight the pale tones of their skin against the saturated colors of their purple, pink, and gray robes, creating a dark and tense atmosphere that freezes a moment of high-stakes drama and emotional crisis.", + "seed": 31535, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_2342_seed_31535_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.014969254843890667, + "vlm_logprob": -0.014969254843890667, + "vlm_yes_no_margin": 5.35, + "vlm_token_credit": -0.014969254843890667, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008061609230935574 + ], + "opposite_logprob_tokens": [ + -7.758061408996582 + ], + "logprob_reward": -0.008061609230935574, + "yes_no_margin_reward": 7.75, + "token_credit_reward": -0.008061609230935574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17076356709003448, + 0.08972735702991486, + 0.025531738996505737, + 0.08460589498281479, + 0.0792512372136116, + 0.05593430623412132, + 0.13648086786270142, + 0.13005544245243073, + 0.16258276998996735, + 0.06506676971912384 + ], + "credit_map": [ + 0.021181456744670868, + 0.01832493394613266, + 0.019442811608314514, + 0.022285962477326393, + 0.016003906726837158, + 0.022163942456245422, + 0.02252584509551525, + 0.028834715485572815, + 0.009282500483095646, + 0.012068451382219791, + 0.017304077744483948, + 0.013697498477995396, + 0.007839946076273918, + 0.009690545499324799, + 0.010466028936207294, + 0.009378313086926937, + 0.0028027123771607876, + 0.0033129423391073942, + 0.002825044794008136, + 0.0027401449624449015, + 0.004186456091701984, + 0.0032110216561704874, + 0.0034580016508698463, + 0.002995414426550269, + 0.00828008633106947, + 0.01016236376017332, + 0.023370685055851936, + 0.006435778923332691, + 0.008458600379526615, + 0.007471373304724693, + 0.009047050029039383, + 0.011379960924386978, + 0.007388027850538492, + 0.010455844923853874, + 0.022483961656689644, + 0.006730457302182913, + 0.0070063029415905476, + 0.0063574961386621, + 0.006845975760370493, + 0.011983167380094528, + 0.006517794914543629, + 0.006978162098675966, + 0.009503716602921486, + 0.004902282729744911, + 0.007204135879874229, + 0.0047154841013252735, + 0.007764532696455717, + 0.008348198607563972, + 0.014648252166807652, + 0.01606590487062931, + 0.024354498833417892, + 0.013229531235992908, + 0.012138848192989826, + 0.015309418551623821, + 0.019490351900458336, + 0.021244054660201073, + 0.014218918047845364, + 0.01614997908473015, + 0.017677370458841324, + 0.013902626000344753, + 0.018936118111014366, + 0.011676273308694363, + 0.017769817262887955, + 0.019724346697330475, + 0.019086593762040138, + 0.020353641360998154, + 0.026012830436229706, + 0.016739385202527046, + 0.014083122834563255, + 0.01684505306184292, + 0.02062058635056019, + 0.028841564431786537, + 0.006919316481798887, + 0.00811829324811697, + 0.012007329612970352, + 0.00663337716832757, + 0.006133226677775383, + 0.007062267977744341, + 0.00940596405416727, + 0.008786998689174652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.008061609230935574 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036311899311840534 + ], + "opposite_logprob_tokens": [ + -5.628631114959717 + ], + "logprob_reward": -0.0036311899311840534, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036311899311840534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2080502212047577, + 0.11834099888801575, + 0.038703348487615585, + 0.07680357992649078, + 0.07394874840974808, + 0.06454543024301529, + 0.1118723675608635, + 0.1511387676000595, + 0.05035175010561943, + 0.10624474287033081 + ], + "credit_map": [ + 0.024869298562407494, + 0.026371661573648453, + 0.026434117928147316, + 0.025298140943050385, + 0.021437620744109154, + 0.028578482568264008, + 0.020235413685441017, + 0.03482547402381897, + 0.01377204991877079, + 0.015381642617285252, + 0.009163948707282543, + 0.014993940480053425, + 0.013415857218205929, + 0.020165035501122475, + 0.015141062438488007, + 0.016307469457387924, + 0.005372696090489626, + 0.004777512513101101, + 0.004217368550598621, + 0.0033923324663192034, + 0.0052426219917833805, + 0.005868224427103996, + 0.004284159280359745, + 0.0055484347976744175, + 0.0086660236120224, + 0.008684364147484303, + 0.01741725206375122, + 0.006371868308633566, + 0.008279794827103615, + 0.009149196557700634, + 0.01007993333041668, + 0.008155141957104206, + 0.009166279807686806, + 0.009385310113430023, + 0.011723803356289864, + 0.005735231097787619, + 0.009032540023326874, + 0.010140332393348217, + 0.008506672456860542, + 0.010258583351969719, + 0.007526368368417025, + 0.007777432911098003, + 0.01152891106903553, + 0.005905960686504841, + 0.007558487821370363, + 0.008274617604911327, + 0.008753565140068531, + 0.007220088504254818, + 0.011096933856606483, + 0.012077944353222847, + 0.0173622015863657, + 0.009402460418641567, + 0.012627399526536465, + 0.013216369785368443, + 0.01575525477528572, + 0.020333800464868546, + 0.01575169526040554, + 0.01675863191485405, + 0.018971381708979607, + 0.020327674224972725, + 0.01388046145439148, + 0.02216462604701519, + 0.0186738483607769, + 0.024610446766018867, + 0.005515232682228088, + 0.006162158213555813, + 0.00814762618392706, + 0.0059183682315051556, + 0.00428373459726572, + 0.005865318235009909, + 0.006194933317601681, + 0.008264380507171154, + 0.010319043882191181, + 0.012811458669602871, + 0.010326450690627098, + 0.013274955563247204, + 0.009341645985841751, + 0.020631637424230576, + 0.011300035752356052, + 0.018239513039588928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.0036311899311840534 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019402741454541683 + ], + "opposite_logprob_tokens": [ + -6.2519402503967285 + ], + "logprob_reward": -0.0019402741454541683, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019402741454541683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3538767993450165, + 0.09918664395809174, + 0.042578693479299545, + 0.04104611650109291, + 0.056085795164108276, + 0.052217528223991394, + 0.1493779867887497, + 0.12961284816265106, + 0.03449440747499466, + 0.04152321442961693 + ], + "credit_map": [ + 0.046203650534152985, + 0.03782196715474129, + 0.059413231909275055, + 0.03299018740653992, + 0.03613044321537018, + 0.042967360466718674, + 0.03568808361887932, + 0.06266188621520996, + 0.011134699918329716, + 0.012001283466815948, + 0.014496524818241596, + 0.01108333095908165, + 0.013625742867588997, + 0.009420967660844326, + 0.012813660316169262, + 0.014610438607633114, + 0.0036865174770355225, + 0.006754678674042225, + 0.0051689608953893185, + 0.006064144428819418, + 0.004378331359475851, + 0.004610613454133272, + 0.007076092530041933, + 0.00483935559168458, + 0.004184054210782051, + 0.00457024248316884, + 0.010969053953886032, + 0.0031572196166962385, + 0.004363753832876682, + 0.004066310822963715, + 0.004509527701884508, + 0.005225955508649349, + 0.006827173754572868, + 0.00561368465423584, + 0.015331350266933441, + 0.004534285515546799, + 0.004948031157255173, + 0.005548520013689995, + 0.005742497276514769, + 0.007540250662714243, + 0.004646540153771639, + 0.005845862440764904, + 0.014687187038362026, + 0.004818908404558897, + 0.005071406252682209, + 0.005231539253145456, + 0.0053812856785953045, + 0.006534799933433533, + 0.013080764561891556, + 0.01853693090379238, + 0.0220994520932436, + 0.015828242525458336, + 0.024012576788663864, + 0.014392212964594364, + 0.02276759408414364, + 0.01866019517183304, + 0.011826064437627792, + 0.016632994636893272, + 0.02200387418270111, + 0.017363669350743294, + 0.011654055677354336, + 0.014192407950758934, + 0.01788211427628994, + 0.01805766671895981, + 0.003862605895847082, + 0.003817819058895111, + 0.005058690439909697, + 0.004509011749178171, + 0.003751444397494197, + 0.0037804257590323687, + 0.004118426702916622, + 0.0055959830060601234, + 0.003945847041904926, + 0.005358205642551184, + 0.006187811493873596, + 0.004008043557405472, + 0.0041432552970945835, + 0.0059686931781470776, + 0.0038999514654278755, + 0.008011406287550926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0019402741454541683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018171383067965508 + ], + "opposite_logprob_tokens": [ + -4.018171310424805 + ], + "logprob_reward": -0.018171383067965508, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018171383067965508, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3563327491283417, + 0.06830613315105438, + 0.06347925215959549, + 0.05451948568224907, + 0.04510442540049553, + 0.04340316727757454, + 0.20006684958934784, + 0.06782395392656326, + 0.05780385062098503, + 0.0431600883603096 + ], + "credit_map": [ + 0.04707745835185051, + 0.03860537335276604, + 0.06163044646382332, + 0.03455023840069771, + 0.04124246537685394, + 0.03701399639248848, + 0.03480483219027519, + 0.06140795722603798, + 0.008495720103383064, + 0.0077472166158258915, + 0.008537521585822105, + 0.0078035262413322926, + 0.007988722063601017, + 0.0078610023483634, + 0.011054006405174732, + 0.008818418718874454, + 0.0053674522787332535, + 0.010225954465568066, + 0.007737994194030762, + 0.007022390607744455, + 0.007316616829484701, + 0.006538521032780409, + 0.009847870096564293, + 0.009422450326383114, + 0.00479405652731657, + 0.006146048195660114, + 0.014737059362232685, + 0.005870115477591753, + 0.005482824984937906, + 0.004777914844453335, + 0.0049874065443873405, + 0.0077240606769919395, + 0.0039253998547792435, + 0.004529740661382675, + 0.009952619671821594, + 0.005035671405494213, + 0.0059455158188939095, + 0.003939984366297722, + 0.007269185967743397, + 0.004506309516727924, + 0.004842681344598532, + 0.0053494274616241455, + 0.007171051111072302, + 0.00384850287809968, + 0.005272652022540569, + 0.004648734349757433, + 0.005033655092120171, + 0.007236463017761707, + 0.01748139224946499, + 0.024787819012999535, + 0.03014403022825718, + 0.02170669287443161, + 0.030967773869633675, + 0.019000792875885963, + 0.032053329050540924, + 0.02392503060400486, + 0.007723307237029076, + 0.008309298194944859, + 0.010527967475354671, + 0.009165638126432896, + 0.005688524805009365, + 0.00812581367790699, + 0.009027132764458656, + 0.009256269782781601, + 0.006834982428699732, + 0.007654232904314995, + 0.011343792080879211, + 0.004800738766789436, + 0.005313417874276638, + 0.006631385535001755, + 0.006460187491029501, + 0.008765118196606636, + 0.004979834426194429, + 0.005621929652988911, + 0.006439718417823315, + 0.004146709106862545, + 0.004855998791754246, + 0.006860143039375544, + 0.004159256350249052, + 0.006096499972045422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.018171383067965508 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304181784391403 + ], + "opposite_logprob_tokens": [ + -3.168041706085205 + ], + "logprob_reward": -0.04304181784391403, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304181784391403, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23459099233150482, + 0.0674290806055069, + 0.030127178877592087, + 0.03902328386902809, + 0.0709943175315857, + 0.05931778997182846, + 0.27773919701576233, + 0.14452555775642395, + 0.026799622923135757, + 0.0494530089199543 + ], + "credit_map": [ + 0.028960855677723885, + 0.027077186852693558, + 0.032440800219774246, + 0.02784632332623005, + 0.023048769682645798, + 0.029396744444966316, + 0.026215827092528343, + 0.03960447385907173, + 0.006264066323637962, + 0.007674959022551775, + 0.008382954634726048, + 0.009641080163419247, + 0.007306381594389677, + 0.007523872423917055, + 0.01297449879348278, + 0.007661270443350077, + 0.0027874992229044437, + 0.003720135660842061, + 0.004063931759446859, + 0.004726187791675329, + 0.004103758372366428, + 0.0026378966867923737, + 0.004802324343472719, + 0.0032854455057531595, + 0.0035178104881197214, + 0.004960149526596069, + 0.009942596778273582, + 0.0032916944473981857, + 0.003993100952357054, + 0.0033621264155954123, + 0.0048043266870081425, + 0.005151478108018637, + 0.008299571461975574, + 0.008090031333267689, + 0.02080461010336876, + 0.006012625992298126, + 0.007302194368094206, + 0.005774164572358131, + 0.005475599784404039, + 0.009235517121851444, + 0.005592829547822475, + 0.006630278658121824, + 0.015968987718224525, + 0.00457890797406435, + 0.006001751869916916, + 0.006507520563900471, + 0.006141820922493935, + 0.007895690388977528, + 0.024262525141239166, + 0.034767527133226395, + 0.04392784833908081, + 0.029429728165268898, + 0.043121155351400375, + 0.025578467175364494, + 0.044955410063266754, + 0.03169653192162514, + 0.014740090817213058, + 0.019442345947027206, + 0.021221008151769638, + 0.01951202191412449, + 0.014122321270406246, + 0.014667099341750145, + 0.02010970376431942, + 0.020710967481136322, + 0.002882069442421198, + 0.003726294497027993, + 0.0037527407985180616, + 0.003130759112536907, + 0.0024319232907146215, + 0.003128646407276392, + 0.003963839262723923, + 0.003783351508900523, + 0.0037919997703284025, + 0.005736305378377438, + 0.006712400354444981, + 0.007376511115580797, + 0.0044157193042337894, + 0.005522756837308407, + 0.006428152788430452, + 0.009469161741435528 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.04304181784391403 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot captures the tense scene, rendered in a clean, graphic animation style against a deep blue void. Thin, sharp white lines radiate outward from an invisible point off-screen, creating a stark web-like pattern that defines the abstract space. In the center, several pale figures gather. The leftmost figure, frozen in mid-shouting, their mouths agape, their faces contorted in pain in silent despair, clad in a heavy, dark purple robe. Just behind them, a second figure in a somber gray robe places a steady hand on the shouting figure's shoulder, a gesture of comfort or restraint, their own expression serious and focused. The relative stillness of the central group contrasts sharply with the surrounding chaos, as blurred, shadowy outlines race through the foreground and background, their forms blurred and ghostly. The lighting is harsh and cool, casting sharp shadows that highlight the pale tones of their skin against the saturated colors of their purple, pink, and gray robes, creating a dark and tense atmosphere that freezes a moment of high-stakes drama and emotional crisis.", + "seed": 31536, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_0/index_2342_seed_31536_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010482885967940092, + "vlm_logprob": -0.010482885967940092, + "vlm_yes_no_margin": 5.55, + "vlm_token_credit": -0.010482885967940092, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008870244957506657 + ], + "opposite_logprob_tokens": [ + -8.383870124816895 + ], + "logprob_reward": -0.008870244957506657, + "yes_no_margin_reward": 8.375, + "token_credit_reward": -0.008870244957506657, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08038414269685745, + 0.056971754878759384, + 0.10470828413963318, + 0.02724464051425457, + 0.03101600706577301, + 0.0316937230527401, + 0.4767283797264099, + 0.02770618535578251, + 0.025639289990067482, + 0.13790760934352875 + ], + "credit_map": [ + 0.01819092035293579, + 0.006142971571534872, + 0.005185579881072044, + 0.0049284216947853565, + 0.005180544685572386, + 0.020291393622756004, + 0.0065969848074018955, + 0.01386732142418623, + 0.010612593032419682, + 0.006697430275380611, + 0.0045990957878530025, + 0.004891436547040939, + 0.005570626817643642, + 0.00789719633758068, + 0.0062781525775790215, + 0.010425223037600517, + 0.024336140602827072, + 0.0066182175651192665, + 0.00679883174598217, + 0.007243500556796789, + 0.00704006664454937, + 0.01438535563647747, + 0.013125068508088589, + 0.025161098688840866, + 0.005103639326989651, + 0.002481617033481598, + 0.002417762763798237, + 0.0023711565881967545, + 0.002848450094461441, + 0.0031746751628816128, + 0.004273918457329273, + 0.004573421087116003, + 0.009314723312854767, + 0.002091056667268276, + 0.002406671643257141, + 0.002478131325915456, + 0.0021545628551393747, + 0.004599996376782656, + 0.0038669705390930176, + 0.004103897139430046, + 0.007322036661207676, + 0.002247961238026619, + 0.0029290500096976757, + 0.0031435396522283554, + 0.0024468281771987677, + 0.004675508011132479, + 0.004379838705062866, + 0.004548960365355015, + 0.16807015240192413, + 0.017342017963528633, + 0.02001393772661686, + 0.014967757277190685, + 0.020641617476940155, + 0.14802362024784088, + 0.015955086797475815, + 0.07171420007944107, + 0.008804578334093094, + 0.0017422157106921077, + 0.0021302516106516123, + 0.0017688384978100657, + 0.0013535083271563053, + 0.005209519062191248, + 0.0022782543674111366, + 0.004419018514454365, + 0.004488190170377493, + 0.0021325049456208944, + 0.0028879677411168814, + 0.0022265820298343897, + 0.0021621407940983772, + 0.00409200694411993, + 0.002998775104060769, + 0.004651122260838747, + 0.03664623573422432, + 0.005140976049005985, + 0.007168269250541925, + 0.006491972599178553, + 0.007268501911312342, + 0.020797057077288628, + 0.009200961329042912, + 0.04519364610314369 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 2 + ], + "selected_reward": -0.008870244957506657 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004108085297048092 + ], + "opposite_logprob_tokens": [ + -5.50410795211792 + ], + "logprob_reward": -0.004108085297048092, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004108085297048092, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1399616301059723, + 0.09156667441129684, + 0.11335346102714539, + 0.07761447876691818, + 0.12319675832986832, + 0.14187610149383545, + 0.20589374005794525, + 0.03481690213084221, + 0.040043678134679794, + 0.031676579266786575 + ], + "credit_map": [ + 0.046895161271095276, + 0.009876820258796215, + 0.007483203895390034, + 0.007119440473616123, + 0.00572296604514122, + 0.02158610336482525, + 0.007796262390911579, + 0.033481668680906296, + 0.017095353454351425, + 0.005771471653133631, + 0.006481145042926073, + 0.007843957282602787, + 0.005400192923843861, + 0.023539742454886436, + 0.009160795249044895, + 0.016274018213152885, + 0.02276669256389141, + 0.010829457081854343, + 0.009406940080225468, + 0.009988063015043736, + 0.010896609164774418, + 0.014835763722658157, + 0.01449784729629755, + 0.02013207972049713, + 0.013290004804730415, + 0.005847149062901735, + 0.007064124569296837, + 0.006342281121760607, + 0.006530065555125475, + 0.011185098439455032, + 0.00822081696242094, + 0.019134938716888428, + 0.04159446060657501, + 0.006488759070634842, + 0.011293698102235794, + 0.007190662436187267, + 0.0070725115947425365, + 0.020625343546271324, + 0.009050348773598671, + 0.019880961626768112, + 0.039696983993053436, + 0.00811982061713934, + 0.012785970233380795, + 0.012019461020827293, + 0.009738363325595856, + 0.024487916380167007, + 0.012911247089505196, + 0.022116338834166527, + 0.0733228325843811, + 0.007504941895604134, + 0.009414145722985268, + 0.006428680382668972, + 0.00857065711170435, + 0.06426612287759781, + 0.006690958049148321, + 0.029695387929677963, + 0.00914055947214365, + 0.002305302768945694, + 0.002532209735363722, + 0.003065553493797779, + 0.0021916169207543135, + 0.0052441698499023914, + 0.0031303667929023504, + 0.007207121700048447, + 0.005777130834758282, + 0.003606777638196945, + 0.0049176630564033985, + 0.003835991257801652, + 0.0030209675896912813, + 0.004286669194698334, + 0.006389823742210865, + 0.008208648301661015, + 0.006613299250602722, + 0.0021203698124736547, + 0.0022805999033153057, + 0.0023371498100459576, + 0.0016279934206977487, + 0.004678943194448948, + 0.0021510215010493994, + 0.009867199696600437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.004108085297048092 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028190179727971554 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.0028190179727971554, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028190179727971554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14815934002399445, + 0.11351627111434937, + 0.07527738064527512, + 0.03978552669286728, + 0.04846537485718727, + 0.03678826987743378, + 0.2791357636451721, + 0.13051274418830872, + 0.04025755822658539, + 0.08810168504714966 + ], + "credit_map": [ + 0.02483268827199936, + 0.013124153017997742, + 0.011356515809893608, + 0.014131447300314903, + 0.012360911816358566, + 0.022145532071590424, + 0.015529049560427666, + 0.034679073840379715, + 0.014387274160981178, + 0.014506934210658073, + 0.011976491659879684, + 0.01462865062057972, + 0.013115419074892998, + 0.014442910440266132, + 0.013187097385525703, + 0.017271509394049644, + 0.01084216870367527, + 0.005742034409195185, + 0.009666097350418568, + 0.008111375384032726, + 0.007289370987564325, + 0.008715753443539143, + 0.015269008465111256, + 0.009641585871577263, + 0.006927209906280041, + 0.0026212562806904316, + 0.003735243808478117, + 0.003827403299510479, + 0.003438333747908473, + 0.005659646820276976, + 0.006389739457517862, + 0.007186697330325842, + 0.007758104242384434, + 0.00401688925921917, + 0.004763535689562559, + 0.005189345683902502, + 0.004802687093615532, + 0.005959552712738514, + 0.006088734604418278, + 0.00988653302192688, + 0.005106091499328613, + 0.004193125758320093, + 0.003415942657738924, + 0.005319900345057249, + 0.0038869797717779875, + 0.0045385658740997314, + 0.004913612734526396, + 0.005414056591689587, + 0.09444881975650787, + 0.012227116152644157, + 0.011459038592875004, + 0.009671139530837536, + 0.012553016655147076, + 0.07076429575681686, + 0.013246912509202957, + 0.05476543679833412, + 0.036751795560121536, + 0.00989453960210085, + 0.01111908070743084, + 0.013025320135056973, + 0.00915308017283678, + 0.021884039044380188, + 0.009917885065078735, + 0.018767017871141434, + 0.005865901242941618, + 0.0037451323587447405, + 0.006463003344833851, + 0.004263199865818024, + 0.0030845210421830416, + 0.004956195596605539, + 0.004055659752339125, + 0.007823950611054897, + 0.019821494817733765, + 0.005787150468677282, + 0.006020415108650923, + 0.0059746927581727505, + 0.005992779042571783, + 0.01589941792190075, + 0.006431925110518932, + 0.022173820063471794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.0028190179727971554 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016051357612013817 + ], + "opposite_logprob_tokens": [ + -4.141051292419434 + ], + "logprob_reward": -0.016051357612013817, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016051357612013817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08639666438102722, + 0.11632655560970306, + 0.2928628921508789, + 0.06675883382558823, + 0.08976907283067703, + 0.056432828307151794, + 0.16037243604660034, + 0.044036444276571274, + 0.03070124424993992, + 0.05634300410747528 + ], + "credit_map": [ + 0.013755615800619125, + 0.012934313155710697, + 0.008451101370155811, + 0.008521487936377525, + 0.007487812079489231, + 0.011368434876203537, + 0.007489350624382496, + 0.0163885485380888, + 0.021615896373987198, + 0.00839077215641737, + 0.009490993805229664, + 0.009468341246247292, + 0.006672592367976904, + 0.02657981775701046, + 0.011358559131622314, + 0.02274959348142147, + 0.062297556549310684, + 0.02171638421714306, + 0.02097894623875618, + 0.020213967189192772, + 0.021713757887482643, + 0.03916541114449501, + 0.031838126480579376, + 0.07493872940540314, + 0.007979104295372963, + 0.0053175692446529865, + 0.009234007447957993, + 0.008056296966969967, + 0.007040882017463446, + 0.00787715706974268, + 0.01162708830088377, + 0.009626727551221848, + 0.013535921461880207, + 0.007001848891377449, + 0.008847985416650772, + 0.010526883415877819, + 0.010245361365377903, + 0.010710155591368675, + 0.016632862389087677, + 0.01226805243641138, + 0.0052329618483781815, + 0.006595894694328308, + 0.005749414209276438, + 0.007359260693192482, + 0.006723000202327967, + 0.00831921398639679, + 0.009393722750246525, + 0.007059359923005104, + 0.03623304143548012, + 0.009176509454846382, + 0.009556778706610203, + 0.0068814936093986034, + 0.011919654905796051, + 0.05655575916171074, + 0.012230902910232544, + 0.017818300053477287, + 0.01281006634235382, + 0.003045432036742568, + 0.003198608523234725, + 0.002859503496438265, + 0.002719088224694133, + 0.00853104516863823, + 0.003139874432235956, + 0.007732827682048082, + 0.00405027624219656, + 0.003971579950302839, + 0.0027220984920859337, + 0.00311355316080153, + 0.0026862546801567078, + 0.0033908032346516848, + 0.004722828045487404, + 0.006043850909918547, + 0.010249366983771324, + 0.005039567593485117, + 0.005045799538493156, + 0.004715284798294306, + 0.004798104055225849, + 0.009833527728915215, + 0.006327630952000618, + 0.010333719663321972 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.016051357612013817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video rendered in a clean, graphic animation style against a deep blue void background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are thin, sharp white lines radiating outward from an off-screen point to create a web-like pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is there a central group of pale figures, including one in a dark purple robe shouting with an agape mouth and contorted face?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a second figure in a somber gray robe visible behind the shouting figure, placing a hand on their shoulder?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are blurred, shadowy outlines racing through the foreground and background while the central group remains relatively still?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02056572400033474 + ], + "opposite_logprob_tokens": [ + -3.8955657482147217 + ], + "logprob_reward": -0.02056572400033474, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02056572400033474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11589960753917694, + 0.19163891673088074, + 0.18415507674217224, + 0.029532700777053833, + 0.037403520196676254, + 0.05563898757100105, + 0.25380539894104004, + 0.06419190019369125, + 0.030971454456448555, + 0.036762434989213943 + ], + "credit_map": [ + 0.023403894156217575, + 0.013326226733624935, + 0.008612392470240593, + 0.007940957322716713, + 0.00904135312885046, + 0.01587940752506256, + 0.011719380505383015, + 0.025975998491048813, + 0.03668905794620514, + 0.013153789564967155, + 0.01260713767260313, + 0.014355653896927834, + 0.013258056715130806, + 0.04129908233880997, + 0.023039326071739197, + 0.037236813455820084, + 0.040734488517045975, + 0.012792563997209072, + 0.012629996053874493, + 0.012213950045406818, + 0.013023658655583858, + 0.0242946557700634, + 0.02059318497776985, + 0.047872573137283325, + 0.005804017186164856, + 0.0017580022104084492, + 0.003294826252385974, + 0.002013548742979765, + 0.002238775137811899, + 0.0037975581362843513, + 0.003617956768721342, + 0.007008016109466553, + 0.007578387390822172, + 0.0026370531413704157, + 0.0034988585393875837, + 0.0036903114523738623, + 0.0036648327950388193, + 0.00454154796898365, + 0.005979336332529783, + 0.005813188850879669, + 0.010399922728538513, + 0.004562655929476023, + 0.006336837075650692, + 0.005331987049430609, + 0.00541006401181221, + 0.007218878250569105, + 0.006298222113400698, + 0.010080419480800629, + 0.07231248915195465, + 0.011423708871006966, + 0.013153726235032082, + 0.009969095699489117, + 0.016940921545028687, + 0.08381572365760803, + 0.014505285769701004, + 0.03168441727757454, + 0.018298259004950523, + 0.0028355142567306757, + 0.0037533424329012632, + 0.00433885445818305, + 0.0028645172715187073, + 0.018685607239603996, + 0.0038249697536230087, + 0.009590832516551018, + 0.005075078457593918, + 0.002008878393098712, + 0.00334717589430511, + 0.002809592755511403, + 0.0022313292138278484, + 0.0058279260993003845, + 0.003720009233802557, + 0.005951463710516691, + 0.007263633888214827, + 0.0027156691066920757, + 0.0028833067044615746, + 0.0025459870230406523, + 0.00320737948641181, + 0.008397186174988747, + 0.003319720970466733, + 0.006429553963243961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 2 + ], + "selected_reward": -0.02056572400033474 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_tensor([6943, 6943])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_tensor([6943, 6943])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..316aaa77702b14dce8d6e9a3ccbbed88185158da --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_tensor([6943, 6943])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a whimsical, surreal cartoon style, an uninterrupted, gliding shot smoothly moves through a dreamlike seed laboratory, where thick frost rapidly crystallizes on every surface. The camera tracks rows of glowing glass vials, placed on ornate brass racks atop polished mahogany workbenches. Within each crystalline vial, a tiny, dormant sapling, its delicately spiraling leaves emitting a soft, warm, golden bioluminescence. The dominant effect is the relentless and rapid bloom of frost; intricate, shimmering ice patterns, like irregular ferns, visibly grow and spread across the glass surface, delicately enveloping the vials. This crystallization is not a harsh freeze, but a magical transformation. The newly formed ice captures and refracts the sapling's inner light, casting a dazzling prismatic shimmer throughout the laboratory. The camera's steady, gliding motion creates a sense of wonder and intimacy, allowing the viewer to witness a silent, enchanting process as the entire dreamlike environment is taken over by a breathtaking, frozen beauty.", + "seed": 28343, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28343_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.019820256251841785, + "vlm_logprob": -0.019820256251841785, + "vlm_yes_no_margin": 4.124999952316284, + "vlm_token_credit": -0.019820256251841785, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03832460194826126 + ], + "opposite_logprob_tokens": [ + -3.4133245944976807 + ], + "logprob_reward": -0.03832460194826126, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03832460194826126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1581311970949173, + 0.07441595941781998, + 0.18139414489269257, + 0.02806480973958969, + 0.09613469988107681, + 0.03917081654071808, + 0.07928340137004852, + 0.1286940723657608, + 0.08422943949699402, + 0.13048142194747925 + ], + "credit_map": [ + 0.03240996226668358, + 0.015508687123656273, + 0.019838323816657066, + 0.01470229309052229, + 0.02310832589864731, + 0.028757549822330475, + 0.01048905961215496, + 0.013317003846168518, + 0.018030891194939613, + 0.016936803236603737, + 0.007089074235409498, + 0.005316214170306921, + 0.008882268331944942, + 0.010290443897247314, + 0.0033630637917667627, + 0.004507202655076981, + 0.04848787933588028, + 0.0640394538640976, + 0.013783281669020653, + 0.014202570542693138, + 0.01316472701728344, + 0.013442142866551876, + 0.008223003707826138, + 0.0060510761104524136, + 0.0042439913377165794, + 0.005094877444207668, + 0.003861323930323124, + 0.0026711008977144957, + 0.004659568425267935, + 0.0028223958797752857, + 0.0019333930686116219, + 0.0027781580574810505, + 0.017702091485261917, + 0.012550861574709415, + 0.01240016333758831, + 0.006783307064324617, + 0.023333590477705002, + 0.011062687262892723, + 0.006935846526175737, + 0.005366145633161068, + 0.00534719368442893, + 0.007892411202192307, + 0.00449106702581048, + 0.0030204588547348976, + 0.006147528998553753, + 0.004757553804665804, + 0.002796594984829426, + 0.004718010779470205, + 0.006446299143135548, + 0.005870260298252106, + 0.009250114671885967, + 0.007848745211958885, + 0.013712536543607712, + 0.010997042991220951, + 0.016313491389155388, + 0.008844919502735138, + 0.008683591149747372, + 0.024419961497187614, + 0.010355901904404163, + 0.01344219408929348, + 0.029863324016332626, + 0.02274286188185215, + 0.011714071035385132, + 0.007472170516848564, + 0.016840949654579163, + 0.010445043444633484, + 0.007189265452325344, + 0.008298203349113464, + 0.014552079141139984, + 0.016426019370555878, + 0.00588650070130825, + 0.004591384902596474, + 0.01661544479429722, + 0.017678936943411827, + 0.0124845951795578, + 0.015427544713020325, + 0.02187231369316578, + 0.02013499289751053, + 0.014457833021879196, + 0.011809759773314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.03832460194826126 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00864796806126833 + ], + "opposite_logprob_tokens": [ + -4.758647918701172 + ], + "logprob_reward": -0.00864796806126833, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00864796806126833, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1800736039876938, + 0.1700361669063568, + 0.16760623455047607, + 0.029287371784448624, + 0.1227857694029808, + 0.035037562251091, + 0.03937460854649544, + 0.10588892549276352, + 0.055410273373126984, + 0.09449951350688934 + ], + "credit_map": [ + 0.04543958976864815, + 0.02030974254012108, + 0.013022453524172306, + 0.01867394521832466, + 0.028314055874943733, + 0.027805663645267487, + 0.013081902638077736, + 0.013426261022686958, + 0.05851755663752556, + 0.0626564472913742, + 0.007486067712306976, + 0.0044934735633432865, + 0.018653985112905502, + 0.0075329444371163845, + 0.0055886623449623585, + 0.005107043776661158, + 0.04631498083472252, + 0.06002696603536606, + 0.010472264140844345, + 0.013063106685876846, + 0.011672784574329853, + 0.011710692197084427, + 0.007804286666214466, + 0.0065411534160375595, + 0.003133955178782344, + 0.005153201520442963, + 0.004729617387056351, + 0.0033756576012820005, + 0.004765411838889122, + 0.0033728652633726597, + 0.0023096606601029634, + 0.0024470023345202208, + 0.011319361627101898, + 0.01030572596937418, + 0.03614131733775139, + 0.01151227205991745, + 0.009190656244754791, + 0.03345963731408119, + 0.006480122450739145, + 0.004376681987196207, + 0.0048539587296545506, + 0.006597054190933704, + 0.004518056288361549, + 0.002596610691398382, + 0.007006992120295763, + 0.0037959059700369835, + 0.0028517195023596287, + 0.0028172628954052925, + 0.0026853238232433796, + 0.0032683941535651684, + 0.0032365128863602877, + 0.003050982253625989, + 0.011281589046120644, + 0.007662655785679817, + 0.005330380517989397, + 0.0028587684500962496, + 0.00962875783443451, + 0.014952016994357109, + 0.012381951324641705, + 0.013139141723513603, + 0.01696508750319481, + 0.019360342994332314, + 0.009419618174433708, + 0.010042010806500912, + 0.012644181959331036, + 0.007826676592230797, + 0.003806316526606679, + 0.0034602300729602575, + 0.00970963854342699, + 0.010120573453605175, + 0.004012236837297678, + 0.003830422880128026, + 0.012071721255779266, + 0.012178346514701843, + 0.010285433381795883, + 0.01468783337622881, + 0.011353750713169575, + 0.012338333763182163, + 0.010418691672384739, + 0.011165405623614788 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.00864796806126833 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076268198899924755 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.0076268198899924755, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076268198899924755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14868362247943878, + 0.1540335863828659, + 0.23034293949604034, + 0.06508126109838486, + 0.08532299846410751, + 0.027887385338544846, + 0.043764229863882065, + 0.07635284960269928, + 0.08619411289691925, + 0.08233706653118134 + ], + "credit_map": [ + 0.02882615476846695, + 0.018254337832331657, + 0.018342018127441406, + 0.020542502403259277, + 0.020535068586468697, + 0.01926000602543354, + 0.012592386454343796, + 0.010331123135983944, + 0.04162724316120148, + 0.04572230949997902, + 0.011510446667671204, + 0.007000905927270651, + 0.018929988145828247, + 0.012873451225459576, + 0.0077974568121135235, + 0.008571765385568142, + 0.06479988247156143, + 0.08399710804224014, + 0.015885762870311737, + 0.017409129068255424, + 0.01596168987452984, + 0.015633584931492805, + 0.00958779826760292, + 0.007067972328513861, + 0.008239350281655788, + 0.011320799589157104, + 0.011039054952561855, + 0.007842911407351494, + 0.011413498781621456, + 0.007303201127797365, + 0.003984265960752964, + 0.003938176669180393, + 0.011244595050811768, + 0.009259562008082867, + 0.018411127850413322, + 0.006524479482322931, + 0.012823362834751606, + 0.018041914328932762, + 0.004680335987359285, + 0.004337603226304054, + 0.0033777731005102396, + 0.006785848643630743, + 0.0039223115891218185, + 0.0025504950899630785, + 0.0038657390978187323, + 0.0025356158148497343, + 0.0017405339749529958, + 0.0031090632546693087, + 0.0031833176035434008, + 0.0038893052842468023, + 0.004563986323773861, + 0.004064903594553471, + 0.011003075167536736, + 0.0069631836377084255, + 0.006817847490310669, + 0.003278604242950678, + 0.007083039730787277, + 0.013248570263385773, + 0.005689837038516998, + 0.0076451911590993404, + 0.017570529133081436, + 0.013997279107570648, + 0.0062704067677259445, + 0.0048479847609996796, + 0.020709654316306114, + 0.01037542149424553, + 0.005968471523374319, + 0.005840555299073458, + 0.015083380974829197, + 0.015554017387330532, + 0.00645688408985734, + 0.006205716170370579, + 0.009709404781460762, + 0.013073031790554523, + 0.011559131555259228, + 0.010510497726500034, + 0.011829161085188389, + 0.010468762367963791, + 0.008675623685121536, + 0.006511440500617027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0076268198899924755 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633226290345192 + ], + "opposite_logprob_tokens": [ + -3.651332378387451 + ], + "logprob_reward": -0.02633226290345192, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633226290345192, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04707765206694603, + 0.19147585332393646, + 0.040116868913173676, + 0.02737228013575077, + 0.036786288022994995, + 0.012708857655525208, + 0.1941465586423874, + 0.09541600197553635, + 0.056310977786779404, + 0.29858869314193726 + ], + "credit_map": [ + 0.006761464290320873, + 0.008591532707214355, + 0.005600674543529749, + 0.004875728394836187, + 0.008291436359286308, + 0.005000601056963205, + 0.00429625203832984, + 0.003659955458715558, + 0.05682602897286415, + 0.06050417199730873, + 0.012316294014453888, + 0.008335488848388195, + 0.020152637735009193, + 0.014767342247068882, + 0.00974457710981369, + 0.008829288184642792, + 0.005568185821175575, + 0.005092211998999119, + 0.00533418171107769, + 0.00567492563277483, + 0.005280538462102413, + 0.0052825333550572395, + 0.0036745541729032993, + 0.004209733568131924, + 0.0032332241535186768, + 0.006849854718893766, + 0.0037273734342306852, + 0.002261305693536997, + 0.005952152889221907, + 0.0024371081963181496, + 0.001496854703873396, + 0.0014144014567136765, + 0.003724179696291685, + 0.003228616900742054, + 0.009899556636810303, + 0.0035270233638584614, + 0.004421067889779806, + 0.007976903580129147, + 0.0024863756261765957, + 0.001522559323348105, + 0.001856982591561973, + 0.001841343124397099, + 0.0017071740003302693, + 0.001106582349166274, + 0.0020941440016031265, + 0.0012360189575701952, + 0.0010369218653067946, + 0.001829689834266901, + 0.016841603443026543, + 0.015227794647216797, + 0.019245246425271034, + 0.01652050018310547, + 0.048405203968286514, + 0.03753475472331047, + 0.025114523246884346, + 0.015256918035447598, + 0.008267288096249104, + 0.015540050342679024, + 0.006256910040974617, + 0.009005263447761536, + 0.02299189195036888, + 0.0181846022605896, + 0.008394688367843628, + 0.006775295827537775, + 0.008652588352560997, + 0.0075179762206971645, + 0.003544832346960902, + 0.0042044189758598804, + 0.008250052109360695, + 0.01708257757127285, + 0.0040479665622115135, + 0.0030105619225651026, + 0.037762001156806946, + 0.043666210025548935, + 0.04436776787042618, + 0.031308479607105255, + 0.036593686789274216, + 0.05003517493605614, + 0.022683434188365936, + 0.03217189013957977 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 1 + ], + "selected_reward": -0.02633226290345192 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018169628456234932 + ], + "opposite_logprob_tokens": [ + -4.018169403076172 + ], + "logprob_reward": -0.018169628456234932, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018169628456234932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.075078584253788, + 0.11592551320791245, + 0.04582798480987549, + 0.04790736362338066, + 0.07924071699380875, + 0.08239862322807312, + 0.17889811098575592, + 0.14540547132492065, + 0.043991126120090485, + 0.18532653152942657 + ], + "credit_map": [ + 0.012209546752274036, + 0.00817889254540205, + 0.008785207755863667, + 0.010771067813038826, + 0.009272860363125801, + 0.011103305965662003, + 0.008756411261856556, + 0.006001288536936045, + 0.023322438821196556, + 0.024524133652448654, + 0.011897560209035873, + 0.00933041237294674, + 0.013421989046037197, + 0.01492504496127367, + 0.009871323592960835, + 0.008632602170109749, + 0.009990392252802849, + 0.012528571300208569, + 0.004854846745729446, + 0.005264642182737589, + 0.003963559865951538, + 0.00338940997608006, + 0.0029571435879915953, + 0.0028794212266802788, + 0.004949498455971479, + 0.008689161390066147, + 0.007287608925253153, + 0.005263091996312141, + 0.009208731353282928, + 0.005477544851601124, + 0.003665274241939187, + 0.003366451943293214, + 0.007075055968016386, + 0.008048112504184246, + 0.02149277925491333, + 0.006724522449076176, + 0.006410321686416864, + 0.021522056311368942, + 0.004161270335316658, + 0.003806592198088765, + 0.011597332544624805, + 0.0162513367831707, + 0.010022871196269989, + 0.006513575091958046, + 0.015894224867224693, + 0.00926545262336731, + 0.0062028756365180016, + 0.006650950293987989, + 0.012797845527529716, + 0.013905901461839676, + 0.011317943222820759, + 0.011370898224413395, + 0.04877876490354538, + 0.034649889916181564, + 0.03155538812279701, + 0.014521468430757523, + 0.013653740286827087, + 0.017951607704162598, + 0.020425597205758095, + 0.021772613748908043, + 0.015311707742512226, + 0.024799766018986702, + 0.015502918511629105, + 0.01598752848803997, + 0.00624749343842268, + 0.007976287975907326, + 0.003920263610780239, + 0.004581904038786888, + 0.007101746741682291, + 0.00862806849181652, + 0.0033946239855140448, + 0.00214073876850307, + 0.02338697575032711, + 0.032602328807115555, + 0.021282022818922997, + 0.021705031394958496, + 0.02252049185335636, + 0.027491925284266472, + 0.016040628775954247, + 0.020297124981880188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.018169628456234932 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a whimsical, surreal cartoon style, an uninterrupted, gliding shot smoothly moves through a dreamlike seed laboratory, where thick frost rapidly crystallizes on every surface. The camera tracks rows of glowing glass vials, placed on ornate brass racks atop polished mahogany workbenches. Within each crystalline vial, a tiny, dormant sapling, its delicately spiraling leaves emitting a soft, warm, golden bioluminescence. The dominant effect is the relentless and rapid bloom of frost; intricate, shimmering ice patterns, like irregular ferns, visibly grow and spread across the glass surface, delicately enveloping the vials. This crystallization is not a harsh freeze, but a magical transformation. The newly formed ice captures and refracts the sapling's inner light, casting a dazzling prismatic shimmer throughout the laboratory. The camera's steady, gliding motion creates a sense of wonder and intimacy, allowing the viewer to witness a silent, enchanting process as the entire dreamlike environment is taken over by a breathtaking, frozen beauty.", + "seed": 28344, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28344_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06260341070592404, + "vlm_logprob": -0.06260341070592404, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.06260341070592404, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14484812319278717 + ], + "opposite_logprob_tokens": [ + -2.019848108291626 + ], + "logprob_reward": -0.14484812319278717, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14484812319278717, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07322977483272552, + 0.11872558295726776, + 0.05762884393334389, + 0.1293206363916397, + 0.07376211136579514, + 0.10409357398748398, + 0.280689537525177, + 0.0472746267914772, + 0.07561224699020386, + 0.03966306522488594 + ], + "credit_map": [ + 0.008038242347538471, + 0.0182154830545187, + 0.0033697178587317467, + 0.004113398492336273, + 0.013236696831882, + 0.009421325288712978, + 0.009275952354073524, + 0.007558960933238268, + 0.02074548788368702, + 0.011307712644338608, + 0.004679702688008547, + 0.012860878370702267, + 0.019470805302262306, + 0.013763036578893661, + 0.019931938499212265, + 0.01596602238714695, + 0.005389702506363392, + 0.005212198942899704, + 0.0034347984474152327, + 0.009248876012861729, + 0.011451602913439274, + 0.005173900164663792, + 0.012511122971773148, + 0.005206642672419548, + 0.007329822052270174, + 0.007905548438429832, + 0.014125331304967403, + 0.04353673383593559, + 0.023978278040885925, + 0.006102483719587326, + 0.009860018268227577, + 0.01648242585361004, + 0.006699929479509592, + 0.009158427827060223, + 0.008652073331177235, + 0.012199144810438156, + 0.01243650633841753, + 0.004570588003844023, + 0.008804736658930779, + 0.01124070305377245, + 0.00743136927485466, + 0.00811599101871252, + 0.009306642226874828, + 0.007070331368595362, + 0.02733139880001545, + 0.015554282814264297, + 0.014844199642539024, + 0.014439361169934273, + 0.01597866788506508, + 0.09621453285217285, + 0.011311860755085945, + 0.009041826240718365, + 0.08536869287490845, + 0.019393589347600937, + 0.023126766085624695, + 0.020253630355000496, + 0.005164142232388258, + 0.006176371593028307, + 0.005744777619838715, + 0.005427198018878698, + 0.006102865096181631, + 0.00508116465061903, + 0.00611122278496623, + 0.007466884329915047, + 0.010194345377385616, + 0.011108268983662128, + 0.006711728870868683, + 0.01260341051965952, + 0.005677393637597561, + 0.010079956613481045, + 0.010430360212922096, + 0.008806778118014336, + 0.003574446542188525, + 0.005827102344483137, + 0.0038126343861222267, + 0.0036753995809704065, + 0.004318567458540201, + 0.0059156715869903564, + 0.003993873950093985, + 0.008545367047190666 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 1 + ], + "selected_reward": -0.14484812319278717 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898624986410141 + ], + "opposite_logprob_tokens": [ + -2.578986167907715 + ], + "logprob_reward": -0.07898624986410141, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898624986410141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09419634938240051, + 0.10312531888484955, + 0.07910899817943573, + 0.4307447075843811, + 0.04234044998884201, + 0.036959417164325714, + 0.06932411342859268, + 0.037004780024290085, + 0.04974835738539696, + 0.05744750425219536 + ], + "credit_map": [ + 0.00960245169699192, + 0.024934817105531693, + 0.007236822508275509, + 0.009203034453094006, + 0.011017756536602974, + 0.009434144012629986, + 0.013171358034014702, + 0.009595959447324276, + 0.013443500734865665, + 0.010062383487820625, + 0.006820926908403635, + 0.017161639407277107, + 0.014630763791501522, + 0.010075309313833714, + 0.01566873863339424, + 0.015262055210769176, + 0.008104955777525902, + 0.00848266389220953, + 0.006151396781206131, + 0.011148713529109955, + 0.019973108544945717, + 0.007082943804562092, + 0.01354860421270132, + 0.004616620484739542, + 0.028497867286205292, + 0.034013669937849045, + 0.049124084413051605, + 0.1253443956375122, + 0.11624785512685776, + 0.024828586727380753, + 0.026245398446917534, + 0.026442890986800194, + 0.0034220304805785418, + 0.00535403611138463, + 0.004533085506409407, + 0.009108607657253742, + 0.005367634352296591, + 0.002905156696215272, + 0.005299083422869444, + 0.006350815761834383, + 0.0041304766200482845, + 0.006452761124819517, + 0.004379861056804657, + 0.0046533867716789246, + 0.0031726956367492676, + 0.0059405784122645855, + 0.003989295102655888, + 0.004240361973643303, + 0.005592275410890579, + 0.02202126942574978, + 0.003615233115851879, + 0.002770774532109499, + 0.018153177574276924, + 0.006004359573125839, + 0.006275375839322805, + 0.004891646094620228, + 0.004633055068552494, + 0.005223126150667667, + 0.0034391351509839296, + 0.003583886194974184, + 0.005309512373059988, + 0.005314055364578962, + 0.004783856216818094, + 0.004718156065791845, + 0.00991661474108696, + 0.009022469632327557, + 0.003099133726209402, + 0.006159885320812464, + 0.004379605408757925, + 0.004715904127806425, + 0.006073793862015009, + 0.00638095336034894, + 0.0033010307233780622, + 0.010005980730056763, + 0.006224900484085083, + 0.005708103068172932, + 0.005033121444284916, + 0.008618134073913097, + 0.004733479116111994, + 0.013822756707668304 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.07898624986410141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187422305345535 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187422305345535, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187422305345535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03405286371707916, + 0.0652947798371315, + 0.15521317720413208, + 0.347446084022522, + 0.059501130133867264, + 0.04371106997132301, + 0.11813103407621384, + 0.028041629120707512, + 0.041180577129125595, + 0.107427679002285 + ], + "credit_map": [ + 0.002896403893828392, + 0.005873494781553745, + 0.0032560578547418118, + 0.004313760902732611, + 0.0052750324830412865, + 0.00401539308950305, + 0.004592115990817547, + 0.003830607980489731, + 0.008763070218265057, + 0.007273451425135136, + 0.003826156957075, + 0.006333108060061932, + 0.015102054923772812, + 0.009026921354234219, + 0.008303294889628887, + 0.00666672270745039, + 0.012650517746806145, + 0.010879839770495892, + 0.004568731412291527, + 0.017414292320609093, + 0.0524735264480114, + 0.01253802515566349, + 0.039229076355695724, + 0.005459175910800695, + 0.022299617528915405, + 0.02769804373383522, + 0.038246408104896545, + 0.10216544568538666, + 0.09331940114498138, + 0.019658664241433144, + 0.020120009779930115, + 0.023938477039337158, + 0.0040992447175085545, + 0.005508306436240673, + 0.005379613488912582, + 0.01134317647665739, + 0.005887189414352179, + 0.009364141151309013, + 0.0054014455527067184, + 0.012518014758825302, + 0.003666863776743412, + 0.004771796986460686, + 0.004537761211395264, + 0.003910217899829149, + 0.008784296922385693, + 0.006392067763954401, + 0.005941139068454504, + 0.005706922151148319, + 0.008594575338065624, + 0.038327667862176895, + 0.004915466997772455, + 0.004164527170360088, + 0.03428031876683235, + 0.008721889927983284, + 0.010004882700741291, + 0.009121712297201157, + 0.0034751840867102146, + 0.00364307826384902, + 0.0042757997289299965, + 0.002549894154071808, + 0.0035654299426823854, + 0.002897857688367367, + 0.003436044091358781, + 0.004198341630399227, + 0.005772107280790806, + 0.006622565910220146, + 0.003183161374181509, + 0.0050078146159648895, + 0.005820457357913256, + 0.004044931847602129, + 0.005594453774392605, + 0.005135086365044117, + 0.007006495259702206, + 0.019602447748184204, + 0.014034559950232506, + 0.00880154874175787, + 0.0074202208779752254, + 0.016022374853491783, + 0.007433204911649227, + 0.027106832712888718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.018187422305345535 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941318929195404 + ], + "opposite_logprob_tokens": [ + -2.929941415786743 + ], + "logprob_reward": -0.054941318929195404, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941318929195404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.010083884000778198, + 0.024375488981604576, + 0.03815614804625511, + 0.10102462023496628, + 0.024026552215218544, + 0.045646410435438156, + 0.6676840782165527, + 0.009976615197956562, + 0.04714784026145935, + 0.03187841922044754 + ], + "credit_map": [ + 0.0011045816354453564, + 0.0013493165606632829, + 0.0007805574568919837, + 0.0009415536187589169, + 0.0018883153097704053, + 0.0013734808890148997, + 0.001448041177354753, + 0.001198036246933043, + 0.0029986160807311535, + 0.003912810701876879, + 0.0012144438223913312, + 0.003591646673157811, + 0.004284515976905823, + 0.003095827763900161, + 0.0023935753852128983, + 0.0028840522281825542, + 0.005611399654299021, + 0.003433654084801674, + 0.0012595639564096928, + 0.003940911032259464, + 0.011526418849825859, + 0.00490894028916955, + 0.004595354665070772, + 0.0028799015562981367, + 0.006159904412925243, + 0.007507519796490669, + 0.010826448909938335, + 0.031454384326934814, + 0.024701932445168495, + 0.005363142117857933, + 0.0058058397844433784, + 0.00920544657856226, + 0.0019296881509944797, + 0.002285270020365715, + 0.0029995276127010584, + 0.005105468910187483, + 0.002189111430197954, + 0.003252876689657569, + 0.002032136544585228, + 0.004232470411807299, + 0.004592519719153643, + 0.004492546431720257, + 0.0038798232562839985, + 0.004778493195772171, + 0.004634151700884104, + 0.012978981249034405, + 0.0041998689994215965, + 0.006090021226555109, + 0.031826604157686234, + 0.2454632669687271, + 0.017581799998879433, + 0.020350908860564232, + 0.20946285128593445, + 0.04098927974700928, + 0.05519728362560272, + 0.04681199789047241, + 0.001270641922019422, + 0.001522694481536746, + 0.0012712954776361585, + 0.0011649371590465307, + 0.0013798174913972616, + 0.0007377863512374461, + 0.0014334690058603883, + 0.0011959715047851205, + 0.004861472640186548, + 0.00785642396658659, + 0.003263686550781131, + 0.008412068709731102, + 0.006886823568493128, + 0.005517812445759773, + 0.005738810636103153, + 0.0046107349917292595, + 0.0019193440675735474, + 0.00650994386523962, + 0.003447512164711952, + 0.00257396069355309, + 0.0021171190310269594, + 0.004975253716111183, + 0.0021863009314984083, + 0.0081489821895957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 8 + ], + "selected_reward": -0.054941318929195404 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605393923819065 + ], + "opposite_logprob_tokens": [ + -4.141054153442383 + ], + "logprob_reward": -0.01605393923819065, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605393923819065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03523080796003342, + 0.08596199750900269, + 0.056562766432762146, + 0.22891545295715332, + 0.080508753657341, + 0.14952363073825836, + 0.22872883081436157, + 0.031705062836408615, + 0.024432340636849403, + 0.07843038439750671 + ], + "credit_map": [ + 0.0033460885751992464, + 0.0068013956770300865, + 0.003302882192656398, + 0.004502162337303162, + 0.003741320688277483, + 0.0034541527274996042, + 0.0064017390832304955, + 0.003681065747514367, + 0.008345300331711769, + 0.013432972133159637, + 0.0043991487473249435, + 0.011607913300395012, + 0.016903413459658623, + 0.011669648811221123, + 0.009745295159518719, + 0.00985830556601286, + 0.00598798505961895, + 0.0051718964241445065, + 0.0018045842880383134, + 0.004456732422113419, + 0.020303623750805855, + 0.00695402268320322, + 0.00883940514177084, + 0.003044515149667859, + 0.016399677842855453, + 0.02327568642795086, + 0.017146166414022446, + 0.0501316636800766, + 0.06911769509315491, + 0.01576613076031208, + 0.016683515161275864, + 0.02039489895105362, + 0.005383042152971029, + 0.005986312869936228, + 0.009593374095857143, + 0.012692702934145927, + 0.004360415507107973, + 0.01561728771775961, + 0.004306216724216938, + 0.022569390013813972, + 0.008918405510485172, + 0.013084683567285538, + 0.011910947039723396, + 0.007192589342594147, + 0.04776160791516304, + 0.015252658165991306, + 0.022576291114091873, + 0.022826438769698143, + 0.010592440143227577, + 0.0842471793293953, + 0.006414737552404404, + 0.006920098327100277, + 0.07224523276090622, + 0.014467339962720871, + 0.018304724246263504, + 0.015537075698375702, + 0.003628592239692807, + 0.0026637562550604343, + 0.004239542409777641, + 0.0042632645927369595, + 0.0040057143196463585, + 0.0028755832463502884, + 0.004713698290288448, + 0.005314908921718597, + 0.0025475078728049994, + 0.003346715122461319, + 0.0018261522054672241, + 0.0039650071412324905, + 0.002321515930816531, + 0.003262723097577691, + 0.004009854048490524, + 0.0031528647523373365, + 0.0046743168495595455, + 0.01476798765361309, + 0.011211979202926159, + 0.006774820387363434, + 0.0047335014678537846, + 0.011446623131632805, + 0.005351392086595297, + 0.019469760358333588 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.01605393923819065 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a whimsical, surreal cartoon style, an uninterrupted, gliding shot smoothly moves through a dreamlike seed laboratory, where thick frost rapidly crystallizes on every surface. The camera tracks rows of glowing glass vials, placed on ornate brass racks atop polished mahogany workbenches. Within each crystalline vial, a tiny, dormant sapling, its delicately spiraling leaves emitting a soft, warm, golden bioluminescence. The dominant effect is the relentless and rapid bloom of frost; intricate, shimmering ice patterns, like irregular ferns, visibly grow and spread across the glass surface, delicately enveloping the vials. This crystallization is not a harsh freeze, but a magical transformation. The newly formed ice captures and refracts the sapling's inner light, casting a dazzling prismatic shimmer throughout the laboratory. The camera's steady, gliding motion creates a sense of wonder and intimacy, allowing the viewer to witness a silent, enchanting process as the entire dreamlike environment is taken over by a breathtaking, frozen beauty.", + "seed": 28345, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28345_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.023872571252286433, + "vlm_logprob": -0.023872571252286433, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.023872571252286433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03649647533893585 + ], + "opposite_logprob_tokens": [ + -3.41149640083313 + ], + "logprob_reward": -0.03649647533893585, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03649647533893585, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1286669373512268, + 0.08950284868478775, + 0.1622328907251358, + 0.09574732929468155, + 0.040748823434114456, + 0.04204164817929268, + 0.25780579447746277, + 0.06705042719841003, + 0.031986430287361145, + 0.08421684801578522 + ], + "credit_map": [ + 0.01128658838570118, + 0.016377735882997513, + 0.011337384581565857, + 0.01663041301071644, + 0.013694108463823795, + 0.01703791506588459, + 0.01791216991841793, + 0.024390632286667824, + 0.009915507398545742, + 0.011630739085376263, + 0.011250092647969723, + 0.007600345648825169, + 0.009826787747442722, + 0.013686184771358967, + 0.011546049267053604, + 0.014047142118215561, + 0.02128446102142334, + 0.02136721834540367, + 0.03264566510915756, + 0.013047847896814346, + 0.015191403217613697, + 0.015232780016958714, + 0.01654324121773243, + 0.026920268312096596, + 0.009759797714650631, + 0.013139341026544571, + 0.011639095842838287, + 0.012640165165066719, + 0.015370974317193031, + 0.010633180849254131, + 0.011767271906137466, + 0.01079749595373869, + 0.006382053252309561, + 0.0071785361506044865, + 0.007396613247692585, + 0.0052900598384439945, + 0.003527553053572774, + 0.0034171161241829395, + 0.004429416265338659, + 0.003127475269138813, + 0.007361107971519232, + 0.005421888083219528, + 0.004381491336971521, + 0.008122127503156662, + 0.004371961113065481, + 0.004334187600761652, + 0.0048200287856161594, + 0.0032288546208292246, + 0.032064288854599, + 0.03282846510410309, + 0.019111495465040207, + 0.04774892330169678, + 0.05556470528244972, + 0.018836554139852524, + 0.03206857293844223, + 0.019582804292440414, + 0.013259687460958958, + 0.009270591661334038, + 0.005255855619907379, + 0.0048019071109592915, + 0.00787260290235281, + 0.011958515271544456, + 0.008962360210716724, + 0.00566891161724925, + 0.003199937753379345, + 0.004716265480965376, + 0.003116264007985592, + 0.004334263503551483, + 0.004010902252048254, + 0.003945317585021257, + 0.005395129788666964, + 0.003268350148573518, + 0.009309428744018078, + 0.010046443901956081, + 0.010782073251903057, + 0.009830988012254238, + 0.016210734844207764, + 0.009974740445613861, + 0.010129307396709919, + 0.007933136075735092 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 0 + ], + "selected_reward": -0.03649647533893585 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02636152133345604 + ], + "opposite_logprob_tokens": [ + -3.6513614654541016 + ], + "logprob_reward": -0.02636152133345604, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02636152133345604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.052630435675382614, + 0.05207310616970062, + 0.0997481495141983, + 0.130179762840271, + 0.5394179224967957, + 0.010246221907436848, + 0.07058966904878616, + 0.01823073998093605, + 0.018452685326337814, + 0.008431359194219112 + ], + "credit_map": [ + 0.004449229221791029, + 0.007579732686281204, + 0.004861714784055948, + 0.0058473688550293446, + 0.005654314998537302, + 0.007854730822145939, + 0.008287684991955757, + 0.00809565931558609, + 0.004887287970632315, + 0.006298697087913752, + 0.0050589339807629585, + 0.004561959765851498, + 0.006108890753239393, + 0.009989305399358273, + 0.006804211996495724, + 0.00836381409317255, + 0.01381839718669653, + 0.015688885003328323, + 0.016020262613892555, + 0.008737565949559212, + 0.010164881125092506, + 0.010647234506905079, + 0.010961988009512424, + 0.013708931393921375, + 0.01668418198823929, + 0.010087089613080025, + 0.018694104626774788, + 0.020100979134440422, + 0.019350705668330193, + 0.011264883913099766, + 0.01653868332505226, + 0.01745912805199623, + 0.16546736657619476, + 0.23500747978687286, + 0.022351223975419998, + 0.024633033201098442, + 0.0322648286819458, + 0.01984255388379097, + 0.02070704847574234, + 0.019144346937537193, + 0.0015582137275487185, + 0.001177654368802905, + 0.0014441213570535183, + 0.0017774119041860104, + 0.0012469340581446886, + 0.00115033402107656, + 0.0009241196676157415, + 0.000967432395555079, + 0.009201069362461567, + 0.00885563064366579, + 0.00527533795684576, + 0.012549439445137978, + 0.01562979444861412, + 0.004948203451931477, + 0.008885237388312817, + 0.0052449507638812065, + 0.0021361433900892735, + 0.0038449298590421677, + 0.0011360695352777839, + 0.001505499822087586, + 0.0022352670785039663, + 0.0030447179451584816, + 0.002698022872209549, + 0.001630087848752737, + 0.0017592567019164562, + 0.002156545640900731, + 0.0018980340100824833, + 0.002432408509775996, + 0.0031837509013712406, + 0.0021170699037611485, + 0.0029284656047821045, + 0.0019771535880863667, + 0.0010102660162374377, + 0.0009611968416720629, + 0.0013441805494949222, + 0.001009198254905641, + 0.001587913022376597, + 0.0009959966409951448, + 0.0006522182957269251, + 0.0008703885832801461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.02636152133345604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537707574665546 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537707574665546, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537707574665546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04562481492757797, + 0.09412354230880737, + 0.13572484254837036, + 0.11482134461402893, + 0.43577438592910767, + 0.02942577190697193, + 0.0824470967054367, + 0.02171684242784977, + 0.026468096300959587, + 0.013873239047825336 + ], + "credit_map": [ + 0.004917695187032223, + 0.005082618910819292, + 0.006164300721138716, + 0.0048534078523516655, + 0.005126070696860552, + 0.006026061251759529, + 0.006329389289021492, + 0.007125271949917078, + 0.009851536713540554, + 0.01184579636901617, + 0.013090158812701702, + 0.005709961988031864, + 0.008274217136204243, + 0.0188190545886755, + 0.011305703781545162, + 0.015227110125124454, + 0.010875863954424858, + 0.013973579742014408, + 0.03982890024781227, + 0.009236764162778854, + 0.008829033002257347, + 0.010056794621050358, + 0.009166804142296314, + 0.033757101744413376, + 0.01464735809713602, + 0.00804997980594635, + 0.014271029271185398, + 0.020270738750696182, + 0.01765376888215542, + 0.008374523371458054, + 0.01593523472547531, + 0.015618720091879368, + 0.13368871808052063, + 0.18950246274471283, + 0.0183467548340559, + 0.0201099943369627, + 0.025995977222919464, + 0.016138773411512375, + 0.016677027568221092, + 0.015314684249460697, + 0.005994686856865883, + 0.0040790969505906105, + 0.0038452576845884323, + 0.004708752501755953, + 0.0031904836650937796, + 0.002754421206191182, + 0.0026404946111142635, + 0.0022125807590782642, + 0.010633841156959534, + 0.00967366248369217, + 0.006570970173925161, + 0.015779614448547363, + 0.016995428130030632, + 0.005930989980697632, + 0.01033573318272829, + 0.006526855286210775, + 0.0031474509742110968, + 0.0030647695530205965, + 0.0013170139864087105, + 0.0024890645872801542, + 0.0037596761249005795, + 0.0036971683148294687, + 0.0022990747820585966, + 0.0019426248036324978, + 0.0029318903107196093, + 0.004225466400384903, + 0.00253951083868742, + 0.003173579927533865, + 0.003115148516371846, + 0.0034556593745946884, + 0.0034877192229032516, + 0.0035391219425946474, + 0.001187982619740069, + 0.001386463176459074, + 0.0017020631348714232, + 0.001471164869144559, + 0.003141646506264806, + 0.0017398110358044505, + 0.001489222515374422, + 0.0017548849573358893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 3 + ], + "selected_reward": -0.012537707574665546 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014189658686518669 + ], + "opposite_logprob_tokens": [ + -4.264189720153809 + ], + "logprob_reward": -0.014189658686518669, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014189658686518669, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11028620600700378, + 0.05452852323651314, + 0.14282649755477905, + 0.27822595834732056, + 0.2932855188846588, + 0.029808664694428444, + 0.020989418029785156, + 0.01837356947362423, + 0.00802971888333559, + 0.04364585503935814 + ], + "credit_map": [ + 0.014399698004126549, + 0.017057500779628754, + 0.013611416332423687, + 0.013169507496058941, + 0.009310874156653881, + 0.00843948032706976, + 0.015484115108847618, + 0.018813613802194595, + 0.005094122141599655, + 0.006124370265752077, + 0.006220338400453329, + 0.003770664567127824, + 0.006963399704545736, + 0.011244472116231918, + 0.006195510737597942, + 0.008915652520954609, + 0.007117307744920254, + 0.006091774441301823, + 0.05376693233847618, + 0.009036320261657238, + 0.004576710518449545, + 0.007108119782060385, + 0.008384442888200283, + 0.046744897961616516, + 0.037236008793115616, + 0.02589590661227703, + 0.03591087833046913, + 0.04376384988427162, + 0.04670017585158348, + 0.025148525834083557, + 0.029005704447627068, + 0.03456496074795723, + 0.08938070386648178, + 0.12593984603881836, + 0.01368307787925005, + 0.013940482400357723, + 0.01799337938427925, + 0.01066275592893362, + 0.011467792093753815, + 0.010217508301138878, + 0.005088782869279385, + 0.00400930130854249, + 0.003088157856836915, + 0.00546066090464592, + 0.003185165813192725, + 0.003967704251408577, + 0.002696319716051221, + 0.002312570810317993, + 0.002811693586409092, + 0.002348275389522314, + 0.0020199373830109835, + 0.0037605417892336845, + 0.004038569051772356, + 0.0017044756095856428, + 0.00248673465102911, + 0.001819191500544548, + 0.0022189118899405003, + 0.0027610829565674067, + 0.0014719057362526655, + 0.002026379806920886, + 0.0023738935124129057, + 0.003260271390900016, + 0.002522216411307454, + 0.0017389097483828664, + 0.0007439942564815283, + 0.001153219840489328, + 0.0008158789714798331, + 0.0010350855300202966, + 0.0008443380356766284, + 0.0010487595573067665, + 0.0012110531097278, + 0.0011773910373449326, + 0.005245485343039036, + 0.005228785332292318, + 0.006185601465404034, + 0.004086872562766075, + 0.007512783166021109, + 0.006254298612475395, + 0.005109010264277458, + 0.0040230234153568745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.014189658686518669 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029777493327856064 + ], + "opposite_logprob_tokens": [ + -3.5297775268554688 + ], + "logprob_reward": -0.029777493327856064, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029777493327856064, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1630469113588333, + 0.04960579797625542, + 0.0412340946495533, + 0.04266084358096123, + 0.571565568447113, + 0.029372166842222214, + 0.02987617254257202, + 0.028820155188441277, + 0.018981005996465683, + 0.024837248027324677 + ], + "credit_map": [ + 0.021040337160229683, + 0.02490534819662571, + 0.013324578292667866, + 0.019680438563227654, + 0.012933055870234966, + 0.01312293391674757, + 0.02649659477174282, + 0.031543634831905365, + 0.005094733089208603, + 0.005899397190660238, + 0.004347929265350103, + 0.004683870356529951, + 0.006220436189323664, + 0.004704789724200964, + 0.008447444066405296, + 0.01020719576627016, + 0.005503675434738398, + 0.004965751897543669, + 0.00662884721532464, + 0.002508321311324835, + 0.004589377902448177, + 0.008503512479364872, + 0.0034563071094453335, + 0.005078298505395651, + 0.006881876848638058, + 0.003943513613194227, + 0.008019719272851944, + 0.005520564038306475, + 0.004546298645436764, + 0.0034961714409291744, + 0.004156517796218395, + 0.006096187513321638, + 0.17647455632686615, + 0.24926254153251648, + 0.022203845903277397, + 0.026318827643990517, + 0.03494652360677719, + 0.02073395438492298, + 0.02146674133837223, + 0.02015860751271248, + 0.004065625835210085, + 0.0038105312269181013, + 0.0037713085766881704, + 0.005459665320813656, + 0.003199951723217964, + 0.004567230120301247, + 0.002526108641177416, + 0.001971745863556862, + 0.00391786266118288, + 0.004083010833710432, + 0.002369675552472472, + 0.005246229935437441, + 0.006409374065697193, + 0.0021899081766605377, + 0.003588817548006773, + 0.0020712935365736485, + 0.004171527456492186, + 0.004063635598868132, + 0.0023727237712591887, + 0.002560361986979842, + 0.00425042724236846, + 0.004211988765746355, + 0.004323558881878853, + 0.002865931950509548, + 0.002808937104418874, + 0.003069888334721327, + 0.0017242387402802706, + 0.0019842227920889854, + 0.0027094860561192036, + 0.002797334222123027, + 0.0022904211655259132, + 0.0015964776976034045, + 0.002561065601184964, + 0.003105026436969638, + 0.003418255364522338, + 0.0025404845364391804, + 0.004397095181047916, + 0.003349607577547431, + 0.00285534281283617, + 0.002610370982438326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.029777493327856064 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a whimsical, surreal cartoon style, an uninterrupted, gliding shot smoothly moves through a dreamlike seed laboratory, where thick frost rapidly crystallizes on every surface. The camera tracks rows of glowing glass vials, placed on ornate brass racks atop polished mahogany workbenches. Within each crystalline vial, a tiny, dormant sapling, its delicately spiraling leaves emitting a soft, warm, golden bioluminescence. The dominant effect is the relentless and rapid bloom of frost; intricate, shimmering ice patterns, like irregular ferns, visibly grow and spread across the glass surface, delicately enveloping the vials. This crystallization is not a harsh freeze, but a magical transformation. The newly formed ice captures and refracts the sapling's inner light, casting a dazzling prismatic shimmer throughout the laboratory. The camera's steady, gliding motion creates a sense of wonder and intimacy, allowing the viewer to witness a silent, enchanting process as the entire dreamlike environment is taken over by a breathtaking, frozen beauty.", + "seed": 28346, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_1/index_6943_seed_28346_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02072169315069914, + "vlm_logprob": -0.02072169315069914, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02072169315069914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.052254803478717804 + ], + "opposite_logprob_tokens": [ + -3.0522549152374268 + ], + "logprob_reward": -0.052254803478717804, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.052254803478717804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12949830293655396, + 0.034685295075178146, + 0.08440867066383362, + 0.08010723441839218, + 0.07917912304401398, + 0.04088933765888214, + 0.027447523549199104, + 0.14322790503501892, + 0.12039851397275925, + 0.26015815138816833 + ], + "credit_map": [ + 0.031285159289836884, + 0.012212739326059818, + 0.011474297381937504, + 0.013681094162166119, + 0.009401308372616768, + 0.007765586022287607, + 0.01239344198256731, + 0.03128465637564659, + 0.0036952162627130747, + 0.0036991580855101347, + 0.003535372205078602, + 0.003551568603143096, + 0.003241237485781312, + 0.004414573311805725, + 0.007420729845762253, + 0.005127437878400087, + 0.012048110365867615, + 0.00667469622567296, + 0.012733586132526398, + 0.013200459070503712, + 0.01090390793979168, + 0.0104045569896698, + 0.008540211245417595, + 0.00990313570946455, + 0.00460027065128088, + 0.006526606157422066, + 0.0045917523093521595, + 0.0036698216572403908, + 0.0062850844115018845, + 0.01814703643321991, + 0.011591959744691849, + 0.024694696068763733, + 0.011249260976910591, + 0.010902928188443184, + 0.012269522063434124, + 0.008340708911418915, + 0.008427743799984455, + 0.013100789859890938, + 0.007515625096857548, + 0.007372533902525902, + 0.005899933632463217, + 0.007736410945653915, + 0.0046266112476587296, + 0.0037361630238592625, + 0.005226311273872852, + 0.005913339555263519, + 0.0035557986702769995, + 0.004194764420390129, + 0.0040334537625312805, + 0.005237164441496134, + 0.0027498325798660517, + 0.002926638349890709, + 0.002840291941538453, + 0.003072480671107769, + 0.0029096179641783237, + 0.0036780403461307287, + 0.03126217797398567, + 0.049346208572387695, + 0.01207401417195797, + 0.013888020068407059, + 0.008963956497609615, + 0.0134192556142807, + 0.006791375577449799, + 0.007482895627617836, + 0.019397776573896408, + 0.015470688231289387, + 0.018281875178217888, + 0.018729345872998238, + 0.014063727110624313, + 0.013895475305616856, + 0.010854947380721569, + 0.00970466248691082, + 0.01482360903173685, + 0.09296160936355591, + 0.06156102195382118, + 0.014662624336779118, + 0.03480274975299835, + 0.01681653968989849, + 0.01257028616964817, + 0.011959657073020935 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.052254803478717804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004104048945009708 + ], + "opposite_logprob_tokens": [ + -5.504104137420654 + ], + "logprob_reward": -0.004104048945009708, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004104048945009708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17645223438739777, + 0.052243683487176895, + 0.08391691744327545, + 0.035004790872335434, + 0.044412482529878616, + 0.052618175745010376, + 0.01739434152841568, + 0.12850837409496307, + 0.1452653408050537, + 0.2641836106777191 + ], + "credit_map": [ + 0.05087036266922951, + 0.013948333449661732, + 0.01227249950170517, + 0.014830050989985466, + 0.014443851076066494, + 0.013484091497957706, + 0.01385080348700285, + 0.042752232402563095, + 0.005834387615323067, + 0.006373245734721422, + 0.009977800771594048, + 0.0069099776446819305, + 0.0038632131181657314, + 0.0054772780276834965, + 0.007166003808379173, + 0.006641776766628027, + 0.011313769035041332, + 0.006960013415664434, + 0.01363640557974577, + 0.011579773388803005, + 0.00911326427012682, + 0.0081375977024436, + 0.015232227742671967, + 0.007943865843117237, + 0.0037720040418207645, + 0.003871473018079996, + 0.004318746272474527, + 0.003909503109753132, + 0.004654812626540661, + 0.005027599632740021, + 0.0037011224776506424, + 0.00574952969327569, + 0.0056397016160190105, + 0.00540130352601409, + 0.007260930258780718, + 0.005142217967659235, + 0.004483556840568781, + 0.005845516454428434, + 0.005286797881126404, + 0.00535245705395937, + 0.011274403892457485, + 0.006633620709180832, + 0.007504396140575409, + 0.0058817146345973015, + 0.003913113847374916, + 0.005820826627314091, + 0.005379355046898127, + 0.006210748106241226, + 0.002982674166560173, + 0.0030431330669671297, + 0.0017640687292441726, + 0.002093210816383362, + 0.0018250350840389729, + 0.0018661218928173184, + 0.0018974555423483253, + 0.0019226439762860537, + 0.026763951405882835, + 0.038313183933496475, + 0.01790686510503292, + 0.01296207495033741, + 0.007873531430959702, + 0.00977612566202879, + 0.0065176645293831825, + 0.008394971489906311, + 0.02339719980955124, + 0.0169406495988369, + 0.018701372668147087, + 0.027954574674367905, + 0.013023524545133114, + 0.02306114137172699, + 0.012504961341619492, + 0.009681916795670986, + 0.016956960782408714, + 0.1051880270242691, + 0.08058168739080429, + 0.008895611390471458, + 0.02500595897436142, + 0.011811218224465847, + 0.0081352973356843, + 0.007608872838318348 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.004104048945009708 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005950829479843378 + ], + "opposite_logprob_tokens": [ + -5.130950927734375 + ], + "logprob_reward": -0.005950829479843378, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005950829479843378, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07843795418739319, + 0.028334612026810646, + 0.091822050511837, + 0.04140317812561989, + 0.06631976366043091, + 0.05000082775950432, + 0.0213149506598711, + 0.0817008763551712, + 0.1182916909456253, + 0.42237409949302673 + ], + "credit_map": [ + 0.008615285158157349, + 0.01086512766778469, + 0.008507698774337769, + 0.010017191991209984, + 0.010354019701480865, + 0.009722123853862286, + 0.010375513695180416, + 0.00998099148273468, + 0.0025883468333631754, + 0.0033086345065385103, + 0.004695235751569271, + 0.0035473089665174484, + 0.002519317902624607, + 0.0032512634061276913, + 0.004180004820227623, + 0.004244497511535883, + 0.013929116539657116, + 0.007921379059553146, + 0.014493186958134174, + 0.010709740221500397, + 0.011215822771191597, + 0.010039303451776505, + 0.014882639050483704, + 0.008630857802927494, + 0.004662499763071537, + 0.0045142159797251225, + 0.004977569915354252, + 0.0034114776644855738, + 0.004320250824093819, + 0.00585786160081625, + 0.0057290298864245415, + 0.007930270396173, + 0.0104595348238945, + 0.008982771076261997, + 0.01061377301812172, + 0.007286950014531612, + 0.00600815424695611, + 0.007898126728832722, + 0.007337337359786034, + 0.007733118254691362, + 0.010445072315633297, + 0.005972486454993486, + 0.005776300095021725, + 0.002896462334319949, + 0.004269135184586048, + 0.01125893834978342, + 0.004754677880555391, + 0.004627755377441645, + 0.004010501317679882, + 0.0040797037072479725, + 0.002149165142327547, + 0.0021709317807108164, + 0.0020126341842114925, + 0.002064470201730728, + 0.002390878740698099, + 0.002436664653941989, + 0.013410982675850391, + 0.010240800678730011, + 0.023369338363409042, + 0.008887476287782192, + 0.007208348251879215, + 0.00825880840420723, + 0.005247273948043585, + 0.005077853333204985, + 0.02148178406059742, + 0.013269471935927868, + 0.020838959142565727, + 0.017914842814207077, + 0.010383512824773788, + 0.01754816249012947, + 0.009046323597431183, + 0.007808632683008909, + 0.026171427220106125, + 0.16851942241191864, + 0.1257665455341339, + 0.01521981693804264, + 0.041583556681871414, + 0.0194978229701519, + 0.013415737077593803, + 0.012199770659208298 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 2 + ], + "selected_reward": -0.005950829479843378 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623388897627592 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623388897627592, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623388897627592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18746592104434967, + 0.035361699759960175, + 0.050989165902137756, + 0.08402819186449051, + 0.1151638925075531, + 0.050187546759843826, + 0.030763275921344757, + 0.13707631826400757, + 0.06580694019794464, + 0.24315707385540009 + ], + "credit_map": [ + 0.05538148805499077, + 0.015710486099123955, + 0.010397224687039852, + 0.013165110722184181, + 0.012812908738851547, + 0.016965195536613464, + 0.014525564387440681, + 0.0485079251229763, + 0.003196656471118331, + 0.004163753241300583, + 0.007037260569632053, + 0.00422243494540453, + 0.0026310179382562637, + 0.003492755116894841, + 0.005386500619351864, + 0.005231321323662996, + 0.004919915925711393, + 0.0036007631570100784, + 0.00935562513768673, + 0.007881156168878078, + 0.004277793224900961, + 0.003833676688373089, + 0.011708605103194714, + 0.0054116323590278625, + 0.011613990180194378, + 0.008789276704192162, + 0.008766470476984978, + 0.006299466826021671, + 0.005456948187202215, + 0.005786666180938482, + 0.0140608511865139, + 0.023254519328475, + 0.01687823422253132, + 0.016773145645856857, + 0.01474866271018982, + 0.011663062497973442, + 0.013106743805110455, + 0.01862889714539051, + 0.010246465913951397, + 0.013118678703904152, + 0.005353330168873072, + 0.010626057162880898, + 0.004196922294795513, + 0.0038955218624323606, + 0.008759617805480957, + 0.007659777533262968, + 0.005999332759529352, + 0.003696984611451626, + 0.004676697310060263, + 0.0061250114813447, + 0.003365365555509925, + 0.003370394231751561, + 0.0032303915359079838, + 0.0025909447576850653, + 0.0038004464004188776, + 0.003604025347158313, + 0.019289005547761917, + 0.014846676029264927, + 0.03808125481009483, + 0.01791895553469658, + 0.01239714864641428, + 0.01506667211651802, + 0.010080846026539803, + 0.009395756758749485, + 0.008529976941645145, + 0.0094155790284276, + 0.00886237621307373, + 0.01109789777547121, + 0.00623719347640872, + 0.010940580628812313, + 0.006126592867076397, + 0.004596743267029524, + 0.015741636976599693, + 0.08899108320474625, + 0.07251689583063126, + 0.01213726494461298, + 0.018547408282756805, + 0.012706040404736996, + 0.012362794019281864, + 0.010153958573937416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.007623388897627592 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the visual style a whimsical, surreal cartoon style?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the setting a seed laboratory with ornate brass racks and polished mahogany workbenches?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are glowing glass vials containing tiny dormant saplings with spiraling leaves visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is thick frost rapidly crystallizing on surfaces, forming intricate ice patterns like irregular ferns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera execute a smooth, uninterrupted gliding shot through the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03367539495229721 + ], + "opposite_logprob_tokens": [ + -3.4086754322052 + ], + "logprob_reward": -0.03367539495229721, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03367539495229721, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18653333187103271, + 0.029330961406230927, + 0.024740753695368767, + 0.03920692950487137, + 0.05167383328080177, + 0.026349928230047226, + 0.021655963733792305, + 0.5181720852851868, + 0.06949326395988464, + 0.032842960208654404 + ], + "credit_map": [ + 0.05246348679065704, + 0.014043967239558697, + 0.011050008237361908, + 0.01728939823806286, + 0.012621216475963593, + 0.015317955985665321, + 0.016898365691304207, + 0.04684893786907196, + 0.0027906165923923254, + 0.003324202261865139, + 0.004105248488485813, + 0.004976114723831415, + 0.0027163196355104446, + 0.0029705625493079424, + 0.004555118270218372, + 0.003892777720466256, + 0.0025094125885516405, + 0.002442781114950776, + 0.0057822358794510365, + 0.002077449345961213, + 0.0021064691245555878, + 0.0017094953218474984, + 0.005993167404085398, + 0.0021197423338890076, + 0.005447127390652895, + 0.005180189851671457, + 0.004947817884385586, + 0.0033583000767976046, + 0.003973404876887798, + 0.008060161024332047, + 0.003555414965376258, + 0.00468451576307416, + 0.006381410639733076, + 0.008092953823506832, + 0.008809457533061504, + 0.005090000107884407, + 0.00459030969068408, + 0.007177611812949181, + 0.005797827150672674, + 0.005734261125326157, + 0.0032582052517682314, + 0.005095301661640406, + 0.002416693139821291, + 0.002239239402115345, + 0.0024596252478659153, + 0.0048479801043868065, + 0.0028585942927747965, + 0.0031742886640131474, + 0.00352382636629045, + 0.002748640952631831, + 0.002485354198142886, + 0.0023048638831824064, + 0.002395040588453412, + 0.0024795960634946823, + 0.0029976745136082172, + 0.0027209692634642124, + 0.1070893332362175, + 0.14937354624271393, + 0.05062204226851463, + 0.055646613240242004, + 0.03441474959254265, + 0.038458116352558136, + 0.03979811817407608, + 0.04276955872774124, + 0.01034488808363676, + 0.00834216084331274, + 0.009535904973745346, + 0.010378316976130009, + 0.006210064049810171, + 0.012237547896802425, + 0.006919613108038902, + 0.005524769891053438, + 0.0025767264887690544, + 0.004385918378829956, + 0.004156873561441898, + 0.003937729634344578, + 0.007179113104939461, + 0.0028708577156066895, + 0.005053223576396704, + 0.0026825203094631433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.03367539495229721 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_tensor([1392, 1392])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_tensor([1392, 1392])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e0d70d14fa17f84365f72484a34ffc5074f4d3f0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_tensor([1392, 1392])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning macro close-up shot, the camera held still, focuses on a pot of clear, scalding hot water, where a square, dry, and firm, pale yellow instant noodle cake quietly soaks. A soft, warm light from above penetrates the rising white steam, illuminating the entire scene and clearly revealing the wavy texture pressed into the noodle cake. As the water seeps in, the edges of the noodle cake begin to round out, and the curled noodles on the outer layer begin to loosen, delicately unfolding like petals. Then, the entire noodle cake steadily and leisurely expands from the inside out. The once tightly packed, wavy noodles break free, floating gracefully and unraveling in the water, resuming their natural curls. Tiny bubbles continuously emerge from the bottom of the pot, clinging to the increasingly translucent noodles, whose moist surfaces reflect an alluring luster in the light. The slow and rhythmic process creates a warm, soothing, and deeply satisfying tranquil atmosphere.", + "seed": 1459, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1459_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05617902716621757, + "vlm_logprob": -0.05617902716621757, + "vlm_yes_no_margin": 4.275, + "vlm_token_credit": -0.05617902716621757, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03315923362970352 + ], + "opposite_logprob_tokens": [ + -4.408159255981445 + ], + "logprob_reward": -0.03315923362970352, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.03315923362970352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0881844162940979, + 0.05983951315283775, + 0.07612840086221695, + 0.08870657533407211, + 0.017438150942325592, + 0.09316196292638779, + 0.33925342559814453, + 0.1119227260351181, + 0.07353310286998749, + 0.0518316850066185 + ], + "credit_map": [ + 0.012315529398620129, + 0.015349670313298702, + 0.00647764652967453, + 0.01515236496925354, + 0.011065925471484661, + 0.009088585153222084, + 0.0119566535577178, + 0.006778046488761902, + 0.0115797845646739, + 0.013090690597891808, + 0.004262410569936037, + 0.0066429064609110355, + 0.007386595010757446, + 0.007069679908454418, + 0.005732572637498379, + 0.004074877593666315, + 0.011549398303031921, + 0.009912199340760708, + 0.005894879810512066, + 0.010244891047477722, + 0.006872898433357477, + 0.01697971485555172, + 0.008477951399981976, + 0.006196466274559498, + 0.013463337905704975, + 0.01224619708955288, + 0.014828014187514782, + 0.010763988830149174, + 0.007340393029153347, + 0.013639162294566631, + 0.008308447897434235, + 0.008117031306028366, + 0.0025828867219388485, + 0.0017884294502437115, + 0.0018579415045678616, + 0.002957194112241268, + 0.0018061515875160694, + 0.0017960448749363422, + 0.0030167116783559322, + 0.0016327922930940986, + 0.021035192534327507, + 0.006265838630497456, + 0.004797244910150766, + 0.011945941485464573, + 0.006604740861803293, + 0.011622399091720581, + 0.025527318939566612, + 0.0053632850758731365, + 0.0496692880988121, + 0.030414870008826256, + 0.03726295009255409, + 0.06553011387586594, + 0.01813868246972561, + 0.027513617649674416, + 0.08810441941022873, + 0.022619500756263733, + 0.018881134688854218, + 0.006933152675628662, + 0.004591984674334526, + 0.02390860766172409, + 0.005420260597020388, + 0.0180676206946373, + 0.027093222364783287, + 0.007026738952845335, + 0.008350534364581108, + 0.005287739913910627, + 0.002608776092529297, + 0.019870348274707794, + 0.003741728374734521, + 0.0049211736768484116, + 0.022100407630205154, + 0.00665239617228508, + 0.007451889105141163, + 0.006349945440888405, + 0.005099879112094641, + 0.007164142560213804, + 0.004462901968508959, + 0.008639253675937653, + 0.005970324389636517, + 0.006693349685519934 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.03315923362970352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019738255068659782 + ], + "opposite_logprob_tokens": [ + -6.251973628997803 + ], + "logprob_reward": -0.0019738255068659782, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019738255068659782, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1338602900505066, + 0.06992422789335251, + 0.07314865291118622, + 0.11479302495718002, + 0.06439923495054245, + 0.10446037352085114, + 0.1675342172384262, + 0.19417153298854828, + 0.04196185991168022, + 0.035746634006500244 + ], + "credit_map": [ + 0.01297756191343069, + 0.03017308935523033, + 0.007362224627286196, + 0.015598271042108536, + 0.0224557276815176, + 0.016465257853269577, + 0.019962452352046967, + 0.00886570941656828, + 0.012027994729578495, + 0.011778103187680244, + 0.0035797543823719025, + 0.010417932644486427, + 0.009258408099412918, + 0.008491721004247665, + 0.009607884101569653, + 0.004762435331940651, + 0.009865486063063145, + 0.008467474021017551, + 0.0031070197001099586, + 0.011964933946728706, + 0.007379875052720308, + 0.015795055776834488, + 0.011373062618076801, + 0.00519575085490942, + 0.029965372756123543, + 0.010858239606022835, + 0.007020534947514534, + 0.013277665711939335, + 0.009526141919195652, + 0.026144852861762047, + 0.010998616926372051, + 0.007001599296927452, + 0.0059375157579779625, + 0.006939275655895472, + 0.003680524183437228, + 0.014049561694264412, + 0.010462979786098003, + 0.005057842470705509, + 0.013802426867187023, + 0.0044691129587590694, + 0.019012680277228355, + 0.007399530615657568, + 0.003650231519713998, + 0.011547745205461979, + 0.013272672891616821, + 0.012061657384037971, + 0.031162938103079796, + 0.006352918688207865, + 0.025070123374462128, + 0.015533681958913803, + 0.017363540828227997, + 0.03178604319691658, + 0.008839217014610767, + 0.015042085200548172, + 0.04356643557548523, + 0.010333090089261532, + 0.0291421040892601, + 0.009002079255878925, + 0.006569073535501957, + 0.04653596878051758, + 0.007091887295246124, + 0.021008603274822235, + 0.06464553624391556, + 0.010176297277212143, + 0.007001122925430536, + 0.0034836363047361374, + 0.0022760562133044004, + 0.008524802513420582, + 0.002651963382959366, + 0.004138852935284376, + 0.008910740725696087, + 0.00497468700632453, + 0.003809430869296193, + 0.004168483894318342, + 0.0021056977566331625, + 0.008112149313092232, + 0.00356920063495636, + 0.0026156126987189054, + 0.008334747515618801, + 0.0030313122551888227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.0019738255068659782 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059676566161215305 + ], + "opposite_logprob_tokens": [ + -5.130967617034912 + ], + "logprob_reward": -0.0059676566161215305, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059676566161215305, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1751420795917511, + 0.08241043239831924, + 0.12319432944059372, + 0.09943094849586487, + 0.04702245444059372, + 0.058062683790922165, + 0.24254845082759857, + 0.11558828502893448, + 0.027089206501841545, + 0.02951115556061268 + ], + "credit_map": [ + 0.01998923160135746, + 0.026440182700753212, + 0.009792719967663288, + 0.02541433461010456, + 0.030664972960948944, + 0.022084137424826622, + 0.026129189878702164, + 0.014627287164330482, + 0.01392216607928276, + 0.012654703110456467, + 0.004147800151258707, + 0.01169427391141653, + 0.01158720813691616, + 0.011036155745387077, + 0.012091993354260921, + 0.005276121664792299, + 0.016805073246359825, + 0.021928176283836365, + 0.00784614123404026, + 0.02062729373574257, + 0.011507656425237656, + 0.015925725921988487, + 0.019800951704382896, + 0.008753303438425064, + 0.02653958648443222, + 0.012241162359714508, + 0.011624726466834545, + 0.009023360908031464, + 0.007101021241396666, + 0.01956438273191452, + 0.007376194000244141, + 0.005960513837635517, + 0.004195269197225571, + 0.0055643245577812195, + 0.004488811828196049, + 0.009890452958643436, + 0.005377310328185558, + 0.0030972848180681467, + 0.00992579199373722, + 0.004483206197619438, + 0.009522421285510063, + 0.0058697545900940895, + 0.0065713198855519295, + 0.010316172614693642, + 0.008804435841739178, + 0.0055304099805653095, + 0.007148243952542543, + 0.0042999181896448135, + 0.039159756153821945, + 0.020239965990185738, + 0.018105287104845047, + 0.048650678247213364, + 0.01311629917472601, + 0.02095949277281761, + 0.06870841979980469, + 0.013608512468636036, + 0.020576097071170807, + 0.005032193381339312, + 0.004069953225553036, + 0.023358002305030823, + 0.004208240192383528, + 0.018116267397999763, + 0.033076588064432144, + 0.0071509284898638725, + 0.004002743400633335, + 0.00167073926422745, + 0.001645006937906146, + 0.007810893002897501, + 0.001480238395743072, + 0.0022910484112799168, + 0.006317903287708759, + 0.001870631822384894, + 0.0043305098079144955, + 0.0033633902203291655, + 0.003990672528743744, + 0.003158829640597105, + 0.0028601239901036024, + 0.003899491624906659, + 0.004282814916223288, + 0.0036253216676414013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.0059676566161215305 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014201881363987923 + ], + "opposite_logprob_tokens": [ + -4.264202117919922 + ], + "logprob_reward": -0.014201881363987923, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014201881363987923, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09585901349782944, + 0.030379703268408775, + 0.10225815325975418, + 0.05522098392248154, + 0.020343957468867302, + 0.06779137253761292, + 0.3858756422996521, + 0.14150889217853546, + 0.024234799668192863, + 0.076527439057827 + ], + "credit_map": [ + 0.013289421796798706, + 0.02359328418970108, + 0.010573293082416058, + 0.009511820040643215, + 0.009177789092063904, + 0.011855477467179298, + 0.009517253376543522, + 0.0083406837657094, + 0.005678981076925993, + 0.00545997591689229, + 0.0033481353893876076, + 0.0047248792834579945, + 0.002851206576451659, + 0.0025636092759668827, + 0.0033908833283931017, + 0.0023620326537638903, + 0.014418373815715313, + 0.019802385941147804, + 0.0056795296259224415, + 0.014740675687789917, + 0.009617728181183338, + 0.018422676250338554, + 0.011911934241652489, + 0.007664850447326899, + 0.006876339204609394, + 0.009124740026891232, + 0.005091214086860418, + 0.007863134145736694, + 0.006678716745227575, + 0.006369859911501408, + 0.008437187410891056, + 0.0047797937877476215, + 0.002369513502344489, + 0.0024454868398606777, + 0.0021622530184686184, + 0.0036032036878168583, + 0.0027446565218269825, + 0.0022019229363650084, + 0.0028698535170406103, + 0.00194706826005131, + 0.010852508246898651, + 0.005642492324113846, + 0.0042397272773087025, + 0.01179244089871645, + 0.007779226638376713, + 0.0070118773728609085, + 0.015654753893613815, + 0.004818345420062542, + 0.05560040473937988, + 0.034194327890872955, + 0.04248920828104019, + 0.07651384174823761, + 0.02190544083714485, + 0.02965512126684189, + 0.10227210074663162, + 0.023245221003890038, + 0.025058623403310776, + 0.0060869744047522545, + 0.003972959239035845, + 0.029611948877573013, + 0.005229493603110313, + 0.020859602838754654, + 0.04238900542259216, + 0.008300283923745155, + 0.003668227232992649, + 0.0030895122326910496, + 0.0018740129889920354, + 0.005427846685051918, + 0.0022517936304211617, + 0.001827558153308928, + 0.0035091449972242117, + 0.0025867025833576918, + 0.00701404083520174, + 0.006994910072535276, + 0.004302981775254011, + 0.013306022621691227, + 0.009048053994774818, + 0.008511418476700783, + 0.020801186561584473, + 0.006548828445374966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 2 + ], + "selected_reward": -0.014201881363987923 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559253871440887 + ], + "opposite_logprob_tokens": [ + -1.6005924940109253 + ], + "logprob_reward": -0.22559253871440887, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559253871440887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07307277619838715, + 0.038447845727205276, + 0.1007687896490097, + 0.09351935237646103, + 0.04332659766077995, + 0.24631348252296448, + 0.15915200114250183, + 0.1010870710015297, + 0.03408799320459366, + 0.11022408306598663 + ], + "credit_map": [ + 0.008086164481937885, + 0.012806751765310764, + 0.005864135921001434, + 0.008750074543058872, + 0.00937286950647831, + 0.009980508126318455, + 0.00850815512239933, + 0.009704113006591797, + 0.006659782491624355, + 0.005436380859464407, + 0.0027205701917409897, + 0.005278666969388723, + 0.005864534992724657, + 0.0038433594163507223, + 0.0062605952844023705, + 0.002383954357355833, + 0.013440091162919998, + 0.0173825453966856, + 0.0062233805656433105, + 0.014566471800208092, + 0.00857776403427124, + 0.01826317608356476, + 0.013148956932127476, + 0.009166399948298931, + 0.022684196010231972, + 0.010548143647611141, + 0.009478071704506874, + 0.009563562460243702, + 0.008194074966013432, + 0.01980448327958584, + 0.006957436911761761, + 0.0062893833965063095, + 0.005328495055437088, + 0.005054195877164602, + 0.006133243441581726, + 0.00551989208906889, + 0.007577292155474424, + 0.004860416986048222, + 0.004778199829161167, + 0.004074864089488983, + 0.03333626314997673, + 0.01745903491973877, + 0.007646678481251001, + 0.02747102454304695, + 0.04430988058447838, + 0.023119987919926643, + 0.07850175350904465, + 0.01446886919438839, + 0.019784262403845787, + 0.018614498898386955, + 0.01520521193742752, + 0.031642768532037735, + 0.009922828525304794, + 0.013914204202592373, + 0.038482170552015305, + 0.011586052365601063, + 0.01854787953197956, + 0.00810482632368803, + 0.006226181518286467, + 0.016085222363471985, + 0.006407949607819319, + 0.012563570402562618, + 0.023280613124370575, + 0.009870825335383415, + 0.004786231089383364, + 0.004721934907138348, + 0.0037822704762220383, + 0.005636902526021004, + 0.002909933915361762, + 0.0033017140813171864, + 0.004572464153170586, + 0.004376545082777739, + 0.013584522530436516, + 0.011330452747642994, + 0.015307779423892498, + 0.013811451382935047, + 0.008519774302840233, + 0.018804237246513367, + 0.015459958463907242, + 0.013405917212367058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 9 + ], + "selected_reward": -0.22559253871440887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up shot, the camera held still, focuses on a pot of clear, scalding hot water, where a square, dry, and firm, pale yellow instant noodle cake quietly soaks. A soft, warm light from above penetrates the rising white steam, illuminating the entire scene and clearly revealing the wavy texture pressed into the noodle cake. As the water seeps in, the edges of the noodle cake begin to round out, and the curled noodles on the outer layer begin to loosen, delicately unfolding like petals. Then, the entire noodle cake steadily and leisurely expands from the inside out. The once tightly packed, wavy noodles break free, floating gracefully and unraveling in the water, resuming their natural curls. Tiny bubbles continuously emerge from the bottom of the pot, clinging to the increasingly translucent noodles, whose moist surfaces reflect an alluring luster in the light. The slow and rhythmic process creates a warm, soothing, and deeply satisfying tranquil atmosphere.", + "seed": 1460, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1460_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07836285457015038, + "vlm_logprob": -0.07836285457015038, + "vlm_yes_no_margin": 3.6500000238418577, + "vlm_token_credit": -0.07836285457015038, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16812077164649963 + ], + "opposite_logprob_tokens": [ + -2.043120861053467 + ], + "logprob_reward": -0.16812077164649963, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.16812077164649963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.227495014667511, + 0.07585157454013824, + 0.026623733341693878, + 0.015327073633670807, + 0.07453320175409317, + 0.04900345206260681, + 0.410499632358551, + 0.03613482415676117, + 0.05826694518327713, + 0.02626453898847103 + ], + "credit_map": [ + 0.01325107179582119, + 0.042225904762744904, + 0.06113264709711075, + 0.024053730070590973, + 0.016871893778443336, + 0.03178561478853226, + 0.017674461007118225, + 0.020499717444181442, + 0.011938358657062054, + 0.009709825739264488, + 0.005645744036883116, + 0.00938049703836441, + 0.0065178838558495045, + 0.01253702025860548, + 0.007948353886604309, + 0.012173895724117756, + 0.0040464517660439014, + 0.0027257162146270275, + 0.0029989262111485004, + 0.0029822203796356916, + 0.0024778791703283787, + 0.004871902987360954, + 0.001869404106400907, + 0.00465123588219285, + 0.001202556537464261, + 0.004986975807696581, + 0.0016735702520236373, + 0.001792094437405467, + 0.0017881460953503847, + 0.0015215998282656074, + 0.0013734247768297791, + 0.000988706131465733, + 0.004593679215759039, + 0.0030165533535182476, + 0.024462535977363586, + 0.007731799501925707, + 0.012336360290646553, + 0.01793866977095604, + 0.0019496716558933258, + 0.0025039385072886944, + 0.007120564114302397, + 0.005040062125772238, + 0.004926536697894335, + 0.010156501084566116, + 0.0036966348998248577, + 0.0045799738727509975, + 0.009923473000526428, + 0.0035597102250903845, + 0.018651088699698448, + 0.14895473420619965, + 0.01153535582125187, + 0.15330101549625397, + 0.026901625096797943, + 0.014777650125324726, + 0.0247146375477314, + 0.01166356448084116, + 0.0028996646869927645, + 0.004766445141285658, + 0.0055513642728328705, + 0.007416706532239914, + 0.003915630746632814, + 0.003030681051313877, + 0.004982779733836651, + 0.0035715585108846426, + 0.004458157811313868, + 0.008247152902185917, + 0.006271739490330219, + 0.007518348284065723, + 0.008949053473770618, + 0.004224136006087065, + 0.012054718099534512, + 0.006543646100908518, + 0.002922494662925601, + 0.0035165008157491684, + 0.0033826599828898907, + 0.003460300387814641, + 0.0031124036759138107, + 0.0027403051499277353, + 0.003388339886441827, + 0.003741537919268012 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.16812077164649963 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004678138066083193 + ], + "opposite_logprob_tokens": [ + -5.379678249359131 + ], + "logprob_reward": -0.004678138066083193, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004678138066083193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3196566104888916, + 0.04619995877146721, + 0.016166705638170242, + 0.01904313452541828, + 0.30926835536956787, + 0.03377490118145943, + 0.1443827897310257, + 0.02639726735651493, + 0.05169784650206566, + 0.03341247886419296 + ], + "credit_map": [ + 0.019127491861581802, + 0.1119443029165268, + 0.05081997439265251, + 0.030637847259640694, + 0.03940604627132416, + 0.029788322746753693, + 0.012709011323750019, + 0.025223620235919952, + 0.005256385076791048, + 0.005808813497424126, + 0.0049645788967609406, + 0.00837739184498787, + 0.006314162164926529, + 0.006060854997485876, + 0.0050887856632471085, + 0.004328983835875988, + 0.0016515791648998857, + 0.0018265078542754054, + 0.0015770334284752607, + 0.003152984892949462, + 0.0023219073191285133, + 0.0019399350276216865, + 0.0023707840591669083, + 0.0013259731931611896, + 0.0015691635198891163, + 0.005247848108410835, + 0.0023039530497044325, + 0.002411159686744213, + 0.002132169436663389, + 0.001795477932319045, + 0.0021764892153441906, + 0.0014068736927583814, + 0.020024556666612625, + 0.011272678151726723, + 0.1138114482164383, + 0.024669155478477478, + 0.023357026278972626, + 0.09821876883506775, + 0.005081407260149717, + 0.012833337299525738, + 0.004987858701497316, + 0.0034581059589982033, + 0.0034854731056839228, + 0.00622872170060873, + 0.002421565121039748, + 0.003338595386594534, + 0.007178411819040775, + 0.002676168456673622, + 0.00666270824149251, + 0.053925544023513794, + 0.003785890294238925, + 0.053985003381967545, + 0.008985340595245361, + 0.00523581076413393, + 0.008001776412129402, + 0.0038007106631994247, + 0.0024155043065547943, + 0.0052530947141349316, + 0.002969467081129551, + 0.0065415785647928715, + 0.0022262092679739, + 0.0018673944287002087, + 0.0028253577183932066, + 0.0022986598778516054, + 0.004573105834424496, + 0.01116189081221819, + 0.003782970132306218, + 0.006706720218062401, + 0.007832232862710953, + 0.0038354371208697557, + 0.009286640211939812, + 0.004518847446888685, + 0.0036062123253941536, + 0.00498407194390893, + 0.003359598573297262, + 0.004250628408044577, + 0.004461200907826424, + 0.0031248696614056826, + 0.0038389586843550205, + 0.00578693812713027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.004678138066083193 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764929736033082 + ], + "opposite_logprob_tokens": [ + -4.8826494216918945 + ], + "logprob_reward": -0.00764929736033082, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764929736033082, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42807257175445557, + 0.02498479373753071, + 0.020991045981645584, + 0.01838505268096924, + 0.12399041652679443, + 0.01839962787926197, + 0.259823739528656, + 0.05744946748018265, + 0.015019196085631847, + 0.032884154468774796 + ], + "credit_map": [ + 0.020703334361314774, + 0.15036720037460327, + 0.08398837596178055, + 0.04017048701643944, + 0.05049819499254227, + 0.043131422251462936, + 0.015504329465329647, + 0.023709174245595932, + 0.002809775061905384, + 0.0031896226573735476, + 0.002637134864926338, + 0.004601568914949894, + 0.0025740182027220726, + 0.003452452365309, + 0.003241084748879075, + 0.0024791336618363857, + 0.002182802651077509, + 0.00218398729339242, + 0.002258794382214546, + 0.0039149802178144455, + 0.0027131964452564716, + 0.0030134557746350765, + 0.002577295759692788, + 0.002146529732272029, + 0.0017008855938911438, + 0.004378998186439276, + 0.0022626114077866077, + 0.002505522919818759, + 0.001994719263166189, + 0.002232432132586837, + 0.0022330647334456444, + 0.0010768164647743106, + 0.004699322860687971, + 0.012769625522196293, + 0.023005254566669464, + 0.0194089375436306, + 0.04580903798341751, + 0.008743739686906338, + 0.004602022934705019, + 0.0049524614587426186, + 0.0019605541601777077, + 0.0026244160253554583, + 0.001983448164537549, + 0.0034958194009959698, + 0.0018254797905683517, + 0.0019932640716433525, + 0.002900873776525259, + 0.0016157700447365642, + 0.013359000906348228, + 0.09629963338375092, + 0.008057720959186554, + 0.09801199287176132, + 0.013985281810164452, + 0.009417866356670856, + 0.013695970177650452, + 0.006996244192123413, + 0.0037695453502237797, + 0.011911298148334026, + 0.006219939794391394, + 0.015657247975468636, + 0.0072246151976287365, + 0.003505144966766238, + 0.0057657258585095406, + 0.0033959452994167805, + 0.0016828997759148479, + 0.0015683532692492008, + 0.0021054940298199654, + 0.002106128493323922, + 0.001897017820738256, + 0.0012520593591034412, + 0.0028101925272494555, + 0.001597049762494862, + 0.0034057083539664745, + 0.005251938011497259, + 0.004822372924536467, + 0.0038798311725258827, + 0.004745058249682188, + 0.0027823399286717176, + 0.0036145569756627083, + 0.004382347222417593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.00764929736033082 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009791442193090916 + ], + "opposite_logprob_tokens": [ + -4.634791374206543 + ], + "logprob_reward": -0.009791442193090916, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009791442193090916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3896026909351349, + 0.09875684231519699, + 0.013875081203877926, + 0.010646023787558079, + 0.20870256423950195, + 0.0650332048535347, + 0.039052609354257584, + 0.12348522990942001, + 0.03405249863862991, + 0.01679324358701706 + ], + "credit_map": [ + 0.021439991891384125, + 0.09449256956577301, + 0.11814102530479431, + 0.028609685599803925, + 0.03200175613164902, + 0.0561084970831871, + 0.01129213161766529, + 0.027517044916749, + 0.015890147536993027, + 0.012417585588991642, + 0.00911406334489584, + 0.008976324461400509, + 0.009921311400830746, + 0.017430303618311882, + 0.008508863858878613, + 0.016498232260346413, + 0.0014566286699846387, + 0.0011310144327580929, + 0.0019038582686334848, + 0.0017564443405717611, + 0.001617856090888381, + 0.0023987647145986557, + 0.0015288887079805136, + 0.0020816256292164326, + 0.0008912578923627734, + 0.002525712363421917, + 0.0012396163074299693, + 0.0013457542518153787, + 0.0012236912734806538, + 0.0009870438370853662, + 0.001264263759367168, + 0.0011686848010867834, + 0.013665915466845036, + 0.005568445194512606, + 0.07775561511516571, + 0.02404646947979927, + 0.013481369242072105, + 0.06093642860651016, + 0.004280703142285347, + 0.008967608213424683, + 0.0079514030367136, + 0.006385121960192919, + 0.006410831585526466, + 0.014774365350604057, + 0.0050842962227761745, + 0.0076356492936611176, + 0.012774177826941013, + 0.004017357248812914, + 0.0027377435471862555, + 0.009854873642325401, + 0.0027122970204800367, + 0.010529232211411, + 0.0035560831893235445, + 0.0028089748229831457, + 0.004694387782365084, + 0.0021590159740298986, + 0.0075250002555549145, + 0.026767367497086525, + 0.013509614393115044, + 0.03258196637034416, + 0.013470510952174664, + 0.00741468695923686, + 0.014471299946308136, + 0.007744777947664261, + 0.003244943916797638, + 0.009499040432274342, + 0.0022458182647824287, + 0.00367892999202013, + 0.007308382075279951, + 0.002229115227237344, + 0.0039396644569933414, + 0.001906600664369762, + 0.002013338031247258, + 0.002625759458169341, + 0.001946027623489499, + 0.0026904400438070297, + 0.002109709894284606, + 0.0015879435231909156, + 0.0022088871337473392, + 0.0016111376462504268 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.009791442193090916 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20157462358474731 + ], + "opposite_logprob_tokens": [ + -1.7015745639801025 + ], + "logprob_reward": -0.20157462358474731, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20157462358474731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08730850368738174, + 0.02636558935046196, + 0.02729293331503868, + 0.020157182589173317, + 0.3187110424041748, + 0.04044974595308304, + 0.3683711588382721, + 0.03202606737613678, + 0.04596063122153282, + 0.03335711732506752 + ], + "credit_map": [ + 0.010516983456909657, + 0.009693257510662079, + 0.02044319175183773, + 0.005846490152180195, + 0.006156164687126875, + 0.012776192277669907, + 0.007005779538303614, + 0.014870445244014263, + 0.003112882375717163, + 0.002947197062894702, + 0.0028173315804451704, + 0.004399165976792574, + 0.0025963461957871914, + 0.0032508750446140766, + 0.0037924321368336678, + 0.0034493589773774147, + 0.0032030486036092043, + 0.0028021852485835552, + 0.003008177736774087, + 0.00446087121963501, + 0.0032178754918277264, + 0.004104907158762217, + 0.0031605453696101904, + 0.0033353231847286224, + 0.0015833729412406683, + 0.0044141472317278385, + 0.0026456820778548717, + 0.0034015486016869545, + 0.0023773889988660812, + 0.0019447625381872058, + 0.0025559451896697283, + 0.001234334777109325, + 0.014059243723750114, + 0.02545854076743126, + 0.06595027446746826, + 0.04579794779419899, + 0.07524377852678299, + 0.07275765389204025, + 0.007762285880744457, + 0.011681296862661839, + 0.004035611636936665, + 0.005771483760327101, + 0.0046951184049248695, + 0.007348199374973774, + 0.004005142487585545, + 0.004466508515179157, + 0.006602638866752386, + 0.0035250403452664614, + 0.017303667962551117, + 0.13087481260299683, + 0.010789805091917515, + 0.1328008621931076, + 0.02665502019226551, + 0.015756845474243164, + 0.022531433030962944, + 0.011658703908324242, + 0.0023786320816725492, + 0.006750083528459072, + 0.0022528315894305706, + 0.006901632063090801, + 0.003942267037928104, + 0.002511007944121957, + 0.004217620473355055, + 0.0030719919595867395, + 0.0047168671153485775, + 0.007075417321175337, + 0.005089932586997747, + 0.007125578820705414, + 0.0065903388895094395, + 0.005599931813776493, + 0.004967559594660997, + 0.004795004613697529, + 0.003581710858270526, + 0.004408047068864107, + 0.003793436335399747, + 0.004141516517847776, + 0.0045118327252566814, + 0.003975145984441042, + 0.004509045742452145, + 0.004436382092535496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.20157462358474731 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up shot, the camera held still, focuses on a pot of clear, scalding hot water, where a square, dry, and firm, pale yellow instant noodle cake quietly soaks. A soft, warm light from above penetrates the rising white steam, illuminating the entire scene and clearly revealing the wavy texture pressed into the noodle cake. As the water seeps in, the edges of the noodle cake begin to round out, and the curled noodles on the outer layer begin to loosen, delicately unfolding like petals. Then, the entire noodle cake steadily and leisurely expands from the inside out. The once tightly packed, wavy noodles break free, floating gracefully and unraveling in the water, resuming their natural curls. Tiny bubbles continuously emerge from the bottom of the pot, clinging to the increasingly translucent noodles, whose moist surfaces reflect an alluring luster in the light. The slow and rhythmic process creates a warm, soothing, and deeply satisfying tranquil atmosphere.", + "seed": 1461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1461_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11224160068668425, + "vlm_logprob": -0.11224160068668425, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.11224160068668425, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062819704413414 + ], + "opposite_logprob_tokens": [ + -3.062819719314575 + ], + "logprob_reward": -0.062819704413414, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.062819704413414, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04504667967557907, + 0.02972874604165554, + 0.028877150267362595, + 0.03302520141005516, + 0.14793473482131958, + 0.2817969024181366, + 0.09426070004701614, + 0.2337362915277481, + 0.03182361274957657, + 0.07376999408006668 + ], + "credit_map": [ + 0.008359890431165695, + 0.005051565822213888, + 0.004728854168206453, + 0.005209040828049183, + 0.005505390930920839, + 0.0038111491594463587, + 0.007181130815297365, + 0.005199657287448645, + 0.006052343174815178, + 0.003772163297981024, + 0.0041822465136647224, + 0.0025678451638668776, + 0.002353884745389223, + 0.0046789199113845825, + 0.0026345138903707266, + 0.003486830275505781, + 0.006383172702044249, + 0.003156286431476474, + 0.004698517266660929, + 0.002728569321334362, + 0.0022430114913731813, + 0.0027605840004980564, + 0.0032630469650030136, + 0.003643963485956192, + 0.006208079867064953, + 0.0037981169298291206, + 0.0035123596899211407, + 0.0045057907700538635, + 0.003159155370667577, + 0.0027146812062710524, + 0.005232175346463919, + 0.0038948410656303167, + 0.013089417479932308, + 0.007423492148518562, + 0.016764257103204727, + 0.00744736660271883, + 0.009994331747293472, + 0.051292020827531815, + 0.008211025036871433, + 0.03371281549334526, + 0.02012990601360798, + 0.01735367253422737, + 0.017434068024158478, + 0.06642588973045349, + 0.024835415184497833, + 0.011913419701159, + 0.04644639417529106, + 0.07725813239812851, + 0.01749926619231701, + 0.010566442273557186, + 0.008047758601605892, + 0.013380181044340134, + 0.011956973001360893, + 0.008658850565552711, + 0.015054735355079174, + 0.00909650418907404, + 0.046207476407289505, + 0.02169002592563629, + 0.03796079754829407, + 0.027565600350499153, + 0.01541275903582573, + 0.015019677579402924, + 0.051763102412223816, + 0.018116865307092667, + 0.0055847796611487865, + 0.0042915488593280315, + 0.004560815170407295, + 0.0031324299052357674, + 0.0028073671273887157, + 0.0027143468614667654, + 0.0042799655348062515, + 0.004452358465641737, + 0.007150474935770035, + 0.02420758083462715, + 0.004632533527910709, + 0.007924087345600128, + 0.00930805504322052, + 0.0061376141384243965, + 0.008641766384243965, + 0.005767889320850372 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.062819704413414 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002229587407782674 + ], + "opposite_logprob_tokens": [ + -6.127229690551758 + ], + "logprob_reward": -0.002229587407782674, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002229587407782674, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05434190481901169, + 0.034847065806388855, + 0.033584944903850555, + 0.03496720269322395, + 0.15573853254318237, + 0.277425616979599, + 0.12650388479232788, + 0.1924317628145218, + 0.049870532006025314, + 0.04028860107064247 + ], + "credit_map": [ + 0.011032612062990665, + 0.006466479506343603, + 0.008034897036850452, + 0.004944539628922939, + 0.005813136231154203, + 0.00666472315788269, + 0.005982627160847187, + 0.005402885377407074, + 0.008846638724207878, + 0.0049695768393576145, + 0.00559123745188117, + 0.003090373007580638, + 0.002807340817525983, + 0.0033934651874005795, + 0.003130841301754117, + 0.0030175899155437946, + 0.004291366320103407, + 0.003660216461867094, + 0.003227327251806855, + 0.0029281415045261383, + 0.0032076414208859205, + 0.004489688668400049, + 0.0050208671018481255, + 0.006759691052138805, + 0.006072648800909519, + 0.003807233413681388, + 0.00401662290096283, + 0.004432591609656811, + 0.003240640740841627, + 0.0035014073364436626, + 0.005912883672863245, + 0.003983170259743929, + 0.013315767981112003, + 0.0072582741267979145, + 0.016346346586942673, + 0.007370316423475742, + 0.010926901362836361, + 0.05457857996225357, + 0.009501676075160503, + 0.0364406518638134, + 0.01997649483382702, + 0.017795421183109283, + 0.018760617822408676, + 0.0687464103102684, + 0.019778897985816002, + 0.012779333628714085, + 0.0406591035425663, + 0.07892931997776031, + 0.012948969379067421, + 0.008159272372722626, + 0.00887158140540123, + 0.017823351547122, + 0.017840053886175156, + 0.010203458368778229, + 0.024341344833374023, + 0.026315838098526, + 0.03858627751469612, + 0.016069535166025162, + 0.03093000501394272, + 0.02339109405875206, + 0.012124243192374706, + 0.012630950659513474, + 0.041893381625413895, + 0.01680626906454563, + 0.006140844896435738, + 0.007579997181892395, + 0.004516443703323603, + 0.005292331799864769, + 0.008243833668529987, + 0.004956502001732588, + 0.005860448349267244, + 0.007280123420059681, + 0.005381316412240267, + 0.007322062738239765, + 0.004456697963178158, + 0.006197643466293812, + 0.004308277741074562, + 0.002968068700283766, + 0.005718861240893602, + 0.003935664892196655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 4 + ], + "selected_reward": -0.002229587407782674 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646339945495129 + ], + "opposite_logprob_tokens": [ + -4.882646560668945 + ], + "logprob_reward": -0.007646339945495129, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646339945495129, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09185905009508133, + 0.038321711122989655, + 0.04020309075713158, + 0.04436187446117401, + 0.05975892394781113, + 0.18487678468227386, + 0.2225726842880249, + 0.07946277409791946, + 0.14217287302017212, + 0.09641029685735703 + ], + "credit_map": [ + 0.016341567039489746, + 0.013225752860307693, + 0.00971295963972807, + 0.009219883009791374, + 0.013119958341121674, + 0.009823049418628216, + 0.01227527391165495, + 0.00814060028642416, + 0.005904542747884989, + 0.0038788130041211843, + 0.0038680348079651594, + 0.004408295266330242, + 0.004786637611687183, + 0.003435828723013401, + 0.008254574611783028, + 0.0037849820218980312, + 0.0056238928809762, + 0.005577778443694115, + 0.0034186942502856255, + 0.00515550933778286, + 0.004549652338027954, + 0.005133109167218208, + 0.006217551417648792, + 0.004526905249804258, + 0.005869960878044367, + 0.005058294162154198, + 0.005379281472414732, + 0.004934382624924183, + 0.006781809497624636, + 0.003809457179158926, + 0.005022278055548668, + 0.007506407331675291, + 0.009299220517277718, + 0.00805795006453991, + 0.008725353516638279, + 0.009443125687539577, + 0.005800782237201929, + 0.0037035965360701084, + 0.009536386467516422, + 0.005192508455365896, + 0.013384181074798107, + 0.011779993772506714, + 0.01203250978142023, + 0.04592886194586754, + 0.014996289275586605, + 0.008650582283735275, + 0.027084561064839363, + 0.051019810140132904, + 0.028619488701224327, + 0.03161623328924179, + 0.014202730730175972, + 0.025434274226427078, + 0.044747136533260345, + 0.016668178141117096, + 0.045034147799015045, + 0.016250498592853546, + 0.015033651143312454, + 0.008872677572071552, + 0.01147159468382597, + 0.008625437505543232, + 0.006176853086799383, + 0.006670344155281782, + 0.009509381838142872, + 0.01310284249484539, + 0.019260717555880547, + 0.017156582325696945, + 0.015829619020223618, + 0.015179614536464214, + 0.01635979861021042, + 0.02046082355082035, + 0.017483387142419815, + 0.020442333072423935, + 0.01251958031207323, + 0.012226111255586147, + 0.013873377814888954, + 0.01466295588761568, + 0.0110393725335598, + 0.009594045579433441, + 0.012049762532114983, + 0.010445088148117065 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.007646339945495129 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014203761704266071 + ], + "opposite_logprob_tokens": [ + -4.2642035484313965 + ], + "logprob_reward": -0.014203761704266071, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014203761704266071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15066201984882355, + 0.09994106739759445, + 0.1006406620144844, + 0.06864418834447861, + 0.09820274263620377, + 0.10154351592063904, + 0.1841438114643097, + 0.07348316162824631, + 0.0477604866027832, + 0.0749783143401146 + ], + "credit_map": [ + 0.031538836658000946, + 0.021535541862249374, + 0.022790154442191124, + 0.016277790069580078, + 0.014670771546661854, + 0.014259948395192623, + 0.015475710853934288, + 0.014113285578787327, + 0.014891757629811764, + 0.008180085569620132, + 0.01160876452922821, + 0.00899073202162981, + 0.007669028360396624, + 0.023644650354981422, + 0.010592669248580933, + 0.014363386668264866, + 0.020753510296344757, + 0.013120065443217754, + 0.007383390795439482, + 0.011267458088696003, + 0.008630367927253246, + 0.007478710729628801, + 0.01695305109024048, + 0.015054110437631607, + 0.016642816364765167, + 0.00832809042185545, + 0.009042616002261639, + 0.006697844248265028, + 0.006152333691716194, + 0.004711496643722057, + 0.009601769037544727, + 0.007467231247574091, + 0.01167207583785057, + 0.0067218877375125885, + 0.012288319878280163, + 0.005807503126561642, + 0.008139810524880886, + 0.02931215427815914, + 0.007816953584551811, + 0.016444042325019836, + 0.016882695257663727, + 0.0066521381959319115, + 0.010857905261218548, + 0.01766456663608551, + 0.007636440452188253, + 0.00691814161837101, + 0.011830413714051247, + 0.02310122922062874, + 0.01839342713356018, + 0.02850222960114479, + 0.01379157043993473, + 0.025590533390641212, + 0.019358379766345024, + 0.01070103794336319, + 0.04709170013666153, + 0.020714962854981422, + 0.0115518718957901, + 0.01037127897143364, + 0.012974223122000694, + 0.008980320766568184, + 0.006629385054111481, + 0.005684391595423222, + 0.012337815947830677, + 0.0049538761377334595, + 0.005602988414466381, + 0.005960934795439243, + 0.006205123849213123, + 0.006164927501231432, + 0.0049551245756447315, + 0.005422885529696941, + 0.008376703597605228, + 0.005071807187050581, + 0.009738370776176453, + 0.009669107384979725, + 0.006690613925457001, + 0.012633766047656536, + 0.008587010204792023, + 0.007375620771199465, + 0.012246331200003624, + 0.008037499152123928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 5 + ], + "selected_reward": -0.014203761704266071 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4743086099624634 + ], + "opposite_logprob_tokens": [ + -0.9743086099624634 + ], + "logprob_reward": -0.4743086099624634, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4743086099624634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06338338553905487, + 0.019658749923110008, + 0.06049244478344917, + 0.036332543939352036, + 0.08811318129301071, + 0.08438899368047714, + 0.19719304144382477, + 0.20105156302452087, + 0.15993350744247437, + 0.08945253491401672 + ], + "credit_map": [ + 0.011585863307118416, + 0.006160007789731026, + 0.008184967562556267, + 0.005521236918866634, + 0.007704508490860462, + 0.007663807366043329, + 0.00958265271037817, + 0.006980345584452152, + 0.0031259432435035706, + 0.0016161331441253424, + 0.0022101467475295067, + 0.002036317950114608, + 0.0021096242126077414, + 0.0023836218751966953, + 0.003631707513704896, + 0.002545255236327648, + 0.012399650178849697, + 0.009781697764992714, + 0.006721543148159981, + 0.00601133331656456, + 0.0060210516676306725, + 0.00516918022185564, + 0.008703061379492283, + 0.005684928502887487, + 0.008143415674567223, + 0.0025763516314327717, + 0.004065028391778469, + 0.0025177139323204756, + 0.0038817659951746464, + 0.004115432035177946, + 0.0059460224583745, + 0.005086815450340509, + 0.009571544826030731, + 0.004760505631566048, + 0.009952874854207039, + 0.005238196812570095, + 0.006695759017020464, + 0.02890685945749283, + 0.005857993382960558, + 0.017129456624388695, + 0.008172493427991867, + 0.006444324739277363, + 0.004791990388184786, + 0.020805343985557556, + 0.005505886394530535, + 0.005080968141555786, + 0.01053776778280735, + 0.023050224408507347, + 0.021253393962979317, + 0.020891420543193817, + 0.014116505160927773, + 0.021442700177431107, + 0.033865850418806076, + 0.015212311409413815, + 0.03217081353068352, + 0.03824004903435707, + 0.04671506956219673, + 0.021720346063375473, + 0.03182431310415268, + 0.026430413126945496, + 0.01492086611688137, + 0.014469203539192677, + 0.02924063429236412, + 0.0157307256013155, + 0.022000955417752266, + 0.02335335686802864, + 0.01889633946120739, + 0.022869883105158806, + 0.013756009750068188, + 0.02312283217906952, + 0.018152866512537003, + 0.01778126321732998, + 0.011019298806786537, + 0.011046755127608776, + 0.013989518396556377, + 0.014289957471191883, + 0.00878115277737379, + 0.009605323895812035, + 0.011669493280351162, + 0.009051037952303886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 8 + ], + "selected_reward": -0.4743086099624634 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning macro close-up shot, the camera held still, focuses on a pot of clear, scalding hot water, where a square, dry, and firm, pale yellow instant noodle cake quietly soaks. A soft, warm light from above penetrates the rising white steam, illuminating the entire scene and clearly revealing the wavy texture pressed into the noodle cake. As the water seeps in, the edges of the noodle cake begin to round out, and the curled noodles on the outer layer begin to loosen, delicately unfolding like petals. Then, the entire noodle cake steadily and leisurely expands from the inside out. The once tightly packed, wavy noodles break free, floating gracefully and unraveling in the water, resuming their natural curls. Tiny bubbles continuously emerge from the bottom of the pot, clinging to the increasingly translucent noodles, whose moist surfaces reflect an alluring luster in the light. The slow and rhythmic process creates a warm, soothing, and deeply satisfying tranquil atmosphere.", + "seed": 1462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_10/index_1392_seed_1462_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11783189829438925, + "vlm_logprob": -0.11783189829438925, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.11783189829438925, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04453849047422409 + ], + "opposite_logprob_tokens": [ + -3.6695384979248047 + ], + "logprob_reward": -0.04453849047422409, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.04453849047422409, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21010564267635345, + 0.06467082351446152, + 0.023856205865740776, + 0.02645506151020527, + 0.019495364278554916, + 0.016903767362236977, + 0.039417073130607605, + 0.22230443358421326, + 0.13390395045280457, + 0.2428877204656601 + ], + "credit_map": [ + 0.030229631811380386, + 0.07345768064260483, + 0.01074205432087183, + 0.04185381159186363, + 0.013345838524401188, + 0.015860719606280327, + 0.016224661841988564, + 0.008391248993575573, + 0.007730762008577585, + 0.01584387756884098, + 0.005170417483896017, + 0.010603122413158417, + 0.007413098588585854, + 0.004927145782858133, + 0.009101306088268757, + 0.0038810966070741415, + 0.0035882769152522087, + 0.003319297218695283, + 0.0019614556804299355, + 0.003123244736343622, + 0.0030429065227508545, + 0.0018805473810061812, + 0.005034813657402992, + 0.0019056635210290551, + 0.004281464032828808, + 0.004753207787871361, + 0.002679027384147048, + 0.002886898349970579, + 0.0031520838383585215, + 0.0021589442621916533, + 0.004012004006654024, + 0.002531431382521987, + 0.0028184116818010807, + 0.003507190616801381, + 0.001790571608580649, + 0.0026499175000935793, + 0.0031038785818964243, + 0.0015473078237846494, + 0.0024678311310708523, + 0.0016102552181109786, + 0.002660056808963418, + 0.001994631253182888, + 0.0015203077346086502, + 0.0018992744153365493, + 0.0025745443999767303, + 0.0018165612127631903, + 0.0025713439099490643, + 0.0018670480931177735, + 0.0050994898192584515, + 0.0032380251213908195, + 0.005917904898524284, + 0.004412275273352861, + 0.007384839933365583, + 0.0027335418853908777, + 0.006652377545833588, + 0.003978620283305645, + 0.05197546258568764, + 0.02822674624621868, + 0.018753185868263245, + 0.024029474705457687, + 0.01958790235221386, + 0.02153647132217884, + 0.03401067480444908, + 0.024184515699744225, + 0.011249805800616741, + 0.023319266736507416, + 0.008697452023625374, + 0.01623496413230896, + 0.04010305926203728, + 0.009368070401251316, + 0.014265679754316807, + 0.010665652342140675, + 0.014540274627506733, + 0.023221580311655998, + 0.006760085001587868, + 0.017883997410535812, + 0.08689317107200623, + 0.011627374216914177, + 0.07300995290279388, + 0.008951280266046524 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.04453849047422409 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025158440694212914 + ], + "opposite_logprob_tokens": [ + -6.00251579284668 + ], + "logprob_reward": -0.0025158440694212914, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025158440694212914, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49299782514572144, + 0.16344839334487915, + 0.05752900615334511, + 0.07751572132110596, + 0.03246195241808891, + 0.0127986129373312, + 0.016066409647464752, + 0.04188210144639015, + 0.08823567628860474, + 0.01706428825855255 + ], + "credit_map": [ + 0.07923523336648941, + 0.18191207945346832, + 0.017846662551164627, + 0.09704972058534622, + 0.028818484395742416, + 0.03683899715542793, + 0.0328407883644104, + 0.018455902114510536, + 0.026407752186059952, + 0.06183820590376854, + 0.006204612087458372, + 0.02496539242565632, + 0.011443491093814373, + 0.014407558366656303, + 0.010353056713938713, + 0.007828333415091038, + 0.006580692250281572, + 0.007666232530027628, + 0.004644433967769146, + 0.010973182506859303, + 0.006090956274420023, + 0.0042887344025075436, + 0.012406203895807266, + 0.004878578241914511, + 0.011145727708935738, + 0.007402463816106319, + 0.005916791968047619, + 0.015165610238909721, + 0.009741823188960552, + 0.007127093151211739, + 0.014961347915232182, + 0.006054877303540707, + 0.008987167850136757, + 0.0034271006006747484, + 0.002483414253219962, + 0.0029930383898317814, + 0.0029349657706916332, + 0.007221920415759087, + 0.0022223584819585085, + 0.0021919875871390104, + 0.0013202977133914828, + 0.001248104264959693, + 0.0009492349345237017, + 0.0019119164207950234, + 0.0020805990789085627, + 0.0015320860547944903, + 0.0028815334662795067, + 0.0008748424588702619, + 0.0022591527085751295, + 0.0015477763954550028, + 0.0017347915563732386, + 0.0024416502565145493, + 0.0030465342570096254, + 0.0012097940780222416, + 0.0025395762640982866, + 0.0012871340150013566, + 0.0093461275100708, + 0.007049954496324062, + 0.002952979877591133, + 0.004436136689037085, + 0.003985424060374498, + 0.0038199189584702253, + 0.0061686281114816666, + 0.004122937563806772, + 0.007472269237041473, + 0.01068347878754139, + 0.0028681354597210884, + 0.017353178933262825, + 0.033480025827884674, + 0.00485921511426568, + 0.007907746359705925, + 0.003611629828810692, + 0.001979343593120575, + 0.0024469110649079084, + 0.0012446909677237272, + 0.002727050567045808, + 0.0038552582263946533, + 0.001319881877861917, + 0.001976526342332363, + 0.0015146259684115648 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0025158440694212914 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007646103389561176 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.007646103389561176, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007646103389561176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5437334179878235, + 0.08316903561353683, + 0.016912709921598434, + 0.016386572271585464, + 0.020353052765130997, + 0.01417132094502449, + 0.03487348556518555, + 0.12372507154941559, + 0.048223208636045456, + 0.0984521135687828 + ], + "credit_map": [ + 0.10258805006742477, + 0.22280070185661316, + 0.011770108714699745, + 0.08986539393663406, + 0.033028777688741684, + 0.0517745241522789, + 0.01809733919799328, + 0.013808554969727993, + 0.011531258001923561, + 0.019473817199468613, + 0.00896140281111002, + 0.00895286537706852, + 0.010347826406359673, + 0.00900181196630001, + 0.008382037281990051, + 0.0065180095843970776, + 0.003767055692151189, + 0.002612051088362932, + 0.0014837911585345864, + 0.0017308358801528811, + 0.001824287697672844, + 0.0014554300578311086, + 0.002512613544240594, + 0.001526643754914403, + 0.0019392161630094051, + 0.0017041197279468179, + 0.0015435314271599054, + 0.0030455824453383684, + 0.002265795599669218, + 0.0012715223710983992, + 0.0032110940665006638, + 0.001405709539540112, + 0.005358047783374786, + 0.00207234057597816, + 0.0015299671795219183, + 0.002629858674481511, + 0.0016667441232129931, + 0.003991615958511829, + 0.0017956577939912677, + 0.0013088197447359562, + 0.001914381980895996, + 0.0019878658931702375, + 0.0015255971811711788, + 0.0013360026059672236, + 0.002030805451795459, + 0.0016928473487496376, + 0.002489125356078148, + 0.001194694428704679, + 0.0039054027292877436, + 0.003770496929064393, + 0.004753102082759142, + 0.0036336134653538465, + 0.005190475843846798, + 0.0024138418957591057, + 0.007411445025354624, + 0.003795105265453458, + 0.02981950156390667, + 0.013819358311593533, + 0.011773278005421162, + 0.01304792519658804, + 0.011068161576986313, + 0.012576204724609852, + 0.01735745184123516, + 0.014263186603784561, + 0.0035195776727050543, + 0.006687016226351261, + 0.0022978337947279215, + 0.006614271085709333, + 0.015976693481206894, + 0.003265013685449958, + 0.005558881442993879, + 0.0043039144948124886, + 0.00880979560315609, + 0.009783243760466576, + 0.007412196137011051, + 0.01679891347885132, + 0.02739202231168747, + 0.007165241055190563, + 0.013985923491418362, + 0.007104774471372366 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.007646103389561176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081933043897152 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081933043897152, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081933043897152, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30012986063957214, + 0.09293655306100845, + 0.03930095583200455, + 0.024320557713508606, + 0.04840246960520744, + 0.027241280302405357, + 0.028715185821056366, + 0.0875684842467308, + 0.23891034722328186, + 0.11247418820858002 + ], + "credit_map": [ + 0.051010582596063614, + 0.111319400370121, + 0.01147475466132164, + 0.045105718076229095, + 0.02006717212498188, + 0.033785562962293625, + 0.013560611754655838, + 0.013806081376969814, + 0.012914339080452919, + 0.02576058730483055, + 0.006628667004406452, + 0.013236274011433125, + 0.007202454376965761, + 0.0061553060077130795, + 0.01358641404658556, + 0.007452522870153189, + 0.01020616665482521, + 0.005906395148485899, + 0.0036325876135379076, + 0.004353254102170467, + 0.003517239121720195, + 0.003041947493329644, + 0.004789050202816725, + 0.0038543203845620155, + 0.0030867510940879583, + 0.0028694949578493834, + 0.001993643119931221, + 0.0024558117147535086, + 0.005427293013781309, + 0.002620929153636098, + 0.0035004899837076664, + 0.0023661472368985415, + 0.01368831004947424, + 0.005547300446778536, + 0.0032184163574129343, + 0.004507812671363354, + 0.004417872987687588, + 0.010158947668969631, + 0.0032874394673854113, + 0.003576375776901841, + 0.004765650723129511, + 0.004023435525596142, + 0.0024268063716590405, + 0.0026001259684562683, + 0.0037483512423932552, + 0.00273837149143219, + 0.003809743793681264, + 0.0031287975143641233, + 0.003250172594562173, + 0.0027631395496428013, + 0.004041191656142473, + 0.0034711607731878757, + 0.003791186725720763, + 0.0017872945172712207, + 0.0060477606020867825, + 0.0035632820799946785, + 0.014107976108789444, + 0.013613112270832062, + 0.008384059183299541, + 0.010795933194458485, + 0.007512889336794615, + 0.007985790260136127, + 0.015461234375834465, + 0.009707508608698845, + 0.02036304399371147, + 0.02472696080803871, + 0.0092151565477252, + 0.04667852446436882, + 0.0854748860001564, + 0.013498423621058464, + 0.026470575481653214, + 0.012482807971537113, + 0.00943627767264843, + 0.01212148368358612, + 0.006668475456535816, + 0.01057224441319704, + 0.03713739290833473, + 0.005627891514450312, + 0.021275639533996582, + 0.00963479932397604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.011081933043897152 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a macro close-up of a pot containing clear, scalding hot water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the initial object a square, dry, firm, pale yellow instant noodle cake?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the noodle cake expand from the inside out while the outer noodles loosen and unravel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny bubbles continuously emerging from the bottom of the pot and clinging to the noodles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held still throughout the scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233771204948425 + ], + "opposite_logprob_tokens": [ + -0.8983771204948425 + ], + "logprob_reward": -0.5233771204948425, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233771204948425, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3511020839214325, + 0.20307783782482147, + 0.04529903456568718, + 0.07269569486379623, + 0.01874527893960476, + 0.02033282071352005, + 0.04422562196850777, + 0.11075157672166824, + 0.06878457218408585, + 0.06498542428016663 + ], + "credit_map": [ + 0.06859845668077469, + 0.14349554479122162, + 0.011077490635216236, + 0.05302901193499565, + 0.023536883294582367, + 0.028220944106578827, + 0.011771619319915771, + 0.011372140608727932, + 0.030147094279527664, + 0.05722256749868393, + 0.018021155148744583, + 0.0215573962777853, + 0.024036703631281853, + 0.023262780159711838, + 0.015075860545039177, + 0.01375427097082138, + 0.0055006276816129684, + 0.00569709949195385, + 0.004895127844065428, + 0.006380667444318533, + 0.0067596640437841415, + 0.004063998814672232, + 0.006413686089217663, + 0.0055881645530462265, + 0.012933369725942612, + 0.011588099412620068, + 0.011589332483708858, + 0.005858464166522026, + 0.011158522218465805, + 0.006370878778398037, + 0.005296020768582821, + 0.007901007309556007, + 0.0035565148573368788, + 0.002603651722893119, + 0.0013173482147976756, + 0.00344611844047904, + 0.0024351554457098246, + 0.0013909967383369803, + 0.0020685268100351095, + 0.001926966244354844, + 0.003183644963428378, + 0.0035560429096221924, + 0.002381294732913375, + 0.0019070389680564404, + 0.0023240193258970976, + 0.0024554000701755285, + 0.0030479624401777983, + 0.0014774188166484237, + 0.0051253619603812695, + 0.004572228994220495, + 0.005106461700052023, + 0.006566320080310106, + 0.007939141243696213, + 0.0030831373296678066, + 0.007800074759870768, + 0.004032897297292948, + 0.027627388015389442, + 0.012459179386496544, + 0.010337982326745987, + 0.011053177528083324, + 0.009259063750505447, + 0.010756680741906166, + 0.01469537615776062, + 0.014562726020812988, + 0.006229553837329149, + 0.00634531956166029, + 0.0033078088890761137, + 0.013263994827866554, + 0.02215449884533882, + 0.004528875928372145, + 0.007950336672365665, + 0.0050041829235851765, + 0.006137627176940441, + 0.006367571651935577, + 0.004939517006278038, + 0.009165112860500813, + 0.016426347196102142, + 0.00511247618123889, + 0.010409870184957981, + 0.0064269015565514565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.5233771204948425 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_tensor([4105, 4105])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_tensor([4105, 4105])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..804f7fea6e70a131de15d0a2beb04510ea196b40 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_tensor([4105, 4105])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "With the camera held perfectly still, the extreme macro shot focuses with microscopic clarity on the sterile, polished stainless steel tip of a hypodermic needle. Casting a sharp, well-defined highlight under a bright clinical light source, a shimmering drop of perfectly transparent liquid slowly and deliberately seeps from the needle's angled opening. The liquid expands, its surface tension creating a perfect, trembling sphere that clings precariously to the metal, refracting light like a tiny crystal lens. The background is a completely blurred, soft-focus, cool blue and white, suggesting a laboratory setting but revealing no specific details. The overall atmosphere is one of intense concentration, quiet anticipation, and tedious precision, as the weight of the drop grows heavier, seemingly at the very moment of its impact.", + "seed": 12587, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_4105_seed_12587_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.11546832044841722, + "vlm_logprob": -0.11546832044841722, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.11546832044841722, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008471623994410038 + ], + "opposite_logprob_tokens": [ + -5.258471488952637 + ], + "logprob_reward": -0.008471623994410038, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008471623994410038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18980681896209717, + 0.060739848762750626, + 0.03764024004340172, + 0.030487697571516037, + 0.041711267083883286, + 0.0283666905015707, + 0.048374369740486145, + 0.18347428739070892, + 0.17984436452388763, + 0.19955438375473022 + ], + "credit_map": [ + 0.01904534362256527, + 0.0258344616740942, + 0.03061983920633793, + 0.020967742428183556, + 0.022851040586829185, + 0.014730231836438179, + 0.02677888423204422, + 0.02897927537560463, + 0.007095927372574806, + 0.00903798546642065, + 0.006376832257956266, + 0.00671376520767808, + 0.006961564067751169, + 0.005416777450591326, + 0.008238382637500763, + 0.01089861523360014, + 0.003899587085470557, + 0.004744083620607853, + 0.003230949630960822, + 0.005580435506999493, + 0.004796657711267471, + 0.003064369549974799, + 0.00678536482155323, + 0.005538790952414274, + 0.0033388661686331034, + 0.005187229253351688, + 0.004670594818890095, + 0.0030982245225459337, + 0.0038517967332154512, + 0.0030203538481146097, + 0.003792017698287964, + 0.0035286122001707554, + 0.0045503610745072365, + 0.00437554158270359, + 0.004614919889718294, + 0.005661656614392996, + 0.005169034469872713, + 0.0043755085207521915, + 0.0061000194400548935, + 0.006864224560558796, + 0.004143265541642904, + 0.004462975077331066, + 0.0033418063540011644, + 0.00222686305642128, + 0.0028068337123841047, + 0.002611703472211957, + 0.0032574390061199665, + 0.00551580497995019, + 0.007341581862419844, + 0.005531097762286663, + 0.004705243278294802, + 0.003789770184084773, + 0.0076131559908390045, + 0.004395819269120693, + 0.007894127629697323, + 0.007103572133928537, + 0.0211492907255888, + 0.02391893044114113, + 0.025364980101585388, + 0.022016339004039764, + 0.022281954064965248, + 0.02048776112496853, + 0.02753419615328312, + 0.020720822736620903, + 0.01820618100464344, + 0.02114059217274189, + 0.026918035000562668, + 0.024445101618766785, + 0.022732721641659737, + 0.021235667169094086, + 0.02550826221704483, + 0.019657794386148453, + 0.02246035449206829, + 0.02376409061253071, + 0.030853893607854843, + 0.0285404734313488, + 0.02110316976904869, + 0.014931404031813145, + 0.03287624567747116, + 0.02502473071217537 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.008471623994410038 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018197957426309586 + ], + "opposite_logprob_tokens": [ + -4.018198013305664 + ], + "logprob_reward": -0.018197957426309586, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018197957426309586, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09488404542207718, + 0.08428849279880524, + 0.04424002021551132, + 0.04790780693292618, + 0.034140586853027344, + 0.03393949195742607, + 0.05265273526310921, + 0.02241579443216324, + 0.27505049109458923, + 0.310480535030365 + ], + "credit_map": [ + 0.010656784288585186, + 0.011511052027344704, + 0.00904290284961462, + 0.015630466863512993, + 0.010014045983552933, + 0.007253497838973999, + 0.014533648267388344, + 0.0162416510283947, + 0.011683107353746891, + 0.011167142540216446, + 0.008149396628141403, + 0.009448201395571232, + 0.009834584780037403, + 0.0073578376322984695, + 0.013026963919401169, + 0.013621269725263119, + 0.0035365025978535414, + 0.0060360427014529705, + 0.008134153671562672, + 0.005706729367375374, + 0.004864368587732315, + 0.0038752760738134384, + 0.0057593220844864845, + 0.006327626761049032, + 0.0066394018940627575, + 0.006141537334769964, + 0.006572813726961613, + 0.005387517623603344, + 0.0049415766261518, + 0.004233615938574076, + 0.006206224672496319, + 0.0077851261012256145, + 0.0035657871048897505, + 0.005181056447327137, + 0.004634101875126362, + 0.004877912346273661, + 0.004460636992007494, + 0.0034185757394880056, + 0.0046538240276277065, + 0.003348695347085595, + 0.004074747208505869, + 0.004342393483966589, + 0.003924074582755566, + 0.00359728978946805, + 0.004400057718157768, + 0.0035200004931539297, + 0.00514467665925622, + 0.004936254117637873, + 0.010308505967259407, + 0.00822825450450182, + 0.005056603811681271, + 0.0041824281215667725, + 0.005784332286566496, + 0.005038536619395018, + 0.006618221290409565, + 0.007435860112309456, + 0.005282782483845949, + 0.0034887942019850016, + 0.0025721671991050243, + 0.0019158952636644244, + 0.002181493677198887, + 0.001990132499486208, + 0.001686297357082367, + 0.0032982337288558483, + 0.03154165297746658, + 0.03461965173482895, + 0.043085914105176926, + 0.03626585379242897, + 0.03209901601076126, + 0.033514201641082764, + 0.036785196512937546, + 0.027139022946357727, + 0.03247850015759468, + 0.03857318311929703, + 0.052520498633384705, + 0.04693872109055519, + 0.030826671048998833, + 0.029273871332406998, + 0.04756993055343628, + 0.032299164682626724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018197957426309586 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026366863399744034 + ], + "opposite_logprob_tokens": [ + -3.651366949081421 + ], + "logprob_reward": -0.026366863399744034, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026366863399744034, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.055624835193157196, + 0.026820870116353035, + 0.01325277891010046, + 0.017002347856760025, + 0.007353212218731642, + 0.043056901544332504, + 0.027279658243060112, + 0.07771354168653488, + 0.12240631133317947, + 0.6094894409179688 + ], + "credit_map": [ + 0.006994177587330341, + 0.006103662773966789, + 0.008518621325492859, + 0.005231558810919523, + 0.005650724284350872, + 0.004600392654538155, + 0.00923824217170477, + 0.009287461638450623, + 0.002779741305857897, + 0.0025982919614762068, + 0.0027489650528877974, + 0.002932096365839243, + 0.002690274966880679, + 0.002342461608350277, + 0.005249947775155306, + 0.005479094106703997, + 0.0012361694825813174, + 0.001423829817213118, + 0.001404639333486557, + 0.0014129923656582832, + 0.0015706211561337113, + 0.001401272602379322, + 0.001987622817978263, + 0.0028156328480690718, + 0.0015035874675959349, + 0.0020858163479715586, + 0.0021973676048219204, + 0.0020422658417373896, + 0.002135166898369789, + 0.0016671812627464533, + 0.003211753210052848, + 0.002159210853278637, + 0.0009243662352673709, + 0.0008261410403065383, + 0.0007840840262360871, + 0.0006637283368036151, + 0.0008363425731658936, + 0.0008965350571088493, + 0.0009827129542827606, + 0.0014393026940524578, + 0.004844220355153084, + 0.004312088247388601, + 0.0047180685214698315, + 0.004446518607437611, + 0.005204397719353437, + 0.004465613514184952, + 0.008400061167776585, + 0.00666593573987484, + 0.0033430280163884163, + 0.003440884407609701, + 0.003211228409782052, + 0.0026391209103167057, + 0.004061233717948198, + 0.0026506660506129265, + 0.004175563808530569, + 0.003757934086024761, + 0.009257436729967594, + 0.010059132240712643, + 0.01047319546341896, + 0.008800075389444828, + 0.009634509682655334, + 0.008198549039661884, + 0.011800081469118595, + 0.009490576572716236, + 0.015187227167189121, + 0.016270698979496956, + 0.018886270001530647, + 0.015941740944981575, + 0.013823200948536396, + 0.014517814852297306, + 0.015590613707900047, + 0.012188764289021492, + 0.06327621638774872, + 0.07669311761856079, + 0.10227973759174347, + 0.08855534344911575, + 0.05870755389332771, + 0.06142456457018852, + 0.09491587430238724, + 0.06363710016012192 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.026366863399744034 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009214208112098277 + ], + "opposite_logprob_tokens": [ + -7.000921249389648 + ], + "logprob_reward": -0.0009214208112098277, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009214208112098277, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1350029855966568, + 0.05438537523150444, + 0.07747681438922882, + 0.05141792446374893, + 0.046437881886959076, + 0.024009566754102707, + 0.058486223220825195, + 0.3126530945301056, + 0.057090844959020615, + 0.18303930759429932 + ], + "credit_map": [ + 0.016541652381420135, + 0.014046497642993927, + 0.022946760058403015, + 0.013489983975887299, + 0.014496473595499992, + 0.012413854710757732, + 0.021662451326847076, + 0.019405297935009003, + 0.007290728855878115, + 0.008735904470086098, + 0.0060845185071229935, + 0.005363333970308304, + 0.006618530489504337, + 0.005299466662108898, + 0.0076909903436899185, + 0.007301902864128351, + 0.00965396873652935, + 0.010094575583934784, + 0.01289933081716299, + 0.006582631729543209, + 0.00746406614780426, + 0.006427583284676075, + 0.009841056540608406, + 0.014513591304421425, + 0.006410364061594009, + 0.009258639998733997, + 0.006747550796717405, + 0.0060236225835978985, + 0.0053815580904483795, + 0.005635243374854326, + 0.005249719135463238, + 0.0067112199030816555, + 0.006593191064894199, + 0.008542333729565144, + 0.0045595550909638405, + 0.003541014390066266, + 0.004812050610780716, + 0.004908456467092037, + 0.005126284901052713, + 0.008354991674423218, + 0.0032584487926214933, + 0.003534112125635147, + 0.003121083602309227, + 0.0021480012219399214, + 0.002354092663154006, + 0.0028810640797019005, + 0.0032757744193077087, + 0.0034369879867881536, + 0.009111444465816021, + 0.008475069887936115, + 0.006251169368624687, + 0.0058683669194579124, + 0.007645760662853718, + 0.005488297436386347, + 0.007410889957100153, + 0.00823521800339222, + 0.0378936342895031, + 0.042945608496665955, + 0.04433714598417282, + 0.03587135300040245, + 0.03685443103313446, + 0.03461012616753578, + 0.04455270245671272, + 0.03558804839849472, + 0.005835185293108225, + 0.0066844020038843155, + 0.008912811987102032, + 0.007502107415348291, + 0.007366943638771772, + 0.006777436472475529, + 0.007818699814379215, + 0.006193250883370638, + 0.02691003493964672, + 0.029927203431725502, + 0.029609179124236107, + 0.021896090358495712, + 0.01800350844860077, + 0.022175375372171402, + 0.019319990649819374, + 0.015197922475636005 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0009214208112098277 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233837366104126 + ], + "opposite_logprob_tokens": [ + -0.8983837366104126 + ], + "logprob_reward": -0.5233837366104126, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233837366104126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06292520463466644, + 0.024814818054437637, + 0.01721939630806446, + 0.024316847324371338, + 0.017355902120471, + 0.049878280609846115, + 0.019218232482671738, + 0.10769674181938171, + 0.5375183820724487, + 0.13905620574951172 + ], + "credit_map": [ + 0.007844613865017891, + 0.0074563296511769295, + 0.009361139498651028, + 0.005673996172845364, + 0.007068755105137825, + 0.004729481413960457, + 0.009224451147019863, + 0.011566439643502235, + 0.002804246498271823, + 0.0035959561355412006, + 0.0030226001981645823, + 0.00271148351021111, + 0.0024276121985167265, + 0.0022724808659404516, + 0.0035668492782860994, + 0.004413589835166931, + 0.0018662598449736834, + 0.002744614612311125, + 0.0024104109033942223, + 0.0017668690998107195, + 0.0016939324559643865, + 0.0017635298427194357, + 0.0021571919787675142, + 0.002816590014845133, + 0.0029566276352852583, + 0.0034814539831131697, + 0.0031769757624715567, + 0.0019973067101091146, + 0.0022180781234055758, + 0.002322572749108076, + 0.003432711586356163, + 0.004731121007353067, + 0.0024968241341412067, + 0.0027304792311042547, + 0.0024717592168599367, + 0.00182779086753726, + 0.001963327405974269, + 0.0013958582421764731, + 0.0026006875559687614, + 0.0018691748846322298, + 0.005694766063243151, + 0.004688010551035404, + 0.0060140956193208694, + 0.005796034354716539, + 0.006397831253707409, + 0.004907780326902866, + 0.008326548151671886, + 0.008053215220570564, + 0.0025001622270792723, + 0.0029740657191723585, + 0.0017762513598427176, + 0.0020240929443389177, + 0.003016226226463914, + 0.0018441214924678206, + 0.002504624892026186, + 0.0025786878541111946, + 0.013855678960680962, + 0.013742844574153423, + 0.013457377441227436, + 0.011962109245359898, + 0.01407236885279417, + 0.011371273547410965, + 0.015057899989187717, + 0.014177191071212292, + 0.05911380052566528, + 0.06723819673061371, + 0.08465877920389175, + 0.07333964854478836, + 0.06312226504087448, + 0.06453127413988113, + 0.07381081581115723, + 0.051703594624996185, + 0.0252239853143692, + 0.023806245997548103, + 0.021059855818748474, + 0.014184980653226376, + 0.01229271199554205, + 0.018198568373918533, + 0.013165762647986412, + 0.01112409308552742 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.5233837366104126 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With the camera held perfectly still, the extreme macro shot focuses with microscopic clarity on the sterile, polished stainless steel tip of a hypodermic needle. Casting a sharp, well-defined highlight under a bright clinical light source, a shimmering drop of perfectly transparent liquid slowly and deliberately seeps from the needle's angled opening. The liquid expands, its surface tension creating a perfect, trembling sphere that clings precariously to the metal, refracting light like a tiny crystal lens. The background is a completely blurred, soft-focus, cool blue and white, suggesting a laboratory setting but revealing no specific details. The overall atmosphere is one of intense concentration, quiet anticipation, and tedious precision, as the weight of the drop grows heavier, seemingly at the very moment of its impact.", + "seed": 12588, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_4105_seed_12588_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11899028203915804, + "vlm_logprob": -0.11899028203915804, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.11899028203915804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004774004686623812 + ], + "opposite_logprob_tokens": [ + -6.25477409362793 + ], + "logprob_reward": -0.004774004686623812, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.004774004686623812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2039785385131836, + 0.07056236267089844, + 0.07104604691267014, + 0.07198218256235123, + 0.12306104600429535, + 0.12155472487211227, + 0.09284605830907822, + 0.14405740797519684, + 0.04136769846081734, + 0.059543952345848083 + ], + "credit_map": [ + 0.02344154752790928, + 0.030642567202448845, + 0.01686214469373226, + 0.035398174077272415, + 0.04995792731642723, + 0.023934287950396538, + 0.012027657590806484, + 0.011714235879480839, + 0.009055905975401402, + 0.01066935807466507, + 0.008512946777045727, + 0.01177195180207491, + 0.011201568879187107, + 0.007548845838755369, + 0.004955376964062452, + 0.006846410222351551, + 0.006241617724299431, + 0.012421796098351479, + 0.0045659299939870834, + 0.008905814960598946, + 0.02314435876905918, + 0.005351185332983732, + 0.004783525597304106, + 0.005631818901747465, + 0.008820765651762486, + 0.009678100235760212, + 0.006414259783923626, + 0.010283218696713448, + 0.014282758347690105, + 0.008019505999982357, + 0.006096051540225744, + 0.008387522771954536, + 0.011488760821521282, + 0.01502692699432373, + 0.013603344559669495, + 0.02835327759385109, + 0.023834645748138428, + 0.015461383387446404, + 0.008056503720581532, + 0.007236207835376263, + 0.014721476472914219, + 0.014426058158278465, + 0.009882684797048569, + 0.02684120461344719, + 0.02610270120203495, + 0.012973453849554062, + 0.008522524498403072, + 0.008084614761173725, + 0.00993728544563055, + 0.010370525531470776, + 0.009476304985582829, + 0.012526705861091614, + 0.02173536643385887, + 0.014564122073352337, + 0.006344526540488005, + 0.007891220971941948, + 0.016685234382748604, + 0.018585287034511566, + 0.016826799139380455, + 0.034798428416252136, + 0.027535464614629745, + 0.014190646819770336, + 0.008341006934642792, + 0.007094542495906353, + 0.004971589893102646, + 0.005507737398147583, + 0.004208599682897329, + 0.00775704113766551, + 0.006794908549636602, + 0.0046668811701238155, + 0.0038870451971888542, + 0.0035738996230065823, + 0.0064717624336481094, + 0.007329108659178019, + 0.007429126650094986, + 0.013421519659459591, + 0.008921951055526733, + 0.007694784551858902, + 0.004202262964099646, + 0.004073436371982098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.004774004686623812 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006749095860868692 + ], + "opposite_logprob_tokens": [ + -5.006749153137207 + ], + "logprob_reward": -0.006749095860868692, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006749095860868692, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20061008632183075, + 0.04261123389005661, + 0.08784810453653336, + 0.08150086551904678, + 0.14945939183235168, + 0.2352430671453476, + 0.06271397322416306, + 0.0336565338075161, + 0.064634308218956, + 0.04172244295477867 + ], + "credit_map": [ + 0.031554073095321655, + 0.024399541318416595, + 0.023175904527306557, + 0.039713192731142044, + 0.03308481350541115, + 0.022789165377616882, + 0.014539288356900215, + 0.01135409064590931, + 0.006073860451579094, + 0.006466716527938843, + 0.004155426751822233, + 0.008076603524386883, + 0.0076209139078855515, + 0.004851137287914753, + 0.0029976193327456713, + 0.0023689523804932833, + 0.009841805323958397, + 0.013381495140492916, + 0.009638526476919651, + 0.022443342953920364, + 0.009474575519561768, + 0.0127714267000556, + 0.005174900870770216, + 0.005122025962918997, + 0.007813621312379837, + 0.011075342074036598, + 0.010908946394920349, + 0.0231737419962883, + 0.00896697398275137, + 0.009471232071518898, + 0.005241746548563242, + 0.004849257413297892, + 0.01657208241522312, + 0.019812706857919693, + 0.011299102567136288, + 0.024544846266508102, + 0.03727070614695549, + 0.01780535653233528, + 0.011859183199703693, + 0.010295403189957142, + 0.030147843062877655, + 0.02769000083208084, + 0.018995976075530052, + 0.04180627316236496, + 0.05659289285540581, + 0.0258918646723032, + 0.01724029704928398, + 0.016877902671694756, + 0.005199525970965624, + 0.009214887395501137, + 0.007178512401878834, + 0.020062753930687904, + 0.007723264861851931, + 0.006327247712761164, + 0.0037885447964072227, + 0.003219237085431814, + 0.0036747283302247524, + 0.003966862335801125, + 0.004470211453735828, + 0.0072929938323795795, + 0.0050865341909229755, + 0.0041848644614219666, + 0.0023820747155696154, + 0.0025982605293393135, + 0.007822657003998756, + 0.008063146844506264, + 0.008088560774922371, + 0.01437125913798809, + 0.0068404795601964, + 0.011658280156552792, + 0.004023869521915913, + 0.0037660510279238224, + 0.005614497233182192, + 0.005176003556698561, + 0.004012896679341793, + 0.008053052239120007, + 0.00576816638931632, + 0.008622247725725174, + 0.00239859102293849, + 0.002076982520520687 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.006749095860868692 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260435398668051 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260435398668051, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260435398668051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11433462798595428, + 0.05385596677660942, + 0.0504988431930542, + 0.07439848780632019, + 0.13312819600105286, + 0.2519218921661377, + 0.2559596300125122, + 0.0186183899641037, + 0.02670266106724739, + 0.020581258460879326 + ], + "credit_map": [ + 0.013653019443154335, + 0.014151359908282757, + 0.013291087001562119, + 0.02198956161737442, + 0.019508304074406624, + 0.014250525273382664, + 0.009329723194241524, + 0.008161044679582119, + 0.006604140158742666, + 0.00834619253873825, + 0.006512830965220928, + 0.010971556417644024, + 0.007768208626657724, + 0.006465333979576826, + 0.0038700688164681196, + 0.0033176364377141, + 0.00511071365326643, + 0.006289258599281311, + 0.005221994128078222, + 0.006722725462168455, + 0.009154037572443485, + 0.009510541334748268, + 0.0036557361017912626, + 0.004833836108446121, + 0.007643167395144701, + 0.009214992634952068, + 0.00834635179489851, + 0.013746597804129124, + 0.012206054292619228, + 0.013164161704480648, + 0.004906651563942432, + 0.005170509684830904, + 0.011779267340898514, + 0.017281321808695793, + 0.011924098245799541, + 0.021408798173069954, + 0.03577708080410957, + 0.015150829218327999, + 0.009333106689155102, + 0.010473700240254402, + 0.022155288606882095, + 0.031377024948596954, + 0.01946096681058407, + 0.033429019153118134, + 0.0764429122209549, + 0.03255448862910271, + 0.016969788819551468, + 0.01953241415321827, + 0.023222070187330246, + 0.03038889355957508, + 0.02340833656489849, + 0.05312445014715195, + 0.060991231352090836, + 0.029961494728922844, + 0.016692891716957092, + 0.01817026548087597, + 0.0023689200170338154, + 0.002193338004872203, + 0.002260628854855895, + 0.004430640023201704, + 0.0033790841698646545, + 0.001678328961133957, + 0.0010720405261963606, + 0.001235409639775753, + 0.0028227041475474834, + 0.003079925896599889, + 0.0027830861508846283, + 0.00398711534217, + 0.004729714710265398, + 0.004688386805355549, + 0.002320502419024706, + 0.002291227923706174, + 0.0030058424454182386, + 0.0024939016439020634, + 0.0018956501735374331, + 0.0035764791537076235, + 0.003170655807480216, + 0.004144687671214342, + 0.00113075936678797, + 0.0011632830137386918 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.005260435398668051 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3330688178539276, + 0.12571606040000916, + 0.10943058878183365, + 0.08783848583698273, + 0.052583880722522736, + 0.0862969234585762, + 0.06542441248893738, + 0.07131590694189072, + 0.041065312922000885, + 0.02725963294506073 + ], + "credit_map": [ + 0.04165779426693916, + 0.046104319393634796, + 0.03125487267971039, + 0.06329576671123505, + 0.0704454854130745, + 0.0396835058927536, + 0.021013807505369186, + 0.019613271579146385, + 0.011900555342435837, + 0.017707642167806625, + 0.012847662903368473, + 0.029580170288681984, + 0.018164435401558876, + 0.017451314255595207, + 0.009782282635569572, + 0.008281983435153961, + 0.016161717474460602, + 0.012276947498321533, + 0.013935286551713943, + 0.01789858750998974, + 0.012479792349040508, + 0.02525864541530609, + 0.005958194378763437, + 0.005461417138576508, + 0.013048894703388214, + 0.010527781210839748, + 0.009843996725976467, + 0.015182227827608585, + 0.012718383222818375, + 0.013387120328843594, + 0.006189905572682619, + 0.006940169259905815, + 0.00668673450127244, + 0.006779178977012634, + 0.00579206645488739, + 0.008371022529900074, + 0.011990937404334545, + 0.005254766438156366, + 0.003873425303027034, + 0.003835746319964528, + 0.009985382668673992, + 0.009699884802103043, + 0.008092671632766724, + 0.017104921862483025, + 0.011949189007282257, + 0.017475375905632973, + 0.0059516034089028835, + 0.006037888582795858, + 0.005981448572129011, + 0.007389097940176725, + 0.006198357790708542, + 0.012262825854122639, + 0.016049450263381004, + 0.008999758400022984, + 0.004383680410683155, + 0.004159791395068169, + 0.005352763459086418, + 0.010345806367695332, + 0.009865496307611465, + 0.022032517939805984, + 0.007048202212899923, + 0.009108726866543293, + 0.004223044961690903, + 0.003339344635605812, + 0.0036757851485162973, + 0.005071595776826143, + 0.005361724644899368, + 0.010206339880824089, + 0.005829835310578346, + 0.004710330627858639, + 0.003382036928087473, + 0.002827663440257311, + 0.004079720471054316, + 0.004259184468537569, + 0.0031720020342618227, + 0.004971775691956282, + 0.003606108482927084, + 0.0031292045023292303, + 0.0020189820788800716, + 0.002022653352469206 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.001943010720424354 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762248635292053 + ], + "opposite_logprob_tokens": [ + -0.8262248635292053 + ], + "logprob_reward": -0.5762248635292053, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762248635292053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.34213516116142273, + 0.1160430982708931, + 0.11507558822631836, + 0.06687530130147934, + 0.06520985811948776, + 0.04440900683403015, + 0.065591961145401, + 0.055429719388484955, + 0.1018824577331543, + 0.02734784595668316 + ], + "credit_map": [ + 0.050338082015514374, + 0.04983654245734215, + 0.032621800899505615, + 0.04929475113749504, + 0.08126972615718842, + 0.04026723653078079, + 0.020256957039237022, + 0.018250038847327232, + 0.016523245722055435, + 0.01526691671460867, + 0.015397977083921432, + 0.022214092314243317, + 0.015769021585583687, + 0.01696433499455452, + 0.008221307769417763, + 0.005686207674443722, + 0.012357581406831741, + 0.017329992726445198, + 0.012727946043014526, + 0.02783736027777195, + 0.014718681573867798, + 0.015444265678524971, + 0.007204090245068073, + 0.0074556609615683556, + 0.00927472859621048, + 0.008520504459738731, + 0.008089499548077583, + 0.013328230008482933, + 0.007098729722201824, + 0.012248696759343147, + 0.004133002366870642, + 0.004181914497166872, + 0.006842819973826408, + 0.008586608804762363, + 0.0070675271563231945, + 0.01240340806543827, + 0.013955740258097649, + 0.006942721549421549, + 0.00465314881876111, + 0.004757888149470091, + 0.004835747182369232, + 0.005798563361167908, + 0.005175202153623104, + 0.010358220897614956, + 0.007440477609634399, + 0.005105534568428993, + 0.002854942111298442, + 0.0028403187170624733, + 0.005513837561011314, + 0.007755720056593418, + 0.0065589542500674725, + 0.01489257626235485, + 0.014581133611500263, + 0.006022030953317881, + 0.004787689074873924, + 0.005480015184730291, + 0.005136545281857252, + 0.0070806159637868404, + 0.005625341087579727, + 0.009617122821509838, + 0.010890239849686623, + 0.009179489687085152, + 0.0034831338562071323, + 0.004417229909449816, + 0.010635560378432274, + 0.012293199077248573, + 0.011927267536520958, + 0.021919168531894684, + 0.01724480837583542, + 0.01407309528440237, + 0.007394298445433378, + 0.006395064294338226, + 0.003779373597353697, + 0.0035495730116963387, + 0.0035462870728224516, + 0.004832877777516842, + 0.0035708157811313868, + 0.0038455845788121223, + 0.0021124552004039288, + 0.002110881032422185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.5762248635292053 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With the camera held perfectly still, the extreme macro shot focuses with microscopic clarity on the sterile, polished stainless steel tip of a hypodermic needle. Casting a sharp, well-defined highlight under a bright clinical light source, a shimmering drop of perfectly transparent liquid slowly and deliberately seeps from the needle's angled opening. The liquid expands, its surface tension creating a perfect, trembling sphere that clings precariously to the metal, refracting light like a tiny crystal lens. The background is a completely blurred, soft-focus, cool blue and white, suggesting a laboratory setting but revealing no specific details. The overall atmosphere is one of intense concentration, quiet anticipation, and tedious precision, as the weight of the drop grows heavier, seemingly at the very moment of its impact.", + "seed": 12589, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_4105_seed_12589_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.23472652193158866, + "vlm_logprob": -0.23472652193158866, + "vlm_yes_no_margin": 1.9999999761581422, + "vlm_token_credit": -0.23472652193158866, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4327895939350128 + ], + "opposite_logprob_tokens": [ + -1.0577895641326904 + ], + "logprob_reward": -0.4327895939350128, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4327895939350128, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2733907103538513, + 0.10362072288990021, + 0.07070183753967285, + 0.026250550523400307, + 0.04885347932577133, + 0.03055247664451599, + 0.0781993493437767, + 0.23974792659282684, + 0.04797496646642685, + 0.08070796728134155 + ], + "credit_map": [ + 0.047867920249700546, + 0.023303726688027382, + 0.04306330531835556, + 0.030825331807136536, + 0.026726119220256805, + 0.03735455498099327, + 0.03203220292925835, + 0.03221758082509041, + 0.013955713249742985, + 0.008030034601688385, + 0.015190002508461475, + 0.012099632993340492, + 0.013571897521615028, + 0.01237932126969099, + 0.015272228978574276, + 0.013121902011334896, + 0.008844705298542976, + 0.006098676007241011, + 0.011263325810432434, + 0.008861374109983444, + 0.007375556975603104, + 0.010231048800051212, + 0.008422798477113247, + 0.009604356251657009, + 0.0034616878256201744, + 0.00228433171287179, + 0.002991018583998084, + 0.002824540250003338, + 0.003571835346519947, + 0.00433805026113987, + 0.002614049706608057, + 0.004165039397776127, + 0.004818385001271963, + 0.00440592784434557, + 0.008502819575369358, + 0.00875628087669611, + 0.00640702061355114, + 0.0058922674506902695, + 0.004321329761296511, + 0.005749452393501997, + 0.0024883609730750322, + 0.002331799827516079, + 0.0029833540320396423, + 0.0043836976401507854, + 0.006000773515552282, + 0.003774177748709917, + 0.004087200853973627, + 0.004503116477280855, + 0.00907810963690281, + 0.009151583537459373, + 0.013441206887364388, + 0.010708625428378582, + 0.009311054833233356, + 0.00931235309690237, + 0.0069945985451340675, + 0.010201826691627502, + 0.024946866557002068, + 0.02291564829647541, + 0.04083148017525673, + 0.02730124071240425, + 0.03767789900302887, + 0.03302488103508949, + 0.025723978877067566, + 0.0273259487003088, + 0.0053030382841825485, + 0.004268546588718891, + 0.005407379940152168, + 0.00427982909604907, + 0.010304635390639305, + 0.00712978420779109, + 0.0056944340467453, + 0.005587324500083923, + 0.008429520763456821, + 0.0068090008571743965, + 0.0084116505458951, + 0.00783627014607191, + 0.012942087836563587, + 0.0136629743501544, + 0.012188575230538845, + 0.010427895933389664 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.4327895939350128 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282236993312836 + ], + "opposite_logprob_tokens": [ + -2.017822265625 + ], + "logprob_reward": -0.14282236993312836, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14282236993312836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3198714554309845, + 0.0815240889787674, + 0.07535386085510254, + 0.04144291579723358, + 0.040370047092437744, + 0.06743674725294113, + 0.06815212219953537, + 0.1352677196264267, + 0.07144230604171753, + 0.09913870692253113 + ], + "credit_map": [ + 0.02712938003242016, + 0.019863450899720192, + 0.02398989349603653, + 0.031106343492865562, + 0.041941169649362564, + 0.060987722128629684, + 0.0609869509935379, + 0.053866587579250336, + 0.010737966746091843, + 0.006824223790317774, + 0.011590836569666862, + 0.010697157122194767, + 0.011795207858085632, + 0.009510516189038754, + 0.008879251778125763, + 0.011488936841487885, + 0.008303945884108543, + 0.006159930489957333, + 0.011077608913183212, + 0.009101101197302341, + 0.00893040094524622, + 0.01183251105248928, + 0.00960561539977789, + 0.010342751629650593, + 0.004170877393335104, + 0.0029577233362942934, + 0.005097805056720972, + 0.0048485416918993, + 0.0050828419625759125, + 0.006873862352222204, + 0.003862945595756173, + 0.008548319339752197, + 0.0044632176868617535, + 0.0033859184477478266, + 0.005643908865749836, + 0.0056683081202209, + 0.0046348050236701965, + 0.006349893752485514, + 0.004106844309717417, + 0.006117156241089106, + 0.006632459349930286, + 0.00530640035867691, + 0.007010305300354958, + 0.009039475582540035, + 0.010948906652629375, + 0.0098983533680439, + 0.007392566651105881, + 0.01120829302817583, + 0.007144450210034847, + 0.006762176286429167, + 0.0090081337839365, + 0.007641688454896212, + 0.010999011807143688, + 0.009110846556723118, + 0.007433627266436815, + 0.010052193887531757, + 0.012344450689852238, + 0.011152567341923714, + 0.01693057455122471, + 0.014617069624364376, + 0.026996424421668053, + 0.019945895299315453, + 0.01762949861586094, + 0.015651246532797813, + 0.007970886304974556, + 0.006409313529729843, + 0.006811860948801041, + 0.006502128671854734, + 0.016366126015782356, + 0.010942964814603329, + 0.008487428538501263, + 0.007951595820486546, + 0.010279778391122818, + 0.008144693449139595, + 0.010346857830882072, + 0.009401179850101471, + 0.016073808073997498, + 0.016734743490815163, + 0.015230539254844189, + 0.012927114963531494 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.14282236993312836 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062082793563604355 + ], + "opposite_logprob_tokens": [ + -2.8120827674865723 + ], + "logprob_reward": -0.062082793563604355, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062082793563604355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22763323783874512, + 0.08173592388629913, + 0.048100557178258896, + 0.03019285760819912, + 0.03904923051595688, + 0.03397642821073532, + 0.052731748670339584, + 0.26115813851356506, + 0.045393750071525574, + 0.18002809584140778 + ], + "credit_map": [ + 0.037602704018354416, + 0.01828024722635746, + 0.032316237688064575, + 0.02425929345190525, + 0.025091568008065224, + 0.02859082631766796, + 0.03023621439933777, + 0.03125614672899246, + 0.010717485100030899, + 0.00526313204318285, + 0.010206916369497776, + 0.00836662482470274, + 0.008714024908840656, + 0.015076509676873684, + 0.010848449543118477, + 0.012542780488729477, + 0.006412971299141645, + 0.004467181395739317, + 0.007306681480258703, + 0.005703268107026815, + 0.00558958575129509, + 0.006242942065000534, + 0.005671162623912096, + 0.0067067635245621204, + 0.0028063422068953514, + 0.002442740835249424, + 0.0029717828147113323, + 0.003616810543462634, + 0.0044829268008470535, + 0.0047808135859668255, + 0.0038074764888733625, + 0.005283965729176998, + 0.004170950036495924, + 0.0030864286236464977, + 0.006374481599777937, + 0.005528552457690239, + 0.004521361086517572, + 0.005525620654225349, + 0.0047966307029128075, + 0.00504520395770669, + 0.0035650189965963364, + 0.002677998738363385, + 0.0032016532495617867, + 0.0038827324751764536, + 0.0046509383246302605, + 0.0061888801865279675, + 0.00421919347718358, + 0.005590013228356838, + 0.00684598833322525, + 0.005489443428814411, + 0.007582168094813824, + 0.005683631170541048, + 0.00731945876032114, + 0.0068483296781778336, + 0.005477370694279671, + 0.007485360372811556, + 0.018149206414818764, + 0.021165303885936737, + 0.023264195770025253, + 0.028728660196065903, + 0.05951656028628349, + 0.03955626115202904, + 0.04053781181573868, + 0.03024015575647354, + 0.0054285768419504166, + 0.003589638974517584, + 0.005439800675958395, + 0.004431173671036959, + 0.00907407607883215, + 0.007227515801787376, + 0.0048454697243869305, + 0.005357501562684774, + 0.019548647105693817, + 0.014479145407676697, + 0.019771894440054893, + 0.017124168574810028, + 0.028143728151917458, + 0.030513368546962738, + 0.026432499289512634, + 0.02401464618742466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.062082793563604355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012541121803224087 + ], + "opposite_logprob_tokens": [ + -4.3875412940979 + ], + "logprob_reward": -0.012541121803224087, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012541121803224087, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5060343146324158, + 0.0623227134346962, + 0.03601634129881859, + 0.024758297950029373, + 0.06930682063102722, + 0.029071977362036705, + 0.06707439571619034, + 0.11673103272914886, + 0.03151693567633629, + 0.057167116552591324 + ], + "credit_map": [ + 0.053800079971551895, + 0.028410758823156357, + 0.040025513619184494, + 0.03753848001360893, + 0.031162308529019356, + 0.12583373486995697, + 0.09682852029800415, + 0.09243495017290115, + 0.006671130657196045, + 0.003905672812834382, + 0.007065617013722658, + 0.005193459335714579, + 0.008000295609235764, + 0.01247668918222189, + 0.008591166697442532, + 0.010418687015771866, + 0.006244730204343796, + 0.0023843266535550356, + 0.004961754195392132, + 0.0034048743546009064, + 0.002483413089066744, + 0.00547759747132659, + 0.0038742769975215197, + 0.007185370661318302, + 0.0032220743596553802, + 0.0021385883446782827, + 0.002808927558362484, + 0.002843548310920596, + 0.002886493457481265, + 0.003765946486964822, + 0.0025455851573497057, + 0.0045471349731087685, + 0.00924764946103096, + 0.004287485498934984, + 0.006475734990090132, + 0.008198530413210392, + 0.006752524059265852, + 0.011617772281169891, + 0.0064034536480903625, + 0.016323678195476532, + 0.002836971776559949, + 0.0024767094291746616, + 0.0029080586973577738, + 0.0040055690333247185, + 0.004454267676919699, + 0.004105820786207914, + 0.0036783504765480757, + 0.004606230650097132, + 0.008175654336810112, + 0.007258385885506868, + 0.01044214703142643, + 0.008677010424435139, + 0.009082658216357231, + 0.008299094624817371, + 0.006033098790794611, + 0.009106348268687725, + 0.010342507623136044, + 0.009078352712094784, + 0.01356950867921114, + 0.012539769522845745, + 0.02327640913426876, + 0.017563441768288612, + 0.015648502856492996, + 0.014712545089423656, + 0.003484123619273305, + 0.002663285005837679, + 0.0030918896663933992, + 0.0028768926858901978, + 0.007146395742893219, + 0.004713806789368391, + 0.003735726233571768, + 0.0038048194255679846, + 0.005773336626589298, + 0.00487195560708642, + 0.006284484639763832, + 0.006000953260809183, + 0.009120524860918522, + 0.00887574814260006, + 0.009209856390953064, + 0.007030260283499956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.012541121803224087 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233967304229736 + ], + "opposite_logprob_tokens": [ + -0.8983967304229736 + ], + "logprob_reward": -0.5233967304229736, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233967304229736, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7315768003463745, + 0.05166376009583473, + 0.018203621730208397, + 0.026431694626808167, + 0.01966080814599991, + 0.012052927166223526, + 0.0338619090616703, + 0.07369332015514374, + 0.016583329066634178, + 0.01627187244594097 + ], + "credit_map": [ + 0.09039086103439331, + 0.04767937213182449, + 0.06699210405349731, + 0.04774760082364082, + 0.06914424151182175, + 0.14748045802116394, + 0.10881143808364868, + 0.15333066880702972, + 0.003914662171155214, + 0.004241385962814093, + 0.006615838501602411, + 0.006101598963141441, + 0.008401037193834782, + 0.009515408426523209, + 0.005393223371356726, + 0.007480603642761707, + 0.0018696734914556146, + 0.0014490553876385093, + 0.0028742013964802027, + 0.002442929893732071, + 0.0023236016277223825, + 0.002358117839321494, + 0.0022401863243430853, + 0.002645856235176325, + 0.0023947046138346195, + 0.001955814892426133, + 0.0027693784795701504, + 0.003084799274802208, + 0.0031796805560588837, + 0.005373873747885227, + 0.003973543178290129, + 0.003699899883940816, + 0.0015393003122881055, + 0.001191395684145391, + 0.001827389351092279, + 0.0024665934033691883, + 0.0020627574995160103, + 0.003968473989516497, + 0.0019327732734382153, + 0.004672124516218901, + 0.0014185301261022687, + 0.0009120183531194925, + 0.001214618212543428, + 0.0016237974632531404, + 0.0016428064554929733, + 0.002251328667625785, + 0.0009965492645278573, + 0.0019932787399739027, + 0.0034971667919307947, + 0.0032345273066312075, + 0.004392458591610193, + 0.0035798472817987204, + 0.005167605821043253, + 0.0050300369039177895, + 0.0039225490763783455, + 0.005037717055529356, + 0.0071586305275559425, + 0.007039757911115885, + 0.011784021742641926, + 0.008663966320455074, + 0.011913128197193146, + 0.010194352827966213, + 0.007986590266227722, + 0.008952876552939415, + 0.001682986505329609, + 0.001431060372851789, + 0.0013961909571662545, + 0.001447899965569377, + 0.0034725109580904245, + 0.002958832774311304, + 0.001907329773530364, + 0.002286517759785056, + 0.001993560465052724, + 0.0012463268358260393, + 0.001667264848947525, + 0.001870849751867354, + 0.002288585063070059, + 0.0025996961630880833, + 0.0018780590035021305, + 0.0027275304310023785 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.5233967304229736 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "With the camera held perfectly still, the extreme macro shot focuses with microscopic clarity on the sterile, polished stainless steel tip of a hypodermic needle. Casting a sharp, well-defined highlight under a bright clinical light source, a shimmering drop of perfectly transparent liquid slowly and deliberately seeps from the needle's angled opening. The liquid expands, its surface tension creating a perfect, trembling sphere that clings precariously to the metal, refracting light like a tiny crystal lens. The background is a completely blurred, soft-focus, cool blue and white, suggesting a laboratory setting but revealing no specific details. The overall atmosphere is one of intense concentration, quiet anticipation, and tedious precision, as the weight of the drop grows heavier, seemingly at the very moment of its impact.", + "seed": 12590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_11/index_4105_seed_12590_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.2513228189200163, + "vlm_logprob": -0.2513228189200163, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.2513228189200163, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.024224713444709778 + ], + "opposite_logprob_tokens": [ + -3.8992247581481934 + ], + "logprob_reward": -0.024224713444709778, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.024224713444709778, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040738243609666824, + 0.013903507962822914, + 0.006325918715447187, + 0.01804708130657673, + 0.016101183369755745, + 0.047542229294776917, + 0.41931408643722534, + 0.3184046447277069, + 0.07670261710882187, + 0.042920444160699844 + ], + "credit_map": [ + 0.003426955547183752, + 0.005627748556435108, + 0.0047070616856217384, + 0.004780895542353392, + 0.007255214266479015, + 0.003323520766571164, + 0.0062959762290120125, + 0.005320876836776733, + 0.0015134277055040002, + 0.0015907409833744168, + 0.001804840168915689, + 0.0013636074727401137, + 0.003181565785780549, + 0.0014427972491830587, + 0.0016920955386012793, + 0.0013144355034455657, + 0.0006975919823162258, + 0.0005533417570404708, + 0.000802883820142597, + 0.000661380763631314, + 0.00100579927675426, + 0.0007823581108823419, + 0.0006542607443407178, + 0.0011683028424158692, + 0.0015976919094100595, + 0.0035717154387384653, + 0.006256509572267532, + 0.0011335057206451893, + 0.0016013482818379998, + 0.0013203948037698865, + 0.0013055528979748487, + 0.001260364893823862, + 0.001289362902753055, + 0.003570102620869875, + 0.005557226948440075, + 0.0009537370060570538, + 0.0011121188290417194, + 0.0011000254889950156, + 0.0009841659339144826, + 0.001534446026198566, + 0.002735322108492255, + 0.014494272880256176, + 0.01840866543352604, + 0.0025778859853744507, + 0.002872243756428361, + 0.0024924916215240955, + 0.0019357046112418175, + 0.00202564662322402, + 0.020135322585701942, + 0.15804389119148254, + 0.1720070242881775, + 0.0146005405113101, + 0.024081513285636902, + 0.012377911247313023, + 0.008940435014665127, + 0.00912746787071228, + 0.016879679635167122, + 0.11214293539524078, + 0.13761547207832336, + 0.01111641712486744, + 0.013729552738368511, + 0.01290365494787693, + 0.007412209175527096, + 0.006604775786399841, + 0.0050349365919828415, + 0.023078005760908127, + 0.03063540905714035, + 0.0030057348776608706, + 0.004240337759256363, + 0.0041181063279509544, + 0.0024995170533657074, + 0.004090585745871067, + 0.004080281592905521, + 0.006199714262038469, + 0.009249499998986721, + 0.0056653995998203754, + 0.00444081099703908, + 0.004800053779035807, + 0.003620608476921916, + 0.004864079412072897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.024224713444709778 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038108546286821365 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.038108546286821365, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038108546286821365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02189379744231701, + 0.00735103664919734, + 0.004493250511586666, + 0.003270057961344719, + 0.009296268224716187, + 0.1277467906475067, + 0.13180387020111084, + 0.4720717966556549, + 0.1745094209909439, + 0.047563761472702026 + ], + "credit_map": [ + 0.001862473669461906, + 0.003767666406929493, + 0.002757604466751218, + 0.0026435174513608217, + 0.003641872899606824, + 0.0018180519109591842, + 0.003186320187523961, + 0.0022162876557558775, + 0.0008501918637193739, + 0.0007620561518706381, + 0.0011527533642947674, + 0.0009249011054635048, + 0.001259552431292832, + 0.0007140807574614882, + 0.0007174770580604672, + 0.0009700234513729811, + 0.0004579833475872874, + 0.0004691957728937268, + 0.0005191521486267447, + 0.00048665422946214676, + 0.0007225374574773014, + 0.0005900163669139147, + 0.0005912396591156721, + 0.0006564711802639067, + 0.000408433610573411, + 0.0003601056814659387, + 0.0006386320455931127, + 0.00040342850843444467, + 0.0004170568136032671, + 0.0003350515617057681, + 0.00031811147346161306, + 0.0003892379463650286, + 0.0008637078572064638, + 0.002108470071107149, + 0.0034029483795166016, + 0.0005739472107961774, + 0.0007528543355874717, + 0.0006437857518903911, + 0.0005258271703496575, + 0.00042472625500522554, + 0.006649111397564411, + 0.04625672474503517, + 0.054216377437114716, + 0.004571964964270592, + 0.004803892225027084, + 0.004753313958644867, + 0.003240601858124137, + 0.0032547989394515753, + 0.006738748867064714, + 0.04419742897152901, + 0.05208788067102432, + 0.011526355519890785, + 0.005488674156367779, + 0.004657619632780552, + 0.0031342972069978714, + 0.003972865641117096, + 0.024156512692570686, + 0.1705458015203476, + 0.20638160407543182, + 0.015398702584207058, + 0.019993530586361885, + 0.01861468143761158, + 0.009834615513682365, + 0.007146309595555067, + 0.009529061615467072, + 0.058159634470939636, + 0.07193604111671448, + 0.007128680590540171, + 0.007952185347676277, + 0.007703846786171198, + 0.004747076891362667, + 0.007352878805249929, + 0.0041620200499892235, + 0.006042805965989828, + 0.008932518772780895, + 0.005213185679167509, + 0.008941812440752983, + 0.004922635853290558, + 0.003781963372603059, + 0.005566817242652178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.038108546286821365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497122183442116 + ], + "opposite_logprob_tokens": [ + -2.929971218109131 + ], + "logprob_reward": -0.05497122183442116, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497122183442116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020313521847128868, + 0.01329710055142641, + 0.007049064617604017, + 0.0048886677250266075, + 0.015707828104496002, + 0.04952358826994896, + 0.3809862732887268, + 0.32531389594078064, + 0.13991129398345947, + 0.04300876706838608 + ], + "credit_map": [ + 0.0014532090863212943, + 0.0037025404162704945, + 0.0024137855507433414, + 0.0025699110701680183, + 0.0036409066524356604, + 0.0015809137839823961, + 0.0026546104345470667, + 0.002297644969075918, + 0.0012490118388086557, + 0.0016439658356830478, + 0.0013969764113426208, + 0.0015645200619474053, + 0.0019787095952779055, + 0.002215357031673193, + 0.0013293417869135737, + 0.0019192170584574342, + 0.0007620555115863681, + 0.0005489816539920866, + 0.0009788491297513247, + 0.000566602626349777, + 0.0010173344053328037, + 0.0008317327592521906, + 0.0011281075421720743, + 0.0012154008727520704, + 0.0005336219328455627, + 0.0005103722796775401, + 0.0008759992197155952, + 0.0003384386363904923, + 0.0008505251025781035, + 0.00046589013072662055, + 0.0004487967526074499, + 0.0008650236413814127, + 0.0011045867577195168, + 0.003228827379643917, + 0.003893381217494607, + 0.0009950457606464624, + 0.0021809053141623735, + 0.0009130562539212406, + 0.0018252114532515407, + 0.0015668126288801432, + 0.0028455997817218304, + 0.016432037577033043, + 0.020562272518873215, + 0.002051775110885501, + 0.0021961559541523457, + 0.0023161976132541895, + 0.0015944308834150434, + 0.0015251152217388153, + 0.018592478707432747, + 0.13023154437541962, + 0.15503180027008057, + 0.030393261462450027, + 0.0169004425406456, + 0.01371001172810793, + 0.008732383139431477, + 0.007394352927803993, + 0.0163568165153265, + 0.11779230087995529, + 0.1412762999534607, + 0.01130689587444067, + 0.014088491909205914, + 0.012686064466834068, + 0.006750228814780712, + 0.0050567868165671825, + 0.008001268841326237, + 0.04487686976790428, + 0.056870974600315094, + 0.006334381643682718, + 0.007526225410401821, + 0.006144749000668526, + 0.0041743796318769455, + 0.005982446949928999, + 0.003406969830393791, + 0.00587431387975812, + 0.007497033104300499, + 0.00503067160025239, + 0.010053008794784546, + 0.004330268129706383, + 0.0034265508875250816, + 0.0033899506088346243 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.05497122183442116 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002197353169322014 + ], + "opposite_logprob_tokens": [ + -6.127197265625 + ], + "logprob_reward": -0.002197353169322014, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002197353169322014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03468622267246246, + 0.01534636877477169, + 0.006962617859244347, + 0.041818294674158096, + 0.011681498028337955, + 0.03092670813202858, + 0.4891265332698822, + 0.14728142321109772, + 0.06846308708190918, + 0.15370725095272064 + ], + "credit_map": [ + 0.0035079375375062227, + 0.006456427276134491, + 0.004346971400082111, + 0.0033962067682296038, + 0.005551811773329973, + 0.002338140970095992, + 0.004473842680454254, + 0.0046148826368153095, + 0.0018750665476545691, + 0.001916665118187666, + 0.0018554995767772198, + 0.0017794305458664894, + 0.002308797324076295, + 0.0023506288416683674, + 0.0010525144170969725, + 0.0022077655885368586, + 0.0006004989263601601, + 0.0007574846385978162, + 0.0009246863191947341, + 0.0006876884144730866, + 0.00118643615860492, + 0.0008020700188353658, + 0.001237058429978788, + 0.0007666941382922232, + 0.00636513764038682, + 0.00538664823397994, + 0.007930389605462551, + 0.004399477504193783, + 0.005033239256590605, + 0.006233430001884699, + 0.002639658283442259, + 0.0038303097244352102, + 0.001497125020250678, + 0.001751119620166719, + 0.0027974823024123907, + 0.0011164661264047027, + 0.0012337021762505174, + 0.0012454898096621037, + 0.0010305438190698624, + 0.001009568921290338, + 0.0020168425980955362, + 0.009435064159333706, + 0.011753827333450317, + 0.0013693160144612193, + 0.0021335904020816088, + 0.0014893893385306, + 0.0013797292485833168, + 0.0013489470584318042, + 0.022817639634013176, + 0.17808017134666443, + 0.20480850338935852, + 0.02838687039911747, + 0.02201402559876442, + 0.015956511721014977, + 0.009212822653353214, + 0.007849996909499168, + 0.007324774283915758, + 0.05258200690150261, + 0.06249207258224487, + 0.005503040738403797, + 0.006825424265116453, + 0.005618434399366379, + 0.003287701169028878, + 0.0036479744594544172, + 0.0038067628629505634, + 0.022764034569263458, + 0.028391726315021515, + 0.00287196459248662, + 0.0035713131073862314, + 0.002949295099824667, + 0.00204645493067801, + 0.002061528153717518, + 0.01612422615289688, + 0.022585570812225342, + 0.03173893317580223, + 0.015318364836275578, + 0.018397286534309387, + 0.024054935202002525, + 0.010608615353703499, + 0.014879295602440834 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.002197353169322014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject the tip of a hypodermic needle made of polished stainless steel?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the shot an extreme macro view with microscopic clarity?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a drop of perfectly transparent liquid seeping from the needle's opening and forming a trembling sphere?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background completely blurred with a soft-focus cool blue and white color palette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1371122598648071 + ], + "opposite_logprob_tokens": [ + -0.38711225986480713 + ], + "logprob_reward": -1.1371122598648071, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1371122598648071, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.019820094108581543, + 0.006096864119172096, + 0.005509776063263416, + 0.006120073609054089, + 0.01950356736779213, + 0.05401051416993141, + 0.619439423084259, + 0.16221986711025238, + 0.04468834772706032, + 0.06259148567914963 + ], + "credit_map": [ + 0.002270299708470702, + 0.0020512084010988474, + 0.002710967557504773, + 0.0030588917434215546, + 0.0032476717606186867, + 0.0018439056584611535, + 0.002132497262209654, + 0.0025046521332114935, + 0.0009556374279782176, + 0.0005769106210209429, + 0.0008080867701210082, + 0.0006005640607327223, + 0.0009780974360182881, + 0.0008100128034129739, + 0.000485293276142329, + 0.0008822616655379534, + 0.0006939027225598693, + 0.0004729393112938851, + 0.0008374254684895277, + 0.0004599537933245301, + 0.0009017008705995977, + 0.0005588358035311103, + 0.0007882124627940357, + 0.000796805543359369, + 0.0006550085381604731, + 0.0007393685518763959, + 0.0013648144667968154, + 0.000439600960817188, + 0.0007538326899521053, + 0.0006763995625078678, + 0.000832480494864285, + 0.000658568344078958, + 0.001546327257528901, + 0.004236113280057907, + 0.006148559972643852, + 0.0011449989397078753, + 0.0022489523980766535, + 0.0012239634525030851, + 0.0014673094265162945, + 0.001487343106418848, + 0.0027469920460134745, + 0.018149029463529587, + 0.021641841158270836, + 0.0025110472925007343, + 0.0023594892118126154, + 0.0023449386935681105, + 0.002333055716007948, + 0.0019241158151999116, + 0.029120629653334618, + 0.23813337087631226, + 0.2621786296367645, + 0.020392313599586487, + 0.03117969073355198, + 0.017241396009922028, + 0.01109668705612421, + 0.010096680372953415, + 0.0088078947737813, + 0.055173806846141815, + 0.06901770830154419, + 0.007957231253385544, + 0.00747305154800415, + 0.0064201075583696365, + 0.004081536550074816, + 0.0032885330729186535, + 0.0024816368240863085, + 0.014449834823608398, + 0.016837768256664276, + 0.0018847159808501601, + 0.003230511676520109, + 0.0019371816888451576, + 0.0016399346059188247, + 0.0022267589811235666, + 0.0044413055293262005, + 0.008303883485496044, + 0.008655423298478127, + 0.006454940419644117, + 0.016913045197725296, + 0.005909625440835953, + 0.0060089752078056335, + 0.005904286168515682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -1.1371122598648071 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_tensor([1846, 1846])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_tensor([1846, 1846])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..be23b3b4bf706c4e8988ca84ed6aa0478275a125 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_tensor([1846, 1846])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic shot begins in medium shot, then moves in a steady, fluid motion, backward and upward. The image initially focuses on a mythical, multicolored phoenix, soaring above a vast sea of clouds. The phoenix's feathers shimmer like silk and brocade: a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail, each feather edged with a shimmering golden thread. Illuminated by the warm golden light of the setting sun, its form appears shimmering. The phoenix flaps its massive wings in a majestic and rhythmic motion, each downward stroke powerful yet effortless, its long, resplendent tail gracefully flowing behind it like a rainbow ribbon. As the camera slowly pulls away, the viewer's perspective expands dramatically, the phoenix's once-dominant grandeur gradually dwarfed by the boundless, swirling cotton candy clouds below. Sunlight penetrates the gaps in the clouds, casting magnificent beams of light and dyeing the sea of clouds into a golden ocean. Ultimately, the phoenix transforms into a dazzling speck of color on the distant skyline, solitary and moving across the vastness of heaven and earth. The entire scene exudes a sense of ultimate epicness and solitary grandeur, a tranquil and sacred atmosphere, and warm and brilliant tones.", + "seed": 24556, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_1846_seed_24556_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021461419714614748, + "vlm_logprob": -0.021461419714614748, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.021461419714614748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01724463701248169 + ], + "opposite_logprob_tokens": [ + -4.142244815826416 + ], + "logprob_reward": -0.01724463701248169, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01724463701248169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.164272740483284, + 0.17356400191783905, + 0.03670347481966019, + 0.035164546221494675, + 0.1222902312874794, + 0.01794883795082569, + 0.016777576878666878, + 0.07613391429185867, + 0.098940409719944, + 0.25820422172546387 + ], + "credit_map": [ + 0.02234867587685585, + 0.03912965580821037, + 0.022791888564825058, + 0.021965617313981056, + 0.017383405938744545, + 0.011279067024588585, + 0.013667883351445198, + 0.01570655032992363, + 0.010972296819090843, + 0.005419256631284952, + 0.04412104934453964, + 0.011753266677260399, + 0.006902168970555067, + 0.07120732963085175, + 0.009950622916221619, + 0.013237995095551014, + 0.004756955895572901, + 0.002889031544327736, + 0.006976092699915171, + 0.0021340115927159786, + 0.00396641343832016, + 0.008902693167328835, + 0.003222514409571886, + 0.003855760209262371, + 0.006241474766284227, + 0.0029975157231092453, + 0.005343433003872633, + 0.005658688023686409, + 0.004291827324777842, + 0.004803880117833614, + 0.0033275301102548838, + 0.0025001978501677513, + 0.014552467502653599, + 0.023089483380317688, + 0.011173700913786888, + 0.013601441867649555, + 0.014274953864514828, + 0.02217850275337696, + 0.009951109066605568, + 0.013468571938574314, + 0.0010609051678329706, + 0.0011150026693940163, + 0.00466433260589838, + 0.0038273376412689686, + 0.0017182447481900454, + 0.0028076458256691694, + 0.0013877245364710689, + 0.001367645338177681, + 0.001743545406498015, + 0.0017443452961742878, + 0.0019397770520299673, + 0.0024119368754327297, + 0.002763730939477682, + 0.0020023270044475794, + 0.0023106373846530914, + 0.0018612765707075596, + 0.005695515312254429, + 0.0034027048386633396, + 0.021393556147813797, + 0.008795215748250484, + 0.012206816114485264, + 0.006182697135955095, + 0.006033141631633043, + 0.012424263171851635, + 0.016815921291708946, + 0.007445657160133123, + 0.016927922144532204, + 0.033837225288152695, + 0.0040323445573449135, + 0.0029722831677645445, + 0.01157282292842865, + 0.005336230620741844, + 0.024673616513609886, + 0.020145336166024208, + 0.05606226250529289, + 0.04599166661500931, + 0.020132724195718765, + 0.0418199859559536, + 0.02791501209139824, + 0.021463628858327866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.01724463701248169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03811703622341156 + ], + "opposite_logprob_tokens": [ + -3.288116931915283 + ], + "logprob_reward": -0.03811703622341156, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03811703622341156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08257696032524109, + 0.1627824604511261, + 0.08520326018333435, + 0.06497043371200562, + 0.05446932464838028, + 0.030898433178663254, + 0.014034051448106766, + 0.10269366204738617, + 0.04556941241025925, + 0.35680198669433594 + ], + "credit_map": [ + 0.010720859281718731, + 0.01698967255651951, + 0.014229662716388702, + 0.00964057631790638, + 0.008825364522635937, + 0.006764405407011509, + 0.007648601662367582, + 0.007757818792015314, + 0.010583272203803062, + 0.005569200497120619, + 0.0430942066013813, + 0.011290479451417923, + 0.006678344216197729, + 0.06595416367053986, + 0.007860862649977207, + 0.011751936748623848, + 0.011771871708333492, + 0.0058581712655723095, + 0.014803165569901466, + 0.0038736588321626186, + 0.009142332710325718, + 0.02247612364590168, + 0.00706520676612854, + 0.010212725028395653, + 0.009679323062300682, + 0.0065689487382769585, + 0.009405023418366909, + 0.00797327607870102, + 0.010257759131491184, + 0.011308891698718071, + 0.00460589537397027, + 0.005171316210180521, + 0.006083375308662653, + 0.010205462574958801, + 0.0059834737330675125, + 0.007551219314336777, + 0.006998712662607431, + 0.007306981366127729, + 0.0046615912579, + 0.005678504705429077, + 0.002356685698032379, + 0.0023009672295302153, + 0.006453626789152622, + 0.005140573251992464, + 0.0029496937058866024, + 0.006414031144231558, + 0.0023178893607109785, + 0.002964964834973216, + 0.0011379002826288342, + 0.0014971679775044322, + 0.0023577422834932804, + 0.0022922223433852196, + 0.0017570813652127981, + 0.0015017808182165027, + 0.0018834843067452312, + 0.0016066714888438582, + 0.007769196759909391, + 0.00428331783041358, + 0.030659785494208336, + 0.011566835455596447, + 0.013524599373340607, + 0.006898207124322653, + 0.008759615942835808, + 0.01923210918903351, + 0.0082288458943367, + 0.0036524508614093065, + 0.007320632226765156, + 0.014475294388830662, + 0.00227077747695148, + 0.0016363495960831642, + 0.005461674649268389, + 0.0025233870837837458, + 0.03371954709291458, + 0.027269350364804268, + 0.07765549421310425, + 0.06296814233064651, + 0.02835770510137081, + 0.059226181358098984, + 0.03788932040333748, + 0.029716281220316887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 7 + ], + "selected_reward": -0.03811703622341156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021940225269645452 + ], + "opposite_logprob_tokens": [ + -6.127193927764893 + ], + "logprob_reward": -0.0021940225269645452, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021940225269645452, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07558664679527283, + 0.06330458074808121, + 0.012900926172733307, + 0.07154817134141922, + 0.10202737897634506, + 0.027370596304535866, + 0.014089586213231087, + 0.1187727227807045, + 0.16035453975200653, + 0.3540448546409607 + ], + "credit_map": [ + 0.01004489604383707, + 0.01756751537322998, + 0.013181864283978939, + 0.009386161342263222, + 0.007899010553956032, + 0.005438061896711588, + 0.00632359879091382, + 0.005745536647737026, + 0.004232869017869234, + 0.0029635692480951548, + 0.016998041421175003, + 0.004580918233841658, + 0.002950341673567891, + 0.023852942511439323, + 0.003408518387004733, + 0.004317375831305981, + 0.0013863032218068838, + 0.001790145062841475, + 0.0017962924903258681, + 0.0012592391576617956, + 0.0014700102619826794, + 0.0021337971556931734, + 0.0019259953405708075, + 0.0011391430161893368, + 0.014123528264462948, + 0.004275480285286903, + 0.017857277765870094, + 0.008703799918293953, + 0.004378541372716427, + 0.009999075904488564, + 0.006103771273046732, + 0.006106698885560036, + 0.011236296035349369, + 0.014653640799224377, + 0.012052075006067753, + 0.012914433144032955, + 0.0098210284486413, + 0.020747017115354538, + 0.008949228562414646, + 0.011653647758066654, + 0.0017918277299031615, + 0.0018841761630028486, + 0.007451667916029692, + 0.0062990919686853886, + 0.0020726968068629503, + 0.0036450934130698442, + 0.0021987794898450375, + 0.002027261070907116, + 0.0013227714225649834, + 0.0015293575124815106, + 0.0015184108633548021, + 0.0016667022136971354, + 0.001962554408237338, + 0.0021748519502580166, + 0.0021896706894040108, + 0.0017252659890800714, + 0.008695520460605621, + 0.005709217395633459, + 0.03801048919558525, + 0.013317132368683815, + 0.011040735058486462, + 0.0082768015563488, + 0.009484312497079372, + 0.024238508194684982, + 0.03153654932975769, + 0.013618270866572857, + 0.02258298359811306, + 0.04976686090230942, + 0.007700867019593716, + 0.006156060378998518, + 0.019279779866337776, + 0.009713164530694485, + 0.03310495242476463, + 0.023482302203774452, + 0.07798776030540466, + 0.05787985026836395, + 0.02924407459795475, + 0.06305795907974243, + 0.03654436394572258, + 0.032743558287620544 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.0021940225269645452 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03368761017918587 + ], + "opposite_logprob_tokens": [ + -3.4086875915527344 + ], + "logprob_reward": -0.03368761017918587, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03368761017918587, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08019664883613586, + 0.07817314565181732, + 0.023103119805455208, + 0.047821931540966034, + 0.3896578848361969, + 0.08515501022338867, + 0.04289505258202553, + 0.04453897103667259, + 0.17319539189338684, + 0.0352628268301487 + ], + "credit_map": [ + 0.01157140638679266, + 0.018534276634454727, + 0.01149372011423111, + 0.011476081795990467, + 0.008127138949930668, + 0.006075198762118816, + 0.006686734966933727, + 0.006232090760022402, + 0.004833409562706947, + 0.005066358018666506, + 0.015894751995801926, + 0.007143829949200153, + 0.00398996053263545, + 0.02975161373615265, + 0.005773528479039669, + 0.005719691514968872, + 0.0026787545066326857, + 0.0020777310710400343, + 0.004954636562615633, + 0.002328014001250267, + 0.0022708994802087545, + 0.003618429647758603, + 0.002016079379245639, + 0.003158576786518097, + 0.007735338062047958, + 0.0038590207695961, + 0.006703411694616079, + 0.009262830950319767, + 0.0046684592962265015, + 0.004684077575802803, + 0.0065067233517766, + 0.004402068443596363, + 0.04705134034156799, + 0.07656779140233994, + 0.03769197314977646, + 0.0439925417304039, + 0.04619025066494942, + 0.06512691080570221, + 0.03154776617884636, + 0.041489291936159134, + 0.006607014685869217, + 0.007303970400243998, + 0.011087799444794655, + 0.01129889115691185, + 0.014481893740594387, + 0.013960769400000572, + 0.011388774029910564, + 0.009025899693369865, + 0.004456028807908297, + 0.004893534816801548, + 0.006692044902592897, + 0.0064549557864665985, + 0.0051042004488408566, + 0.004720397759228945, + 0.004294371698051691, + 0.006279516499489546, + 0.0036595913115888834, + 0.0022343758028000593, + 0.013024752959609032, + 0.004586905241012573, + 0.004601075313985348, + 0.0026963904965668917, + 0.005420902743935585, + 0.008314977400004864, + 0.027720032259821892, + 0.012495441362261772, + 0.030083060264587402, + 0.06003889814019203, + 0.007176585961133242, + 0.00572190573439002, + 0.020422259345650673, + 0.009537212550640106, + 0.0037873873952776194, + 0.0030554074328392744, + 0.009346893057227135, + 0.005704488139599562, + 0.002330763963982463, + 0.003503476968035102, + 0.00447037722915411, + 0.0030640331096947193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.03368761017918587 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016063792631030083 + ], + "opposite_logprob_tokens": [ + -4.141063690185547 + ], + "logprob_reward": -0.016063792631030083, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016063792631030083, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03168165683746338, + 0.21479454636573792, + 0.01811262220144272, + 0.02111184038221836, + 0.032058827579021454, + 0.03213132545351982, + 0.023740693926811218, + 0.07637912780046463, + 0.533481776714325, + 0.01650761254131794 + ], + "credit_map": [ + 0.0036717832554131746, + 0.004636559635400772, + 0.006057394668459892, + 0.0044295950792729855, + 0.0032448379788547754, + 0.002925967564806342, + 0.003301761345937848, + 0.003413755213841796, + 0.013788070529699326, + 0.005356135778129101, + 0.054507505148649216, + 0.013930303044617176, + 0.008372140116989613, + 0.09048216044902802, + 0.012172448448836803, + 0.016185766085982323, + 0.002298906445503235, + 0.0019017203012481332, + 0.002685047220438719, + 0.0018708843272179365, + 0.002425581682473421, + 0.003345468547195196, + 0.0019012122647836804, + 0.0016838028095662594, + 0.0030861487612128258, + 0.0015713580651208758, + 0.0028151250444352627, + 0.003679183777421713, + 0.0024142905604094267, + 0.002912547206506133, + 0.002797665074467659, + 0.0018355220090597868, + 0.0038212849758565426, + 0.0046120909973979, + 0.005112576298415661, + 0.00474883709102869, + 0.0033482674043625593, + 0.00445155706256628, + 0.003200063481926918, + 0.0027641484048217535, + 0.0025924162473529577, + 0.003246949752792716, + 0.004564282950013876, + 0.004680545534938574, + 0.005051250569522381, + 0.005100040230900049, + 0.0036625463981181383, + 0.003233292605727911, + 0.0016548701096326113, + 0.0032290818635374308, + 0.004383475985378027, + 0.0031976664904505014, + 0.004511131905019283, + 0.001828636508435011, + 0.0024401983246207237, + 0.002495633438229561, + 0.004800771828740835, + 0.004022980574518442, + 0.0158506091684103, + 0.006803040858358145, + 0.013990347273647785, + 0.007597924210131168, + 0.010474500246345997, + 0.012838955037295818, + 0.0880725234746933, + 0.03866364806890488, + 0.09555986523628235, + 0.18437211215496063, + 0.02270890772342682, + 0.015100385993719101, + 0.061336979269981384, + 0.027667339891195297, + 0.0012866501929238439, + 0.0015439693816006184, + 0.003681729082018137, + 0.0032193518709391356, + 0.0011595282703638077, + 0.0015681396471336484, + 0.0020404241513460875, + 0.002007820410653949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 7 + ], + "selected_reward": -0.016063792631030083 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot begins in medium shot, then moves in a steady, fluid motion, backward and upward. The image initially focuses on a mythical, multicolored phoenix, soaring above a vast sea of clouds. The phoenix's feathers shimmer like silk and brocade: a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail, each feather edged with a shimmering golden thread. Illuminated by the warm golden light of the setting sun, its form appears shimmering. The phoenix flaps its massive wings in a majestic and rhythmic motion, each downward stroke powerful yet effortless, its long, resplendent tail gracefully flowing behind it like a rainbow ribbon. As the camera slowly pulls away, the viewer's perspective expands dramatically, the phoenix's once-dominant grandeur gradually dwarfed by the boundless, swirling cotton candy clouds below. Sunlight penetrates the gaps in the clouds, casting magnificent beams of light and dyeing the sea of clouds into a golden ocean. Ultimately, the phoenix transforms into a dazzling speck of color on the distant skyline, solitary and moving across the vastness of heaven and earth. The entire scene exudes a sense of ultimate epicness and solitary grandeur, a tranquil and sacred atmosphere, and warm and brilliant tones.", + "seed": 24557, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_1846_seed_24557_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016244227159768344, + "vlm_logprob": -0.016244227159768344, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.016244227159768344, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021493833512067795 + ], + "opposite_logprob_tokens": [ + -3.896493911743164 + ], + "logprob_reward": -0.021493833512067795, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021493833512067795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06358301639556885, + 0.05853775143623352, + 0.027469318360090256, + 0.04084407538175583, + 0.03283580765128136, + 0.02380308136343956, + 0.06011350452899933, + 0.03495095297694206, + 0.3347424268722534, + 0.3231201171875 + ], + "credit_map": [ + 0.006681915372610092, + 0.0067518833093345165, + 0.007521192543208599, + 0.009286308661103249, + 0.008524427190423012, + 0.005331582855433226, + 0.010740647092461586, + 0.008745062164962292, + 0.005965310614556074, + 0.005400627385824919, + 0.010522758588194847, + 0.009432143531739712, + 0.005643235519528389, + 0.006120258942246437, + 0.007397442124783993, + 0.008055972866714, + 0.0028457206208258867, + 0.0035175587981939316, + 0.003986262250691652, + 0.0035839215852320194, + 0.002461902564391494, + 0.002263909438624978, + 0.004283118527382612, + 0.004526924807578325, + 0.00526199908927083, + 0.004858631175011396, + 0.004606383386999369, + 0.005761373788118362, + 0.003353486768901348, + 0.003616605419665575, + 0.003753422759473324, + 0.0096321702003479, + 0.0031077659223228693, + 0.00231356848962605, + 0.009796176105737686, + 0.008887181058526039, + 0.0021628134418278933, + 0.0018116961000487208, + 0.0022877748124301434, + 0.0024688320700079203, + 0.0020438190549612045, + 0.0023703076876699924, + 0.0037636274937540293, + 0.0034785836469382048, + 0.002359469421207905, + 0.0023290631361305714, + 0.002747149672359228, + 0.004711061716079712, + 0.0065301563590765, + 0.017489375546574593, + 0.00688902335241437, + 0.007265107240527868, + 0.005088489037007093, + 0.004923260770738125, + 0.005898008123040199, + 0.006030085030943155, + 0.0032360213808715343, + 0.003910311497747898, + 0.004902257584035397, + 0.00512720039114356, + 0.004112287424504757, + 0.007294639479368925, + 0.003190672257915139, + 0.003177562030032277, + 0.012955285608768463, + 0.01394047774374485, + 0.013848789036273956, + 0.09713206440210342, + 0.020939350128173828, + 0.03822914510965347, + 0.11866562068462372, + 0.01903168112039566, + 0.011610664427280426, + 0.009144199080765247, + 0.00665862113237381, + 0.0060717337764799595, + 0.13984332978725433, + 0.13552986085414886, + 0.006916459649801254, + 0.007345241494476795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.021493833512067795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020597022026777267 + ], + "opposite_logprob_tokens": [ + -3.895596981048584 + ], + "logprob_reward": -0.020597022026777267, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020597022026777267, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15938030183315277, + 0.08215578645467758, + 0.0684957504272461, + 0.06069290637969971, + 0.06168113648891449, + 0.07021260261535645, + 0.04153314232826233, + 0.059973686933517456, + 0.19387951493263245, + 0.20199522376060486 + ], + "credit_map": [ + 0.023772140964865685, + 0.018244879320263863, + 0.020644890144467354, + 0.020093603059649467, + 0.01528296060860157, + 0.014542261138558388, + 0.022657906636595726, + 0.024141645058989525, + 0.008664916269481182, + 0.009768659248948097, + 0.011852903291583061, + 0.011301708407700062, + 0.0102543281391263, + 0.008855938911437988, + 0.01016960572451353, + 0.011287718079984188, + 0.00695151137188077, + 0.00926932878792286, + 0.010870331898331642, + 0.009049953892827034, + 0.00627750251442194, + 0.006106370594352484, + 0.009340445511043072, + 0.010630298405885696, + 0.008656556718051434, + 0.006651988252997398, + 0.006400557700544596, + 0.008477892726659775, + 0.004700470715761185, + 0.004848542623221874, + 0.006037644110620022, + 0.014919252134859562, + 0.005621338728815317, + 0.005234723445028067, + 0.016601935029029846, + 0.013815529644489288, + 0.0056775836274027824, + 0.004231583792716265, + 0.004822087939828634, + 0.00567635428160429, + 0.006893605459481478, + 0.008584314957261086, + 0.007546187844127417, + 0.008218914270401001, + 0.008462616242468357, + 0.010204308666288853, + 0.010507914237678051, + 0.009794740937650204, + 0.004051157273352146, + 0.004349075257778168, + 0.006485643796622753, + 0.006541458424180746, + 0.0034790874924510717, + 0.008782180026173592, + 0.0038841047789901495, + 0.003960428759455681, + 0.006222644355148077, + 0.006574260536581278, + 0.008710425347089767, + 0.011485788971185684, + 0.005362238269299269, + 0.008460302837193012, + 0.008026132360100746, + 0.005131890997290611, + 0.010913305915892124, + 0.011031627655029297, + 0.012043217197060585, + 0.04988565668463707, + 0.01620166376233101, + 0.031094614416360855, + 0.05102850869297981, + 0.011680902913212776, + 0.00809814315289259, + 0.009245780296623707, + 0.007374450098723173, + 0.007283075246959925, + 0.08180586993694305, + 0.07549834996461868, + 0.007127894554287195, + 0.005561647471040487 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.020597022026777267 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005247035529464483 + ], + "opposite_logprob_tokens": [ + -5.255247116088867 + ], + "logprob_reward": -0.005247035529464483, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005247035529464483, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.012281467206776142, + 0.009458563290536404, + 0.011074166744947433, + 0.018717311322689056, + 0.011772628873586655, + 0.008046909235417843, + 0.009068749845027924, + 0.008996382355690002, + 0.10080283880233765, + 0.8097809553146362 + ], + "credit_map": [ + 0.00197645858861506, + 0.0014380773063749075, + 0.0014576570829376578, + 0.001624893513508141, + 0.0014198923017829657, + 0.0012458739802241325, + 0.001784338615834713, + 0.0013342763995751739, + 0.0009968981612473726, + 0.000890866038389504, + 0.0018470860086381435, + 0.0016262305434793234, + 0.0010475954040884972, + 0.0008216550340875983, + 0.0011607599444687366, + 0.0010674720397219062, + 0.0013129226863384247, + 0.0011831800220534205, + 0.002025581430643797, + 0.0023467543069273233, + 0.0006551143014803529, + 0.0007802839390933514, + 0.0013649303000420332, + 0.001405400107614696, + 0.0027856710366904736, + 0.0028835495468229055, + 0.002809533616527915, + 0.0036012562923133373, + 0.0012023710878565907, + 0.0011064307764172554, + 0.0017904401756823063, + 0.002538059139624238, + 0.0012503807665780187, + 0.0007252328796312213, + 0.004048280883580446, + 0.0033504641614854336, + 0.0005178962601348758, + 0.0005095293745398521, + 0.0006720786914229393, + 0.000698765623383224, + 0.00133999134413898, + 0.0011209883959963918, + 0.000882960855960846, + 0.0009924821788445115, + 0.0007022118661552668, + 0.0005612592794932425, + 0.001021405216306448, + 0.001425609807483852, + 0.0008928323513828218, + 0.0018130821408703923, + 0.0016561007359996438, + 0.001483178697526455, + 0.0006136794690974057, + 0.0007961065275594592, + 0.0009226223337464035, + 0.0008911470067687333, + 0.0010474073933437467, + 0.0010710787028074265, + 0.001100946799851954, + 0.0014683695044368505, + 0.0007496695034205914, + 0.001641805749386549, + 0.0010881987400352955, + 0.0008289064280688763, + 0.004960448481142521, + 0.005692071281373501, + 0.0051861354149878025, + 0.02523503080010414, + 0.007787061855196953, + 0.016187047585844994, + 0.030109066516160965, + 0.00564597686752677, + 0.02866831235587597, + 0.022302111610770226, + 0.014377020299434662, + 0.014647883363068104, + 0.35412827134132385, + 0.34106695652008057, + 0.017533376812934875, + 0.017057014629244804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.005247035529464483 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789643362164497 + ], + "opposite_logprob_tokens": [ + -3.529789686203003 + ], + "logprob_reward": -0.029789643362164497, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789643362164497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04649440571665764, + 0.03241317719221115, + 0.03934745863080025, + 0.020511923357844353, + 0.051571544259786606, + 0.023944687098264694, + 0.07849310338497162, + 0.11116167902946472, + 0.26206091046333313, + 0.3340011537075043 + ], + "credit_map": [ + 0.006768946070224047, + 0.004381848964840174, + 0.005282405763864517, + 0.006947522051632404, + 0.004992800764739513, + 0.003970686811953783, + 0.0075812023133039474, + 0.006568987853825092, + 0.003946772776544094, + 0.0026363783981651068, + 0.00467213848605752, + 0.0066809747368097305, + 0.002518459688872099, + 0.0024632641579955816, + 0.003944374155253172, + 0.005550813395529985, + 0.004205398727208376, + 0.0043024043552577496, + 0.006750872824341059, + 0.007040147669613361, + 0.003122120862826705, + 0.0032803991343826056, + 0.005020817741751671, + 0.005625292658805847, + 0.002438562223687768, + 0.0023651106748729944, + 0.00217186426743865, + 0.003529874375090003, + 0.0013994999462738633, + 0.0013404203345999122, + 0.0026835023891180754, + 0.00458308681845665, + 0.006862512789666653, + 0.004303935915231705, + 0.015205143950879574, + 0.013015667907893658, + 0.0025932337157428265, + 0.003227352863177657, + 0.0034033535048365593, + 0.0029603412840515375, + 0.0022747234907001257, + 0.002905121771618724, + 0.002513580722734332, + 0.0020994171500205994, + 0.0031939882319420576, + 0.004238917492330074, + 0.003300474025309086, + 0.0034184621181339025, + 0.007938304916024208, + 0.017395587638020515, + 0.009657735005021095, + 0.008057039231061935, + 0.009109252132475376, + 0.007090815808624029, + 0.00975733995437622, + 0.009487024508416653, + 0.01149197481572628, + 0.015455280430614948, + 0.012222285382449627, + 0.01189789455384016, + 0.014353468082845211, + 0.02541392482817173, + 0.00985321681946516, + 0.010473622009158134, + 0.012313791550695896, + 0.01340320985764265, + 0.011925860308110714, + 0.07497250288724899, + 0.018579872325062752, + 0.03347953036427498, + 0.08355443924665451, + 0.01383167877793312, + 0.012315143831074238, + 0.00990897137671709, + 0.007656889967620373, + 0.008147253654897213, + 0.14008715748786926, + 0.138163223862648, + 0.008834267035126686, + 0.008888235315680504 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.029789643362164497 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093601368367672 + ], + "opposite_logprob_tokens": [ + -5.504093647003174 + ], + "logprob_reward": -0.004093601368367672, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093601368367672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.018199997022747993, + 0.01375581230968237, + 0.019825227558612823, + 0.022085899487137794, + 0.017241228371858597, + 0.015526078641414642, + 0.0683860331773758, + 0.04774966835975647, + 0.22949554026126862, + 0.5477344989776611 + ], + "credit_map": [ + 0.0030880626291036606, + 0.002117605647072196, + 0.0023287164513021708, + 0.002452593995258212, + 0.0016971180448308587, + 0.0018344586715102196, + 0.0028168498538434505, + 0.0018645919626578689, + 0.001291100401431322, + 0.001569282147102058, + 0.001720933592878282, + 0.0018647987162694335, + 0.002079031430184841, + 0.0011541028507053852, + 0.0020292364060878754, + 0.00204732664860785, + 0.002191185252740979, + 0.0019187013385817409, + 0.002860057633370161, + 0.0034271704498678446, + 0.00194897863548249, + 0.0015796443913131952, + 0.0035613582003861666, + 0.0023381314240396023, + 0.0025178599171340466, + 0.003801154438406229, + 0.0027093759272247553, + 0.00390731543302536, + 0.0017648086650297046, + 0.001736397622153163, + 0.0023705840576440096, + 0.0032784035429358482, + 0.00275829597376287, + 0.0021608001552522182, + 0.0017249681986868382, + 0.0020870387088507414, + 0.0015561311738565564, + 0.001419802545569837, + 0.0024522067978978157, + 0.0030819850508123636, + 0.0028135485481470823, + 0.0021013752557337284, + 0.0014428449794650078, + 0.0015945574268698692, + 0.0014346666866913438, + 0.0016037051100283861, + 0.0020188242197036743, + 0.0025165569968521595, + 0.005550599191337824, + 0.01889757812023163, + 0.007961905561387539, + 0.0076926397159695625, + 0.007582829799503088, + 0.007073812652379274, + 0.00721453782171011, + 0.0064121391624212265, + 0.003553630318492651, + 0.006966954562813044, + 0.005247101653367281, + 0.006471110042184591, + 0.006021846551448107, + 0.00956847332417965, + 0.005728176794946194, + 0.004192371852695942, + 0.009286927059292793, + 0.010269955731928349, + 0.009166480042040348, + 0.06786036491394043, + 0.014592951163649559, + 0.02681455761194229, + 0.07897105067968369, + 0.012533250264823437, + 0.01964055933058262, + 0.014982013031840324, + 0.010840845294296741, + 0.009910563006997108, + 0.2373885065317154, + 0.23033533990383148, + 0.01204506866633892, + 0.01259161438792944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 6 + ], + "selected_reward": -0.004093601368367672 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot begins in medium shot, then moves in a steady, fluid motion, backward and upward. The image initially focuses on a mythical, multicolored phoenix, soaring above a vast sea of clouds. The phoenix's feathers shimmer like silk and brocade: a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail, each feather edged with a shimmering golden thread. Illuminated by the warm golden light of the setting sun, its form appears shimmering. The phoenix flaps its massive wings in a majestic and rhythmic motion, each downward stroke powerful yet effortless, its long, resplendent tail gracefully flowing behind it like a rainbow ribbon. As the camera slowly pulls away, the viewer's perspective expands dramatically, the phoenix's once-dominant grandeur gradually dwarfed by the boundless, swirling cotton candy clouds below. Sunlight penetrates the gaps in the clouds, casting magnificent beams of light and dyeing the sea of clouds into a golden ocean. Ultimately, the phoenix transforms into a dazzling speck of color on the distant skyline, solitary and moving across the vastness of heaven and earth. The entire scene exudes a sense of ultimate epicness and solitary grandeur, a tranquil and sacred atmosphere, and warm and brilliant tones.", + "seed": 24558, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_1846_seed_24558_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.019192208838649093, + "vlm_logprob": -0.019192208838649093, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.019192208838649093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015237473882734776 + ], + "opposite_logprob_tokens": [ + -4.265237331390381 + ], + "logprob_reward": -0.015237473882734776, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015237473882734776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1044006198644638, + 0.047493137419223785, + 0.09948267042636871, + 0.1465095430612564, + 0.06680600345134735, + 0.11946796625852585, + 0.15100474655628204, + 0.043732963502407074, + 0.0319163016974926, + 0.18918609619140625 + ], + "credit_map": [ + 0.00966699793934822, + 0.012397317215800285, + 0.016542261466383934, + 0.018119603395462036, + 0.013199223205447197, + 0.007274908944964409, + 0.015826338902115822, + 0.011373961344361305, + 0.006418644450604916, + 0.005225688219070435, + 0.011223872192203999, + 0.0056143514811992645, + 0.00603788485750556, + 0.004325569607317448, + 0.004032651428133249, + 0.004614468663930893, + 0.009723210707306862, + 0.009177502244710922, + 0.0239077340811491, + 0.01589355245232582, + 0.006133805960416794, + 0.006727647501975298, + 0.014321647584438324, + 0.01359756663441658, + 0.014682231470942497, + 0.0161180067807436, + 0.024793650954961777, + 0.019471535459160805, + 0.01738511584699154, + 0.013441930525004864, + 0.020423073321580887, + 0.020193982869386673, + 0.008471508510410786, + 0.008124635554850101, + 0.011772123165428638, + 0.010952125303447247, + 0.0069053396582603455, + 0.0058033219538629055, + 0.00698675774037838, + 0.007790188770741224, + 0.012864789925515652, + 0.012682647444307804, + 0.018310369923710823, + 0.018206220120191574, + 0.014465555548667908, + 0.01570730283856392, + 0.014843405224382877, + 0.012387670576572418, + 0.014403428882360458, + 0.03146683797240257, + 0.013889072462916374, + 0.013999557122588158, + 0.012717248871922493, + 0.008272765204310417, + 0.021514957770705223, + 0.0347408726811409, + 0.00794901978224516, + 0.006973897572606802, + 0.004856480751186609, + 0.004085622262209654, + 0.005124371033161879, + 0.005061514209955931, + 0.006030008662492037, + 0.0036520431749522686, + 0.0025061180349439383, + 0.0036780918017029762, + 0.003573925234377384, + 0.003785975743085146, + 0.0045938813127577305, + 0.004532069433480501, + 0.006540987174957991, + 0.002705252729356289, + 0.004029040224850178, + 0.005563931539654732, + 0.013272015377879143, + 0.08174067735671997, + 0.0055060433223843575, + 0.0034210013691335917, + 0.06928066164255142, + 0.006372709758579731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 3 + ], + "selected_reward": -0.015237473882734776 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020596204325556755 + ], + "opposite_logprob_tokens": [ + -3.8955962657928467 + ], + "logprob_reward": -0.020596204325556755, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020596204325556755, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08460009843111038, + 0.04625227302312851, + 0.10333719104528427, + 0.03516006842255592, + 0.03588119521737099, + 0.07673545181751251, + 0.12748393416404724, + 0.07173126190900803, + 0.02919498272240162, + 0.3896235525608063 + ], + "credit_map": [ + 0.00927815679460764, + 0.00928953755646944, + 0.012258946895599365, + 0.015613196417689323, + 0.009805893525481224, + 0.005673773121088743, + 0.013348954729735851, + 0.009331641718745232, + 0.0061336904764175415, + 0.004574652761220932, + 0.011127524077892303, + 0.005538645666092634, + 0.006020094733685255, + 0.003426595125347376, + 0.005098411347717047, + 0.0043326569721102715, + 0.009888668544590473, + 0.011236623860895634, + 0.024907799437642097, + 0.013314440846443176, + 0.009008988738059998, + 0.0135099021717906, + 0.008475449867546558, + 0.012995309196412563, + 0.004119868390262127, + 0.004495416302233934, + 0.007102606352418661, + 0.004898407030850649, + 0.0047237020917236805, + 0.002536386251449585, + 0.002952530048787594, + 0.004331153351813555, + 0.003343603340908885, + 0.006549604702740908, + 0.0034220474772155285, + 0.004453080706298351, + 0.003222418949007988, + 0.0038121570833027363, + 0.004568713717162609, + 0.006509570404887199, + 0.008644210174679756, + 0.008482941426336765, + 0.011968133971095085, + 0.011673099361360073, + 0.009230428375303745, + 0.009680918417870998, + 0.008908089250326157, + 0.00814763456583023, + 0.011705136857926846, + 0.03182962164282799, + 0.00451228441670537, + 0.007245227228850126, + 0.010780616663396358, + 0.005383306182920933, + 0.0196286179125309, + 0.036399129778146744, + 0.012688538990914822, + 0.010906181298196316, + 0.0094614801928401, + 0.007252656854689121, + 0.007847685366868973, + 0.007711600977927446, + 0.009404967539012432, + 0.006458154879510403, + 0.002373801777139306, + 0.004041044041514397, + 0.0033843640703707933, + 0.003707882482558489, + 0.004427841864526272, + 0.004128217231482267, + 0.004813614767044783, + 0.002318217186257243, + 0.00791905913501978, + 0.012646813876926899, + 0.030257074162364006, + 0.1689440757036209, + 0.011081808246672153, + 0.0059585124254226685, + 0.13949674367904663, + 0.013319456949830055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 2 + ], + "selected_reward": -0.020596204325556755 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017108578467741609 + ], + "opposite_logprob_tokens": [ + -6.376710891723633 + ], + "logprob_reward": -0.0017108578467741609, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017108578467741609, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08146781474351883, + 0.01892007514834404, + 0.05488651618361473, + 0.07553347200155258, + 0.0750756561756134, + 0.06517156213521957, + 0.23249953985214233, + 0.13688696920871735, + 0.07616317272186279, + 0.18339525163173676 + ], + "credit_map": [ + 0.006798516493290663, + 0.01075061410665512, + 0.012424956075847149, + 0.01438699197024107, + 0.01049219723790884, + 0.006223972421139479, + 0.011544176377356052, + 0.008846396580338478, + 0.0034069756511598825, + 0.00242198770865798, + 0.0026323748752474785, + 0.0018244475359097123, + 0.0029811463318765163, + 0.0015488952631130815, + 0.0017801322974264622, + 0.0023241140879690647, + 0.006274712271988392, + 0.004024202469736338, + 0.011329173110425472, + 0.007253130432218313, + 0.00528473686426878, + 0.0043711368925869465, + 0.008204001933336258, + 0.008145417086780071, + 0.008692718110978603, + 0.008204715326428413, + 0.015574941411614418, + 0.010573657229542732, + 0.0060887387953698635, + 0.005555962678045034, + 0.010584832169115543, + 0.010257896967232227, + 0.009498700499534607, + 0.014853231608867645, + 0.008333385922014713, + 0.009185577742755413, + 0.004796230234205723, + 0.005747654475271702, + 0.008011089637875557, + 0.014649788849055767, + 0.006767910439521074, + 0.006562408991158009, + 0.010342211462557316, + 0.009770338423550129, + 0.006830104161053896, + 0.008248290047049522, + 0.008468764834105968, + 0.008181534707546234, + 0.02082831598818302, + 0.0567912794649601, + 0.006772263441234827, + 0.01430713851004839, + 0.02147291600704193, + 0.009299235418438911, + 0.03529930114746094, + 0.06772909313440323, + 0.02255878783762455, + 0.021654341369867325, + 0.01708269491791725, + 0.01432910654693842, + 0.013384713791310787, + 0.015080457553267479, + 0.019249698147177696, + 0.013547173701226711, + 0.0073454310186207294, + 0.011246111243963242, + 0.008876970037817955, + 0.008968241512775421, + 0.010584249161183834, + 0.010672524571418762, + 0.012477628886699677, + 0.005992008838802576, + 0.005763772409409285, + 0.00925891287624836, + 0.009522814303636551, + 0.06563202291727066, + 0.009000132791697979, + 0.006980325095355511, + 0.06589148193597794, + 0.011345778591930866 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 7 + ], + "selected_reward": -0.0017108578467741609 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04863899573683739 + ], + "opposite_logprob_tokens": [ + -3.0486390590667725 + ], + "logprob_reward": -0.04863899573683739, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04863899573683739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0516706220805645, + 0.10206738114356995, + 0.05770672485232353, + 0.0938311517238617, + 0.061045002192258835, + 0.05620897188782692, + 0.3431629240512848, + 0.14153575897216797, + 0.027635572478175163, + 0.06513587385416031 + ], + "credit_map": [ + 0.004525462165474892, + 0.004136273171752691, + 0.01328203920274973, + 0.007762816734611988, + 0.00639681052416563, + 0.002811998128890991, + 0.006532583385705948, + 0.006222644820809364, + 0.016949785873293877, + 0.007543511688709259, + 0.03578861057758331, + 0.009482004679739475, + 0.010747348889708519, + 0.005969404708594084, + 0.006596599705517292, + 0.008990115486085415, + 0.0054139080457389355, + 0.005500038154423237, + 0.013196216896176338, + 0.007953583262860775, + 0.0053402273915708065, + 0.007208639290183783, + 0.006466715130954981, + 0.006627403199672699, + 0.009384240955114365, + 0.008515453897416592, + 0.012255603447556496, + 0.012726455926895142, + 0.009293636307120323, + 0.009241343475878239, + 0.01924751326441765, + 0.01316690631210804, + 0.007157085929065943, + 0.00828028004616499, + 0.009126397781074047, + 0.008361823856830597, + 0.006028615403920412, + 0.006602972745895386, + 0.006383949890732765, + 0.009103880263864994, + 0.004976336844265461, + 0.004809889476746321, + 0.007990642450749874, + 0.007404498290270567, + 0.006790845654904842, + 0.008115844801068306, + 0.007375188637524843, + 0.008745732717216015, + 0.02943948470056057, + 0.07094646990299225, + 0.027694664895534515, + 0.02645338699221611, + 0.03335517272353172, + 0.01873152144253254, + 0.053809236735105515, + 0.08273299038410187, + 0.02469184435904026, + 0.0227871872484684, + 0.018093544989824295, + 0.014260551892220974, + 0.014436101540923119, + 0.014519767835736275, + 0.019460570067167282, + 0.01328620407730341, + 0.0029343434143811464, + 0.003183382097631693, + 0.0033295361790806055, + 0.0034548104740679264, + 0.004683017730712891, + 0.0037827317137271166, + 0.003424328751862049, + 0.002843423280864954, + 0.0031581686343997717, + 0.006435931660234928, + 0.005673572886735201, + 0.020488714799284935, + 0.004542882554233074, + 0.004111398942768574, + 0.016804777085781097, + 0.0039204321801662445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.04863899573683739 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009777512401342392 + ], + "opposite_logprob_tokens": [ + -4.634777545928955 + ], + "logprob_reward": -0.009777512401342392, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009777512401342392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.028785228729248047, + 0.03558061644434929, + 0.03707633167505264, + 0.05123075470328331, + 0.05105745419859886, + 0.03658967465162277, + 0.19319464266300201, + 0.10290917009115219, + 0.048766348510980606, + 0.41480979323387146 + ], + "credit_map": [ + 0.0031047603115439415, + 0.003979828674346209, + 0.008546850644052029, + 0.003371840575709939, + 0.0027742269448935986, + 0.0020008496940135956, + 0.001540384255349636, + 0.003466489026322961, + 0.004371296148747206, + 0.003016549628227949, + 0.01132075022906065, + 0.0037932354025542736, + 0.00402420898899436, + 0.001880743307992816, + 0.003216245211660862, + 0.00395759055390954, + 0.004996315110474825, + 0.0043343063443899155, + 0.005227954126894474, + 0.0064943027682602406, + 0.003522335784509778, + 0.004924453794956207, + 0.003144312184303999, + 0.004432351794093847, + 0.00488021457567811, + 0.00398607412353158, + 0.009859452955424786, + 0.007282714359462261, + 0.004462279379367828, + 0.0041468446142971516, + 0.008606940507888794, + 0.008006238378584385, + 0.004449001979082823, + 0.010923219844698906, + 0.004967876244336367, + 0.004501785151660442, + 0.0050630103796720505, + 0.004403094295412302, + 0.004714427515864372, + 0.012035045772790909, + 0.004083401523530483, + 0.004464571364223957, + 0.005494958255439997, + 0.004328956827521324, + 0.00472941342741251, + 0.0043054549023509026, + 0.005024779122322798, + 0.0041581448167562485, + 0.01709877699613571, + 0.04042868688702583, + 0.01288959663361311, + 0.015282181091606617, + 0.019942529499530792, + 0.009807146154344082, + 0.029016776010394096, + 0.04872896149754524, + 0.018000870943069458, + 0.015653148293495178, + 0.013010908849537373, + 0.011163090355694294, + 0.010585864074528217, + 0.009783206507563591, + 0.014850067906081676, + 0.009862020611763, + 0.005470365285873413, + 0.007742334622889757, + 0.0061887591145932674, + 0.0060346429236233234, + 0.005304959136992693, + 0.005832732655107975, + 0.008363730274140835, + 0.003828829387202859, + 0.008709229528903961, + 0.012333705089986324, + 0.03081776760518551, + 0.18123890459537506, + 0.01119887363165617, + 0.00643161591142416, + 0.14998063445091248, + 0.014099061489105225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.009777512401342392 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic shot begins in medium shot, then moves in a steady, fluid motion, backward and upward. The image initially focuses on a mythical, multicolored phoenix, soaring above a vast sea of clouds. The phoenix's feathers shimmer like silk and brocade: a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail, each feather edged with a shimmering golden thread. Illuminated by the warm golden light of the setting sun, its form appears shimmering. The phoenix flaps its massive wings in a majestic and rhythmic motion, each downward stroke powerful yet effortless, its long, resplendent tail gracefully flowing behind it like a rainbow ribbon. As the camera slowly pulls away, the viewer's perspective expands dramatically, the phoenix's once-dominant grandeur gradually dwarfed by the boundless, swirling cotton candy clouds below. Sunlight penetrates the gaps in the clouds, casting magnificent beams of light and dyeing the sea of clouds into a golden ocean. Ultimately, the phoenix transforms into a dazzling speck of color on the distant skyline, solitary and moving across the vastness of heaven and earth. The entire scene exudes a sense of ultimate epicness and solitary grandeur, a tranquil and sacred atmosphere, and warm and brilliant tones.", + "seed": 24559, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_12/index_1846_seed_24559_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.010909693175926804, + "vlm_logprob": -0.010909693175926804, + "vlm_yes_no_margin": 4.975, + "vlm_token_credit": -0.010909693175926804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021290214732289314 + ], + "opposite_logprob_tokens": [ + -3.8962903022766113 + ], + "logprob_reward": -0.021290214732289314, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021290214732289314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0801868811249733, + 0.06122104823589325, + 0.07078107446432114, + 0.2509515881538391, + 0.41668590903282166, + 0.03847622498869896, + 0.025837112218141556, + 0.027554266154766083, + 0.014131145551800728, + 0.014174675568938255 + ], + "credit_map": [ + 0.014760204590857029, + 0.013981360010802746, + 0.010682065039873123, + 0.00812375359237194, + 0.010273450054228306, + 0.007927178405225277, + 0.0053034392185509205, + 0.009135439060628414, + 0.005341819953173399, + 0.0062920101918280125, + 0.011764843948185444, + 0.008799374103546143, + 0.008039119653403759, + 0.010260909795761108, + 0.0052101570181548595, + 0.005512821488082409, + 0.013150286860764027, + 0.007157218176871538, + 0.02346319705247879, + 0.007061411626636982, + 0.00360044464468956, + 0.007181962486356497, + 0.003824039828032255, + 0.005342516582459211, + 0.05323035642504692, + 0.02273590676486492, + 0.05051286891102791, + 0.029175536707043648, + 0.01694651134312153, + 0.03346763551235199, + 0.015792638063430786, + 0.02909013070166111, + 0.04307473823428154, + 0.0898912101984024, + 0.07505098730325699, + 0.02183619700372219, + 0.034917544573545456, + 0.08625951409339905, + 0.020994383841753006, + 0.04466136172413826, + 0.0026070778258144855, + 0.0038372548297047615, + 0.003889397019520402, + 0.003646576777100563, + 0.008934576995670795, + 0.005747128743678331, + 0.006907720118761063, + 0.002906498033553362, + 0.0017952980706468225, + 0.004554317332804203, + 0.002012274693697691, + 0.002579557243734598, + 0.0032887039706110954, + 0.002007177332416177, + 0.005861078854650259, + 0.0037387050688266754, + 0.002540531335398555, + 0.0030142932664602995, + 0.003912894520908594, + 0.004450174048542976, + 0.002982118632644415, + 0.0033648558892309666, + 0.003445653012022376, + 0.00384374987334013, + 0.001094927778467536, + 0.003569574560970068, + 0.0012742318212985992, + 0.0016257158713415265, + 0.0011439910158514977, + 0.002739396644756198, + 0.0014469462912529707, + 0.0012363618006929755, + 0.00102125178091228, + 0.0017509959870949388, + 0.0012830732157453895, + 0.0022062351927161217, + 0.0018915277905762196, + 0.0021996500436216593, + 0.0016886490629985929, + 0.0021332940086722374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.021290214732289314 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077987849712372 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016077987849712372, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077987849712372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08190406858921051, + 0.06849721819162369, + 0.39113664627075195, + 0.18425223231315613, + 0.10654007643461227, + 0.05450308695435524, + 0.028216350823640823, + 0.026936985552310944, + 0.031518567353487015, + 0.02649477683007717 + ], + "credit_map": [ + 0.01065751165151596, + 0.011344109661877155, + 0.013323716819286346, + 0.008557263761758804, + 0.006464674137532711, + 0.011515381745994091, + 0.007661364506930113, + 0.01238004770129919, + 0.005720712244510651, + 0.01123074907809496, + 0.011385305784642696, + 0.007178034633398056, + 0.010726789012551308, + 0.010250537656247616, + 0.00479750894010067, + 0.007207577116787434, + 0.07338298112154007, + 0.036263447254896164, + 0.11885721236467361, + 0.04388970881700516, + 0.023967763409018517, + 0.041822560131549835, + 0.01900266855955124, + 0.033950336277484894, + 0.02955731935799122, + 0.014949269592761993, + 0.04508284106850624, + 0.02532421052455902, + 0.012852686457335949, + 0.025172045454382896, + 0.011545747518539429, + 0.01976810395717621, + 0.007823680527508259, + 0.016241276636719704, + 0.012079235166311264, + 0.010374078527092934, + 0.011730373837053776, + 0.0186016783118248, + 0.012362769804894924, + 0.017326969653367996, + 0.004076563753187656, + 0.005456180777400732, + 0.006933704949915409, + 0.00779899675399065, + 0.008511539548635483, + 0.006446426268666983, + 0.009676670655608177, + 0.005603004712611437, + 0.0018603572389110923, + 0.005006602965295315, + 0.002717534312978387, + 0.0029842578805983067, + 0.003187898313626647, + 0.0019422388868406415, + 0.007374298758804798, + 0.00314316269941628, + 0.001966642215847969, + 0.0038820714689791203, + 0.0019554393365979195, + 0.004041103180497885, + 0.004233567509800196, + 0.002465296071022749, + 0.005344356410205364, + 0.003048509592190385, + 0.0025731439236551523, + 0.009142023511230946, + 0.002294576494023204, + 0.0036702011711895466, + 0.002973367925733328, + 0.0051558492705225945, + 0.0031395191326737404, + 0.0025698866229504347, + 0.0020650699734687805, + 0.003664134070277214, + 0.002713876310735941, + 0.005126649979501963, + 0.00198856252245605, + 0.0023247820790857077, + 0.004474428482353687, + 0.004137272946536541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.016077987849712372 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010389650706201792 + ], + "opposite_logprob_tokens": [ + -6.876039028167725 + ], + "logprob_reward": -0.0010389650706201792, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010389650706201792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09898869693279266, + 0.04617100581526756, + 0.029249655082821846, + 0.05772557109594345, + 0.4185939431190491, + 0.08382157236337662, + 0.0798485204577446, + 0.05778897553682327, + 0.03135837987065315, + 0.0964537039399147 + ], + "credit_map": [ + 0.02100684680044651, + 0.015002976171672344, + 0.013012103736400604, + 0.010200364515185356, + 0.014815484173595905, + 0.011536592617630959, + 0.005703786853700876, + 0.007710534147918224, + 0.0050661321729421616, + 0.006888015661388636, + 0.006112820003181696, + 0.00716570857912302, + 0.005022366531193256, + 0.007159929256886244, + 0.003907295875251293, + 0.00484874052926898, + 0.004693716298788786, + 0.004505859222263098, + 0.005016935057938099, + 0.0031877385918051004, + 0.003509976901113987, + 0.003299379488453269, + 0.0022812255192548037, + 0.0027548233047127724, + 0.011266320012509823, + 0.005056216847151518, + 0.011156235821545124, + 0.006493269000202417, + 0.0038242670707404613, + 0.006911186035722494, + 0.004073369316756725, + 0.008944706059992313, + 0.04274638742208481, + 0.09089161455631256, + 0.08497276157140732, + 0.021500419825315475, + 0.03068387322127819, + 0.09540381282567978, + 0.015919353812932968, + 0.036475736647844315, + 0.008542225696146488, + 0.010195172391831875, + 0.009841741994023323, + 0.011279498226940632, + 0.013245548121631145, + 0.011807164177298546, + 0.011129721067845821, + 0.0077805048786103725, + 0.005243369843810797, + 0.01411378663033247, + 0.005990976933389902, + 0.007339193485677242, + 0.015353950671851635, + 0.007065295707434416, + 0.01745249517261982, + 0.0072894468903541565, + 0.00506219919770956, + 0.009455333463847637, + 0.004864582326263189, + 0.0067752632312476635, + 0.010654781013727188, + 0.008442657999694347, + 0.0066866581328213215, + 0.005847502965480089, + 0.003541426733136177, + 0.0047880555503070354, + 0.0032680302392691374, + 0.004772720858454704, + 0.004371552262455225, + 0.003543594852089882, + 0.0032111522741615772, + 0.003861846635118127, + 0.006108738016337156, + 0.011643250472843647, + 0.009404661133885384, + 0.016394617035984993, + 0.011792229488492012, + 0.011872353032231331, + 0.013491783291101456, + 0.01574607379734516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 9 + ], + "selected_reward": -0.0010389650706201792 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012529231607913971 + ], + "opposite_logprob_tokens": [ + -4.387529373168945 + ], + "logprob_reward": -0.012529231607913971, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012529231607913971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048623643815517426, + 0.023961879312992096, + 0.08416738361120224, + 0.08706247806549072, + 0.34195107221603394, + 0.11062251031398773, + 0.18589302897453308, + 0.05314629152417183, + 0.019862130284309387, + 0.04470958560705185 + ], + "credit_map": [ + 0.006828846875578165, + 0.009674374014139175, + 0.00708431750535965, + 0.0060495841316878796, + 0.007134859450161457, + 0.004511380102485418, + 0.0032222128938883543, + 0.004118066746741533, + 0.0019632200710475445, + 0.004402352962642908, + 0.0032026078552007675, + 0.0033433963544666767, + 0.003147521987557411, + 0.002903435844928026, + 0.0018611425766721368, + 0.003138202242553234, + 0.014596337452530861, + 0.008145848289132118, + 0.023331167176365852, + 0.011486676521599293, + 0.006216248031705618, + 0.009066388942301273, + 0.00417700782418251, + 0.0071477084420621395, + 0.01943063735961914, + 0.006224895361810923, + 0.01970522105693817, + 0.01021672785282135, + 0.004885263275355101, + 0.01252180989831686, + 0.005447720643132925, + 0.008630204014480114, + 0.03693360090255737, + 0.07142917811870575, + 0.0571853369474411, + 0.01773584634065628, + 0.026839997619390488, + 0.07464256882667542, + 0.016735771670937538, + 0.04044877365231514, + 0.006702270824462175, + 0.011650231666862965, + 0.011126847006380558, + 0.011771089397370815, + 0.023546341806650162, + 0.01281792763620615, + 0.024054814130067825, + 0.008952988311648369, + 0.011985351331532001, + 0.022752437740564346, + 0.0151585154235363, + 0.022209500893950462, + 0.025288591161370277, + 0.016621597111225128, + 0.050383131951093674, + 0.02149391360580921, + 0.005065590608865023, + 0.007088140118867159, + 0.004220480564981699, + 0.00790343713015318, + 0.00983452145010233, + 0.00706803472712636, + 0.0063209994696080685, + 0.005645087454468012, + 0.002164745470508933, + 0.004338447470217943, + 0.0016464589862152934, + 0.002485276199877262, + 0.0018216656753793359, + 0.0034526572562754154, + 0.002201960189267993, + 0.0017509175231680274, + 0.003404053160920739, + 0.005529565270990133, + 0.004984559025615454, + 0.00840207189321518, + 0.003994933795183897, + 0.004350363742560148, + 0.0068225865252316, + 0.007221451960504055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 5 + ], + "selected_reward": -0.012529231607913971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a mythical phoenix with a vermilion crest, emerald green ruff, royal blue wings, imperial yellow belly, and pure white tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the phoenix's feathers edged with shimmering golden thread and does it flap its massive wings in a rhythmic motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a vast sea of clouds illuminated by the warm golden light of a setting sun?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera begin in a medium shot and move in a steady, fluid motion backward and upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the video show the phoenix transforming into a distant speck of color while beams of sunlight penetrate the clouds?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036120666190981865 + ], + "opposite_logprob_tokens": [ + -5.628612041473389 + ], + "logprob_reward": -0.0036120666190981865, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036120666190981865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04025197774171829, + 0.041714347898960114, + 0.10072542726993561, + 0.04243851825594902, + 0.4435334801673889, + 0.07033570855855942, + 0.11912701278924942, + 0.04824953153729439, + 0.03155503049492836, + 0.06206892803311348 + ], + "credit_map": [ + 0.005043061915785074, + 0.008429448120296001, + 0.005179284140467644, + 0.0038943488616496325, + 0.007084173616021872, + 0.0046651316806674, + 0.0027642962522804737, + 0.003192237578332424, + 0.004372829105705023, + 0.006980538368225098, + 0.00549297733232379, + 0.005385035648941994, + 0.005335536319762468, + 0.0058090598322451115, + 0.003188756760209799, + 0.005149617325514555, + 0.017890136688947678, + 0.009232320822775364, + 0.02655072882771492, + 0.01209560688585043, + 0.008291531354188919, + 0.011539686471223831, + 0.0054611689411103725, + 0.009664258919656277, + 0.007721405941992998, + 0.0029800443444401026, + 0.010733164846897125, + 0.006353321019560099, + 0.0026854088064283133, + 0.005278829485177994, + 0.002475337591022253, + 0.004211007617413998, + 0.04558233171701431, + 0.09502791613340378, + 0.08250410854816437, + 0.022436534985899925, + 0.03424857556819916, + 0.10141497105360031, + 0.019046075642108917, + 0.043272990733385086, + 0.005170522257685661, + 0.007660993374884129, + 0.008902834728360176, + 0.00591079518198967, + 0.012648263946175575, + 0.01216962467879057, + 0.012709243223071098, + 0.005163442809134722, + 0.007399219088256359, + 0.022416358813643456, + 0.009546310640871525, + 0.0160386860370636, + 0.014871503226459026, + 0.008522158488631248, + 0.026759253814816475, + 0.01357352640479803, + 0.005601521581411362, + 0.00606903899461031, + 0.0032753569539636374, + 0.00733167864382267, + 0.007486963178962469, + 0.008528893813490868, + 0.004879865329712629, + 0.005076216533780098, + 0.0035448563285171986, + 0.004422126337885857, + 0.0024218489415943623, + 0.0032855148892849684, + 0.004720884375274181, + 0.006092727184295654, + 0.0033478273544460535, + 0.0037192460149526596, + 0.006540766917169094, + 0.007848058827221394, + 0.005138568114489317, + 0.00901663489639759, + 0.005061442032456398, + 0.009811462834477425, + 0.011701938696205616, + 0.006950065027922392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.0036120666190981865 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_tensor([9350, 9350])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_tensor([9350, 9350])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..188019773678585ae6c586eaf2779700a2ddec77 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_tensor([9350, 9350])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady medium shot, shot from a level angle, places an elderly man, with white hair and youthful features, a gaunt complexion, yet piercing eyes, firmly in the center of the frame. He wears a loose, traditional white silk training suit, tied with a dark cloth belt at his waist, making him stand out against the backdrop of a quaint courtyard. It is afternoon, and the gentle, slanting sun filters through the gnarled branches of a nearby ancient tree, casting dappled shadows on a ground covered in golden and crimson leaves. The elderly man plants his feet firmly, his body slightly lowered. After taking a deep breath, he slowly and steadily thrusts his right palm forward from his waist. As he thrusts his palm forward, an invisible internal energy seems to surge from his palm. The still, dry leaves on the ground directly in front of him begin to tremble restlessly, then are swept up by a gentle yet powerful current, forming a miniature vortex that spirals upward around his arm, the leaves tumbling and dancing in the air. The entire picture presents a warm, slightly faded nostalgic film texture with slight grain and soft vignetting. The sunlight captures the leaves flying in the air, making them shimmer with golden light, creating a profound artistic conception of tranquility, concentration and restrained power.", + "seed": 14299, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14299_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04608167982660234, + "vlm_logprob": -0.04608167982660234, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.04608167982660234, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016529520973563194 + ], + "opposite_logprob_tokens": [ + -5.016529560089111 + ], + "logprob_reward": -0.016529520973563194, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.016529520973563194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21522945165634155, + 0.06888588517904282, + 0.052317239344120026, + 0.05147416517138481, + 0.07609996199607849, + 0.05507094785571098, + 0.07012747973203659, + 0.11013221740722656, + 0.10519418120384216, + 0.19546841084957123 + ], + "credit_map": [ + 0.01847510226070881, + 0.024847056716680527, + 0.020293377339839935, + 0.031039873138070107, + 0.04013016074895859, + 0.021907834336161613, + 0.023866189643740654, + 0.03466988354921341, + 0.006242647301405668, + 0.0076451473869383335, + 0.00959644466638565, + 0.008522860705852509, + 0.010827930644154549, + 0.006192432250827551, + 0.00973598100244999, + 0.010122441686689854, + 0.004644246771931648, + 0.006228448357433081, + 0.010031383484601974, + 0.005835922434926033, + 0.008766472339630127, + 0.005188828334212303, + 0.005157896317541599, + 0.006464048288762569, + 0.0030356640927493572, + 0.003882067743688822, + 0.005135260056704283, + 0.008485590107738972, + 0.009759735316038132, + 0.006072930060327053, + 0.007416191045194864, + 0.0076867323368787766, + 0.007040741387754679, + 0.007603753823786974, + 0.009282058104872704, + 0.014691985212266445, + 0.013967309147119522, + 0.00901885237544775, + 0.008216338232159615, + 0.006278934422880411, + 0.003212876617908478, + 0.004113465081900358, + 0.00984044000506401, + 0.012457753531634808, + 0.007805990520864725, + 0.006127511151134968, + 0.00502332067117095, + 0.006489596329629421, + 0.00974601786583662, + 0.007621086202561855, + 0.010499983094632626, + 0.006507963873445988, + 0.009464398957788944, + 0.009927121922373772, + 0.008382548578083515, + 0.007978365756571293, + 0.011997959576547146, + 0.018176928162574768, + 0.015448169782757759, + 0.010531949810683727, + 0.01238065492361784, + 0.017049776390194893, + 0.008775387890636921, + 0.015771402046084404, + 0.009242895059287548, + 0.011740700341761112, + 0.012560365721583366, + 0.011916733346879482, + 0.015524757094681263, + 0.0166483111679554, + 0.015272028744220734, + 0.012288399040699005, + 0.011815660633146763, + 0.014606935903429985, + 0.01917574554681778, + 0.01219362672418356, + 0.018852896988391876, + 0.01818135753273964, + 0.04966343939304352, + 0.05097876861691475 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016529520973563194 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009790497832000256 + ], + "opposite_logprob_tokens": [ + -4.634790420532227 + ], + "logprob_reward": -0.009790497832000256, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009790497832000256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24578283727169037, + 0.047220610082149506, + 0.037558555603027344, + 0.07925116270780563, + 0.059211473912000656, + 0.03916844725608826, + 0.048342399299144745, + 0.04623451083898544, + 0.26347601413726807, + 0.1337539553642273 + ], + "credit_map": [ + 0.02649499475955963, + 0.03920721262693405, + 0.042916394770145416, + 0.023598982021212578, + 0.026664147153496742, + 0.029162971302866936, + 0.030242500826716423, + 0.027495667338371277, + 0.006802324205636978, + 0.004154975991696119, + 0.009815026074647903, + 0.006584251765161753, + 0.005023302044719458, + 0.005160263739526272, + 0.004310659132897854, + 0.005369814578443766, + 0.004550186451524496, + 0.005029390100389719, + 0.005285772029310465, + 0.004674156662076712, + 0.0050308010540902615, + 0.0040323990397155285, + 0.003958302550017834, + 0.004997550509870052, + 0.004729124251753092, + 0.007522938773036003, + 0.00729780737310648, + 0.012482100166380405, + 0.012609473429620266, + 0.013292097486555576, + 0.009816906414926052, + 0.011500722728669643, + 0.005590149201452732, + 0.0073825097642838955, + 0.008923625573515892, + 0.008601156063377857, + 0.00957388337701559, + 0.007549630012363195, + 0.005918811075389385, + 0.005671708844602108, + 0.0021649168338626623, + 0.0029767851810902357, + 0.0073819756507873535, + 0.011389296501874924, + 0.004082809202373028, + 0.00355907971970737, + 0.00400881189852953, + 0.003604775294661522, + 0.006574205122888088, + 0.005020408425480127, + 0.005371556617319584, + 0.0037492970004677773, + 0.004095262382179499, + 0.006072561256587505, + 0.007727661170065403, + 0.00973145104944706, + 0.006409757304936647, + 0.006668003275990486, + 0.008372818119823933, + 0.004999212920665741, + 0.0036473851650953293, + 0.0052737584337592125, + 0.005610584747046232, + 0.005252995062619448, + 0.021884558722376823, + 0.030620215460658073, + 0.021827425807714462, + 0.02446797490119934, + 0.0620240792632103, + 0.04124412313103676, + 0.03151726350188255, + 0.0298903938382864, + 0.0073058330453932285, + 0.008222188800573349, + 0.010101237334311008, + 0.0066076721996068954, + 0.013509117066860199, + 0.01452548522502184, + 0.03435070067644119, + 0.03913171961903572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.009790497832000256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579854026436806 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579854026436806, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579854026436806, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11845580488443375, + 0.03303518146276474, + 0.024716882035136223, + 0.11861615628004074, + 0.14679951965808868, + 0.2713596522808075, + 0.03868013247847557, + 0.04046466574072838, + 0.12870468199253082, + 0.07916730642318726 + ], + "credit_map": [ + 0.016428407281637192, + 0.014557726681232452, + 0.01660338044166565, + 0.013692716136574745, + 0.01654938794672489, + 0.012051915749907494, + 0.0146094411611557, + 0.013962828554213047, + 0.003626336809247732, + 0.002868502866476774, + 0.005775741767138243, + 0.005622513592243195, + 0.005115718115121126, + 0.0030239205807447433, + 0.0035109228920191526, + 0.003491528332233429, + 0.0031715636141598225, + 0.0026952982880175114, + 0.0020953244529664516, + 0.0035141294356435537, + 0.005264518782496452, + 0.002518847119063139, + 0.002798527479171753, + 0.002658675191923976, + 0.009553533047437668, + 0.011919282376766205, + 0.01533714309334755, + 0.017213676124811172, + 0.017543688416481018, + 0.01785183884203434, + 0.013627438805997372, + 0.01556956022977829, + 0.009005606174468994, + 0.008832571096718311, + 0.010133329778909683, + 0.04518532752990723, + 0.023658234626054764, + 0.012133389711380005, + 0.02069263905286789, + 0.017158422619104385, + 0.018050117418169975, + 0.018633943051099777, + 0.047692831605672836, + 0.08795315772294998, + 0.02302146703004837, + 0.024458713829517365, + 0.02616937831044197, + 0.025380056351423264, + 0.0020011537708342075, + 0.002662951359525323, + 0.0031905262731015682, + 0.0056620812974870205, + 0.004099570214748383, + 0.00403731781989336, + 0.008566395379602909, + 0.008460139855742455, + 0.0038650876376777887, + 0.006567244417965412, + 0.003974759951233864, + 0.0033709451090544462, + 0.008315387181937695, + 0.0037381360307335854, + 0.006344456225633621, + 0.004288650583475828, + 0.009884597733616829, + 0.014494530856609344, + 0.01437620259821415, + 0.014434284530580044, + 0.028678875416517258, + 0.015381642617285252, + 0.0159488245844841, + 0.015505722723901272, + 0.005851428024470806, + 0.004559348803013563, + 0.005228871945291758, + 0.00537040363997221, + 0.016240093857049942, + 0.012985252775251865, + 0.01958334818482399, + 0.00934856478124857 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.020579854026436806 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989253997802734 + ], + "opposite_logprob_tokens": [ + -1.8048925399780273 + ], + "logprob_reward": -0.17989253997802734, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989253997802734, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19607976078987122, + 0.031379058957099915, + 0.05306302383542061, + 0.04278406500816345, + 0.07312171906232834, + 0.0790918618440628, + 0.03434034809470177, + 0.0862116739153862, + 0.3344017267227173, + 0.06952673941850662 + ], + "credit_map": [ + 0.03362417221069336, + 0.018338149413466454, + 0.020507048815488815, + 0.019709542393684387, + 0.02633536420762539, + 0.024923542514443398, + 0.027959126979112625, + 0.024682816118001938, + 0.004005162976682186, + 0.002989361761137843, + 0.0038932536263018847, + 0.006710616406053305, + 0.00427732989192009, + 0.0027176951989531517, + 0.0033815710339695215, + 0.003404070157557726, + 0.006422283593565226, + 0.004173783119767904, + 0.004346726927906275, + 0.007883196696639061, + 0.010322722606360912, + 0.004728215746581554, + 0.008925380185246468, + 0.0062607149593532085, + 0.0021869251504540443, + 0.003703084774315357, + 0.004011397249996662, + 0.00703155854716897, + 0.008434105664491653, + 0.006630047224462032, + 0.005378557834774256, + 0.005408388562500477, + 0.00948045402765274, + 0.005515377502888441, + 0.008636371232569218, + 0.009726341813802719, + 0.010830762796103954, + 0.011640815064311028, + 0.006490551400929689, + 0.010801039636135101, + 0.00821878295391798, + 0.005487990565598011, + 0.016075054183602333, + 0.0163723174482584, + 0.010595130734145641, + 0.007094045635312796, + 0.006027041003108025, + 0.009221500717103481, + 0.002843244932591915, + 0.0028790158685296774, + 0.005113912280648947, + 0.00684511661529541, + 0.003796182107180357, + 0.003740127431228757, + 0.004710326436907053, + 0.004412420559674501, + 0.009157133288681507, + 0.008609842509031296, + 0.011158313602209091, + 0.007131133694201708, + 0.01662384904921055, + 0.008116954006254673, + 0.013735453598201275, + 0.011678989045321941, + 0.02671530656516552, + 0.045089561492204666, + 0.02303563989698887, + 0.03352423384785652, + 0.07776147127151489, + 0.04370680823922157, + 0.040470413863658905, + 0.04409828037023544, + 0.005062940996140242, + 0.004382599610835314, + 0.005315892864018679, + 0.005370340310037136, + 0.014928700402379036, + 0.010966422036290169, + 0.015620212070643902, + 0.007879635319113731 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.17989253997802734 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036159863229840994 + ], + "opposite_logprob_tokens": [ + -5.628615856170654 + ], + "logprob_reward": -0.0036159863229840994, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036159863229840994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21454653143882751, + 0.08630838245153427, + 0.09295108169317245, + 0.07574910670518875, + 0.06755544990301132, + 0.11421190202236176, + 0.057657599449157715, + 0.058093614876270294, + 0.13720369338989258, + 0.09572262316942215 + ], + "credit_map": [ + 0.02920551598072052, + 0.017353860661387444, + 0.022847890853881836, + 0.03217676281929016, + 0.037274885922670364, + 0.020650817081332207, + 0.027564022690057755, + 0.02747277542948723, + 0.012667432427406311, + 0.008080949075520039, + 0.009600721299648285, + 0.01178654097020626, + 0.012546254321932793, + 0.008656478486955166, + 0.012889481149613857, + 0.010080530308187008, + 0.011836319230496883, + 0.01348455436527729, + 0.011386928148567677, + 0.009456364437937737, + 0.010605129413306713, + 0.008736133575439453, + 0.016466550529003143, + 0.010979094542562962, + 0.007534509524703026, + 0.008766730315983295, + 0.012000661343336105, + 0.008927829563617706, + 0.01051916740834713, + 0.009560061618685722, + 0.0077575440518558025, + 0.010682599619030952, + 0.00819857232272625, + 0.007635034155100584, + 0.005851031746715307, + 0.012416056357324123, + 0.014588144607841969, + 0.0067011103965342045, + 0.0076362695544958115, + 0.004529230762273073, + 0.009121639654040337, + 0.008167621679604053, + 0.02903396263718605, + 0.02131621725857258, + 0.01303311437368393, + 0.011591494083404541, + 0.009397033601999283, + 0.012550810351967812, + 0.004662254359573126, + 0.005660561844706535, + 0.004349575378000736, + 0.010442194528877735, + 0.007780969608575106, + 0.004842054098844528, + 0.01026095449924469, + 0.009659035131335258, + 0.005146604962646961, + 0.005200055427849293, + 0.007257615681737661, + 0.005354844965040684, + 0.011281142011284828, + 0.006345170550048351, + 0.010485644452273846, + 0.0070225391536951065, + 0.01267925463616848, + 0.014713989570736885, + 0.011115203611552715, + 0.011987816542387009, + 0.01883431151509285, + 0.015018894337117672, + 0.03149525448679924, + 0.021358981728553772, + 0.01034973468631506, + 0.009780537337064743, + 0.010639254935085773, + 0.006402087397873402, + 0.013297506608068943, + 0.012000611051917076, + 0.01726166531443596, + 0.015991223976016045 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.0036159863229840994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, places an elderly man, with white hair and youthful features, a gaunt complexion, yet piercing eyes, firmly in the center of the frame. He wears a loose, traditional white silk training suit, tied with a dark cloth belt at his waist, making him stand out against the backdrop of a quaint courtyard. It is afternoon, and the gentle, slanting sun filters through the gnarled branches of a nearby ancient tree, casting dappled shadows on a ground covered in golden and crimson leaves. The elderly man plants his feet firmly, his body slightly lowered. After taking a deep breath, he slowly and steadily thrusts his right palm forward from his waist. As he thrusts his palm forward, an invisible internal energy seems to surge from his palm. The still, dry leaves on the ground directly in front of him begin to tremble restlessly, then are swept up by a gentle yet powerful current, forming a miniature vortex that spirals upward around his arm, the leaves tumbling and dancing in the air. The entire picture presents a warm, slightly faded nostalgic film texture with slight grain and soft vignetting. The sunlight captures the leaves flying in the air, making them shimmer with golden light, creating a profound artistic conception of tranquility, concentration and restrained power.", + "seed": 14300, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14300_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.039378436654806136, + "vlm_logprob": -0.039378436654806136, + "vlm_yes_no_margin": 3.9249999284744264, + "vlm_token_credit": -0.039378436654806136, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01633673533797264 + ], + "opposite_logprob_tokens": [ + -4.516336917877197 + ], + "logprob_reward": -0.01633673533797264, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01633673533797264, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0669059306383133, + 0.04738808423280716, + 0.06874606758356094, + 0.34625449776649475, + 0.05480743572115898, + 0.03292611613869667, + 0.1070248931646347, + 0.07879938930273056, + 0.1481759250164032, + 0.048971641808748245 + ], + "credit_map": [ + 0.008507795631885529, + 0.008172821253538132, + 0.01003742590546608, + 0.006961148697882891, + 0.006028852891176939, + 0.0068517401814460754, + 0.008788692764937878, + 0.011557451449334621, + 0.004064604640007019, + 0.008443323895335197, + 0.0053518242202699184, + 0.00592917762696743, + 0.005729525815695524, + 0.005204080604016781, + 0.00700813764706254, + 0.005657409783452749, + 0.009209291078150272, + 0.0096850311383605, + 0.009414220228791237, + 0.007420499809086323, + 0.010358688421547413, + 0.008036353625357151, + 0.007044342812150717, + 0.007577643264085054, + 0.03223221004009247, + 0.019910799339413643, + 0.09198357909917831, + 0.10994458198547363, + 0.01860528439283371, + 0.019898848608136177, + 0.03667859733104706, + 0.017000596970319748, + 0.005870104767382145, + 0.004144101403653622, + 0.014504424296319485, + 0.009710327722132206, + 0.0036097769625484943, + 0.006120704580098391, + 0.005136602558195591, + 0.0057113938964903355, + 0.003459898754954338, + 0.0031966902315616608, + 0.005106746219098568, + 0.0027991924434900284, + 0.004014324862509966, + 0.0064538344740867615, + 0.003467312315478921, + 0.0044281152077019215, + 0.014985064044594765, + 0.00496314512565732, + 0.02975444868206978, + 0.027868380770087242, + 0.004909393843263388, + 0.012864691205322742, + 0.006685966160148382, + 0.004993809852749109, + 0.005940116010606289, + 0.007947862148284912, + 0.01716558262705803, + 0.01704157330095768, + 0.008993254043161869, + 0.009669119492173195, + 0.006904385052621365, + 0.005137493368238211, + 0.009981026872992516, + 0.026160888373851776, + 0.010480416938662529, + 0.01672835275530815, + 0.036190226674079895, + 0.018319331109523773, + 0.01788708195090294, + 0.012428603135049343, + 0.003660701448097825, + 0.008176811039447784, + 0.008050046861171722, + 0.008695312775671482, + 0.005800746846944094, + 0.005334655288606882, + 0.004455565940588713, + 0.00479779951274395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 6 + ], + "selected_reward": -0.01633673533797264 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978813599795103 + ], + "opposite_logprob_tokens": [ + -4.6347880363464355 + ], + "logprob_reward": -0.00978813599795103, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978813599795103, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03951409086585045, + 0.04872070625424385, + 0.024748247116804123, + 0.5993133187294006, + 0.024491261690855026, + 0.04273223876953125, + 0.013189435005187988, + 0.0216579157859087, + 0.13215072453022003, + 0.053482070565223694 + ], + "credit_map": [ + 0.00472973333671689, + 0.004311493132263422, + 0.005881925579160452, + 0.0037616093177348375, + 0.003379436209797859, + 0.006236598826944828, + 0.0036225307267159224, + 0.00759076327085495, + 0.004516664892435074, + 0.003892894834280014, + 0.004159200005233288, + 0.005270121153444052, + 0.0048943571746349335, + 0.005047336220741272, + 0.013401425443589687, + 0.007538700941950083, + 0.003047909587621689, + 0.0026398745831102133, + 0.0031332960352301598, + 0.003283308120444417, + 0.0032369636464864016, + 0.003276247764006257, + 0.0028956171590834856, + 0.0032350283581763506, + 0.056121472269296646, + 0.02870185673236847, + 0.16760626435279846, + 0.19523262977600098, + 0.029175514355301857, + 0.03171377256512642, + 0.06376894563436508, + 0.02699286863207817, + 0.0025913964491337538, + 0.002007716102525592, + 0.005084515083581209, + 0.005718656815588474, + 0.0017798723420128226, + 0.0020606336183845997, + 0.0022308139596134424, + 0.0030176565051078796, + 0.004649080336093903, + 0.0033186168875545263, + 0.00605426449328661, + 0.0036446170415729284, + 0.005556155927479267, + 0.009862273000180721, + 0.0045471759513020515, + 0.005100053269416094, + 0.0017668291693553329, + 0.0012265173718333244, + 0.0017374817980453372, + 0.0014439242659136653, + 0.0017934562638401985, + 0.0026092343032360077, + 0.0015402131248265505, + 0.0010717774275690317, + 0.003300492186099291, + 0.0016463261563330889, + 0.0032500033266842365, + 0.0032400859054178, + 0.0020309214014559984, + 0.003877296345308423, + 0.0021991939283907413, + 0.0021135970018804073, + 0.012477822601795197, + 0.014554898254573345, + 0.015636330470442772, + 0.017252495512366295, + 0.02079283632338047, + 0.021881384775042534, + 0.010476517491042614, + 0.019078420475125313, + 0.011511729098856449, + 0.006675895303487778, + 0.006418060511350632, + 0.006167571991682053, + 0.004717712756246328, + 0.007091580890119076, + 0.004430558066815138, + 0.0064689600840210915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 9 + ], + "selected_reward": -0.00978813599795103 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018182972446084023 + ], + "opposite_logprob_tokens": [ + -4.018182754516602 + ], + "logprob_reward": -0.018182972446084023, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.018182972446084023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03736606612801552, + 0.020701931789517403, + 0.016173867508769035, + 0.44901150465011597, + 0.04351387172937393, + 0.09847051650285721, + 0.2149951010942459, + 0.043374959379434586, + 0.03176450356841087, + 0.044627685099840164 + ], + "credit_map": [ + 0.003757791593670845, + 0.004266466945409775, + 0.0047470275312662125, + 0.004247140605002642, + 0.0026326358783990145, + 0.007528786081820726, + 0.004647103603929281, + 0.005539115518331528, + 0.003457612358033657, + 0.003546635154634714, + 0.0025389574002474546, + 0.002510922262445092, + 0.00221425318159163, + 0.0020034939516335726, + 0.002360776998102665, + 0.002069280482828617, + 0.002000481355935335, + 0.0018397829262539744, + 0.0022098124027252197, + 0.0020850934088230133, + 0.002268963260576129, + 0.002146360697224736, + 0.0017333339201286435, + 0.0018900400027632713, + 0.04415028169751167, + 0.018890490755438805, + 0.1305890530347824, + 0.15212087333202362, + 0.019164105877280235, + 0.020216571167111397, + 0.04752662405371666, + 0.016353502869606018, + 0.004151768982410431, + 0.004588101990520954, + 0.006667672656476498, + 0.006907055154442787, + 0.003170953830704093, + 0.00730762118473649, + 0.005790177267044783, + 0.0049305204302072525, + 0.009771963581442833, + 0.00855265837162733, + 0.011156885884702206, + 0.008080477826297283, + 0.01932353898882866, + 0.017886538058519363, + 0.012286866083741188, + 0.01141158863902092, + 0.01325138844549656, + 0.009735377505421638, + 0.0710197165608406, + 0.06376761198043823, + 0.012115146964788437, + 0.024269938468933105, + 0.01241842657327652, + 0.008417501114308834, + 0.003914884757250547, + 0.0037521349731832743, + 0.006491534877568483, + 0.008361960761249065, + 0.0066764564253389835, + 0.004990590736269951, + 0.0062951939180493355, + 0.0028922001365572214, + 0.0024067459162324667, + 0.003161672968417406, + 0.005727089010179043, + 0.003154545556753874, + 0.0038889560382813215, + 0.0044480981305241585, + 0.002958959434181452, + 0.006018440239131451, + 0.005509634967893362, + 0.005494949407875538, + 0.0052367267198860645, + 0.00702626584097743, + 0.004356018733233213, + 0.00749873137101531, + 0.004163037054240704, + 0.005342322401702404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 5 + ], + "selected_reward": -0.018182972446084023 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280375838279724 + ], + "opposite_logprob_tokens": [ + -2.01780366897583 + ], + "logprob_reward": -0.14280375838279724, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280375838279724, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.027588976547122, + 0.020594626665115356, + 0.011986429803073406, + 0.7236891388893127, + 0.022340847179293633, + 0.04059714078903198, + 0.023213569074869156, + 0.028466682881116867, + 0.05066654086112976, + 0.05085613206028938 + ], + "credit_map": [ + 0.0032263747416436672, + 0.003980017267167568, + 0.0034512230195105076, + 0.0025103280786424875, + 0.0021719583310186863, + 0.005426149349659681, + 0.002972535090520978, + 0.003850391600281, + 0.0024395810905843973, + 0.003507039276883006, + 0.00219117384403944, + 0.0021324814297258854, + 0.0017806417308747768, + 0.0015521291643381119, + 0.0042648338712751865, + 0.0027267467230558395, + 0.0012431541690602899, + 0.0015262224478647113, + 0.0015716617926955223, + 0.0013611209578812122, + 0.0019166691927239299, + 0.0016141110099852085, + 0.0014060953399166465, + 0.0013473951257765293, + 0.06680331379175186, + 0.034322939813137054, + 0.20325422286987305, + 0.24387243390083313, + 0.032975103706121445, + 0.03734392300248146, + 0.07626362890005112, + 0.0288535226136446, + 0.002037281868979335, + 0.002103441162034869, + 0.0034432292450219393, + 0.0028361366130411625, + 0.0014335036976262927, + 0.003988913260400295, + 0.003989162854850292, + 0.0025091778952628374, + 0.00434537511318922, + 0.00336388498544693, + 0.005237605422735214, + 0.003138030646368861, + 0.006358105223625898, + 0.008186470717191696, + 0.005056244786828756, + 0.0049114222638309, + 0.0022580765653401613, + 0.0016135568730533123, + 0.00539463572204113, + 0.0042050681076943874, + 0.0024805546272546053, + 0.0033845463767647743, + 0.0024113659746944904, + 0.0014657648280262947, + 0.002096991054713726, + 0.0030846938025206327, + 0.003400030080229044, + 0.004148127511143684, + 0.004216189030557871, + 0.003945129457861185, + 0.004429199732840061, + 0.003146320814266801, + 0.005375910550355911, + 0.005664062220603228, + 0.006523268762975931, + 0.005617221817374229, + 0.00595219898968935, + 0.009412882849574089, + 0.004237110260874033, + 0.007883887737989426, + 0.007203774992376566, + 0.005164585076272488, + 0.006453851703554392, + 0.006924993824213743, + 0.005102623254060745, + 0.009378107264637947, + 0.004255019593983889, + 0.006373180542141199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.14280375838279724 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978058110922575 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.00978058110922575, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978058110922575, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08042450249195099, + 0.03705742955207825, + 0.05459132418036461, + 0.30087247490882874, + 0.08666108548641205, + 0.04518039897084236, + 0.03793342784047127, + 0.04562061280012131, + 0.15023718774318695, + 0.16142156720161438 + ], + "credit_map": [ + 0.011677826754748821, + 0.006988062057644129, + 0.011968303471803665, + 0.007103681564331055, + 0.007203255780041218, + 0.014536525122821331, + 0.00751191982999444, + 0.013434930704534054, + 0.007043966092169285, + 0.005288528744131327, + 0.005440074484795332, + 0.004147583618760109, + 0.0033047066535800695, + 0.004043085500597954, + 0.0037728105671703815, + 0.0040166741237044334, + 0.007449433207511902, + 0.0053815096616744995, + 0.0071549974381923676, + 0.0054193525575101376, + 0.00731247290968895, + 0.010960030369460583, + 0.005236092023551464, + 0.005677436012774706, + 0.030525822192430496, + 0.017051907256245613, + 0.08081082999706268, + 0.0902724638581276, + 0.016963614150881767, + 0.01603233814239502, + 0.03197677806019783, + 0.017238711938261986, + 0.007955791428685188, + 0.01467669103294611, + 0.00826212763786316, + 0.018003065139055252, + 0.006245021242648363, + 0.0073647950775921345, + 0.013850400224328041, + 0.010303189978003502, + 0.004867375362664461, + 0.007640737108886242, + 0.0060342238284647465, + 0.006678462959825993, + 0.003828084794804454, + 0.005322661250829697, + 0.00534684956073761, + 0.005462002940475941, + 0.003616918344050646, + 0.0031055952422320843, + 0.006947475951164961, + 0.008920056745409966, + 0.004859305918216705, + 0.0043453313410282135, + 0.003254196373745799, + 0.0028845483902841806, + 0.005326970014721155, + 0.005444166716188192, + 0.005449559539556503, + 0.006860740017145872, + 0.005963327828794718, + 0.005266222637146711, + 0.006044834386557341, + 0.005264792125672102, + 0.00953506026417017, + 0.029307974502444267, + 0.008692016825079918, + 0.032669562846422195, + 0.029663708060979843, + 0.016888944432139397, + 0.01572721265256405, + 0.007752707228064537, + 0.010553346946835518, + 0.026344185695052147, + 0.04025976359844208, + 0.028968224301934242, + 0.014188599772751331, + 0.014531048014760017, + 0.012250375933945179, + 0.014326030388474464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 8 + ], + "selected_reward": -0.00978058110922575 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, places an elderly man, with white hair and youthful features, a gaunt complexion, yet piercing eyes, firmly in the center of the frame. He wears a loose, traditional white silk training suit, tied with a dark cloth belt at his waist, making him stand out against the backdrop of a quaint courtyard. It is afternoon, and the gentle, slanting sun filters through the gnarled branches of a nearby ancient tree, casting dappled shadows on a ground covered in golden and crimson leaves. The elderly man plants his feet firmly, his body slightly lowered. After taking a deep breath, he slowly and steadily thrusts his right palm forward from his waist. As he thrusts his palm forward, an invisible internal energy seems to surge from his palm. The still, dry leaves on the ground directly in front of him begin to tremble restlessly, then are swept up by a gentle yet powerful current, forming a miniature vortex that spirals upward around his arm, the leaves tumbling and dancing in the air. The entire picture presents a warm, slightly faded nostalgic film texture with slight grain and soft vignetting. The sunlight captures the leaves flying in the air, making them shimmer with golden light, creating a profound artistic conception of tranquility, concentration and restrained power.", + "seed": 14301, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14301_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04686998790130019, + "vlm_logprob": -0.04686998790130019, + "vlm_yes_no_margin": 3.75, + "vlm_token_credit": -0.04686998790130019, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016637155786156654 + ], + "opposite_logprob_tokens": [ + -4.391637325286865 + ], + "logprob_reward": -0.016637155786156654, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.016637155786156654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20369577407836914, + 0.09546767920255661, + 0.06341689079999924, + 0.06503809988498688, + 0.07829402387142181, + 0.1734100580215454, + 0.08005256950855255, + 0.07479153573513031, + 0.07054800540208817, + 0.09528538584709167 + ], + "credit_map": [ + 0.015893319621682167, + 0.024068623781204224, + 0.023865729570388794, + 0.030842285603284836, + 0.0275846179574728, + 0.026515742763876915, + 0.028955040499567986, + 0.025970399379730225, + 0.006520209368318319, + 0.020583581179380417, + 0.008309119381010532, + 0.017439810559153557, + 0.010493192821741104, + 0.01253081951290369, + 0.012164890766143799, + 0.0074260481633245945, + 0.0034283201675862074, + 0.004679916892200708, + 0.005539201200008392, + 0.008704087696969509, + 0.009944647550582886, + 0.008846329525113106, + 0.013604494743049145, + 0.00866988766938448, + 0.005871991626918316, + 0.010226497426629066, + 0.006662555038928986, + 0.007957774214446545, + 0.005037618800997734, + 0.009448789060115814, + 0.00893833301961422, + 0.0108945332467556, + 0.010001124814152718, + 0.010597631335258484, + 0.012035507708787918, + 0.01225720252841711, + 0.007977786473929882, + 0.00914854183793068, + 0.009092835709452629, + 0.007183389272540808, + 0.006638358347117901, + 0.009583168663084507, + 0.013451635837554932, + 0.011221984401345253, + 0.017299925908446312, + 0.05019429698586464, + 0.014117058366537094, + 0.05090361461043358, + 0.009546109475195408, + 0.015449853613972664, + 0.008795238099992275, + 0.00842630211263895, + 0.012097754515707493, + 0.0074882120825350285, + 0.010168015025556087, + 0.00808107852935791, + 0.004697511438280344, + 0.006870629265904427, + 0.009994426742196083, + 0.007511606905609369, + 0.014135987497866154, + 0.012970787473022938, + 0.009782712906599045, + 0.008827862329781055, + 0.007414978463202715, + 0.009796530939638615, + 0.008935684338212013, + 0.0073141297325491905, + 0.011503363028168678, + 0.008678545244038105, + 0.00872401800006628, + 0.008180747739970684, + 0.011269735172390938, + 0.017947448417544365, + 0.013537277467548847, + 0.009464253671467304, + 0.011051996611058712, + 0.013568849302828312, + 0.0071623800322413445, + 0.011283433996140957 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.016637155786156654 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012543476186692715 + ], + "opposite_logprob_tokens": [ + -4.387543678283691 + ], + "logprob_reward": -0.012543476186692715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012543476186692715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.140221506357193, + 0.1640983521938324, + 0.07658792287111282, + 0.07542511075735092, + 0.1288912445306778, + 0.13583020865917206, + 0.028720442205667496, + 0.08445104956626892, + 0.09262315928936005, + 0.07315094769001007 + ], + "credit_map": [ + 0.01111058983951807, + 0.025810208171606064, + 0.01952839083969593, + 0.028581462800502777, + 0.011162389069795609, + 0.01701321452856064, + 0.013980092480778694, + 0.01303516048938036, + 0.013226291164755821, + 0.04252834990620613, + 0.015247649513185024, + 0.03233536332845688, + 0.011737621389329433, + 0.02356851100921631, + 0.011537914164364338, + 0.013916656374931335, + 0.008616045117378235, + 0.017640165984630585, + 0.007880696095526218, + 0.01037515327334404, + 0.0056950608268380165, + 0.010906611569225788, + 0.0074848150834441185, + 0.007989375852048397, + 0.008765039965510368, + 0.017369452863931656, + 0.007761097047477961, + 0.010165322571992874, + 0.0049750544130802155, + 0.011110554449260235, + 0.007180876098573208, + 0.008097711019217968, + 0.013624557293951511, + 0.03119296208024025, + 0.010602145455777645, + 0.014377264305949211, + 0.016732951626181602, + 0.01848938874900341, + 0.011399354785680771, + 0.012472626753151417, + 0.009546317160129547, + 0.016518747434020042, + 0.01675017550587654, + 0.01073354110121727, + 0.011792103759944439, + 0.028115298599004745, + 0.009438146837055683, + 0.03293588012456894, + 0.0038088462315499783, + 0.0034643858671188354, + 0.003949206322431564, + 0.00233678100630641, + 0.003489505732432008, + 0.0037585748359560966, + 0.003268488682806492, + 0.004644656553864479, + 0.00846694502979517, + 0.011563004925847054, + 0.011466286145150661, + 0.010412286035716534, + 0.008345907554030418, + 0.009958939626812935, + 0.008423893712460995, + 0.015813792124390602, + 0.014919561333954334, + 0.016509726643562317, + 0.014097164385020733, + 0.00788896530866623, + 0.009296886622905731, + 0.010848479345440865, + 0.008174262940883636, + 0.010888110846281052, + 0.008673221804201603, + 0.010834376327693462, + 0.01157395076006651, + 0.006643280852586031, + 0.007128818426281214, + 0.011337030678987503, + 0.005732901394367218, + 0.011227368377149105 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.012543476186692715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808479383587837 + ], + "opposite_logprob_tokens": [ + -3.2880847454071045 + ], + "logprob_reward": -0.03808479383587837, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808479383587837, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20974907279014587, + 0.1387224644422531, + 0.17791959643363953, + 0.08410607278347015, + 0.04301625117659569, + 0.08247962594032288, + 0.040864743292331696, + 0.04930895194411278, + 0.039132244884967804, + 0.1347009539604187 + ], + "credit_map": [ + 0.019300738349556923, + 0.04715941101312637, + 0.022759314626455307, + 0.033557578921318054, + 0.016902977600693703, + 0.03067450411617756, + 0.01940947398543358, + 0.01998506858944893, + 0.013609937392175198, + 0.026806801557540894, + 0.013966143131256104, + 0.025108207017183304, + 0.016979629173874855, + 0.018266303464770317, + 0.013064131140708923, + 0.010921323671936989, + 0.011164757423102856, + 0.016578299924731255, + 0.014431716874241829, + 0.01717217080295086, + 0.035184383392333984, + 0.023234974592924118, + 0.037332482635974884, + 0.022820817306637764, + 0.008901454508304596, + 0.014894038438796997, + 0.008178049698472023, + 0.008275388740003109, + 0.011685897596180439, + 0.01162265706807375, + 0.012492785230278969, + 0.008055802434682846, + 0.0023719514720141888, + 0.005137477535754442, + 0.004175547510385513, + 0.004525644704699516, + 0.00944647565484047, + 0.005252963397651911, + 0.008505687117576599, + 0.003600501688197255, + 0.004643463995307684, + 0.005321533419191837, + 0.010089059360325336, + 0.006872958038002253, + 0.006429123226553202, + 0.016854891553521156, + 0.010172237642109394, + 0.02209634892642498, + 0.00290277018211782, + 0.007529051508754492, + 0.0058719017542898655, + 0.004721862263977528, + 0.006034026388078928, + 0.0045930263586342335, + 0.005120972171425819, + 0.004091132897883654, + 0.0035104562994092703, + 0.0038794069550931454, + 0.007047169841825962, + 0.004392402712255716, + 0.007711910177022219, + 0.00746449688449502, + 0.007957693189382553, + 0.007345416117459536, + 0.0028972311411052942, + 0.003826658008620143, + 0.0035971926990896463, + 0.004341890104115009, + 0.005987366195768118, + 0.004253771621733904, + 0.0072702644392848015, + 0.0069578709080815315, + 0.01587781310081482, + 0.026453018188476562, + 0.01214259210973978, + 0.011756904423236847, + 0.021414807066321373, + 0.019285501912236214, + 0.013004409149289131, + 0.014765914529561996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.03808479383587837 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035033762454987 + ], + "opposite_logprob_tokens": [ + -1.9103503227233887 + ], + "logprob_reward": -0.16035033762454987, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035033762454987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1908690482378006, + 0.12561628222465515, + 0.1230212077498436, + 0.0755511224269867, + 0.07589685916900635, + 0.19009646773338318, + 0.05016427859663963, + 0.05794142931699753, + 0.04061005264520645, + 0.07023322582244873 + ], + "credit_map": [ + 0.018828319385647774, + 0.038722772151231766, + 0.02177003026008606, + 0.026760775595903397, + 0.019144335761666298, + 0.025819510221481323, + 0.017708241939544678, + 0.02211506851017475, + 0.012960867024958134, + 0.02818593569099903, + 0.015331056900322437, + 0.02177463099360466, + 0.011467961594462395, + 0.017941579222679138, + 0.0078699616715312, + 0.010084278881549835, + 0.0054658325389027596, + 0.012281347066164017, + 0.014937681145966053, + 0.013472693040966988, + 0.019182294607162476, + 0.016458895057439804, + 0.024251075461506844, + 0.016971392557024956, + 0.00899632927030325, + 0.015885230153799057, + 0.00990232266485691, + 0.008361905813217163, + 0.006283389404416084, + 0.010639684274792671, + 0.007982797920703888, + 0.007499466184526682, + 0.006352377124130726, + 0.013789277523756027, + 0.011920887045562267, + 0.007598195690661669, + 0.010051648132503033, + 0.009731976315379143, + 0.00922373216599226, + 0.00722876563668251, + 0.006634450517594814, + 0.007474619895219803, + 0.01563130132853985, + 0.008482370525598526, + 0.016691580414772034, + 0.056496694684028625, + 0.014462225139141083, + 0.06422321498394012, + 0.0038975030183792114, + 0.009989487007260323, + 0.007701475638896227, + 0.005762934219092131, + 0.006410466507077217, + 0.0058405110612511635, + 0.0056606498546898365, + 0.004901251755654812, + 0.004898983519524336, + 0.006128378212451935, + 0.006735260598361492, + 0.007129092235118151, + 0.00696538295596838, + 0.007418460678309202, + 0.007498616818338633, + 0.011167255230247974, + 0.004299570340663195, + 0.007152896374464035, + 0.003431859891861677, + 0.004064877051860094, + 0.007554243318736553, + 0.004122387617826462, + 0.005686439573764801, + 0.004297780804336071, + 0.006343602202832699, + 0.01257429551333189, + 0.00745078781619668, + 0.006217322777956724, + 0.008029714226722717, + 0.00771741708740592, + 0.00950261764228344, + 0.012397464364767075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.16035033762454987 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734176073223352 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734176073223352, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734176073223352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22141706943511963, + 0.1214200109243393, + 0.12634144723415375, + 0.0541556216776371, + 0.05988776311278343, + 0.10925435274839401, + 0.048930633813142776, + 0.10839852690696716, + 0.09458071738481522, + 0.05561386048793793 + ], + "credit_map": [ + 0.03817686811089516, + 0.03640288487076759, + 0.027161652222275734, + 0.03466745465993881, + 0.016398416832089424, + 0.027011457830667496, + 0.02205297164618969, + 0.019545355811715126, + 0.013248048722743988, + 0.02222546562552452, + 0.012450695969164371, + 0.028989402577280998, + 0.00917026400566101, + 0.016988229006528854, + 0.010160058736801147, + 0.008187847211956978, + 0.008388538844883442, + 0.01506825815886259, + 0.008665517903864384, + 0.011497555300593376, + 0.026510315015912056, + 0.015415905974805355, + 0.02596641518175602, + 0.014828932471573353, + 0.002734356326982379, + 0.004710428416728973, + 0.007243485655635595, + 0.008956372737884521, + 0.00651911087334156, + 0.006863184738904238, + 0.007058895193040371, + 0.010069784708321095, + 0.006656591780483723, + 0.011120962910354137, + 0.00679697236046195, + 0.007728958036750555, + 0.007497649174183607, + 0.0077946363016963005, + 0.007388215512037277, + 0.004903774708509445, + 0.00802606251090765, + 0.007972568273544312, + 0.010090715251863003, + 0.006356789264827967, + 0.007981954142451286, + 0.030964642763137817, + 0.007538272999227047, + 0.030323345214128494, + 0.006084438879042864, + 0.010754154995083809, + 0.0050683761946856976, + 0.005195400677621365, + 0.005025795660912991, + 0.005502260290086269, + 0.006158004980534315, + 0.005142204463481903, + 0.008226489648222923, + 0.013027295470237732, + 0.014585205353796482, + 0.010048811323940754, + 0.020268037915229797, + 0.017165139317512512, + 0.011623803526163101, + 0.013453739695250988, + 0.008157184347510338, + 0.011887350119650364, + 0.011595349758863449, + 0.010143679566681385, + 0.013665047474205494, + 0.012779711745679379, + 0.011986684054136276, + 0.014365713112056255, + 0.005351801868528128, + 0.010889536701142788, + 0.00696787191554904, + 0.00632598577067256, + 0.005364136304706335, + 0.006951211951673031, + 0.006916124373674393, + 0.006847187411040068 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.006734176073223352 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady medium shot, shot from a level angle, places an elderly man, with white hair and youthful features, a gaunt complexion, yet piercing eyes, firmly in the center of the frame. He wears a loose, traditional white silk training suit, tied with a dark cloth belt at his waist, making him stand out against the backdrop of a quaint courtyard. It is afternoon, and the gentle, slanting sun filters through the gnarled branches of a nearby ancient tree, casting dappled shadows on a ground covered in golden and crimson leaves. The elderly man plants his feet firmly, his body slightly lowered. After taking a deep breath, he slowly and steadily thrusts his right palm forward from his waist. As he thrusts his palm forward, an invisible internal energy seems to surge from his palm. The still, dry leaves on the ground directly in front of him begin to tremble restlessly, then are swept up by a gentle yet powerful current, forming a miniature vortex that spirals upward around his arm, the leaves tumbling and dancing in the air. The entire picture presents a warm, slightly faded nostalgic film texture with slight grain and soft vignetting. The sunlight captures the leaves flying in the air, making them shimmer with golden light, creating a profound artistic conception of tranquility, concentration and restrained power.", + "seed": 14302, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_13/index_9350_seed_14302_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08962069796398282, + "vlm_logprob": -0.08962069796398282, + "vlm_yes_no_margin": 3.375000011920929, + "vlm_token_credit": -0.08962069796398282, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01504948828369379 + ], + "opposite_logprob_tokens": [ + -4.640049457550049 + ], + "logprob_reward": -0.01504948828369379, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01504948828369379, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2222393900156021, + 0.044259652495384216, + 0.02694428525865078, + 0.05366596207022667, + 0.02958511747419834, + 0.04725511372089386, + 0.07571300864219666, + 0.08559994399547577, + 0.06297571957111359, + 0.35176175832748413 + ], + "credit_map": [ + 0.04928642883896828, + 0.019851073622703552, + 0.033181045204401016, + 0.02975422516465187, + 0.014938135631382465, + 0.020717604085803032, + 0.020458023995161057, + 0.034052859991788864, + 0.008859287016093731, + 0.0031981770880520344, + 0.0099015012383461, + 0.005374275613576174, + 0.0028884774073958397, + 0.004304917994886637, + 0.004203558433800936, + 0.005529456306248903, + 0.0032750291284173727, + 0.004680123180150986, + 0.0036471763160079718, + 0.0035886161495000124, + 0.003128619398921728, + 0.00216822256334126, + 0.002465680008754134, + 0.003990817815065384, + 0.01158040203154087, + 0.005116186570376158, + 0.008147252723574638, + 0.005420581437647343, + 0.0035817082971334457, + 0.007866084575653076, + 0.0046771494671702385, + 0.007276597432792187, + 0.004141445271670818, + 0.0048657190054655075, + 0.0052909208461642265, + 0.0051850625313818455, + 0.0029198750853538513, + 0.0014247748767957091, + 0.003196975914761424, + 0.0025603449903428555, + 0.006317051127552986, + 0.005998426582664251, + 0.0056784688495099545, + 0.006508314050734043, + 0.0040971869602799416, + 0.0042443773709237576, + 0.005289345048367977, + 0.009121942333877087, + 0.010487169027328491, + 0.011076709255576134, + 0.009580429643392563, + 0.011425809003412724, + 0.007255980744957924, + 0.009057058952748775, + 0.007983352988958359, + 0.00884649995714426, + 0.013446611352264881, + 0.009077290073037148, + 0.011762682348489761, + 0.006403404753655195, + 0.00779447378590703, + 0.016202658414840698, + 0.007871986366808414, + 0.013040835037827492, + 0.005846306681632996, + 0.006400684360414743, + 0.0074942088685929775, + 0.009932837449014187, + 0.008516416884958744, + 0.007745753508061171, + 0.008014975115656853, + 0.009024540893733501, + 0.06321680545806885, + 0.02698681876063347, + 0.06200329586863518, + 0.02112157642841339, + 0.026780184358358383, + 0.0699455514550209, + 0.043023254722356796, + 0.03868427127599716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.01504948828369379 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370974212884903 + ], + "opposite_logprob_tokens": [ + -3.408709764480591 + ], + "logprob_reward": -0.03370974212884903, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370974212884903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27486154437065125, + 0.04010747745633125, + 0.040965884923934937, + 0.0924658551812172, + 0.020043207332491875, + 0.05372842028737068, + 0.06685832887887955, + 0.27123332023620605, + 0.036306288093328476, + 0.10342966020107269 + ], + "credit_map": [ + 0.047560449689626694, + 0.012999823316931725, + 0.09299016743898392, + 0.03082304634153843, + 0.011651343666017056, + 0.016423748806118965, + 0.016898438334465027, + 0.04551452398300171, + 0.008332604542374611, + 0.0036040989216417074, + 0.006373279727995396, + 0.003996440675109625, + 0.003205584129318595, + 0.004839002154767513, + 0.0031561849173158407, + 0.006600282154977322, + 0.0050536650232970715, + 0.005466990172863007, + 0.0060337805189192295, + 0.006758672650903463, + 0.004407448228448629, + 0.00404994236305356, + 0.003924384713172913, + 0.005270997062325478, + 0.02312585711479187, + 0.004793624859303236, + 0.01569155417382717, + 0.004777661524713039, + 0.0055669876746833324, + 0.019188053905963898, + 0.0049954368732869625, + 0.01432668138295412, + 0.0034129281993955374, + 0.001906006713397801, + 0.003331706626340747, + 0.0026084822602570057, + 0.0015556232538074255, + 0.0016461561899632215, + 0.002033211523666978, + 0.0035490926820784807, + 0.008091919124126434, + 0.00394268287345767, + 0.010295230895280838, + 0.005052079446613789, + 0.004194773733615875, + 0.0061624664813280106, + 0.005667184479534626, + 0.0103220846503973, + 0.011963861994445324, + 0.006358432583510876, + 0.010124345310032368, + 0.005028566811233759, + 0.006367622409015894, + 0.010298078879714012, + 0.005192083772271872, + 0.011525333859026432, + 0.04364858195185661, + 0.025370771065354347, + 0.037963107228279114, + 0.021319488063454628, + 0.024079982191324234, + 0.05211647227406502, + 0.025874052196741104, + 0.040860865265131, + 0.0035059943329542875, + 0.0038962289690971375, + 0.0048772841691970825, + 0.005019954405725002, + 0.0031782896257936954, + 0.0066877431236207485, + 0.0038405172526836395, + 0.0053002736531198025, + 0.02227483130991459, + 0.006640582345426083, + 0.016772828996181488, + 0.008139587938785553, + 0.009477231651544571, + 0.01637478917837143, + 0.011562461033463478, + 0.012187337502837181 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.03370974212884903 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043055519461631775 + ], + "opposite_logprob_tokens": [ + -3.168055534362793 + ], + "logprob_reward": -0.043055519461631775, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043055519461631775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3249497711658478, + 0.06265314668416977, + 0.05509043112397194, + 0.08858077228069305, + 0.0882091224193573, + 0.1353309005498886, + 0.05602078139781952, + 0.031003188341856003, + 0.048566680401563644, + 0.10959520190954208 + ], + "credit_map": [ + 0.09054455161094666, + 0.01409880444407463, + 0.12475671619176865, + 0.026379579678177834, + 0.015148131176829338, + 0.014062904752790928, + 0.01558996643871069, + 0.02436915412545204, + 0.008690619841217995, + 0.006956434343010187, + 0.010529600083827972, + 0.009016921743750572, + 0.005883023142814636, + 0.006830544676631689, + 0.008644713088870049, + 0.006101295351982117, + 0.007533916272222996, + 0.007668602745980024, + 0.008760190568864346, + 0.009436813183128834, + 0.006067104171961546, + 0.0038658208213746548, + 0.0072262948378920555, + 0.004531696438789368, + 0.010551882907748222, + 0.010056347586214542, + 0.013788687996566296, + 0.014676794409751892, + 0.00897340476512909, + 0.004206493496894836, + 0.019645674154162407, + 0.006681486498564482, + 0.010048829019069672, + 0.01047021709382534, + 0.012828373350203037, + 0.015853801742196083, + 0.007027497980743647, + 0.006206422578543425, + 0.014380021952092648, + 0.011393964290618896, + 0.02267550490796566, + 0.014728868380188942, + 0.02174178510904312, + 0.01692258007824421, + 0.010057928040623665, + 0.010283384472131729, + 0.018089162185788155, + 0.020831700414419174, + 0.00788236502557993, + 0.006057862192392349, + 0.008608444593846798, + 0.007069922052323818, + 0.004879281856119633, + 0.007212671916931868, + 0.006073846016079187, + 0.00823639053851366, + 0.005111339967697859, + 0.0037018468137830496, + 0.004744764883071184, + 0.0035160693805664778, + 0.0021966188214719296, + 0.00396161712706089, + 0.0034836139529943466, + 0.004287319723516703, + 0.004817511420696974, + 0.004878746811300516, + 0.006497259251773357, + 0.008328203111886978, + 0.00420704809948802, + 0.006983147468417883, + 0.0058230930007994175, + 0.0070316726341843605, + 0.01499724481254816, + 0.012448963709175587, + 0.015697069466114044, + 0.01325536984950304, + 0.019103363156318665, + 0.010749832727015018, + 0.012483171187341213, + 0.010860204696655273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.043055519461631775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486568331718445 + ], + "opposite_logprob_tokens": [ + -1.2236568927764893 + ], + "logprob_reward": -0.3486568331718445, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.3486568331718445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24590986967086792, + 0.053535912185907364, + 0.05387752875685692, + 0.023258483037352562, + 0.04625708609819412, + 0.13646066188812256, + 0.04301213473081589, + 0.03384394198656082, + 0.1534997522830963, + 0.21034467220306396 + ], + "credit_map": [ + 0.0649569183588028, + 0.015278959646821022, + 0.06738781929016113, + 0.018008898943662643, + 0.014307450503110886, + 0.016369828954339027, + 0.01800541579723358, + 0.03159457445144653, + 0.007732627913355827, + 0.004423480946570635, + 0.009488239884376526, + 0.006499382201582193, + 0.004437734838575125, + 0.006826534401625395, + 0.006128267850726843, + 0.007999643683433533, + 0.006877908483147621, + 0.006422801874577999, + 0.008892040699720383, + 0.009501781314611435, + 0.005308277904987335, + 0.004876525141298771, + 0.005167729686945677, + 0.006830464582890272, + 0.0032633778173476458, + 0.002439627656713128, + 0.0029670901130884886, + 0.0028710574842989445, + 0.0019701372366398573, + 0.0025763660669326782, + 0.004138157237321138, + 0.0030326684936881065, + 0.00734779704362154, + 0.005163347814232111, + 0.008158238604664803, + 0.007423255126923323, + 0.0030661574564874172, + 0.0033185877837240696, + 0.0056542325764894485, + 0.006125468760728836, + 0.020831122994422913, + 0.01315482147037983, + 0.02297273278236389, + 0.017879296094179153, + 0.012714018113911152, + 0.011754509061574936, + 0.013276996091008186, + 0.023877177387475967, + 0.005674409214407206, + 0.004013151861727238, + 0.006922694388777018, + 0.005250470247119665, + 0.00439298665151, + 0.0067263636738061905, + 0.0040871636010706425, + 0.005944893229752779, + 0.004699525889009237, + 0.00326026719994843, + 0.004567421972751617, + 0.0032949652522802353, + 0.0026507643051445484, + 0.005104606039822102, + 0.0038731354288756847, + 0.0063932547345757484, + 0.01753009855747223, + 0.011478717438876629, + 0.019425736740231514, + 0.03232842683792114, + 0.013652908615767956, + 0.015545157715678215, + 0.02152221091091633, + 0.022016491740942, + 0.04105604067444801, + 0.013681421987712383, + 0.03489336371421814, + 0.015838978812098503, + 0.01684834435582161, + 0.041272230446338654, + 0.023310111835598946, + 0.023444190621376038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.3486568331718445 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with white hair and a gaunt complexion wearing a loose white silk training suit with a dark cloth belt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a quaint courtyard with gnarled tree branches and ground covered in golden and crimson leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the man plant his feet, lower his body, and slowly thrust his right palm forward from his waist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do dry leaves on the ground tremble and form a miniature vortex spiraling upward around the man's arm?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a steady medium shot from a level angle with a warm, slightly faded nostalgic film texture including slight grain and soft vignetting?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007631906773895025 + ], + "opposite_logprob_tokens": [ + -4.882631778717041 + ], + "logprob_reward": -0.007631906773895025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007631906773895025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4883120656013489, + 0.05961683392524719, + 0.0774407610297203, + 0.03209173306822777, + 0.04875373840332031, + 0.06359993666410446, + 0.053537726402282715, + 0.05929289385676384, + 0.0365561842918396, + 0.08079814165830612 + ], + "credit_map": [ + 0.10346498340368271, + 0.019908178597688675, + 0.19213822484016418, + 0.04922990873456001, + 0.019396407529711723, + 0.017945382744073868, + 0.02667124755680561, + 0.05955773964524269, + 0.010628590360283852, + 0.0049782367423176765, + 0.013662684708833694, + 0.006528499536216259, + 0.005262749269604683, + 0.005343235097825527, + 0.005936342291533947, + 0.007276496849954128, + 0.009975258260965347, + 0.010772415436804295, + 0.013690115883946419, + 0.013328499160706997, + 0.008895881474018097, + 0.004597421269863844, + 0.008248386904597282, + 0.007932784967124462, + 0.003584307385608554, + 0.003315242938697338, + 0.005274716764688492, + 0.005157641135156155, + 0.002719591837376356, + 0.0026887929998338223, + 0.004532854072749615, + 0.004818589426577091, + 0.008635476231575012, + 0.006137229967862368, + 0.008103703148663044, + 0.008479387499392033, + 0.0031787254847586155, + 0.0033475738018751144, + 0.00442893011495471, + 0.006442718207836151, + 0.007937509566545486, + 0.004162487108260393, + 0.00957528781145811, + 0.006917682476341724, + 0.005762428976595402, + 0.007914958521723747, + 0.007764827460050583, + 0.013564756140112877, + 0.005861018784344196, + 0.00796957965940237, + 0.005187200848013163, + 0.00893094576895237, + 0.005473417229950428, + 0.005835378542542458, + 0.0072161005809903145, + 0.0070640891790390015, + 0.009428928606212139, + 0.0048927851021289825, + 0.008199319243431091, + 0.004693943541496992, + 0.004864418879151344, + 0.011995629407465458, + 0.0060661169700324535, + 0.009151753969490528, + 0.003159352345392108, + 0.003546816296875477, + 0.005345293786376715, + 0.006046687252819538, + 0.002679981291294098, + 0.006319178733974695, + 0.004508985672146082, + 0.004949891474097967, + 0.009623132646083832, + 0.0118327671661973, + 0.013402293436229229, + 0.008808388374745846, + 0.007908916100859642, + 0.00970165990293026, + 0.009365108795464039, + 0.010155875235795975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.007631906773895025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_tensor([2409, 2409])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_tensor([2409, 2409])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4d6abc651de01adf897e6f515cfd6d10c107dfe1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_tensor([2409, 2409])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a dreamlike close-up, the camera performing a slow, continuous roll, causing the entire world to gently rotate. Within this swirling vision, a handmade rag doll lies quietly on a fluffy white pillow, bathed in the warm, rich golden afternoon sun. Its body is sewn from rough, beige linen, the sunlight highlighting the texture of the fabric. Its eyes are two dark buttons of varying sizes: one a simple four-hole plastic button, the other slightly larger and patinaed with wood. A simple smile, embroidered in red thread, is frozen on its face. As the camera rotates slowly, the rag doll seems to succumb to gravity in slow motion and begins a silent slide: its body slides down from the top of the pillow very slowly, the fabric rubbing slightly against the cotton surface of the pillowcase, and finally its soft body stuffed with cotton slides silently off the edge of the pillow. The whole process, amidst the swirling golden halo and flying dust, creates a nostalgic atmosphere that is both peaceful and slightly sad, like a fading childhood memory.", + "seed": 5833, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_2409_seed_5833_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1393347446806729, + "vlm_logprob": -0.1393347446806729, + "vlm_yes_no_margin": 2.95, + "vlm_token_credit": -0.1393347446806729, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4356514513492584 + ], + "opposite_logprob_tokens": [ + -1.060651421546936 + ], + "logprob_reward": -0.4356514513492584, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4356514513492584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04407374560832977, + 0.04272863641381264, + 0.09721168875694275, + 0.09095999598503113, + 0.07434304803609848, + 0.41590428352355957, + 0.06755027920007706, + 0.017497753724455833, + 0.03806304931640625, + 0.11166758090257645 + ], + "credit_map": [ + 0.004944070242345333, + 0.004747925791889429, + 0.007373705971986055, + 0.006548223551362753, + 0.0041709039360284805, + 0.005333986598998308, + 0.004667429253458977, + 0.006287493743002415, + 0.0049058361910283566, + 0.007002435624599457, + 0.004355826415121555, + 0.007520198822021484, + 0.005205304361879826, + 0.006651496049016714, + 0.003291144734248519, + 0.0037963909562677145, + 0.01194535568356514, + 0.011375700123608112, + 0.012720372527837753, + 0.013085169717669487, + 0.01321211364120245, + 0.0110872657969594, + 0.009074659086763859, + 0.014711044728755951, + 0.009263377636671066, + 0.013949712738394737, + 0.008303256705403328, + 0.012632913887500763, + 0.01202323380857706, + 0.010084249079227448, + 0.009148988872766495, + 0.015554259531199932, + 0.010694748722016811, + 0.009199082851409912, + 0.0064443619921803474, + 0.007751720491796732, + 0.008397803641855717, + 0.011268412694334984, + 0.011268503963947296, + 0.009318404830992222, + 0.12093943357467651, + 0.17251329123973846, + 0.013951967470347881, + 0.03437723591923714, + 0.03457794338464737, + 0.019484680145978928, + 0.009636851027607918, + 0.010422858409583569, + 0.0065500978380441666, + 0.006026431918144226, + 0.009004856459796429, + 0.006495482288300991, + 0.01003983337432146, + 0.008282069116830826, + 0.010762727819383144, + 0.010388773865997791, + 0.002124432008713484, + 0.0024584485217928886, + 0.002152706030756235, + 0.001985242823138833, + 0.0018945073243230581, + 0.0015196127351373434, + 0.0018098477739840746, + 0.0035529565066099167, + 0.004173962865024805, + 0.0037107784301042557, + 0.0053065563552081585, + 0.005141245201230049, + 0.004070901311933994, + 0.006408435758203268, + 0.004447943065315485, + 0.004803221207112074, + 0.0058216555044054985, + 0.015603396110236645, + 0.006377859972417355, + 0.037439532577991486, + 0.00514059467241168, + 0.008280354551970959, + 0.013345537707209587, + 0.01965864561498165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 2 + ], + "selected_reward": -0.4356514513492584 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268024746328592 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005268024746328592, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268024746328592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08608854562044144, + 0.027310149744153023, + 0.16522280871868134, + 0.10173613578081131, + 0.09307820349931717, + 0.28833210468292236, + 0.025039518252015114, + 0.025972241535782814, + 0.04779153689742088, + 0.1394287347793579 + ], + "credit_map": [ + 0.010754821822047234, + 0.010473314672708511, + 0.010170376859605312, + 0.012255609035491943, + 0.008647111244499683, + 0.00951923057436943, + 0.01133278850466013, + 0.012935287319123745, + 0.0034767785109579563, + 0.0032991552725434303, + 0.00395627785474062, + 0.0037361839786171913, + 0.00304767326451838, + 0.004266179166734219, + 0.002868293086066842, + 0.0026596076786518097, + 0.016725627705454826, + 0.0192254725843668, + 0.017242399975657463, + 0.025726864114403725, + 0.02210659720003605, + 0.020369302481412888, + 0.018031084910035133, + 0.025795454159379005, + 0.010583454743027687, + 0.01268196851015091, + 0.007282564882189035, + 0.015373175032436848, + 0.010178080759942532, + 0.012803099118173122, + 0.014273438602685928, + 0.018560342490673065, + 0.011559829115867615, + 0.010403245687484741, + 0.0064344629645347595, + 0.010738515295088291, + 0.014720706269145012, + 0.010884272865951061, + 0.019225943833589554, + 0.009111226536333561, + 0.08205712586641312, + 0.12171327322721481, + 0.008800354786217213, + 0.023534826934337616, + 0.024482736364006996, + 0.014377132058143616, + 0.006675136741250753, + 0.006691519636660814, + 0.0030100401490926743, + 0.002337656682357192, + 0.004817677661776543, + 0.0028047373052686453, + 0.0028486980590969324, + 0.0037611399311572313, + 0.002712799469009042, + 0.002746767131611705, + 0.0026919471565634012, + 0.003250249195843935, + 0.003967136610299349, + 0.005848450120538473, + 0.0019844654016196728, + 0.0022272043861448765, + 0.002711679320782423, + 0.0032911060843616724, + 0.004312573466449976, + 0.003986984957009554, + 0.006961367558687925, + 0.006273621693253517, + 0.006053314544260502, + 0.00701752956956625, + 0.005661599338054657, + 0.007524540647864342, + 0.0062912097200751305, + 0.016040269285440445, + 0.005446531809866428, + 0.04954962059855461, + 0.00504412641748786, + 0.008142419159412384, + 0.014229669235646725, + 0.034684885293245316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.005268024746328592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2015988826751709 + ], + "opposite_logprob_tokens": [ + -1.701598882675171 + ], + "logprob_reward": -0.2015988826751709, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.2015988826751709, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11702229827642441, + 0.06233296915888786, + 0.1814698725938797, + 0.04319875314831734, + 0.1420454978942871, + 0.22190970182418823, + 0.0314767025411129, + 0.061714764684438705, + 0.04348551109433174, + 0.09534386545419693 + ], + "credit_map": [ + 0.012476371601223946, + 0.013135337270796299, + 0.01916959322988987, + 0.01984146051108837, + 0.009580074809491634, + 0.013432924635708332, + 0.012962915934622288, + 0.016423627734184265, + 0.007900911383330822, + 0.00971928983926773, + 0.005840871948748827, + 0.010942752473056316, + 0.006343490909785032, + 0.012333543039858341, + 0.0040029860101640224, + 0.00524913240224123, + 0.020166192203760147, + 0.021809855476021767, + 0.01790577732026577, + 0.02670872211456299, + 0.025958241894841194, + 0.021560750901699066, + 0.01830434799194336, + 0.029056016355752945, + 0.004863397218286991, + 0.005539593752473593, + 0.003475993173196912, + 0.00673424918204546, + 0.006152269896119833, + 0.004780194256454706, + 0.004373075906187296, + 0.007279981393367052, + 0.012438417412340641, + 0.01665152981877327, + 0.010555626824498177, + 0.01522716972976923, + 0.02479695715010166, + 0.01663973554968834, + 0.02811506576836109, + 0.017621001228690147, + 0.060404691845178604, + 0.09272411465644836, + 0.006330756936222315, + 0.018317753449082375, + 0.01900405064225197, + 0.012306641787290573, + 0.007184277754276991, + 0.005637416150420904, + 0.0031155350152403116, + 0.0019122990779578686, + 0.0052467091009020805, + 0.00480993464589119, + 0.0023293232079595327, + 0.0036329119466245174, + 0.007207230664789677, + 0.00322276190854609, + 0.00643044151365757, + 0.007558485958725214, + 0.01003765594214201, + 0.009863742627203465, + 0.005389684345573187, + 0.004296706058084965, + 0.008229616098105907, + 0.009908437728881836, + 0.003893428947776556, + 0.004364043474197388, + 0.007478616666048765, + 0.005497738718986511, + 0.0064615970477461815, + 0.0061018881388008595, + 0.004680704791098833, + 0.005007497500628233, + 0.0067878649570047855, + 0.012809955514967442, + 0.006899684201925993, + 0.031924255192279816, + 0.0037422648165374994, + 0.0068881576880812645, + 0.015378273092210293, + 0.0109134241938591 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.2015988826751709 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03809649869799614 + ], + "opposite_logprob_tokens": [ + -3.2880964279174805 + ], + "logprob_reward": -0.03809649869799614, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03809649869799614, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02339964732527733, + 0.0115100909024477, + 0.18656617403030396, + 0.02543039806187153, + 0.04751553013920784, + 0.5742830634117126, + 0.0393458753824234, + 0.03649256005883217, + 0.030039655044674873, + 0.025417091324925423 + ], + "credit_map": [ + 0.0028489460237324238, + 0.0035796030424535275, + 0.0024273497983813286, + 0.005078535992652178, + 0.0022735237143933773, + 0.002991837216541171, + 0.002258345717564225, + 0.0019415069837123156, + 0.0012511267559602857, + 0.0016751298680901527, + 0.001411919016391039, + 0.0019234027713537216, + 0.00118377769831568, + 0.0013462717179208994, + 0.0010185742285102606, + 0.001699889195151627, + 0.020136622712016106, + 0.022080618888139725, + 0.018797410652041435, + 0.028084533289074898, + 0.026120737195014954, + 0.022449200972914696, + 0.019242189824581146, + 0.029654860496520996, + 0.0045561217702925205, + 0.002780649811029434, + 0.002192788990214467, + 0.0031108572147786617, + 0.0021534643601626158, + 0.004055649973452091, + 0.002495435532182455, + 0.004085431341081858, + 0.007529089227318764, + 0.005063881166279316, + 0.0041642505675554276, + 0.004084334243088961, + 0.007148801349103451, + 0.00858330074697733, + 0.006771382875740528, + 0.004170487634837627, + 0.1652940958738327, + 0.2435818463563919, + 0.01709771528840065, + 0.046348851174116135, + 0.04810303449630737, + 0.028585797175765038, + 0.012945804744958878, + 0.012325855903327465, + 0.0035036723129451275, + 0.0041078538633883, + 0.00761849619448185, + 0.00515690678730607, + 0.0038572687190026045, + 0.005119024775922298, + 0.005731512792408466, + 0.004251136910170317, + 0.004170485772192478, + 0.003432078519836068, + 0.005235977936536074, + 0.007566844578832388, + 0.002209528349339962, + 0.002334756078198552, + 0.004903905559331179, + 0.006638980470597744, + 0.0024963002651929855, + 0.0027172830887138844, + 0.0065549686551094055, + 0.004000730812549591, + 0.00309424614533782, + 0.0030180432368069887, + 0.003514696843922138, + 0.0046433862298727036, + 0.0012628171825781465, + 0.003960584755986929, + 0.0011968491598963737, + 0.008495798334479332, + 0.0012068870710209012, + 0.0017811181023716927, + 0.0030300035141408443, + 0.004483034368604422 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.03809649869799614 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016058865934610367 + ], + "opposite_logprob_tokens": [ + -4.141058921813965 + ], + "logprob_reward": -0.016058865934610367, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016058865934610367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0434735082089901, + 0.02943556010723114, + 0.1670287698507309, + 0.01822911947965622, + 0.16520297527313232, + 0.3466181755065918, + 0.036184605211019516, + 0.10466424375772476, + 0.045561570674180984, + 0.04360147565603256 + ], + "credit_map": [ + 0.005152308382093906, + 0.005404054652899504, + 0.007227972615510225, + 0.007595574948936701, + 0.0035195343662053347, + 0.005650815553963184, + 0.004516710992902517, + 0.004406533669680357, + 0.003993664868175983, + 0.004591129254549742, + 0.0022208858281373978, + 0.004907107912003994, + 0.003355439519509673, + 0.004501553252339363, + 0.0030099600553512573, + 0.0028558201156556606, + 0.01779782958328724, + 0.020256847143173218, + 0.015804769471287727, + 0.026029929518699646, + 0.023585854098200798, + 0.020909490063786507, + 0.016751952469348907, + 0.0258921068161726, + 0.0018611940322443843, + 0.0025102521758526564, + 0.0017045323038473725, + 0.002784556010738015, + 0.0027958385180681944, + 0.001956518739461899, + 0.0017256629653275013, + 0.0028905633371323347, + 0.03174654394388199, + 0.018500706180930138, + 0.015345749445259571, + 0.016323700547218323, + 0.019162463024258614, + 0.032055776566267014, + 0.019651221111416817, + 0.012416815385222435, + 0.09904063493013382, + 0.14662358164787292, + 0.010351639240980148, + 0.028125157579779625, + 0.028849957510828972, + 0.017059441655874252, + 0.00855094101279974, + 0.008016816340386868, + 0.003619207302108407, + 0.0035057603381574154, + 0.004850363824516535, + 0.00464647775515914, + 0.004592750687152147, + 0.0054075391963124275, + 0.005495980381965637, + 0.004066525958478451, + 0.010426818393170834, + 0.013931303285062313, + 0.01816880702972412, + 0.016525283455848694, + 0.008126795291900635, + 0.006439608987420797, + 0.012275240384042263, + 0.018770387396216393, + 0.0053377230651676655, + 0.004588407464325428, + 0.005010815337300301, + 0.0059485724195837975, + 0.004939229693263769, + 0.006680995225906372, + 0.006858999375253916, + 0.006196826230734587, + 0.002530808327719569, + 0.0025768086779862642, + 0.0022922265343368053, + 0.013491853140294552, + 0.0023048899602144957, + 0.002881463151425123, + 0.0039107403717935085, + 0.013612685725092888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 4 + ], + "selected_reward": -0.016058865934610367 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dreamlike close-up, the camera performing a slow, continuous roll, causing the entire world to gently rotate. Within this swirling vision, a handmade rag doll lies quietly on a fluffy white pillow, bathed in the warm, rich golden afternoon sun. Its body is sewn from rough, beige linen, the sunlight highlighting the texture of the fabric. Its eyes are two dark buttons of varying sizes: one a simple four-hole plastic button, the other slightly larger and patinaed with wood. A simple smile, embroidered in red thread, is frozen on its face. As the camera rotates slowly, the rag doll seems to succumb to gravity in slow motion and begins a silent slide: its body slides down from the top of the pillow very slowly, the fabric rubbing slightly against the cotton surface of the pillowcase, and finally its soft body stuffed with cotton slides silently off the edge of the pillow. The whole process, amidst the swirling golden halo and flying dust, creates a nostalgic atmosphere that is both peaceful and slightly sad, like a fading childhood memory.", + "seed": 5834, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_2409_seed_5834_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.09806553404778243, + "vlm_logprob": -0.09806553404778243, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.09806553404778243, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21218426525592804 + ], + "opposite_logprob_tokens": [ + -1.7121843099594116 + ], + "logprob_reward": -0.21218426525592804, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.21218426525592804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1631186604499817, + 0.05182316526770592, + 0.06168258562684059, + 0.054947979748249054, + 0.03968668729066849, + 0.09418702125549316, + 0.11456242948770523, + 0.27297666668891907, + 0.0982532799243927, + 0.0487615168094635 + ], + "credit_map": [ + 0.01341431774199009, + 0.018038278445601463, + 0.018612070009112358, + 0.030194411054253578, + 0.012681513093411922, + 0.017817547544836998, + 0.02945024147629738, + 0.02291027456521988, + 0.006843355484306812, + 0.006073832977563143, + 0.003593647852540016, + 0.0088056530803442, + 0.004587145056575537, + 0.005915143992751837, + 0.009553973563015461, + 0.006450410932302475, + 0.010703553445637226, + 0.006094025913625956, + 0.006961478851735592, + 0.00966812577098608, + 0.006056716665625572, + 0.010313001461327076, + 0.006086675450205803, + 0.005799009930342436, + 0.009740279987454414, + 0.006168999709188938, + 0.008701042272150517, + 0.004934671800583601, + 0.006912663578987122, + 0.007009712979197502, + 0.006596305873245001, + 0.004884305410087109, + 0.0046690115705132484, + 0.0038108339067548513, + 0.004877523519098759, + 0.004761971533298492, + 0.005182189401239157, + 0.005601976532489061, + 0.005077109672129154, + 0.005706070922315121, + 0.01145959459245205, + 0.007932750508189201, + 0.017064396291971207, + 0.010553122498095036, + 0.016519859433174133, + 0.012080544605851173, + 0.011274904944002628, + 0.007301854435354471, + 0.016135716810822487, + 0.011447072960436344, + 0.011348748579621315, + 0.015356125310063362, + 0.01573244296014309, + 0.012393785640597343, + 0.021672019734978676, + 0.010476517491042614, + 0.027555106207728386, + 0.01659206487238407, + 0.03609725460410118, + 0.03064832091331482, + 0.042557112872600555, + 0.030570143833756447, + 0.050706684589385986, + 0.038249991834163666, + 0.008663155138492584, + 0.007442358881235123, + 0.012279599905014038, + 0.007396393921226263, + 0.01696346141397953, + 0.018556766211986542, + 0.012366222217679024, + 0.014585316181182861, + 0.004226842429488897, + 0.004723882302641869, + 0.00649366807192564, + 0.004808004014194012, + 0.010145965963602066, + 0.008100666105747223, + 0.005368530750274658, + 0.004893958568572998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.21218426525592804 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012564548291265965 + ], + "opposite_logprob_tokens": [ + -4.387564659118652 + ], + "logprob_reward": -0.012564548291265965, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012564548291265965, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10114786773920059, + 0.28250885009765625, + 0.05031260848045349, + 0.03691667318344116, + 0.03482464700937271, + 0.06832228600978851, + 0.11773748695850372, + 0.2198745161294937, + 0.050364043563604355, + 0.03799106553196907 + ], + "credit_map": [ + 0.009465212002396584, + 0.013526765629649162, + 0.010046717710793018, + 0.01602979563176632, + 0.013700596056878567, + 0.009315293282270432, + 0.01399300992488861, + 0.015070473775267601, + 0.023023026064038277, + 0.042938049882650375, + 0.017884526401758194, + 0.04464986175298691, + 0.05695902556180954, + 0.03400034084916115, + 0.03320469334721565, + 0.029849326238036156, + 0.004829506389796734, + 0.006188326049596071, + 0.002774935681372881, + 0.0066274432465434074, + 0.007360030896961689, + 0.003932849038392305, + 0.008065108209848404, + 0.01053441222757101, + 0.006047612521797419, + 0.004942469298839569, + 0.004806064069271088, + 0.0037765023298561573, + 0.0052322340197861195, + 0.004801608622074127, + 0.004664233420044184, + 0.0026459491346031427, + 0.004295939113944769, + 0.0039003086276352406, + 0.0024860878475010395, + 0.004569424781948328, + 0.006095669232308865, + 0.004826634656637907, + 0.0033785756677389145, + 0.005272008944302797, + 0.007985669188201427, + 0.0070933080278337, + 0.011351191438734531, + 0.009213201701641083, + 0.011028299108147621, + 0.009351327084004879, + 0.007380667608231306, + 0.004918624646961689, + 0.013746451586484909, + 0.009234143421053886, + 0.014999107457697392, + 0.019210893660783768, + 0.014128059148788452, + 0.012437590397894382, + 0.018300766125321388, + 0.015680471435189247, + 0.0223382655531168, + 0.017898082733154297, + 0.013881869614124298, + 0.020121939480304718, + 0.06752340495586395, + 0.0373072549700737, + 0.030742885544896126, + 0.010060802102088928, + 0.004383440129458904, + 0.0037678165826946497, + 0.007859673351049423, + 0.005366531200706959, + 0.008001082576811314, + 0.008038387633860111, + 0.006442470010370016, + 0.006504642777144909, + 0.0036439006216824055, + 0.0038820134941488504, + 0.0035146495793014765, + 0.003774558659642935, + 0.009258170612156391, + 0.005854870658367872, + 0.0036650546826422215, + 0.004397848155349493 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 6 + ], + "selected_reward": -0.012564548291265965 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428583264350891 + ], + "opposite_logprob_tokens": [ + -2.0178582668304443 + ], + "logprob_reward": -0.1428583264350891, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428583264350891, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1774102747440338, + 0.08951269090175629, + 0.12477770447731018, + 0.06716297566890717, + 0.06473413854837418, + 0.08051329851150513, + 0.08240076899528503, + 0.20507113635540009, + 0.07663172483444214, + 0.03178524971008301 + ], + "credit_map": [ + 0.013548883609473705, + 0.014696452766656876, + 0.013282731175422668, + 0.030524248257279396, + 0.013988220132887363, + 0.022161493077874184, + 0.03265945985913277, + 0.036548785865306854, + 0.009775848127901554, + 0.012000879272818565, + 0.006307365372776985, + 0.016582127660512924, + 0.010863185860216618, + 0.008123601786792278, + 0.01666773110628128, + 0.009191958233714104, + 0.016565507277846336, + 0.01217446755617857, + 0.00694115087389946, + 0.019248507916927338, + 0.015922680497169495, + 0.017957087606191635, + 0.013095238246023655, + 0.02287306636571884, + 0.008742693811655045, + 0.006300706882029772, + 0.009555804543197155, + 0.007293857168406248, + 0.010902712121605873, + 0.009537787176668644, + 0.008719001896679401, + 0.006110418122261763, + 0.006910290569067001, + 0.00791085697710514, + 0.004282985348254442, + 0.007702856790274382, + 0.015116022899746895, + 0.008626335300505161, + 0.007153014652431011, + 0.007031779736280441, + 0.009207571856677532, + 0.006790356244891882, + 0.010175442323088646, + 0.01344191376119852, + 0.009056735783815384, + 0.01185267698019743, + 0.011765684932470322, + 0.00822292361408472, + 0.01159981545060873, + 0.007945538498461246, + 0.010007926262915134, + 0.012913274578750134, + 0.008879347704350948, + 0.009474718943238258, + 0.012153922580182552, + 0.009426232427358627, + 0.021030765026807785, + 0.01595878228545189, + 0.016734952107071877, + 0.01972113363444805, + 0.05787237361073494, + 0.03146255761384964, + 0.02707142010331154, + 0.015219169668853283, + 0.008439366705715656, + 0.009946275502443314, + 0.012738250195980072, + 0.005934466607868671, + 0.011505572125315666, + 0.008877771906554699, + 0.007877827621996403, + 0.011312203481793404, + 0.002833036705851555, + 0.003324309131130576, + 0.00591241056099534, + 0.0023506104480475187, + 0.004678858444094658, + 0.00476324325427413, + 0.004258747678250074, + 0.0036640381440520287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 2 + ], + "selected_reward": -0.1428583264350891 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11294171959161758 + ], + "opposite_logprob_tokens": [ + -2.2379417419433594 + ], + "logprob_reward": -0.11294171959161758, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11294171959161758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.031024277210235596, + 0.039791230112314224, + 0.0684695914387703, + 0.022762108594179153, + 0.027326403185725212, + 0.09026527404785156, + 0.04325072467327118, + 0.5111234188079834, + 0.12825356423854828, + 0.03773338347673416 + ], + "credit_map": [ + 0.0025311135686933994, + 0.0025338155683130026, + 0.0027869788464158773, + 0.004705457482486963, + 0.002514277584850788, + 0.0033749754074960947, + 0.00536281568929553, + 0.0072148446924984455, + 0.004367426503449678, + 0.006042194552719593, + 0.0025709534529596567, + 0.0063111851923167706, + 0.00781531073153019, + 0.004441710654646158, + 0.0038924594409763813, + 0.0043499949388206005, + 0.010186650790274143, + 0.006855162791907787, + 0.004639484453946352, + 0.009529013186693192, + 0.008330905809998512, + 0.012081606313586235, + 0.007097463123500347, + 0.00974931288510561, + 0.002891962183639407, + 0.0022267207968980074, + 0.0036769204307347536, + 0.002794709987938404, + 0.0025716417003422976, + 0.0037761188577860594, + 0.0028288178145885468, + 0.0019952189177274704, + 0.003033158602192998, + 0.0030750767327845097, + 0.0030784648843109608, + 0.0032247721683233976, + 0.004809064324945211, + 0.002871248172596097, + 0.002576455706730485, + 0.004658165853470564, + 0.010295598767697811, + 0.008175142109394073, + 0.018189070746302605, + 0.010872348211705685, + 0.014773602597415447, + 0.01142227090895176, + 0.010012807324528694, + 0.006524443626403809, + 0.00557348458096385, + 0.0035388933029025793, + 0.006266110111027956, + 0.0066771069541573524, + 0.006036004051566124, + 0.00442295940592885, + 0.006778690032660961, + 0.003957481123507023, + 0.05341534689068794, + 0.04001086577773094, + 0.03669929876923561, + 0.04876483231782913, + 0.1499665528535843, + 0.08391086012125015, + 0.07490291446447372, + 0.02345280535519123, + 0.012147440575063229, + 0.009351351298391819, + 0.014302979223430157, + 0.00993172824382782, + 0.022331014275550842, + 0.02436237968504429, + 0.015859749168157578, + 0.01996692456305027, + 0.0035243842285126448, + 0.0035161746200174093, + 0.003396053798496723, + 0.003891731146723032, + 0.009145128540694714, + 0.005907735787332058, + 0.0041064489632844925, + 0.004245728254318237 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 5 + ], + "selected_reward": -0.11294171959161758 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009778810665011406 + ], + "opposite_logprob_tokens": [ + -4.63477897644043 + ], + "logprob_reward": -0.009778810665011406, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009778810665011406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.043571095913648605, + 0.05826784297823906, + 0.05233106389641762, + 0.022703953087329865, + 0.044032156467437744, + 0.08004496246576309, + 0.03661303222179413, + 0.43738290667533875, + 0.09798247367143631, + 0.12707048654556274 + ], + "credit_map": [ + 0.004271665122359991, + 0.0051537128165364265, + 0.006708880886435509, + 0.007585558574646711, + 0.003665817203000188, + 0.004372762516140938, + 0.005750165320932865, + 0.0060625383630394936, + 0.004939376842230558, + 0.007824664004147053, + 0.002924117026850581, + 0.0065789236687123775, + 0.01111008320003748, + 0.008581978268921375, + 0.009253690019249916, + 0.007055013440549374, + 0.007346765138208866, + 0.005533210001885891, + 0.0034153261221945286, + 0.006973950657993555, + 0.006069514900445938, + 0.010118624195456505, + 0.005086774472147226, + 0.00778690492734313, + 0.0032035005278885365, + 0.002455502049997449, + 0.0030987493228167295, + 0.0017483759438619018, + 0.002786185359582305, + 0.0044577778317034245, + 0.002888144925236702, + 0.0020657202694565058, + 0.005383254960179329, + 0.005074767861515284, + 0.0033291198778897524, + 0.004200778901576996, + 0.008295672945678234, + 0.005929222796112299, + 0.00421948591247201, + 0.007599857170134783, + 0.012654504738748074, + 0.008648923598229885, + 0.015494804829359055, + 0.010307550430297852, + 0.006959203165024519, + 0.010218051262199879, + 0.008077586069703102, + 0.007684338837862015, + 0.004606665577739477, + 0.003581528551876545, + 0.0040086847729980946, + 0.004328158218413591, + 0.005324413999915123, + 0.004214100074023008, + 0.006898793391883373, + 0.00365069299004972, + 0.04478691518306732, + 0.03367336466908455, + 0.02593154087662697, + 0.04014185070991516, + 0.13171672821044922, + 0.07672557979822159, + 0.06439118832349777, + 0.020015740767121315, + 0.0076829297468066216, + 0.008687753230333328, + 0.011703054420650005, + 0.010569771751761436, + 0.016232773661613464, + 0.013773291371762753, + 0.013914454728364944, + 0.015418454073369503, + 0.012582698836922646, + 0.012856816872954369, + 0.01332010142505169, + 0.015917733311653137, + 0.022917527705430984, + 0.01774158515036106, + 0.015004904940724373, + 0.01672913134098053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.009778810665011406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dreamlike close-up, the camera performing a slow, continuous roll, causing the entire world to gently rotate. Within this swirling vision, a handmade rag doll lies quietly on a fluffy white pillow, bathed in the warm, rich golden afternoon sun. Its body is sewn from rough, beige linen, the sunlight highlighting the texture of the fabric. Its eyes are two dark buttons of varying sizes: one a simple four-hole plastic button, the other slightly larger and patinaed with wood. A simple smile, embroidered in red thread, is frozen on its face. As the camera rotates slowly, the rag doll seems to succumb to gravity in slow motion and begins a silent slide: its body slides down from the top of the pillow very slowly, the fabric rubbing slightly against the cotton surface of the pillowcase, and finally its soft body stuffed with cotton slides silently off the edge of the pillow. The whole process, amidst the swirling golden halo and flying dust, creates a nostalgic atmosphere that is both peaceful and slightly sad, like a fading childhood memory.", + "seed": 5835, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_2409_seed_5835_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.17775304559618235, + "vlm_logprob": -0.17775304559618235, + "vlm_yes_no_margin": 2.824999976158142, + "vlm_token_credit": -0.17775304559618235, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.6997384428977966 + ], + "opposite_logprob_tokens": [ + -0.6997384428977966 + ], + "logprob_reward": -0.6997384428977966, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.6997384428977966, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.08549446612596512, + 0.06059899181127548, + 0.05964933708310127, + 0.04766876995563507, + 0.07286572456359863, + 0.05336197093129158, + 0.08015426993370056, + 0.11319653689861298, + 0.31007468700408936, + 0.11693524569272995 + ], + "credit_map": [ + 0.01365925557911396, + 0.010794766247272491, + 0.009912878274917603, + 0.010247386060655117, + 0.012776218354701996, + 0.011315650306642056, + 0.007982529699802399, + 0.008805778808891773, + 0.016293978318572044, + 0.014328175224363804, + 0.004833698738366365, + 0.006069171242415905, + 0.00564174260944128, + 0.004712867084890604, + 0.004249492660164833, + 0.0044698636047542095, + 0.010223815217614174, + 0.00829445943236351, + 0.011267551220953465, + 0.005195113364607096, + 0.006056615151464939, + 0.005378855392336845, + 0.007349615450948477, + 0.0058833155781030655, + 0.008129646070301533, + 0.005560869351029396, + 0.012179384008049965, + 0.004909471608698368, + 0.004098453093320131, + 0.004954536445438862, + 0.004616474732756615, + 0.0032199339475482702, + 0.010157987475395203, + 0.010916076600551605, + 0.005374043248593807, + 0.009817076846957207, + 0.007438225671648979, + 0.008575042709708214, + 0.010636205784976482, + 0.00995106715708971, + 0.008159123361110687, + 0.00877829734236002, + 0.005507187452167273, + 0.007483169436454773, + 0.005994405597448349, + 0.004709888249635696, + 0.007685224059969187, + 0.005044671706855297, + 0.008953812532126904, + 0.009853080846369267, + 0.008318041451275349, + 0.015921492129564285, + 0.009110014885663986, + 0.006955455057322979, + 0.01236424408853054, + 0.008678126148879528, + 0.01941513642668724, + 0.014700161293148994, + 0.01467687077820301, + 0.011582763865590096, + 0.01768077164888382, + 0.016451271250844002, + 0.009384670294821262, + 0.00930488109588623, + 0.02647963911294937, + 0.04093563184142113, + 0.04289756342768669, + 0.05045470595359802, + 0.04019271954894066, + 0.027739737182855606, + 0.03949004039168358, + 0.0418846420943737, + 0.015089783817529678, + 0.016645697876811028, + 0.01386908907443285, + 0.015462605282664299, + 0.01544209010899067, + 0.015086484141647816, + 0.012130352668464184, + 0.013209152035415173 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.6997384428977966 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060624770820141 + ], + "opposite_logprob_tokens": [ + -3.895606279373169 + ], + "logprob_reward": -0.02060624770820141, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060624770820141, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18890129029750824, + 0.04382126405835152, + 0.037949416786432266, + 0.06919236481189728, + 0.04579775780439377, + 0.05192261189222336, + 0.1001565158367157, + 0.08731672912836075, + 0.23482514917850494, + 0.1401168555021286 + ], + "credit_map": [ + 0.03135232627391815, + 0.02638549543917179, + 0.02334067039191723, + 0.02646520547568798, + 0.029259968549013138, + 0.01876714639365673, + 0.02034926414489746, + 0.012981226667761803, + 0.007912038825452328, + 0.007292548660188913, + 0.004897665232419968, + 0.004906987305730581, + 0.004546527285128832, + 0.005068548023700714, + 0.004956327844411135, + 0.004240620415657759, + 0.006003550719469786, + 0.004404854495078325, + 0.0062216781079769135, + 0.004116535186767578, + 0.005054977722465992, + 0.0035786188673228025, + 0.005278876516968012, + 0.0032903256360441446, + 0.012127864174544811, + 0.010882418602705002, + 0.006862143520265818, + 0.005874216090887785, + 0.0070559680461883545, + 0.013104439713060856, + 0.006269897799938917, + 0.00701542105525732, + 0.006852972321212292, + 0.004770117811858654, + 0.00451795756816864, + 0.0039052804931998253, + 0.007391755934804678, + 0.007187583949416876, + 0.0068169548176229, + 0.004355135373771191, + 0.00755127239972353, + 0.004797236528247595, + 0.005414871498942375, + 0.007197758182883263, + 0.004532167688012123, + 0.004888136405497789, + 0.00862991064786911, + 0.008911261335015297, + 0.010596895590424538, + 0.009197232313454151, + 0.00945054180920124, + 0.016360772773623466, + 0.010092129930853844, + 0.008101197890937328, + 0.016316184774041176, + 0.020041564479470253, + 0.008786724880337715, + 0.00915156863629818, + 0.014082541689276695, + 0.011428444646298885, + 0.009468009695410728, + 0.012133831158280373, + 0.009907284751534462, + 0.012358329258859158, + 0.01898197829723358, + 0.02394251711666584, + 0.026882143691182137, + 0.04487159103155136, + 0.04132461175322533, + 0.018289683386683464, + 0.0287147406488657, + 0.031817901879549026, + 0.009125998243689537, + 0.007092393934726715, + 0.045517802238464355, + 0.013041670434176922, + 0.00782101135700941, + 0.03893710672855377, + 0.006453275680541992, + 0.012127608060836792 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02060624770820141 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12708909809589386 + ], + "opposite_logprob_tokens": [ + -2.127089023590088 + ], + "logprob_reward": -0.12708909809589386, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12708909809589386, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13049888610839844, + 0.1284381002187729, + 0.15062102675437927, + 0.08865758031606674, + 0.028268951922655106, + 0.08987384289503098, + 0.11251404881477356, + 0.04745912551879883, + 0.11672325432300568, + 0.10694516450166702 + ], + "credit_map": [ + 0.018701277673244476, + 0.01490798220038414, + 0.01954817771911621, + 0.0148007245734334, + 0.024662455543875694, + 0.011155538260936737, + 0.016172775998711586, + 0.010549965314567089, + 0.03683565557003021, + 0.03208480030298233, + 0.016019150614738464, + 0.009368459694087505, + 0.011362330056726933, + 0.008294133469462395, + 0.007344112731516361, + 0.007129461970180273, + 0.027671238407492638, + 0.024337653070688248, + 0.019243178889155388, + 0.012840795330703259, + 0.021036434918642044, + 0.016913458704948425, + 0.016629334539175034, + 0.011948940344154835, + 0.013529622927308083, + 0.009464188478887081, + 0.007119708228856325, + 0.00977634359151125, + 0.01155058853328228, + 0.017167218029499054, + 0.011863403022289276, + 0.008186506107449532, + 0.004542901646345854, + 0.0036787153221666813, + 0.005014135502278805, + 0.003303367644548416, + 0.0026946438010782003, + 0.0026975146029144526, + 0.0034169144928455353, + 0.0029207589104771614, + 0.014054293744266033, + 0.014523939229547977, + 0.006615510676056147, + 0.008147621527314186, + 0.010687628760933876, + 0.01052841916680336, + 0.016935687512159348, + 0.008380744606256485, + 0.019138630479574203, + 0.009622083976864815, + 0.015620511956512928, + 0.015110496431589127, + 0.008404831402003765, + 0.011308890767395496, + 0.01752936653792858, + 0.01577923260629177, + 0.006316204555332661, + 0.004496704787015915, + 0.0049087414517998695, + 0.0054589384235441685, + 0.0073577906005084515, + 0.00703325355425477, + 0.0043421415612101555, + 0.007545349188148975, + 0.01153453066945076, + 0.018794436007738113, + 0.014727921225130558, + 0.01980101875960827, + 0.017359692603349686, + 0.010486762039363384, + 0.011237043887376785, + 0.012781851924955845, + 0.008409565314650536, + 0.013727830722928047, + 0.015573518350720406, + 0.013344532810151577, + 0.013231668621301651, + 0.021553710103034973, + 0.01157552283257246, + 0.009528820402920246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.12708909809589386 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370651230216026 + ], + "opposite_logprob_tokens": [ + -3.4087064266204834 + ], + "logprob_reward": -0.03370651230216026, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370651230216026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07617125660181046, + 0.0688583254814148, + 0.04704756289720535, + 0.043605681508779526, + 0.03494584560394287, + 0.11166007071733475, + 0.16757042706012726, + 0.09199761599302292, + 0.24520327150821686, + 0.1129399761557579 + ], + "credit_map": [ + 0.013206126168370247, + 0.008897881023585796, + 0.006948454771190882, + 0.007678802125155926, + 0.014940210618078709, + 0.008672415278851986, + 0.008889380842447281, + 0.006937979254871607, + 0.022744692862033844, + 0.01530251931399107, + 0.006668176501989365, + 0.004275037907063961, + 0.0049422685988247395, + 0.005509380716830492, + 0.005424551200121641, + 0.0039916918613016605, + 0.008793083019554615, + 0.006669973023235798, + 0.006230585277080536, + 0.006585485301911831, + 0.005742336623370647, + 0.004033239092677832, + 0.004608927760273218, + 0.004383928142488003, + 0.007188905496150255, + 0.005590404849499464, + 0.006266718730330467, + 0.004531967453658581, + 0.004490924067795277, + 0.005431339610368013, + 0.006401117891073227, + 0.00370429758913815, + 0.004884316585958004, + 0.005717010702937841, + 0.003444431349635124, + 0.004285317845642567, + 0.0033329620491713285, + 0.004256078042089939, + 0.004113917239010334, + 0.004911808762699366, + 0.029194120317697525, + 0.033029910176992416, + 0.006434592884033918, + 0.007305867969989777, + 0.007658811751753092, + 0.009128080680966377, + 0.012566650286316872, + 0.006342033855617046, + 0.017323443666100502, + 0.016532590612769127, + 0.017731189727783203, + 0.03066481277346611, + 0.018397031351923943, + 0.012495417147874832, + 0.030283577740192413, + 0.024142349138855934, + 0.013026714324951172, + 0.013654771260917187, + 0.015488523058593273, + 0.011693703010678291, + 0.010599445551633835, + 0.010510854423046112, + 0.010118232108652592, + 0.00690536480396986, + 0.024018101394176483, + 0.04065766930580139, + 0.03311387822031975, + 0.029977386817336082, + 0.03866005688905716, + 0.02377147786319256, + 0.024423029273748398, + 0.030581651255488396, + 0.009855911135673523, + 0.01885845884680748, + 0.016216807067394257, + 0.013165222480893135, + 0.01277709286659956, + 0.01358079630881548, + 0.010687876492738724, + 0.017797809094190598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 9 + ], + "selected_reward": -0.03370651230216026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076249269768595695 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.0076249269768595695, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076249269768595695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1027321070432663, + 0.035379353910684586, + 0.016846803948283195, + 0.03056548349559307, + 0.041547250002622604, + 0.045018043369054794, + 0.06298511475324631, + 0.27008509635925293, + 0.11827286332845688, + 0.27656787633895874 + ], + "credit_map": [ + 0.022027045488357544, + 0.013601039536297321, + 0.009607743471860886, + 0.012592993676662445, + 0.01632375828921795, + 0.012368546798825264, + 0.008951753377914429, + 0.007259230595082045, + 0.006998032331466675, + 0.007727163378149271, + 0.0037874977570027113, + 0.0027812246698886156, + 0.0039651826955378056, + 0.00336181977763772, + 0.0034708306193351746, + 0.003287601750344038, + 0.0034111165441572666, + 0.0031459257006645203, + 0.0016969672869890928, + 0.0017370916903018951, + 0.0018840152770280838, + 0.0018871190259233117, + 0.0016861517215147614, + 0.0013984176330268383, + 0.004832710139453411, + 0.0041808453388512135, + 0.0037448385264724493, + 0.0032390744891017675, + 0.0035583830904215574, + 0.0034361251164227724, + 0.005313629750162363, + 0.0022598758805543184, + 0.006406672764569521, + 0.005605102516710758, + 0.0031561560463160276, + 0.004391447175294161, + 0.004014891106635332, + 0.004582852125167847, + 0.00687013752758503, + 0.006519991438835859, + 0.0061578224413096905, + 0.004770929459482431, + 0.00411670608446002, + 0.005868825130164623, + 0.005397470202296972, + 0.0045453584752976894, + 0.008641956374049187, + 0.005518975201994181, + 0.009426066651940346, + 0.006207613740116358, + 0.0062384940683841705, + 0.0123677346855402, + 0.005444536451250315, + 0.008055581711232662, + 0.01038454007357359, + 0.004860545974224806, + 0.038216520100831985, + 0.03847408667206764, + 0.046102993190288544, + 0.02627747319638729, + 0.028180014342069626, + 0.03421563282608986, + 0.0269999448210001, + 0.031618405133485794, + 0.010597769170999527, + 0.02225547842681408, + 0.012569903396070004, + 0.01775183714926243, + 0.01552650611847639, + 0.012607681564986706, + 0.011392447166144848, + 0.015571244060993195, + 0.021249879151582718, + 0.026700245216488838, + 0.0716024711728096, + 0.03251359239220619, + 0.025518417358398438, + 0.05169107764959335, + 0.017552757635712624, + 0.029739420861005783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0076249269768595695 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a dreamlike close-up, the camera performing a slow, continuous roll, causing the entire world to gently rotate. Within this swirling vision, a handmade rag doll lies quietly on a fluffy white pillow, bathed in the warm, rich golden afternoon sun. Its body is sewn from rough, beige linen, the sunlight highlighting the texture of the fabric. Its eyes are two dark buttons of varying sizes: one a simple four-hole plastic button, the other slightly larger and patinaed with wood. A simple smile, embroidered in red thread, is frozen on its face. As the camera rotates slowly, the rag doll seems to succumb to gravity in slow motion and begins a silent slide: its body slides down from the top of the pillow very slowly, the fabric rubbing slightly against the cotton surface of the pillowcase, and finally its soft body stuffed with cotton slides silently off the edge of the pillow. The whole process, amidst the swirling golden halo and flying dust, creates a nostalgic atmosphere that is both peaceful and slightly sad, like a fading childhood memory.", + "seed": 5836, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_14/index_2409_seed_5836_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.19147545425221324, + "vlm_logprob": -0.19147545425221324, + "vlm_yes_no_margin": 2.699999976158142, + "vlm_token_credit": -0.19147545425221324, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7052862048149109 + ], + "opposite_logprob_tokens": [ + -0.7052862048149109 + ], + "logprob_reward": -0.7052862048149109, + "yes_no_margin_reward": 0.0, + "token_credit_reward": -0.7052862048149109, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.07393182814121246, + 0.10013105720281601, + 0.03719010204076767, + 0.1238442212343216, + 0.0412101075053215, + 0.050770051777362823, + 0.03672592341899872, + 0.11064478009939194, + 0.1279478371143341, + 0.2976040542125702 + ], + "credit_map": [ + 0.013112698681652546, + 0.006860516034066677, + 0.011471616104245186, + 0.009350254200398922, + 0.005424808245152235, + 0.011829407885670662, + 0.008141495287418365, + 0.007741031236946583, + 0.010242683812975883, + 0.012274380773305893, + 0.014534812420606613, + 0.017601316794753075, + 0.009828455746173859, + 0.012052114121615887, + 0.012102705426514149, + 0.011494586244225502, + 0.0032404186204075813, + 0.005722605623304844, + 0.002963491016998887, + 0.007113560102880001, + 0.0035196857061237097, + 0.0049919746816158295, + 0.004623362794518471, + 0.005015004891902208, + 0.021702000871300697, + 0.01014441903680563, + 0.008072951808571815, + 0.014278597198426723, + 0.011993505991995335, + 0.038462091237306595, + 0.009725401178002357, + 0.00946525577455759, + 0.0065641277469694614, + 0.0050912234000861645, + 0.0021178785245865583, + 0.008243965916335583, + 0.0031595700420439243, + 0.007887826301157475, + 0.0039061252027750015, + 0.004239391069859266, + 0.008171802386641502, + 0.004563450813293457, + 0.004156370181590319, + 0.006405266467481852, + 0.004698761738836765, + 0.008319022133946419, + 0.006988754030317068, + 0.007466623559594154, + 0.003954252228140831, + 0.0023145750164985657, + 0.001505408901721239, + 0.0030710475984960794, + 0.0027918231207877398, + 0.006067792419344187, + 0.006191544234752655, + 0.010829479433596134, + 0.008570176549255848, + 0.014084357768297195, + 0.007242176216095686, + 0.01660306751728058, + 0.010197279043495655, + 0.017504755407571793, + 0.021164605394005775, + 0.015278367325663567, + 0.011019869707524776, + 0.011196575127542019, + 0.010925721377134323, + 0.021020617336034775, + 0.01729513332247734, + 0.013681676238775253, + 0.022122971713542938, + 0.02068527601659298, + 0.01728709414601326, + 0.10843678563833237, + 0.006989053916186094, + 0.008702187798917294, + 0.1132068783044815, + 0.024162646383047104, + 0.007476358208805323, + 0.011343044228851795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 3 + ], + "selected_reward": -0.7052862048149109 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012565255165100098 + ], + "opposite_logprob_tokens": [ + -4.3875651359558105 + ], + "logprob_reward": -0.012565255165100098, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012565255165100098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31978532671928406, + 0.055532120168209076, + 0.04920676350593567, + 0.04053616523742676, + 0.036154165863990784, + 0.03344417363405228, + 0.068100206553936, + 0.07897354662418365, + 0.036895737051963806, + 0.28137174248695374 + ], + "credit_map": [ + 0.07079572230577469, + 0.025654396042227745, + 0.026273362338542938, + 0.028423696756362915, + 0.024963194504380226, + 0.07349362969398499, + 0.030603395774960518, + 0.039577923715114594, + 0.00521107716485858, + 0.008788548409938812, + 0.00793148297816515, + 0.009393790736794472, + 0.006159642711281776, + 0.005845340434461832, + 0.006613902747631073, + 0.0055883340537548065, + 0.00465416070073843, + 0.007860648445785046, + 0.005109633319079876, + 0.007447417825460434, + 0.004028825089335442, + 0.006568416021764278, + 0.0065060569904744625, + 0.007031605578958988, + 0.003866341896355152, + 0.0038364860229194164, + 0.003764079650864005, + 0.007702216040343046, + 0.0036901417188346386, + 0.006408126093447208, + 0.005422079935669899, + 0.005846692714840174, + 0.007623988203704357, + 0.003427644958719611, + 0.0018933097599074244, + 0.003630751743912697, + 0.0023368047550320625, + 0.010901868343353271, + 0.0027756213676184416, + 0.0035641766153275967, + 0.006555468775331974, + 0.0024215823505073786, + 0.0018303408287465572, + 0.0035072893369942904, + 0.0024541786406189203, + 0.010623181238770485, + 0.0023577988613396883, + 0.003694333368912339, + 0.009096425957977772, + 0.0053043058142066, + 0.005850383546203375, + 0.008912241086363792, + 0.006064539775252342, + 0.00749131478369236, + 0.012173544615507126, + 0.013207451440393925, + 0.0066592092625796795, + 0.010194743983447552, + 0.005830573383718729, + 0.009607733227312565, + 0.012935408391058445, + 0.013098427094519138, + 0.010488847270607948, + 0.010158603079617023, + 0.0034633998293429613, + 0.005312243476510048, + 0.002786163939163089, + 0.004247134085744619, + 0.004085295833647251, + 0.004057017620652914, + 0.00801845546811819, + 0.00492602726444602, + 0.02037864923477173, + 0.0950280949473381, + 0.01174736674875021, + 0.01484330091625452, + 0.09104269742965698, + 0.019711237400770187, + 0.01399078220129013, + 0.014629621990025043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.012565255165100098 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284881949424744 + ], + "opposite_logprob_tokens": [ + -2.0178487300872803 + ], + "logprob_reward": -0.14284881949424744, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14284881949424744, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21005761623382568, + 0.16125503182411194, + 0.058950573205947876, + 0.13970249891281128, + 0.0960678830742836, + 0.050700701773166656, + 0.05439995229244232, + 0.08382225036621094, + 0.0744527280330658, + 0.07059070467948914 + ], + "credit_map": [ + 0.048756469041109085, + 0.0200473815202713, + 0.01717943139374256, + 0.016014302149415016, + 0.015106038190424442, + 0.051783036440610886, + 0.01955728605389595, + 0.021613676100969315, + 0.015129083767533302, + 0.020235184580087662, + 0.022826161235570908, + 0.025034235790371895, + 0.011493023484945297, + 0.01952645182609558, + 0.022010600194334984, + 0.025000305846333504, + 0.005130565259605646, + 0.011302987113595009, + 0.005400736816227436, + 0.009765413589775562, + 0.0069854483008384705, + 0.007268447428941727, + 0.005958213936537504, + 0.007138766348361969, + 0.029788078740239143, + 0.005149379838258028, + 0.007082166150212288, + 0.011913694441318512, + 0.016016272827982903, + 0.05373343452811241, + 0.007391209248453379, + 0.008628271520137787, + 0.022327523678541183, + 0.00593235157430172, + 0.005202639847993851, + 0.009908908978104591, + 0.0061570098623633385, + 0.03188810124993324, + 0.006127682980149984, + 0.008523677475750446, + 0.006244384218007326, + 0.007029305212199688, + 0.0030641613993793726, + 0.00970561895519495, + 0.0038654447998851538, + 0.00452880235388875, + 0.006673403084278107, + 0.009589588269591331, + 0.005513061303645372, + 0.003998222760856152, + 0.0026105984579771757, + 0.005917516071349382, + 0.004078300669789314, + 0.008247815072536469, + 0.008055518381297588, + 0.015978926792740822, + 0.0065678260289132595, + 0.009280904196202755, + 0.006721579935401678, + 0.016356762498617172, + 0.0077852471731603146, + 0.008012097328901291, + 0.018494965508580208, + 0.010602869093418121, + 0.00683517474681139, + 0.007332397624850273, + 0.006620064377784729, + 0.009926602244377136, + 0.007232409901916981, + 0.008119010366499424, + 0.015520147047936916, + 0.012866928242146969, + 0.009998584166169167, + 0.013307394459843636, + 0.005442652385681868, + 0.00738231698051095, + 0.010246786288917065, + 0.008756048046052456, + 0.007372304797172546, + 0.008084617555141449 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.14284881949424744 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904993534088135 + ], + "opposite_logprob_tokens": [ + -2.464049816131592 + ], + "logprob_reward": -0.08904993534088135, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904993534088135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10776541382074356, + 0.0587431974709034, + 0.03619759529829025, + 0.09692797809839249, + 0.04569089785218239, + 0.031301744282245636, + 0.05650866776704788, + 0.3206656277179718, + 0.051343463361263275, + 0.19485540688037872 + ], + "credit_map": [ + 0.02289547771215439, + 0.008004125207662582, + 0.010062959976494312, + 0.007318384945392609, + 0.0076390597969293594, + 0.026206225156784058, + 0.0103618074208498, + 0.015277357771992683, + 0.006299818400293589, + 0.007715573068708181, + 0.0071637993678450584, + 0.009664144366979599, + 0.005349932238459587, + 0.007014692761003971, + 0.008035492151975632, + 0.007499741390347481, + 0.002662506652995944, + 0.003883116412907839, + 0.0032605843152850866, + 0.0042915805242955685, + 0.00404676329344511, + 0.00443927850574255, + 0.006477878894656897, + 0.007135882508009672, + 0.016183920204639435, + 0.007473415695130825, + 0.005672842729836702, + 0.01063708309084177, + 0.01005837693810463, + 0.029745519161224365, + 0.008463151752948761, + 0.00869365967810154, + 0.008883066475391388, + 0.004234781488776207, + 0.002022692235186696, + 0.00426216796040535, + 0.00341390585526824, + 0.012651574797928333, + 0.004462942481040955, + 0.005759758874773979, + 0.005204922053962946, + 0.0025014104321599007, + 0.0018540583550930023, + 0.003973405808210373, + 0.002463918412104249, + 0.007191814482212067, + 0.0037584155797958374, + 0.0043537975288927555, + 0.005595936439931393, + 0.0071091437712311745, + 0.0036830950994044542, + 0.011013875715434551, + 0.005403673741966486, + 0.00535374041646719, + 0.011501996777951717, + 0.0068472023122012615, + 0.021782008931040764, + 0.035653844475746155, + 0.021149683743715286, + 0.04864000901579857, + 0.039240218698978424, + 0.04945322498679161, + 0.06157386302947998, + 0.04317273944616318, + 0.004556769039481878, + 0.004908564034849405, + 0.0046289050951600075, + 0.009833070449531078, + 0.005513640586286783, + 0.005492432974278927, + 0.011117984540760517, + 0.005292088259011507, + 0.013304444961249828, + 0.06024543195962906, + 0.009767657145857811, + 0.012399670667946339, + 0.05906428024172783, + 0.013829552568495274, + 0.012925568968057632, + 0.01331877801567316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.08904993534088135 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera performing a slow, continuous roll causing the scene to rotate?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a handmade rag doll made of rough beige linen visible on a fluffy white pillow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the doll's eyes consist of two dark buttons of varying sizes, one plastic and one wood-patinaed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the doll slide slowly off the edge of the pillow in slow motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by warm golden afternoon sun with visible flying dust?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007627056445926428 + ], + "opposite_logprob_tokens": [ + -4.882627010345459 + ], + "logprob_reward": -0.007627056445926428, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007627056445926428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13152943551540375, + 0.04292699694633484, + 0.026956165209412575, + 0.16889642179012299, + 0.01646302454173565, + 0.05138328671455383, + 0.11545305699110031, + 0.15097099542617798, + 0.060384850949048996, + 0.23503577709197998 + ], + "credit_map": [ + 0.02789325825870037, + 0.01093731727451086, + 0.011211314238607883, + 0.012173699215054512, + 0.01015402376651764, + 0.028387514874339104, + 0.013286968693137169, + 0.017485344782471657, + 0.004844839219003916, + 0.008247054181993008, + 0.0048304456286132336, + 0.005603988654911518, + 0.00341181643307209, + 0.00422136764973402, + 0.006192108616232872, + 0.005575378425419331, + 0.002107840497046709, + 0.004740649834275246, + 0.002929739421233535, + 0.00432039238512516, + 0.003442268120124936, + 0.003439859254285693, + 0.0026669593062251806, + 0.003308459185063839, + 0.03399018943309784, + 0.007126135751605034, + 0.0087149478495121, + 0.01508361753076315, + 0.0192574355751276, + 0.06340330094099045, + 0.010964157991111279, + 0.010356632992625237, + 0.002481263829395175, + 0.0018748757429420948, + 0.0013595068594440818, + 0.002474480541422963, + 0.001444850000552833, + 0.0019727193284779787, + 0.0024047340266406536, + 0.0024505951441824436, + 0.007402104325592518, + 0.007113713771104813, + 0.002292649121955037, + 0.0071788690984249115, + 0.003370828926563263, + 0.010641700588166714, + 0.0069389911368489265, + 0.006444435101002455, + 0.010434985160827637, + 0.011623041704297066, + 0.0060563599690794945, + 0.017014574259519577, + 0.011198123916983604, + 0.014064988121390343, + 0.01851826347410679, + 0.026542730629444122, + 0.010947192087769508, + 0.016866391524672508, + 0.014898479916155338, + 0.029241137206554413, + 0.012621612288057804, + 0.014453054405748844, + 0.030732033774256706, + 0.02121109887957573, + 0.005847888998687267, + 0.007896331138908863, + 0.004786068107932806, + 0.008224236778914928, + 0.007307812105864286, + 0.007131796795874834, + 0.010781110264360905, + 0.008409609086811543, + 0.01214380469173193, + 0.0880863219499588, + 0.0036467397585511208, + 0.005421228241175413, + 0.09141679853200912, + 0.019719894975423813, + 0.009188530035316944, + 0.005412461701780558 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 7 + ], + "selected_reward": -0.007627056445926428 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_tensor([8006, 8006])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_tensor([8006, 8006])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f1357c8948db0fd289ba8aec63de2f5df1858722 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_tensor([8006, 8006])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, slightly low-angled, medium shot captures a figure seated in perfect lotus pose on a simple, dark cushion within a tranquil meditation chamber. They are draped in a magnificent, deep indigo silk robe, its heavy fabric falling in elegant folds. The setting consists of dark, polished wooden floors, and against the muted backdrop, thin incense smoke curls upward from a small altar. A shaft of dusty afternoon light filters through the tall, latticed windows, illuminating the figure and creating a dramatic play of light and shadow. The figure's chest rises and falls in a slow, low, rhythmic rhythm, the heavy silk of the robe shifting subtly with each steady breath. As this tranquil movement continues, the camera begins a slow, deliberate zoom, sliding forward to focus intimately on the robe's chest area. The lens draws the viewer's gaze from the overall form to the breathtaking microcosm of the intricate embroidery, revealing the texture of each gold and silver metallic stitch, forming swirling celestial dragons and dark clouds, creating a profound atmosphere of deep concentration and spiritual tranquility.", + "seed": 4925, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_8006_seed_4925_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.017930667102336883, + "vlm_logprob": -0.017930667102336883, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.017930667102336883, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03099866583943367 + ], + "opposite_logprob_tokens": [ + -3.530998706817627 + ], + "logprob_reward": -0.03099866583943367, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.03099866583943367, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27060630917549133, + 0.19420179724693298, + 0.15600134432315826, + 0.09001221507787704, + 0.023978566750884056, + 0.045500729233026505, + 0.048806071281433105, + 0.08762533217668533, + 0.0367487296462059, + 0.046518851071596146 + ], + "credit_map": [ + 0.03250860795378685, + 0.04066099971532822, + 0.03288601338863373, + 0.05493004992604256, + 0.01851685903966427, + 0.025589274242520332, + 0.029825422912836075, + 0.035689111799001694, + 0.011551697738468647, + 0.0245283842086792, + 0.012749237939715385, + 0.01622036285698414, + 0.00929348822683096, + 0.06731726974248886, + 0.015464934520423412, + 0.0370764285326004, + 0.009690056554973125, + 0.016034647822380066, + 0.010869472287595272, + 0.008488476276397705, + 0.010308829136192799, + 0.017001019790768623, + 0.04273177683353424, + 0.040877074003219604, + 0.015819882974028587, + 0.014342600479722023, + 0.008561128750443459, + 0.011090505868196487, + 0.0068741156719625, + 0.007673133630305529, + 0.00621328130364418, + 0.019437581300735474, + 0.004275357350707054, + 0.0033243088982999325, + 0.0032625312451273203, + 0.0029943790286779404, + 0.0016521825455129147, + 0.0036665014922618866, + 0.0028434821870177984, + 0.0019598251674324274, + 0.003755734534934163, + 0.0061248112469911575, + 0.011552338488399982, + 0.0032896266784518957, + 0.003716502571478486, + 0.006304086651653051, + 0.005876646377146244, + 0.004880983382463455, + 0.005113840568810701, + 0.006310734432190657, + 0.006116229109466076, + 0.0034439843147993088, + 0.006923914887011051, + 0.009184681810438633, + 0.007189879193902016, + 0.0045228078961372375, + 0.008555347099900246, + 0.016928408294916153, + 0.014506883919239044, + 0.008527431637048721, + 0.010157138109207153, + 0.009664350189268589, + 0.011519416235387325, + 0.007766354829072952, + 0.0037349334452301264, + 0.005186771042644978, + 0.005369191057980061, + 0.004239861853420734, + 0.004935069475322962, + 0.006156470626592636, + 0.0027835567016154528, + 0.0043428754433989525, + 0.004975864198058844, + 0.00690614664927125, + 0.00646234443411231, + 0.0069588604383170605, + 0.006405724212527275, + 0.005272107198834419, + 0.005306896287947893, + 0.004230913240462542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.03099866583943367 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005968722980469465 + ], + "opposite_logprob_tokens": [ + -5.1309685707092285 + ], + "logprob_reward": -0.005968722980469465, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005968722980469465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0842348039150238, + 0.1364295482635498, + 0.23655322194099426, + 0.061021994799375534, + 0.03024464100599289, + 0.048057105392217636, + 0.13720442354679108, + 0.1326410174369812, + 0.09745246917009354, + 0.03616075962781906 + ], + "credit_map": [ + 0.005508732050657272, + 0.011961019597947598, + 0.015469594858586788, + 0.012344264425337315, + 0.00886539090424776, + 0.006741147954016924, + 0.010139351710677147, + 0.013205303810536861, + 0.008987885899841785, + 0.016427479684352875, + 0.020210815593600273, + 0.017948662862181664, + 0.008327958174049854, + 0.02010393515229225, + 0.012811440043151379, + 0.031611368060112, + 0.011399374343454838, + 0.022710228338837624, + 0.02464590221643448, + 0.018021319061517715, + 0.027089105919003487, + 0.016300318762660027, + 0.05986277014017105, + 0.05652419105172157, + 0.005781095940619707, + 0.006289156619459391, + 0.0050301929004490376, + 0.0033305957913398743, + 0.01275234017521143, + 0.00816411804407835, + 0.00978702213615179, + 0.009887473657727242, + 0.0044313096441328526, + 0.0045157792046666145, + 0.003805005457252264, + 0.004431859590113163, + 0.003588438732549548, + 0.0029799030162394047, + 0.0036870951298624277, + 0.002805253490805626, + 0.004204330965876579, + 0.005349150858819485, + 0.01476900652050972, + 0.004437387455254793, + 0.0028457038570195436, + 0.004599384497851133, + 0.006836536806076765, + 0.0050156074576079845, + 0.009561488404870033, + 0.021416662260890007, + 0.014078649692237377, + 0.011628195643424988, + 0.016246184706687927, + 0.03762666508555412, + 0.016365963965654373, + 0.01028062030673027, + 0.012007378041744232, + 0.019727736711502075, + 0.019203154370188713, + 0.014874051325023174, + 0.022224457934498787, + 0.016099555417895317, + 0.015032793395221233, + 0.013471892103552818, + 0.017889847978949547, + 0.013798484578728676, + 0.013530116528272629, + 0.010172096081078053, + 0.010389095172286034, + 0.010781632736325264, + 0.011528674513101578, + 0.009362515062093735, + 0.0035239725839346647, + 0.005429420154541731, + 0.00456571951508522, + 0.005062838550657034, + 0.004733797628432512, + 0.003743206150829792, + 0.004721793346107006, + 0.004380012396723032 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 1 + ], + "selected_reward": -0.005968722980469465 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008636740036308765 + ], + "opposite_logprob_tokens": [ + -4.758636951446533 + ], + "logprob_reward": -0.008636740036308765, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008636740036308765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21514183282852173, + 0.07776113599538803, + 0.23077432811260223, + 0.0346667617559433, + 0.07627472281455994, + 0.09344850480556488, + 0.04890532046556473, + 0.0742371454834938, + 0.0917075052857399, + 0.05708274617791176 + ], + "credit_map": [ + 0.02140183374285698, + 0.023235158994793892, + 0.039950963109731674, + 0.020102614536881447, + 0.015342212282121181, + 0.02989419735968113, + 0.020492171868681908, + 0.04472270980477333, + 0.006734350696206093, + 0.012889223173260689, + 0.01149899885058403, + 0.014434938319027424, + 0.0056818448938429356, + 0.01161870639771223, + 0.008362982422113419, + 0.006540094502270222, + 0.02897150069475174, + 0.026051588356494904, + 0.032666001468896866, + 0.02014945074915886, + 0.029802031815052032, + 0.032059118151664734, + 0.03821870684623718, + 0.022855939343571663, + 0.004489650018513203, + 0.004974646028131247, + 0.006375707220286131, + 0.003808331210166216, + 0.0041861021891236305, + 0.003901985241100192, + 0.004539811983704567, + 0.002390531124547124, + 0.009709101170301437, + 0.014520864002406597, + 0.007896855473518372, + 0.00856711808592081, + 0.006625210400670767, + 0.011849397793412209, + 0.007540638092905283, + 0.009565538726747036, + 0.008117725141346455, + 0.009350335225462914, + 0.030181115493178368, + 0.004942282568663359, + 0.006821276620030403, + 0.006356863770633936, + 0.02290467917919159, + 0.004774236585944891, + 0.004226076882332563, + 0.007575210649520159, + 0.005869687534868717, + 0.004231128375977278, + 0.004919788334518671, + 0.01170846726745367, + 0.00573037238791585, + 0.004644594620913267, + 0.007588742766529322, + 0.011361708864569664, + 0.009691473096609116, + 0.010973452590405941, + 0.011360454373061657, + 0.007636798545718193, + 0.008618528954684734, + 0.0070059942081570625, + 0.011266285553574562, + 0.014137404039502144, + 0.011691111139953136, + 0.009989911690354347, + 0.012744995765388012, + 0.013561137020587921, + 0.010185524821281433, + 0.008131138049066067, + 0.008730805478990078, + 0.008608012460172176, + 0.0065916008315980434, + 0.008534172549843788, + 0.007438660599291325, + 0.005343747790902853, + 0.007816916331648827, + 0.004018832929432392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.008636740036308765 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038104183971881866 + ], + "opposite_logprob_tokens": [ + -3.288104295730591 + ], + "logprob_reward": -0.038104183971881866, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038104183971881866, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10360091179609299, + 0.06187193840742111, + 0.46970123052597046, + 0.024017268791794777, + 0.05749247968196869, + 0.0514955148100853, + 0.0827176421880722, + 0.04185296222567558, + 0.06360600143671036, + 0.04364413022994995 + ], + "credit_map": [ + 0.01418763492256403, + 0.019578032195568085, + 0.009970292448997498, + 0.013656463474035263, + 0.011052162386476994, + 0.011605237610638142, + 0.013008621521294117, + 0.010542460717260838, + 0.004745140206068754, + 0.00571388378739357, + 0.0115168746560812, + 0.007168011274188757, + 0.004941469524055719, + 0.016100263223052025, + 0.006962839979678392, + 0.004723448306322098, + 0.0579356923699379, + 0.05310685187578201, + 0.06315617263317108, + 0.03951529785990715, + 0.06008639931678772, + 0.07380934804677963, + 0.0786239504814148, + 0.04346749186515808, + 0.0017405168619006872, + 0.0036963806487619877, + 0.002534343861043453, + 0.0035673819947987795, + 0.003732529701665044, + 0.002847501076757908, + 0.004241548012942076, + 0.0016570657026022673, + 0.005014418158680201, + 0.00987653061747551, + 0.009910017251968384, + 0.007008923217654228, + 0.005575773771852255, + 0.007555610500276089, + 0.007581479847431183, + 0.004969724919646978, + 0.006715202238410711, + 0.0070462278090417385, + 0.014362204819917679, + 0.00384182739071548, + 0.0033781863749027252, + 0.0032560580875724554, + 0.009063417091965675, + 0.0038323907647281885, + 0.011350004933774471, + 0.013684522360563278, + 0.011984148062765598, + 0.012022402137517929, + 0.007220772560685873, + 0.01050824485719204, + 0.009625622071325779, + 0.006321916822344065, + 0.003996789455413818, + 0.006749078165739775, + 0.006244643125683069, + 0.0060006240382790565, + 0.005034944042563438, + 0.004698509816080332, + 0.005579106509685516, + 0.0035492656752467155, + 0.009290100075304508, + 0.010850233025848866, + 0.007746024522930384, + 0.005912098102271557, + 0.00908129196614027, + 0.008575337938964367, + 0.006060696206986904, + 0.006090215872973204, + 0.004928808659315109, + 0.006548772566020489, + 0.00486981263384223, + 0.008208850398659706, + 0.005557062104344368, + 0.003908467944711447, + 0.005728712305426598, + 0.003893639426678419 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.038104183971881866 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.059987250715494156, + 0.14953270554542542, + 0.07012087106704712, + 0.031388964504003525, + 0.04045618698000908, + 0.12050926685333252, + 0.20027422904968262, + 0.07394088804721832, + 0.12009330093860626, + 0.13369637727737427 + ], + "credit_map": [ + 0.007198613602668047, + 0.006944982334971428, + 0.007314222399145365, + 0.009115190245211124, + 0.004119016695767641, + 0.009918849915266037, + 0.004870779346674681, + 0.01050559338182211, + 0.007603948935866356, + 0.01679915189743042, + 0.01259618904441595, + 0.01890115812420845, + 0.013430005870759487, + 0.0389464795589447, + 0.031285155564546585, + 0.00997060164809227, + 0.005577473435550928, + 0.006998494733124971, + 0.006807966623455286, + 0.005102161783725023, + 0.007223017048090696, + 0.005350424908101559, + 0.018921850249171257, + 0.014139479026198387, + 0.002835145452991128, + 0.005189761519432068, + 0.00342230754904449, + 0.005839752033352852, + 0.004568044096231461, + 0.004278387408703566, + 0.003403079928830266, + 0.0018524841871112585, + 0.004058845806866884, + 0.005610652267932892, + 0.005238172598183155, + 0.004621845670044422, + 0.003919357433915138, + 0.007223061285912991, + 0.005369748920202255, + 0.004414496012032032, + 0.014207095839083195, + 0.013381654396653175, + 0.026247862726449966, + 0.010926441289484501, + 0.009014085866510868, + 0.014832333661615849, + 0.020787885412573814, + 0.011111891828477383, + 0.02644979953765869, + 0.03488047048449516, + 0.02482215315103531, + 0.0251674372702837, + 0.01775399036705494, + 0.039546091109514236, + 0.017856713384389877, + 0.013797548599541187, + 0.006438903976231813, + 0.011006913147866726, + 0.010614216327667236, + 0.009959202259778976, + 0.011868243105709553, + 0.0059652868658304214, + 0.01091143861413002, + 0.007176676299422979, + 0.011904769577085972, + 0.020625565201044083, + 0.016893355175852776, + 0.012037845328450203, + 0.018060650676488876, + 0.015179356560111046, + 0.011257450096309185, + 0.014134305529296398, + 0.016423117369413376, + 0.020846620202064514, + 0.018290452659130096, + 0.018541863188147545, + 0.017341209575533867, + 0.014101136475801468, + 0.013308921828866005, + 0.014843044802546501 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 9 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, medium shot captures a figure seated in perfect lotus pose on a simple, dark cushion within a tranquil meditation chamber. They are draped in a magnificent, deep indigo silk robe, its heavy fabric falling in elegant folds. The setting consists of dark, polished wooden floors, and against the muted backdrop, thin incense smoke curls upward from a small altar. A shaft of dusty afternoon light filters through the tall, latticed windows, illuminating the figure and creating a dramatic play of light and shadow. The figure's chest rises and falls in a slow, low, rhythmic rhythm, the heavy silk of the robe shifting subtly with each steady breath. As this tranquil movement continues, the camera begins a slow, deliberate zoom, sliding forward to focus intimately on the robe's chest area. The lens draws the viewer's gaze from the overall form to the breathtaking microcosm of the intricate embroidery, revealing the texture of each gold and silver metallic stitch, forming swirling celestial dragons and dark clouds, creating a profound atmosphere of deep concentration and spiritual tranquility.", + "seed": 4926, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_8006_seed_4926_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017965007200837137, + "vlm_logprob": -0.017965007200837137, + "vlm_yes_no_margin": 4.6, + "vlm_token_credit": -0.017965007200837137, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008087978698313236 + ], + "opposite_logprob_tokens": [ + -5.008088111877441 + ], + "logprob_reward": -0.008087978698313236, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.008087978698313236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21109548211097717, + 0.07478539645671844, + 0.05536230653524399, + 0.04936414584517479, + 0.16206374764442444, + 0.05912510305643082, + 0.17896708846092224, + 0.02698872610926628, + 0.051690734922885895, + 0.13055726885795593 + ], + "credit_map": [ + 0.031668029725551605, + 0.026177437976002693, + 0.018094085156917572, + 0.033980656415224075, + 0.03575450927019119, + 0.015735715627670288, + 0.03080558590590954, + 0.018879452720284462, + 0.01864471845328808, + 0.004879738669842482, + 0.0032462403178215027, + 0.019272776320576668, + 0.012768065556883812, + 0.002808015327900648, + 0.009452697820961475, + 0.003713147481903434, + 0.012343618087470531, + 0.0032568583264946938, + 0.003816585522145033, + 0.017300721257925034, + 0.003995272796601057, + 0.004353824537247419, + 0.0071364277973771095, + 0.0031589989084750414, + 0.0044202180579304695, + 0.005589406006038189, + 0.004429613705724478, + 0.006190725602209568, + 0.012480822391808033, + 0.005155967548489571, + 0.0068453773856163025, + 0.004252013750374317, + 0.019423509016633034, + 0.019983608275651932, + 0.017227232456207275, + 0.028823208063840866, + 0.01674569956958294, + 0.01890186406672001, + 0.023424796760082245, + 0.01753382943570614, + 0.0039716013707220554, + 0.007461870554834604, + 0.007467373739928007, + 0.007179100066423416, + 0.006841070018708706, + 0.005741380620747805, + 0.011072022840380669, + 0.009390685707330704, + 0.018538616597652435, + 0.010761097073554993, + 0.021318864077329636, + 0.014858005568385124, + 0.014358130283653736, + 0.0245219673961401, + 0.022031499072909355, + 0.05257890000939369, + 0.0021678446792066097, + 0.004015351179987192, + 0.0037511158734560013, + 0.0031834631226956844, + 0.002849269425496459, + 0.0038213920779526234, + 0.002817857777699828, + 0.004382433835417032, + 0.003908551763743162, + 0.0030462874565273523, + 0.006536115426570177, + 0.007160881534218788, + 0.0077553098089993, + 0.006836057640612125, + 0.006975539494305849, + 0.009471993893384933, + 0.006035903934389353, + 0.008661838248372078, + 0.01566101238131523, + 0.015902359038591385, + 0.020462503656744957, + 0.017818056046962738, + 0.025350911542773247, + 0.020664682611823082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.008087978698313236 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005269210785627365 + ], + "opposite_logprob_tokens": [ + -5.2552690505981445 + ], + "logprob_reward": -0.005269210785627365, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005269210785627365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07683820277452469, + 0.0727294310927391, + 0.032310161739587784, + 0.06939050555229187, + 0.32168883085250854, + 0.09064934402704239, + 0.07953766733407974, + 0.02771998755633831, + 0.08673734962940216, + 0.14239850640296936 + ], + "credit_map": [ + 0.009200602769851685, + 0.010409251786768436, + 0.007109930273145437, + 0.012733513489365578, + 0.013710934668779373, + 0.0062027424573898315, + 0.01031583733856678, + 0.007155396975576878, + 0.018720874562859535, + 0.004338180646300316, + 0.002685107057914138, + 0.02192896232008934, + 0.0090779485180974, + 0.0024200694169849157, + 0.009752597659826279, + 0.0038056918419897556, + 0.0048200516030192375, + 0.0024629635736346245, + 0.002740842057392001, + 0.008245119825005531, + 0.003844495862722397, + 0.002607357455417514, + 0.005106555763632059, + 0.002482776762917638, + 0.005249525886029005, + 0.007855216972529888, + 0.0058482070453464985, + 0.008921058848500252, + 0.018458032980561256, + 0.005945986602455378, + 0.009868664667010307, + 0.00724381860345602, + 0.039345983415842056, + 0.040139090269804, + 0.033088669180870056, + 0.053581587970256805, + 0.03638897091150284, + 0.03976058587431908, + 0.044479407370090485, + 0.03490455821156502, + 0.009342296980321407, + 0.011481475085020065, + 0.014310779981315136, + 0.011281908489763737, + 0.012797816656529903, + 0.00767302792519331, + 0.012799191288650036, + 0.010962851345539093, + 0.008994179777801037, + 0.005578531417995691, + 0.0095356535166502, + 0.00688885198906064, + 0.00845383945852518, + 0.01172224897891283, + 0.009376301430165768, + 0.01898805983364582, + 0.002695787465199828, + 0.003076926339417696, + 0.003634799737483263, + 0.0037064654752612114, + 0.0039301528595387936, + 0.004365723114460707, + 0.00334922899492085, + 0.002960903337225318, + 0.007533591706305742, + 0.005633077118545771, + 0.010717558674514294, + 0.010957034304738045, + 0.013029548339545727, + 0.013159947469830513, + 0.01144672092050314, + 0.014259873889386654, + 0.007894272916018963, + 0.010437730699777603, + 0.018082991242408752, + 0.01903834566473961, + 0.030429361388087273, + 0.018967358395457268, + 0.020282849669456482, + 0.01726558618247509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 5 + ], + "selected_reward": -0.005269210785627365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1314925253391266, + 0.08648909628391266, + 0.10805905610322952, + 0.062015362083911896, + 0.21111415326595306, + 0.09001927822828293, + 0.12539145350456238, + 0.02338498830795288, + 0.06550562381744385, + 0.09652841836214066 + ], + "credit_map": [ + 0.019489102065563202, + 0.01230744831264019, + 0.010464563965797424, + 0.02169949933886528, + 0.02405134029686451, + 0.014905819669365883, + 0.016256557777523994, + 0.012318192981183529, + 0.019718576222658157, + 0.006995938718318939, + 0.009542956948280334, + 0.019725749269127846, + 0.01089772954583168, + 0.004229705780744553, + 0.009873345494270325, + 0.0055050915107131, + 0.01778951846063137, + 0.007110167294740677, + 0.013891139067709446, + 0.02556057833135128, + 0.010924828238785267, + 0.00778757082298398, + 0.015546783804893494, + 0.009448467753827572, + 0.00710698775947094, + 0.005957586690783501, + 0.008331780321896076, + 0.007164518814533949, + 0.008603029884397984, + 0.009731032885611057, + 0.008214106783270836, + 0.006906319409608841, + 0.028425374999642372, + 0.026201525703072548, + 0.022824782878160477, + 0.03577137738466263, + 0.02370336465537548, + 0.027121776714920998, + 0.025143634527921677, + 0.02192232757806778, + 0.009751041419804096, + 0.011126892641186714, + 0.011998087167739868, + 0.012062755413353443, + 0.010740493424236774, + 0.006161934696137905, + 0.017624536529183388, + 0.010553528554737568, + 0.010593953542411327, + 0.006431710906326771, + 0.014458593912422657, + 0.011650379747152328, + 0.013488217256963253, + 0.016382332891225815, + 0.014851014129817486, + 0.03753525763750076, + 0.0022469128016382456, + 0.0028717918321490288, + 0.002728518098592758, + 0.0026137179229408503, + 0.0027998252771794796, + 0.0033295603934675455, + 0.003404242917895317, + 0.003390419762581587, + 0.00313565949909389, + 0.005287409294396639, + 0.008274990133941174, + 0.009214178659021854, + 0.011662312783300877, + 0.009673486463725567, + 0.010012831538915634, + 0.008244756609201431, + 0.005088488105684519, + 0.007272948510944843, + 0.011191715486347675, + 0.010677402839064598, + 0.01657167449593544, + 0.014277057722210884, + 0.019363349303603172, + 0.012085780501365662 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 6 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06205030158162117 + ], + "opposite_logprob_tokens": [ + -2.8120503425598145 + ], + "logprob_reward": -0.06205030158162117, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06205030158162117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10124234110116959, + 0.1390215903520584, + 0.04066520184278488, + 0.07056213915348053, + 0.14088137447834015, + 0.22174449265003204, + 0.1264052838087082, + 0.03190692514181137, + 0.053100619465112686, + 0.07447006553411484 + ], + "credit_map": [ + 0.015734679996967316, + 0.012589573860168457, + 0.010544108226895332, + 0.013018226251006126, + 0.017189444974064827, + 0.01269976794719696, + 0.01082905475050211, + 0.008637478575110435, + 0.028378864750266075, + 0.010947278700768948, + 0.010394626297056675, + 0.030115962028503418, + 0.022727372124791145, + 0.010015022940933704, + 0.016942869871854782, + 0.009499597363173962, + 0.004032905213534832, + 0.0032494456972926855, + 0.006544320844113827, + 0.006411112379282713, + 0.005439626518636942, + 0.004119439981877804, + 0.00580764701589942, + 0.00506070489063859, + 0.010142994113266468, + 0.0066101085394620895, + 0.011436596512794495, + 0.008885745890438557, + 0.010088218376040459, + 0.008911756798624992, + 0.008550017140805721, + 0.005936702247709036, + 0.017485234886407852, + 0.017592448741197586, + 0.015323699451982975, + 0.024682966992259026, + 0.014766762033104897, + 0.018040457740426064, + 0.017866171896457672, + 0.015123624354600906, + 0.023435676470398903, + 0.027917593717575073, + 0.029304470866918564, + 0.031195776537060738, + 0.026557905599474907, + 0.014753379859030247, + 0.04043515399098396, + 0.028144538402557373, + 0.01697673834860325, + 0.011954422108829021, + 0.016575820744037628, + 0.013004839420318604, + 0.011580434627830982, + 0.016982337459921837, + 0.01709180697798729, + 0.022238874807953835, + 0.0034272295888513327, + 0.003426353679969907, + 0.004062880761921406, + 0.004186754114925861, + 0.004370096605271101, + 0.004127404652535915, + 0.003145211609080434, + 0.005160992965102196, + 0.003924211021512747, + 0.005418289918452501, + 0.007549835368990898, + 0.0065121715888381, + 0.009257582016289234, + 0.0065109822899103165, + 0.0048761190846562386, + 0.009051425382494926, + 0.0038084913976490498, + 0.005490908864885569, + 0.00825241208076477, + 0.009232455864548683, + 0.013286047615110874, + 0.010855997912585735, + 0.015307565219700336, + 0.008236187510192394 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 1 + ], + "selected_reward": -0.06205030158162117 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636727273464203 + ], + "opposite_logprob_tokens": [ + -5.379636764526367 + ], + "logprob_reward": -0.004636727273464203, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636727273464203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15607281029224396, + 0.0477929525077343, + 0.0219431072473526, + 0.018673723563551903, + 0.12281473726034164, + 0.0386735238134861, + 0.1366228610277176, + 0.10468930751085281, + 0.1604948341846466, + 0.19222216308116913 + ], + "credit_map": [ + 0.014342065900564194, + 0.016004392877221107, + 0.009185017086565495, + 0.03144556283950806, + 0.03654716908931732, + 0.009992571547627449, + 0.02652234211564064, + 0.012033688835799694, + 0.004097769036889076, + 0.004722794517874718, + 0.0033537240233272314, + 0.007714488077908754, + 0.013783029280602932, + 0.003100642701610923, + 0.006718814373016357, + 0.0043016900308430195, + 0.0028072199784219265, + 0.001842755707912147, + 0.0036474145017564297, + 0.003328660735860467, + 0.0033332491293549538, + 0.0019084130180999637, + 0.002675618277862668, + 0.0023997763637453318, + 0.0021290930453687906, + 0.0019666554871946573, + 0.00208244938403368, + 0.0026874663308262825, + 0.0035856314934790134, + 0.0014367456315085292, + 0.002239430556073785, + 0.0025462519843131304, + 0.011908001266419888, + 0.015814993530511856, + 0.012272683903574944, + 0.02266019396483898, + 0.012648316100239754, + 0.012339271605014801, + 0.020373472943902016, + 0.014797802083194256, + 0.0036449518520385027, + 0.004763263743370771, + 0.005508003756403923, + 0.004746052902191877, + 0.004629912786185741, + 0.0040363771840929985, + 0.006287120282649994, + 0.005057839211076498, + 0.01158576924353838, + 0.010825121775269508, + 0.017417918890714645, + 0.01117938756942749, + 0.01282296609133482, + 0.019194304943084717, + 0.014356262050569057, + 0.039241135120391846, + 0.008020485751330853, + 0.01166408509016037, + 0.016013214364647865, + 0.014053295366466045, + 0.015671493485569954, + 0.013370617292821407, + 0.014829721301794052, + 0.011066393926739693, + 0.00830407440662384, + 0.013150215148925781, + 0.021235905587673187, + 0.020880497992038727, + 0.02774711512029171, + 0.0202843826264143, + 0.023168686777353287, + 0.025723960250616074, + 0.007868940941989422, + 0.016333797946572304, + 0.02020174264907837, + 0.03527484089136124, + 0.02563750185072422, + 0.02535693347454071, + 0.039704579859972, + 0.02184382639825344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.004636727273464203 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, medium shot captures a figure seated in perfect lotus pose on a simple, dark cushion within a tranquil meditation chamber. They are draped in a magnificent, deep indigo silk robe, its heavy fabric falling in elegant folds. The setting consists of dark, polished wooden floors, and against the muted backdrop, thin incense smoke curls upward from a small altar. A shaft of dusty afternoon light filters through the tall, latticed windows, illuminating the figure and creating a dramatic play of light and shadow. The figure's chest rises and falls in a slow, low, rhythmic rhythm, the heavy silk of the robe shifting subtly with each steady breath. As this tranquil movement continues, the camera begins a slow, deliberate zoom, sliding forward to focus intimately on the robe's chest area. The lens draws the viewer's gaze from the overall form to the breathtaking microcosm of the intricate embroidery, revealing the texture of each gold and silver metallic stitch, forming swirling celestial dragons and dark clouds, creating a profound atmosphere of deep concentration and spiritual tranquility.", + "seed": 4927, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_8006_seed_4927_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012210834329016507, + "vlm_logprob": -0.012210834329016507, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.012210834329016507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021516816690564156 + ], + "opposite_logprob_tokens": [ + -3.896516799926758 + ], + "logprob_reward": -0.021516816690564156, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.021516816690564156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2626264691352844, + 0.02152727171778679, + 0.017208106815814972, + 0.02568674087524414, + 0.07330404967069626, + 0.11894974112510681, + 0.101557157933712, + 0.23815709352493286, + 0.10694711655378342, + 0.034036241471767426 + ], + "credit_map": [ + 0.024823373183608055, + 0.015819627791643143, + 0.016479501500725746, + 0.035918619483709335, + 0.04042964428663254, + 0.03654235601425171, + 0.05101999640464783, + 0.04159337654709816, + 0.003153532510623336, + 0.0012450573267415166, + 0.00221957266330719, + 0.0025400130543857813, + 0.0019077614415436983, + 0.0025796634145081043, + 0.0032201805151998997, + 0.0046614911407232285, + 0.0022451081313192844, + 0.0013582640094682574, + 0.002836997387930751, + 0.0017591181676834822, + 0.0011771336430683732, + 0.0032723171170800924, + 0.0017061493126675487, + 0.002853019628673792, + 0.0032208601478487253, + 0.001894546439871192, + 0.0021951692178845406, + 0.0046926154755055904, + 0.0022249193862080574, + 0.003475107718259096, + 0.004695544485002756, + 0.0032879775390028954, + 0.008227190002799034, + 0.007766438648104668, + 0.010533250868320465, + 0.01054043136537075, + 0.00879524927586317, + 0.009953444823622704, + 0.007014693692326546, + 0.01047335285693407, + 0.013440875336527824, + 0.020974494516849518, + 0.011807860806584358, + 0.015593014657497406, + 0.01599820889532566, + 0.01353723369538784, + 0.015988092869520187, + 0.011609961278736591, + 0.008398492820560932, + 0.008700129576027393, + 0.01273626834154129, + 0.015131344087421894, + 0.011648556217551231, + 0.012026078999042511, + 0.0198536179959774, + 0.013062666170299053, + 0.027292711660265923, + 0.022356640547513962, + 0.04027147591114044, + 0.030329616740345955, + 0.015012591145932674, + 0.03827008977532387, + 0.027881711721420288, + 0.03674224019050598, + 0.008496533147990704, + 0.012161687016487122, + 0.020414896309375763, + 0.014495613053441048, + 0.008771499618887901, + 0.011762315407395363, + 0.012995775789022446, + 0.017848799005150795, + 0.004667072091251612, + 0.0037953811697661877, + 0.004125606268644333, + 0.004188240505754948, + 0.0030748178251087666, + 0.0036159066949039698, + 0.005502883344888687, + 0.005066332872956991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.021516816690564156 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019521719077602029 + ], + "opposite_logprob_tokens": [ + -6.251952171325684 + ], + "logprob_reward": -0.0019521719077602029, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019521719077602029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1946573704481125, + 0.07780791074037552, + 0.04491032287478447, + 0.08403905481100082, + 0.23320892453193665, + 0.049181386828422546, + 0.054048802703619, + 0.10554303228855133, + 0.09717196226119995, + 0.059431273490190506 + ], + "credit_map": [ + 0.015430600382387638, + 0.011392013169825077, + 0.013168994337320328, + 0.01839507184922695, + 0.02069074846804142, + 0.01728002168238163, + 0.04775995388627052, + 0.05053996294736862, + 0.011718948371708393, + 0.009364789351820946, + 0.008741968311369419, + 0.00707095954567194, + 0.005418653599917889, + 0.009917453862726688, + 0.008873146027326584, + 0.016701988875865936, + 0.00512410094961524, + 0.0031650580931454897, + 0.006786568555980921, + 0.005359834525734186, + 0.00667624780908227, + 0.006686999928206205, + 0.0036142372991889715, + 0.007497277110815048, + 0.007819274440407753, + 0.004642376210540533, + 0.006814286112785339, + 0.012467671185731888, + 0.006962332874536514, + 0.01633085310459137, + 0.011935117654502392, + 0.017067141830921173, + 0.013622426427900791, + 0.012178231030702591, + 0.027010221034288406, + 0.023443687707185745, + 0.022634131833910942, + 0.05697951093316078, + 0.0161447711288929, + 0.061195939779281616, + 0.00906870886683464, + 0.008693477138876915, + 0.0035412744618952274, + 0.007067953702062368, + 0.005857502575963736, + 0.005084224510937929, + 0.004424611572176218, + 0.005443629343062639, + 0.005614741239696741, + 0.004888713825494051, + 0.005283016245812178, + 0.008096743375062943, + 0.007339926436543465, + 0.006441442295908928, + 0.007409179583191872, + 0.008975040167570114, + 0.013069580309092999, + 0.011030692607164383, + 0.015230205841362476, + 0.016676127910614014, + 0.007182343862950802, + 0.013502346351742744, + 0.012810947373509407, + 0.016040781512856483, + 0.007529499474912882, + 0.012681241147220135, + 0.01891348883509636, + 0.013152343221008778, + 0.00606361823156476, + 0.010493308305740356, + 0.011587265878915787, + 0.01675119251012802, + 0.007535694632679224, + 0.007672193925827742, + 0.006189355161041021, + 0.008433954790234566, + 0.004871974233537912, + 0.006858957931399345, + 0.00913933478295803, + 0.008729812689125538 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.0019521719077602029 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007624453864991665 + ], + "opposite_logprob_tokens": [ + -4.882624626159668 + ], + "logprob_reward": -0.007624453864991665, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007624453864991665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22290267050266266, + 0.1644188016653061, + 0.031048908829689026, + 0.07384598255157471, + 0.13839805126190186, + 0.07447222620248795, + 0.05520198494195938, + 0.08334948122501373, + 0.1151614785194397, + 0.0412004292011261 + ], + "credit_map": [ + 0.018727904185652733, + 0.015083557926118374, + 0.013430069200694561, + 0.01769464835524559, + 0.024011611938476562, + 0.026565156877040863, + 0.04907539114356041, + 0.05831431224942207, + 0.020234204828739166, + 0.013709207996726036, + 0.023045672103762627, + 0.01576547883450985, + 0.010513967834413052, + 0.020959900692105293, + 0.022296054288744926, + 0.037894319742918015, + 0.0029482850804924965, + 0.0026648289058357477, + 0.004995355382561684, + 0.004557440988719463, + 0.0028515460435301065, + 0.004664210602641106, + 0.0035816181916743517, + 0.004785624798387289, + 0.006744264625012875, + 0.004887323826551437, + 0.007510526571422815, + 0.009551110677421093, + 0.005869767162948847, + 0.01408524252474308, + 0.010328831151127815, + 0.014868911355733871, + 0.01177321095019579, + 0.01028905063867569, + 0.013089173473417759, + 0.014688833616673946, + 0.005284289363771677, + 0.03577301278710365, + 0.008015492931008339, + 0.039484985172748566, + 0.009255565702915192, + 0.008106660097837448, + 0.009214391000568867, + 0.010370990261435509, + 0.008186700753867626, + 0.009343847632408142, + 0.011518369428813457, + 0.00847569853067398, + 0.005101837683469057, + 0.004593710880726576, + 0.007400756701827049, + 0.008372285403311253, + 0.0066656144335865974, + 0.006344001740217209, + 0.009323034435510635, + 0.00740074273198843, + 0.007295240182429552, + 0.008098701015114784, + 0.008226117119193077, + 0.01308346726000309, + 0.006344394292682409, + 0.015248028561472893, + 0.01252737082540989, + 0.012526160106062889, + 0.007904556579887867, + 0.00883297435939312, + 0.034473977982997894, + 0.012644975446164608, + 0.004452656954526901, + 0.008046140894293785, + 0.008000283502042294, + 0.03080592304468155, + 0.006266055628657341, + 0.006366702262312174, + 0.005601534154266119, + 0.005648903548717499, + 0.0030336009804159403, + 0.0042195627465844154, + 0.004648290574550629, + 0.0054157814010977745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.007624453864991665 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026347355917096138 + ], + "opposite_logprob_tokens": [ + -3.6513473987579346 + ], + "logprob_reward": -0.026347355917096138, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026347355917096138, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27437078952789307, + 0.2492510825395584, + 0.049302052706480026, + 0.06784462183713913, + 0.13184905052185059, + 0.05821133032441139, + 0.04416147246956825, + 0.04121122136712074, + 0.049402035772800446, + 0.034396324306726456 + ], + "credit_map": [ + 0.024670438840985298, + 0.012775584124028683, + 0.017597509548068047, + 0.014510637149214745, + 0.01228458434343338, + 0.02693387120962143, + 0.07156696170568466, + 0.09403122961521149, + 0.05155413597822189, + 0.03222276270389557, + 0.033631425350904465, + 0.023914320394396782, + 0.020316321402788162, + 0.023925229907035828, + 0.025273198261857033, + 0.03841371461749077, + 0.0058868080377578735, + 0.0033685702364891768, + 0.013586717657744884, + 0.005135027691721916, + 0.002937343204393983, + 0.005825134459882975, + 0.0043885535560548306, + 0.008173906244337559, + 0.007032206282019615, + 0.005068079102784395, + 0.007318357937037945, + 0.007839814759790897, + 0.004955505020916462, + 0.013123104348778725, + 0.007656401488929987, + 0.014851158484816551, + 0.01071716845035553, + 0.01063542626798153, + 0.012078474275767803, + 0.012804806232452393, + 0.007373489439487457, + 0.03346695005893707, + 0.005926864687353373, + 0.03884589299559593, + 0.011588509194552898, + 0.006688370369374752, + 0.005467959214001894, + 0.007783513516187668, + 0.004847302101552486, + 0.007971908897161484, + 0.006554963067173958, + 0.007308808621019125, + 0.004652934614568949, + 0.004204666707664728, + 0.004829137120395899, + 0.007492803502827883, + 0.0034269217867404222, + 0.004948386922478676, + 0.004299547057598829, + 0.010307076387107372, + 0.003607406746596098, + 0.004847608506679535, + 0.005983198527246714, + 0.006601567380130291, + 0.002570562297478318, + 0.0060261571779847145, + 0.005364073906093836, + 0.006210655439645052, + 0.004784190561622381, + 0.0054541900753974915, + 0.011771013960242271, + 0.005459848325699568, + 0.0025140775833278894, + 0.0038414117880165577, + 0.0034161030780524015, + 0.012161205522716045, + 0.004881621804088354, + 0.004730619955807924, + 0.004944234620779753, + 0.003987508825957775, + 0.002693664748221636, + 0.003481572028249502, + 0.005106149706989527, + 0.00457095867022872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.026347355917096138 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003613373264670372 + ], + "opposite_logprob_tokens": [ + -5.628613471984863 + ], + "logprob_reward": -0.003613373264670372, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003613373264670372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1405179500579834, + 0.028192812576889992, + 0.026329703629016876, + 0.07687842845916748, + 0.060979560017585754, + 0.03485790640115738, + 0.13546502590179443, + 0.1427483856678009, + 0.14841492474079132, + 0.20561528205871582 + ], + "credit_map": [ + 0.013519568368792534, + 0.0094708651304245, + 0.008559542708098888, + 0.021074211224913597, + 0.01965847797691822, + 0.0152049008756876, + 0.031397245824337006, + 0.02163313701748848, + 0.0054743411019444466, + 0.002444836078211665, + 0.004966581240296364, + 0.002635825891047716, + 0.0021963962353765965, + 0.002917665522545576, + 0.0029183970764279366, + 0.004638770595192909, + 0.0038479757495224476, + 0.0025816019624471664, + 0.0025563829112797976, + 0.0027842498384416103, + 0.0031261048279702663, + 0.004570502322167158, + 0.002770150313153863, + 0.0040927366353571415, + 0.007927747443318367, + 0.004172380547970533, + 0.006074868142604828, + 0.010902932845056057, + 0.0074765924364328384, + 0.014386119320988655, + 0.010397748090326786, + 0.0155400475487113, + 0.005032066255807877, + 0.004914695862680674, + 0.01161732617765665, + 0.007048991974443197, + 0.005685744807124138, + 0.010459323413670063, + 0.004627803340554237, + 0.01159361656755209, + 0.004735514987260103, + 0.0031734926160424948, + 0.004720388911664486, + 0.004841334652155638, + 0.004136224742978811, + 0.004263714421540499, + 0.003531693248078227, + 0.005455546546727419, + 0.015090977773070335, + 0.01365650538355112, + 0.01725822128355503, + 0.0229791309684515, + 0.013154826126992702, + 0.011403340846300125, + 0.017458634451031685, + 0.024463394656777382, + 0.009761893190443516, + 0.012560557574033737, + 0.024770047515630722, + 0.025404371321201324, + 0.008576416410505772, + 0.025037478655576706, + 0.014870512299239635, + 0.02176710218191147, + 0.009492103941738605, + 0.00798494927585125, + 0.04347900301218033, + 0.014920125715434551, + 0.006005471106618643, + 0.011886019259691238, + 0.011062058620154858, + 0.04358521103858948, + 0.02635018341243267, + 0.01864687353372574, + 0.02120969630777836, + 0.028038015589118004, + 0.016395702958106995, + 0.017251206561923027, + 0.047521088272333145, + 0.030202528461813927 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.003613373264670372 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, slightly low-angled, medium shot captures a figure seated in perfect lotus pose on a simple, dark cushion within a tranquil meditation chamber. They are draped in a magnificent, deep indigo silk robe, its heavy fabric falling in elegant folds. The setting consists of dark, polished wooden floors, and against the muted backdrop, thin incense smoke curls upward from a small altar. A shaft of dusty afternoon light filters through the tall, latticed windows, illuminating the figure and creating a dramatic play of light and shadow. The figure's chest rises and falls in a slow, low, rhythmic rhythm, the heavy silk of the robe shifting subtly with each steady breath. As this tranquil movement continues, the camera begins a slow, deliberate zoom, sliding forward to focus intimately on the robe's chest area. The lens draws the viewer's gaze from the overall form to the breathtaking microcosm of the intricate embroidery, revealing the texture of each gold and silver metallic stitch, forming swirling celestial dragons and dark clouds, creating a profound atmosphere of deep concentration and spiritual tranquility.", + "seed": 4928, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_15/index_8006_seed_4928_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.013725146558135748, + "vlm_logprob": -0.013725146558135748, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.013725146558135748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004446025937795639 + ], + "opposite_logprob_tokens": [ + -6.004446029663086 + ], + "logprob_reward": -0.004446025937795639, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.004446025937795639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10498085618019104, + 0.04040229320526123, + 0.03869045898318291, + 0.2953091859817505, + 0.20495682954788208, + 0.032392021268606186, + 0.06339506059885025, + 0.11475904285907745, + 0.05543449893593788, + 0.049679771065711975 + ], + "credit_map": [ + 0.006892576813697815, + 0.008095640689134598, + 0.03236730769276619, + 0.009803089313209057, + 0.01096286065876484, + 0.006649985443800688, + 0.009352343156933784, + 0.02085704542696476, + 0.004679995123296976, + 0.004500562325119972, + 0.005765085108578205, + 0.004835250321775675, + 0.00359205505810678, + 0.0052338154055178165, + 0.00448751263320446, + 0.007308015134185553, + 0.002875345526263118, + 0.0028582410886883736, + 0.011842470616102219, + 0.005056245252490044, + 0.0027599073946475983, + 0.002089311135932803, + 0.00525714410468936, + 0.005951785948127508, + 0.01648401841521263, + 0.010408650152385235, + 0.029149308800697327, + 0.022427281364798546, + 0.016059428453445435, + 0.010049879550933838, + 0.1074507012963295, + 0.08327988535165787, + 0.04712824523448944, + 0.007683067116886377, + 0.011794605292379856, + 0.0064969356171786785, + 0.007451778277754784, + 0.06511066854000092, + 0.014801543205976486, + 0.044489964842796326, + 0.006450065411627293, + 0.005279728677123785, + 0.00365467625670135, + 0.0031827562488615513, + 0.0046728006564080715, + 0.002725900150835514, + 0.003936894703656435, + 0.002489191247150302, + 0.0052499049343168736, + 0.008898450061678886, + 0.00932303350418806, + 0.005936514586210251, + 0.011761513538658619, + 0.0037445309571921825, + 0.012241820804774761, + 0.006239281501621008, + 0.007714945822954178, + 0.008839437738060951, + 0.02365591935813427, + 0.012251966632902622, + 0.014713332056999207, + 0.007474996615201235, + 0.027675552293658257, + 0.012432876043021679, + 0.005703128408640623, + 0.004512494429945946, + 0.010682893916964531, + 0.007163351401686668, + 0.005684849806129932, + 0.007112907245755196, + 0.008879556320607662, + 0.0056953090243041515, + 0.005680704023689032, + 0.00511453440412879, + 0.00831130426377058, + 0.006878388114273548, + 0.005463555455207825, + 0.005946374498307705, + 0.00566514115780592, + 0.006619764491915703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 7 + ], + "selected_reward": -0.004446025937795639 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00764752272516489 + ], + "opposite_logprob_tokens": [ + -4.882647514343262 + ], + "logprob_reward": -0.00764752272516489, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00764752272516489, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10451095551252365, + 0.03656960651278496, + 0.04446006938815117, + 0.07679814845323563, + 0.20034697651863098, + 0.15139134228229523, + 0.10888694226741791, + 0.07028307765722275, + 0.06892140954732895, + 0.1378314197063446 + ], + "credit_map": [ + 0.008081031031906605, + 0.008041437715291977, + 0.028134485706686974, + 0.012620656751096249, + 0.010062722489237785, + 0.007315079681575298, + 0.011115665547549725, + 0.019139882177114487, + 0.004359961953014135, + 0.0030183305498212576, + 0.0059868330135941505, + 0.004426825791597366, + 0.0037524744402617216, + 0.006133623421192169, + 0.00433011818677187, + 0.004561443347483873, + 0.0037828178610652685, + 0.004818468354642391, + 0.005552626680582762, + 0.011027047410607338, + 0.0051643382757902145, + 0.003047577803954482, + 0.007236956153064966, + 0.0038302394095808268, + 0.004406729247421026, + 0.0035847641993314028, + 0.007156203035265207, + 0.0058296783827245235, + 0.004579372238367796, + 0.003192137461155653, + 0.026840338483452797, + 0.021208930760622025, + 0.04939860478043556, + 0.0061353049241006374, + 0.013302423059940338, + 0.005843733437359333, + 0.007455169223248959, + 0.07067018747329712, + 0.008049140684306622, + 0.03949242830276489, + 0.042231958359479904, + 0.012978424318134785, + 0.014841921627521515, + 0.006797978188842535, + 0.011620758101344109, + 0.01156105287373066, + 0.012140565551817417, + 0.03921869769692421, + 0.010369425639510155, + 0.011071777902543545, + 0.0164582971483469, + 0.011927791871130466, + 0.01753098890185356, + 0.006177143659442663, + 0.02179756760597229, + 0.01355394721031189, + 0.004629813134670258, + 0.00612375745549798, + 0.014263557270169258, + 0.008932298049330711, + 0.008316464722156525, + 0.004992340691387653, + 0.013403505086898804, + 0.009621347300708294, + 0.006666443310678005, + 0.003237458411604166, + 0.013099472038447857, + 0.007831034250557423, + 0.006704029161483049, + 0.012527134269475937, + 0.009615056216716766, + 0.009240781888365746, + 0.013543952256441116, + 0.006730351131409407, + 0.03006567247211933, + 0.010924655944108963, + 0.009974976070225239, + 0.03150659054517746, + 0.009237630292773247, + 0.02584759145975113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.00764752272516489 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006734294816851616 + ], + "opposite_logprob_tokens": [ + -5.006734371185303 + ], + "logprob_reward": -0.006734294816851616, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006734294816851616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2591644525527954, + 0.1094181090593338, + 0.0365966334939003, + 0.0691528394818306, + 0.2730846405029297, + 0.02312905341386795, + 0.10204676538705826, + 0.036740824580192566, + 0.03345822915434837, + 0.05720842629671097 + ], + "credit_map": [ + 0.020906655117869377, + 0.01473704632371664, + 0.0867057517170906, + 0.026185963302850723, + 0.025427209213376045, + 0.015011901967227459, + 0.02178288996219635, + 0.048407044261693954, + 0.011773372069001198, + 0.00967982318252325, + 0.024395637214183807, + 0.014280421659350395, + 0.009413477964699268, + 0.012530189007520676, + 0.014473410323262215, + 0.01287178136408329, + 0.004580946639180183, + 0.0034642501268535852, + 0.005835537798702717, + 0.005952739622443914, + 0.0034766346216201782, + 0.005332453176379204, + 0.00450977124273777, + 0.0034443035256117582, + 0.005257966462522745, + 0.0030837554950267076, + 0.0067564393393695354, + 0.005007252097129822, + 0.004237002693116665, + 0.0037422513123601675, + 0.0220332033932209, + 0.019034966826438904, + 0.05906476080417633, + 0.008827200159430504, + 0.014405248686671257, + 0.006916203536093235, + 0.00964389182627201, + 0.08952067792415619, + 0.01846572570502758, + 0.06624092161655426, + 0.003140182700008154, + 0.003928566351532936, + 0.0030043451115489006, + 0.00242982292547822, + 0.0029532192274928093, + 0.0023833387531340122, + 0.002916386118158698, + 0.002373192459344864, + 0.011951766908168793, + 0.011994177475571632, + 0.02202010527253151, + 0.010168126784265041, + 0.013784762471914291, + 0.006051456090062857, + 0.0154793756082654, + 0.01059699896723032, + 0.0024585004430264235, + 0.003930384758859873, + 0.006311430130153894, + 0.0038387924432754517, + 0.005358364433050156, + 0.002862849971279502, + 0.0074830204248428345, + 0.004497478250414133, + 0.002694925758987665, + 0.0026611145585775375, + 0.0050631677731871605, + 0.0037676666397601366, + 0.004190992563962936, + 0.005514326971024275, + 0.004792522173374891, + 0.0047735124826431274, + 0.0071197268553078175, + 0.005643144249916077, + 0.011413116939365864, + 0.006597419735044241, + 0.006266121752560139, + 0.00758034223690629, + 0.004971891641616821, + 0.007616662420332432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.006734294816851616 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306454211473465 + ], + "opposite_logprob_tokens": [ + -3.168064594268799 + ], + "logprob_reward": -0.04306454211473465, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306454211473465, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17066697776317596, + 0.16125108301639557, + 0.06678039580583572, + 0.16596141457557678, + 0.22257168591022491, + 0.06738919764757156, + 0.048660971224308014, + 0.03874398022890091, + 0.02988276071846485, + 0.028091605752706528 + ], + "credit_map": [ + 0.022463088855147362, + 0.01782849244773388, + 0.030269119888544083, + 0.016878174617886543, + 0.017492791637778282, + 0.027187973260879517, + 0.021099787205457687, + 0.01744753122329712, + 0.021895138546824455, + 0.007841025479137897, + 0.026059450581669807, + 0.011974246241152287, + 0.01184158492833376, + 0.033156868070364, + 0.011500157415866852, + 0.036982595920562744, + 0.004680583719164133, + 0.0024645186495035887, + 0.0254308320581913, + 0.006066877860575914, + 0.0035412050783634186, + 0.004525915719568729, + 0.007941583171486855, + 0.012128875590860844, + 0.010554933920502663, + 0.004672732204198837, + 0.015314248390495777, + 0.01159636490046978, + 0.008376236073672771, + 0.007919392548501492, + 0.057864509522914886, + 0.049662988632917404, + 0.05970441922545433, + 0.007171127013862133, + 0.012966816313564777, + 0.0063696797005832195, + 0.007894185371696949, + 0.07446101307868958, + 0.013115651905536652, + 0.04088877886533737, + 0.017562171444296837, + 0.004583573900163174, + 0.0045857816003263, + 0.0024655754677951336, + 0.003892150241881609, + 0.003637810004875064, + 0.007754494436085224, + 0.022907637059688568, + 0.005726266652345657, + 0.005105680786073208, + 0.00724627124145627, + 0.004900707863271236, + 0.0057494500651955605, + 0.004391648806631565, + 0.010347398929297924, + 0.005193546414375305, + 0.0027001535054296255, + 0.0052149598486721516, + 0.007232567761093378, + 0.0040743728168308735, + 0.004405676387250423, + 0.0031175585463643074, + 0.0070095425471663475, + 0.004989145323634148, + 0.0033667644020169973, + 0.0019109654240310192, + 0.006526128854602575, + 0.003049298422411084, + 0.0030543303582817316, + 0.004253623075783253, + 0.003940546419471502, + 0.003781101666390896, + 0.003883176716044545, + 0.0020995857194066048, + 0.006090542301535606, + 0.0025829579681158066, + 0.002734755864366889, + 0.003591198241338134, + 0.00242812093347311, + 0.004681264981627464 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.04306454211473465 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a figure seated in a lotus pose on a dark cushion visible in the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a deep indigo silk robe with heavy fabric and elegant folds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene include dark polished wooden floors and an altar with thin incense smoke curling upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a shaft of dusty afternoon light filtering through tall latticed windows creating light and shadow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform a slow deliberate zoom forward to focus on the robe's chest area revealing gold and silver embroidery?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006733347196131945 + ], + "opposite_logprob_tokens": [ + -5.006733417510986 + ], + "logprob_reward": -0.006733347196131945, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006733347196131945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0547507219016552, + 0.02269909344613552, + 0.026703067123889923, + 0.056236691772937775, + 0.40777286887168884, + 0.05510741472244263, + 0.08011990785598755, + 0.09228304028511047, + 0.06438813358545303, + 0.1399390697479248 + ], + "credit_map": [ + 0.004537568893283606, + 0.008069938980042934, + 0.008451545611023903, + 0.00862210150808096, + 0.008022603578865528, + 0.00466346088796854, + 0.0076880138367414474, + 0.00469549186527729, + 0.002254633931443095, + 0.0022739761043339968, + 0.0034585341345518827, + 0.003324387827888131, + 0.0025803428143262863, + 0.0029476857744157314, + 0.00305986893363297, + 0.0027996611315757036, + 0.0020369982812553644, + 0.0027739210054278374, + 0.0036680814810097218, + 0.005424759816378355, + 0.003638466354459524, + 0.002534178551286459, + 0.003947707824409008, + 0.002678952179849148, + 0.005025600548833609, + 0.0027939171995967627, + 0.004755679983645678, + 0.004085578955709934, + 0.0034140530042350292, + 0.002555230399593711, + 0.017006047070026398, + 0.016600584611296654, + 0.052281659096479416, + 0.013965632766485214, + 0.021466068923473358, + 0.01283437293022871, + 0.014986914582550526, + 0.12765014171600342, + 0.030953720211982727, + 0.13363435864448547, + 0.015270505100488663, + 0.0035248356871306896, + 0.004836814943701029, + 0.0025245118886232376, + 0.0038685405161231756, + 0.0024685256648808718, + 0.003835000330582261, + 0.01877867802977562, + 0.007652196101844311, + 0.010865440592169762, + 0.01418338343501091, + 0.007095980457961559, + 0.013817883096635342, + 0.004599099978804588, + 0.01511048898100853, + 0.006795437540858984, + 0.006966365035623312, + 0.00902778934687376, + 0.011274457909166813, + 0.009969476610422134, + 0.011544788256287575, + 0.007520034443587065, + 0.022701255977153778, + 0.01327887736260891, + 0.007371385581791401, + 0.007733824197202921, + 0.009209022857248783, + 0.009001464582979679, + 0.007414764259010553, + 0.007495081517845392, + 0.008986348286271095, + 0.007176242768764496, + 0.023296166211366653, + 0.014110312797129154, + 0.02380537986755371, + 0.017176833003759384, + 0.015793608501553535, + 0.01378121692687273, + 0.014226417057216167, + 0.017749140039086342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.006733347196131945 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_tensor([1248, 1248])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_tensor([1248, 1248])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..da584133248ba39cc7db6904f32c6b71be8d64e6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_tensor([1248, 1248])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot is fixed on a vibrant, small, red plastic toy sailboat, lying perfectly still in the glassy, clear water that fills a pristine white ceramic bathtub. Soft, warm overhead lighting illuminates the scene, creating gentle highlights on the sleek hull and casting subtle, shimmering reflections on the water, creating an atmosphere of quiet concentration. A silent ripple begins to move across the water, its concentric rings methodically approaching the toy. As the wave reaches the boat, it deliberately lifts one side, causing it to sway precariously. The boat's simple, white plastic mast begins to slowly and inexorably tilt toward the water, triggering a capsize. In one swift, graceful motion, the boat flips completely, mast and sail disappearing beneath the surface as it turns. At the end of the motion, the toy floats motionless, completely upside down, its bright red hull now face-up, gently rocking in the dissipating aftermath.", + "seed": 19189, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19189_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.02304655686020851, + "vlm_logprob": -0.02304655686020851, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.02304655686020851, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017026441171765327 + ], + "opposite_logprob_tokens": [ + -4.517026424407959 + ], + "logprob_reward": -0.017026441171765327, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017026441171765327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.032937098294496536, + 0.05206424742937088, + 0.025252295657992363, + 0.294927716255188, + 0.35546666383743286, + 0.14126402139663696, + 0.030251920223236084, + 0.010395889170467854, + 0.005375804845243692, + 0.052064359188079834 + ], + "credit_map": [ + 0.003615826601162553, + 0.007022877689450979, + 0.0038719179574400187, + 0.002152440370991826, + 0.004639931954443455, + 0.004586255177855492, + 0.00301833963021636, + 0.00402950681746006, + 0.004506328608840704, + 0.007391245570033789, + 0.005503011867403984, + 0.004251442849636078, + 0.009127754718065262, + 0.008312540128827095, + 0.00767796253785491, + 0.005293963942676783, + 0.002836328698322177, + 0.003782039973884821, + 0.005812477320432663, + 0.0019154491601511836, + 0.0028077331371605396, + 0.0031081228516995907, + 0.0025332404766231775, + 0.0024569055531173944, + 0.007322337944060564, + 0.00946192815899849, + 0.010082685388624668, + 0.006925092078745365, + 0.012153062969446182, + 0.11863914132118225, + 0.011341109871864319, + 0.11900234967470169, + 0.02149091847240925, + 0.02034647762775421, + 0.025481397286057472, + 0.030941303819417953, + 0.04140082374215126, + 0.06915999948978424, + 0.06072583794593811, + 0.08591988682746887, + 0.0036132116802036762, + 0.0032606124877929688, + 0.005274217575788498, + 0.0032137639354914427, + 0.008004020899534225, + 0.05979849025607109, + 0.0021182361524552107, + 0.055981479585170746, + 0.0024237793404608965, + 0.003333380213007331, + 0.004503516014665365, + 0.0030242649372667074, + 0.006036557722836733, + 0.003021489828824997, + 0.00398766016587615, + 0.003921270836144686, + 0.0012011565268039703, + 0.0014628159115090966, + 0.0014741539489477873, + 0.0010782280005514622, + 0.0013753448147326708, + 0.0013767988421022892, + 0.0010106696281582117, + 0.0014167207991704345, + 0.00045638694427907467, + 0.0006471641245298088, + 0.0009159219334833324, + 0.0007541675586253405, + 0.0006783163407817483, + 0.0006602107896469533, + 0.0006894610123708844, + 0.0005741758504882455, + 0.0034903017804026604, + 0.0038271804805845022, + 0.014395220205187798, + 0.007930678315460682, + 0.0031621085945516825, + 0.010680170729756355, + 0.002674233866855502, + 0.005904464982450008 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 5 + ], + "selected_reward": -0.017026441171765327 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011089949868619442 + ], + "opposite_logprob_tokens": [ + -4.51108980178833 + ], + "logprob_reward": -0.011089949868619442, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011089949868619442, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08520013839006424, + 0.053621843457221985, + 0.08231554925441742, + 0.05749509111046791, + 0.33462658524513245, + 0.2118251621723175, + 0.03219621628522873, + 0.0490303635597229, + 0.016233069822192192, + 0.07745592296123505 + ], + "credit_map": [ + 0.007854367606341839, + 0.02036241814494133, + 0.012472293339669704, + 0.004130769520998001, + 0.012949297204613686, + 0.013714012689888477, + 0.00486443005502224, + 0.00885255727916956, + 0.004703196696937084, + 0.007687023840844631, + 0.00602623401209712, + 0.00481270719319582, + 0.009925223886966705, + 0.007221767213195562, + 0.00872125942260027, + 0.004524431191384792, + 0.004887270741164684, + 0.005467424634844065, + 0.010734242387115955, + 0.002792158629745245, + 0.005248864181339741, + 0.023279165849089622, + 0.0052694291807711124, + 0.024636993184685707, + 0.002355054719373584, + 0.0037686116993427277, + 0.0036012884229421616, + 0.0023776611778885126, + 0.0034900973550975323, + 0.01882234215736389, + 0.0033683273941278458, + 0.019711710512638092, + 0.019051039591431618, + 0.019891511648893356, + 0.0454123429954052, + 0.020670097321271896, + 0.030598537996411324, + 0.07427217066287994, + 0.03716539219021797, + 0.08756548911333084, + 0.0051370649598538876, + 0.0047494578175246716, + 0.008938722312450409, + 0.004838946275413036, + 0.011849949136376381, + 0.08899915218353271, + 0.0027260021306574345, + 0.08458586782217026, + 0.002608840586617589, + 0.003810921683907509, + 0.004461181350052357, + 0.0034251571632921696, + 0.006263151299208403, + 0.003681801725178957, + 0.0029003615491092205, + 0.005044803023338318, + 0.005120929796248674, + 0.004965939559042454, + 0.00913600716739893, + 0.006384919863194227, + 0.004570537246763706, + 0.007132600527256727, + 0.003631971310824156, + 0.008087456226348877, + 0.0017315121367573738, + 0.0018513533286750317, + 0.0034103267826139927, + 0.0024737834464758635, + 0.0016072713769972324, + 0.0019394602859392762, + 0.001867437269538641, + 0.0013519253116101027, + 0.005020648706704378, + 0.00645371014252305, + 0.020758382976055145, + 0.012369012460112572, + 0.00567989144474268, + 0.01524439174681902, + 0.00507945055142045, + 0.006850440055131912 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.011089949868619442 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867635294795036 + ], + "opposite_logprob_tokens": [ + -3.0486762523651123 + ], + "logprob_reward": -0.04867635294795036, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867635294795036, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10608132183551788, + 0.09577653557062149, + 0.06004346162080765, + 0.12337897717952728, + 0.17955821752548218, + 0.27578237652778625, + 0.06556447595357895, + 0.021234041079878807, + 0.022444484755396843, + 0.05013606697320938 + ], + "credit_map": [ + 0.010041958652436733, + 0.0235991720110178, + 0.014042532071471214, + 0.006224498618394136, + 0.013280307874083519, + 0.01805446669459343, + 0.006825270596891642, + 0.014013111591339111, + 0.006505143828690052, + 0.014194900169968605, + 0.01034961361438036, + 0.004941319115459919, + 0.010195814073085785, + 0.022404378280043602, + 0.008442259393632412, + 0.018743105232715607, + 0.003591489978134632, + 0.00445178896188736, + 0.007467884104698896, + 0.002109694527462125, + 0.003490090137347579, + 0.0167793370783329, + 0.004142433870583773, + 0.01801074482500553, + 0.004028061404824257, + 0.005265130195766687, + 0.0058147902600467205, + 0.004941131453961134, + 0.007267784792929888, + 0.044713862240314484, + 0.00841844454407692, + 0.04292977973818779, + 0.014278330840170383, + 0.013870809227228165, + 0.015241296961903572, + 0.02165384031832218, + 0.028475170955061913, + 0.017622068524360657, + 0.04314829409122467, + 0.025268403813242912, + 0.00703280046582222, + 0.006754279136657715, + 0.010128196328878403, + 0.00660994928330183, + 0.01489583682268858, + 0.11531718075275421, + 0.004177003633230925, + 0.11086712032556534, + 0.004190847277641296, + 0.006915288977324963, + 0.008301278576254845, + 0.006544756703078747, + 0.015667054802179337, + 0.010424220934510231, + 0.006615857128053904, + 0.006905169691890478, + 0.002699644770473242, + 0.0020352983847260475, + 0.0029085148125886917, + 0.002188813639804721, + 0.002389252884313464, + 0.002692392561584711, + 0.003980718553066254, + 0.0023394052404910326, + 0.0022389015648514032, + 0.0022917699534446, + 0.0036085208412259817, + 0.002864826936274767, + 0.0026287934742867947, + 0.0032150561455637217, + 0.0038316063582897186, + 0.001765010179951787, + 0.0031721361447125673, + 0.0055120112374424934, + 0.010202862322330475, + 0.010608864016830921, + 0.0028811213560402393, + 0.008284174837172031, + 0.0042418004013597965, + 0.0052330950275063515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 3 + ], + "selected_reward": -0.04867635294795036 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642413653433323 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642413653433323, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642413653433323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1092640608549118, + 0.020383479073643684, + 0.04609393700957298, + 0.13612133264541626, + 0.07606878876686096, + 0.3850700259208679, + 0.08366860449314117, + 0.06438564509153366, + 0.027750197798013687, + 0.05119385942816734 + ], + "credit_map": [ + 0.009954620152711868, + 0.026070214807987213, + 0.014630316756665707, + 0.006627348251640797, + 0.016262641176581383, + 0.015408572740852833, + 0.007863064296543598, + 0.012447288259863853, + 0.0019316200632601976, + 0.0039147622883319855, + 0.00323985586874187, + 0.0016934442101046443, + 0.00284918793477118, + 0.0023928231094032526, + 0.002334516029804945, + 0.0020272708497941494, + 0.002039798302575946, + 0.0024618564639240503, + 0.004054239485412836, + 0.0017326701199635863, + 0.0023729680106043816, + 0.01593729481101036, + 0.002097995253279805, + 0.015397117473185062, + 0.0033485752064734697, + 0.00458809407427907, + 0.004518526140600443, + 0.003728491021320224, + 0.006187646649777889, + 0.0545283742249012, + 0.005144433584064245, + 0.05407719314098358, + 0.004770196508616209, + 0.006076121237128973, + 0.013623859733343124, + 0.005767799448221922, + 0.007296422030776739, + 0.015472841449081898, + 0.007166076917201281, + 0.015895478427410126, + 0.00968227069824934, + 0.008502752520143986, + 0.013270623050630093, + 0.009017408825457096, + 0.021071700379252434, + 0.16199879348278046, + 0.006143912672996521, + 0.15538252890110016, + 0.005858339369297028, + 0.007975572720170021, + 0.012169481255114079, + 0.009367182850837708, + 0.019352106377482414, + 0.009210068732500076, + 0.010351492092013359, + 0.00938436109572649, + 0.0059749470092356205, + 0.011440840549767017, + 0.009138863533735275, + 0.005870298016816378, + 0.009280079044401646, + 0.007896704599261284, + 0.007338495459407568, + 0.007445415481925011, + 0.002266604918986559, + 0.0024484568275511265, + 0.0034486742224544287, + 0.0039090546779334545, + 0.003862702753394842, + 0.0033386978320777416, + 0.00503149488940835, + 0.003444512141868472, + 0.004336549434810877, + 0.00597779918462038, + 0.009789485484361649, + 0.009240496903657913, + 0.004441494587808847, + 0.0076331645250320435, + 0.005758344661444426, + 0.00401652418076992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.008642413653433323 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0297976266592741 + ], + "opposite_logprob_tokens": [ + -3.5297975540161133 + ], + "logprob_reward": -0.0297976266592741, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.0297976266592741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11061207950115204, + 0.05620148032903671, + 0.04448447376489639, + 0.3633018136024475, + 0.06826244294643402, + 0.14401787519454956, + 0.07410121709108353, + 0.03168170154094696, + 0.02088562771677971, + 0.08645128458738327 + ], + "credit_map": [ + 0.012370231561362743, + 0.027262581512331963, + 0.015418053604662418, + 0.006239111069589853, + 0.014920162968337536, + 0.01300792582333088, + 0.008179747499525547, + 0.01321427058428526, + 0.003722998546436429, + 0.006407817825675011, + 0.00500362878665328, + 0.003498358651995659, + 0.00777278421446681, + 0.013078362680971622, + 0.0063146199099719524, + 0.010402911342680454, + 0.003815433708950877, + 0.006346722133457661, + 0.008672050200402737, + 0.004143334459513426, + 0.005716576240956783, + 0.005859041120857, + 0.005608887877315283, + 0.004322427324950695, + 0.008803455159068108, + 0.011422715149819851, + 0.012497679330408573, + 0.008111449889838696, + 0.014225063845515251, + 0.1490076184272766, + 0.011135807260870934, + 0.14809802174568176, + 0.004235554952174425, + 0.009807076305150986, + 0.009149244986474514, + 0.006751856300979853, + 0.008606303483247757, + 0.007567824330180883, + 0.011969772167503834, + 0.010174808092415333, + 0.0031750721391290426, + 0.003270229324698448, + 0.0059473165310919285, + 0.0055335997603833675, + 0.0074747777543962, + 0.05861089751124382, + 0.0025207982398569584, + 0.057485181838274, + 0.005791174713522196, + 0.010449660941958427, + 0.009545200504362583, + 0.0090503990650177, + 0.018459338694810867, + 0.008110450580716133, + 0.005561541300266981, + 0.007133448030799627, + 0.0026118364185094833, + 0.003952502738684416, + 0.0038207184989005327, + 0.0040458375588059425, + 0.0034635153133422136, + 0.004113410599529743, + 0.006631846074014902, + 0.003042031079530716, + 0.001962613081559539, + 0.0024058425333350897, + 0.0038590047042816877, + 0.004063761793076992, + 0.0020192209631204605, + 0.002075034426525235, + 0.002090295311063528, + 0.0024098572321236134, + 0.0061308215372264385, + 0.008027235046029091, + 0.0202332753688097, + 0.014949360862374306, + 0.005505315959453583, + 0.013584423810243607, + 0.006996676791459322, + 0.011024178005754948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.0297976266592741 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot is fixed on a vibrant, small, red plastic toy sailboat, lying perfectly still in the glassy, clear water that fills a pristine white ceramic bathtub. Soft, warm overhead lighting illuminates the scene, creating gentle highlights on the sleek hull and casting subtle, shimmering reflections on the water, creating an atmosphere of quiet concentration. A silent ripple begins to move across the water, its concentric rings methodically approaching the toy. As the wave reaches the boat, it deliberately lifts one side, causing it to sway precariously. The boat's simple, white plastic mast begins to slowly and inexorably tilt toward the water, triggering a capsize. In one swift, graceful motion, the boat flips completely, mast and sail disappearing beneath the surface as it turns. At the end of the motion, the toy floats motionless, completely upside down, its bright red hull now face-up, gently rocking in the dissipating aftermath.", + "seed": 19190, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19190_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.017080309754237532, + "vlm_logprob": -0.017080309754237532, + "vlm_yes_no_margin": 4.65, + "vlm_token_credit": -0.017080309754237532, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008738137781620026 + ], + "opposite_logprob_tokens": [ + -5.508738040924072 + ], + "logprob_reward": -0.008738137781620026, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008738137781620026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23303578794002533, + 0.07126487791538239, + 0.07928021252155304, + 0.09085434675216675, + 0.044142406433820724, + 0.046507976949214935, + 0.07701064646244049, + 0.038879524916410446, + 0.25317248702049255, + 0.06585173308849335 + ], + "credit_map": [ + 0.02913709357380867, + 0.030631650239229202, + 0.02301100455224514, + 0.02222292684018612, + 0.03885805234313011, + 0.035306256264448166, + 0.025792328640818596, + 0.028076477348804474, + 0.007133376784622669, + 0.010591501370072365, + 0.006052518729120493, + 0.005155561957508326, + 0.014510718174278736, + 0.011584615334868431, + 0.0071852728724479675, + 0.009051315486431122, + 0.009226474910974503, + 0.01132966112345457, + 0.006179408635944128, + 0.0065501779317855835, + 0.014611116610467434, + 0.014777972362935543, + 0.007213549222797155, + 0.00939185917377472, + 0.012191160582005978, + 0.013446350581943989, + 0.013204338029026985, + 0.006091307383030653, + 0.009610739536583424, + 0.023606877774000168, + 0.006846715696156025, + 0.005856869742274284, + 0.004064488224685192, + 0.006322171539068222, + 0.006893223151564598, + 0.003428684314712882, + 0.008861234411597252, + 0.006887458730489016, + 0.0035711571108549833, + 0.004113992676138878, + 0.0049930051900446415, + 0.008279417641460896, + 0.00565970316529274, + 0.005688174162060022, + 0.005411694757640362, + 0.0056444620713591576, + 0.0053588286973536015, + 0.005472694989293814, + 0.012355237267911434, + 0.008359402418136597, + 0.010490489192306995, + 0.009816568344831467, + 0.008990535512566566, + 0.01055016741156578, + 0.008210333995521069, + 0.008237916976213455, + 0.004860479384660721, + 0.003383220871910453, + 0.007491180673241615, + 0.0059320200234651566, + 0.0053411382250487804, + 0.005316609516739845, + 0.0032789651304483414, + 0.003275912255048752, + 0.02357851155102253, + 0.03655267879366875, + 0.030933015048503876, + 0.029219001531600952, + 0.04305769503116608, + 0.0399213507771492, + 0.026688119396567345, + 0.02322212979197502, + 0.008399547077715397, + 0.012202025391161442, + 0.004716199357062578, + 0.006542851682752371, + 0.008524573408067226, + 0.013614688999950886, + 0.006174602545797825, + 0.005677253007888794 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.008738137781620026 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954147316515446 + ], + "opposite_logprob_tokens": [ + -5.130954265594482 + ], + "logprob_reward": -0.005954147316515446, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954147316515446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1748095005750656, + 0.04037576541304588, + 0.043622661381959915, + 0.04761580750346184, + 0.045209307223558426, + 0.038817163556814194, + 0.06682660430669785, + 0.16379417479038239, + 0.32264500856399536, + 0.05628398060798645 + ], + "credit_map": [ + 0.02871382236480713, + 0.024824652820825577, + 0.016746485605835915, + 0.01306785736232996, + 0.021899374201893806, + 0.03354049474000931, + 0.011233244091272354, + 0.02478357031941414, + 0.0053167883306741714, + 0.005441660061478615, + 0.0034672259353101254, + 0.0029175863601267338, + 0.004865099675953388, + 0.00863045547157526, + 0.002921380801126361, + 0.006815568543970585, + 0.005564291495829821, + 0.005886246915906668, + 0.0050442893989384174, + 0.004180539399385452, + 0.007660869043320417, + 0.006856425199657679, + 0.004323168657720089, + 0.00410682987421751, + 0.0066107772290706635, + 0.006086559034883976, + 0.008496686816215515, + 0.0028292250353842974, + 0.006415471434593201, + 0.00800368282943964, + 0.0027602759655565023, + 0.00641312962397933, + 0.004492395557463169, + 0.007256085053086281, + 0.00506176520138979, + 0.002243617083877325, + 0.004337876569479704, + 0.01597702130675316, + 0.0018057299312204123, + 0.0040348125621676445, + 0.004799635149538517, + 0.006163426674902439, + 0.004493411164730787, + 0.0035945188719779253, + 0.00411957036703825, + 0.004239584784954786, + 0.0042736949399113655, + 0.007133320905268192, + 0.012972847558557987, + 0.008936391212046146, + 0.01074723619967699, + 0.005615896545350552, + 0.004271089099347591, + 0.00993175059556961, + 0.006991616450250149, + 0.007359769195318222, + 0.031873539090156555, + 0.024275651201605797, + 0.02129935659468174, + 0.013059761375188828, + 0.017738541588187218, + 0.021681802347302437, + 0.012831850908696651, + 0.021033676341176033, + 0.03368423134088516, + 0.08136855810880661, + 0.029892457649111748, + 0.023979559540748596, + 0.03082740120589733, + 0.08275499194860458, + 0.021346958354115486, + 0.01879085786640644, + 0.007834146730601788, + 0.006901264190673828, + 0.0057147229090332985, + 0.006550522055476904, + 0.00743058929219842, + 0.006972592789679766, + 0.006818688940256834, + 0.008061454631388187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.005954147316515446 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08011104166507721, + 0.059382300823926926, + 0.03990429639816284, + 0.07717335969209671, + 0.23324719071388245, + 0.13519340753555298, + 0.10402930527925491, + 0.06235387176275253, + 0.12764964997768402, + 0.08095552772283554 + ], + "credit_map": [ + 0.01177208498120308, + 0.011885428801178932, + 0.010154505260288715, + 0.007931185886263847, + 0.011085686273872852, + 0.012175733223557472, + 0.006846525706350803, + 0.00825989805161953, + 0.008429407142102718, + 0.009352128021419048, + 0.006484509911388159, + 0.004293850157409906, + 0.006868093740195036, + 0.016158903017640114, + 0.003389781340956688, + 0.004405628424137831, + 0.005066197365522385, + 0.005818834528326988, + 0.004341160878539085, + 0.002740540076047182, + 0.007429849356412888, + 0.006622257176786661, + 0.0027070259675383568, + 0.00517843384295702, + 0.010391632094979286, + 0.008453195914626122, + 0.012997503392398357, + 0.005204763263463974, + 0.011148620396852493, + 0.014038057066500187, + 0.007402136456221342, + 0.007537461817264557, + 0.023432893678545952, + 0.03907180204987526, + 0.026172932237386703, + 0.009915727190673351, + 0.023395666852593422, + 0.0863848403096199, + 0.007375697605311871, + 0.017497655004262924, + 0.018568944185972214, + 0.028202194720506668, + 0.01299126073718071, + 0.009053466841578484, + 0.014960214495658875, + 0.01618596911430359, + 0.009951776824891567, + 0.025279590860009193, + 0.015676692128181458, + 0.011210892349481583, + 0.010618417523801327, + 0.011374634690582752, + 0.01619325578212738, + 0.014026155695319176, + 0.014697565697133541, + 0.010231697000563145, + 0.00878407433629036, + 0.00862898025661707, + 0.010065324604511261, + 0.00557193486019969, + 0.008863173425197601, + 0.008163785561919212, + 0.004792473278939724, + 0.007484128698706627, + 0.011358599178493023, + 0.030015453696250916, + 0.014505504630506039, + 0.007901359349489212, + 0.013647863641381264, + 0.0326337106525898, + 0.008481806144118309, + 0.009105360135436058, + 0.011497754603624344, + 0.00933449249714613, + 0.00589510565623641, + 0.01106901466846466, + 0.013042612001299858, + 0.0108340410515666, + 0.01013928558677435, + 0.009143227711319923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 8 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028183048125356436 + ], + "opposite_logprob_tokens": [ + -5.8778181076049805 + ], + "logprob_reward": -0.0028183048125356436, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028183048125356436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11435770243406296, + 0.047480519860982895, + 0.0409722775220871, + 0.09908076375722885, + 0.11042530089616776, + 0.07999015599489212, + 0.1726742684841156, + 0.07213706523180008, + 0.1917218416929245, + 0.07116010785102844 + ], + "credit_map": [ + 0.013008872047066689, + 0.014230205677449703, + 0.014043567702174187, + 0.010519802570343018, + 0.014704725705087185, + 0.02347467467188835, + 0.011594580486416817, + 0.012781260535120964, + 0.00572784710675478, + 0.007418286986649036, + 0.005472387187182903, + 0.0034412634558975697, + 0.005721032153815031, + 0.011302722617983818, + 0.003573969705030322, + 0.004823011811822653, + 0.00436557037755847, + 0.006072954274713993, + 0.003792222822085023, + 0.0030149072408676147, + 0.006184516940265894, + 0.011006293818354607, + 0.002924627158790827, + 0.003611188381910324, + 0.010498402640223503, + 0.01155893225222826, + 0.018803348764777184, + 0.006718384101986885, + 0.012548703700304031, + 0.024687370285391808, + 0.007221937645226717, + 0.007043680641800165, + 0.010707775130867958, + 0.021270405501127243, + 0.015581604093313217, + 0.004961664788424969, + 0.0092414366081357, + 0.03750885650515556, + 0.004632346797734499, + 0.006521207746118307, + 0.012709956616163254, + 0.011142579838633537, + 0.011855925433337688, + 0.007707458455115557, + 0.007378117181360722, + 0.009201915003359318, + 0.0051962146535515785, + 0.01479798648506403, + 0.025009188801050186, + 0.021736755967140198, + 0.018920516595244408, + 0.018748929724097252, + 0.024034036323428154, + 0.022924436256289482, + 0.022060152143239975, + 0.01924026384949684, + 0.01279413141310215, + 0.011330146342515945, + 0.01033114269375801, + 0.005814812611788511, + 0.010136152617633343, + 0.009202562272548676, + 0.005508922040462494, + 0.00701919523999095, + 0.014929800294339657, + 0.044233184307813644, + 0.022035157307982445, + 0.01555741298943758, + 0.024070292711257935, + 0.04278179630637169, + 0.014521004632115364, + 0.013593189418315887, + 0.006228218320757151, + 0.017699314281344414, + 0.003819379024207592, + 0.005922403652220964, + 0.009600667282938957, + 0.01676306501030922, + 0.005720735993236303, + 0.005406325217336416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.0028183048125356436 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038087546825408936 + ], + "opposite_logprob_tokens": [ + -3.2880876064300537 + ], + "logprob_reward": -0.038087546825408936, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038087546825408936, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09524384886026382, + 0.04480797424912453, + 0.06283654272556305, + 0.1286868304014206, + 0.023548297584056854, + 0.0485917367041111, + 0.09255386143922806, + 0.07950926572084427, + 0.2843247652053833, + 0.13989683985710144 + ], + "credit_map": [ + 0.014426296576857567, + 0.013543947599828243, + 0.010385354980826378, + 0.0065986220724880695, + 0.01599176786839962, + 0.017162444069981575, + 0.007546406239271164, + 0.009589024819433689, + 0.00733961770310998, + 0.005936440080404282, + 0.006591942626982927, + 0.002256362233310938, + 0.0059104664251208305, + 0.010968904010951519, + 0.0021142931655049324, + 0.003689949167892337, + 0.008239582180976868, + 0.008938878774642944, + 0.007790624629706144, + 0.004884733352810144, + 0.007113368716090918, + 0.017348535358905792, + 0.0045203547924757, + 0.004000470042228699, + 0.010573836974799633, + 0.01603342406451702, + 0.018103305250406265, + 0.0064373621717095375, + 0.013817775063216686, + 0.04218961298465729, + 0.007301376201212406, + 0.014230149798095226, + 0.0030395593494176865, + 0.0026365441735833883, + 0.003141261637210846, + 0.0019182435935363173, + 0.00310137658379972, + 0.005619969684630632, + 0.0013991103041917086, + 0.0026922342367470264, + 0.006162981502711773, + 0.005279760807752609, + 0.007498602848500013, + 0.004281895700842142, + 0.005138931330293417, + 0.0061934650875627995, + 0.004789097234606743, + 0.009247004985809326, + 0.017721595242619514, + 0.013752228580415249, + 0.01564640738070011, + 0.00842317845672369, + 0.006233114283531904, + 0.013469532132148743, + 0.008113349787890911, + 0.009194458834826946, + 0.012495039030909538, + 0.010816090740263462, + 0.011328422464430332, + 0.008706762455403805, + 0.009339172393083572, + 0.009988230653107166, + 0.008044889196753502, + 0.008790655992925167, + 0.01879142038524151, + 0.10427835583686829, + 0.009474040009081364, + 0.009428986348211765, + 0.01695411093533039, + 0.10282427817583084, + 0.009414786472916603, + 0.013158818706870079, + 0.010899404995143414, + 0.037211839109659195, + 0.00696605583652854, + 0.009337354451417923, + 0.015850424766540527, + 0.04203183203935623, + 0.00938283558934927, + 0.0082171019166708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 3 + ], + "selected_reward": -0.038087546825408936 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot is fixed on a vibrant, small, red plastic toy sailboat, lying perfectly still in the glassy, clear water that fills a pristine white ceramic bathtub. Soft, warm overhead lighting illuminates the scene, creating gentle highlights on the sleek hull and casting subtle, shimmering reflections on the water, creating an atmosphere of quiet concentration. A silent ripple begins to move across the water, its concentric rings methodically approaching the toy. As the wave reaches the boat, it deliberately lifts one side, causing it to sway precariously. The boat's simple, white plastic mast begins to slowly and inexorably tilt toward the water, triggering a capsize. In one swift, graceful motion, the boat flips completely, mast and sail disappearing beneath the surface as it turns. At the end of the motion, the toy floats motionless, completely upside down, its bright red hull now face-up, gently rocking in the dissipating aftermath.", + "seed": 19191, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19191_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.017013972392305732, + "vlm_logprob": -0.017013972392305732, + "vlm_yes_no_margin": 4.425, + "vlm_token_credit": -0.017013972392305732, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.030347559601068497 + ], + "opposite_logprob_tokens": [ + -3.9053475856781006 + ], + "logprob_reward": -0.030347559601068497, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.030347559601068497, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08764971792697906, + 0.1348274201154709, + 0.0771825984120369, + 0.12358979880809784, + 0.027671897783875465, + 0.033847466111183167, + 0.3659120798110962, + 0.10029882192611694, + 0.03560152277350426, + 0.013418730348348618 + ], + "credit_map": [ + 0.008127979934215546, + 0.015042897313833237, + 0.013966887257993221, + 0.008014913648366928, + 0.012442473322153091, + 0.009949637576937675, + 0.01114536914974451, + 0.008959557861089706, + 0.010894950479269028, + 0.027206160128116608, + 0.016605105251073837, + 0.012392938137054443, + 0.017119010910391808, + 0.013316511176526546, + 0.024205153807997704, + 0.013087580911815166, + 0.007507971487939358, + 0.0126531096175313, + 0.015120182186365128, + 0.006990053225308657, + 0.01110875979065895, + 0.007538086734712124, + 0.008723062463104725, + 0.007541363593190908, + 0.005716551095247269, + 0.005052156280726194, + 0.005758317653089762, + 0.0020755392033606768, + 0.011597485281527042, + 0.051858313381671906, + 0.002617516089230776, + 0.03891391307115555, + 0.0033130773808807135, + 0.002728072926402092, + 0.004850741941481829, + 0.002825844567269087, + 0.0030026191379874945, + 0.004114649724215269, + 0.0028002518229186535, + 0.0040366388857364655, + 0.0033679120242595673, + 0.003940097521990538, + 0.0033795468043535948, + 0.0033633760176599026, + 0.004968356341123581, + 0.005624805111438036, + 0.0027749082073569298, + 0.00642846105620265, + 0.024536598473787308, + 0.037817589938640594, + 0.12303493916988373, + 0.026205360889434814, + 0.018083684146404266, + 0.09955964982509613, + 0.014967971481382847, + 0.02170622907578945, + 0.008101466111838818, + 0.028358494862914085, + 0.012348011136054993, + 0.007684047799557447, + 0.008056610822677612, + 0.019086675718426704, + 0.009140384383499622, + 0.007523121312260628, + 0.003551011672243476, + 0.012744421139359474, + 0.008456461131572723, + 0.0015763469273224473, + 0.004154651425778866, + 0.0019183659460395575, + 0.0018827586900442839, + 0.00131750269792974, + 0.0017091312911361456, + 0.0020444439724087715, + 0.001456016325391829, + 0.0011776124592870474, + 0.0017734174616634846, + 0.00146548415068537, + 0.0021359985694289207, + 0.0016566249541938305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 3 + ], + "selected_reward": -0.030347559601068497 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076459851115942 + ], + "opposite_logprob_tokens": [ + -4.882646083831787 + ], + "logprob_reward": -0.0076459851115942, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076459851115942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09555167704820633, + 0.040935218334198, + 0.053697653114795685, + 0.12908989191055298, + 0.06848521530628204, + 0.07316123694181442, + 0.1412409394979477, + 0.0718449130654335, + 0.28342339396476746, + 0.042569849640131 + ], + "credit_map": [ + 0.011241086758673191, + 0.011155031621456146, + 0.01112488191574812, + 0.007114850915968418, + 0.01263098418712616, + 0.01606210693717003, + 0.00960707850754261, + 0.01661565713584423, + 0.006200648844242096, + 0.004642021376639605, + 0.00403337087482214, + 0.00321779097430408, + 0.005828627850860357, + 0.006957806646823883, + 0.004218817688524723, + 0.005836129188537598, + 0.004922649823129177, + 0.0064938366413116455, + 0.01242148783057928, + 0.0031471436377614737, + 0.006076208781450987, + 0.006970871239900589, + 0.005448745097965002, + 0.008216713555157185, + 0.014737285673618317, + 0.008656777441501617, + 0.010944237932562828, + 0.005253961309790611, + 0.009864449501037598, + 0.039813827723264694, + 0.006336106453090906, + 0.03348323702812195, + 0.009669954888522625, + 0.008143877610564232, + 0.011202874593436718, + 0.005418608896434307, + 0.007500042207539082, + 0.011208686046302319, + 0.007327491417527199, + 0.008013684302568436, + 0.0036231703124940395, + 0.00474324868991971, + 0.005865339655429125, + 0.0022893876302987337, + 0.00645019393414259, + 0.02401544153690338, + 0.002978845965117216, + 0.023195603862404823, + 0.010689557529985905, + 0.01602114364504814, + 0.04364745318889618, + 0.010308003053069115, + 0.007447392214089632, + 0.036286935210227966, + 0.006558465305715799, + 0.010281983762979507, + 0.0045729405246675014, + 0.024285152554512024, + 0.005792031064629555, + 0.003213218878954649, + 0.004776779096573591, + 0.021426314488053322, + 0.004045284818857908, + 0.0037331923376768827, + 0.027906926348805428, + 0.11528731882572174, + 0.039668191224336624, + 0.00676565058529377, + 0.02716219052672386, + 0.05094796419143677, + 0.005562033038586378, + 0.010123122483491898, + 0.0044812364503741264, + 0.00942514929920435, + 0.008741350844502449, + 0.0033895382657647133, + 0.0048347883857786655, + 0.0036915976088494062, + 0.002927142893895507, + 0.0050790440291166306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 3 + ], + "selected_reward": -0.0076459851115942 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020595736801624298 + ], + "opposite_logprob_tokens": [ + -3.8955957889556885 + ], + "logprob_reward": -0.020595736801624298, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020595736801624298, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056503813713788986, + 0.01764703169465065, + 0.019609306007623672, + 0.036617182195186615, + 0.037180639803409576, + 0.07118717581033707, + 0.26014360785484314, + 0.10853907465934753, + 0.3788703978061676, + 0.01370181143283844 + ], + "credit_map": [ + 0.005820937920361757, + 0.0064684427343308926, + 0.00878747459501028, + 0.005473184399306774, + 0.006392254494130611, + 0.006569500081241131, + 0.009509057737886906, + 0.007482962682843208, + 0.00240086717531085, + 0.0021152545232325792, + 0.0019322545267641544, + 0.001558836316689849, + 0.0023468604777008295, + 0.002555284183472395, + 0.0024538857396692038, + 0.0022837899159640074, + 0.002371805487200618, + 0.0021463450975716114, + 0.00453057000413537, + 0.0010269720805808902, + 0.0020284969359636307, + 0.0027034920640289783, + 0.0017211855156347156, + 0.0030804392881691456, + 0.0036528646014630795, + 0.008430742658674717, + 0.0036240159533917904, + 0.003282388439401984, + 0.005334316287189722, + 0.003981652669608593, + 0.0035396041348576546, + 0.004771596752107143, + 0.0047737364657223225, + 0.005714464001357555, + 0.005629623308777809, + 0.0035729012452065945, + 0.0034006524365395308, + 0.004998547490686178, + 0.003945276141166687, + 0.005145437549799681, + 0.0037576761096715927, + 0.005282960832118988, + 0.005461608059704304, + 0.007790383882820606, + 0.006244782358407974, + 0.020351259037852287, + 0.004349644761532545, + 0.01794886775314808, + 0.017007015645503998, + 0.022116050124168396, + 0.08467584103345871, + 0.014033147133886814, + 0.014932477846741676, + 0.07883783429861069, + 0.011155839078128338, + 0.017385397106409073, + 0.008111433126032352, + 0.030042147263884544, + 0.014814168214797974, + 0.006132849026471376, + 0.0077340686693787575, + 0.02378026582300663, + 0.006482491735368967, + 0.011441655457019806, + 0.033719975501298904, + 0.14187227189540863, + 0.0429924875497818, + 0.012729158625006676, + 0.030706360936164856, + 0.08900444209575653, + 0.010906628333032131, + 0.016939062625169754, + 0.0013288308400660753, + 0.0028357014525681734, + 0.0018691755831241608, + 0.0009621861390769482, + 0.001729472423903644, + 0.0016351764788851142, + 0.0012489057844504714, + 0.0020923633128404617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 7 + ], + "selected_reward": -0.020595736801624298 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031953013967722654 + ], + "opposite_logprob_tokens": [ + -5.753195285797119 + ], + "logprob_reward": -0.0031953013967722654, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031953013967722654, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07089143246412277, + 0.021918263286352158, + 0.017046162858605385, + 0.08228301256895065, + 0.04696999490261078, + 0.06447246670722961, + 0.3174789547920227, + 0.1440475434064865, + 0.2079596072435379, + 0.026932546868920326 + ], + "credit_map": [ + 0.007439759559929371, + 0.006330259144306183, + 0.011187640950083733, + 0.006088377442210913, + 0.009077562019228935, + 0.009688271209597588, + 0.01054003182798624, + 0.01053953543305397, + 0.0021449963096529245, + 0.0034537329338490963, + 0.0023681630846112967, + 0.0013481974601745605, + 0.0023918093647807837, + 0.0036290381103754044, + 0.0029323999769985676, + 0.0036499258130788803, + 0.0017077510710805655, + 0.0019432940753176808, + 0.0034210733138024807, + 0.001009689294733107, + 0.002008101437240839, + 0.0018564420752227306, + 0.0020499008242040873, + 0.0030499130953103304, + 0.004167433362454176, + 0.007858292199671268, + 0.005165449343621731, + 0.002616954268887639, + 0.00808151438832283, + 0.028262341395020485, + 0.0030960626900196075, + 0.023034969344735146, + 0.0059874411672353745, + 0.004950983915477991, + 0.009682773612439632, + 0.0038957770448178053, + 0.004608994349837303, + 0.006754482164978981, + 0.005165934097021818, + 0.0059236097149550915, + 0.0030061963479965925, + 0.0029626241885125637, + 0.005090869963169098, + 0.0038840591441839933, + 0.0045939357951283455, + 0.022449469193816185, + 0.0023291436955332756, + 0.020156171172857285, + 0.021169112995266914, + 0.032920949161052704, + 0.1073400005698204, + 0.020631643012166023, + 0.015981748700141907, + 0.086344413459301, + 0.013655324466526508, + 0.01943575218319893, + 0.009813104756176472, + 0.048098575323820114, + 0.01581035926938057, + 0.007298734970390797, + 0.00980584230273962, + 0.03571091219782829, + 0.008048610761761665, + 0.009461416862905025, + 0.018784217536449432, + 0.08105681091547012, + 0.023628616705536842, + 0.0057569099590182304, + 0.016549386084079742, + 0.04924577474594116, + 0.005085031036287546, + 0.007852871902287006, + 0.0018838575342670083, + 0.010002273134887218, + 0.0033657990861684084, + 0.0012326763244345784, + 0.0027120031882077456, + 0.004959987476468086, + 0.0014578172704204917, + 0.0013181358808651567 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.0031953013967722654 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232852790504694 + ], + "opposite_logprob_tokens": [ + -3.773285388946533 + ], + "logprob_reward": -0.0232852790504694, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232852790504694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05467945337295532, + 0.027161935344338417, + 0.018697815015912056, + 0.15219378471374512, + 0.043574996292591095, + 0.07381552457809448, + 0.49208036065101624, + 0.058476172387599945, + 0.06114614009857178, + 0.018173858523368835 + ], + "credit_map": [ + 0.006440804805606604, + 0.005947403609752655, + 0.009294590912759304, + 0.004782584961503744, + 0.006553939543664455, + 0.006114436779171228, + 0.007771302945911884, + 0.007774386089295149, + 0.002690665191039443, + 0.004501338582485914, + 0.004169785417616367, + 0.0018504696199670434, + 0.003633193438872695, + 0.003247829619795084, + 0.004084161948412657, + 0.002984491642564535, + 0.0017542167333886027, + 0.0020925370045006275, + 0.0043371329084038734, + 0.0012753002811223269, + 0.0020176146645098925, + 0.002311031799763441, + 0.001427551032975316, + 0.0034824293106794357, + 0.007789668161422014, + 0.012509320862591267, + 0.007553594186902046, + 0.003906155936419964, + 0.015242132358253002, + 0.057420916855335236, + 0.0037920773029327393, + 0.04397990554571152, + 0.005738755222409964, + 0.006825683172792196, + 0.006649186369031668, + 0.0035937486682087183, + 0.004198993090540171, + 0.006876736879348755, + 0.003596905618906021, + 0.006094984710216522, + 0.00368033186532557, + 0.005313978996127844, + 0.00512702576816082, + 0.004824083298444748, + 0.006975811906158924, + 0.023397546261548996, + 0.003790303599089384, + 0.02070644125342369, + 0.03213461861014366, + 0.0498049333691597, + 0.17056137323379517, + 0.029173484072089195, + 0.023441363126039505, + 0.13953888416290283, + 0.018779076635837555, + 0.028646577149629593, + 0.004749678540974855, + 0.014903942123055458, + 0.009192530065774918, + 0.0035481383092701435, + 0.00470776529982686, + 0.010765411891043186, + 0.003913247492164373, + 0.006695450749248266, + 0.006149131339043379, + 0.010685769841074944, + 0.009356176480650902, + 0.004731900990009308, + 0.008073429577052593, + 0.011603409424424171, + 0.003937589470297098, + 0.006608732510358095, + 0.001518737175501883, + 0.005631729960441589, + 0.0018697044579312205, + 0.0008287096861749887, + 0.001476661884225905, + 0.004500750917941332, + 0.0009814051445573568, + 0.0013661585981026292 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 5 + ], + "selected_reward": -0.0232852790504694 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot is fixed on a vibrant, small, red plastic toy sailboat, lying perfectly still in the glassy, clear water that fills a pristine white ceramic bathtub. Soft, warm overhead lighting illuminates the scene, creating gentle highlights on the sleek hull and casting subtle, shimmering reflections on the water, creating an atmosphere of quiet concentration. A silent ripple begins to move across the water, its concentric rings methodically approaching the toy. As the wave reaches the boat, it deliberately lifts one side, causing it to sway precariously. The boat's simple, white plastic mast begins to slowly and inexorably tilt toward the water, triggering a capsize. In one swift, graceful motion, the boat flips completely, mast and sail disappearing beneath the surface as it turns. At the end of the motion, the toy floats motionless, completely upside down, its bright red hull now face-up, gently rocking in the dissipating aftermath.", + "seed": 19192, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_2/index_1248_seed_19192_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020401790458709, + "vlm_logprob": -0.020401790458709, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.020401790458709, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01870925910770893 + ], + "opposite_logprob_tokens": [ + -4.518709182739258 + ], + "logprob_reward": -0.01870925910770893, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01870925910770893, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11777746677398682, + 0.04202539101243019, + 0.14670521020889282, + 0.021226422861218452, + 0.016898643225431442, + 0.01892203465104103, + 0.01810440979897976, + 0.40609222650527954, + 0.08867679536342621, + 0.12357143312692642 + ], + "credit_map": [ + 0.012019087560474873, + 0.01805609092116356, + 0.017045820131897926, + 0.012679077684879303, + 0.016555262729525566, + 0.013673159293830395, + 0.013713774271309376, + 0.014035198837518692, + 0.004564890172332525, + 0.005997522734105587, + 0.006342853419482708, + 0.003046479308977723, + 0.004786103498190641, + 0.005492337513715029, + 0.003760149935260415, + 0.008035054430365562, + 0.00990941934287548, + 0.010193743743002415, + 0.05242573842406273, + 0.005468130577355623, + 0.0072194612585008144, + 0.05216454342007637, + 0.004038282670080662, + 0.005285885650664568, + 0.003274547401815653, + 0.002668156521394849, + 0.003429894335567951, + 0.0013779089786112309, + 0.0015977838775143027, + 0.002957301214337349, + 0.0017299304017797112, + 0.004190900828689337, + 0.0019970370922237635, + 0.0028891172260046005, + 0.002853906014934182, + 0.0014133488293737173, + 0.0014708336675539613, + 0.00182616978418082, + 0.0016232794150710106, + 0.0028249516617506742, + 0.0022155423648655415, + 0.003006807528436184, + 0.0022879126481711864, + 0.0016261492855846882, + 0.002390707843005657, + 0.0024176877923309803, + 0.001664479961618781, + 0.0033127455972135067, + 0.002199755981564522, + 0.0013719421112909913, + 0.0028195723425596952, + 0.0017338639590889215, + 0.0025638160295784473, + 0.0025319859851151705, + 0.0017586118774488568, + 0.0031248629093170166, + 0.059353966265916824, + 0.038843706250190735, + 0.05256574600934982, + 0.035613298416137695, + 0.038840144872665405, + 0.09286634624004364, + 0.04127192497253418, + 0.04673708975315094, + 0.010836795903742313, + 0.015175714157521725, + 0.01079030055552721, + 0.015315770171582699, + 0.008036581799387932, + 0.009626444429159164, + 0.008084466680884361, + 0.010810725390911102, + 0.013201156631112099, + 0.01485889870673418, + 0.02030538208782673, + 0.011416886001825333, + 0.012866110540926456, + 0.021656040102243423, + 0.013675954192876816, + 0.015591016039252281 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.01870925910770893 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004651203285902739 + ], + "opposite_logprob_tokens": [ + -5.379651069641113 + ], + "logprob_reward": -0.004651203285902739, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004651203285902739, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08582950383424759, + 0.02332991175353527, + 0.16031776368618011, + 0.010699110105633736, + 0.01672600954771042, + 0.030719006434082985, + 0.04791752249002457, + 0.053438715636730194, + 0.3935038149356842, + 0.17751862108707428 + ], + "credit_map": [ + 0.010442698374390602, + 0.013567614369094372, + 0.012923890724778175, + 0.008309894241392612, + 0.009825770743191242, + 0.009884858503937721, + 0.01001500990241766, + 0.010859767906367779, + 0.002498987130820751, + 0.003190547227859497, + 0.003705258946865797, + 0.002286248840391636, + 0.00276395445689559, + 0.002701649907976389, + 0.002721769269555807, + 0.0034614959731698036, + 0.009151432663202286, + 0.008100127801299095, + 0.06143198907375336, + 0.004174019210040569, + 0.006474494934082031, + 0.06212461739778519, + 0.002852580975741148, + 0.006008508615195751, + 0.0013098771451041102, + 0.001302790828049183, + 0.0020198726560920477, + 0.0006680316873826087, + 0.001187272951938212, + 0.0014403632376343012, + 0.0008168112835846841, + 0.0019540905486792326, + 0.0018942035967484117, + 0.00205906480550766, + 0.002632791642099619, + 0.0009745083516463637, + 0.0012969160452485085, + 0.0025813551619648933, + 0.0015149483224377036, + 0.0037722233682870865, + 0.00342297344468534, + 0.005191939417272806, + 0.004161819349974394, + 0.0018596681766211987, + 0.00315251387655735, + 0.0037210085429251194, + 0.0022424091584980488, + 0.00696667330339551, + 0.005334582645446062, + 0.004480956122279167, + 0.007115721236914396, + 0.0050589898601174355, + 0.007871252484619617, + 0.005963412579149008, + 0.00427665701135993, + 0.007815955206751823, + 0.007314328569918871, + 0.006301084067672491, + 0.006045355461537838, + 0.005704315844923258, + 0.006335153244435787, + 0.00830696802586317, + 0.006585895083844662, + 0.006845618598163128, + 0.039391014724969864, + 0.05841531231999397, + 0.05308694392442703, + 0.10631323605775833, + 0.02748570404946804, + 0.0492926761507988, + 0.025513112545013428, + 0.034005824476480484, + 0.026407526805996895, + 0.026686346158385277, + 0.03728426620364189, + 0.020020587369799614, + 0.015825355425477028, + 0.017565632238984108, + 0.01921095885336399, + 0.014517958275973797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 2 + ], + "selected_reward": -0.004651203285902739 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016077987849712372 + ], + "opposite_logprob_tokens": [ + -4.141077995300293 + ], + "logprob_reward": -0.016077987849712372, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016077987849712372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06989455223083496, + 0.02273706905543804, + 0.2772415578365326, + 0.017981786280870438, + 0.012951367534697056, + 0.029734507203102112, + 0.06430062651634216, + 0.16290222108364105, + 0.2492627650499344, + 0.0929936096072197 + ], + "credit_map": [ + 0.008383470587432384, + 0.010786298662424088, + 0.011073821224272251, + 0.005665276665240526, + 0.007750838063657284, + 0.010178160853683949, + 0.005977709777653217, + 0.010078974068164825, + 0.0035595057997852564, + 0.0029914602637290955, + 0.004217078443616629, + 0.0015644714003428817, + 0.0017013902543112636, + 0.0028866934590041637, + 0.0022583443205803633, + 0.0035581248812377453, + 0.01744874380528927, + 0.015771690756082535, + 0.10053064674139023, + 0.008375849574804306, + 0.01245447713881731, + 0.10473237931728363, + 0.007621256168931723, + 0.010306507349014282, + 0.0025332316290587187, + 0.0024023842997848988, + 0.00363975390791893, + 0.0011948616011068225, + 0.0015370836481451988, + 0.0022058184258639812, + 0.001361314789392054, + 0.003107337048277259, + 0.0013137077912688255, + 0.0014270319370552897, + 0.002237352542579174, + 0.0005745019880123436, + 0.000770506274420768, + 0.0022547245025634766, + 0.000950367480982095, + 0.0034231748431921005, + 0.0026979674585163593, + 0.007134180516004562, + 0.00501447357237339, + 0.0017821844667196274, + 0.00385101861320436, + 0.003882939461618662, + 0.0018542056204751134, + 0.0035175394732505083, + 0.006234989035874605, + 0.0056525026448071, + 0.008788417093455791, + 0.00969274528324604, + 0.005766549613326788, + 0.00774313323199749, + 0.007366742007434368, + 0.013055549003183842, + 0.022233279421925545, + 0.018661143258213997, + 0.019609207287430763, + 0.013384630903601646, + 0.01864764466881752, + 0.03304107114672661, + 0.0187271349132061, + 0.018598107621073723, + 0.01243889331817627, + 0.08999642729759216, + 0.08194125443696976, + 0.023392945528030396, + 0.0107420077547431, + 0.014999392442405224, + 0.006129671819508076, + 0.009622172452509403, + 0.012612112797796726, + 0.012734011746942997, + 0.019506243988871574, + 0.009217317216098309, + 0.006405428983271122, + 0.017275281250476837, + 0.00797310285270214, + 0.007270114030689001 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 7 + ], + "selected_reward": -0.016077987849712372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762906763702631 + ], + "opposite_logprob_tokens": [ + -4.882628917694092 + ], + "logprob_reward": -0.00762906763702631, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762906763702631, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0772625058889389, + 0.04432150349020958, + 0.3956727385520935, + 0.02614288032054901, + 0.024000613018870354, + 0.027896540239453316, + 0.0509013794362545, + 0.09331769496202469, + 0.1561284065246582, + 0.10435570031404495 + ], + "credit_map": [ + 0.010898824781179428, + 0.010536936111748219, + 0.010194248519837856, + 0.005548102781176567, + 0.007677637506276369, + 0.010393866337835789, + 0.00791837740689516, + 0.014094509184360504, + 0.004905534442514181, + 0.005507640540599823, + 0.006657260004431009, + 0.003938707988709211, + 0.004582415800541639, + 0.005594619549810886, + 0.005273387301713228, + 0.007861936464905739, + 0.02138001285493374, + 0.018802819773554802, + 0.15591928362846375, + 0.009885738603770733, + 0.016271770000457764, + 0.15700331330299377, + 0.005763555411249399, + 0.0106462761759758, + 0.004095356911420822, + 0.004052877426147461, + 0.004314290825277567, + 0.0018624975346028805, + 0.0026287997607141733, + 0.003275931579992175, + 0.0019744383171200752, + 0.00393868749961257, + 0.0028950171545147896, + 0.0029153504874557257, + 0.00375083414837718, + 0.0012363290879875422, + 0.0016265587182715535, + 0.0035546307917684317, + 0.0017596029210835695, + 0.006262290291488171, + 0.00366788636893034, + 0.006195921916514635, + 0.0034290289040654898, + 0.0020461278036236763, + 0.0037741223350167274, + 0.003264643484726548, + 0.002089706715196371, + 0.0034291022457182407, + 0.006509964819997549, + 0.004419743083417416, + 0.005715435370802879, + 0.00695603946223855, + 0.006985324900597334, + 0.007308548782020807, + 0.005034624133259058, + 0.007971694692969322, + 0.012440729886293411, + 0.01093789841979742, + 0.011345591396093369, + 0.00868490431457758, + 0.01048235408961773, + 0.017003966495394707, + 0.010595343075692654, + 0.011826903559267521, + 0.016714442521333694, + 0.027695365250110626, + 0.01953144557774067, + 0.03127695992588997, + 0.016201980412006378, + 0.017300991341471672, + 0.009132197126746178, + 0.018275022506713867, + 0.015504215843975544, + 0.014474104158580303, + 0.02444593608379364, + 0.01442955993115902, + 0.008555975742638111, + 0.011670755222439766, + 0.006913139950484037, + 0.008362006396055222 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 9 + ], + "selected_reward": -0.00762906763702631 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a small red plastic toy sailboat with a white mast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set in a pristine white ceramic bathtub filled with clear water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the video show a silent ripple approaching the boat and causing it to capsize?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the boat end the motion floating upside down with its red hull facing up?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, eye-level medium shot that remains fixed on the boat?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054941434413194656 + ], + "opposite_logprob_tokens": [ + -2.929941415786743 + ], + "logprob_reward": -0.054941434413194656, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054941434413194656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.025496579706668854, + 0.025470446795225143, + 0.7503766417503357, + 0.023401767015457153, + 0.01557779312133789, + 0.009563826955854893, + 0.04286959394812584, + 0.03998005762696266, + 0.046181026846170425, + 0.021082336083054543 + ], + "credit_map": [ + 0.003667172510176897, + 0.0029767965897917747, + 0.0041418238542973995, + 0.002002446446567774, + 0.0023835450410842896, + 0.0034599362406879663, + 0.0021436833776533604, + 0.004721176810562611, + 0.0034961048513650894, + 0.0033881585113704205, + 0.004100058227777481, + 0.0023464749101549387, + 0.0025354200042784214, + 0.0030010377522557974, + 0.0024421196430921555, + 0.004161073360592127, + 0.04125264659523964, + 0.03823985159397125, + 0.2861190140247345, + 0.020400395616889, + 0.03176848962903023, + 0.2956399917602539, + 0.01438050251454115, + 0.022575777024030685, + 0.0024450235068798065, + 0.002803446026518941, + 0.004785117693245411, + 0.0015027704648673534, + 0.002680483739823103, + 0.004150389228016138, + 0.002069835551083088, + 0.0029647003393620253, + 0.0012925132177770138, + 0.002094387076795101, + 0.002186901168897748, + 0.0007330747903324664, + 0.001338196569122374, + 0.00269237975589931, + 0.0010112739400938153, + 0.004229067359119654, + 0.0010090498253703117, + 0.0016348767094314098, + 0.00142922333907336, + 0.0006575743318535388, + 0.001299423398450017, + 0.0009915815899148583, + 0.0009900890290737152, + 0.001552008674480021, + 0.00413385359570384, + 0.0038941868115216494, + 0.006301314570009708, + 0.00479043647646904, + 0.006979337427765131, + 0.005886957049369812, + 0.00397225096821785, + 0.006911261007189751, + 0.005425082985311747, + 0.004581774584949017, + 0.004664188716560602, + 0.0034717984963208437, + 0.005018102470785379, + 0.007457857485860586, + 0.004812648519873619, + 0.0045486073940992355, + 0.005976567044854164, + 0.007734263315796852, + 0.005684821400791407, + 0.008859354071319103, + 0.004747696686536074, + 0.005132604856044054, + 0.004021754022687674, + 0.004023971501737833, + 0.0026775053702294827, + 0.0028523006476461887, + 0.00445216428488493, + 0.0022843361366540194, + 0.002234864514321089, + 0.0028322949074208736, + 0.001817299984395504, + 0.0019315698882564902 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 6 + ], + "selected_reward": -0.054941434413194656 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_tensor([7167, 7167])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_tensor([7167, 7167])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9c1d29e42449f947e050fbfb99b91430a7b5c879 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_tensor([7167, 7167])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium-length shot, rendered in a surreal Unreal Engine style, captures a moment of pure camaraderie between two young men on a bright, sun-drenched city street corner. A young white male with tousled, sun-bleached blond hair and a friendly smile, dressed in a soft white cotton T-shirt and a textured blue denim jacket, faces a young black male with neatly braided hair and a warm, genuine smile, dressed in a heather-gray pullover with an intricate weave. The scene is captured at a crisp, high frame rate, and their synchronized movements unfold with perfect fluidity as they smoothly extend their arms and bump fists together. The connection is firm and solid, a testament to their bond, and their shared smile grows even brighter in the moment of connection. Intense, high-angle sunlight bathes the scene, creating strong rim lighting on their shoulders and hair, and producing subtle, aesthetically pleasing lens flares that drift across the frame. In the background, the gentle bustle of city life—blurred pedestrians and soft-focus vehicles—is rendered with a shallow depth of field, focusing all attention on the genuine, joyful interactions, meticulously depicting the realistic textures of their skin and clothing.", + "seed": 32247, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32247_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.026675822772085668, + "vlm_logprob": -0.026675822772085668, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.026675822772085668, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011507445015013218 + ], + "opposite_logprob_tokens": [ + -4.636507511138916 + ], + "logprob_reward": -0.011507445015013218, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.011507445015013218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12497122585773468, + 0.13993486762046814, + 0.03131571039557457, + 0.05210750550031662, + 0.33188939094543457, + 0.08211494982242584, + 0.0846807211637497, + 0.04285641014575958, + 0.07289206981658936, + 0.03723708540201187 + ], + "credit_map": [ + 0.02080520987510681, + 0.013110755011439323, + 0.021138407289981842, + 0.021123426035046577, + 0.010193735361099243, + 0.018597053363919258, + 0.00946986023336649, + 0.010532776825129986, + 0.009091298095881939, + 0.009525162167847157, + 0.04391627013683319, + 0.010289322584867477, + 0.04369965195655823, + 0.011108978651463985, + 0.007448722142726183, + 0.0048554642125964165, + 0.003974210470914841, + 0.004377160686999559, + 0.005638483911752701, + 0.004170306492596865, + 0.003600118216127157, + 0.003345811041072011, + 0.0030366028659045696, + 0.0031730164773762226, + 0.005113330669701099, + 0.005594638641923666, + 0.00797668844461441, + 0.007081152871251106, + 0.005145486444234848, + 0.008719761855900288, + 0.006807240191847086, + 0.005669207312166691, + 0.023815060034394264, + 0.015363375656306744, + 0.08851016312837601, + 0.13073177635669708, + 0.02506253495812416, + 0.016715899109840393, + 0.016152629628777504, + 0.015537966974079609, + 0.004511678125709295, + 0.007742239627987146, + 0.009820263832807541, + 0.026867864653468132, + 0.005102368537336588, + 0.007219197694212198, + 0.01286669634282589, + 0.007984647527337074, + 0.008557315915822983, + 0.0065355561673641205, + 0.0060500591062009335, + 0.008226274512708187, + 0.01896953582763672, + 0.02622879482805729, + 0.00592059176415205, + 0.00419258838519454, + 0.006191345397382975, + 0.00785838346928358, + 0.003863802645355463, + 0.003815185511484742, + 0.003131094854325056, + 0.004415201488882303, + 0.009761162102222443, + 0.0038202370051294565, + 0.009557519108057022, + 0.020224804058670998, + 0.006418583914637566, + 0.004353338852524757, + 0.012682688422501087, + 0.006361240055412054, + 0.006364527624100447, + 0.00692936684936285, + 0.00681473221629858, + 0.007808414287865162, + 0.008820751681923866, + 0.002206162316724658, + 0.002809526864439249, + 0.002963288454338908, + 0.0028548138216137886, + 0.0029593969229608774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.011507445015013218 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903630077838898 + ], + "opposite_logprob_tokens": [ + -2.464036226272583 + ], + "logprob_reward": -0.08903630077838898, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903630077838898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054902106523513794, + 0.07221904397010803, + 0.04653023183345795, + 0.08671160042285919, + 0.0959642231464386, + 0.43704360723495483, + 0.08733180910348892, + 0.04127918556332588, + 0.045304395258426666, + 0.032713793218135834 + ], + "credit_map": [ + 0.00779991177842021, + 0.005872746929526329, + 0.01071415189653635, + 0.007143153343349695, + 0.00519562466070056, + 0.008413885720074177, + 0.00487129669636488, + 0.004891331307590008, + 0.007496235426515341, + 0.011835617013275623, + 0.008446866646409035, + 0.008412866853177547, + 0.008925042115151882, + 0.012717904523015022, + 0.007175924256443977, + 0.007208586670458317, + 0.006019426044076681, + 0.009154603816568851, + 0.006578944623470306, + 0.004396114964038134, + 0.006503907963633537, + 0.004884406924247742, + 0.004290022421628237, + 0.004702808801084757, + 0.006901138927787542, + 0.006626958027482033, + 0.011150367558002472, + 0.019752833992242813, + 0.014829120598733425, + 0.008974005468189716, + 0.012833420187234879, + 0.0056437538005411625, + 0.006895586848258972, + 0.006052082870155573, + 0.024521004408597946, + 0.03203757479786873, + 0.0044108182191848755, + 0.00776366051286459, + 0.007044378202408552, + 0.007239120546728373, + 0.026304414495825768, + 0.05638932064175606, + 0.049420490860939026, + 0.16150982677936554, + 0.0305887833237648, + 0.04526582360267639, + 0.04080716148018837, + 0.026757804676890373, + 0.006660917773842812, + 0.007530359085649252, + 0.007318977266550064, + 0.009615378454327583, + 0.021877117455005646, + 0.018578659743070602, + 0.007746949326246977, + 0.008003449998795986, + 0.004585091955959797, + 0.00865014735609293, + 0.004928684793412685, + 0.0033635657746344805, + 0.0037510255351662636, + 0.004437113180756569, + 0.007574367802590132, + 0.003989191260188818, + 0.005922711454331875, + 0.005504220258444548, + 0.006722319405525923, + 0.005505504086613655, + 0.007452504243701696, + 0.003927447833120823, + 0.0064605893567204475, + 0.003809101413935423, + 0.004833371844142675, + 0.006905753165483475, + 0.008628721348941326, + 0.002063087187707424, + 0.0027358666993677616, + 0.0025225181598216295, + 0.0024035261012613773, + 0.002620949177071452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 6 + ], + "selected_reward": -0.08903630077838898 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003610759973526001 + ], + "opposite_logprob_tokens": [ + -5.628610610961914 + ], + "logprob_reward": -0.003610759973526001, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003610759973526001, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09962347894906998, + 0.07887441664934158, + 0.057479310780763626, + 0.023862015455961227, + 0.029307816177606583, + 0.027052314952015877, + 0.15058133006095886, + 0.07780296355485916, + 0.4384443461894989, + 0.01697205938398838 + ], + "credit_map": [ + 0.012452887371182442, + 0.010318257845938206, + 0.013458037748932838, + 0.012695493176579475, + 0.013619163073599339, + 0.016444042325019836, + 0.011629913002252579, + 0.009005681611597538, + 0.005395686719566584, + 0.004020702559500933, + 0.0245478767901659, + 0.004851675592362881, + 0.022541094571352005, + 0.007549424655735493, + 0.006895601283758879, + 0.0030723530799150467, + 0.008269794285297394, + 0.00773288169875741, + 0.007492105942219496, + 0.005569767206907272, + 0.007669405080378056, + 0.009741024114191532, + 0.0062091113068163395, + 0.004795217886567116, + 0.0034281129483133554, + 0.002867834409698844, + 0.0036700726486742496, + 0.002676967065781355, + 0.0027852824423462152, + 0.002582933520898223, + 0.004026530776172876, + 0.0018242780352011323, + 0.00255555915646255, + 0.0021000057458877563, + 0.007014892995357513, + 0.008389792405068874, + 0.002245972165837884, + 0.003066869219765067, + 0.0019028517417609692, + 0.002031870884820819, + 0.0016821736935526133, + 0.005007957573980093, + 0.002816908759996295, + 0.006651627831161022, + 0.002535713603720069, + 0.0024471476208418608, + 0.003597159869968891, + 0.002313621575012803, + 0.011821971274912357, + 0.014853554777801037, + 0.01291667390614748, + 0.02066643163561821, + 0.027630414813756943, + 0.03479662910103798, + 0.012186535634100437, + 0.015709102153778076, + 0.006716880947351456, + 0.021757742390036583, + 0.009169467724859715, + 0.005293005611747503, + 0.011230436153709888, + 0.006481970194727182, + 0.011458590626716614, + 0.005694860126823187, + 0.032322704792022705, + 0.1425314098596573, + 0.03959793224930763, + 0.04916788265109062, + 0.08511320501565933, + 0.029697207733988762, + 0.032354213297367096, + 0.027659747749567032, + 0.003788204863667488, + 0.002313511911779642, + 0.002737975912168622, + 0.0013754012761637568, + 0.00171723123639822, + 0.0016605640994384885, + 0.001494149211794138, + 0.0018850191263481975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.003610759973526001 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945852026343346 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945852026343346, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945852026343346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19294887781143188, + 0.0793428122997284, + 0.11570536345243454, + 0.024989182129502296, + 0.0478719100356102, + 0.03012937493622303, + 0.261089026927948, + 0.07090073078870773, + 0.1357174515724182, + 0.04130532965064049 + ], + "credit_map": [ + 0.028727855533361435, + 0.02352488972246647, + 0.022563733160495758, + 0.017216572538018227, + 0.022227907553315163, + 0.047066956758499146, + 0.014951979741454124, + 0.01666896976530552, + 0.008168833330273628, + 0.005105422809720039, + 0.019959604367613792, + 0.007371807470917702, + 0.013035165145993233, + 0.011953618377447128, + 0.004826750606298447, + 0.008921608328819275, + 0.015710141509771347, + 0.02394685707986355, + 0.019649282097816467, + 0.010978770442306995, + 0.012116600759327412, + 0.011372197419404984, + 0.01214864943176508, + 0.009782867506146431, + 0.0019096919568255544, + 0.0016674874350428581, + 0.004738105461001396, + 0.003446513321250677, + 0.0030649967957288027, + 0.0032821735367178917, + 0.004519355483353138, + 0.0023608573246747255, + 0.004151647444814444, + 0.005932270549237728, + 0.006315081380307674, + 0.013246603310108185, + 0.006309601478278637, + 0.005714682396501303, + 0.0028856880962848663, + 0.0033163365442305803, + 0.0016524706734344363, + 0.003244020976126194, + 0.002597648650407791, + 0.004945188760757446, + 0.0038809366524219513, + 0.003806492779403925, + 0.006152560468763113, + 0.0038500558584928513, + 0.019666185602545738, + 0.03112347424030304, + 0.027463922277092934, + 0.047901347279548645, + 0.03234877437353134, + 0.04185837134718895, + 0.025851840153336525, + 0.03487511724233627, + 0.009591182693839073, + 0.011189086362719536, + 0.008454280905425549, + 0.0058390311896800995, + 0.006505182012915611, + 0.007367071229964495, + 0.015178508125245571, + 0.006776388268917799, + 0.02231467515230179, + 0.028263164684176445, + 0.018188998103141785, + 0.009151673875749111, + 0.016406256705522537, + 0.011546019464731216, + 0.020345427095890045, + 0.009501240216195583, + 0.005675541702657938, + 0.008739604614675045, + 0.008480444550514221, + 0.0020468963775783777, + 0.0038747249636799097, + 0.003913016989827156, + 0.0035130572505295277, + 0.005062043201178312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.005945852026343346 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02327875606715679 + ], + "opposite_logprob_tokens": [ + -3.7732787132263184 + ], + "logprob_reward": -0.02327875606715679, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02327875606715679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09064934402704239, + 0.042786285281181335, + 0.05341166630387306, + 0.03142421320080757, + 0.055486954748630524, + 0.06859444081783295, + 0.16955463588237762, + 0.08778455853462219, + 0.3610590994358063, + 0.03924877196550369 + ], + "credit_map": [ + 0.014040092006325722, + 0.009164002723991871, + 0.015341924503445625, + 0.013091430999338627, + 0.011456580832600594, + 0.010382004082202911, + 0.007820408791303635, + 0.009352898225188255, + 0.0038476577028632164, + 0.0042778681963682175, + 0.009149222634732723, + 0.004879174754023552, + 0.006808545906096697, + 0.005884875543415546, + 0.0027946217451244593, + 0.005144320894032717, + 0.0077798874117434025, + 0.015524650923907757, + 0.008501826785504818, + 0.0038000738713890314, + 0.005296564195305109, + 0.006252532824873924, + 0.0036358472425490618, + 0.002620282582938671, + 0.0027102024760097265, + 0.0030816805083304644, + 0.0038453005254268646, + 0.002719103591516614, + 0.004838399589061737, + 0.00726071884855628, + 0.004305928014218807, + 0.0026628782507032156, + 0.004309256561100483, + 0.0036074784584343433, + 0.00918836984783411, + 0.011759879998862743, + 0.006205152720212936, + 0.008087855763733387, + 0.006074930541217327, + 0.006254029460251331, + 0.004537767730653286, + 0.007791366893798113, + 0.007361351512372494, + 0.022697359323501587, + 0.005053315311670303, + 0.006661828141659498, + 0.008703558705747128, + 0.005787892732769251, + 0.012243104167282581, + 0.021583043038845062, + 0.01770464889705181, + 0.029368288815021515, + 0.021260127425193787, + 0.02619875594973564, + 0.01767776720225811, + 0.023518914356827736, + 0.009957757778465748, + 0.008426861837506294, + 0.011949105188250542, + 0.010676528327167034, + 0.009079077281057835, + 0.008264810778200626, + 0.022469529882073402, + 0.006960882339626551, + 0.026494568213820457, + 0.11467500030994415, + 0.03313804790377617, + 0.048267386853694916, + 0.06599980592727661, + 0.021598434075713158, + 0.0254339762032032, + 0.02545185759663582, + 0.005957168992608786, + 0.010288373567163944, + 0.009149079211056232, + 0.0018515271367505193, + 0.003270031651481986, + 0.003063254989683628, + 0.0024142730981111526, + 0.0032550632022321224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.02327875606715679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-length shot, rendered in a surreal Unreal Engine style, captures a moment of pure camaraderie between two young men on a bright, sun-drenched city street corner. A young white male with tousled, sun-bleached blond hair and a friendly smile, dressed in a soft white cotton T-shirt and a textured blue denim jacket, faces a young black male with neatly braided hair and a warm, genuine smile, dressed in a heather-gray pullover with an intricate weave. The scene is captured at a crisp, high frame rate, and their synchronized movements unfold with perfect fluidity as they smoothly extend their arms and bump fists together. The connection is firm and solid, a testament to their bond, and their shared smile grows even brighter in the moment of connection. Intense, high-angle sunlight bathes the scene, creating strong rim lighting on their shoulders and hair, and producing subtle, aesthetically pleasing lens flares that drift across the frame. In the background, the gentle bustle of city life—blurred pedestrians and soft-focus vehicles—is rendered with a shallow depth of field, focusing all attention on the genuine, joyful interactions, meticulously depicting the realistic textures of their skin and clothing.", + "seed": 32248, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32248_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.021356187853962184, + "vlm_logprob": -0.021356187853962184, + "vlm_yes_no_margin": 4.325, + "vlm_token_credit": -0.021356187853962184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007469463627785444 + ], + "opposite_logprob_tokens": [ + -5.382469654083252 + ], + "logprob_reward": -0.007469463627785444, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.007469463627785444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31542667746543884, + 0.041352421045303345, + 0.03700356185436249, + 0.07286018133163452, + 0.03925740346312523, + 0.14237426221370697, + 0.05630303546786308, + 0.06929323077201843, + 0.09283972531557083, + 0.13328947126865387 + ], + "credit_map": [ + 0.038038067519664764, + 0.030689271166920662, + 0.03239177167415619, + 0.029018104076385498, + 0.04161499813199043, + 0.08382512629032135, + 0.020543526858091354, + 0.039305806159973145, + 0.005179605912417173, + 0.004274788312613964, + 0.005312999244779348, + 0.0060754697769880295, + 0.005509702488780022, + 0.005684285424649715, + 0.005567394196987152, + 0.0037481747567653656, + 0.003908728715032339, + 0.0067484877072274685, + 0.0038255550898611546, + 0.00548272579908371, + 0.004046435002237558, + 0.004280359949916601, + 0.0036704347003251314, + 0.0050408379174768925, + 0.009691585786640644, + 0.010624715127050877, + 0.005556969903409481, + 0.012476460076868534, + 0.010087518952786922, + 0.00895011518150568, + 0.006561441347002983, + 0.00891137309372425, + 0.003790401853621006, + 0.006770158652216196, + 0.0037673949263989925, + 0.005077201407402754, + 0.005424439441412687, + 0.004588071256875992, + 0.005666361190378666, + 0.004173374734818935, + 0.01367957703769207, + 0.017348960041999817, + 0.008590709418058395, + 0.009174313396215439, + 0.03658238798379898, + 0.021827811375260353, + 0.01578100584447384, + 0.019389500841498375, + 0.005160118918865919, + 0.009968605823814869, + 0.007123468443751335, + 0.006954769138246775, + 0.006986507214605808, + 0.007177270948886871, + 0.006554711144417524, + 0.006377583369612694, + 0.005054199602454901, + 0.0079495282843709, + 0.007530878763645887, + 0.004466743674129248, + 0.013761473819613457, + 0.012111996300518513, + 0.009225152432918549, + 0.00919326301664114, + 0.011311573907732964, + 0.011879409663379192, + 0.008239052258431911, + 0.00940346997231245, + 0.01609445922076702, + 0.015488429926335812, + 0.008393689058721066, + 0.012029644101858139, + 0.010177225805819035, + 0.01056120079010725, + 0.01075067650526762, + 0.013676485046744347, + 0.022263847291469574, + 0.011967264115810394, + 0.020471179857850075, + 0.0334215983748436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.007469463627785444 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054949671030044556 + ], + "opposite_logprob_tokens": [ + -2.9299497604370117 + ], + "logprob_reward": -0.054949671030044556, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054949671030044556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13299810886383057, + 0.033925604075193405, + 0.08297131210565567, + 0.08841224759817123, + 0.05728325992822647, + 0.233642116189003, + 0.1083984225988388, + 0.05789489671587944, + 0.11988651752471924, + 0.08458752185106277 + ], + "credit_map": [ + 0.014163521118462086, + 0.013945887796580791, + 0.021238848567008972, + 0.01994074508547783, + 0.02249833196401596, + 0.018783198669552803, + 0.010185999795794487, + 0.012241578660905361, + 0.0054437629878520966, + 0.003907919395714998, + 0.004452758003026247, + 0.0037044130731374025, + 0.0046130456030368805, + 0.004031143616884947, + 0.004591449163854122, + 0.003181111067533493, + 0.00837003905326128, + 0.011374646797776222, + 0.006467647384852171, + 0.01298584509640932, + 0.00748472148552537, + 0.009645744226872921, + 0.014041299931704998, + 0.012601366266608238, + 0.006070103961974382, + 0.01541837491095066, + 0.008990269154310226, + 0.015331885777413845, + 0.007650146726518869, + 0.010383588261902332, + 0.012361457571387291, + 0.012206408195197582, + 0.007507997564971447, + 0.006641326937824488, + 0.007794768549501896, + 0.007907628081738949, + 0.005008479580283165, + 0.009511265903711319, + 0.006759575102478266, + 0.006152215879410505, + 0.02456933818757534, + 0.030603433027863503, + 0.01589428074657917, + 0.016694389283657074, + 0.05299236997961998, + 0.03470000624656677, + 0.028357863426208496, + 0.02983042784035206, + 0.017277972772717476, + 0.010236633010208607, + 0.006600989494472742, + 0.006127545610070229, + 0.020803198218345642, + 0.012056674808263779, + 0.017356466501951218, + 0.017938928678631783, + 0.0047029415145516396, + 0.008062507025897503, + 0.005340568255633116, + 0.004457211587578058, + 0.006458701100200415, + 0.011724167503416538, + 0.008240731433033943, + 0.008908065035939217, + 0.01445751078426838, + 0.019029526039958, + 0.010046354494988918, + 0.015510535798966885, + 0.015850571915507317, + 0.018360260874032974, + 0.008764860220253468, + 0.017866894602775574, + 0.006454132962971926, + 0.007057079114019871, + 0.006469063926488161, + 0.010532489977777004, + 0.014918755739927292, + 0.006256821099668741, + 0.011027994565665722, + 0.021871179342269897 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.054949671030044556 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005942297168076038 + ], + "opposite_logprob_tokens": [ + -5.130942344665527 + ], + "logprob_reward": -0.005942297168076038, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005942297168076038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33916816115379333, + 0.03881261497735977, + 0.11771149933338165, + 0.20268471539020538, + 0.03190191835165024, + 0.07929572463035583, + 0.03424419090151787, + 0.03617729991674423, + 0.08078968524932861, + 0.03921422362327576 + ], + "credit_map": [ + 0.05605668947100639, + 0.04330328479409218, + 0.06377587467432022, + 0.05502888560295105, + 0.04724700003862381, + 0.024453725665807724, + 0.023928778246045113, + 0.0253739133477211, + 0.009320291690528393, + 0.00376272015273571, + 0.0036049068439751863, + 0.003731770673766732, + 0.005145993083715439, + 0.004894292447715998, + 0.005383470095694065, + 0.0029691713862121105, + 0.011950269341468811, + 0.015245764516294003, + 0.012335996143519878, + 0.01177167147397995, + 0.014760036952793598, + 0.013471849262714386, + 0.021518738940358162, + 0.016657168045639992, + 0.05008087307214737, + 0.02158273570239544, + 0.018887337297201157, + 0.017071593552827835, + 0.024918917566537857, + 0.02483845129609108, + 0.023791754618287086, + 0.02151305601000786, + 0.004616937134414911, + 0.00470592500641942, + 0.004447381943464279, + 0.00415780209004879, + 0.002925524953752756, + 0.003772973082959652, + 0.004040469415485859, + 0.0032349044922739267, + 0.007707125041633844, + 0.009739990346133709, + 0.003843373619019985, + 0.00526779517531395, + 0.01816112920641899, + 0.013425547629594803, + 0.009974343702197075, + 0.011176424100995064, + 0.006082418840378523, + 0.003757342230528593, + 0.0033267270773649216, + 0.0024611307308077812, + 0.0047730994410812855, + 0.0036854699719697237, + 0.004681991878896952, + 0.00547600956633687, + 0.0043051461689174175, + 0.005156972445547581, + 0.0029809605330228806, + 0.0025742510333657265, + 0.004313165787607431, + 0.006353732664138079, + 0.0040739248506724834, + 0.006419147364795208, + 0.011768555268645287, + 0.011312827467918396, + 0.007988041266798973, + 0.014192521572113037, + 0.007615583948791027, + 0.0082219447940588, + 0.004341483116149902, + 0.015348727814853191, + 0.008903246372938156, + 0.0035722649190574884, + 0.003594431094825268, + 0.003817574819549918, + 0.006576234940439463, + 0.004372422117739916, + 0.004280915018171072, + 0.004097132943570614 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.005942297168076038 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634377270936966 + ], + "opposite_logprob_tokens": [ + -4.758634567260742 + ], + "logprob_reward": -0.008634377270936966, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634377270936966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23794792592525482, + 0.029670067131519318, + 0.04339643940329552, + 0.03229806572198868, + 0.020565815269947052, + 0.36373522877693176, + 0.09285259246826172, + 0.026507828384637833, + 0.11319439858198166, + 0.039831697940826416 + ], + "credit_map": [ + 0.03491055220365524, + 0.037281882017850876, + 0.045756105333566666, + 0.04078703001141548, + 0.02439766563475132, + 0.020033346489071846, + 0.01844310387969017, + 0.016338223591446877, + 0.003786653047427535, + 0.0037522362545132637, + 0.0025167723651975393, + 0.003087414661422372, + 0.005581249482929707, + 0.004947073291987181, + 0.002973552793264389, + 0.003025114769116044, + 0.0043685161508619785, + 0.006369067821651697, + 0.003778852056711912, + 0.005501129198819399, + 0.005270711146295071, + 0.004765737801790237, + 0.007037630304694176, + 0.006304790731519461, + 0.004932841751724482, + 0.004713912028819323, + 0.0023983546998351812, + 0.0033309750724583864, + 0.0038810735568404198, + 0.0037528451066464186, + 0.004826903808861971, + 0.004461158998310566, + 0.0020116446539759636, + 0.0031384574249386787, + 0.0028130298014730215, + 0.0033103893510997295, + 0.0023864805698394775, + 0.0023820861242711544, + 0.0025671557523310184, + 0.0019565708935260773, + 0.03698688745498657, + 0.05003810301423073, + 0.018069183453917503, + 0.022711891680955887, + 0.09521247446537018, + 0.052797604352235794, + 0.04884791374206543, + 0.03907114267349243, + 0.014003679156303406, + 0.00889386422932148, + 0.006551796104758978, + 0.004670547787100077, + 0.022595657035708427, + 0.010723743587732315, + 0.012312382459640503, + 0.013100924901664257, + 0.0027301639784127474, + 0.0028594755567610264, + 0.0023603513836860657, + 0.0020797844044864178, + 0.003881776938214898, + 0.005479204468429089, + 0.0025404610205441713, + 0.004576610866934061, + 0.014847815036773682, + 0.020132243633270264, + 0.015857050195336342, + 0.014461717568337917, + 0.01174965687096119, + 0.011449995450675488, + 0.006146817933768034, + 0.018549101427197456, + 0.009962162934243679, + 0.004852079786360264, + 0.003460115985944867, + 0.0033989730291068554, + 0.0051323226653039455, + 0.002740240655839443, + 0.0024949759244918823, + 0.007790826261043549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.008634377270936966 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02978513017296791 + ], + "opposite_logprob_tokens": [ + -3.52978515625 + ], + "logprob_reward": -0.02978513017296791, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02978513017296791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2672743499279022, + 0.0615963339805603, + 0.057297658175230026, + 0.055180251598358154, + 0.08379974216222763, + 0.1404121071100235, + 0.07877305150032043, + 0.03787430375814438, + 0.14946702122688293, + 0.06832511723041534 + ], + "credit_map": [ + 0.03990066051483154, + 0.03429331257939339, + 0.05273737758398056, + 0.0414348728954792, + 0.03853059560060501, + 0.022323599085211754, + 0.019146492704749107, + 0.01890743337571621, + 0.009914235211908817, + 0.00642215833067894, + 0.008766449056565762, + 0.006154890637844801, + 0.012945465743541718, + 0.007055066060274839, + 0.005259308032691479, + 0.00507876044139266, + 0.005533700808882713, + 0.005428902339190245, + 0.006400043144822121, + 0.0052867792546749115, + 0.0065617915242910385, + 0.007153651211410761, + 0.010234792716801167, + 0.01069799903780222, + 0.007849209010601044, + 0.009006415493786335, + 0.004548944067209959, + 0.006648357957601547, + 0.006357244681566954, + 0.00586226861923933, + 0.008205246180295944, + 0.006702560465782881, + 0.012730032205581665, + 0.013257890939712524, + 0.010103406384587288, + 0.010501391254365444, + 0.00866264384239912, + 0.009905417449772358, + 0.010227266699075699, + 0.008411694318056107, + 0.014133911579847336, + 0.019845211878418922, + 0.008321162313222885, + 0.01020251214504242, + 0.03574239835143089, + 0.020159659907221794, + 0.01681927777826786, + 0.015187980607151985, + 0.013677152805030346, + 0.006950360722839832, + 0.004958285018801689, + 0.004173355642706156, + 0.016587980091571808, + 0.00917516928166151, + 0.011428453028202057, + 0.0118222925812006, + 0.0038739561568945646, + 0.004322746768593788, + 0.0038002661895006895, + 0.003184447530657053, + 0.006040450185537338, + 0.006956002209335566, + 0.005099361762404442, + 0.004597073420882225, + 0.02224033512175083, + 0.018626857548952103, + 0.016664063557982445, + 0.017741505056619644, + 0.013249214738607407, + 0.020545268431305885, + 0.008849503472447395, + 0.03155027702450752, + 0.013706894591450691, + 0.00705175893381238, + 0.004925871267914772, + 0.008143629878759384, + 0.00882288534194231, + 0.00508183054625988, + 0.004140343517065048, + 0.01645190641283989 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.02978513017296791 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-length shot, rendered in a surreal Unreal Engine style, captures a moment of pure camaraderie between two young men on a bright, sun-drenched city street corner. A young white male with tousled, sun-bleached blond hair and a friendly smile, dressed in a soft white cotton T-shirt and a textured blue denim jacket, faces a young black male with neatly braided hair and a warm, genuine smile, dressed in a heather-gray pullover with an intricate weave. The scene is captured at a crisp, high frame rate, and their synchronized movements unfold with perfect fluidity as they smoothly extend their arms and bump fists together. The connection is firm and solid, a testament to their bond, and their shared smile grows even brighter in the moment of connection. Intense, high-angle sunlight bathes the scene, creating strong rim lighting on their shoulders and hair, and producing subtle, aesthetically pleasing lens flares that drift across the frame. In the background, the gentle bustle of city life—blurred pedestrians and soft-focus vehicles—is rendered with a shallow depth of field, focusing all attention on the genuine, joyful interactions, meticulously depicting the realistic textures of their skin and clothing.", + "seed": 32249, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32249_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.035388314817100766, + "vlm_logprob": -0.035388314817100766, + "vlm_yes_no_margin": 3.725, + "vlm_token_credit": -0.035388314817100766, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.025017715990543365 + ], + "opposite_logprob_tokens": [ + -3.775017738342285 + ], + "logprob_reward": -0.025017715990543365, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.025017715990543365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10474861413240433, + 0.07788456231355667, + 0.04192624241113663, + 0.5604596734046936, + 0.030940620228648186, + 0.03284326568245888, + 0.02790842205286026, + 0.035934947431087494, + 0.0388069711625576, + 0.04854670539498329 + ], + "credit_map": [ + 0.010683941654860973, + 0.007930414751172066, + 0.009538211859762669, + 0.013715150766074657, + 0.020787280052900314, + 0.01148592121899128, + 0.007261506747454405, + 0.023346180096268654, + 0.008213989436626434, + 0.011936589144170284, + 0.004720358643680811, + 0.0060046399012207985, + 0.020516287535429, + 0.005578318145126104, + 0.004278568085283041, + 0.01663581095635891, + 0.007156417705118656, + 0.003946097567677498, + 0.003676393534988165, + 0.0023047092836350203, + 0.0037802010774612427, + 0.0056495871394872665, + 0.0030463410075753927, + 0.01236649602651596, + 0.06486068665981293, + 0.028015756979584694, + 0.01640225015580654, + 0.032488271594047546, + 0.17562894523143768, + 0.03303373605012894, + 0.03470223397016525, + 0.17532779276371002, + 0.0042920042760670185, + 0.005347062833607197, + 0.004668394103646278, + 0.0029451686423271894, + 0.0023205874022096395, + 0.004604509100317955, + 0.002387429354712367, + 0.004375464282929897, + 0.004419583827257156, + 0.008536649867892265, + 0.004468183033168316, + 0.002409040229395032, + 0.003561372170224786, + 0.004282668232917786, + 0.0025481958873569965, + 0.0026175708044320345, + 0.005042495671659708, + 0.003271665656939149, + 0.0030175952706485987, + 0.00279274582862854, + 0.005580225959420204, + 0.0038308496586978436, + 0.0028839053120464087, + 0.00148894137237221, + 0.004667207133024931, + 0.0038426637183874846, + 0.0031080199405550957, + 0.0029489800799638033, + 0.004507905337959528, + 0.007638015318661928, + 0.005102834198623896, + 0.004119319841265678, + 0.010896923020482063, + 0.0030858248937875032, + 0.003308393992483616, + 0.0037285450380295515, + 0.0029448827262967825, + 0.011097948998212814, + 0.0019618598744273186, + 0.0017825948307290673, + 0.009602400474250317, + 0.004235198255628347, + 0.00881059281527996, + 0.004911492578685284, + 0.004410297144204378, + 0.010189413093030453, + 0.0031947148963809013, + 0.0031925986986607313 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.025017715990543365 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031387954950333 + ], + "opposite_logprob_tokens": [ + -2.350313901901245 + ], + "logprob_reward": -0.10031387954950333, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031387954950333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12968860566616058, + 0.0675891786813736, + 0.08513857424259186, + 0.1548498123884201, + 0.08453503251075745, + 0.11771126091480255, + 0.05189860239624977, + 0.07498209178447723, + 0.09756067395210266, + 0.1360461562871933 + ], + "credit_map": [ + 0.01820029318332672, + 0.01545818243175745, + 0.016659410670399666, + 0.014170745387673378, + 0.011871392838656902, + 0.01512699294835329, + 0.011526839807629585, + 0.026674756780266762, + 0.011458632536232471, + 0.010340429842472076, + 0.006605314556509256, + 0.008323435671627522, + 0.005788139067590237, + 0.007876761257648468, + 0.010169350542128086, + 0.007027111016213894, + 0.015526195988059044, + 0.006731550209224224, + 0.007113437168300152, + 0.00508631719276309, + 0.006137412041425705, + 0.012090113013982773, + 0.00548141822218895, + 0.026972128078341484, + 0.0180269256234169, + 0.00782992783933878, + 0.005364949349313974, + 0.009091895073652267, + 0.046917274594306946, + 0.010285255499184132, + 0.009182078763842583, + 0.04815150052309036, + 0.016385653987526894, + 0.00965374056249857, + 0.01718238554894924, + 0.0062790485098958015, + 0.008481534197926521, + 0.010397462174296379, + 0.006629291921854019, + 0.009525912813842297, + 0.015084699727594852, + 0.025660425424575806, + 0.014433320611715317, + 0.013423953205347061, + 0.0085078664124012, + 0.013275641947984695, + 0.01574668101966381, + 0.011578668840229511, + 0.005761237349361181, + 0.005930823273956776, + 0.00657027680426836, + 0.006343870889395475, + 0.006229548715054989, + 0.00859936885535717, + 0.006797970272600651, + 0.005665507633239031, + 0.00937939528375864, + 0.007088974118232727, + 0.004279831424355507, + 0.005748725961893797, + 0.0153586370870471, + 0.021754497662186623, + 0.006598798558115959, + 0.0047732326202094555, + 0.03214987367391586, + 0.00685900216922164, + 0.007105393335223198, + 0.00784812681376934, + 0.0051646470092237, + 0.031031226739287376, + 0.0031711619812995195, + 0.004231240600347519, + 0.01621820218861103, + 0.012350516393780708, + 0.01030711829662323, + 0.010362395085394382, + 0.027124881744384766, + 0.01360070239752531, + 0.023802338168025017, + 0.0222800113260746 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 0 + ], + "selected_reward": -0.10031387954950333 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191421680152416 + ], + "opposite_logprob_tokens": [ + -4.264191627502441 + ], + "logprob_reward": -0.014191421680152416, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191421680152416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20763260126113892, + 0.11938630789518356, + 0.0653369277715683, + 0.1756535917520523, + 0.11852659285068512, + 0.12814870476722717, + 0.02326039969921112, + 0.02205505780875683, + 0.11193018406629562, + 0.028069593012332916 + ], + "credit_map": [ + 0.020755723118782043, + 0.015640975907444954, + 0.0178414024412632, + 0.022615518420934677, + 0.042684152722358704, + 0.024063337594270706, + 0.015327989123761654, + 0.04870351031422615, + 0.011628126725554466, + 0.014398012310266495, + 0.004718185868114233, + 0.006675883196294308, + 0.03632914647459984, + 0.011528262868523598, + 0.0067090922966599464, + 0.02739959955215454, + 0.0062638442032039165, + 0.00779585400596261, + 0.006347606424242258, + 0.006258385255932808, + 0.007759573869407177, + 0.00882687233388424, + 0.013479004614055157, + 0.008605794049799442, + 0.01973215863108635, + 0.00939044076949358, + 0.005523847881704569, + 0.01025218516588211, + 0.05502452701330185, + 0.012081457301974297, + 0.011523687280714512, + 0.05212530121207237, + 0.019847895950078964, + 0.01150000374764204, + 0.014420775696635246, + 0.008403885178267956, + 0.010431117378175259, + 0.026845362037420273, + 0.011952998116612434, + 0.01512457337230444, + 0.01590397208929062, + 0.017071448266506195, + 0.012904690578579903, + 0.010509908199310303, + 0.01374400407075882, + 0.02123788557946682, + 0.026611488312482834, + 0.010165303945541382, + 0.00287457462400198, + 0.0036710400599986315, + 0.002590551972389221, + 0.0031851292587816715, + 0.002210574923083186, + 0.003024109173566103, + 0.0033934132661670446, + 0.002311006188392639, + 0.0027447198517620564, + 0.0018090206431224942, + 0.0016010142862796783, + 0.002243076916784048, + 0.003853714093565941, + 0.0058836303651332855, + 0.0016867838567122817, + 0.002233098726719618, + 0.03504496067762375, + 0.009566575288772583, + 0.008061252534389496, + 0.007970552891492844, + 0.005105592310428619, + 0.03602035716176033, + 0.004392893053591251, + 0.005768002010881901, + 0.00537321250885725, + 0.0020768139511346817, + 0.0034324522130191326, + 0.0018627071985974908, + 0.003190376562997699, + 0.007846054621040821, + 0.0021210804115980864, + 0.002166896825656295 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.014191421680152416 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762729300186038 + ], + "opposite_logprob_tokens": [ + -4.882627487182617 + ], + "logprob_reward": -0.00762729300186038, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762729300186038, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.152226984500885, + 0.2259284406900406, + 0.06336621195077896, + 0.3337801396846771, + 0.04659940302371979, + 0.0331769660115242, + 0.019920293241739273, + 0.016022633761167526, + 0.07142042368650436, + 0.03755848854780197 + ], + "credit_map": [ + 0.01497089583426714, + 0.013871896080672741, + 0.012333987280726433, + 0.01144983246922493, + 0.03944312408566475, + 0.015223450027406216, + 0.011305452324450016, + 0.03362833708524704, + 0.022766951471567154, + 0.025491179898381233, + 0.008403705433011055, + 0.012218319810926914, + 0.0695185661315918, + 0.022473536431789398, + 0.013008562847971916, + 0.05204762518405914, + 0.009188810363411903, + 0.0060263690538704395, + 0.00417614309117198, + 0.004170294385403395, + 0.007590520195662975, + 0.00912440288811922, + 0.006917350925505161, + 0.016172319650650024, + 0.03910931944847107, + 0.01716216653585434, + 0.009245915338397026, + 0.01912154257297516, + 0.10354381799697876, + 0.021280037239193916, + 0.020295878872275352, + 0.10402144491672516, + 0.005152498837560415, + 0.0058366390876472, + 0.00551066268235445, + 0.0049352217465639114, + 0.007684079464524984, + 0.00845190417021513, + 0.0035381705965846777, + 0.00549022713676095, + 0.005000345874577761, + 0.006119434721767902, + 0.004731365013867617, + 0.0034658389631658792, + 0.0025872020050883293, + 0.003616344416514039, + 0.0033013098873198032, + 0.004355123732239008, + 0.00343713047914207, + 0.003061537630856037, + 0.0020460509695112705, + 0.002436252310872078, + 0.0031497818417847157, + 0.0028683894779533148, + 0.0015552521217614412, + 0.0013658968964591622, + 0.003194308141246438, + 0.001689849072135985, + 0.001347373821772635, + 0.0017635496333241463, + 0.001682937960140407, + 0.0031415696721524, + 0.0017173985252156854, + 0.0014856461202725768, + 0.024388080462813377, + 0.007205527275800705, + 0.005048919934779406, + 0.003555451985448599, + 0.004144038539379835, + 0.021843427792191505, + 0.002777522662654519, + 0.002457450842484832, + 0.008023194037377834, + 0.003071741433814168, + 0.002835986902937293, + 0.0018897723639383912, + 0.005012067500501871, + 0.012389391660690308, + 0.002401071833446622, + 0.0019352649105712771 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.00762729300186038 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979126386344433 + ], + "opposite_logprob_tokens": [ + -3.5297913551330566 + ], + "logprob_reward": -0.02979126386344433, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979126386344433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11455243080854416, + 0.05203574523329735, + 0.05713430419564247, + 0.5016680359840393, + 0.06126774847507477, + 0.03495635464787483, + 0.04205373302102089, + 0.05071767419576645, + 0.046788815408945084, + 0.03882509097456932 + ], + "credit_map": [ + 0.017528660595417023, + 0.012253367342054844, + 0.009969848208129406, + 0.016429103910923004, + 0.015680694952607155, + 0.010353743098676205, + 0.008893072605133057, + 0.023443950340151787, + 0.007584328297525644, + 0.006281680427491665, + 0.003998034633696079, + 0.005213848780840635, + 0.007887011393904686, + 0.00804436020553112, + 0.005624856799840927, + 0.007401630748063326, + 0.008854224346578121, + 0.007617207244038582, + 0.004861605353653431, + 0.003848855849355459, + 0.006496269255876541, + 0.0077589331194758415, + 0.00456945551559329, + 0.013127765618264675, + 0.05973382666707039, + 0.023893428966403008, + 0.014014230109751225, + 0.028504692018032074, + 0.1557571440935135, + 0.030005108565092087, + 0.030268371105194092, + 0.15949125587940216, + 0.008633705787360668, + 0.006286453455686569, + 0.00930003635585308, + 0.006468239706009626, + 0.009226171299815178, + 0.007832400500774384, + 0.005584294907748699, + 0.00793644692748785, + 0.004242633935064077, + 0.003938828594982624, + 0.003755594603717327, + 0.0034979996271431446, + 0.0042586554773151875, + 0.004347964655607939, + 0.006459823343902826, + 0.004454857669770718, + 0.007996058091521263, + 0.005408524069935083, + 0.00366818904876709, + 0.004259963985532522, + 0.008624322712421417, + 0.005308685824275017, + 0.004352615214884281, + 0.0024353787302970886, + 0.00756983133032918, + 0.004186674486845732, + 0.0035635908134281635, + 0.0034377386327832937, + 0.009909084998071194, + 0.0156524907797575, + 0.0034798907581716776, + 0.002918376587331295, + 0.012956508435308933, + 0.004945611581206322, + 0.005247863009572029, + 0.005086794029921293, + 0.00335629191249609, + 0.010647044517099857, + 0.00271694827824831, + 0.001831760979257524, + 0.005387226119637489, + 0.003005390055477619, + 0.008137631230056286, + 0.004787901416420937, + 0.005628463812172413, + 0.00519309239462018, + 0.004177065100520849, + 0.002508322475478053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.02979126386344433 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium-length shot, rendered in a surreal Unreal Engine style, captures a moment of pure camaraderie between two young men on a bright, sun-drenched city street corner. A young white male with tousled, sun-bleached blond hair and a friendly smile, dressed in a soft white cotton T-shirt and a textured blue denim jacket, faces a young black male with neatly braided hair and a warm, genuine smile, dressed in a heather-gray pullover with an intricate weave. The scene is captured at a crisp, high frame rate, and their synchronized movements unfold with perfect fluidity as they smoothly extend their arms and bump fists together. The connection is firm and solid, a testament to their bond, and their shared smile grows even brighter in the moment of connection. Intense, high-angle sunlight bathes the scene, creating strong rim lighting on their shoulders and hair, and producing subtle, aesthetically pleasing lens flares that drift across the frame. In the background, the gentle bustle of city life—blurred pedestrians and soft-focus vehicles—is rendered with a shallow depth of field, focusing all attention on the genuine, joyful interactions, meticulously depicting the realistic textures of their skin and clothing.", + "seed": 32250, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_3/index_7167_seed_32250_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.017953686974942686, + "vlm_logprob": -0.017953686974942686, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.017953686974942686, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008087861351668835 + ], + "opposite_logprob_tokens": [ + -5.258087635040283 + ], + "logprob_reward": -0.008087861351668835, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.008087861351668835, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23506923019886017, + 0.0770130604505539, + 0.08665009588003159, + 0.06261716037988663, + 0.10519684106111526, + 0.09472404420375824, + 0.042605750262737274, + 0.11893755942583084, + 0.047697681933641434, + 0.12948861718177795 + ], + "credit_map": [ + 0.02635323256254196, + 0.03834880515933037, + 0.03372221067547798, + 0.04210303723812103, + 0.023087117820978165, + 0.025036035105586052, + 0.023107916116714478, + 0.02331087365746498, + 0.007137841545045376, + 0.015342244878411293, + 0.008673508651554585, + 0.00864316988736391, + 0.009994289837777615, + 0.007418156135827303, + 0.014364330098032951, + 0.005439515691250563, + 0.006525541190057993, + 0.024466069415211678, + 0.0158409271389246, + 0.005368659272789955, + 0.008928482420742512, + 0.009568101726472378, + 0.011407916434109211, + 0.004544398747384548, + 0.0065366351045668125, + 0.009555656462907791, + 0.009803918190300465, + 0.007318892981857061, + 0.0066818539053201675, + 0.007696002256125212, + 0.007567713037133217, + 0.0074564870446920395, + 0.014590676873922348, + 0.02753443643450737, + 0.012835399247705936, + 0.010432840324938297, + 0.01076369546353817, + 0.014099294319748878, + 0.009444257244467735, + 0.005496248137205839, + 0.009147081524133682, + 0.010511914268136024, + 0.019428692758083344, + 0.008293554186820984, + 0.008020663633942604, + 0.012901845388114452, + 0.02049766480922699, + 0.005922623910009861, + 0.0042736465111374855, + 0.00426487484946847, + 0.007607877720147371, + 0.005282557103782892, + 0.006069476716220379, + 0.006410365924239159, + 0.00569766853004694, + 0.002999285003170371, + 0.007884891703724861, + 0.014676437713205814, + 0.020351171493530273, + 0.011142952367663383, + 0.016023116186261177, + 0.012514966540038586, + 0.013688559643924236, + 0.022655462846159935, + 0.005225192755460739, + 0.007463535293936729, + 0.0051689245738089085, + 0.004608547780662775, + 0.004995943047106266, + 0.007234550546854734, + 0.007694983389228582, + 0.005306003615260124, + 0.008723913691937923, + 0.03774663433432579, + 0.012024324387311935, + 0.024923793971538544, + 0.011077984236180782, + 0.017614055424928665, + 0.006774506065994501, + 0.010603400878608227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.008087861351668835 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023297740146517754 + ], + "opposite_logprob_tokens": [ + -3.7732977867126465 + ], + "logprob_reward": -0.023297740146517754, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023297740146517754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10875105112791061, + 0.034224048256874084, + 0.11784612387418747, + 0.07762988656759262, + 0.42912378907203674, + 0.0636226236820221, + 0.062083899974823, + 0.053799115121364594, + 0.020663226023316383, + 0.03225630521774292 + ], + "credit_map": [ + 0.013543330132961273, + 0.019271157681941986, + 0.01899753324687481, + 0.019041020423173904, + 0.010704057291150093, + 0.008304710499942303, + 0.011590422131121159, + 0.007298820652067661, + 0.0031244056299328804, + 0.006790132727473974, + 0.005144250579178333, + 0.0038089067675173283, + 0.0029931350145488977, + 0.003914788831025362, + 0.005611105356365442, + 0.002837324980646372, + 0.00919750239700079, + 0.037510909140110016, + 0.021538855507969856, + 0.008186607621610165, + 0.0094343526288867, + 0.012770682573318481, + 0.01344316266477108, + 0.005764042492955923, + 0.007630550302565098, + 0.010831515304744244, + 0.010266818106174469, + 0.008671816438436508, + 0.009595304727554321, + 0.011322909966111183, + 0.008911761455237865, + 0.010399209335446358, + 0.04983042553067207, + 0.04841335490345955, + 0.0720856636762619, + 0.04970452934503555, + 0.045041732490062714, + 0.08353259414434433, + 0.05197354406118393, + 0.02854197844862938, + 0.0074469177052378654, + 0.005864822771400213, + 0.01676492765545845, + 0.004656850825995207, + 0.005148523021489382, + 0.0056820339523255825, + 0.014225487597286701, + 0.0038330615498125553, + 0.0034844104666262865, + 0.004159117583185434, + 0.015421578660607338, + 0.005884460639208555, + 0.02188895083963871, + 0.003653203137218952, + 0.004305411595851183, + 0.0032867658883333206, + 0.003398364642634988, + 0.0069030094891786575, + 0.007399057969450951, + 0.004713543225079775, + 0.0073903766460716724, + 0.0057403454557061195, + 0.007967608980834484, + 0.010286807082593441, + 0.0018751182360574603, + 0.005890555679798126, + 0.0023503140546381474, + 0.00246255355887115, + 0.0029023527167737484, + 0.0017987677128985524, + 0.0018690229626372457, + 0.0015145394718274474, + 0.002616730984300375, + 0.00807593297213316, + 0.00434274785220623, + 0.004322059918195009, + 0.0034058690071105957, + 0.004339227452874184, + 0.002206906210631132, + 0.0029468315187841654 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 0 + ], + "selected_reward": -0.023297740146517754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016059570014476776 + ], + "opposite_logprob_tokens": [ + -4.141059398651123 + ], + "logprob_reward": -0.016059570014476776, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016059570014476776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12115951627492905, + 0.0818561241030693, + 0.15762506425380707, + 0.04652814567089081, + 0.1715022772550583, + 0.05827339366078377, + 0.04813312366604805, + 0.2626390755176544, + 0.015363846905529499, + 0.0369194820523262 + ], + "credit_map": [ + 0.008855093270540237, + 0.015714328736066818, + 0.013119705952703953, + 0.019322890788316727, + 0.019877010956406593, + 0.0146919135004282, + 0.020305901765823364, + 0.009272674098610878, + 0.007000800222158432, + 0.012684069573879242, + 0.010057142935693264, + 0.008133633993566036, + 0.010492278262972832, + 0.00954122468829155, + 0.01627356745302677, + 0.007673407904803753, + 0.014346756041049957, + 0.023085076361894608, + 0.022495919838547707, + 0.011416350491344929, + 0.03914611041545868, + 0.018384398892521858, + 0.0202077217400074, + 0.008542733266949654, + 0.003452940611168742, + 0.0074904123321175575, + 0.0139014832675457, + 0.0030817524529993534, + 0.0036960989236831665, + 0.009143486618995667, + 0.00307364109903574, + 0.002688330365344882, + 0.02036701701581478, + 0.022157981991767883, + 0.030677763745188713, + 0.018362320959568024, + 0.01754380390048027, + 0.03218600153923035, + 0.019456559792160988, + 0.010750828310847282, + 0.006803687661886215, + 0.006232007406651974, + 0.013086685910820961, + 0.005552753806114197, + 0.0069591812789440155, + 0.008134549483656883, + 0.008261511102318764, + 0.0032430170103907585, + 0.0030409544706344604, + 0.003172661643475294, + 0.01085102092474699, + 0.004441891331225634, + 0.014042646624147892, + 0.0040975636802613735, + 0.004825929179787636, + 0.0036604590713977814, + 0.020185358822345734, + 0.021061640232801437, + 0.0361555740237236, + 0.03877069428563118, + 0.05492619052529335, + 0.01899825967848301, + 0.03446689993143082, + 0.038074471056461334, + 0.0011386512778699398, + 0.002483542077243328, + 0.002917703939601779, + 0.0017656490672379732, + 0.001753428834490478, + 0.0018754183547571301, + 0.001655406435020268, + 0.0017740473849698901, + 0.00244558765552938, + 0.004612496122717857, + 0.003889417741447687, + 0.005524139851331711, + 0.004189672414213419, + 0.003041194984689355, + 0.009767971001565456, + 0.0034490018151700497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 2 + ], + "selected_reward": -0.016059570014476776 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14855876564979553, + 0.07562380284070969, + 0.14435799419879913, + 0.0722828283905983, + 0.1940910369157791, + 0.08684121817350388, + 0.03088472969830036, + 0.11138265579938889, + 0.04084167256951332, + 0.09513527899980545 + ], + "credit_map": [ + 0.01593279093503952, + 0.021835116669535637, + 0.019904423505067825, + 0.026580896228551865, + 0.020508665591478348, + 0.01629733107984066, + 0.01606554538011551, + 0.011434003710746765, + 0.007382362615317106, + 0.011244303546845913, + 0.009680061601102352, + 0.010756057687103748, + 0.011344566941261292, + 0.008733081631362438, + 0.009816646575927734, + 0.0066667236387729645, + 0.01330037135630846, + 0.04743003472685814, + 0.02048574388027191, + 0.0070470478385686874, + 0.02076571248471737, + 0.014833508059382439, + 0.013939048163592815, + 0.006556529086083174, + 0.004585398826748133, + 0.009178335778415203, + 0.011744877323508263, + 0.0074333143420517445, + 0.007190587930381298, + 0.01962694153189659, + 0.006546610966324806, + 0.005976759362965822, + 0.02234821580350399, + 0.024624565616250038, + 0.0305411908775568, + 0.022891614586114883, + 0.019517870619893074, + 0.036718614399433136, + 0.02510407753288746, + 0.01234489493072033, + 0.013131645508110523, + 0.007163516711443663, + 0.02393423207104206, + 0.0070799654349684715, + 0.00703929690644145, + 0.010278844274580479, + 0.01228596456348896, + 0.005927749909460545, + 0.002574407961219549, + 0.0025820874143391848, + 0.004070680122822523, + 0.002831771271303296, + 0.006397982127964497, + 0.005469081923365593, + 0.0031269376631826162, + 0.003831778885796666, + 0.007009054999798536, + 0.014583436772227287, + 0.014186346903443336, + 0.016134807839989662, + 0.013691338710486889, + 0.014530199579894543, + 0.019230183213949203, + 0.01201728917658329, + 0.00391051871702075, + 0.009871446527540684, + 0.005925927776843309, + 0.004876889754086733, + 0.004487517289817333, + 0.005169818177819252, + 0.002708146581426263, + 0.003891405649483204, + 0.006336688995361328, + 0.015661772340536118, + 0.010910957120358944, + 0.011386038735508919, + 0.013433237560093403, + 0.010061237029731274, + 0.018643230199813843, + 0.008702117949724197 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.012535471469163895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two young men visible in the scene, one with blond hair and a white denim jacket, and the other with braided hair and a gray pullover?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Do the two men extend their arms to bump fists together in a synchronized motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a bright, sun-drenched city street corner with blurred pedestrians and vehicles in the background?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature intense high-angle sunlight creating strong rim lighting on the subjects' shoulders and hair?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the video rendered in a surreal Unreal Engine style with a shallow depth of field focusing on the characters?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029787791892886162 + ], + "opposite_logprob_tokens": [ + -3.52978777885437 + ], + "logprob_reward": -0.029787791892886162, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029787791892886162, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1855926364660263, + 0.1270064115524292, + 0.06330206245183945, + 0.08625643700361252, + 0.19410717487335205, + 0.08211531490087509, + 0.06660807877779007, + 0.08680397272109985, + 0.047832150012254715, + 0.06037580221891403 + ], + "credit_map": [ + 0.02034093625843525, + 0.03554803878068924, + 0.029083089902997017, + 0.026837801560759544, + 0.03144115209579468, + 0.01382426917552948, + 0.017291339114308357, + 0.0112259890884161, + 0.014120775274932384, + 0.020801300182938576, + 0.016193106770515442, + 0.021395035088062286, + 0.013051679357886314, + 0.01054450310766697, + 0.021634645760059357, + 0.009265354834496975, + 0.003821644699200988, + 0.009080902673304081, + 0.010669851675629616, + 0.005195676349103451, + 0.010669125244021416, + 0.009444337338209152, + 0.008918856270611286, + 0.00550166517496109, + 0.006280696019530296, + 0.01348776463419199, + 0.0174147617071867, + 0.007035989314317703, + 0.009327970445156097, + 0.014669150114059448, + 0.00841846689581871, + 0.009621632285416126, + 0.024326074868440628, + 0.026871925219893456, + 0.03628704324364662, + 0.01821715757250786, + 0.019810151308774948, + 0.03705937787890434, + 0.019587356597185135, + 0.011948080733418465, + 0.008831843733787537, + 0.005462525878101587, + 0.02258363366127014, + 0.005956206936389208, + 0.007282906211912632, + 0.00847680028527975, + 0.01905200071632862, + 0.004469394683837891, + 0.004430690780282021, + 0.007448089774698019, + 0.009055407717823982, + 0.005138247739523649, + 0.02049373835325241, + 0.007026993203908205, + 0.00572977215051651, + 0.00728512741625309, + 0.006318775471299887, + 0.009267904795706272, + 0.008866013027727604, + 0.01670939102768898, + 0.012068885378539562, + 0.013544931076467037, + 0.01200331375002861, + 0.008024749346077442, + 0.0047802794724702835, + 0.013500012457370758, + 0.006494984496384859, + 0.006886341143399477, + 0.004996096715331078, + 0.004106746055185795, + 0.003827918553724885, + 0.003239770419895649, + 0.005058083217591047, + 0.009363390505313873, + 0.006777349393814802, + 0.009700234048068523, + 0.006789469625800848, + 0.008254620246589184, + 0.006068019662052393, + 0.008364635519683361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.029787791892886162 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_tensor([2543, 2543])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_tensor([2543, 2543])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e7a7259a61c36f1dab5a9feec677d12163e34fae --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_tensor([2543, 2543])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, medium shot, held at eye level, frames a lightweight wooden door, its surface revealing the fine grain of aged pine beneath a layer of cracked and flaking white paint, slightly ajar in a dark, worn frame. The scene is a dimly lit, silent corridor. A cool, gray light from an unseen window filters across the dusty wooden floorboards, illuminating the dancing dust motes in the still air. An unseen draft gently applies pressure to the door, causing it to swing silently inward on its tarnished brass hinges in a wide, ghostly arc, revealing more of the profound darkness of the room beyond. It then reverses direction, beginning a slow, hypnotic back-and-forth oscillation. With each passage, the faint, dry creak of the hinges softens as the door's momentum steadily fades, until it finally rests in complete silence, precisely returning to its original, slightly ajar position, creating a haunting and suspenseful atmosphere.", + "seed": 3387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_2543_seed_3387_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.12993693863973022, + "vlm_logprob": -0.12993693863973022, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.12993693863973022, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.280563622713089 + ], + "opposite_logprob_tokens": [ + -1.6555635929107666 + ], + "logprob_reward": -0.280563622713089, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.280563622713089, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11281219869852066, + 0.041525356471538544, + 0.355785995721817, + 0.05678069591522217, + 0.13911785185337067, + 0.07469207793474197, + 0.06652748584747314, + 0.0659877210855484, + 0.026848204433918, + 0.059922438114881516 + ], + "credit_map": [ + 0.013503842987120152, + 0.017096493393182755, + 0.012247119098901749, + 0.013428530655801296, + 0.013097154907882214, + 0.012818039394915104, + 0.014345690608024597, + 0.016275331377983093, + 0.003133149817585945, + 0.00686563178896904, + 0.004055729601532221, + 0.0038543411064893007, + 0.00708318967372179, + 0.0024716018233448267, + 0.010895829647779465, + 0.0031658841762691736, + 0.04007330164313316, + 0.041275639086961746, + 0.09412572532892227, + 0.020731084048748016, + 0.06710276752710342, + 0.02759096957743168, + 0.045662667602300644, + 0.019223840907216072, + 0.004742915742099285, + 0.004823903553187847, + 0.006415320560336113, + 0.011815459467470646, + 0.008891608566045761, + 0.0044471463188529015, + 0.007770069874823093, + 0.007874269969761372, + 0.03261358290910721, + 0.01049946341663599, + 0.01686408370733261, + 0.018743742257356644, + 0.017706435173749924, + 0.013026121072471142, + 0.017350096255540848, + 0.012314327992498875, + 0.012731361202895641, + 0.007352430373430252, + 0.00887790136039257, + 0.007106155157089233, + 0.013708231970667839, + 0.008712499402463436, + 0.010108140297234058, + 0.0060953604988753796, + 0.007757069543004036, + 0.005544958170503378, + 0.009012429974973202, + 0.011479291133582592, + 0.010283784940838814, + 0.007278892677277327, + 0.007868937216699123, + 0.0073021212592720985, + 0.00571503397077322, + 0.005338615272194147, + 0.009184890426695347, + 0.007495117373764515, + 0.013398151844739914, + 0.008404255844652653, + 0.008496223948895931, + 0.007955432869493961, + 0.002525224583223462, + 0.001971067162230611, + 0.0032499448861926794, + 0.0025284436997026205, + 0.0039058509282767773, + 0.004092215560376644, + 0.0027646077796816826, + 0.005810847505927086, + 0.006700261030346155, + 0.0055336011573672295, + 0.00725577725097537, + 0.0061327810399234295, + 0.014431789517402649, + 0.004798024892807007, + 0.009107586927711964, + 0.005962618626654148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.280563622713089 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20159615576267242 + ], + "opposite_logprob_tokens": [ + -1.7015961408615112 + ], + "logprob_reward": -0.20159615576267242, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20159615576267242, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10485228896141052, + 0.18959935009479523, + 0.14907799661159515, + 0.09264639765024185, + 0.11054326593875885, + 0.030785787850618362, + 0.06990192830562592, + 0.10176187753677368, + 0.07114865630865097, + 0.07968241721391678 + ], + "credit_map": [ + 0.012637658044695854, + 0.01918656751513481, + 0.010141505859792233, + 0.009889238514006138, + 0.0129281310364604, + 0.010193823836743832, + 0.013223652727901936, + 0.016651716083288193, + 0.013238423503935337, + 0.022793874144554138, + 0.014918938279151917, + 0.018942449241876602, + 0.03411658853292465, + 0.01178512442857027, + 0.059290625154972076, + 0.014513319358229637, + 0.01712430641055107, + 0.012774726375937462, + 0.038646455854177475, + 0.006997264456003904, + 0.03193259239196777, + 0.009950301609933376, + 0.021705808117985725, + 0.009946527890861034, + 0.00911638792604208, + 0.006795897614210844, + 0.010575570166110992, + 0.012861744500696659, + 0.018405461683869362, + 0.0089683523401618, + 0.015856217592954636, + 0.010066773742437363, + 0.013538386672735214, + 0.007520449813455343, + 0.016319533810019493, + 0.009438980370759964, + 0.021496856585144997, + 0.012846669182181358, + 0.018285613507032394, + 0.011096776463091373, + 0.004878433886915445, + 0.002406650222837925, + 0.0036067983601242304, + 0.005025291815400124, + 0.003854970447719097, + 0.0028164030518382788, + 0.0037513847928494215, + 0.004445855040103197, + 0.007950570434331894, + 0.005406601820141077, + 0.013041310012340546, + 0.008300995454192162, + 0.010037476196885109, + 0.008462153375148773, + 0.00883611012250185, + 0.007866710424423218, + 0.007651308085769415, + 0.00724960258230567, + 0.010591099038720131, + 0.013066536746919155, + 0.014623388648033142, + 0.020410221070051193, + 0.012628566473722458, + 0.015541154891252518, + 0.007654088549315929, + 0.004948201589286327, + 0.009389991872012615, + 0.008654060773551464, + 0.014285735785961151, + 0.007809730246663094, + 0.008192749693989754, + 0.01021409872919321, + 0.010507911443710327, + 0.0074595119804143906, + 0.009774420410394669, + 0.011478095315396786, + 0.010170153342187405, + 0.009576298296451569, + 0.011180944740772247, + 0.009535086341202259 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.20159615576267242 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067511084489524364 + ], + "opposite_logprob_tokens": [ + -5.00675106048584 + ], + "logprob_reward": -0.0067511084489524364, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067511084489524364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09062797576189041, + 0.10222483426332474, + 0.15271948277950287, + 0.07227522134780884, + 0.22751453518867493, + 0.1007392629981041, + 0.053416937589645386, + 0.04987948760390282, + 0.04773692041635513, + 0.10286541283130646 + ], + "credit_map": [ + 0.013100587762892246, + 0.010592730715870857, + 0.010566678829491138, + 0.010564588010311127, + 0.0116200502961874, + 0.007874244824051857, + 0.015002526342868805, + 0.011306564323604107, + 0.007050493732094765, + 0.012175087817013264, + 0.008912649936974049, + 0.012980478815734386, + 0.016787322238087654, + 0.007069636136293411, + 0.026010455563664436, + 0.01123870350420475, + 0.018612874671816826, + 0.020013708621263504, + 0.016715634614229202, + 0.011373322457075119, + 0.03268546983599663, + 0.010365809313952923, + 0.03081575781106949, + 0.012136892415583134, + 0.008547492325305939, + 0.0062309978529810905, + 0.011395275592803955, + 0.014802800491452217, + 0.00804754439741373, + 0.006602924317121506, + 0.006157269701361656, + 0.010490911081433296, + 0.03255348652601242, + 0.015814494341611862, + 0.02749566361308098, + 0.021758340299129486, + 0.04415970668196678, + 0.021628834307193756, + 0.03571738675236702, + 0.028386596590280533, + 0.01660894975066185, + 0.008026201277971268, + 0.01388637162744999, + 0.013441786170005798, + 0.010501069016754627, + 0.00956406444311142, + 0.013716924004256725, + 0.01499389111995697, + 0.004692739807069302, + 0.005105412099510431, + 0.008329386822879314, + 0.004850470926612616, + 0.012841005809605122, + 0.006346867419779301, + 0.0059805926866829395, + 0.005270455963909626, + 0.005929029546678066, + 0.0045961057767271996, + 0.006277837324887514, + 0.003990625496953726, + 0.011520612053573132, + 0.005014325492084026, + 0.005285772494971752, + 0.007265176624059677, + 0.0055566937662661076, + 0.0046896119602024555, + 0.005738089792430401, + 0.0047033545561134815, + 0.013821041211485863, + 0.003465111833065748, + 0.006503273267298937, + 0.0032597375102341175, + 0.012282751500606537, + 0.00733278738334775, + 0.019239580258727074, + 0.013304186053574085, + 0.015917286276817322, + 0.012685153633356094, + 0.009916684590280056, + 0.012186978943645954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.0067511084489524364 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03372795134782791 + ], + "opposite_logprob_tokens": [ + -3.4087278842926025 + ], + "logprob_reward": -0.03372795134782791, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03372795134782791, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09597250074148178, + 0.0567932091653347, + 0.07304549217224121, + 0.07751871645450592, + 0.08638311177492142, + 0.19617953896522522, + 0.10003688186407089, + 0.0391555093228817, + 0.05781644955277443, + 0.21709860861301422 + ], + "credit_map": [ + 0.01089868787676096, + 0.013802515342831612, + 0.00958153698593378, + 0.00960127916187048, + 0.013503557071089745, + 0.00819273293018341, + 0.01779850386083126, + 0.012593683786690235, + 0.0044717625714838505, + 0.006387563422322273, + 0.005324693862348795, + 0.008868813514709473, + 0.008011344820261002, + 0.0035599961411207914, + 0.014845376834273338, + 0.005323657765984535, + 0.006851287093013525, + 0.012593726627528667, + 0.012419312261044979, + 0.005719675216823816, + 0.012725935317575932, + 0.006424469407647848, + 0.010236025787889957, + 0.006075060460716486, + 0.011343106627464294, + 0.006424640771001577, + 0.009914666414260864, + 0.01221790723502636, + 0.011287190020084381, + 0.007567327469587326, + 0.007497246377170086, + 0.011266635730862617, + 0.01636689528822899, + 0.00625318568199873, + 0.014822833240032196, + 0.009820319712162018, + 0.013731769286096096, + 0.0077291750349104404, + 0.010478704236447811, + 0.007180226966738701, + 0.022992417216300964, + 0.01775296777486801, + 0.024171385914087296, + 0.026625724509358406, + 0.02783321402966976, + 0.02459314651787281, + 0.026954885572195053, + 0.02525578998029232, + 0.011202854104340076, + 0.009794416837394238, + 0.016936691477894783, + 0.0072485655546188354, + 0.02916794829070568, + 0.007482536602765322, + 0.012616612017154694, + 0.005587266758084297, + 0.004060336854308844, + 0.0032783334609121084, + 0.005514222197234631, + 0.003383627161383629, + 0.008703883737325668, + 0.005882111378014088, + 0.0042580305598676205, + 0.004074965137988329, + 0.005962925963103771, + 0.004733677953481674, + 0.006387881003320217, + 0.005355184897780418, + 0.014491950161755085, + 0.007113055791705847, + 0.007375792600214481, + 0.006395974662154913, + 0.023528246209025383, + 0.016793707385659218, + 0.03239684924483299, + 0.024127203971147537, + 0.039685532450675964, + 0.02680928260087967, + 0.028482452034950256, + 0.025275327265262604 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 6 + ], + "selected_reward": -0.03372795134782791 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12704585492610931 + ], + "opposite_logprob_tokens": [ + -2.1270458698272705 + ], + "logprob_reward": -0.12704585492610931, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12704585492610931, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08735424280166626, + 0.06657813489437103, + 0.07991097122430801, + 0.051829226315021515, + 0.14190003275871277, + 0.21426568925380707, + 0.1862603724002838, + 0.06717207282781601, + 0.02476048655807972, + 0.07996879518032074 + ], + "credit_map": [ + 0.009271527640521526, + 0.01481303945183754, + 0.009044783189892769, + 0.008693233132362366, + 0.011601464822888374, + 0.00938353966921568, + 0.00936205219477415, + 0.01518460176885128, + 0.005683851428329945, + 0.01569960080087185, + 0.004674260504543781, + 0.006812435109168291, + 0.010239369235932827, + 0.004634252283722162, + 0.012911212630569935, + 0.005923152435570955, + 0.006875734776258469, + 0.00745428167283535, + 0.016540518030524254, + 0.00607918668538332, + 0.01733994670212269, + 0.0050249285995960236, + 0.015000812709331512, + 0.005595562048256397, + 0.008302279748022556, + 0.004356925841420889, + 0.006822200957685709, + 0.007434110157191753, + 0.007517414167523384, + 0.004255257546901703, + 0.007643267046660185, + 0.005497771315276623, + 0.017028408125042915, + 0.01002916693687439, + 0.015124348923563957, + 0.012200682424008846, + 0.025266343727707863, + 0.018314089626073837, + 0.02659762278199196, + 0.017339367419481277, + 0.02283405140042305, + 0.022007877007126808, + 0.02452748827636242, + 0.020528841763734818, + 0.039656683802604675, + 0.032066620886325836, + 0.029358383268117905, + 0.023285740986466408, + 0.017035312950611115, + 0.017422398552298546, + 0.02536884881556034, + 0.0173379834741354, + 0.04087240248918533, + 0.025366278365254402, + 0.026317190378904343, + 0.016539959236979485, + 0.006289833690971136, + 0.005771356634795666, + 0.007580254226922989, + 0.005978721659630537, + 0.012932296842336655, + 0.00784470234066248, + 0.008261663839221, + 0.012513245455920696, + 0.0023161207791417837, + 0.0021570331882685423, + 0.0033832762856036425, + 0.0032674765679985285, + 0.0039055293891578913, + 0.0036190354730933905, + 0.0029579519759863615, + 0.003154063830152154, + 0.008577059954404831, + 0.006358258426189423, + 0.013420382514595985, + 0.011044428683817387, + 0.012558226473629475, + 0.008176781237125397, + 0.010469509288668633, + 0.009364149533212185 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.12704585492610931 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, frames a lightweight wooden door, its surface revealing the fine grain of aged pine beneath a layer of cracked and flaking white paint, slightly ajar in a dark, worn frame. The scene is a dimly lit, silent corridor. A cool, gray light from an unseen window filters across the dusty wooden floorboards, illuminating the dancing dust motes in the still air. An unseen draft gently applies pressure to the door, causing it to swing silently inward on its tarnished brass hinges in a wide, ghostly arc, revealing more of the profound darkness of the room beyond. It then reverses direction, beginning a slow, hypnotic back-and-forth oscillation. With each passage, the faint, dry creak of the hinges softens as the door's momentum steadily fades, until it finally rests in complete silence, precisely returning to its original, slightly ajar position, creating a haunting and suspenseful atmosphere.", + "seed": 3388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_2543_seed_3388_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.17707177698612214, + "vlm_logprob": -0.17707177698612214, + "vlm_yes_no_margin": 2.275, + "vlm_token_credit": -0.17707177698612214, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2805551588535309 + ], + "opposite_logprob_tokens": [ + -1.6555551290512085 + ], + "logprob_reward": -0.2805551588535309, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2805551588535309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06410045176744461, + 0.04956474527716637, + 0.021982472389936447, + 0.14014215767383575, + 0.09683863818645477, + 0.0480220690369606, + 0.17961905896663666, + 0.21251052618026733, + 0.12469156086444855, + 0.06252828240394592 + ], + "credit_map": [ + 0.004380745347589254, + 0.011655199341475964, + 0.004608581308275461, + 0.013340504840016365, + 0.009820962324738503, + 0.005539739970117807, + 0.00968402810394764, + 0.0050706942565739155, + 0.0036874227225780487, + 0.004507801495492458, + 0.0029940460808575153, + 0.006477448623627424, + 0.01025467086583376, + 0.0034681297838687897, + 0.014010879211127758, + 0.004164345096796751, + 0.0018715029582381248, + 0.0028211174067109823, + 0.0022280781995505095, + 0.004143360070884228, + 0.0038904943503439426, + 0.002126747276633978, + 0.0032222489826381207, + 0.0016789245419204235, + 0.009166287258267403, + 0.013626720756292343, + 0.011427895165979862, + 0.03497319296002388, + 0.028205279260873795, + 0.010188207030296326, + 0.02215808816254139, + 0.010396486148238182, + 0.009386425837874413, + 0.008672409690916538, + 0.012433575466275215, + 0.017902467399835587, + 0.016739094629883766, + 0.01038429792970419, + 0.010976349003612995, + 0.010344009846448898, + 0.007119587156921625, + 0.005120652727782726, + 0.007201306987553835, + 0.006297789514064789, + 0.008757059462368488, + 0.0047444915398955345, + 0.004728248808532953, + 0.004052930511534214, + 0.015070642344653606, + 0.01873644068837166, + 0.026810510084033012, + 0.02970992587506771, + 0.034229159355163574, + 0.01140613853931427, + 0.03388600051403046, + 0.00977024994790554, + 0.019292650744318962, + 0.040949515998363495, + 0.0281313918530941, + 0.02802741527557373, + 0.030931614339351654, + 0.018620721995830536, + 0.02707279846072197, + 0.01948443241417408, + 0.009041295386850834, + 0.007708302233368158, + 0.0282723568379879, + 0.015570273622870445, + 0.03314647451043129, + 0.009645442478358746, + 0.01405379083007574, + 0.0072536226361989975, + 0.004319167695939541, + 0.005754636600613594, + 0.006451782304793596, + 0.0072127352468669415, + 0.009807386435568333, + 0.014183424413204193, + 0.008069961331784725, + 0.006729189306497574 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 3 + ], + "selected_reward": -0.2805551588535309 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289393126964569 + ], + "opposite_logprob_tokens": [ + -1.0539393424987793 + ], + "logprob_reward": -0.4289393126964569, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289393126964569, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19298721849918365, + 0.10187491029500961, + 0.09736649692058563, + 0.1353074014186859, + 0.08077261596918106, + 0.05090760067105293, + 0.12634828686714172, + 0.06929288804531097, + 0.0922199934720993, + 0.05292254313826561 + ], + "credit_map": [ + 0.01511178445070982, + 0.028649859130382538, + 0.018453873693943024, + 0.034384362399578094, + 0.028024934232234955, + 0.01885855756700039, + 0.03249020874500275, + 0.017013657838106155, + 0.007077495567500591, + 0.016616739332675934, + 0.00891156680881977, + 0.025238288566470146, + 0.01329225767403841, + 0.009268348105251789, + 0.014497573487460613, + 0.006972644943743944, + 0.009169560857117176, + 0.013776694424450397, + 0.012397358193993568, + 0.014437973499298096, + 0.012081436812877655, + 0.011258633807301521, + 0.011592217721045017, + 0.012652615085244179, + 0.007823396474123001, + 0.014539710246026516, + 0.010198858566582203, + 0.02174435742199421, + 0.03548917546868324, + 0.011828899383544922, + 0.02433422952890396, + 0.009348787367343903, + 0.005889921449124813, + 0.00853796023875475, + 0.007390615530312061, + 0.016321662813425064, + 0.01878664828836918, + 0.006836395710706711, + 0.009907920844852924, + 0.007101486902683973, + 0.0081435926258564, + 0.0058475518599152565, + 0.006737279240041971, + 0.007728496566414833, + 0.007390595972537994, + 0.0048453183844685555, + 0.004922282882034779, + 0.0052924854680895805, + 0.018659887835383415, + 0.0165322944521904, + 0.018486451357603073, + 0.0173155777156353, + 0.012762083671987057, + 0.008552190847694874, + 0.018636789172887802, + 0.015403011813759804, + 0.006020840257406235, + 0.01246404368430376, + 0.010835989378392696, + 0.011160627007484436, + 0.009668050333857536, + 0.005441346205770969, + 0.009162051603198051, + 0.004539933521300554, + 0.0065118721686303616, + 0.008405729196965694, + 0.009227609261870384, + 0.013601219281554222, + 0.014197742566466331, + 0.014208480715751648, + 0.019268250092864037, + 0.006799089256674051, + 0.004634968936443329, + 0.004871294368058443, + 0.006527808960527182, + 0.00724619859829545, + 0.009436611086130142, + 0.007495509926229715, + 0.008550391532480717, + 0.004159759264439344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 6 + ], + "selected_reward": -0.4289393126964569 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219980686903 + ], + "opposite_logprob_tokens": [ + -4.264219760894775 + ], + "logprob_reward": -0.014219980686903, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219980686903, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10469557344913483, + 0.06009388715028763, + 0.07451976090669632, + 0.18590036034584045, + 0.08299092203378677, + 0.10685855150222778, + 0.08724615722894669, + 0.10828753560781479, + 0.10162056982517242, + 0.08778662234544754 + ], + "credit_map": [ + 0.009161832742393017, + 0.015591463074088097, + 0.01042504794895649, + 0.016526274383068085, + 0.014947343617677689, + 0.00954664871096611, + 0.014951766468584538, + 0.013545196503400803, + 0.0037808313500136137, + 0.012458642944693565, + 0.0045481049455702305, + 0.009827260859310627, + 0.00951885525137186, + 0.0054711271077394485, + 0.009711501188576221, + 0.0047775558196008205, + 0.0074464972130954266, + 0.00781532097607851, + 0.009850861504673958, + 0.013016526587307453, + 0.009372526779770851, + 0.00820097979158163, + 0.008348381146788597, + 0.01046866923570633, + 0.0127544105052948, + 0.018850842490792274, + 0.013463628478348255, + 0.030946291983127594, + 0.048235055059194565, + 0.01569829322397709, + 0.030866973102092743, + 0.015084854327142239, + 0.006054732482880354, + 0.007776953279972076, + 0.00826916191726923, + 0.014644830487668514, + 0.015092258341610432, + 0.009820971637964249, + 0.010085845366120338, + 0.011246164329349995, + 0.011176158674061298, + 0.008963200263679028, + 0.013645472005009651, + 0.016167761757969856, + 0.024859940633177757, + 0.009475773200392723, + 0.016129184514284134, + 0.006441067438572645, + 0.010774686001241207, + 0.014073748141527176, + 0.014230038970708847, + 0.012645430862903595, + 0.010455993004143238, + 0.006966457702219486, + 0.008306095376610756, + 0.009793706238269806, + 0.007648142520338297, + 0.01614253968000412, + 0.012484150007367134, + 0.016382642090320587, + 0.013613645918667316, + 0.01370641402900219, + 0.01768527366220951, + 0.010624725371599197, + 0.006394478492438793, + 0.006772838998585939, + 0.018919436261057854, + 0.01596599631011486, + 0.02337731048464775, + 0.010462104342877865, + 0.015280919149518013, + 0.004447488114237785, + 0.006000367924571037, + 0.006648726295679808, + 0.0110054612159729, + 0.013093355111777782, + 0.015193406492471695, + 0.015357556752860546, + 0.0128892557695508, + 0.007598488591611385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.014219980686903 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04870473966002464 + ], + "opposite_logprob_tokens": [ + -3.0487046241760254 + ], + "logprob_reward": -0.04870473966002464, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04870473966002464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1174432709813118, + 0.03628015145659447, + 0.032553475350141525, + 0.1100534126162529, + 0.08183068037033081, + 0.11208634823560715, + 0.18111911416053772, + 0.07683337479829788, + 0.19019468128681183, + 0.06160552427172661 + ], + "credit_map": [ + 0.009560898877680302, + 0.024598991498351097, + 0.008984586223959923, + 0.022804535925388336, + 0.015311913564801216, + 0.009676486253738403, + 0.01683550886809826, + 0.009670346044003963, + 0.0026486332062631845, + 0.003928528632968664, + 0.0033882195129990578, + 0.005953685380518436, + 0.006262208800762892, + 0.002945432672277093, + 0.007107565179467201, + 0.004045877605676651, + 0.003596123307943344, + 0.0032888648565858603, + 0.004192466381937265, + 0.004795600660145283, + 0.004793883301317692, + 0.00299850944429636, + 0.004975794814527035, + 0.0039122323505580425, + 0.006830963306128979, + 0.010196467861533165, + 0.008392173796892166, + 0.021942095831036568, + 0.024311702698469162, + 0.009116633795201778, + 0.019660847261548042, + 0.009602533653378487, + 0.0068204281851649284, + 0.007987264543771744, + 0.008089314214885235, + 0.01584755629301071, + 0.016071291640400887, + 0.00906780268996954, + 0.009380124509334564, + 0.008566894568502903, + 0.012690328992903233, + 0.013126353733241558, + 0.01590030826628208, + 0.012563538737595081, + 0.020465370267629623, + 0.009871600195765495, + 0.01668732427060604, + 0.010781529359519482, + 0.032926324754953384, + 0.021829383447766304, + 0.0320969820022583, + 0.022163936868309975, + 0.017352478578686714, + 0.013475597836077213, + 0.015637559816241264, + 0.025636853650212288, + 0.008475502021610737, + 0.008893093094229698, + 0.00966602936387062, + 0.011229985393583775, + 0.012974487617611885, + 0.006470398046076298, + 0.0132995480671525, + 0.005824328400194645, + 0.012636580504477024, + 0.01207314245402813, + 0.02049366943538189, + 0.03203083574771881, + 0.025836898013949394, + 0.032799433916807175, + 0.035669032484292984, + 0.018655072897672653, + 0.004759049043059349, + 0.004029612522572279, + 0.008171671070158482, + 0.006799096241593361, + 0.010039037093520164, + 0.015607559122145176, + 0.005787268280982971, + 0.00641222856938839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 6, + 0 + ], + "selected_reward": -0.04870473966002464 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11293969303369522 + ], + "opposite_logprob_tokens": [ + -2.2379395961761475 + ], + "logprob_reward": -0.11293969303369522, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11293969303369522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09628573805093765, + 0.04151751101016998, + 0.04193924739956856, + 0.1618623435497284, + 0.1254504919052124, + 0.05580814555287361, + 0.10667765140533447, + 0.13679960370063782, + 0.14429661631584167, + 0.08936262875795364 + ], + "credit_map": [ + 0.010943721979856491, + 0.014208264648914337, + 0.010202902369201183, + 0.014743325300514698, + 0.011505920439958572, + 0.00921943224966526, + 0.015473836101591587, + 0.009988335892558098, + 0.003167273011058569, + 0.00969435553997755, + 0.004086126573383808, + 0.007146607153117657, + 0.005235916003584862, + 0.0028693191707134247, + 0.005596852861344814, + 0.003721063258126378, + 0.0029792231507599354, + 0.006199695635586977, + 0.0034190472215414047, + 0.008911839686334133, + 0.007168003357946873, + 0.0035680995788425207, + 0.00612504780292511, + 0.0035682921297848225, + 0.010603288188576698, + 0.017444361001253128, + 0.01428581029176712, + 0.03527504578232765, + 0.03645540401339531, + 0.012610799632966518, + 0.025630788877606392, + 0.009556838311254978, + 0.012074965052306652, + 0.011327007785439491, + 0.013199874199926853, + 0.022693907842040062, + 0.02708602510392666, + 0.01092789601534605, + 0.01599043793976307, + 0.012150371447205544, + 0.0067350538447499275, + 0.006589076481759548, + 0.008987100794911385, + 0.009713883511722088, + 0.006938074249774218, + 0.004602533765137196, + 0.007216231431812048, + 0.00502619007602334, + 0.014044362120330334, + 0.01024218462407589, + 0.01542411744594574, + 0.017430312931537628, + 0.016109691932797432, + 0.005950137507170439, + 0.019292360171675682, + 0.00818448606878519, + 0.012633280828595161, + 0.02901376597583294, + 0.019536461681127548, + 0.02074465900659561, + 0.020533768460154533, + 0.007554921787232161, + 0.017600908875465393, + 0.009181838482618332, + 0.009299959056079388, + 0.009092950262129307, + 0.020398525521159172, + 0.01939498819410801, + 0.027705801650881767, + 0.017573535442352295, + 0.02862958423793316, + 0.012201271951198578, + 0.006027872208505869, + 0.006608542054891586, + 0.011887476779520512, + 0.010642080567777157, + 0.015594344586133957, + 0.01638394594192505, + 0.013367668725550175, + 0.008850703947246075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 7 + ], + "selected_reward": -0.11293969303369522 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, frames a lightweight wooden door, its surface revealing the fine grain of aged pine beneath a layer of cracked and flaking white paint, slightly ajar in a dark, worn frame. The scene is a dimly lit, silent corridor. A cool, gray light from an unseen window filters across the dusty wooden floorboards, illuminating the dancing dust motes in the still air. An unseen draft gently applies pressure to the door, causing it to swing silently inward on its tarnished brass hinges in a wide, ghostly arc, revealing more of the profound darkness of the room beyond. It then reverses direction, beginning a slow, hypnotic back-and-forth oscillation. With each passage, the faint, dry creak of the hinges softens as the door's momentum steadily fades, until it finally rests in complete silence, precisely returning to its original, slightly ajar position, creating a haunting and suspenseful atmosphere.", + "seed": 3389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_2543_seed_3389_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.11944999769330025, + "vlm_logprob": -0.11944999769330025, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.11944999769330025, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.21929877996444702 + ], + "opposite_logprob_tokens": [ + -1.8442988395690918 + ], + "logprob_reward": -0.21929877996444702, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.21929877996444702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1703682243824005, + 0.12920241057872772, + 0.10756345838308334, + 0.05086009204387665, + 0.06133536994457245, + 0.018579497933387756, + 0.16497570276260376, + 0.2528057098388672, + 0.034857768565416336, + 0.009451826103031635 + ], + "credit_map": [ + 0.009588646702468395, + 0.017738599330186844, + 0.014339893124997616, + 0.02652806043624878, + 0.030815627425909042, + 0.008871735073626041, + 0.04944543167948723, + 0.013040228746831417, + 0.006938803009688854, + 0.012437460944056511, + 0.007229556795209646, + 0.03290732204914093, + 0.01662597805261612, + 0.00539248576387763, + 0.03869255632162094, + 0.008978236466646194, + 0.011381545104086399, + 0.03580137714743614, + 0.007548396475613117, + 0.013189239427447319, + 0.016465622931718826, + 0.005678522866219282, + 0.01060721930116415, + 0.00689152255654335, + 0.002269666874781251, + 0.006385435350239277, + 0.0058136554434895515, + 0.017367668449878693, + 0.005069561768323183, + 0.0016971076838672161, + 0.009491701610386372, + 0.0027652920689433813, + 0.0065575395710766315, + 0.009121939539909363, + 0.0063643986359238625, + 0.007485225796699524, + 0.007461192552000284, + 0.004044157452881336, + 0.013545168563723564, + 0.00675574317574501, + 0.002544945105910301, + 0.0018660599598661065, + 0.00414531072601676, + 0.002266779076308012, + 0.0017453574109822512, + 0.0019112827721983194, + 0.001946820761077106, + 0.00215293955989182, + 0.012429364956915379, + 0.014340955764055252, + 0.016278987750411034, + 0.012997587211430073, + 0.026614004746079445, + 0.04607313498854637, + 0.015757745131850243, + 0.02048390917479992, + 0.013220188207924366, + 0.052065733820199966, + 0.01565374806523323, + 0.015854131430387497, + 0.01909102126955986, + 0.08148481696844101, + 0.015176688320934772, + 0.04025934264063835, + 0.0035643638111650944, + 0.004268820863217115, + 0.005495632532984018, + 0.0027621411718428135, + 0.005629654508084059, + 0.00364688690751791, + 0.005653567612171173, + 0.003836697433143854, + 0.0014126485912129283, + 0.0008779190247878432, + 0.001353709609247744, + 0.0009070672676898539, + 0.001537035102955997, + 0.0011349086416885257, + 0.001327641773968935, + 0.000900895509403199 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.21929877996444702 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707734107971191 + ], + "opposite_logprob_tokens": [ + -2.127077341079712 + ], + "logprob_reward": -0.12707734107971191, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707734107971191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05703339725732803, + 0.14663942158222198, + 0.1890869289636612, + 0.0198172889649868, + 0.040002401918172836, + 0.14443062245845795, + 0.15433426201343536, + 0.21032239496707916, + 0.017934219911694527, + 0.020399102941155434 + ], + "credit_map": [ + 0.004021567292511463, + 0.006894598249346018, + 0.0047182803973555565, + 0.008547368459403515, + 0.009131988510489464, + 0.00384361669421196, + 0.015127227641642094, + 0.004748750012367964, + 0.009307709522545338, + 0.01645294949412346, + 0.011973217129707336, + 0.03439955413341522, + 0.017469419166445732, + 0.0053449030965566635, + 0.0379946194589138, + 0.013697044923901558, + 0.023503514006733894, + 0.07451975345611572, + 0.006957599893212318, + 0.0191473588347435, + 0.029749669134616852, + 0.009725771844387054, + 0.01403703261166811, + 0.011446220800280571, + 0.001558941206894815, + 0.0032514033373445272, + 0.0022965348325669765, + 0.002264035400003195, + 0.0034993221051990986, + 0.001551315886899829, + 0.002878824481740594, + 0.0025169113650918007, + 0.003148987190797925, + 0.007410367485135794, + 0.004389118868857622, + 0.006270811427384615, + 0.006525529082864523, + 0.0032217674888670444, + 0.0048585133627057076, + 0.004177305847406387, + 0.02630234695971012, + 0.0060948156751692295, + 0.02410735934972763, + 0.010282391682267189, + 0.008953983895480633, + 0.039843879640102386, + 0.008825049735605717, + 0.020020807161927223, + 0.010158436372876167, + 0.01141397189348936, + 0.011102098040282726, + 0.008964918553829193, + 0.029552647843956947, + 0.05070418864488602, + 0.011240208521485329, + 0.021197790279984474, + 0.011214915663003922, + 0.04813741520047188, + 0.008570295758545399, + 0.009831296280026436, + 0.01567142829298973, + 0.07193504273891449, + 0.01027172151952982, + 0.034690290689468384, + 0.00355928554199636, + 0.0020854466129094362, + 0.002895526820793748, + 0.0014140326529741287, + 0.002166741993278265, + 0.001677436986938119, + 0.0022094144951552153, + 0.0019263334106653929, + 0.0016427416121587157, + 0.0022379769943654537, + 0.001397938234731555, + 0.0011799705680459738, + 0.006780191324651241, + 0.0037852912209928036, + 0.0017815844621509314, + 0.0015934089897200465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.12707734107971191 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011092189699411392 + ], + "opposite_logprob_tokens": [ + -4.511092185974121 + ], + "logprob_reward": -0.011092189699411392, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011092189699411392, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20680738985538483, + 0.09271150082349777, + 0.09709236025810242, + 0.029901983216404915, + 0.05826706811785698, + 0.1785292774438858, + 0.1960129588842392, + 0.10289648920297623, + 0.024360522627830505, + 0.013420470058918 + ], + "credit_map": [ + 0.011944680474698544, + 0.021587137132883072, + 0.017300797626376152, + 0.034394823014736176, + 0.03668751195073128, + 0.011469739489257336, + 0.057415761053562164, + 0.016006944701075554, + 0.006866055075079203, + 0.008992933668196201, + 0.00714325113222003, + 0.01666453294456005, + 0.012508345767855644, + 0.007665029726922512, + 0.022705551236867905, + 0.01016579382121563, + 0.01252311933785677, + 0.04146306961774826, + 0.0037725185975432396, + 0.009587766602635384, + 0.014704862609505653, + 0.004731711931526661, + 0.0057994103990495205, + 0.004509902559220791, + 0.0025495807640254498, + 0.0036220704205334187, + 0.004025258123874664, + 0.0034895355347543955, + 0.0063398247584700584, + 0.002618726110085845, + 0.003852525260299444, + 0.0034044620115309954, + 0.004779924638569355, + 0.01028480101376772, + 0.00776264164596796, + 0.00779091939330101, + 0.006785944104194641, + 0.004237114917486906, + 0.009648054838180542, + 0.0069776615127921104, + 0.033095959573984146, + 0.00809844397008419, + 0.037933144718408585, + 0.014170631766319275, + 0.010928581468760967, + 0.03863214701414108, + 0.013031675480306149, + 0.022638704627752304, + 0.012704630382359028, + 0.017057809978723526, + 0.013258075341582298, + 0.010862275026738644, + 0.040767181664705276, + 0.062263328582048416, + 0.014583119191229343, + 0.02451653592288494, + 0.006125204265117645, + 0.02064145915210247, + 0.004784532822668552, + 0.005338125396519899, + 0.007245172746479511, + 0.03527813404798508, + 0.006759142503142357, + 0.0167247224599123, + 0.0035393587313592434, + 0.002607687609270215, + 0.00309391668997705, + 0.0027508079074323177, + 0.002855981234461069, + 0.0032152917701750994, + 0.0038697849959135056, + 0.002427695319056511, + 0.0016568641876801848, + 0.0016842836048454046, + 0.0014340050984174013, + 0.0011562186991795897, + 0.002365201013162732, + 0.001876060152426362, + 0.0017932784976437688, + 0.001454558689147234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.011092189699411392 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014210343360900879 + ], + "opposite_logprob_tokens": [ + -4.264210224151611 + ], + "logprob_reward": -0.014210343360900879, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014210343360900879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10105953365564346, + 0.14696760475635529, + 0.11101717501878738, + 0.04040800407528877, + 0.11031937599182129, + 0.24855586886405945, + 0.03116135112941265, + 0.05883248522877693, + 0.10051153600215912, + 0.05116705596446991 + ], + "credit_map": [ + 0.009754512459039688, + 0.013727257028222084, + 0.01324416883289814, + 0.013126731850206852, + 0.013747063465416431, + 0.007943212054669857, + 0.016549039632081985, + 0.012967551127076149, + 0.008814338594675064, + 0.014387007802724838, + 0.01352867390960455, + 0.031686294823884964, + 0.019260631874203682, + 0.006108869332820177, + 0.03996006399393082, + 0.013221712782979012, + 0.013941139914095402, + 0.0455005057156086, + 0.004279447719454765, + 0.011857685633003712, + 0.016751481220126152, + 0.005678344517946243, + 0.007804068271070719, + 0.005204500164836645, + 0.0027909239288419485, + 0.005876658484339714, + 0.005262947641313076, + 0.006169862113893032, + 0.007176960352808237, + 0.0028914539143443108, + 0.005277964752167463, + 0.00496123218908906, + 0.011293619871139526, + 0.0183956827968359, + 0.01592610962688923, + 0.015374129638075829, + 0.014151773415505886, + 0.00935237854719162, + 0.014080189168453217, + 0.011745491065084934, + 0.04743734374642372, + 0.00949510931968689, + 0.07341540604829788, + 0.022566137835383415, + 0.013758281245827675, + 0.03357023745775223, + 0.019593264907598495, + 0.028720077127218246, + 0.0028310795314610004, + 0.004153032321482897, + 0.0034804295282810926, + 0.005493568256497383, + 0.0048807463608682156, + 0.0034907651133835316, + 0.0042502982541918755, + 0.0025814336258918047, + 0.0041343094781041145, + 0.010135261341929436, + 0.005643624346703291, + 0.006876213941723108, + 0.0053755491971969604, + 0.01296992227435112, + 0.00746590830385685, + 0.006231692619621754, + 0.018590213730931282, + 0.011489545926451683, + 0.019107550382614136, + 0.008381073363125324, + 0.009809145703911781, + 0.00993840117007494, + 0.01129868719726801, + 0.01189691573381424, + 0.004301932640373707, + 0.004641904961317778, + 0.005429997108876705, + 0.004745077807456255, + 0.010209386236965656, + 0.007654253859072924, + 0.008751346729695797, + 0.00543315801769495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.014210343360900879 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22557133436203003 + ], + "opposite_logprob_tokens": [ + -1.6005713939666748 + ], + "logprob_reward": -0.22557133436203003, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22557133436203003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1100902408361435, + 0.13724805414676666, + 0.12572376430034637, + 0.049707524478435516, + 0.08647847920656204, + 0.143980011343956, + 0.19537433981895447, + 0.0850764736533165, + 0.05241301655769348, + 0.013908067718148232 + ], + "credit_map": [ + 0.007315563969314098, + 0.01352149248123169, + 0.009963639080524445, + 0.014116339385509491, + 0.018285229802131653, + 0.007344375364482403, + 0.031163090839982033, + 0.008380508981645107, + 0.009672578424215317, + 0.020305046811699867, + 0.014692491851747036, + 0.025418046861886978, + 0.01598784700036049, + 0.006712287664413452, + 0.02919192984700203, + 0.015267823822796345, + 0.013353063724935055, + 0.050190623849630356, + 0.006463530007749796, + 0.01394215039908886, + 0.018680650740861893, + 0.005631599575281143, + 0.010479805059731007, + 0.006982340011745691, + 0.003643942531198263, + 0.007406050339341164, + 0.005974938161671162, + 0.009243864566087723, + 0.007334472145885229, + 0.003321955446153879, + 0.008271087892353535, + 0.004511210136115551, + 0.00663788104429841, + 0.021229906007647514, + 0.008041026070713997, + 0.009313124231994152, + 0.015693265944719315, + 0.006862875539809465, + 0.010198253206908703, + 0.008502152748405933, + 0.02391047775745392, + 0.0077229393646121025, + 0.007303968537598848, + 0.009914586320519447, + 0.010127370245754719, + 0.05421795696020126, + 0.00951943825930357, + 0.021263273432850838, + 0.010914653539657593, + 0.008824685588479042, + 0.0102393738925457, + 0.012713031843304634, + 0.034314606338739395, + 0.07522936910390854, + 0.011550741270184517, + 0.0315878726541996, + 0.005409983918070793, + 0.0203135646879673, + 0.0034749978221952915, + 0.0042223394848406315, + 0.007202729117125273, + 0.028045644983649254, + 0.003695415100082755, + 0.012711797840893269, + 0.0071954443119466305, + 0.006659321486949921, + 0.005737834144383669, + 0.005724212620407343, + 0.007354240398854017, + 0.006294431164860725, + 0.008088620379567146, + 0.0053589134477078915, + 0.0014489191817119718, + 0.001351335784420371, + 0.0014212599489837885, + 0.0008861172827892005, + 0.003952429164201021, + 0.0022104571107774973, + 0.0013581524835899472, + 0.0012793968198820949 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.22557133436203003 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, medium shot, held at eye level, frames a lightweight wooden door, its surface revealing the fine grain of aged pine beneath a layer of cracked and flaking white paint, slightly ajar in a dark, worn frame. The scene is a dimly lit, silent corridor. A cool, gray light from an unseen window filters across the dusty wooden floorboards, illuminating the dancing dust motes in the still air. An unseen draft gently applies pressure to the door, causing it to swing silently inward on its tarnished brass hinges in a wide, ghostly arc, revealing more of the profound darkness of the room beyond. It then reverses direction, beginning a slow, hypnotic back-and-forth oscillation. With each passage, the faint, dry creak of the hinges softens as the door's momentum steadily fades, until it finally rests in complete silence, precisely returning to its original, slightly ajar position, creating a haunting and suspenseful atmosphere.", + "seed": 3390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_4/index_2543_seed_3390_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10023174155503511, + "vlm_logprob": -0.10023174155503511, + "vlm_yes_no_margin": 3.1, + "vlm_token_credit": -0.10023174155503511, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17203263938426971 + ], + "opposite_logprob_tokens": [ + -2.297032594680786 + ], + "logprob_reward": -0.17203263938426971, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.17203263938426971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11614353954792023, + 0.04814708232879639, + 0.11070740967988968, + 0.08157997578382492, + 0.05773264169692993, + 0.03918188065290451, + 0.236429363489151, + 0.23870012164115906, + 0.044740770012140274, + 0.026637237519025803 + ], + "credit_map": [ + 0.01360643096268177, + 0.023899588733911514, + 0.015917427837848663, + 0.02065252512693405, + 0.012309244833886623, + 0.006608462426811457, + 0.013581385836005211, + 0.009568477980792522, + 0.006072267424315214, + 0.007501905784010887, + 0.006037281826138496, + 0.006607910618185997, + 0.008289732038974762, + 0.0031447613146156073, + 0.006306792143732309, + 0.004186432342976332, + 0.012869508005678654, + 0.012192917987704277, + 0.016299346461892128, + 0.01590396650135517, + 0.02235010825097561, + 0.006505667231976986, + 0.016025763005018234, + 0.008560140617191792, + 0.008089105598628521, + 0.018006933853030205, + 0.0063944729045033455, + 0.007583444006741047, + 0.0120195047929883, + 0.0028153981547802687, + 0.021131189540028572, + 0.005539936013519764, + 0.00829130969941616, + 0.012667838484048843, + 0.004742708057165146, + 0.005492925178259611, + 0.006869925651699305, + 0.003559017786756158, + 0.006379624363034964, + 0.0097292959690094, + 0.003588627325370908, + 0.005373585037887096, + 0.005797799676656723, + 0.006834746338427067, + 0.004777115769684315, + 0.002883198903873563, + 0.004235788248479366, + 0.0056910207495093346, + 0.010737009346485138, + 0.02106865867972374, + 0.01151825487613678, + 0.020053986459970474, + 0.013136404566466808, + 0.07473335415124893, + 0.012728533707559109, + 0.07245316356420517, + 0.014153959229588509, + 0.013838880695402622, + 0.008348183706402779, + 0.008184544742107391, + 0.011178947053849697, + 0.08978961408138275, + 0.010988196358084679, + 0.08221781253814697, + 0.005014884751290083, + 0.006942656822502613, + 0.005043527577072382, + 0.004614898934960365, + 0.004588388372212648, + 0.006137175019830465, + 0.006745144724845886, + 0.005654098466038704, + 0.003226357977837324, + 0.00444186432287097, + 0.0026307208463549614, + 0.002009299583733082, + 0.002397005446255207, + 0.00458933413028717, + 0.002525638323277235, + 0.0048170192167162895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.17203263938426971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498385801911354 + ], + "opposite_logprob_tokens": [ + -2.9299838542938232 + ], + "logprob_reward": -0.05498385801911354, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498385801911354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11015887558460236, + 0.026699848473072052, + 0.0341225191950798, + 0.34769853949546814, + 0.025781871750950813, + 0.013365919701755047, + 0.03273311257362366, + 0.36850371956825256, + 0.026800479739904404, + 0.014135085046291351 + ], + "credit_map": [ + 0.013210580684244633, + 0.01934053748846054, + 0.01592540368437767, + 0.01729453168809414, + 0.015355728566646576, + 0.006432648748159409, + 0.014977122657001019, + 0.007622325327247381, + 0.003469674149528146, + 0.004649274051189423, + 0.0036305526737123728, + 0.0035606578458100557, + 0.003560106037184596, + 0.0018475381657481194, + 0.002872150158509612, + 0.0031098956242203712, + 0.005454737693071365, + 0.00471879355609417, + 0.003615935333073139, + 0.004658313002437353, + 0.005380569025874138, + 0.0018585481448099017, + 0.004607372917234898, + 0.0038282505702227354, + 0.030545208603143692, + 0.08912244439125061, + 0.012013377621769905, + 0.02129059098660946, + 0.056926969438791275, + 0.006740176118910313, + 0.11686339974403381, + 0.014196386560797691, + 0.002543635666370392, + 0.004045342095196247, + 0.004217693116515875, + 0.005937707144767046, + 0.002241709502413869, + 0.001421515829861164, + 0.0025460876058787107, + 0.0028281817212700844, + 0.001078161527402699, + 0.0020440390799194574, + 0.0012743864208459854, + 0.0015637048054486513, + 0.0023639483842998743, + 0.0010906063253059983, + 0.0027244314551353455, + 0.001226642751134932, + 0.0022774264216423035, + 0.006618110928684473, + 0.002668544417247176, + 0.004087650682777166, + 0.002267000963911414, + 0.009180375374853611, + 0.0027958943974226713, + 0.002838108455762267, + 0.013054675422608852, + 0.02496241219341755, + 0.008719107136130333, + 0.005420802626758814, + 0.02004314586520195, + 0.1615917682647705, + 0.010516172274947166, + 0.12419562041759491, + 0.0029673981480300426, + 0.003360491245985031, + 0.003071741433814168, + 0.002591085387393832, + 0.0031486342195421457, + 0.003807567059993744, + 0.004319378640502691, + 0.0035341838374733925, + 0.001676943153142929, + 0.0018843243597075343, + 0.0021048346534371376, + 0.0019328538328409195, + 0.0010571122402325273, + 0.0018778828671202064, + 0.001995439874008298, + 0.0016056954627856612 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 0 + ], + "selected_reward": -0.05498385801911354 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005958531983196735 + ], + "opposite_logprob_tokens": [ + -5.130958557128906 + ], + "logprob_reward": -0.005958531983196735, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005958531983196735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.02986639179289341, + 0.01289381179958582, + 0.008208140730857849, + 0.15151849389076233, + 0.013659153133630753, + 0.02428252063691616, + 0.03038448840379715, + 0.7073904871940613, + 0.01598113402724266, + 0.005815379787236452 + ], + "credit_map": [ + 0.002996977185830474, + 0.006390766240656376, + 0.0030334226321429014, + 0.005028851330280304, + 0.0036612635012716055, + 0.0017060622340068221, + 0.005056014284491539, + 0.0019930326379835606, + 0.0019092874135822058, + 0.0024261530488729477, + 0.0019856123253703117, + 0.002148277126252651, + 0.0015786918811500072, + 0.0006382692954503, + 0.001338101108558476, + 0.0008694198913872242, + 0.001042217598296702, + 0.0013283712323755026, + 0.0011220533633604646, + 0.0011725527001544833, + 0.0013671043561771512, + 0.00046296464279294014, + 0.001059869071468711, + 0.0006530079408548772, + 0.013341748155653477, + 0.0387447364628315, + 0.005502596497535706, + 0.0096044996753335, + 0.024550337344408035, + 0.00297777378000319, + 0.05062028020620346, + 0.006176525726914406, + 0.0017730040708556771, + 0.0031227560248225927, + 0.0015720700612291694, + 0.0020844193641096354, + 0.0012543242191895843, + 0.0007478379411622882, + 0.001263099373318255, + 0.0018416409147903323, + 0.002192428568378091, + 0.005821448750793934, + 0.0015881438739597797, + 0.0023562158457934856, + 0.0018881598953157663, + 0.0017710812389850616, + 0.003066571429371834, + 0.005598471499979496, + 0.0015490086516365409, + 0.002744735451415181, + 0.0016268000472337008, + 0.002362130442634225, + 0.0019410743843764067, + 0.009680471383035183, + 0.002110330853611231, + 0.008369936607778072, + 0.024979624897241592, + 0.04806579649448395, + 0.016730818897485733, + 0.009447823278605938, + 0.03850054368376732, + 0.3114760220050812, + 0.018924575299024582, + 0.23926526308059692, + 0.0018084724433720112, + 0.0014683647314086556, + 0.0019750306382775307, + 0.0016201824182644486, + 0.0019376900745555758, + 0.0023615355603396893, + 0.002538737142458558, + 0.0022711199708282948, + 0.00046335571096278727, + 0.0007505235844291747, + 0.0006663055391982198, + 0.0005968346376903355, + 0.0006049827206879854, + 0.0009286229615099728, + 0.0006148954853415489, + 0.001189859234727919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.005958531983196735 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608303189277649 + ], + "opposite_logprob_tokens": [ + -4.141083240509033 + ], + "logprob_reward": -0.01608303189277649, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608303189277649, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03228072449564934, + 0.015835534781217575, + 0.010749147273600101, + 0.20505155622959137, + 0.04333583638072014, + 0.015433951281011105, + 0.07334063202142715, + 0.5722456574440002, + 0.016110943630337715, + 0.015616015531122684 + ], + "credit_map": [ + 0.0034495946019887924, + 0.008073476143181324, + 0.0039024013094604015, + 0.004917139187455177, + 0.003974524326622486, + 0.0022684596478939056, + 0.003495733952149749, + 0.00219939649105072, + 0.0018669628771021962, + 0.0027007150929421186, + 0.002321514766663313, + 0.0018096968997269869, + 0.001837555319070816, + 0.0013570570154115558, + 0.0016532621812075377, + 0.0022887703962624073, + 0.001322498545050621, + 0.001220732112415135, + 0.0017360288184136152, + 0.0018112101824954152, + 0.0017401906661689281, + 0.0007144123082980514, + 0.0012606899254024029, + 0.0009433850063942373, + 0.01786903664469719, + 0.05274106562137604, + 0.007173600606620312, + 0.012245281599462032, + 0.03361596167087555, + 0.0038043803069740534, + 0.06932754069566727, + 0.008274685591459274, + 0.0060311732813715935, + 0.008831508457660675, + 0.005523200146853924, + 0.007063563913106918, + 0.003837536321952939, + 0.002491467399522662, + 0.004152360837906599, + 0.005405028350651264, + 0.0015287770656868815, + 0.0025311470963060856, + 0.001952211488969624, + 0.0025415492709726095, + 0.001634804648347199, + 0.000984117272309959, + 0.0023336983285844326, + 0.001927645760588348, + 0.003086645621806383, + 0.0069007533602416515, + 0.0031523839570581913, + 0.0034016689751297235, + 0.003500739810988307, + 0.028289010748267174, + 0.0029765882063657045, + 0.0220328476279974, + 0.02062820829451084, + 0.037619128823280334, + 0.013102272525429726, + 0.007985299453139305, + 0.03071194514632225, + 0.25143373012542725, + 0.015316136181354523, + 0.19544896483421326, + 0.001507705426774919, + 0.0014871180756017566, + 0.0014798977645114064, + 0.0013661329867318273, + 0.002169335028156638, + 0.002654130570590496, + 0.0024378166999667883, + 0.0030088063795119524, + 0.001968915108591318, + 0.002020652173087001, + 0.0018385376315563917, + 0.0011561027495190501, + 0.0018647238612174988, + 0.0026633120141923428, + 0.0016586967976763844, + 0.0024450747296214104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.01608303189277649 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a static medium shot framed at eye level?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a lightweight wooden door with aged pine grain and cracked white paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the door slightly ajar in a dark, worn frame within a dimly lit corridor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does cool gray light filter across dusty wooden floorboards illuminating dancing dust motes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the door swing silently inward on tarnished brass hinges in a wide arc before oscillating back and forth?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521006464958191 + ], + "opposite_logprob_tokens": [ + -1.5021007061004639 + ], + "logprob_reward": -0.2521006464958191, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521006464958191, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05114785209298134, + 0.01854727417230606, + 0.01792367920279503, + 0.20478512346744537, + 0.04252393916249275, + 0.02501879632472992, + 0.14386233687400818, + 0.4432067275047302, + 0.025481348857283592, + 0.027502991259098053 + ], + "credit_map": [ + 0.005593340378254652, + 0.009956044144928455, + 0.008466538041830063, + 0.00676762918010354, + 0.006250658538192511, + 0.003479549428448081, + 0.006590248551219702, + 0.00404384033754468, + 0.0022407101932913065, + 0.0026580130215734243, + 0.0030609623063355684, + 0.0023364750668406487, + 0.0023809587582945824, + 0.0014493619091808796, + 0.002045916859060526, + 0.0023748744279146194, + 0.0020041929092258215, + 0.0022310216445475817, + 0.0019479646580293775, + 0.0025254583451896906, + 0.0032584494911134243, + 0.0013374526752159, + 0.0025378603022545576, + 0.0020812773145735264, + 0.01738317497074604, + 0.05175493657588959, + 0.007808405440300703, + 0.012240974232554436, + 0.033603258430957794, + 0.004130774177610874, + 0.0692521259188652, + 0.008611448109149933, + 0.004813824780285358, + 0.006173302885144949, + 0.006502798292785883, + 0.004447276704013348, + 0.0053399293683469296, + 0.003297403920441866, + 0.006607098504900932, + 0.0053423019126057625, + 0.001975212711840868, + 0.0035802728962153196, + 0.0021334600169211626, + 0.0020232119131833315, + 0.0037315820809453726, + 0.002749665640294552, + 0.0031637134961783886, + 0.00566167663782835, + 0.008531923405826092, + 0.01412380114197731, + 0.009354130364954472, + 0.01694321259856224, + 0.00964678730815649, + 0.038634445518255234, + 0.012657152488827705, + 0.03397088125348091, + 0.015596650540828705, + 0.03142751380801201, + 0.011787346564233303, + 0.007373978849500418, + 0.024258187040686607, + 0.19201408326625824, + 0.01498360838741064, + 0.1457652896642685, + 0.0029748764354735613, + 0.002350614871829748, + 0.0036536052357405424, + 0.0030813561752438545, + 0.002904718043282628, + 0.0035498756915330887, + 0.0036408614832907915, + 0.003325438592582941, + 0.0028336774557828903, + 0.003856923896819353, + 0.0031530160922557116, + 0.0023251348175108433, + 0.0029143511783331633, + 0.004890132695436478, + 0.0031294396612793207, + 0.004400312434881926 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 6 + ], + "selected_reward": -0.2521006464958191 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_tensor([7688, 7688])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_tensor([7688, 7688])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..abadfaa243cfeb65747c674dccaf1c8c7299e0f6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_tensor([7688, 7688])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Told from a first-person perspective, the camera drifts forward with a slow, weightless, dreamlike stability through the silent, abyssal canyons of the submerged cityscape, a scene rendered with the surrealist aesthetic of Makoto Shinkai. The vibrant neon pinks and electric blues of flickering, water-distorting holographic advertisements cast long, oscillating reflections on the algae-stained glass and steel facades of the submerged skyscrapers, their light piercing the deep blue waters as an ethereal godlight. In the vast open spaces between these massive structures, enormous bell-shaped creatures, their bodies composed of a translucent, gelatinous material, perform a mesmerizing, silent courtship dance. Their delicate, fiber-optic tentacles trail behind them, emitting a soft, shifting inner light from within, pulsating rhythmically from ethereal cyan to warm fuchsia. As the camera glides past, one creature gracefully retracts its bell, slowly ascending as it spins around a vast, darkened building, while its partner responds by emitting bioluminescence in a synchronized, silent beat, their movements creating gentle, almost imperceptible currents in the clear water. The atmosphere is filled with profound, melancholic beauty and silent awe, the organic light of the creatures' ballet contrasting sharply with the ghostly, artificial glow of a lost civilization beneath the waves.", + "seed": 5845, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_7688_seed_5845_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.01779683418571949, + "vlm_logprob": -0.01779683418571949, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.01779683418571949, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04058091342449188 + ], + "opposite_logprob_tokens": [ + -3.290580987930298 + ], + "logprob_reward": -0.04058091342449188, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04058091342449188, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048710547387599945, + 0.0318945087492466, + 0.15340377390384674, + 0.14275150001049042, + 0.06124623119831085, + 0.17251573503017426, + 0.1021309345960617, + 0.13027159869670868, + 0.13400456309318542, + 0.023070629686117172 + ], + "credit_map": [ + 0.004479488357901573, + 0.007583407685160637, + 0.0030228449031710625, + 0.005574410315603018, + 0.006943322252482176, + 0.009087691083550453, + 0.004745112266391516, + 0.007274268195033073, + 0.0033679634798318148, + 0.003794414224103093, + 0.002818501088768244, + 0.0030334971379488707, + 0.0038977153599262238, + 0.003551892237737775, + 0.0038172942586243153, + 0.007613231893628836, + 0.005366823170334101, + 0.011194366030395031, + 0.005469504278153181, + 0.006893728394061327, + 0.017287516966462135, + 0.0623222254216671, + 0.03500752151012421, + 0.009862089529633522, + 0.010464770719408989, + 0.016739871352910995, + 0.00849249679595232, + 0.02106381393969059, + 0.028375724330544472, + 0.02298276126384735, + 0.01619042083621025, + 0.018441637977957726, + 0.008755702525377274, + 0.018435781821608543, + 0.006393030751496553, + 0.007518093567341566, + 0.009208142757415771, + 0.003371266881003976, + 0.004325784742832184, + 0.00323842978104949, + 0.02168336510658264, + 0.03150234371423721, + 0.011820576153695583, + 0.02892402559518814, + 0.034124430269002914, + 0.014947027899324894, + 0.014876089990139008, + 0.014637880027294159, + 0.014653148129582405, + 0.017813673242926598, + 0.010555271059274673, + 0.013033909723162651, + 0.013626281172037125, + 0.011499706655740738, + 0.00953639019280672, + 0.011412550695240498, + 0.016301369294524193, + 0.026177730411291122, + 0.012988812290132046, + 0.013150481507182121, + 0.01779855042695999, + 0.012904065661132336, + 0.017024680972099304, + 0.013925908133387566, + 0.011819345876574516, + 0.01132271345704794, + 0.009275462478399277, + 0.012416447512805462, + 0.01864301785826683, + 0.0393398180603981, + 0.01639273390173912, + 0.014795024879276752, + 0.0018600610783323646, + 0.001957514090463519, + 0.0021690952125936747, + 0.0012738262303173542, + 0.005929277278482914, + 0.0032252392265945673, + 0.004461185075342655, + 0.002194432308897376 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 3 + ], + "selected_reward": -0.04058091342449188 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642885833978653 + ], + "opposite_logprob_tokens": [ + -4.758642673492432 + ], + "logprob_reward": -0.008642885833978653, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642885833978653, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.030699051916599274, + 0.02452942728996277, + 0.1492895781993866, + 0.05322541296482086, + 0.06298381090164185, + 0.3509649634361267, + 0.10417003929615021, + 0.11414659023284912, + 0.09863603860139847, + 0.011355037800967693 + ], + "credit_map": [ + 0.0034755109809339046, + 0.004179149866104126, + 0.0024459250271320343, + 0.0024596182629466057, + 0.004912179429084063, + 0.00500098429620266, + 0.0034228048752993345, + 0.004802878014743328, + 0.002461755648255348, + 0.004590009339153767, + 0.001946694334037602, + 0.001995260827243328, + 0.0036636998411267996, + 0.002257334301248193, + 0.00262837135232985, + 0.004986304324120283, + 0.005233981646597385, + 0.011350657790899277, + 0.005322256591171026, + 0.006852929014712572, + 0.016435937955975533, + 0.059093307703733444, + 0.034031275659799576, + 0.010969233699142933, + 0.006075019016861916, + 0.007240361534059048, + 0.003188897855579853, + 0.006861487403512001, + 0.006203170865774155, + 0.010012869723141193, + 0.007481817156076431, + 0.006161792669445276, + 0.007106895092874765, + 0.01659693941473961, + 0.0052014607936143875, + 0.013178362511098385, + 0.010634319856762886, + 0.0028978611808270216, + 0.004953663796186447, + 0.0024143061600625515, + 0.04171617329120636, + 0.060640543699264526, + 0.02536003477871418, + 0.06588809937238693, + 0.0671585202217102, + 0.025622066110372543, + 0.03835025802254677, + 0.026229271665215492, + 0.012744363397359848, + 0.01864440180361271, + 0.012384466826915741, + 0.014176902361214161, + 0.016151579096913338, + 0.011143186129629612, + 0.009382191114127636, + 0.009542946703732014, + 0.015285857021808624, + 0.02158260904252529, + 0.01162855513393879, + 0.011658675037324429, + 0.0147662702947855, + 0.011398551985621452, + 0.014173775911331177, + 0.01365229394286871, + 0.00983599852770567, + 0.007804154418408871, + 0.006681096740067005, + 0.008692190051078796, + 0.013899222947657108, + 0.028972897678613663, + 0.011746719479560852, + 0.011003758758306503, + 0.0013619855744764209, + 0.0010201118420809507, + 0.0010523262899369001, + 0.0006303925183601677, + 0.0028087066020816565, + 0.001250306610018015, + 0.002455031033605337, + 0.0007761774468235672 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.008642885833978653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016073646023869514 + ], + "opposite_logprob_tokens": [ + -4.141073703765869 + ], + "logprob_reward": -0.016073646023869514, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016073646023869514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05847128853201866, + 0.03226820379495621, + 0.1078958585858345, + 0.1879362165927887, + 0.06407514959573746, + 0.05954767391085625, + 0.1776265650987625, + 0.14215144515037537, + 0.11493328213691711, + 0.05509434640407562 + ], + "credit_map": [ + 0.0038455731701105833, + 0.008233251050114632, + 0.004777199123054743, + 0.006077909376472235, + 0.006710882764309645, + 0.012955007143318653, + 0.004951645154505968, + 0.01091981865465641, + 0.0028275202494114637, + 0.004575617145746946, + 0.002509038895368576, + 0.0032741969916969538, + 0.0060971458442509174, + 0.004222513176500797, + 0.0037981278728693724, + 0.0049640401266515255, + 0.004381358157843351, + 0.009079746901988983, + 0.005017517134547234, + 0.006458352319896221, + 0.012137816287577152, + 0.03883923962712288, + 0.021932188421487808, + 0.010049632750451565, + 0.006956321652978659, + 0.021807730197906494, + 0.006484676618129015, + 0.014665977098047733, + 0.0673469603061676, + 0.05001531541347504, + 0.009936779737472534, + 0.010722443461418152, + 0.005880780052393675, + 0.022061480209231377, + 0.004696227144449949, + 0.008901979774236679, + 0.010564272291958332, + 0.0038927672430872917, + 0.005108989309519529, + 0.002968648448586464, + 0.006872989237308502, + 0.01394208800047636, + 0.006353630218654871, + 0.010151213966310024, + 0.008120485581457615, + 0.004087658599019051, + 0.005898868199437857, + 0.004120740108191967, + 0.017925182357430458, + 0.03417310118675232, + 0.025106268003582954, + 0.021918296813964844, + 0.025772033259272575, + 0.021246429532766342, + 0.013634992763400078, + 0.017850248143076897, + 0.020906962454319, + 0.02873925492167473, + 0.015898387879133224, + 0.013361359015107155, + 0.01634766161441803, + 0.014793346635997295, + 0.01609872095286846, + 0.016005732119083405, + 0.009717412292957306, + 0.008108088746666908, + 0.007857849821448326, + 0.009625334292650223, + 0.01589241623878479, + 0.03595169633626938, + 0.015164818614721298, + 0.012615655548870564, + 0.0032681189477443695, + 0.0025931133423000574, + 0.0034113717265427113, + 0.0017626176122575998, + 0.0172031931579113, + 0.007367451209574938, + 0.01525313500314951, + 0.004235338885337114 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 7 + ], + "selected_reward": -0.016073646023869514 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625045254826546 + ], + "opposite_logprob_tokens": [ + -4.882625102996826 + ], + "logprob_reward": -0.007625045254826546, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625045254826546, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06045279651880264, + 0.048467449843883514, + 0.20446407794952393, + 0.10690755397081375, + 0.044136568903923035, + 0.169729083776474, + 0.12744088470935822, + 0.1284075826406479, + 0.06575583666563034, + 0.0442381277680397 + ], + "credit_map": [ + 0.008166925981640816, + 0.009350918233394623, + 0.005482079461216927, + 0.008375408127903938, + 0.007574098184704781, + 0.008374376222491264, + 0.005916961468756199, + 0.007212027907371521, + 0.006666079163551331, + 0.006617147941142321, + 0.005796587094664574, + 0.004833515267819166, + 0.006105940323323011, + 0.006577738560736179, + 0.0035299533046782017, + 0.008340490981936455, + 0.009158018045127392, + 0.015237764455378056, + 0.008356261998414993, + 0.008367807604372501, + 0.0241324994713068, + 0.08234181255102158, + 0.04178668186068535, + 0.015083224512636662, + 0.010034090839326382, + 0.014179442077875137, + 0.008579444140195847, + 0.01084940880537033, + 0.022897448390722275, + 0.016265347599983215, + 0.01031354907900095, + 0.013788815587759018, + 0.004799955990165472, + 0.00809558480978012, + 0.007855495437979698, + 0.006261104252189398, + 0.005341150797903538, + 0.005125090945512056, + 0.0036638271994888783, + 0.0029943608678877354, + 0.01363039668649435, + 0.02952071651816368, + 0.011367822997272015, + 0.02714630402624607, + 0.028603753075003624, + 0.016833564266562462, + 0.021980956196784973, + 0.020645564422011375, + 0.015363221988081932, + 0.023041022941470146, + 0.01604573242366314, + 0.013290819711983204, + 0.02150639332830906, + 0.012740569189190865, + 0.013079242780804634, + 0.012373876757919788, + 0.014106505550444126, + 0.02041027881205082, + 0.019388053566217422, + 0.012400981038808823, + 0.018371522426605225, + 0.0153118297457695, + 0.01609821990132332, + 0.012320187874138355, + 0.006293799262493849, + 0.007022919598966837, + 0.007311624940484762, + 0.008296467363834381, + 0.009702711366117, + 0.011651800945401192, + 0.008831066079437733, + 0.0066454471088945866, + 0.004039991647005081, + 0.004667332861572504, + 0.006140110082924366, + 0.003521854290738702, + 0.008843429386615753, + 0.006862541660666466, + 0.0064317756332457066, + 0.003731091506779194 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.007625045254826546 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016061680391430855 + ], + "opposite_logprob_tokens": [ + -4.141061782836914 + ], + "logprob_reward": -0.016061680391430855, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016061680391430855, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09863369166851044, + 0.05182493478059769, + 0.07596450299024582, + 0.131023108959198, + 0.05845775082707405, + 0.401271790266037, + 0.07323980331420898, + 0.05492214858531952, + 0.026149144396185875, + 0.02851315774023533 + ], + "credit_map": [ + 0.008268596604466438, + 0.013782531023025513, + 0.009010480716824532, + 0.010837570764124393, + 0.011359672993421555, + 0.021246591582894325, + 0.007871797308325768, + 0.01625644415616989, + 0.009179255925118923, + 0.007211060728877783, + 0.006880840752273798, + 0.0051587894558906555, + 0.005670545622706413, + 0.005475946236401796, + 0.0038542351685464382, + 0.008394261822104454, + 0.0044053844176232815, + 0.0066859726794064045, + 0.0037051464896649122, + 0.004241765011101961, + 0.008553491905331612, + 0.02760661207139492, + 0.015225172974169254, + 0.005540959537029266, + 0.007715157698839903, + 0.01197720505297184, + 0.00882233027368784, + 0.01638372428715229, + 0.02448858506977558, + 0.026980923488736153, + 0.01921277865767479, + 0.015442402102053165, + 0.006274404004216194, + 0.010708666406571865, + 0.004723489284515381, + 0.007134640123695135, + 0.008694014512002468, + 0.008655611425638199, + 0.0048182662576437, + 0.007448656018823385, + 0.03980978578329086, + 0.0676184594631195, + 0.029088985174894333, + 0.07088743150234222, + 0.07365205883979797, + 0.03225554898381233, + 0.04730093479156494, + 0.040658578276634216, + 0.006930600386112928, + 0.01013320405036211, + 0.009716368280351162, + 0.010082309134304523, + 0.01176547259092331, + 0.008418516255915165, + 0.010157241486012936, + 0.006036089267581701, + 0.006568607874214649, + 0.011247165501117706, + 0.007601051591336727, + 0.005276654381304979, + 0.006076428107917309, + 0.006600504741072655, + 0.006336356047540903, + 0.005215376149863005, + 0.002163945697247982, + 0.0025935391895473003, + 0.0022807864006608725, + 0.002449305495247245, + 0.00367392017506063, + 0.0047376397997140884, + 0.00496048666536808, + 0.0032895233016461134, + 0.0035152158234268427, + 0.0018964391201734543, + 0.003711644560098648, + 0.0013247192837297916, + 0.007939890027046204, + 0.0036693536676466465, + 0.004858022090047598, + 0.0015978732844814658 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 0 + ], + "selected_reward": -0.016061680391430855 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a first-person perspective, the camera drifts forward with a slow, weightless, dreamlike stability through the silent, abyssal canyons of the submerged cityscape, a scene rendered with the surrealist aesthetic of Makoto Shinkai. The vibrant neon pinks and electric blues of flickering, water-distorting holographic advertisements cast long, oscillating reflections on the algae-stained glass and steel facades of the submerged skyscrapers, their light piercing the deep blue waters as an ethereal godlight. In the vast open spaces between these massive structures, enormous bell-shaped creatures, their bodies composed of a translucent, gelatinous material, perform a mesmerizing, silent courtship dance. Their delicate, fiber-optic tentacles trail behind them, emitting a soft, shifting inner light from within, pulsating rhythmically from ethereal cyan to warm fuchsia. As the camera glides past, one creature gracefully retracts its bell, slowly ascending as it spins around a vast, darkened building, while its partner responds by emitting bioluminescence in a synchronized, silent beat, their movements creating gentle, almost imperceptible currents in the clear water. The atmosphere is filled with profound, melancholic beauty and silent awe, the organic light of the creatures' ballet contrasting sharply with the ghostly, artificial glow of a lost civilization beneath the waves.", + "seed": 5846, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_7688_seed_5846_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.018951823841780423, + "vlm_logprob": -0.018951823841780423, + "vlm_yes_no_margin": 4.15, + "vlm_token_credit": -0.018951823841780423, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03305463120341301 + ], + "opposite_logprob_tokens": [ + -3.6580545902252197 + ], + "logprob_reward": -0.03305463120341301, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03305463120341301, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23937523365020752, + 0.10362963378429413, + 0.040103036910295486, + 0.081660695374012, + 0.05808170884847641, + 0.0895623192191124, + 0.13087494671344757, + 0.09203414618968964, + 0.061047717928886414, + 0.10363048315048218 + ], + "credit_map": [ + 0.029673604294657707, + 0.03644711151719093, + 0.025501936674118042, + 0.022420698776841164, + 0.05074925348162651, + 0.02834341861307621, + 0.02520410344004631, + 0.021035129204392433, + 0.012688341550529003, + 0.02673390693962574, + 0.006230928935110569, + 0.014918164350092411, + 0.007373211905360222, + 0.010278313420712948, + 0.014628047123551369, + 0.01077872421592474, + 0.00490712933242321, + 0.003971134312450886, + 0.004544650670140982, + 0.005714367143809795, + 0.002605770481750369, + 0.006476917769759893, + 0.007605001330375671, + 0.004278064239770174, + 0.01081776525825262, + 0.009546024724841118, + 0.013388746418058872, + 0.011622074991464615, + 0.009487710893154144, + 0.00996796041727066, + 0.007801772560924292, + 0.009028637781739235, + 0.005074632354080677, + 0.00557542871683836, + 0.015499264001846313, + 0.007681174669414759, + 0.003926097881048918, + 0.011965923942625523, + 0.0036069191992282867, + 0.004752264823764563, + 0.013140314258635044, + 0.010157493874430656, + 0.010903306305408478, + 0.009126045741140842, + 0.01086137443780899, + 0.014152192510664463, + 0.009919433854520321, + 0.011302159167826176, + 0.016377750784158707, + 0.013206600211560726, + 0.021090522408485413, + 0.01928868144750595, + 0.015732545405626297, + 0.015808558091521263, + 0.011691365391016006, + 0.017678938806056976, + 0.012274939566850662, + 0.010432509705424309, + 0.010349245741963387, + 0.00952454749494791, + 0.01910717412829399, + 0.012324873358011246, + 0.009059537202119827, + 0.008961321786046028, + 0.007160086650401354, + 0.007525801658630371, + 0.008540731854736805, + 0.005640801973640919, + 0.006888472940772772, + 0.009688164107501507, + 0.006641501095145941, + 0.008962154388427734, + 0.011988529935479164, + 0.010652617551386356, + 0.012978097423911095, + 0.01483763474971056, + 0.007120246067643166, + 0.015597037971019745, + 0.020139064639806747, + 0.010317259468138218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.03305463120341301 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012550304643809795 + ], + "opposite_logprob_tokens": [ + -4.387550354003906 + ], + "logprob_reward": -0.012550304643809795, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012550304643809795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05707710236310959, + 0.0731670930981636, + 0.038490474224090576, + 0.2188267707824707, + 0.031041031703352928, + 0.18855540454387665, + 0.03779183328151703, + 0.19081029295921326, + 0.0625867024064064, + 0.10165330767631531 + ], + "credit_map": [ + 0.006863055285066366, + 0.011052246205508709, + 0.0076139853335917, + 0.007930956780910492, + 0.00717573519796133, + 0.00568621838465333, + 0.005027020815759897, + 0.0057278843596577644, + 0.005601866170763969, + 0.01926240883767605, + 0.005581849254667759, + 0.009965067729353905, + 0.010416777804493904, + 0.0080498568713665, + 0.007639972493052483, + 0.006649291608482599, + 0.005277123302221298, + 0.00391697883605957, + 0.007952217012643814, + 0.004658494144678116, + 0.0031463182531297207, + 0.005129391793161631, + 0.004347638692706823, + 0.004062311723828316, + 0.02881241962313652, + 0.0430435948073864, + 0.037184592336416245, + 0.03180304542183876, + 0.01765185222029686, + 0.023143788799643517, + 0.01783275231719017, + 0.01935473084449768, + 0.0035381510388106108, + 0.0030134543776512146, + 0.00752889784052968, + 0.0039290087297558784, + 0.0015508553478866816, + 0.0038277937564998865, + 0.004186650738120079, + 0.003466220572590828, + 0.022188184782862663, + 0.024503160268068314, + 0.017719555646181107, + 0.019907262176275253, + 0.04035608097910881, + 0.02056930586695671, + 0.01871935836970806, + 0.02459249086678028, + 0.0045413849875330925, + 0.004364904016256332, + 0.004529913421720266, + 0.004487384110689163, + 0.005563549231737852, + 0.004795853979885578, + 0.005249867215752602, + 0.004258973989635706, + 0.02360313944518566, + 0.018365047872066498, + 0.02736075222492218, + 0.021119412034749985, + 0.022715426981449127, + 0.024314269423484802, + 0.023898987099528313, + 0.02943326346576214, + 0.0078739570453763, + 0.006840317510068417, + 0.007674453314393759, + 0.0061546373181045055, + 0.01060558296740055, + 0.009293058887124062, + 0.00588991167023778, + 0.008254783228039742, + 0.010707119479775429, + 0.012309643439948559, + 0.014728800393640995, + 0.020317506045103073, + 0.007625540718436241, + 0.011010023765265942, + 0.01677708886563778, + 0.008177589625120163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 5 + ], + "selected_reward": -0.012550304643809795 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026354322209954262 + ], + "opposite_logprob_tokens": [ + -3.6513543128967285 + ], + "logprob_reward": -0.026354322209954262, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026354322209954262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13082528114318848, + 0.125387504696846, + 0.08292146772146225, + 0.07047565281391144, + 0.0411713570356369, + 0.10244271159172058, + 0.05401229113340378, + 0.1279297024011612, + 0.06961610168218613, + 0.19521796703338623 + ], + "credit_map": [ + 0.018864912912249565, + 0.025073450058698654, + 0.01750177890062332, + 0.010323102585971355, + 0.011217329651117325, + 0.017597196623682976, + 0.014199381694197655, + 0.016048116609454155, + 0.011035057716071606, + 0.031434252858161926, + 0.00728653185069561, + 0.025486906990408897, + 0.008670816197991371, + 0.011511766351759434, + 0.014425003901124, + 0.015537170693278313, + 0.008250528015196323, + 0.00894953217357397, + 0.009113803505897522, + 0.009972501546144485, + 0.013058070093393326, + 0.014210679568350315, + 0.009897680953145027, + 0.009468669071793556, + 0.007811251562088728, + 0.010468339547514915, + 0.015367408283054829, + 0.009411873295903206, + 0.006985559593886137, + 0.007788748946040869, + 0.005670459475368261, + 0.0069720144383609295, + 0.003064842661842704, + 0.004403308965265751, + 0.007228228263556957, + 0.004409044981002808, + 0.0037736075464636087, + 0.01113286241889, + 0.002577953040599823, + 0.004581510554999113, + 0.011705742217600346, + 0.011781183071434498, + 0.01059885136783123, + 0.012928377836942673, + 0.017594050616025925, + 0.012410447001457214, + 0.01162276603281498, + 0.01380128227174282, + 0.004927922040224075, + 0.005251321475952864, + 0.006264024879783392, + 0.005442477762699127, + 0.01057166513055563, + 0.006480274256318808, + 0.006947303656488657, + 0.008127303794026375, + 0.013947236351668835, + 0.011672279797494411, + 0.01839001104235649, + 0.015334099531173706, + 0.02151229791343212, + 0.017763648182153702, + 0.014736535958945751, + 0.014573597349226475, + 0.008960976265370846, + 0.00869305245578289, + 0.008842955343425274, + 0.006921881344169378, + 0.01060597226023674, + 0.009307135827839375, + 0.009314712136983871, + 0.006969416979700327, + 0.020073302090168, + 0.018376745283603668, + 0.028414016589522362, + 0.029203714802861214, + 0.017154313623905182, + 0.021821526810526848, + 0.04060499370098114, + 0.01956935040652752 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.026354322209954262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735833827406168 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735833827406168, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735833827406168, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1314910650253296, + 0.06732102483510971, + 0.051556117832660675, + 0.0373692587018013, + 0.11827987432479858, + 0.21445488929748535, + 0.10145597159862518, + 0.08843610435724258, + 0.09377565234899521, + 0.09586004167795181 + ], + "credit_map": [ + 0.019335726276040077, + 0.017571832984685898, + 0.03311847895383835, + 0.014012444764375687, + 0.011366356164216995, + 0.011845946311950684, + 0.010948891751468182, + 0.013291393406689167, + 0.009550296701490879, + 0.011642258614301682, + 0.006230755243450403, + 0.009743943810462952, + 0.006636365316808224, + 0.010330699384212494, + 0.006561800371855497, + 0.006624906789511442, + 0.004588973708450794, + 0.003597227856516838, + 0.005415925290435553, + 0.0061020636931061745, + 0.003476956393569708, + 0.01256733387708664, + 0.010371848940849304, + 0.005435785744339228, + 0.0038421018980443478, + 0.005447216331958771, + 0.005904734134674072, + 0.0049764234572649, + 0.0050589581951498985, + 0.004526217933744192, + 0.0037566181272268295, + 0.003856985829770565, + 0.012429927475750446, + 0.010680137202143669, + 0.03418997675180435, + 0.0148147689178586, + 0.008556295186281204, + 0.014699112623929977, + 0.006817768327891827, + 0.016091886907815933, + 0.024416837841272354, + 0.024885190650820732, + 0.017751766368746758, + 0.019681919366121292, + 0.04892805591225624, + 0.028026392683386803, + 0.025022482499480247, + 0.02574223466217518, + 0.013477983884513378, + 0.014400621876120567, + 0.013515382073819637, + 0.015886405482888222, + 0.011632222682237625, + 0.012505725026130676, + 0.009460093453526497, + 0.010577536188066006, + 0.006793804932385683, + 0.008835967630147934, + 0.009606275707483292, + 0.009512623772025108, + 0.01698450930416584, + 0.011126348748803139, + 0.011842101812362671, + 0.013734469190239906, + 0.014585074968636036, + 0.014646382071077824, + 0.008507722988724709, + 0.007613212801516056, + 0.00808520708233118, + 0.018340634182095528, + 0.009847499430179596, + 0.012149922549724579, + 0.013439998961985111, + 0.012478902004659176, + 0.015077714808285236, + 0.015396966598927975, + 0.005775028839707375, + 0.012414682656526566, + 0.012442239560186863, + 0.008834512904286385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.006735833827406168 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016064027324318886 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.016064027324318886, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016064027324318886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11036374419927597, + 0.1999259889125824, + 0.09440560638904572, + 0.08408833295106888, + 0.09778063744306564, + 0.05017738416790962, + 0.14276273548603058, + 0.10978004336357117, + 0.06173717975616455, + 0.0489782877266407 + ], + "credit_map": [ + 0.014350054785609245, + 0.015878021717071533, + 0.00949829537421465, + 0.009288801811635494, + 0.016623759642243385, + 0.01770464889705181, + 0.01241425983607769, + 0.01460591796785593, + 0.018817933276295662, + 0.05360648036003113, + 0.009956690482795238, + 0.037314411252737045, + 0.017226317897439003, + 0.020552989095449448, + 0.018322501331567764, + 0.02412869967520237, + 0.006037243641912937, + 0.015448004938662052, + 0.006158645264804363, + 0.009969050996005535, + 0.005187253002077341, + 0.025933438912034035, + 0.015154271386563778, + 0.010517705231904984, + 0.012174739502370358, + 0.014357348904013634, + 0.009784780442714691, + 0.012909071519970894, + 0.006676880642771721, + 0.010551029816269875, + 0.007821095176041126, + 0.00981339905411005, + 0.009102691896259785, + 0.0065178717486560345, + 0.028280217200517654, + 0.01321255974471569, + 0.006708529312163591, + 0.016183950006961823, + 0.0063374838791787624, + 0.011437343433499336, + 0.00535387871786952, + 0.005558614619076252, + 0.005184587091207504, + 0.007292930036783218, + 0.007708884309977293, + 0.005867471918463707, + 0.005980313755571842, + 0.007230709306895733, + 0.020265523344278336, + 0.0206376351416111, + 0.021035211160779, + 0.018863040953874588, + 0.01334332674741745, + 0.01810004748404026, + 0.012923681177198887, + 0.01759430207312107, + 0.010283960960805416, + 0.009467365220189095, + 0.01414432842284441, + 0.010494731366634369, + 0.025100775063037872, + 0.014280927367508411, + 0.011016052216291428, + 0.014991921372711658, + 0.009941067546606064, + 0.006017129868268967, + 0.009293588809669018, + 0.005091029219329357, + 0.0048960233107209206, + 0.012572948820888996, + 0.0049974750727415085, + 0.008927928283810616, + 0.00614415155723691, + 0.00400639558210969, + 0.00798551831394434, + 0.005857698153704405, + 0.004080806393176317, + 0.005654324311763048, + 0.0102403424680233, + 0.005009055603295565 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.016064027324318886 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a first-person perspective, the camera drifts forward with a slow, weightless, dreamlike stability through the silent, abyssal canyons of the submerged cityscape, a scene rendered with the surrealist aesthetic of Makoto Shinkai. The vibrant neon pinks and electric blues of flickering, water-distorting holographic advertisements cast long, oscillating reflections on the algae-stained glass and steel facades of the submerged skyscrapers, their light piercing the deep blue waters as an ethereal godlight. In the vast open spaces between these massive structures, enormous bell-shaped creatures, their bodies composed of a translucent, gelatinous material, perform a mesmerizing, silent courtship dance. Their delicate, fiber-optic tentacles trail behind them, emitting a soft, shifting inner light from within, pulsating rhythmically from ethereal cyan to warm fuchsia. As the camera glides past, one creature gracefully retracts its bell, slowly ascending as it spins around a vast, darkened building, while its partner responds by emitting bioluminescence in a synchronized, silent beat, their movements creating gentle, almost imperceptible currents in the clear water. The atmosphere is filled with profound, melancholic beauty and silent awe, the organic light of the creatures' ballet contrasting sharply with the ghostly, artificial glow of a lost civilization beneath the waves.", + "seed": 5847, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_7688_seed_5847_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022141399374231697, + "vlm_logprob": -0.022141399374231697, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.022141399374231697, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05852658301591873 + ], + "opposite_logprob_tokens": [ + -2.9335265159606934 + ], + "logprob_reward": -0.05852658301591873, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05852658301591873, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06123030558228493, + 0.01541709154844284, + 0.033100470900535583, + 0.2170383632183075, + 0.08449169993400574, + 0.026132455095648766, + 0.15950877964496613, + 0.059601329267024994, + 0.3272888958454132, + 0.0161905474960804 + ], + "credit_map": [ + 0.004924426786601543, + 0.00926742423325777, + 0.004101383499801159, + 0.008254995569586754, + 0.006623023189604282, + 0.011352325789630413, + 0.007839231751859188, + 0.008867497555911541, + 0.0027137245051562786, + 0.0019611110910773277, + 0.0014003043761476874, + 0.0016967381816357374, + 0.0018596117151901126, + 0.0021772957406938076, + 0.0014705691719427705, + 0.0021377368830144405, + 0.0031612450256943703, + 0.004435224458575249, + 0.0025772852823138237, + 0.005552968010306358, + 0.002974235685542226, + 0.0032324176281690598, + 0.003923981916159391, + 0.007243112660944462, + 0.01146511361002922, + 0.02143257111310959, + 0.008169426582753658, + 0.08505415916442871, + 0.007080241572111845, + 0.008101865649223328, + 0.010827748104929924, + 0.0649072453379631, + 0.006234521511942148, + 0.027550095692276955, + 0.0272828359156847, + 0.004378206562250853, + 0.005325056612491608, + 0.002776106121018529, + 0.004447216168045998, + 0.006497658789157867, + 0.002906834939494729, + 0.002408496802672744, + 0.003965427167713642, + 0.005991877056658268, + 0.002465049270540476, + 0.0018979336600750685, + 0.004219737835228443, + 0.0022770999930799007, + 0.05676952376961708, + 0.06103665381669998, + 0.006913449615240097, + 0.006110716611146927, + 0.014486100524663925, + 0.004008878488093615, + 0.006045137066394091, + 0.004138334654271603, + 0.003953839652240276, + 0.0034893506672233343, + 0.01906825602054596, + 0.002948024543002248, + 0.0203013326972723, + 0.0029600667767226696, + 0.003856599098071456, + 0.0030238600447773933, + 0.02613500878214836, + 0.035404022783041, + 0.033885467797517776, + 0.052919428795576096, + 0.04708196967840195, + 0.03563733771443367, + 0.05497298762202263, + 0.041252683848142624, + 0.0011985941091552377, + 0.0018939707661047578, + 0.00173175148665905, + 0.0037137006875127554, + 0.001982390880584717, + 0.002116353949531913, + 0.0019077608594670892, + 0.0016460244078189135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 6 + ], + "selected_reward": -0.05852658301591873 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014204702340066433 + ], + "opposite_logprob_tokens": [ + -4.264204502105713 + ], + "logprob_reward": -0.014204702340066433, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014204702340066433, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.054711442440748215, + 0.021808447316288948, + 0.04162640869617462, + 0.1476972997188568, + 0.1059856191277504, + 0.03930988162755966, + 0.11174971610307693, + 0.2689875662326813, + 0.17595311999320984, + 0.03217046707868576 + ], + "credit_map": [ + 0.0047287470661103725, + 0.006596478167921305, + 0.0037570272106677294, + 0.009192298166453838, + 0.006788048427551985, + 0.010617145337164402, + 0.006061617750674486, + 0.00697008240967989, + 0.002371395705267787, + 0.0026111193001270294, + 0.0025883582420647144, + 0.003150886157527566, + 0.002884902758523822, + 0.0030318491626530886, + 0.002563120098784566, + 0.0026068168226629496, + 0.003929737955331802, + 0.006470961030572653, + 0.0032908390276134014, + 0.005772617645561695, + 0.0059882537461817265, + 0.003491400508210063, + 0.005483971443027258, + 0.007198627106845379, + 0.00866527110338211, + 0.014954776503145695, + 0.00645890086889267, + 0.05376729369163513, + 0.005284835118800402, + 0.007862242870032787, + 0.008816535584628582, + 0.041887447237968445, + 0.009595997631549835, + 0.033188946545124054, + 0.0271555557847023, + 0.009799656458199024, + 0.007709656376391649, + 0.0036197882145643234, + 0.007672802545130253, + 0.007243219297379255, + 0.003387916600331664, + 0.0043776985257864, + 0.00521072605624795, + 0.010375121608376503, + 0.003083843272179365, + 0.0024679244961589575, + 0.006913110613822937, + 0.00349353882484138, + 0.02504640445113182, + 0.025047186762094498, + 0.012946460396051407, + 0.012006215751171112, + 0.009895027615129948, + 0.00631007319316268, + 0.011759248562157154, + 0.008739110082387924, + 0.016706695780158043, + 0.02352207899093628, + 0.10014882683753967, + 0.00981962587684393, + 0.08137611299753189, + 0.011579038575291634, + 0.015280261635780334, + 0.010554933920502663, + 0.01300700195133686, + 0.01913565583527088, + 0.01927950792014599, + 0.029644252732396126, + 0.02600935846567154, + 0.01995999366044998, + 0.025212155655026436, + 0.023705197498202324, + 0.0033818159718066454, + 0.003478137543424964, + 0.003946384415030479, + 0.007029398810118437, + 0.004270879551768303, + 0.004675190895795822, + 0.003293772228062153, + 0.002094888361170888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 3 + ], + "selected_reward": -0.014204702340066433 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011081814765930176 + ], + "opposite_logprob_tokens": [ + -4.511081695556641 + ], + "logprob_reward": -0.011081814765930176, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011081814765930176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28080669045448303, + 0.04794227331876755, + 0.055363334715366364, + 0.1432504653930664, + 0.12488596886396408, + 0.0371473953127861, + 0.10736250877380371, + 0.08782897889614105, + 0.061926309019327164, + 0.053486019372940063 + ], + "credit_map": [ + 0.015271937474608421, + 0.03464336693286896, + 0.016589924693107605, + 0.0784999281167984, + 0.03307802975177765, + 0.03146946057677269, + 0.03206764534115791, + 0.039186425507068634, + 0.00467339064925909, + 0.007035046350210905, + 0.005767422262579203, + 0.008321018889546394, + 0.004770644940435886, + 0.005943760741502047, + 0.005629450548440218, + 0.0058015440590679646, + 0.005530622787773609, + 0.008527001366019249, + 0.0030368189327418804, + 0.007883689366281033, + 0.007555450778454542, + 0.005541677586734295, + 0.007919813506305218, + 0.009368263185024261, + 0.007960034534335136, + 0.013418219052255154, + 0.006293355021625757, + 0.05445491522550583, + 0.005591955967247486, + 0.005902842152863741, + 0.007854793220758438, + 0.041774362325668335, + 0.010962187312543392, + 0.0425405316054821, + 0.03731529787182808, + 0.006400648970156908, + 0.008755658753216267, + 0.0037522141356021166, + 0.006432946305721998, + 0.00872648786753416, + 0.003300981130450964, + 0.004201299976557493, + 0.004370462615042925, + 0.007552756927907467, + 0.004047667607665062, + 0.003945199307054281, + 0.0058578383177518845, + 0.0038711943197995424, + 0.02962641790509224, + 0.031152499839663506, + 0.011603999882936478, + 0.010440272279083729, + 0.007862427271902561, + 0.004532740451395512, + 0.007650112267583609, + 0.004494039807468653, + 0.01062101311981678, + 0.009448365308344364, + 0.01807848922908306, + 0.009235741570591927, + 0.01737007312476635, + 0.005677841138094664, + 0.011137692257761955, + 0.006259775720536709, + 0.004015570506453514, + 0.011471954174339771, + 0.00709103187546134, + 0.011257415637373924, + 0.007846932858228683, + 0.006311739794909954, + 0.005670195911079645, + 0.008261473849415779, + 0.005775851663202047, + 0.00512148579582572, + 0.006371110212057829, + 0.007032301742583513, + 0.006695673801004887, + 0.007406020537018776, + 0.008952927775681019, + 0.006130646914243698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.011081814765930176 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003614442190155387 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.003614442190155387, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003614442190155387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0820297822356224, + 0.021451327949762344, + 0.023057159036397934, + 0.45446527004241943, + 0.11215152591466904, + 0.023526232689619064, + 0.09077031910419464, + 0.04530287906527519, + 0.09768493473529816, + 0.049560513347387314 + ], + "credit_map": [ + 0.0055568162351846695, + 0.012377733364701271, + 0.0054939487017691135, + 0.014529221691191196, + 0.014216416515409946, + 0.01044226810336113, + 0.00923885963857174, + 0.010174519382417202, + 0.0025624611880630255, + 0.0027232260908931494, + 0.0017696040449663997, + 0.0034576214384287596, + 0.002081782091408968, + 0.0038881762884557247, + 0.0020453990437090397, + 0.002923058345913887, + 0.0026808674447238445, + 0.003212790936231613, + 0.0025381064042448997, + 0.0030819776002317667, + 0.0025244704447686672, + 0.002872859127819538, + 0.002347828121855855, + 0.0037982589565217495, + 0.023439576849341393, + 0.04084963724017143, + 0.014667605981230736, + 0.18046581745147705, + 0.014337442815303802, + 0.017135154455900192, + 0.022864338010549545, + 0.14070574939250946, + 0.00844915583729744, + 0.03520556539297104, + 0.03483368828892708, + 0.006619610358029604, + 0.007764384616166353, + 0.0031595174223184586, + 0.0068361880257725716, + 0.009283427149057388, + 0.002776279579848051, + 0.002876858925446868, + 0.002843360183760524, + 0.003577891504392028, + 0.002040799940004945, + 0.002321914304047823, + 0.004039800725877285, + 0.003049330087378621, + 0.03020535781979561, + 0.031142663210630417, + 0.004697042051702738, + 0.006528615485876799, + 0.00725807249546051, + 0.0031586100813001394, + 0.004993955604732037, + 0.002786004915833473, + 0.005289810709655285, + 0.009161651134490967, + 0.007880649529397488, + 0.004200449213385582, + 0.005893469788134098, + 0.0034443174954503775, + 0.005316032096743584, + 0.004116500727832317, + 0.0075338417664170265, + 0.011083230376243591, + 0.010233043693006039, + 0.016979670152068138, + 0.014192618429660797, + 0.011479291133582592, + 0.015347822569310665, + 0.010835433378815651, + 0.00530311930924654, + 0.006433673668652773, + 0.007103160489350557, + 0.011669798754155636, + 0.005879436619579792, + 0.005018291063606739, + 0.005154769401997328, + 0.0029982684645801783 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 8 + ], + "selected_reward": -0.003614442190155387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023279454559087753 + ], + "opposite_logprob_tokens": [ + -3.7732794284820557 + ], + "logprob_reward": -0.023279454559087753, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023279454559087753, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0742843896150589, + 0.03192988410592079, + 0.044919922947883606, + 0.10733184218406677, + 0.2179095447063446, + 0.025687219575047493, + 0.20673182606697083, + 0.04236360639333725, + 0.19455668330192566, + 0.054285094141960144 + ], + "credit_map": [ + 0.007555460091680288, + 0.009266131557524204, + 0.005254996474832296, + 0.018539810553193092, + 0.00799439288675785, + 0.010059054009616375, + 0.006231528241187334, + 0.009383014403283596, + 0.003071082988753915, + 0.003957472741603851, + 0.004128752741962671, + 0.006050196010619402, + 0.003559256438165903, + 0.0038602855056524277, + 0.003173466306179762, + 0.004129374399781227, + 0.006224992219358683, + 0.005632650572806597, + 0.0042497580870985985, + 0.0055203866213560104, + 0.003500201040878892, + 0.005986113101243973, + 0.0071711367927491665, + 0.0066346838138997555, + 0.007160139735788107, + 0.01289376150816679, + 0.008084685541689396, + 0.03477708622813225, + 0.005691139493137598, + 0.005514192394912243, + 0.00779362628236413, + 0.025417214259505272, + 0.01710589788854122, + 0.07378148287534714, + 0.07037907838821411, + 0.010470468550920486, + 0.013850617222487926, + 0.0059234369546175, + 0.010738003067672253, + 0.01566055789589882, + 0.0037192406598478556, + 0.004413391463458538, + 0.003614237764850259, + 0.003212497103959322, + 0.0020807036198675632, + 0.0026349469553679228, + 0.003246174193918705, + 0.002766028046607971, + 0.07358694821596146, + 0.07727867364883423, + 0.008305278606712818, + 0.007254370953887701, + 0.017560908570885658, + 0.009068701416254044, + 0.00785327423363924, + 0.00582366855815053, + 0.006799677386879921, + 0.0050491453148424625, + 0.004814688116312027, + 0.006026985589414835, + 0.007256579585373402, + 0.0031297849491238594, + 0.006156086456030607, + 0.003130657598376274, + 0.015398587100207806, + 0.020389297977089882, + 0.021723691374063492, + 0.032558687031269073, + 0.02880375273525715, + 0.02030293457210064, + 0.03291657194495201, + 0.02246316522359848, + 0.004333381075412035, + 0.006344913505017757, + 0.004729339387267828, + 0.007584382314234972, + 0.009882849641144276, + 0.00951522309333086, + 0.006218344438821077, + 0.0056766606867313385 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 8 + ], + "selected_reward": -0.023279454559087753 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Told from a first-person perspective, the camera drifts forward with a slow, weightless, dreamlike stability through the silent, abyssal canyons of the submerged cityscape, a scene rendered with the surrealist aesthetic of Makoto Shinkai. The vibrant neon pinks and electric blues of flickering, water-distorting holographic advertisements cast long, oscillating reflections on the algae-stained glass and steel facades of the submerged skyscrapers, their light piercing the deep blue waters as an ethereal godlight. In the vast open spaces between these massive structures, enormous bell-shaped creatures, their bodies composed of a translucent, gelatinous material, perform a mesmerizing, silent courtship dance. Their delicate, fiber-optic tentacles trail behind them, emitting a soft, shifting inner light from within, pulsating rhythmically from ethereal cyan to warm fuchsia. As the camera glides past, one creature gracefully retracts its bell, slowly ascending as it spins around a vast, darkened building, while its partner responds by emitting bioluminescence in a synchronized, silent beat, their movements creating gentle, almost imperceptible currents in the clear water. The atmosphere is filled with profound, melancholic beauty and silent awe, the organic light of the creatures' ballet contrasting sharply with the ghostly, artificial glow of a lost civilization beneath the waves.", + "seed": 5848, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_5/index_7688_seed_5848_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01800149492919445, + "vlm_logprob": -0.01800149492919445, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.01800149492919445, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05143929645419121 + ], + "opposite_logprob_tokens": [ + -3.0514392852783203 + ], + "logprob_reward": -0.05143929645419121, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05143929645419121, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23266123235225677, + 0.1097080409526825, + 0.03004400245845318, + 0.1047288179397583, + 0.19679735600948334, + 0.02177548035979271, + 0.07114063203334808, + 0.03474698215723038, + 0.042130809277296066, + 0.15626665949821472 + ], + "credit_map": [ + 0.03084268607199192, + 0.01842726580798626, + 0.030232161283493042, + 0.015815436840057373, + 0.05650222674012184, + 0.034065570682287216, + 0.024402188137173653, + 0.022373687475919724, + 0.015201595611870289, + 0.0057048918679356575, + 0.0054268487729132175, + 0.005162752233445644, + 0.014919807203114033, + 0.040071677416563034, + 0.008912783116102219, + 0.014307689853012562, + 0.003183273132890463, + 0.002153111854568124, + 0.0019582558888942003, + 0.0026604237500578165, + 0.0027412264607846737, + 0.0038578263483941555, + 0.009252257645130157, + 0.00423762621358037, + 0.02142927795648575, + 0.005120707675814629, + 0.008684842847287655, + 0.0056540523655712605, + 0.005644884891808033, + 0.01425138209015131, + 0.012238173745572567, + 0.031705498695373535, + 0.016383184120059013, + 0.009484687820076942, + 0.0529770664870739, + 0.025271793827414513, + 0.008420288562774658, + 0.0284670889377594, + 0.033682681620121, + 0.02211054414510727, + 0.0023018005304038525, + 0.001768812770023942, + 0.001621341798454523, + 0.0022462147753685713, + 0.0020963973365724087, + 0.004083107225596905, + 0.0037204567342996597, + 0.003937350586056709, + 0.0027180653996765614, + 0.003119109431281686, + 0.00425840076059103, + 0.003725352929905057, + 0.0045915814116597176, + 0.02029142715036869, + 0.00702962139621377, + 0.02540707029402256, + 0.00296335993334651, + 0.004487049300223589, + 0.001822135760448873, + 0.0037625799886882305, + 0.005836508236825466, + 0.004676596727222204, + 0.004680556710809469, + 0.006518196314573288, + 0.0059488373808562756, + 0.004111660644412041, + 0.006277520675212145, + 0.004380547441542149, + 0.004132066387683153, + 0.007500145584344864, + 0.004154919646680355, + 0.005625109653919935, + 0.012379122897982597, + 0.012097222730517387, + 0.011129711754620075, + 0.033005841076374054, + 0.01221931166946888, + 0.033143818378448486, + 0.015602204948663712, + 0.0266894344240427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.05143929645419121 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009789552539587021 + ], + "opposite_logprob_tokens": [ + -4.63478946685791 + ], + "logprob_reward": -0.009789552539587021, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009789552539587021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07835347950458527, + 0.05726556479930878, + 0.0911497101187706, + 0.06068866327404976, + 0.3364156186580658, + 0.038069069385528564, + 0.06765782088041306, + 0.05905776470899582, + 0.1329321563243866, + 0.07841013371944427 + ], + "credit_map": [ + 0.009515254758298397, + 0.006909748539328575, + 0.009458613581955433, + 0.005481726955622435, + 0.016566388309001923, + 0.011692916974425316, + 0.010083526372909546, + 0.008645305410027504, + 0.008298475295305252, + 0.00487934285774827, + 0.0034974059090018272, + 0.004457224626094103, + 0.008720070123672485, + 0.01587379164993763, + 0.004067809320986271, + 0.007471444550901651, + 0.008024778217077255, + 0.012441815808415413, + 0.006689064670354128, + 0.009051944129168987, + 0.0112161198630929, + 0.007567033637315035, + 0.022083086892962456, + 0.01407586969435215, + 0.007157175336033106, + 0.006248531397432089, + 0.008852556347846985, + 0.006003919523209333, + 0.007869122549891472, + 0.008986369706690311, + 0.006393587216734886, + 0.00917740073055029, + 0.027781937271356583, + 0.024705151095986366, + 0.07333666831254959, + 0.07531953603029251, + 0.020510390400886536, + 0.02336449734866619, + 0.05486254394054413, + 0.036534909158945084, + 0.004367196932435036, + 0.002865562215447426, + 0.0026202802546322346, + 0.005753055214881897, + 0.003182018641382456, + 0.00480700982734561, + 0.008043318055570126, + 0.0064306301064789295, + 0.0037306868471205235, + 0.005442165769636631, + 0.004677231889218092, + 0.004113268107175827, + 0.005537084303796291, + 0.0177067331969738, + 0.005818950943648815, + 0.020631704479455948, + 0.004942994099110365, + 0.008481074124574661, + 0.005029716063290834, + 0.007589489687234163, + 0.006115158554166555, + 0.006744795478880405, + 0.009078496135771275, + 0.011076042428612709, + 0.009467115625739098, + 0.007057543843984604, + 0.010366079397499561, + 0.03324063494801521, + 0.006096778903156519, + 0.04373774304986, + 0.009860404767096043, + 0.013105854392051697, + 0.009320449084043503, + 0.007194233126938343, + 0.007590899243950844, + 0.012717814184725285, + 0.008171741850674152, + 0.010594367049634457, + 0.011748174205422401, + 0.011072455905377865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.009789552539587021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008642058819532394 + ], + "opposite_logprob_tokens": [ + -4.758642196655273 + ], + "logprob_reward": -0.008642058819532394, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008642058819532394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05529152229428291, + 0.07390861213207245, + 0.050616636872291565, + 0.07014475762844086, + 0.19086489081382751, + 0.08448391407728195, + 0.23270748555660248, + 0.04831138625741005, + 0.12701329588890076, + 0.06665752083063126 + ], + "credit_map": [ + 0.007551711518317461, + 0.004975429270416498, + 0.00519556412473321, + 0.005081148352473974, + 0.007697804365307093, + 0.010995919816195965, + 0.006955369375646114, + 0.006838575471192598, + 0.010403119958937168, + 0.004905638750642538, + 0.005288372281938791, + 0.005940533243119717, + 0.013989477418363094, + 0.017245812341570854, + 0.007284501567482948, + 0.008851153776049614, + 0.004572932608425617, + 0.006258363835513592, + 0.004048979375511408, + 0.004022679291665554, + 0.005208670627325773, + 0.007091735024005175, + 0.013546348549425602, + 0.005866927560418844, + 0.006899981759488583, + 0.009693016298115253, + 0.008005212992429733, + 0.007321212440729141, + 0.006214969791471958, + 0.007903517223894596, + 0.012210413813591003, + 0.01189642958343029, + 0.017245905473828316, + 0.01334484200924635, + 0.04370155185461044, + 0.04019438475370407, + 0.010743374936282635, + 0.013593212701380253, + 0.03213446959853172, + 0.019907144829630852, + 0.006268209777772427, + 0.004432796034961939, + 0.004520748741924763, + 0.006543208844959736, + 0.005217770580202341, + 0.019496848806738853, + 0.02009960450232029, + 0.017904730513691902, + 0.01929178275167942, + 0.03054121881723404, + 0.03356059268116951, + 0.023321358487010002, + 0.024834519252181053, + 0.041930004954338074, + 0.026496686041355133, + 0.0327313169836998, + 0.004874791484326124, + 0.005798682104796171, + 0.004073418211191893, + 0.0045231725089251995, + 0.003788835136219859, + 0.0066660502925515175, + 0.007377293426543474, + 0.011209146119654179, + 0.014490856789052486, + 0.011749467812478542, + 0.01814374141395092, + 0.023156510666012764, + 0.010876256972551346, + 0.025748716667294502, + 0.009960559196770191, + 0.012887179851531982, + 0.0047387671656906605, + 0.005147107411175966, + 0.005041017197072506, + 0.014792977832257748, + 0.005990719888359308, + 0.014882346615195274, + 0.0059194425120949745, + 0.01014514360576868 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 8 + ], + "selected_reward": -0.008642058819532394 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005945615470409393 + ], + "opposite_logprob_tokens": [ + -5.130945682525635 + ], + "logprob_reward": -0.005945615470409393, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005945615470409393, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12832964956760406, + 0.0599856935441494, + 0.023320766165852547, + 0.12023478001356125, + 0.2772614359855652, + 0.0439460463821888, + 0.05485352873802185, + 0.07919362932443619, + 0.11804954707622528, + 0.09482493996620178 + ], + "credit_map": [ + 0.014495598152279854, + 0.01288518961519003, + 0.011839649640023708, + 0.01114843599498272, + 0.029920749366283417, + 0.02183878794312477, + 0.012749345041811466, + 0.013451887294650078, + 0.008173728361725807, + 0.004455633461475372, + 0.004713171161711216, + 0.00564261619001627, + 0.009823194704949856, + 0.013430384919047356, + 0.006425087805837393, + 0.0073218755424022675, + 0.002965350868180394, + 0.0032795665320008993, + 0.0018479267600923777, + 0.0023337749298661947, + 0.0024328185245394707, + 0.0029932979959994555, + 0.003704857314005494, + 0.0037631732411682606, + 0.028519803658127785, + 0.0062261177226901054, + 0.00586461229249835, + 0.007116101216524839, + 0.006963799707591534, + 0.010384348221123219, + 0.012538299895823002, + 0.04262169823050499, + 0.024017374962568283, + 0.019874708727002144, + 0.060476042330265045, + 0.0553756020963192, + 0.015950823202729225, + 0.026959192007780075, + 0.04529814422130585, + 0.029309537261724472, + 0.0031056045554578304, + 0.003384077688679099, + 0.003530909540131688, + 0.007045949809253216, + 0.0031622976530343294, + 0.007050350774079561, + 0.0084677217528224, + 0.008199131116271019, + 0.0028157627675682306, + 0.0034869809169322252, + 0.004475865513086319, + 0.004387824330478907, + 0.004294611979275942, + 0.013379393145442009, + 0.005153051111847162, + 0.016860036179423332, + 0.004933874122798443, + 0.008364113047719002, + 0.004030441399663687, + 0.0052909525111317635, + 0.011631251312792301, + 0.012114566750824451, + 0.01363511011004448, + 0.019193317741155624, + 0.013116770423948765, + 0.007678727153688669, + 0.012217984534800053, + 0.02293327823281288, + 0.008495138958096504, + 0.02861734852194786, + 0.01139007043093443, + 0.013600227423012257, + 0.01180068776011467, + 0.009119887836277485, + 0.008423663675785065, + 0.01583562232553959, + 0.008836553432047367, + 0.012669211253523827, + 0.012547139078378677, + 0.015592172741889954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.005945615470409393 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a submerged cityscape with skyscrapers featuring algae-stained glass and steel facades?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are enormous bell-shaped creatures made of translucent, gelatinous material visible performing a courtship dance?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the creatures have fiber-optic tentacles emitting a shifting inner light that pulses from cyan to fuchsia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera drift forward with a slow, weightless, dreamlike stability through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are vibrant neon pinks and electric blues from holographic advertisements casting oscillating reflections on the submerged structures?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16477970778942108, + 0.11877518892288208, + 0.08274480700492859, + 0.1965113878250122, + 0.1356469690799713, + 0.03570980578660965, + 0.11228033900260925, + 0.049543336033821106, + 0.048595454543828964, + 0.05541301518678665 + ], + "credit_map": [ + 0.021531471982598305, + 0.012514675967395306, + 0.019166816025972366, + 0.010554811917245388, + 0.04311255365610123, + 0.031390029937028885, + 0.01174594834446907, + 0.01476338692009449, + 0.016130900010466576, + 0.006582039408385754, + 0.008141945116221905, + 0.006747346371412277, + 0.015060985460877419, + 0.03731529787182808, + 0.013061793521046638, + 0.01573486253619194, + 0.008417541161179543, + 0.007205551024526358, + 0.00697069289162755, + 0.005691356025636196, + 0.006599378772079945, + 0.013642147183418274, + 0.009306719526648521, + 0.024911411106586456, + 0.06438672542572021, + 0.009565835818648338, + 0.015554746612906456, + 0.007628298830240965, + 0.00953007023781538, + 0.01598375476896763, + 0.023287560790777206, + 0.05057438090443611, + 0.011791720055043697, + 0.01019873283803463, + 0.028174325823783875, + 0.028145845979452133, + 0.008370448835194111, + 0.011665626429021358, + 0.021147239953279495, + 0.016153009608387947, + 0.0039473846554756165, + 0.002939653815701604, + 0.003792512696236372, + 0.00572115508839488, + 0.0036539025604724884, + 0.005007800180464983, + 0.004968380555510521, + 0.005679011810570955, + 0.006412453483790159, + 0.005479938816279173, + 0.006523406598716974, + 0.005385659635066986, + 0.008366426452994347, + 0.03574656695127487, + 0.009252178482711315, + 0.03511369228363037, + 0.0028949109837412834, + 0.004582987166941166, + 0.002455668291077018, + 0.005386115983128548, + 0.008368664421141148, + 0.006680582650005817, + 0.007692829240113497, + 0.011481576599180698, + 0.0033854825887829065, + 0.0036638458259403706, + 0.0043776067905128, + 0.007364481221884489, + 0.007590810302644968, + 0.00844371784478426, + 0.008474976755678654, + 0.005294528324157, + 0.006245227064937353, + 0.005414888728410006, + 0.005399846006184816, + 0.010520726442337036, + 0.004449322819709778, + 0.008964469656348228, + 0.0045415256172418594, + 0.00987700279802084 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.014190951362252235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_tensor([8066, 8066])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_tensor([8066, 8066])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c018c9ce2552e8a419410f31ee5046edda3ab87c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_tensor([8066, 8066])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The scene unfolds in a medium shot at eye level, the camera beginning a slow, almost imperceptible pan to a young man and woman sharing a moment on a large, weathered wooden swing suspended from the sturdy branches of an old tree. The man, with slightly tousled dark brown hair and a comfortable olive green Henley shirt, leans earnestly toward his companion, his voice low and sincere. The woman, her long honey-blond hair softly illuminated by the setting sun, wears a loose-fitting cream-colored knit cardigan over a simple floral dress; she listens intently, her head slightly tilted, a look of deep understanding on her face. They sit in a lush, verdant backyard in golden hour, the background melting into a beautiful bokeh of distant, out-of-focus foliage. Between them, on the swing's wooden planks, their hands are mere inches apart, his hand motionless and open, her fingers slightly curled, the tiny space between them crackling with unspoken emotion. As the camera continues its steady, deliberate close-in, the composition tightens, excluding the wider surroundings to focus purely on the vibrant space between their faces and hands, capturing the quiet intimacy and heartfelt seriousness of their conversation in an atmosphere of tranquility, warmth, and deep romance.", + "seed": 8253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8253_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.018142165336757898, + "vlm_logprob": -0.018142165336757898, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.018142165336757898, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006917811930179596 + ], + "opposite_logprob_tokens": [ + -5.881917953491211 + ], + "logprob_reward": -0.006917811930179596, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.006917811930179596, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1380821317434311, + 0.029822997748851776, + 0.05507679656147957, + 0.051132481545209885, + 0.11139608174562454, + 0.18872544169425964, + 0.13763613998889923, + 0.10482344776391983, + 0.13647745549678802, + 0.0468270406126976 + ], + "credit_map": [ + 0.02035382203757763, + 0.017738057300448418, + 0.016408585011959076, + 0.01522475853562355, + 0.018120411783456802, + 0.01992340013384819, + 0.012443158775568008, + 0.017869923263788223, + 0.0043511101976037025, + 0.004029640927910805, + 0.0038834000006318092, + 0.0026452599558979273, + 0.0037131812423467636, + 0.004527961369603872, + 0.0019949437119066715, + 0.004677500110119581, + 0.010409701615571976, + 0.0033530446235090494, + 0.017735518515110016, + 0.003617977723479271, + 0.00486636022105813, + 0.0032400526106357574, + 0.003239950630813837, + 0.008614188060164452, + 0.0062763202004134655, + 0.007012456655502319, + 0.0061918278224766254, + 0.004227363970130682, + 0.008904585614800453, + 0.00653360690921545, + 0.003865714417770505, + 0.00812060758471489, + 0.03195272758603096, + 0.0057060676626861095, + 0.039127666503190994, + 0.011480535380542278, + 0.006374948658049107, + 0.003920806106179953, + 0.005663703195750713, + 0.0071696327067911625, + 0.009151842445135117, + 0.012083278968930244, + 0.01587967574596405, + 0.06212329491972923, + 0.00937386229634285, + 0.019686568528413773, + 0.01002981886267662, + 0.050397105515003204, + 0.01061940286308527, + 0.01759503036737442, + 0.015075263567268848, + 0.02622833661735058, + 0.021604103967547417, + 0.015258296392858028, + 0.01331005897372961, + 0.01794564351439476, + 0.007184523157775402, + 0.01114303432404995, + 0.015385561622679234, + 0.013457010500133038, + 0.02090482786297798, + 0.02261464297771454, + 0.006641268264502287, + 0.007492575328797102, + 0.013086915016174316, + 0.011581309139728546, + 0.018626797944307327, + 0.01795952022075653, + 0.023102451115846634, + 0.02398941107094288, + 0.012834165245294571, + 0.015296892262995243, + 0.006346897222101688, + 0.007791893556714058, + 0.005500510800629854, + 0.007796086836606264, + 0.005509973503649235, + 0.0047428058460354805, + 0.004462249111384153, + 0.004676624201238155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.006917811930179596 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019564551766961813 + ], + "opposite_logprob_tokens": [ + -6.251956462860107 + ], + "logprob_reward": -0.0019564551766961813, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019564551766961813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13419882953166962, + 0.03746446594595909, + 0.04714272543787956, + 0.059415485709905624, + 0.18553172051906586, + 0.08268015086650848, + 0.15166425704956055, + 0.05300696939229965, + 0.1306162327528, + 0.11827924847602844 + ], + "credit_map": [ + 0.017632635310292244, + 0.01925663836300373, + 0.021946284919977188, + 0.019080517813563347, + 0.014319187961518764, + 0.013346967287361622, + 0.01497549470514059, + 0.013641104102134705, + 0.005334491841495037, + 0.004462392535060644, + 0.004839665722101927, + 0.002873888472095132, + 0.00737534137442708, + 0.0047475434839725494, + 0.003943204414099455, + 0.003887936007231474, + 0.00740106450393796, + 0.003573743859305978, + 0.01220260001718998, + 0.0033912486396729946, + 0.004395931493490934, + 0.003773803822696209, + 0.004134721588343382, + 0.008269605226814747, + 0.00713011110201478, + 0.00963167380541563, + 0.006172319408506155, + 0.007315652444958687, + 0.006948123686015606, + 0.008634083904325962, + 0.0053678047843277454, + 0.008215712383389473, + 0.06191263347864151, + 0.007254430558532476, + 0.06828641891479492, + 0.017159219831228256, + 0.008198628202080727, + 0.00791209563612938, + 0.006040288135409355, + 0.008767984807491302, + 0.004758069291710854, + 0.0063730161637067795, + 0.009224405512213707, + 0.02560409903526306, + 0.006529046222567558, + 0.00831084605306387, + 0.007008866872638464, + 0.014871793799102306, + 0.0065148696303367615, + 0.014387287199497223, + 0.013138698413968086, + 0.05286689102649689, + 0.0088175805285573, + 0.01237540878355503, + 0.010039239190518856, + 0.03352426737546921, + 0.0032869011629372835, + 0.007092856802046299, + 0.0058771767653524876, + 0.006120095495134592, + 0.015705931931734085, + 0.0054497020319104195, + 0.004029558505862951, + 0.005444743670523167, + 0.011820146813988686, + 0.0093418899923563, + 0.01293072197586298, + 0.01984429731965065, + 0.017703773453831673, + 0.034044429659843445, + 0.009672047570347786, + 0.015258915722370148, + 0.012266255915164948, + 0.01958559826016426, + 0.014631080441176891, + 0.022534426301717758, + 0.015041079372167587, + 0.009895802475512028, + 0.010142798535525799, + 0.014182205311954021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.0019564551766961813 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15541008114814758, + 0.11014936119318008, + 0.06321430206298828, + 0.09391455352306366, + 0.17916318774223328, + 0.16957074403762817, + 0.0744507908821106, + 0.06398844718933105, + 0.04896058514714241, + 0.04117801412940025 + ], + "credit_map": [ + 0.02208392135798931, + 0.017738666385412216, + 0.024152422323822975, + 0.016771724447607994, + 0.023933541029691696, + 0.017355667427182198, + 0.01631215400993824, + 0.017061974853277206, + 0.017305247485637665, + 0.016179021447896957, + 0.0174208115786314, + 0.009225896559655666, + 0.010543117299675941, + 0.020373022183775902, + 0.005615163128823042, + 0.013487076386809349, + 0.009609225206077099, + 0.004955977667123079, + 0.01500844769179821, + 0.004599872976541519, + 0.0072680567391216755, + 0.004989346489310265, + 0.0039879498071968555, + 0.012795417569577694, + 0.01091455016285181, + 0.014242911711335182, + 0.014765086583793163, + 0.009301888756453991, + 0.012298429384827614, + 0.014186867512762547, + 0.007561656180769205, + 0.010643156245350838, + 0.0618739128112793, + 0.008940869942307472, + 0.06632685661315918, + 0.016046805307269096, + 0.005854359362274408, + 0.006879846099764109, + 0.005850666668266058, + 0.007389839272946119, + 0.007316266186535358, + 0.010926131159067154, + 0.013748319819569588, + 0.05332339182496071, + 0.008056728169322014, + 0.016028909012675285, + 0.009633130393922329, + 0.0505378395318985, + 0.005684450268745422, + 0.008440556935966015, + 0.00723810913041234, + 0.007958401925861835, + 0.012613445520401001, + 0.012532607652246952, + 0.008890660479664803, + 0.01109254639595747, + 0.005611184053122997, + 0.006704708561301231, + 0.0070102633908391, + 0.008361897431313992, + 0.010651852004230022, + 0.01702146977186203, + 0.004018121398985386, + 0.004608944058418274, + 0.005258801393210888, + 0.0036942358128726482, + 0.006808810867369175, + 0.008277243934571743, + 0.007202691398561001, + 0.00784419383853674, + 0.00425636675208807, + 0.005618235561996698, + 0.0040471553802490234, + 0.005249924026429653, + 0.005836857948452234, + 0.006271714344620705, + 0.005481121130287647, + 0.007317259907722473, + 0.0027523820754140615, + 0.004221595358103514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.0036162238102406263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023287026211619377 + ], + "opposite_logprob_tokens": [ + -3.773287057876587 + ], + "logprob_reward": -0.023287026211619377, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023287026211619377, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13144566118717194, + 0.026692623272538185, + 0.02763528563082218, + 0.051766905933618546, + 0.24163371324539185, + 0.09300937503576279, + 0.10109309107065201, + 0.07962460070848465, + 0.09682372957468033, + 0.15027500689029694 + ], + "credit_map": [ + 0.01740744151175022, + 0.015512371435761452, + 0.020423999056220055, + 0.020983722060918808, + 0.01819441467523575, + 0.012546968646347523, + 0.013833921402692795, + 0.012542814947664738, + 0.0038221587892621756, + 0.002692079171538353, + 0.00464610755443573, + 0.0026683879550546408, + 0.003928366117179394, + 0.0025440289173275232, + 0.0026166681200265884, + 0.003774826880544424, + 0.004462918732315302, + 0.0023263159673660994, + 0.006262569222599268, + 0.002221104921773076, + 0.0030163698829710484, + 0.002880038460716605, + 0.002673804759979248, + 0.0037921646144241095, + 0.005646754987537861, + 0.00530316773802042, + 0.0067445058375597, + 0.008367895148694515, + 0.009015602990984917, + 0.005266223568469286, + 0.004399669822305441, + 0.007023086305707693, + 0.08438204973936081, + 0.010172064416110516, + 0.08905026316642761, + 0.02151207998394966, + 0.010618236847221851, + 0.00846353080123663, + 0.007006854750216007, + 0.010428638197481632, + 0.005043407436460257, + 0.006880901288241148, + 0.009511915035545826, + 0.025630544871091843, + 0.008384104818105698, + 0.012395331636071205, + 0.006222864612936974, + 0.018940309062600136, + 0.005381514318287373, + 0.010045859031379223, + 0.0100982915610075, + 0.030106423422694206, + 0.006961531471461058, + 0.01387198455631733, + 0.007430045399814844, + 0.01719743385910988, + 0.006300780922174454, + 0.009451835416257381, + 0.009838827885687351, + 0.011459252797067165, + 0.015419794246554375, + 0.014810299500823021, + 0.005465751979500055, + 0.006878060754388571, + 0.009286334738135338, + 0.009859858080744743, + 0.009182408452033997, + 0.012778913602232933, + 0.014342417940497398, + 0.018682390451431274, + 0.009211351163685322, + 0.013480052351951599, + 0.01748507097363472, + 0.024001499637961388, + 0.01548799593001604, + 0.03355926275253296, + 0.015844471752643585, + 0.013524001464247704, + 0.013587852008640766, + 0.016784854233264923 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 0 + ], + "selected_reward": -0.023287026211619377 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493330955505371 + ], + "opposite_logprob_tokens": [ + -2.9299333095550537 + ], + "logprob_reward": -0.05493330955505371, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493330955505371, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05505136400461197, + 0.047950100153684616, + 0.046518474817276, + 0.02681894600391388, + 0.23977859318256378, + 0.05704160034656525, + 0.14677290618419647, + 0.05986331030726433, + 0.13639652729034424, + 0.18380814790725708 + ], + "credit_map": [ + 0.0077045210637152195, + 0.007396963890641928, + 0.006198914255946875, + 0.006610835436731577, + 0.00782072450965643, + 0.006663383916020393, + 0.006021772511303425, + 0.006634248420596123, + 0.00781329907476902, + 0.0057984148152172565, + 0.005629288963973522, + 0.003762901294976473, + 0.007055093999952078, + 0.006968675646930933, + 0.00416581192985177, + 0.006756609305739403, + 0.006067686714231968, + 0.0033454797230660915, + 0.009520264342427254, + 0.0032613989897072315, + 0.006422028876841068, + 0.0037742145359516144, + 0.003635738044977188, + 0.010491662658751011, + 0.0033353001344949007, + 0.0027762698009610176, + 0.0034152783919125795, + 0.004584059584885836, + 0.003221849212422967, + 0.0028014760464429855, + 0.0030820025131106377, + 0.0036027103196829557, + 0.08320758491754532, + 0.008779378607869148, + 0.0934268906712532, + 0.022409703582525253, + 0.007296810392290354, + 0.008189263753592968, + 0.006733769550919533, + 0.009735179133713245, + 0.0045341975055634975, + 0.005624442361295223, + 0.005516572389751673, + 0.010158946737647057, + 0.007891031913459301, + 0.009025363251566887, + 0.005967843346297741, + 0.008323200978338718, + 0.007407383527606726, + 0.015757165849208832, + 0.013350146822631359, + 0.051326557993888855, + 0.007462367881089449, + 0.010836745612323284, + 0.00884188711643219, + 0.031790655106306076, + 0.004971515387296677, + 0.00570592749863863, + 0.006635281722992659, + 0.00930178351700306, + 0.009119423106312752, + 0.015215729363262653, + 0.0038356271106749773, + 0.0050780256278812885, + 0.013745809905230999, + 0.013229619711637497, + 0.01311908382922411, + 0.018619833514094353, + 0.025313254445791245, + 0.019748913124203682, + 0.013247721828520298, + 0.01937228813767433, + 0.021444454789161682, + 0.030557820573449135, + 0.02233182266354561, + 0.039805103093385696, + 0.020616207271814346, + 0.013860787265002728, + 0.01640978641808033, + 0.018782176077365875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 6 + ], + "selected_reward": -0.05493330955505371 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium shot at eye level, the camera beginning a slow, almost imperceptible pan to a young man and woman sharing a moment on a large, weathered wooden swing suspended from the sturdy branches of an old tree. The man, with slightly tousled dark brown hair and a comfortable olive green Henley shirt, leans earnestly toward his companion, his voice low and sincere. The woman, her long honey-blond hair softly illuminated by the setting sun, wears a loose-fitting cream-colored knit cardigan over a simple floral dress; she listens intently, her head slightly tilted, a look of deep understanding on her face. They sit in a lush, verdant backyard in golden hour, the background melting into a beautiful bokeh of distant, out-of-focus foliage. Between them, on the swing's wooden planks, their hands are mere inches apart, his hand motionless and open, her fingers slightly curled, the tiny space between them crackling with unspoken emotion. As the camera continues its steady, deliberate close-in, the composition tightens, excluding the wider surroundings to focus purely on the vibrant space between their faces and hands, capturing the quiet intimacy and heartfelt seriousness of their conversation in an atmosphere of tranquility, warmth, and deep romance.", + "seed": 8254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8254_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016885483474470674, + "vlm_logprob": -0.016885483474470674, + "vlm_yes_no_margin": 4.825, + "vlm_token_credit": -0.016885483474470674, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010588772594928741 + ], + "opposite_logprob_tokens": [ + -5.385588645935059 + ], + "logprob_reward": -0.010588772594928741, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.010588772594928741, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08437783271074295, + 0.1394677609205246, + 0.056919876486063004, + 0.05528683960437775, + 0.05783737823367119, + 0.029043223708868027, + 0.4746655821800232, + 0.03569291904568672, + 0.04309680312871933, + 0.023611783981323242 + ], + "credit_map": [ + 0.008969085291028023, + 0.007315639406442642, + 0.011306541971862316, + 0.009355533868074417, + 0.010099535807967186, + 0.018234962597489357, + 0.009035836905241013, + 0.010060698725283146, + 0.01112582441419363, + 0.004238881170749664, + 0.04767051711678505, + 0.007658786606043577, + 0.004459409974515438, + 0.008173038251698017, + 0.011104682460427284, + 0.045036617666482925, + 0.00542473467066884, + 0.0052197654731571674, + 0.008381813764572144, + 0.00827872846275568, + 0.008648551069200039, + 0.005432602949440479, + 0.0072952620685100555, + 0.008238421753048897, + 0.005330977030098438, + 0.004165175836533308, + 0.014696763828396797, + 0.002960951765999198, + 0.003779730526730418, + 0.00407613068819046, + 0.00628524599596858, + 0.013991864398121834, + 0.004771059844642878, + 0.006732186768203974, + 0.004719156306236982, + 0.006891873199492693, + 0.006381336599588394, + 0.005516772158443928, + 0.013542977161705494, + 0.009282011538743973, + 0.004056189674884081, + 0.0029478981159627438, + 0.0049517396837472916, + 0.0031196284107863903, + 0.002799795940518379, + 0.002547134645283222, + 0.0043975720182061195, + 0.004223265219479799, + 0.04060495272278786, + 0.03161443769931793, + 0.028616346418857574, + 0.11112193018198013, + 0.04402803257107735, + 0.05208110064268112, + 0.049979522824287415, + 0.11661926656961441, + 0.0034412655513733625, + 0.0035886222030967474, + 0.0049506607465445995, + 0.006264891009777784, + 0.004890307318419218, + 0.0033912474755197763, + 0.003554981667548418, + 0.005610941909253597, + 0.004349264781922102, + 0.008642587810754776, + 0.005510980263352394, + 0.00655282661318779, + 0.0034880114253610373, + 0.0028223637491464615, + 0.00570242665708065, + 0.006028345786035061, + 0.0022641473915427923, + 0.0028513569850474596, + 0.002134242095053196, + 0.005037439987063408, + 0.0026073195040225983, + 0.003023692173883319, + 0.0025657007936388254, + 0.0031278852839022875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.010588772594928741 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011972171487286687 + ], + "opposite_logprob_tokens": [ + -6.751197338104248 + ], + "logprob_reward": -0.0011972171487286687, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011972171487286687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1514522135257721, + 0.19565197825431824, + 0.08305884897708893, + 0.08964703977108002, + 0.10760975629091263, + 0.09397529810667038, + 0.1476041078567505, + 0.033666376024484634, + 0.04196983203291893, + 0.05536460876464844 + ], + "credit_map": [ + 0.011187460273504257, + 0.010700307786464691, + 0.0220518559217453, + 0.017597079277038574, + 0.023460665717720985, + 0.022784339264035225, + 0.029368164017796516, + 0.014302339404821396, + 0.015710530802607536, + 0.007531014271080494, + 0.06944884359836578, + 0.01227300800383091, + 0.008141640573740005, + 0.011149738915264606, + 0.01975918374955654, + 0.051638029515743256, + 0.0065655698999762535, + 0.006299007683992386, + 0.010577664710581303, + 0.007245148066431284, + 0.012208255007863045, + 0.006671166513115168, + 0.012386557646095753, + 0.021105479449033737, + 0.009610285982489586, + 0.008211083710193634, + 0.018224429339170456, + 0.007578801363706589, + 0.011751162819564342, + 0.008450151421129704, + 0.010265485383570194, + 0.01555564347654581, + 0.008085222914814949, + 0.01054651103913784, + 0.0076217991299927235, + 0.010480836033821106, + 0.011922583915293217, + 0.00906694121658802, + 0.028774444013834, + 0.021111415699124336, + 0.010832623578608036, + 0.01567845046520233, + 0.009172096848487854, + 0.011344034224748611, + 0.013353150337934494, + 0.009258933365345001, + 0.009705755859613419, + 0.014630250632762909, + 0.01276413258165121, + 0.01427316665649414, + 0.009067462757229805, + 0.02988271601498127, + 0.014756711199879646, + 0.01819179207086563, + 0.01606159284710884, + 0.03260653465986252, + 0.0026609948836266994, + 0.00342852296307683, + 0.0024685533717274666, + 0.003882177174091339, + 0.0063926586881279945, + 0.004628196358680725, + 0.005290440283715725, + 0.004914829973131418, + 0.0038336787838488817, + 0.00789094902575016, + 0.0026856977492570877, + 0.007518136408179998, + 0.004849554505199194, + 0.004093896131962538, + 0.005187066271901131, + 0.005910851527005434, + 0.005575450602918863, + 0.009486464783549309, + 0.004662164021283388, + 0.008809954859316349, + 0.005078278947621584, + 0.004903221037238836, + 0.007215116638690233, + 0.009633958339691162 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0011972171487286687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005252371542155743 + ], + "opposite_logprob_tokens": [ + -5.255252361297607 + ], + "logprob_reward": -0.005252371542155743, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005252371542155743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11677449196577072, + 0.1536560356616974, + 0.06978821754455566, + 0.07178962975740433, + 0.06800330430269241, + 0.04192979633808136, + 0.3757585883140564, + 0.034139443188905716, + 0.0408049114048481, + 0.027355635538697243 + ], + "credit_map": [ + 0.008762993849813938, + 0.00862049963325262, + 0.030394189059734344, + 0.013685829006135464, + 0.008785807527601719, + 0.021276824176311493, + 0.011697077192366123, + 0.013551270589232445, + 0.011885070241987705, + 0.007317818235605955, + 0.04641880467534065, + 0.014111892320215702, + 0.007917055860161781, + 0.010696856305003166, + 0.014814894646406174, + 0.04049363359808922, + 0.006975071504712105, + 0.006531447637826204, + 0.009183445014059544, + 0.00503101572394371, + 0.009842067956924438, + 0.006155722308903933, + 0.010545059107244015, + 0.01552438922226429, + 0.004352923482656479, + 0.0052797566168010235, + 0.019206199795007706, + 0.004403953906148672, + 0.005720216315239668, + 0.005444647744297981, + 0.009178351610898972, + 0.018203584477305412, + 0.006614929996430874, + 0.005033410154283047, + 0.004552359227091074, + 0.006170849781483412, + 0.005033434368669987, + 0.005798484198749065, + 0.02045946568250656, + 0.01434036809951067, + 0.003984540235251188, + 0.007403775118291378, + 0.003483743406832218, + 0.005850745365023613, + 0.006373178213834763, + 0.0036129478830844164, + 0.005456386599689722, + 0.005764477886259556, + 0.0315665565431118, + 0.024424897506833076, + 0.022002797573804855, + 0.08692438900470734, + 0.03605879470705986, + 0.0410040058195591, + 0.03988353908061981, + 0.09389359503984451, + 0.003257335629314184, + 0.002906949259340763, + 0.0071205999702215195, + 0.004138140473514795, + 0.004014152102172375, + 0.002726539270952344, + 0.004939116537570953, + 0.005036613903939724, + 0.004601289052516222, + 0.007392809726297855, + 0.0052140443585813046, + 0.005013900808990002, + 0.00438676355406642, + 0.0031237960793077946, + 0.0037941213231533766, + 0.0072781844064593315, + 0.00278921565040946, + 0.004077952820807695, + 0.0023410224821418524, + 0.005083200056105852, + 0.003218925092369318, + 0.0027563527692109346, + 0.003132820362225175, + 0.0039561474695801735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 0 + ], + "selected_reward": -0.005252371542155743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370225057005882 + ], + "opposite_logprob_tokens": [ + -3.4087021350860596 + ], + "logprob_reward": -0.03370225057005882, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370225057005882, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13892006874084473, + 0.23670341074466705, + 0.10974302887916565, + 0.04887049272656441, + 0.054482217878103256, + 0.06668935716152191, + 0.13240286707878113, + 0.06391838192939758, + 0.11200783401727676, + 0.03626236319541931 + ], + "credit_map": [ + 0.011660559102892876, + 0.008120867423713207, + 0.038572121411561966, + 0.011871668510138988, + 0.012064016424119473, + 0.033817704766988754, + 0.01316562294960022, + 0.009647509083151817, + 0.018866633996367455, + 0.009596546180546284, + 0.08520862460136414, + 0.018024016171693802, + 0.008586708456277847, + 0.013470884412527084, + 0.023208482190966606, + 0.059741511940956116, + 0.011306761763989925, + 0.013249128125607967, + 0.013916530646383762, + 0.014236310496926308, + 0.015776334330439568, + 0.01130910124629736, + 0.014767499640583992, + 0.015181354247033596, + 0.00414890144020319, + 0.0054159099236130714, + 0.007066224701702595, + 0.005388342775404453, + 0.006979783996939659, + 0.004150024615228176, + 0.008306806907057762, + 0.007414498366415501, + 0.006476986221969128, + 0.005395965650677681, + 0.0043899258598685265, + 0.005570508074015379, + 0.006650757044553757, + 0.006036624778062105, + 0.012399870902299881, + 0.007561578880995512, + 0.0057081785053014755, + 0.006445737089961767, + 0.010238464921712875, + 0.007369906175881624, + 0.00676214462146163, + 0.00868451688438654, + 0.011258793994784355, + 0.010221616365015507, + 0.012162535451352596, + 0.013812066987156868, + 0.00813711155205965, + 0.04025433957576752, + 0.010433525778353214, + 0.007278186269104481, + 0.011564692482352257, + 0.028760414570569992, + 0.007574834860861301, + 0.006681075319647789, + 0.008029961958527565, + 0.012574339285492897, + 0.006418926641345024, + 0.004907027818262577, + 0.007052995264530182, + 0.010679219849407673, + 0.01700393669307232, + 0.016032196581363678, + 0.01522684097290039, + 0.016470011323690414, + 0.01097873691469431, + 0.008683927357196808, + 0.012335040606558323, + 0.015277151018381119, + 0.004341596737504005, + 0.005956151522696018, + 0.0030346254352480173, + 0.0055009848438203335, + 0.0037600169889628887, + 0.004092333372682333, + 0.004364862106740475, + 0.00521179148927331 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.03370225057005882 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0336868055164814 + ], + "opposite_logprob_tokens": [ + -3.408686876296997 + ], + "logprob_reward": -0.0336868055164814, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.0336868055164814, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0697093978524208, + 0.12691323459148407, + 0.057986851781606674, + 0.0470365546643734, + 0.0780714601278305, + 0.05487769842147827, + 0.29941031336784363, + 0.11827708780765533, + 0.11207272112369537, + 0.03564465045928955 + ], + "credit_map": [ + 0.004802652634680271, + 0.0043433536775410175, + 0.01466080266982317, + 0.0075432839803397655, + 0.00705455057322979, + 0.013884971849620342, + 0.01069877203553915, + 0.006721015553921461, + 0.008334648795425892, + 0.0056928531266748905, + 0.02911338023841381, + 0.007981494069099426, + 0.005761088337749243, + 0.010042238980531693, + 0.014857920818030834, + 0.04512961208820343, + 0.005499985069036484, + 0.005658349487930536, + 0.00686892494559288, + 0.006925360765308142, + 0.008033821359276772, + 0.006720434874296188, + 0.0076461974531412125, + 0.010633779689669609, + 0.0048941923305392265, + 0.007147693540900946, + 0.005160166881978512, + 0.0049437289126217365, + 0.00505032017827034, + 0.004504076670855284, + 0.007669251877814531, + 0.007667128462344408, + 0.005539561156183481, + 0.006828058045357466, + 0.006285870913416147, + 0.007708143908530474, + 0.00517983827739954, + 0.005797363817691803, + 0.02297516167163849, + 0.017757467925548553, + 0.004472923465073109, + 0.004801657050848007, + 0.0041168248280882835, + 0.00669438298791647, + 0.009769986383616924, + 0.007757607381790876, + 0.009759764187037945, + 0.007504554931074381, + 0.04099353402853012, + 0.026478281244635582, + 0.024414680898189545, + 0.05412554740905762, + 0.027390291914343834, + 0.023796524852514267, + 0.029600240290164948, + 0.0726112350821495, + 0.01021733507514, + 0.012042290531098843, + 0.011901489458978176, + 0.027740448713302612, + 0.013464570045471191, + 0.00810064934194088, + 0.011309522204101086, + 0.02350078523159027, + 0.010873689316213131, + 0.02916969358921051, + 0.007367509417235851, + 0.016224540770053864, + 0.007986838929355145, + 0.010199137963354588, + 0.01327704917639494, + 0.01697426475584507, + 0.003234722651541233, + 0.00546702416613698, + 0.003526417538523674, + 0.008216996677219868, + 0.0035560273099690676, + 0.003100341185927391, + 0.0034045639913529158, + 0.005138557404279709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.0336868055164814 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium shot at eye level, the camera beginning a slow, almost imperceptible pan to a young man and woman sharing a moment on a large, weathered wooden swing suspended from the sturdy branches of an old tree. The man, with slightly tousled dark brown hair and a comfortable olive green Henley shirt, leans earnestly toward his companion, his voice low and sincere. The woman, her long honey-blond hair softly illuminated by the setting sun, wears a loose-fitting cream-colored knit cardigan over a simple floral dress; she listens intently, her head slightly tilted, a look of deep understanding on her face. They sit in a lush, verdant backyard in golden hour, the background melting into a beautiful bokeh of distant, out-of-focus foliage. Between them, on the swing's wooden planks, their hands are mere inches apart, his hand motionless and open, her fingers slightly curled, the tiny space between them crackling with unspoken emotion. As the camera continues its steady, deliberate close-in, the composition tightens, excluding the wider surroundings to focus purely on the vibrant space between their faces and hands, capturing the quiet intimacy and heartfelt seriousness of their conversation in an atmosphere of tranquility, warmth, and deep romance.", + "seed": 8255, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8255_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01524583324790001, + "vlm_logprob": -0.01524583324790001, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.01524583324790001, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01928742229938507 + ], + "opposite_logprob_tokens": [ + -4.394287586212158 + ], + "logprob_reward": -0.01928742229938507, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01928742229938507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10908234864473343, + 0.03794202208518982, + 0.038941558450460434, + 0.0348576195538044, + 0.13607056438922882, + 0.019755655899643898, + 0.0518907755613327, + 0.5223085880279541, + 0.029796304181218147, + 0.01935453899204731 + ], + "credit_map": [ + 0.007962006144225597, + 0.010383826680481434, + 0.016128458082675934, + 0.021017931401729584, + 0.015860317274928093, + 0.010818484239280224, + 0.011031528003513813, + 0.01587980054318905, + 0.0036228892859071493, + 0.004748943727463484, + 0.008658451028168201, + 0.005262626800686121, + 0.003624667413532734, + 0.0033859426621347666, + 0.003099620807915926, + 0.005538880359381437, + 0.0031335623934865, + 0.0034509962424635887, + 0.003728100098669529, + 0.007700440473854542, + 0.0041130948811769485, + 0.0035178884863853455, + 0.002333509735763073, + 0.010963968932628632, + 0.00301027437672019, + 0.004600717686116695, + 0.004729097243398428, + 0.005221521481871605, + 0.004392567090690136, + 0.003562027122825384, + 0.004631285089999437, + 0.004710126202553511, + 0.025253569707274437, + 0.014862359501421452, + 0.0280840415507555, + 0.023644961416721344, + 0.01249345950782299, + 0.008396587334573269, + 0.010131267830729485, + 0.013204333372414112, + 0.0026480078231543303, + 0.002453991211950779, + 0.003725086571648717, + 0.002856076695024967, + 0.0020888305734843016, + 0.0025978086050599813, + 0.0011689780512824655, + 0.002216876018792391, + 0.01179460808634758, + 0.005455507431179285, + 0.010839412920176983, + 0.008641485124826431, + 0.003714613150805235, + 0.00496502174064517, + 0.002357954392209649, + 0.0041221734136343, + 0.03177338466048241, + 0.03637033328413963, + 0.01976696588099003, + 0.12348447740077972, + 0.01810687780380249, + 0.19815000891685486, + 0.017595971003174782, + 0.07706054300069809, + 0.0036608469672501087, + 0.00507736811414361, + 0.003821729915216565, + 0.004144732374697924, + 0.002734168665483594, + 0.00601125368848443, + 0.00222500367090106, + 0.002121200319379568, + 0.0038211564533412457, + 0.002710010390728712, + 0.002432091161608696, + 0.0025115024764090776, + 0.0013490448473021388, + 0.003627136116847396, + 0.0014942992711439729, + 0.0014092979254201055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.01928742229938507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013610394671559334 + ], + "opposite_logprob_tokens": [ + -6.626360893249512 + ], + "logprob_reward": -0.0013610394671559334, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013610394671559334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6221685409545898, + 0.025071309879422188, + 0.029525188729166985, + 0.043621838092803955, + 0.012616755440831184, + 0.022296398878097534, + 0.08262335509061813, + 0.06994354724884033, + 0.050651587545871735, + 0.04148149490356445 + ], + "credit_map": [ + 0.039032209664583206, + 0.03333623707294464, + 0.06287344545125961, + 0.23952685296535492, + 0.02496449090540409, + 0.0399947352707386, + 0.02795519307255745, + 0.15448534488677979, + 0.0022635036148130894, + 0.0037588695995509624, + 0.005572437774389982, + 0.002904771128669381, + 0.0024994006380438805, + 0.0019979961216449738, + 0.0024177669547498226, + 0.00365656241774559, + 0.0031117110047489405, + 0.004023004788905382, + 0.0034675823990255594, + 0.004510648548603058, + 0.004819071386009455, + 0.0021657906472682953, + 0.004032263532280922, + 0.0033951164223253727, + 0.004131863359361887, + 0.011928330175578594, + 0.006531840655952692, + 0.004315756727010012, + 0.005022866651415825, + 0.0034277287777513266, + 0.00442708982154727, + 0.003836363088339567, + 0.0012597140157595277, + 0.00226199091412127, + 0.0018641032511368394, + 0.0013128193095326424, + 0.0015921536833047867, + 0.0010191441979259253, + 0.0017631564987823367, + 0.001543673686683178, + 0.0027591169346123934, + 0.0026529894676059484, + 0.0040597375482320786, + 0.0035903037060052156, + 0.0025132799055427313, + 0.002308802679181099, + 0.0013428609818220139, + 0.003069306258112192, + 0.01562479417771101, + 0.006524978205561638, + 0.026485411450266838, + 0.014842617325484753, + 0.004112301394343376, + 0.005663721356540918, + 0.004020345862954855, + 0.005349189043045044, + 0.006398065946996212, + 0.006428959779441357, + 0.005505881737917662, + 0.007309786509722471, + 0.004581828601658344, + 0.018168644979596138, + 0.003186688059940934, + 0.01836368814110756, + 0.005142025649547577, + 0.009333116933703423, + 0.005338830407708883, + 0.009103294461965561, + 0.0049113929271698, + 0.010342111811041832, + 0.0031018333975225687, + 0.0033789812587201595, + 0.004874023608863354, + 0.005953459069132805, + 0.005407403688877821, + 0.0061684236861765385, + 0.004811582155525684, + 0.006920012179762125, + 0.004849362187087536, + 0.002497229026630521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 7 + ], + "selected_reward": -0.0013610394671559334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024920618161559105 + ], + "opposite_logprob_tokens": [ + -6.0024919509887695 + ], + "logprob_reward": -0.0024920618161559105, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024920618161559105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2677415609359741, + 0.041630275547504425, + 0.2598118185997009, + 0.03918301686644554, + 0.049003761261701584, + 0.011232244782149792, + 0.08063255995512009, + 0.21750827133655548, + 0.019324183464050293, + 0.013932283967733383 + ], + "credit_map": [ + 0.01721011847257614, + 0.01647680439054966, + 0.029293818399310112, + 0.10349873453378677, + 0.012224612757563591, + 0.017114195972681046, + 0.01382856722921133, + 0.05809469521045685, + 0.003641959745436907, + 0.006858639884740114, + 0.008961736224591732, + 0.006096153520047665, + 0.004604280926287174, + 0.004002715460956097, + 0.003953878302127123, + 0.0035109028685837984, + 0.020561786368489265, + 0.03396967053413391, + 0.025096071884036064, + 0.0516754612326622, + 0.03440159559249878, + 0.02444758266210556, + 0.021180635318160057, + 0.04847900569438934, + 0.00361652672290802, + 0.006528081372380257, + 0.005917576607316732, + 0.004006228409707546, + 0.004913664422929287, + 0.004376564174890518, + 0.005155145190656185, + 0.004669226706027985, + 0.007788170129060745, + 0.007953284308314323, + 0.00901490543037653, + 0.007899428717792034, + 0.004866717383265495, + 0.003836060408502817, + 0.003624562406912446, + 0.004020632244646549, + 0.0015134414425119758, + 0.0021813337225466967, + 0.0018944708863273263, + 0.001095804269425571, + 0.0012069025542587042, + 0.0010363704059273005, + 0.000977255986072123, + 0.001326664350926876, + 0.017566606402397156, + 0.009047260507941246, + 0.013413648121058941, + 0.011440799571573734, + 0.006248758174479008, + 0.009756423532962799, + 0.003938495181500912, + 0.009220561012625694, + 0.014026992954313755, + 0.015360575169324875, + 0.008603865280747414, + 0.07843174785375595, + 0.008345472626388073, + 0.07180957496166229, + 0.008139708079397678, + 0.01279031578451395, + 0.0026658184360712767, + 0.002213326282799244, + 0.002701491815969348, + 0.0026258232537657022, + 0.001986087765544653, + 0.003799150697886944, + 0.0020341600757092237, + 0.0012983237393200397, + 0.002104071667417884, + 0.002155318856239319, + 0.0014864944387227297, + 0.002049581613391638, + 0.0014200186124071479, + 0.0021958404686301947, + 0.001525635365396738, + 0.0009953201515600085 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.0024920618161559105 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02329576015472412 + ], + "opposite_logprob_tokens": [ + -3.7732958793640137 + ], + "logprob_reward": -0.02329576015472412, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02329576015472412, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16720925271511078, + 0.05428073927760124, + 0.0858963280916214, + 0.059316590428352356, + 0.0337764210999012, + 0.06136773154139519, + 0.0901021733880043, + 0.20643053948879242, + 0.1509353667497635, + 0.09068487584590912 + ], + "credit_map": [ + 0.009305963292717934, + 0.0328366756439209, + 0.024467848241329193, + 0.022869478911161423, + 0.01833399571478367, + 0.014877351000905037, + 0.013005835004150867, + 0.031512100249528885, + 0.004211845807731152, + 0.00958381686359644, + 0.006459505762904882, + 0.008348710834980011, + 0.0068390583619475365, + 0.005087433382868767, + 0.006769764702767134, + 0.006980604957789183, + 0.004867649171501398, + 0.010576045140624046, + 0.010037465952336788, + 0.021900514140725136, + 0.010054342448711395, + 0.009333498775959015, + 0.00861238781362772, + 0.010514420457184315, + 0.006451714783906937, + 0.008210865780711174, + 0.01052559819072485, + 0.006537551060318947, + 0.008454080671072006, + 0.005223229061812162, + 0.00596211850643158, + 0.007951428182423115, + 0.0037722191773355007, + 0.005381113849580288, + 0.006263261195272207, + 0.004061048850417137, + 0.005536451004445553, + 0.0027178514283150434, + 0.003598345210775733, + 0.002446131082251668, + 0.005904842168092728, + 0.008564109914004803, + 0.011625239625573158, + 0.00764379370957613, + 0.006397754419595003, + 0.008272248320281506, + 0.004232988692820072, + 0.008726757951080799, + 0.01746680773794651, + 0.008441084995865822, + 0.02518482133746147, + 0.016384750604629517, + 0.006421568337827921, + 0.005820297170430422, + 0.004711133427917957, + 0.005671706050634384, + 0.017166929319500923, + 0.01474129967391491, + 0.009641431272029877, + 0.06627634912729263, + 0.010930025950074196, + 0.02260342426598072, + 0.011138774454593658, + 0.0539323054254055, + 0.01722688600420952, + 0.029316971078515053, + 0.01486909482628107, + 0.025788381695747375, + 0.014184310100972652, + 0.029749659821391106, + 0.010285879485309124, + 0.009514193050563335, + 0.011917827650904655, + 0.01479237899184227, + 0.006641815882176161, + 0.011389018967747688, + 0.017692821100354195, + 0.015460861846804619, + 0.007467438466846943, + 0.005322709213942289 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.02329576015472412 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979288250207901 + ], + "opposite_logprob_tokens": [ + -3.5297927856445312 + ], + "logprob_reward": -0.02979288250207901, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979288250207901, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27927708625793457, + 0.02473016269505024, + 0.07583214342594147, + 0.05038480460643768, + 0.020161794498562813, + 0.031144119799137115, + 0.04623446241021156, + 0.3189525008201599, + 0.05250260978937149, + 0.10078027099370956 + ], + "credit_map": [ + 0.01719701662659645, + 0.02262917160987854, + 0.027980821207165718, + 0.09758877009153366, + 0.014747442677617073, + 0.01752678118646145, + 0.014473396353423595, + 0.06713370233774185, + 0.0025369394570589066, + 0.0035195008385926485, + 0.005506171379238367, + 0.003270806511864066, + 0.0031152372248470783, + 0.001653345418162644, + 0.003333164844661951, + 0.0017949966713786125, + 0.00553573714569211, + 0.01295112818479538, + 0.008123481646180153, + 0.012372416444122791, + 0.014212074689567089, + 0.007143716793507338, + 0.006163857411593199, + 0.009329731576144695, + 0.004210871644318104, + 0.009800144471228123, + 0.007556537631899118, + 0.004919332917779684, + 0.008414406329393387, + 0.0036553635727614164, + 0.007388292346149683, + 0.004439860116690397, + 0.0020854140166193247, + 0.00386770605109632, + 0.002719779033213854, + 0.002927713328972459, + 0.002679224591702223, + 0.0018236087635159492, + 0.002239689463749528, + 0.0018186604138463736, + 0.002254299819469452, + 0.0034549415577203035, + 0.004293832927942276, + 0.003060801886022091, + 0.003004908561706543, + 0.006346363574266434, + 0.0023057395592331886, + 0.0064232321456074715, + 0.007608100306242704, + 0.005571542773395777, + 0.008037122897803783, + 0.006695388816297054, + 0.0034868784714490175, + 0.0055083767510950565, + 0.0030729298014193773, + 0.006254120264202356, + 0.019914260134100914, + 0.02346564084291458, + 0.012729405425488949, + 0.11030657589435577, + 0.012707621790468693, + 0.10946010798215866, + 0.011916262097656727, + 0.018452612683176994, + 0.007705459836870432, + 0.008622991852462292, + 0.006340821273624897, + 0.009393655695021152, + 0.005511599592864513, + 0.008510364219546318, + 0.003295314032584429, + 0.0031224016565829515, + 0.013585959561169147, + 0.01600712165236473, + 0.007914329878985882, + 0.013149629347026348, + 0.01821061596274376, + 0.01811491884291172, + 0.00821438618004322, + 0.005583304446190596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.02979288250207901 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The scene unfolds in a medium shot at eye level, the camera beginning a slow, almost imperceptible pan to a young man and woman sharing a moment on a large, weathered wooden swing suspended from the sturdy branches of an old tree. The man, with slightly tousled dark brown hair and a comfortable olive green Henley shirt, leans earnestly toward his companion, his voice low and sincere. The woman, her long honey-blond hair softly illuminated by the setting sun, wears a loose-fitting cream-colored knit cardigan over a simple floral dress; she listens intently, her head slightly tilted, a look of deep understanding on her face. They sit in a lush, verdant backyard in golden hour, the background melting into a beautiful bokeh of distant, out-of-focus foliage. Between them, on the swing's wooden planks, their hands are mere inches apart, his hand motionless and open, her fingers slightly curled, the tiny space between them crackling with unspoken emotion. As the camera continues its steady, deliberate close-in, the composition tightens, excluding the wider surroundings to focus purely on the vibrant space between their faces and hands, capturing the quiet intimacy and heartfelt seriousness of their conversation in an atmosphere of tranquility, warmth, and deep romance.", + "seed": 8256, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_6/index_8066_seed_8256_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.016919501451775433, + "vlm_logprob": -0.016919501451775433, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.016919501451775433, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008768152445554733 + ], + "opposite_logprob_tokens": [ + -5.883768081665039 + ], + "logprob_reward": -0.008768152445554733, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.008768152445554733, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15801380574703217, + 0.09440220147371292, + 0.04523462429642677, + 0.03509781137108803, + 0.06792354583740234, + 0.3503536283969879, + 0.039622437208890915, + 0.10767384618520737, + 0.04525594785809517, + 0.05642213672399521 + ], + "credit_map": [ + 0.018351882696151733, + 0.01677607372403145, + 0.017014607787132263, + 0.0301045048981905, + 0.017747392877936363, + 0.017333636060357094, + 0.01978246681392193, + 0.020903250202536583, + 0.007044042460620403, + 0.011065692640841007, + 0.01202456932514906, + 0.013973742723464966, + 0.01062097866088152, + 0.011470618657767773, + 0.019084850326180458, + 0.009117715992033482, + 0.0021618215832859278, + 0.005594182759523392, + 0.00575514230877161, + 0.007893542759120464, + 0.0039472999051213264, + 0.004210962448269129, + 0.009943039156496525, + 0.00572863407433033, + 0.002372587798163295, + 0.004531082231551409, + 0.0029230115469545126, + 0.007135177031159401, + 0.0033620980102568865, + 0.0033473405055701733, + 0.005729065742343664, + 0.0056974501349031925, + 0.006767211481928825, + 0.01316136121749878, + 0.007657886482775211, + 0.007665120996534824, + 0.008340229280292988, + 0.004060101695358753, + 0.00891427043825388, + 0.011357374489307404, + 0.03644581511616707, + 0.03169439360499382, + 0.14263516664505005, + 0.03110687993466854, + 0.009179776534438133, + 0.009915566071867943, + 0.016491541638970375, + 0.0728844702243805, + 0.005605186801403761, + 0.003510496113449335, + 0.005235634278506041, + 0.006142629310488701, + 0.0037885415367782116, + 0.0029255528934299946, + 0.004467165097594261, + 0.00794723629951477, + 0.013253011740744114, + 0.011856140568852425, + 0.018504060804843903, + 0.011837215162813663, + 0.009564824402332306, + 0.008951498195528984, + 0.015112738125026226, + 0.018594374880194664, + 0.004966805223375559, + 0.004526646342128515, + 0.00410196278244257, + 0.00518222339451313, + 0.01110776700079441, + 0.004653502255678177, + 0.0053660511039197445, + 0.005350995808839798, + 0.006185291334986687, + 0.007864031009376049, + 0.005178732797503471, + 0.008910564705729485, + 0.005864955484867096, + 0.007355234120041132, + 0.007729143835604191, + 0.0073341913521289825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.008768152445554733 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011991222854703665 + ], + "opposite_logprob_tokens": [ + -6.751199245452881 + ], + "logprob_reward": -0.0011991222854703665, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011991222854703665, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14424274861812592, + 0.12523595988750458, + 0.040448639541864395, + 0.035617195069789886, + 0.0541425459086895, + 0.10431573539972305, + 0.1891723871231079, + 0.12441319972276688, + 0.08119219541549683, + 0.10121942311525345 + ], + "credit_map": [ + 0.011057189665734768, + 0.016200309619307518, + 0.018136143684387207, + 0.023976953700184822, + 0.014512608759105206, + 0.017514478415250778, + 0.023410717025399208, + 0.01943434774875641, + 0.012897289358079433, + 0.016434159129858017, + 0.015595697797834873, + 0.02678452432155609, + 0.012736529111862183, + 0.008512187749147415, + 0.018165746703743935, + 0.014109823852777481, + 0.0027110606897622347, + 0.005046527367085218, + 0.004726026207208633, + 0.00579869607463479, + 0.004455892369151115, + 0.004927594214677811, + 0.0061117541044950485, + 0.006671086885035038, + 0.0017864333931356668, + 0.00543634220957756, + 0.0031493655405938625, + 0.004972604103386402, + 0.002974950009956956, + 0.0033465048763900995, + 0.005344402510672808, + 0.008606593124568462, + 0.004258549306541681, + 0.008072661235928535, + 0.003976792097091675, + 0.008047152310609818, + 0.012924612499773502, + 0.004635271616280079, + 0.006359828170388937, + 0.005867680534720421, + 0.009488537907600403, + 0.007001182995736599, + 0.02926732413470745, + 0.012916551902890205, + 0.003756763879209757, + 0.004917108453810215, + 0.006539809051901102, + 0.030428459867835045, + 0.015043153427541256, + 0.04379279166460037, + 0.010167565196752548, + 0.01725977100431919, + 0.01538949552923441, + 0.01054367795586586, + 0.017283078283071518, + 0.059692852199077606, + 0.01308103185147047, + 0.01846284046769142, + 0.016631372272968292, + 0.014585664495825768, + 0.016322452574968338, + 0.008115547709167004, + 0.014468222856521606, + 0.022746063768863678, + 0.01027593668550253, + 0.010369307361543179, + 0.004401572979986668, + 0.009539460763335228, + 0.019394664093852043, + 0.008900980465114117, + 0.00821467861533165, + 0.010095593519508839, + 0.009608159773051739, + 0.009905325248837471, + 0.009470205754041672, + 0.017943011596798897, + 0.011468496173620224, + 0.015935804694890976, + 0.013807772658765316, + 0.01308064442127943 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.0011991222854703665 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006739504635334015 + ], + "opposite_logprob_tokens": [ + -5.006739616394043 + ], + "logprob_reward": -0.006739504635334015, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006739504635334015, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11352252960205078, + 0.2257494479417801, + 0.05916300043463707, + 0.07509870082139969, + 0.08779144287109375, + 0.1500108391046524, + 0.10220429301261902, + 0.09622049331665039, + 0.042314883321523666, + 0.04792431369423866 + ], + "credit_map": [ + 0.011610206216573715, + 0.01311908196657896, + 0.018853411078453064, + 0.017239369451999664, + 0.010836356319487095, + 0.014299070462584496, + 0.014487744309008121, + 0.01307728048413992, + 0.022982317954301834, + 0.030502939596772194, + 0.02745169959962368, + 0.04801276698708534, + 0.02236083149909973, + 0.014576834626495838, + 0.03300762176513672, + 0.026854414492845535, + 0.004159039817750454, + 0.00768256513401866, + 0.006070369388908148, + 0.011848856694996357, + 0.006464471574872732, + 0.006988192908465862, + 0.008445778861641884, + 0.007503726053982973, + 0.004919497296214104, + 0.009686063043773174, + 0.007134179584681988, + 0.01446416787803173, + 0.008525214157998562, + 0.005568025633692741, + 0.011830604635179043, + 0.012970954179763794, + 0.008940468542277813, + 0.012994673103094101, + 0.00934920459985733, + 0.013785219751298428, + 0.017036613076925278, + 0.005459583830088377, + 0.009631949476897717, + 0.010593732818961143, + 0.01682974211871624, + 0.013280690647661686, + 0.05509228631854057, + 0.011133135296404362, + 0.008881107904016972, + 0.007152197416871786, + 0.0068168239668011665, + 0.030824854969978333, + 0.009761848486959934, + 0.010826442390680313, + 0.008860562928020954, + 0.020902233198285103, + 0.004886470269411802, + 0.006898491643369198, + 0.010446316562592983, + 0.029621928930282593, + 0.012436456978321075, + 0.01197186391800642, + 0.01218400802463293, + 0.013768093660473824, + 0.012664489448070526, + 0.007381519302725792, + 0.012504881247878075, + 0.0133091751486063, + 0.005870508030056953, + 0.005515914410352707, + 0.0030078263953328133, + 0.00502055836841464, + 0.010119743645191193, + 0.00421674782410264, + 0.003822121536359191, + 0.004741464275866747, + 0.0048537300899624825, + 0.00537117337808013, + 0.0041604433208703995, + 0.007865697145462036, + 0.005122920498251915, + 0.008046085014939308, + 0.0059314086101949215, + 0.0065728528425097466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.006739504635334015 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029803412035107613 + ], + "opposite_logprob_tokens": [ + -3.529803514480591 + ], + "logprob_reward": -0.029803412035107613, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029803412035107613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1864020675420761, + 0.13007047772407532, + 0.029374293982982635, + 0.05120304226875305, + 0.1147676631808281, + 0.07584138959646225, + 0.12281132489442825, + 0.12260516732931137, + 0.0893792137503624, + 0.07754536718130112 + ], + "credit_map": [ + 0.016089944168925285, + 0.02206653729081154, + 0.022425951436161995, + 0.0374571792781353, + 0.017444835975766182, + 0.020495306700468063, + 0.024205412715673447, + 0.026216892525553703, + 0.014615579508244991, + 0.017881207168102264, + 0.014997916296124458, + 0.02806992642581463, + 0.013360430486500263, + 0.008044583722949028, + 0.0195778738707304, + 0.013522958382964134, + 0.0019376884447410703, + 0.003617200767621398, + 0.004244745243340731, + 0.005264353938400745, + 0.003736908081918955, + 0.003457248443737626, + 0.003647063160315156, + 0.003469085320830345, + 0.0024685917887836695, + 0.009121017530560493, + 0.004737980663776398, + 0.004528001416474581, + 0.004447948653250933, + 0.004458106122910976, + 0.007134167943149805, + 0.01430722326040268, + 0.009452512487769127, + 0.013149168342351913, + 0.00864094216376543, + 0.01683366484940052, + 0.02607751451432705, + 0.012737371027469635, + 0.014153710566461086, + 0.013722777366638184, + 0.007636259775608778, + 0.007965002208948135, + 0.011655829846858978, + 0.009038908407092094, + 0.006086882669478655, + 0.0075753675773739815, + 0.009240331128239632, + 0.01664280891418457, + 0.011665002442896366, + 0.02468884550035, + 0.01051511149853468, + 0.013254250399768353, + 0.00705417525023222, + 0.00798415020108223, + 0.012914018705487251, + 0.03473576903343201, + 0.018820013850927353, + 0.016809485852718353, + 0.01526215486228466, + 0.01896335743367672, + 0.010301959700882435, + 0.007259792648255825, + 0.015929371118545532, + 0.019259028136730194, + 0.012892634607851505, + 0.01541096344590187, + 0.006937777623534203, + 0.011770465411245823, + 0.015709957107901573, + 0.007744866888970137, + 0.0074777682311832905, + 0.011434776708483696, + 0.007934663444757462, + 0.010600507259368896, + 0.0062255109660327435, + 0.010586736723780632, + 0.00906744971871376, + 0.012441749684512615, + 0.010134226642549038, + 0.010554528795182705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.029803412035107613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a lush, verdant backyard during golden hour?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are a young man with dark brown hair and an olive green Henley shirt and a young woman with honey-blond hair and a cream-colored knit cardigan visible?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are the man and woman sitting on a large, weathered wooden swing suspended from an old tree?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the man's hand and the woman's hand positioned mere inches apart on the swing's wooden planks?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera begin in a medium shot at eye level and slowly pan/zoom in to a close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808731585741043 + ], + "opposite_logprob_tokens": [ + -3.2880873680114746 + ], + "logprob_reward": -0.03808731585741043, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808731585741043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07464572787284851, + 0.0722387358546257, + 0.03389592841267586, + 0.03700510784983635, + 0.09724785387516022, + 0.30544382333755493, + 0.08723832666873932, + 0.09955710172653198, + 0.14406760036945343, + 0.048659782856702805 + ], + "credit_map": [ + 0.006223446223884821, + 0.00904489029198885, + 0.007550277281552553, + 0.016740169376134872, + 0.008885531686246395, + 0.005804507527500391, + 0.011185687966644764, + 0.009211212396621704, + 0.004150348715484142, + 0.01064169593155384, + 0.00711483508348465, + 0.014153470285236835, + 0.007877898402512074, + 0.008437195792794228, + 0.011214302852749825, + 0.008648980408906937, + 0.0036464582663029432, + 0.004749536979943514, + 0.002792652929201722, + 0.0052742911502718925, + 0.004264817573130131, + 0.003064410062506795, + 0.003788175992667675, + 0.0063155824318528175, + 0.0028663061093539, + 0.007116641383618116, + 0.0051213339902460575, + 0.0063470397144556046, + 0.003551571164280176, + 0.0029690288938581944, + 0.003693864680826664, + 0.0053393226116895676, + 0.009270282462239265, + 0.018259314820170403, + 0.011203251779079437, + 0.013195144012570381, + 0.014741304330527782, + 0.009280126541852951, + 0.009006869979202747, + 0.012291565537452698, + 0.031930942088365555, + 0.02198898233473301, + 0.11722628027200699, + 0.017281796783208847, + 0.007398948073387146, + 0.009761057794094086, + 0.015033667907118797, + 0.0848221480846405, + 0.00822577252984047, + 0.01937134750187397, + 0.007206510286778212, + 0.019397584721446037, + 0.008360017091035843, + 0.003908208571374416, + 0.011256981641054153, + 0.009511907584965229, + 0.014261594042181969, + 0.013052030466496944, + 0.02236497960984707, + 0.015850523486733437, + 0.007637998089194298, + 0.00628034770488739, + 0.01012509223073721, + 0.009984537027776241, + 0.02050749771296978, + 0.012995353899896145, + 0.0073099285364151, + 0.012454374693334103, + 0.04211829602718353, + 0.015213845297694206, + 0.015988536179065704, + 0.017479779198765755, + 0.004766613710671663, + 0.005604038015007973, + 0.003825252875685692, + 0.007937089540064335, + 0.00738532142713666, + 0.007674365770071745, + 0.005377774126827717, + 0.006089329719543457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.03808731585741043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_tensor([1739, 1739])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_tensor([1739, 1739])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..03b3615a778f70186f7a76ff4716565579194deb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_tensor([1739, 1739])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, wide shot opens the scene, the camera violently panning backward to reveal the cataclysmic events unfolding in a surreal, Jodorowsky-inspired vision. The Empress is violently ejected from a vast, geometrically perfect crystal throne, ablaze with an icy inner light. Her body, adorned with a dramatic, jeweled velvet robe of crimson and royal purple, tumbles backward in a metaphysical void. Her alabaster face, a mask of shocked terror, her Kohl-rimmed eyes open, the Empress is caught in a chaotic, slow-motion ballet, her ornate golden headdress cast aside, her limbs flailing. The heavy, jeweled fabric of her robe billows around her, catching and shattering the light emanating from the throne and the surrounding cosmic dreamscape—a swirling indigo cloud and fleeting, glowing alchemical symbols. The entire scene flickers erratically with strobe-like intensity and digital glitches, as if the very fabric of reality is being short-circuited by the sheer, unseen force of impact, creating a deeply unsettling and visually saturated atmosphere of sacred, psychedelic violence.", + "seed": 24511, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24511_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09766023885458708, + "vlm_logprob": -0.09766023885458708, + "vlm_yes_no_margin": 3.3, + "vlm_token_credit": -0.09766023885458708, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3998366892337799 + ], + "opposite_logprob_tokens": [ + -1.1498366594314575 + ], + "logprob_reward": -0.3998366892337799, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3998366892337799, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3675764501094818, + 0.14010831713676453, + 0.05528751760721207, + 0.061584338545799255, + 0.11070386320352554, + 0.02804272063076496, + 0.03652750328183174, + 0.06001749634742737, + 0.04900135099887848, + 0.09115050733089447 + ], + "credit_map": [ + 0.05327010154724121, + 0.0447469986975193, + 0.08441856503486633, + 0.03331872820854187, + 0.040920011699199677, + 0.0414346307516098, + 0.038715559989213943, + 0.03075183928012848, + 0.014145447872579098, + 0.019380349665880203, + 0.02104746550321579, + 0.01727081649005413, + 0.018510272726416588, + 0.013502265326678753, + 0.023118704557418823, + 0.013132994063198566, + 0.008005587384104729, + 0.007124095223844051, + 0.005550907459110022, + 0.006772405467927456, + 0.007902860641479492, + 0.0069337706081569195, + 0.007695646025240421, + 0.00530224060639739, + 0.0057108355686068535, + 0.007749109994620085, + 0.005534577649086714, + 0.007603566162288189, + 0.009777297265827656, + 0.006871491204947233, + 0.010435330681502819, + 0.007902124896645546, + 0.007665758021175861, + 0.008948951959609985, + 0.013293798081576824, + 0.014094055630266666, + 0.012933304533362389, + 0.02084190584719181, + 0.016551462933421135, + 0.016374627128243446, + 0.002789567457512021, + 0.002627183450385928, + 0.0031965370289981365, + 0.004424066282808781, + 0.004513672087341547, + 0.003738600993528962, + 0.002997027477249503, + 0.003756066784262657, + 0.0031575821340084076, + 0.0034531787969172, + 0.0036476803943514824, + 0.004694859031587839, + 0.006529737263917923, + 0.007218914106488228, + 0.0030132385436445475, + 0.00481231277808547, + 0.014703656546771526, + 0.00657792529091239, + 0.004625752568244934, + 0.005243923049420118, + 0.009302493184804916, + 0.008493544533848763, + 0.006507873069494963, + 0.00456232437863946, + 0.00663822190836072, + 0.004205200355499983, + 0.00492647010833025, + 0.004280013032257557, + 0.01071393582969904, + 0.007993442937731743, + 0.004718788433820009, + 0.005525278393179178, + 0.013767522759735584, + 0.008802417665719986, + 0.01304208766669035, + 0.006379955448210239, + 0.026048827916383743, + 0.008040506392717361, + 0.007364680524915457, + 0.007704505696892738 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.3998366892337799 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014225857332348824 + ], + "opposite_logprob_tokens": [ + -4.264225959777832 + ], + "logprob_reward": -0.014225857332348824, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014225857332348824, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14321082830429077, + 0.0946701392531395, + 0.08848713338375092, + 0.11810962110757828, + 0.08302821218967438, + 0.04847151041030884, + 0.1013643741607666, + 0.188175231218338, + 0.07542777806520462, + 0.059055145829916 + ], + "credit_map": [ + 0.020354507490992546, + 0.019030539318919182, + 0.02545863389968872, + 0.014809872955083847, + 0.016438305377960205, + 0.014723148196935654, + 0.017945323139429092, + 0.014450494199991226, + 0.011767621152102947, + 0.014020849950611591, + 0.012573204934597015, + 0.007996316067874432, + 0.014970481395721436, + 0.009488750249147415, + 0.013972537592053413, + 0.009880376979708672, + 0.010262901894748211, + 0.009175894781947136, + 0.010277800261974335, + 0.016808247193694115, + 0.01222673524171114, + 0.008363216184079647, + 0.013354931026697159, + 0.00801740400493145, + 0.00922003947198391, + 0.010171054862439632, + 0.014304913580417633, + 0.011991779319941998, + 0.020668592303991318, + 0.01344339456409216, + 0.022532185539603233, + 0.015777667984366417, + 0.00682412926107645, + 0.011702042073011398, + 0.011361801065504551, + 0.01045257318764925, + 0.010299183428287506, + 0.01271593663841486, + 0.011693463660776615, + 0.007979084737598896, + 0.006071070209145546, + 0.0035756619181483984, + 0.005090357270091772, + 0.00545567786321044, + 0.005752197001129389, + 0.0066378479823470116, + 0.007375252433121204, + 0.008513445034623146, + 0.01032569631934166, + 0.009079545736312866, + 0.011833369731903076, + 0.014916915446519852, + 0.01820727251470089, + 0.014274203218519688, + 0.010926656424999237, + 0.011800719425082207, + 0.02507198601961136, + 0.023497357964515686, + 0.02680589072406292, + 0.017692407593131065, + 0.03586061671376228, + 0.021275578066706657, + 0.022020025178790092, + 0.01595136895775795, + 0.007184159941971302, + 0.011286532506346703, + 0.010141123086214066, + 0.01309703104197979, + 0.007555969059467316, + 0.011354297399520874, + 0.0064450157806277275, + 0.008363653905689716, + 0.007027231622487307, + 0.00775145972147584, + 0.010115614160895348, + 0.00903900247067213, + 0.008492781780660152, + 0.0038841108325868845, + 0.006306909956037998, + 0.006438035052269697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 3 + ], + "selected_reward": -0.014225857332348824 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810407221317291 + ], + "opposite_logprob_tokens": [ + -3.2881040573120117 + ], + "logprob_reward": -0.03810407221317291, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810407221317291, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30049726366996765, + 0.18323469161987305, + 0.03394855186343193, + 0.09274570643901825, + 0.11810684204101562, + 0.0504351370036602, + 0.036845237016677856, + 0.07341771572828293, + 0.03960910812020302, + 0.071159727871418 + ], + "credit_map": [ + 0.04405540972948074, + 0.034045662730932236, + 0.06719458848237991, + 0.028194064274430275, + 0.03103899024426937, + 0.03483368456363678, + 0.03406452015042305, + 0.02707032300531864, + 0.01960836537182331, + 0.027093812823295593, + 0.025387944653630257, + 0.01721908338367939, + 0.03307291865348816, + 0.016264909878373146, + 0.02395346388220787, + 0.020634207874536514, + 0.0037474215496331453, + 0.004331968259066343, + 0.0027915749233216047, + 0.0047018881887197495, + 0.004241730086505413, + 0.004861800465732813, + 0.0040204268880188465, + 0.00525174243375659, + 0.005897053983062506, + 0.008497467264533043, + 0.007852386683225632, + 0.00973542034626007, + 0.018571199849247932, + 0.00877110194414854, + 0.023879170417785645, + 0.009541901759803295, + 0.008188996464014053, + 0.00924626737833023, + 0.014228545129299164, + 0.01604718528687954, + 0.014412730932235718, + 0.02135070413351059, + 0.019096791744232178, + 0.015535624697804451, + 0.0057486481964588165, + 0.004277902189642191, + 0.0050900960341095924, + 0.005461088381707668, + 0.006841057445853949, + 0.00722131272777915, + 0.006247400771826506, + 0.009547632187604904, + 0.004525090102106333, + 0.0026279541198164225, + 0.004317839629948139, + 0.00459282984957099, + 0.0059958565980196, + 0.006471623666584492, + 0.0039437864907085896, + 0.004370256792753935, + 0.015537048690021038, + 0.007806207984685898, + 0.006506738252937794, + 0.0072725070640444756, + 0.01235476229339838, + 0.010201329365372658, + 0.007536427583545446, + 0.006202691234648228, + 0.004716414026916027, + 0.0031680832616984844, + 0.004193748813122511, + 0.004210366867482662, + 0.009582558646798134, + 0.005910571664571762, + 0.0041913315653800964, + 0.0036360325757414103, + 0.00833199918270111, + 0.007592152804136276, + 0.008707177825272083, + 0.00724024185910821, + 0.012517075054347515, + 0.007043079938739538, + 0.009069517254829407, + 0.01065848022699356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.03810407221317291 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009781762026250362 + ], + "opposite_logprob_tokens": [ + -4.634781837463379 + ], + "logprob_reward": -0.009781762026250362, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009781762026250362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15538011491298676, + 0.14186272025108337, + 0.15782926976680756, + 0.0880054160952568, + 0.06872516870498657, + 0.02109929546713829, + 0.0743306577205658, + 0.1238706186413765, + 0.10612209886312485, + 0.06277462840080261 + ], + "credit_map": [ + 0.01599595509469509, + 0.022204482927918434, + 0.01827869936823845, + 0.02574240230023861, + 0.01963784731924534, + 0.009728162549436092, + 0.029995644465088844, + 0.013796908780932426, + 0.01458695437759161, + 0.02057758718729019, + 0.02049187570810318, + 0.012868587858974934, + 0.025624364614486694, + 0.013395211659371853, + 0.01805873028934002, + 0.01625942252576351, + 0.018676986917853355, + 0.019831683486700058, + 0.016353534534573555, + 0.010742521844804287, + 0.027830613777041435, + 0.026649925857782364, + 0.013756992295384407, + 0.02398701012134552, + 0.007999209687113762, + 0.011541874147951603, + 0.01017533615231514, + 0.010417062789201736, + 0.012658984400331974, + 0.0075066713616251945, + 0.01903342828154564, + 0.008672845549881458, + 0.00416598841547966, + 0.005839290097355843, + 0.005715701729059219, + 0.00925005879253149, + 0.011365571059286594, + 0.012345383875072002, + 0.013793340884149075, + 0.006249837577342987, + 0.0021281049121171236, + 0.0019113050075247884, + 0.004259241279214621, + 0.003072371706366539, + 0.00275596184656024, + 0.0024536459241062403, + 0.002252324251458049, + 0.0022663401905447245, + 0.008340202271938324, + 0.005341279786080122, + 0.008314267732203007, + 0.009929032064974308, + 0.008933989331126213, + 0.01680736429989338, + 0.004636673256754875, + 0.01202785037457943, + 0.025743087753653526, + 0.011346166022121906, + 0.011129207909107208, + 0.012023135088384151, + 0.021101243793964386, + 0.015879500657320023, + 0.014927666634321213, + 0.011720611713826656, + 0.011612940579652786, + 0.00792781077325344, + 0.00968901440501213, + 0.00899694487452507, + 0.023591823875904083, + 0.017327027395367622, + 0.00856668222695589, + 0.018409861251711845, + 0.008930978365242481, + 0.006399338133633137, + 0.008776772767305374, + 0.0054464382119476795, + 0.011504576541483402, + 0.005547224078327417, + 0.006931097246706486, + 0.009238208644092083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.009781762026250362 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026352813467383385 + ], + "opposite_logprob_tokens": [ + -3.651352882385254 + ], + "logprob_reward": -0.026352813467383385, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026352813467383385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1286955624818802, + 0.2873399555683136, + 0.04608374089002609, + 0.09007514268159866, + 0.14149612188339233, + 0.022657593712210655, + 0.030240235850214958, + 0.07402624934911728, + 0.06303776055574417, + 0.11634773015975952 + ], + "credit_map": [ + 0.020018937066197395, + 0.01400435995310545, + 0.019071144983172417, + 0.01670149900019169, + 0.017706213518977165, + 0.013289454393088818, + 0.01590909995138645, + 0.011994840577244759, + 0.033523935824632645, + 0.041847843676805496, + 0.0375429205596447, + 0.035229139029979706, + 0.03395974636077881, + 0.025774367153644562, + 0.051301512867212296, + 0.02816043607890606, + 0.0077966973185539246, + 0.006687301676720381, + 0.00437620235607028, + 0.003768408438190818, + 0.007453571073710918, + 0.004526377189904451, + 0.004530795384198427, + 0.00694437837228179, + 0.008871524594724178, + 0.010204941965639591, + 0.013071756809949875, + 0.011720240116119385, + 0.013736473396420479, + 0.009831933304667473, + 0.012615885585546494, + 0.010022375732660294, + 0.009631501510739326, + 0.01145488116890192, + 0.01607627607882023, + 0.01798100396990776, + 0.016159672290086746, + 0.026975682005286217, + 0.02273484133183956, + 0.020482223480939865, + 0.0027032040525227785, + 0.0020215711556375027, + 0.0032162938732653856, + 0.0032526953145861626, + 0.0033166047651320696, + 0.0025279601104557514, + 0.0024768118746578693, + 0.0031424500048160553, + 0.0036318895872682333, + 0.0032104721758514643, + 0.0034497701562941074, + 0.003029487794265151, + 0.006752864923328161, + 0.004378254525363445, + 0.002471636515110731, + 0.003315855050459504, + 0.014123323373496532, + 0.006965328939259052, + 0.006907844450324774, + 0.00833433773368597, + 0.013697811402380466, + 0.009503161534667015, + 0.007937943562865257, + 0.006556489039212465, + 0.005853326991200447, + 0.005219893995672464, + 0.004688277840614319, + 0.005222178064286709, + 0.011400080285966396, + 0.010175702162086964, + 0.005984547082334757, + 0.01449374109506607, + 0.017282575368881226, + 0.011283203959465027, + 0.016070811077952385, + 0.007980098016560078, + 0.030554762110114098, + 0.010752275586128235, + 0.009502663277089596, + 0.012921323999762535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -0.026352813467383385 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot opens the scene, the camera violently panning backward to reveal the cataclysmic events unfolding in a surreal, Jodorowsky-inspired vision. The Empress is violently ejected from a vast, geometrically perfect crystal throne, ablaze with an icy inner light. Her body, adorned with a dramatic, jeweled velvet robe of crimson and royal purple, tumbles backward in a metaphysical void. Her alabaster face, a mask of shocked terror, her Kohl-rimmed eyes open, the Empress is caught in a chaotic, slow-motion ballet, her ornate golden headdress cast aside, her limbs flailing. The heavy, jeweled fabric of her robe billows around her, catching and shattering the light emanating from the throne and the surrounding cosmic dreamscape—a swirling indigo cloud and fleeting, glowing alchemical symbols. The entire scene flickers erratically with strobe-like intensity and digital glitches, as if the very fabric of reality is being short-circuited by the sheer, unseen force of impact, creating a deeply unsettling and visually saturated atmosphere of sacred, psychedelic violence.", + "seed": 24512, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24512_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2811280240304768, + "vlm_logprob": -0.2811280240304768, + "vlm_yes_no_margin": 2.625, + "vlm_token_credit": -0.2811280240304768, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8320468068122864 + ], + "opposite_logprob_tokens": [ + -0.5820468068122864 + ], + "logprob_reward": -0.8320468068122864, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8320468068122864, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.3991739749908447, + 0.11350496858358383, + 0.04472971707582474, + 0.04051600769162178, + 0.0536607950925827, + 0.07313328236341476, + 0.09168241173028946, + 0.03336198627948761, + 0.02489287592470646, + 0.12534399330615997 + ], + "credit_map": [ + 0.04015638306736946, + 0.017546674236655235, + 0.12506964802742004, + 0.04567883536219597, + 0.015612085349857807, + 0.07682143896818161, + 0.0284164659678936, + 0.04987242445349693, + 0.013770151883363724, + 0.016171379014849663, + 0.01800766959786415, + 0.01496746577322483, + 0.01291379053145647, + 0.00853013526648283, + 0.013788283802568913, + 0.015356088988482952, + 0.006006237585097551, + 0.00465836375951767, + 0.0069447411224246025, + 0.003938379697501659, + 0.00634283572435379, + 0.003657395951449871, + 0.005191190633922815, + 0.007990571670234203, + 0.005471659358590841, + 0.003905808087438345, + 0.005846963729709387, + 0.003977518994361162, + 0.00452765729278326, + 0.004833871964365244, + 0.0045395344495773315, + 0.007412994746118784, + 0.0068032569251954556, + 0.007145598065108061, + 0.011071626096963882, + 0.005492073018103838, + 0.006019826978445053, + 0.006602503824979067, + 0.004738153889775276, + 0.005787755828350782, + 0.016023728996515274, + 0.007753991521894932, + 0.02025189995765686, + 0.006361198611557484, + 0.007277359254658222, + 0.006892398465424776, + 0.004398860968649387, + 0.004173841793090105, + 0.012468489818274975, + 0.01132652536034584, + 0.008465487509965897, + 0.011465105228126049, + 0.013813287019729614, + 0.016858771443367004, + 0.009932140819728374, + 0.007352603133767843, + 0.003498632926493883, + 0.0025755467358976603, + 0.0034744825679808855, + 0.00386000145226717, + 0.004868673626333475, + 0.006088898051530123, + 0.005245563108474016, + 0.0037501889746636152, + 0.0025395245756953955, + 0.002213768195360899, + 0.004173578694462776, + 0.0032715501729398966, + 0.004522199276834726, + 0.0032389662228524685, + 0.0031674234196543694, + 0.0017658657161518931, + 0.020317796617746353, + 0.009439385496079922, + 0.01039887685328722, + 0.013572393916547298, + 0.04123587906360626, + 0.010378941893577576, + 0.00983965303748846, + 0.010161071084439754 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.8320468068122864 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006764014717191458 + ], + "opposite_logprob_tokens": [ + -5.006763935089111 + ], + "logprob_reward": -0.006764014717191458, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006764014717191458, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16016711294651031, + 0.09690914303064346, + 0.1072825938463211, + 0.08900316804647446, + 0.09814655035734177, + 0.1219405084848404, + 0.04280475154519081, + 0.10718192905187607, + 0.053707305341959, + 0.12285691499710083 + ], + "credit_map": [ + 0.01641157828271389, + 0.010268719866871834, + 0.015719836577773094, + 0.020685603842139244, + 0.010618572123348713, + 0.03657296299934387, + 0.014867275021970272, + 0.0350225605070591, + 0.011347942054271698, + 0.010860279202461243, + 0.019140902906656265, + 0.008931905031204224, + 0.010454080067574978, + 0.011756930500268936, + 0.010106245055794716, + 0.014310858212411404, + 0.01476572547107935, + 0.014826441183686256, + 0.013259072788059711, + 0.008800644427537918, + 0.020825214684009552, + 0.008204909041523933, + 0.009455793537199497, + 0.017144791781902313, + 0.02106519229710102, + 0.00612808670848608, + 0.015520279295742512, + 0.006607391405850649, + 0.007783986162394285, + 0.007271939888596535, + 0.006972422357648611, + 0.017653873190283775, + 0.009769825264811516, + 0.014325459487736225, + 0.01055030059069395, + 0.011339882388710976, + 0.013878688216209412, + 0.015996575355529785, + 0.013638883829116821, + 0.00864693894982338, + 0.017781376838684082, + 0.016989538446068764, + 0.021412761881947517, + 0.011731300503015518, + 0.01333562284708023, + 0.01703648641705513, + 0.010457835160195827, + 0.013195586390793324, + 0.006289995275437832, + 0.0051564620807766914, + 0.004335537552833557, + 0.004963361658155918, + 0.006272774655371904, + 0.006776893045753241, + 0.004746316932141781, + 0.004263410810381174, + 0.011315646581351757, + 0.008274609223008156, + 0.01832306943833828, + 0.01140639279037714, + 0.01291002705693245, + 0.025504499673843384, + 0.009178397245705128, + 0.010269290767610073, + 0.00549617363139987, + 0.003944262396544218, + 0.0075177694670856, + 0.006763744633644819, + 0.008840242400765419, + 0.005099578760564327, + 0.010221203789114952, + 0.0058243307285010815, + 0.01178450882434845, + 0.00994347408413887, + 0.013614241033792496, + 0.02426307462155819, + 0.027125537395477295, + 0.010277255438268185, + 0.012051655910909176, + 0.013797171413898468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.006764014717191458 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523337721824646 + ], + "opposite_logprob_tokens": [ + -0.898337721824646 + ], + "logprob_reward": -0.523337721824646, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523337721824646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10310007631778717, + 0.08901448547840118, + 0.031812090426683426, + 0.02215452864766121, + 0.17826054990291595, + 0.20590931177139282, + 0.06780935823917389, + 0.04192601516842842, + 0.03241714462637901, + 0.2275964319705963 + ], + "credit_map": [ + 0.007514163386076689, + 0.0073797451332211494, + 0.00966758280992508, + 0.01067335158586502, + 0.005851104389876127, + 0.028435977175831795, + 0.008460408076643944, + 0.025117741897702217, + 0.009836142882704735, + 0.011131029576063156, + 0.011282399296760559, + 0.012848575599491596, + 0.010194485075771809, + 0.007092463783919811, + 0.012748051434755325, + 0.013881338760256767, + 0.0030834374483674765, + 0.004516429267823696, + 0.0030511191580444574, + 0.00293792225420475, + 0.006009192671626806, + 0.002763070398941636, + 0.005631588399410248, + 0.0038193310610949993, + 0.003997156862169504, + 0.0020646543707698584, + 0.003558791009709239, + 0.0024935349356383085, + 0.0021681387443095446, + 0.0021140764001756907, + 0.0022734387312084436, + 0.0034847380593419075, + 0.024978728964924812, + 0.017532264813780785, + 0.04277778044342995, + 0.015681035816669464, + 0.017350075766444206, + 0.02534399926662445, + 0.012480659410357475, + 0.022116001695394516, + 0.04680950939655304, + 0.02197631262242794, + 0.056206539273262024, + 0.014933738857507706, + 0.019920893013477325, + 0.021053345873951912, + 0.013551107607781887, + 0.011457856744527817, + 0.007626086473464966, + 0.009883364662528038, + 0.005737883970141411, + 0.008303612470626831, + 0.012338219210505486, + 0.008019606582820415, + 0.009671822190284729, + 0.006228758487850428, + 0.003867135848850012, + 0.0030970002990216017, + 0.006917534861713648, + 0.004223697353154421, + 0.006344051565974951, + 0.008995089679956436, + 0.004638691432774067, + 0.003842814126983285, + 0.0036560767330229282, + 0.003264759434387088, + 0.0031641884706914425, + 0.004262118134647608, + 0.004969136323779821, + 0.0039068059995770454, + 0.005965943913906813, + 0.0032281172461807728, + 0.04242807254195213, + 0.019838843494653702, + 0.009792277589440346, + 0.009536927565932274, + 0.09309535473585129, + 0.021685350686311722, + 0.019005443900823593, + 0.01221416424959898 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 4 + ], + "selected_reward": -0.523337721824646 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009784949012100697 + ], + "opposite_logprob_tokens": [ + -4.634785175323486 + ], + "logprob_reward": -0.009784949012100697, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009784949012100697, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16954664885997772, + 0.08084191381931305, + 0.07411251217126846, + 0.07142046838998795, + 0.10229046642780304, + 0.026041196659207344, + 0.029483119025826454, + 0.1119007095694542, + 0.06813053041696548, + 0.266232430934906 + ], + "credit_map": [ + 0.019209539517760277, + 0.009710034355521202, + 0.032179221510887146, + 0.020942939445376396, + 0.010669196955859661, + 0.02299259789288044, + 0.014017194509506226, + 0.0398259311914444, + 0.00954130757600069, + 0.006136453244835138, + 0.01582288183271885, + 0.006144647020846605, + 0.005816206801682711, + 0.011012006551027298, + 0.007385574281215668, + 0.018982838839292526, + 0.009984122589230537, + 0.007393415551632643, + 0.011227049864828587, + 0.0056474353186786175, + 0.007753557059913874, + 0.009506525471806526, + 0.008753838948905468, + 0.013846557587385178, + 0.01226697489619255, + 0.004952327813953161, + 0.010795043781399727, + 0.006244131363928318, + 0.00702936016023159, + 0.006884449161589146, + 0.008431436493992805, + 0.014816739596426487, + 0.015691163018345833, + 0.013115378096699715, + 0.018471719697117805, + 0.008166025392711163, + 0.009289694018661976, + 0.023769507184624672, + 0.007158051710575819, + 0.006628927309066057, + 0.0054501499980688095, + 0.0029105583671480417, + 0.0035144377034157515, + 0.0021534657571464777, + 0.002637696685269475, + 0.003990244586020708, + 0.002175348112359643, + 0.003209293121472001, + 0.003618431743234396, + 0.0035711629316210747, + 0.0033080235589295626, + 0.0037826166953891516, + 0.003601549891754985, + 0.0058542657643556595, + 0.002838622545823455, + 0.002908445429056883, + 0.011496571823954582, + 0.008330709300935268, + 0.01985878311097622, + 0.010679904371500015, + 0.010811452753841877, + 0.03256980702280998, + 0.008709078654646873, + 0.009444402530789375, + 0.008073830977082253, + 0.006520743481814861, + 0.007465039845556021, + 0.0068996683694422245, + 0.01042214222252369, + 0.014414837583899498, + 0.006695031654089689, + 0.007639233022928238, + 0.044499270617961884, + 0.021697040647268295, + 0.013902882114052773, + 0.021731706336140633, + 0.10203428566455841, + 0.02387363836169243, + 0.02179405651986599, + 0.016699528321623802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.009784949012100697 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033706627786159515 + ], + "opposite_logprob_tokens": [ + -3.4087066650390625 + ], + "logprob_reward": -0.033706627786159515, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033706627786159515, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41958072781562805, + 0.15790440142154694, + 0.04324987903237343, + 0.05268530547618866, + 0.03671367093920708, + 0.06393047422170639, + 0.0448116809129715, + 0.03709881380200386, + 0.029522882774472237, + 0.11450216919183731 + ], + "credit_map": [ + 0.04891061410307884, + 0.01610344462096691, + 0.14214132726192474, + 0.043119676411151886, + 0.015200885012745857, + 0.07981589436531067, + 0.028339875862002373, + 0.045949000865221024, + 0.019626330584287643, + 0.015687137842178345, + 0.02378993108868599, + 0.015651049092411995, + 0.012852231040596962, + 0.01761814020574093, + 0.015784043818712234, + 0.03689553961157799, + 0.006240184418857098, + 0.0034928712993860245, + 0.005627748556435108, + 0.00422971136868, + 0.003355416702106595, + 0.004240542650222778, + 0.0055632274597883224, + 0.010500175878405571, + 0.011824461631476879, + 0.0038274163380265236, + 0.007743809837847948, + 0.004449837375432253, + 0.0037102261558175087, + 0.004396084696054459, + 0.004695897456258535, + 0.012037570588290691, + 0.006448699627071619, + 0.00454364251345396, + 0.00398766016587615, + 0.0027126693166792393, + 0.003916196059435606, + 0.0092463418841362, + 0.0029245263431221247, + 0.0029339345637708902, + 0.007420391775667667, + 0.010539405979216099, + 0.005856383126229048, + 0.008019705303013325, + 0.011309622786939144, + 0.010149751789867878, + 0.006392871029675007, + 0.004242344293743372, + 0.007087004370987415, + 0.0054434072226285934, + 0.004397110547870398, + 0.004472095984965563, + 0.006060468032956123, + 0.008388849906623363, + 0.0048228902742266655, + 0.0041398536413908005, + 0.003586192848160863, + 0.0029832073487341404, + 0.005147069226950407, + 0.004539002664387226, + 0.004928248934447765, + 0.007361987140029669, + 0.004128957632929087, + 0.0044241491705179214, + 0.0037338342517614365, + 0.0028508531395345926, + 0.002491068560630083, + 0.002594286808744073, + 0.004371088929474354, + 0.006178747396916151, + 0.003145852591842413, + 0.004157152026891708, + 0.01139168068766594, + 0.009334026835858822, + 0.01176817249506712, + 0.02084856666624546, + 0.030011745169758797, + 0.009828646667301655, + 0.009154553525149822, + 0.012164771556854248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.033706627786159515 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot opens the scene, the camera violently panning backward to reveal the cataclysmic events unfolding in a surreal, Jodorowsky-inspired vision. The Empress is violently ejected from a vast, geometrically perfect crystal throne, ablaze with an icy inner light. Her body, adorned with a dramatic, jeweled velvet robe of crimson and royal purple, tumbles backward in a metaphysical void. Her alabaster face, a mask of shocked terror, her Kohl-rimmed eyes open, the Empress is caught in a chaotic, slow-motion ballet, her ornate golden headdress cast aside, her limbs flailing. The heavy, jeweled fabric of her robe billows around her, catching and shattering the light emanating from the throne and the surrounding cosmic dreamscape—a swirling indigo cloud and fleeting, glowing alchemical symbols. The entire scene flickers erratically with strobe-like intensity and digital glitches, as if the very fabric of reality is being short-circuited by the sheer, unseen force of impact, creating a deeply unsettling and visually saturated atmosphere of sacred, psychedelic violence.", + "seed": 24513, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24513_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10531492792069912, + "vlm_logprob": -0.10531492792069912, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.10531492792069912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3564462959766388 + ], + "opposite_logprob_tokens": [ + -1.2314462661743164 + ], + "logprob_reward": -0.3564462959766388, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3564462959766388, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17872853577136993, + 0.25211790204048157, + 0.03969257324934006, + 0.053972773253917694, + 0.06247871369123459, + 0.12040461599826813, + 0.062466904520988464, + 0.070148766040802, + 0.14173592627048492, + 0.018253283575177193 + ], + "credit_map": [ + 0.018565570935606956, + 0.02360718511044979, + 0.025473523885011673, + 0.025970200076699257, + 0.019744308665394783, + 0.020206240937113762, + 0.025437429547309875, + 0.01972406730055809, + 0.056312549859285355, + 0.016492730006575584, + 0.030452242121100426, + 0.012983361259102821, + 0.009139342233538628, + 0.027570627629756927, + 0.014839877374470234, + 0.08432717621326447, + 0.004807616584002972, + 0.004525417927652597, + 0.005376941524446011, + 0.006203124299645424, + 0.005813163239508867, + 0.004561553243547678, + 0.005028909537941217, + 0.003375847125425935, + 0.006683763582259417, + 0.00435193907469511, + 0.013530424796044827, + 0.006190450396388769, + 0.007555859163403511, + 0.006133209448307753, + 0.005228025838732719, + 0.00429910235106945, + 0.008712153881788254, + 0.005531019996851683, + 0.006614548619836569, + 0.010032065212726593, + 0.01016251090914011, + 0.007729914505034685, + 0.007845458574593067, + 0.005851042922586203, + 0.017606202512979507, + 0.011706369929015636, + 0.013297132216393948, + 0.009719721041619778, + 0.018850062042474747, + 0.023086145520210266, + 0.012973100878298283, + 0.013165883719921112, + 0.0077482652850449085, + 0.005396033171564341, + 0.005588202271610498, + 0.006105361506342888, + 0.013528147712349892, + 0.00947801023721695, + 0.009616642259061337, + 0.005006243009120226, + 0.007378161419183016, + 0.006854947656393051, + 0.009492451325058937, + 0.008987604640424252, + 0.010700293816626072, + 0.008252366445958614, + 0.006890549790114164, + 0.011592392809689045, + 0.014822319149971008, + 0.020313700661063194, + 0.0161292664706707, + 0.02137669548392296, + 0.01550990343093872, + 0.019564755260944366, + 0.018103748559951782, + 0.01591554470360279, + 0.002222484676167369, + 0.0029933657497167587, + 0.0011333659058436751, + 0.002560204127803445, + 0.00315211433917284, + 0.0022339976858347654, + 0.0022332114167511463, + 0.0017245406052097678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 8 + ], + "selected_reward": -0.3564462959766388 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014214809983968735 + ], + "opposite_logprob_tokens": [ + -4.264214992523193 + ], + "logprob_reward": -0.014214809983968735, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014214809983968735, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31086912751197815, + 0.07038604468107224, + 0.06887239962816238, + 0.0839996188879013, + 0.1466137319803238, + 0.11149774491786957, + 0.05081211030483246, + 0.05364326015114784, + 0.023258434608578682, + 0.08004758507013321 + ], + "credit_map": [ + 0.047787219285964966, + 0.04467034712433815, + 0.047851838171482086, + 0.032179731875658035, + 0.024919597432017326, + 0.0323704294860363, + 0.041959311813116074, + 0.03913065046072006, + 0.006997664459049702, + 0.01104852743446827, + 0.00912004429847002, + 0.011874659918248653, + 0.006898802239447832, + 0.007158386055380106, + 0.010274088941514492, + 0.007013869471848011, + 0.009099054150283337, + 0.00765392417088151, + 0.007622382137924433, + 0.009264080785214901, + 0.013084239326417446, + 0.00795687548816204, + 0.0077236108481884, + 0.006468229927122593, + 0.01254611648619175, + 0.008682750165462494, + 0.008892061188817024, + 0.012660007923841476, + 0.01451770681887865, + 0.01019215863198042, + 0.011364182457327843, + 0.005144636612385511, + 0.02102339267730713, + 0.01588478498160839, + 0.010187387466430664, + 0.02148388884961605, + 0.029332946985960007, + 0.0159696564078331, + 0.02195323444902897, + 0.01077844388782978, + 0.020376086235046387, + 0.00675535062327981, + 0.010273141786456108, + 0.009128110483288765, + 0.016321441158652306, + 0.02898935042321682, + 0.010908180847764015, + 0.008746086619794369, + 0.00691095320507884, + 0.004079763311892748, + 0.004142765421420336, + 0.006870864890515804, + 0.00804633554071188, + 0.007332044653594494, + 0.009432375431060791, + 0.003997008316218853, + 0.004683012142777443, + 0.006241964641958475, + 0.005033304914832115, + 0.007998221553862095, + 0.010238250717520714, + 0.006969337817281485, + 0.007330593653023243, + 0.005148573778569698, + 0.002235082210972905, + 0.002325410721823573, + 0.0031485434155911207, + 0.0036453164648264647, + 0.00380463944748044, + 0.0029228911735117435, + 0.0024781315587460995, + 0.002698419848456979, + 0.007120386231690645, + 0.010079454630613327, + 0.007384218741208315, + 0.009920146316289902, + 0.013118725270032883, + 0.009643097408115864, + 0.012245716527104378, + 0.010535836219787598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.014214809983968735 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030859708786011 + ], + "opposite_logprob_tokens": [ + -2.350308656692505 + ], + "logprob_reward": -0.10030859708786011, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030859708786011, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1924598664045334, + 0.21614672243595123, + 0.03981133550405502, + 0.06344632059335709, + 0.0636768713593483, + 0.13063660264015198, + 0.12890809774398804, + 0.04536338150501251, + 0.0886399894952774, + 0.030910799279808998 + ], + "credit_map": [ + 0.021595913916826248, + 0.023709643632173538, + 0.038279373198747635, + 0.019782692193984985, + 0.01646837405860424, + 0.017879251390695572, + 0.022480864077806473, + 0.032263755798339844, + 0.0420570932328701, + 0.014401326887309551, + 0.025442710146307945, + 0.01560281217098236, + 0.011931831948459148, + 0.02016291953623295, + 0.01616974174976349, + 0.07037828862667084, + 0.0045800358057022095, + 0.003857614705339074, + 0.0051945107989013195, + 0.006022298242896795, + 0.0053044515661895275, + 0.005320786964148283, + 0.0045749349519610405, + 0.00495670223608613, + 0.010306110605597496, + 0.004540849477052689, + 0.010247712023556232, + 0.007062258664518595, + 0.010981831699609756, + 0.007916243746876717, + 0.0071217711083590984, + 0.005269545596092939, + 0.008481449447572231, + 0.006693608593195677, + 0.006478669121861458, + 0.011810618452727795, + 0.01040628831833601, + 0.0065834904089570045, + 0.008220762014389038, + 0.005001986399292946, + 0.021585511043667793, + 0.010362837463617325, + 0.01662859134376049, + 0.01233881339430809, + 0.012108924798667431, + 0.030331172049045563, + 0.01320298295468092, + 0.01407778449356556, + 0.015376497060060501, + 0.011852259747684002, + 0.014190460555255413, + 0.016767075285315514, + 0.022540783509612083, + 0.017432430759072304, + 0.020067494362592697, + 0.010681101121008396, + 0.00568513385951519, + 0.0050691161304712296, + 0.004855084698647261, + 0.004776632878929377, + 0.0075065698474645615, + 0.006663697771728039, + 0.005184020381420851, + 0.005623128730803728, + 0.010981831699609756, + 0.011886410415172577, + 0.010396250523626804, + 0.012183398939669132, + 0.01050218939781189, + 0.011109196580946445, + 0.01095374021679163, + 0.01062697358429432, + 0.003710699500516057, + 0.004958165809512138, + 0.002419108524918556, + 0.004636016674339771, + 0.004820290952920914, + 0.0037051704712212086, + 0.003228883258998394, + 0.0034324657171964645 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.10030859708786011 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012539708986878395 + ], + "opposite_logprob_tokens": [ + -4.387539863586426 + ], + "logprob_reward": -0.012539708986878395, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012539708986878395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08147064596414566, + 0.25538933277130127, + 0.04866023361682892, + 0.06720215827226639, + 0.05320744588971138, + 0.13016566634178162, + 0.21881097555160522, + 0.07515440881252289, + 0.01623990759253502, + 0.05369917303323746 + ], + "credit_map": [ + 0.006983858998864889, + 0.013516816310584545, + 0.010504554025828838, + 0.009627491235733032, + 0.010523948818445206, + 0.011475975625216961, + 0.010370932519435883, + 0.008467074483633041, + 0.05839751660823822, + 0.01826677843928337, + 0.030129235237836838, + 0.013055248185992241, + 0.009296122938394547, + 0.029807817190885544, + 0.0141121381893754, + 0.08232448250055313, + 0.0070080566219985485, + 0.004716772120445967, + 0.005569295957684517, + 0.0038880598731338978, + 0.00429324246942997, + 0.00860048457980156, + 0.004733610432595015, + 0.009850713424384594, + 0.015115832909941673, + 0.004473282489925623, + 0.010156803764402866, + 0.006281705107539892, + 0.0057188658975064754, + 0.01385919377207756, + 0.00554210739210248, + 0.006054371129721403, + 0.008724656887352467, + 0.004774670582264662, + 0.009216675534844398, + 0.004345222841948271, + 0.00691142026335001, + 0.008551765233278275, + 0.004250607918947935, + 0.006432429421693087, + 0.022413646802306175, + 0.009736924432218075, + 0.017681054770946503, + 0.01279341708868742, + 0.014129544608294964, + 0.03033594787120819, + 0.013618613593280315, + 0.009456533007323742, + 0.030427953228354454, + 0.018259616568684578, + 0.02743166871368885, + 0.025468923151493073, + 0.029266465455293655, + 0.03569118678569794, + 0.029068302363157272, + 0.023196890950202942, + 0.006431865971535444, + 0.011442337185144424, + 0.00506043853238225, + 0.009654179215431213, + 0.01783108338713646, + 0.007826440036296844, + 0.00868975929915905, + 0.008218308910727501, + 0.0018819408724084496, + 0.0017093552742153406, + 0.00253210193477571, + 0.0029591317288577557, + 0.0015243174275383353, + 0.002061177510768175, + 0.0016916635213419795, + 0.0018802194390445948, + 0.008554134517908096, + 0.004565629176795483, + 0.008243399672210217, + 0.003130638273432851, + 0.004517307039350271, + 0.011537968181073666, + 0.0034466651268303394, + 0.009703435003757477 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.012539708986878395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306522756814957 + ], + "opposite_logprob_tokens": [ + -3.168065309524536 + ], + "logprob_reward": -0.04306522756814957, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306522756814957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42063140869140625, + 0.2637636661529541, + 0.03777236118912697, + 0.05207795277237892, + 0.02335195802152157, + 0.038704294711351395, + 0.07314207404851913, + 0.0377938449382782, + 0.041411593556404114, + 0.011350825428962708 + ], + "credit_map": [ + 0.059475794434547424, + 0.04907215014100075, + 0.07430163025856018, + 0.04740608111023903, + 0.040984101593494415, + 0.047734688967466354, + 0.05464886128902435, + 0.04700811207294464, + 0.05957901105284691, + 0.0171758271753788, + 0.03140375018119812, + 0.013691280037164688, + 0.00903324969112873, + 0.030038263648748398, + 0.015567753463983536, + 0.08727455884218216, + 0.005497426260262728, + 0.003344116033986211, + 0.0035677074920386076, + 0.004814833402633667, + 0.003744073212146759, + 0.006057850085198879, + 0.004650522954761982, + 0.006095833610743284, + 0.010122811421751976, + 0.004326858092099428, + 0.012049570679664612, + 0.0037053455598652363, + 0.003506389679387212, + 0.010434613563120365, + 0.003187140915542841, + 0.004745228681713343, + 0.003385972697287798, + 0.0025497751776129007, + 0.002443766687065363, + 0.0030027180910110474, + 0.004944249056279659, + 0.002772099804133177, + 0.002583174966275692, + 0.0016702032880857587, + 0.005480355583131313, + 0.003309233346953988, + 0.0047035664319992065, + 0.004735137801617384, + 0.005492154508829117, + 0.007149247918277979, + 0.0051794578321278095, + 0.002655142918229103, + 0.009892201982438564, + 0.00687157828360796, + 0.00886395014822483, + 0.00901440717279911, + 0.009171879850327969, + 0.009868401102721691, + 0.010349119082093239, + 0.00911053828895092, + 0.004742087796330452, + 0.0045734187588095665, + 0.0038113717455416918, + 0.004855121951550245, + 0.006365205626934767, + 0.004730671644210815, + 0.004803100135177374, + 0.003912868443876505, + 0.0044479090720415115, + 0.004585512448102236, + 0.00507349194958806, + 0.004651249852031469, + 0.005494754295796156, + 0.007607482839375734, + 0.00428370013833046, + 0.005267500411719084, + 0.000978862983174622, + 0.0015084045007824898, + 0.0013095996109768748, + 0.0017282756743952632, + 0.0014873965410515666, + 0.0011103268479928374, + 0.0013171228347346187, + 0.001910837716422975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.04306522756814957 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, wide shot opens the scene, the camera violently panning backward to reveal the cataclysmic events unfolding in a surreal, Jodorowsky-inspired vision. The Empress is violently ejected from a vast, geometrically perfect crystal throne, ablaze with an icy inner light. Her body, adorned with a dramatic, jeweled velvet robe of crimson and royal purple, tumbles backward in a metaphysical void. Her alabaster face, a mask of shocked terror, her Kohl-rimmed eyes open, the Empress is caught in a chaotic, slow-motion ballet, her ornate golden headdress cast aside, her limbs flailing. The heavy, jeweled fabric of her robe billows around her, catching and shattering the light emanating from the throne and the surrounding cosmic dreamscape—a swirling indigo cloud and fleeting, glowing alchemical symbols. The entire scene flickers erratically with strobe-like intensity and digital glitches, as if the very fabric of reality is being short-circuited by the sheer, unseen force of impact, creating a deeply unsettling and visually saturated atmosphere of sacred, psychedelic violence.", + "seed": 24514, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_7/index_1739_seed_24514_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08360313102602959, + "vlm_logprob": -0.08360313102602959, + "vlm_yes_no_margin": 3.6000000119209288, + "vlm_token_credit": -0.08360313102602959, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.354705274105072 + ], + "opposite_logprob_tokens": [ + -1.2297053337097168 + ], + "logprob_reward": -0.354705274105072, + "yes_no_margin_reward": 0.8750000596046448, + "token_credit_reward": -0.354705274105072, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09995532780885696, + 0.027116121724247932, + 0.08235488086938858, + 0.039592914283275604, + 0.037624113261699677, + 0.026115400716662407, + 0.5563237071037292, + 0.026724770665168762, + 0.07045245915651321, + 0.03374028205871582 + ], + "credit_map": [ + 0.006318029947578907, + 0.014769913628697395, + 0.011119913309812546, + 0.012645200826227665, + 0.01610545814037323, + 0.00656048534438014, + 0.015569139271974564, + 0.01686718314886093, + 0.002542993286624551, + 0.0026804080698639154, + 0.003760168794542551, + 0.002988607157021761, + 0.0023338240571320057, + 0.006051856558769941, + 0.003685516305267811, + 0.003072747029364109, + 0.0104426434263587, + 0.006298225373029709, + 0.018309222534298897, + 0.00804819818586111, + 0.006529031787067652, + 0.012916605919599533, + 0.01067032665014267, + 0.009140631183981895, + 0.004188506864011288, + 0.0045366412959992886, + 0.0048248195089399815, + 0.005657399073243141, + 0.0042332252487540245, + 0.005604782607406378, + 0.003964537288993597, + 0.006583002395927906, + 0.010050281882286072, + 0.00407878402620554, + 0.003049723571166396, + 0.0027088443748652935, + 0.0032384705264121294, + 0.009011240676045418, + 0.0029937573708593845, + 0.0024930136278271675, + 0.0021983536425977945, + 0.00414383877068758, + 0.0037062005139887333, + 0.003458269638940692, + 0.0026540637481957674, + 0.0027136977296322584, + 0.002734303940087557, + 0.004506673198193312, + 0.04015606641769409, + 0.009334642440080643, + 0.020710250362753868, + 0.009964880533516407, + 0.04206861928105354, + 0.23917056620121002, + 0.011433429084718227, + 0.1834852695465088, + 0.0029483644757419825, + 0.004744996782392263, + 0.002626059576869011, + 0.003549041459336877, + 0.003169320523738861, + 0.0033372733741998672, + 0.003176954109221697, + 0.0031727594323456287, + 0.0058277021162211895, + 0.009023396298289299, + 0.00927584245800972, + 0.019693391397595406, + 0.006572827696800232, + 0.006802245508879423, + 0.007375174667686224, + 0.005881876684725285, + 0.0025040393229573965, + 0.0036471912171691656, + 0.005641290917992592, + 0.003813808551058173, + 0.005271071568131447, + 0.0030569680966436863, + 0.006576387211680412, + 0.0032295254059135914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.354705274105072 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007652727887034416 + ], + "opposite_logprob_tokens": [ + -4.882652759552002 + ], + "logprob_reward": -0.007652727887034416, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007652727887034416, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1751275360584259, + 0.10814934968948364, + 0.14773020148277283, + 0.1005634292960167, + 0.17986644804477692, + 0.059835199266672134, + 0.054690342396497726, + 0.04940018430352211, + 0.07064686715602875, + 0.053990453481674194 + ], + "credit_map": [ + 0.016746625304222107, + 0.023964112624526024, + 0.025232238695025444, + 0.019461939111351967, + 0.02339962311089039, + 0.02205468900501728, + 0.020958714187145233, + 0.023309584707021713, + 0.008324590511620045, + 0.009336848743259907, + 0.012613184750080109, + 0.026212230324745178, + 0.006839510053396225, + 0.013673664070665836, + 0.018123168498277664, + 0.013026141561567783, + 0.01882561668753624, + 0.01020727027207613, + 0.04149412736296654, + 0.016251524910330772, + 0.011968061327934265, + 0.014367890544235706, + 0.017737502232193947, + 0.016878189519047737, + 0.015855582430958748, + 0.010367188602685928, + 0.008593968115746975, + 0.00939919613301754, + 0.010530691593885422, + 0.025013547390699387, + 0.008987758308649063, + 0.011815490201115608, + 0.027912288904190063, + 0.022627942264080048, + 0.017993099987506866, + 0.014533258974552155, + 0.021238885819911957, + 0.02446981705725193, + 0.02347119338810444, + 0.02761993743479252, + 0.007290472276508808, + 0.0106905996799469, + 0.005556654185056686, + 0.00866735726594925, + 0.007076159119606018, + 0.005505863111466169, + 0.006865230854600668, + 0.008182855322957039, + 0.004595140926539898, + 0.006381371524184942, + 0.003079627873376012, + 0.011848646216094494, + 0.0054011270403862, + 0.009189308620989323, + 0.00730332313105464, + 0.006891796365380287, + 0.006135757081210613, + 0.007671622093766928, + 0.008688981644809246, + 0.0055398400872945786, + 0.00393680902197957, + 0.005505770910531282, + 0.005506028886884451, + 0.00641536433249712, + 0.008462400175631046, + 0.007736147381365299, + 0.006929101888090372, + 0.013770993798971176, + 0.012509404681622982, + 0.006965491455048323, + 0.008024315349757671, + 0.006249001249670982, + 0.004893050994724035, + 0.006753438618034124, + 0.010155541822314262, + 0.006899404339492321, + 0.005756835453212261, + 0.005300142336636782, + 0.008140329271554947, + 0.006091704126447439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.007652727887034416 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02980376034975052 + ], + "opposite_logprob_tokens": [ + -3.52980375289917 + ], + "logprob_reward": -0.02980376034975052, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02980376034975052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06254062056541443, + 0.08204267174005508, + 0.05843493714928627, + 0.03345457464456558, + 0.039593134075403214, + 0.03605462238192558, + 0.4691741168498993, + 0.026351917535066605, + 0.12294907122850418, + 0.06940436363220215 + ], + "credit_map": [ + 0.0043016839772462845, + 0.008211917243897915, + 0.005603954195976257, + 0.00986326765269041, + 0.008556857705116272, + 0.004639158491045237, + 0.010324068367481232, + 0.011039717122912407, + 0.006235078442841768, + 0.00627194344997406, + 0.012240582145750523, + 0.02207750454545021, + 0.004524987190961838, + 0.014081248082220554, + 0.009002932347357273, + 0.007608403451740742, + 0.008136431686580181, + 0.003687832737341523, + 0.012152823619544506, + 0.005448933690786362, + 0.004400904756039381, + 0.01199161447584629, + 0.0066488212905824184, + 0.005967576056718826, + 0.003690962214022875, + 0.0030187596566975117, + 0.0043739196844398975, + 0.003373474581167102, + 0.0033934249076992273, + 0.004028382711112499, + 0.0038596484810113907, + 0.00771600566804409, + 0.012008666060864925, + 0.00219014217145741, + 0.006779810413718224, + 0.002337774960324168, + 0.002008593874052167, + 0.00887750182300806, + 0.0018529529916122556, + 0.0035376902669668198, + 0.005713994149118662, + 0.006707185413688421, + 0.003177550621330738, + 0.005592348985373974, + 0.003140899119898677, + 0.0040528178215026855, + 0.0038647183682769537, + 0.003805109765380621, + 0.03408902883529663, + 0.008778720162808895, + 0.01595117524266243, + 0.008336691185832024, + 0.03552764654159546, + 0.20206543803215027, + 0.010258677415549755, + 0.15416677296161652, + 0.002955142641440034, + 0.003603978781029582, + 0.004672212991863489, + 0.002338959136977792, + 0.0025614702608436346, + 0.0043978504836559296, + 0.001996139530092478, + 0.003826164873316884, + 0.008941291831433773, + 0.016307415440678596, + 0.015428995713591576, + 0.03437398374080658, + 0.01101033203303814, + 0.011981344781816006, + 0.014313728548586369, + 0.010591977275907993, + 0.005580116529017687, + 0.008458919823169708, + 0.011009227484464645, + 0.00837160088121891, + 0.008008080534636974, + 0.006253364030271769, + 0.013747048564255238, + 0.007976008579134941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 1 + ], + "selected_reward": -0.02980376034975052 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783178567886353 + ], + "opposite_logprob_tokens": [ + -4.6347832679748535 + ], + "logprob_reward": -0.009783178567886353, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783178567886353, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21832187473773956, + 0.10327897220849991, + 0.12997910380363464, + 0.16745489835739136, + 0.06311298906803131, + 0.049297768622636795, + 0.11698949337005615, + 0.04989033192396164, + 0.02972470223903656, + 0.07194984704256058 + ], + "credit_map": [ + 0.016886208206415176, + 0.037869907915592194, + 0.021463926881551743, + 0.03375166282057762, + 0.03639078140258789, + 0.014570591039955616, + 0.02903815172612667, + 0.028350653126835823, + 0.009557042270898819, + 0.008449158631265163, + 0.02027149498462677, + 0.012612618505954742, + 0.008362001739442348, + 0.010921124368906021, + 0.01936952769756317, + 0.013736008666455746, + 0.019970856606960297, + 0.008303569629788399, + 0.035901814699172974, + 0.011900398880243301, + 0.009092097170650959, + 0.012355656363070011, + 0.014186696149408817, + 0.01826801896095276, + 0.015622232109308243, + 0.020458374172449112, + 0.018203597515821457, + 0.01800784096121788, + 0.020471008494496346, + 0.022289857268333435, + 0.019033877179026604, + 0.03336811438202858, + 0.007945958524942398, + 0.005664220079779625, + 0.015977496281266212, + 0.005174766294658184, + 0.006940419320017099, + 0.008703862316906452, + 0.005262019578367472, + 0.00744424294680357, + 0.00475815637037158, + 0.007109436672180891, + 0.007049212697893381, + 0.006467347964644432, + 0.003244140651077032, + 0.005202033556997776, + 0.0055253892205655575, + 0.00994205567985773, + 0.008187943138182163, + 0.007208602037280798, + 0.006459361873567104, + 0.007139424793422222, + 0.007621711120009422, + 0.043073177337646484, + 0.0048789517022669315, + 0.0324203260242939, + 0.006264029070734978, + 0.006003221962600946, + 0.007592391688376665, + 0.004052780568599701, + 0.0045571415685117245, + 0.00840190052986145, + 0.005144387483596802, + 0.007874480448663235, + 0.003045914461836219, + 0.003123147878795862, + 0.003008725354447961, + 0.004402270540595055, + 0.004690683912485838, + 0.003432381432503462, + 0.004350452683866024, + 0.003671124577522278, + 0.010087469592690468, + 0.007750018034130335, + 0.011591089889407158, + 0.0090145543217659, + 0.009390008635818958, + 0.008219176903367043, + 0.009557398967444897, + 0.006340136285871267 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.009783178567886353 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a low-angle wide shot that pans violently backward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a character identified as the Empress visible, wearing a jeweled velvet robe of crimson and royal purple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the Empress shown being violently ejected from a vast, geometrically perfect crystal throne ablaze with icy inner light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a surreal cosmic dreamscape featuring a swirling indigo cloud and glowing alchemical symbols?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the scene exhibit erratic flickering with strobe-like intensity and digital glitches?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070714220404625 + ], + "opposite_logprob_tokens": [ + -4.14107084274292 + ], + "logprob_reward": -0.016070714220404625, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070714220404625, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3473747670650482, + 0.07617617398500443, + 0.05335250124335289, + 0.052781518548727036, + 0.03979096934199333, + 0.06027832254767418, + 0.2839352488517761, + 0.026991749182343483, + 0.02094287984073162, + 0.038375988602638245 + ], + "credit_map": [ + 0.03760354220867157, + 0.03358878195285797, + 0.06625839322805405, + 0.04897896572947502, + 0.04174044728279114, + 0.042565859854221344, + 0.04191584512591362, + 0.03472285345196724, + 0.0054662469774484634, + 0.004441157449036837, + 0.00850258395075798, + 0.02123194746673107, + 0.003410419449210167, + 0.017035681754350662, + 0.008658581413328648, + 0.007429542951285839, + 0.007424972485750914, + 0.004431849345564842, + 0.007050829008221626, + 0.005288905464112759, + 0.003919470589607954, + 0.009211446158587933, + 0.008238028734922409, + 0.007786993868649006, + 0.005669370759278536, + 0.009191729128360748, + 0.005639579147100449, + 0.0070700570940971375, + 0.006391061935573816, + 0.007615775801241398, + 0.006211008410900831, + 0.004992929752916098, + 0.008006088435649872, + 0.004518750123679638, + 0.007432353682816029, + 0.0045298379845917225, + 0.0034433561377227306, + 0.0039135911501944065, + 0.0037592840380966663, + 0.004187702666968107, + 0.004741228185594082, + 0.007780191954225302, + 0.009987820871174335, + 0.0054379054345190525, + 0.00612618587911129, + 0.00599611084908247, + 0.00585975544527173, + 0.014349115081131458, + 0.02074909582734108, + 0.00556746032088995, + 0.010866129770874977, + 0.005303696729242802, + 0.020688040181994438, + 0.12137370556592941, + 0.006029795855283737, + 0.09335728734731674, + 0.0030844362918287516, + 0.003444601083174348, + 0.004414109047502279, + 0.002599285217002034, + 0.0024895495735108852, + 0.0031637900974601507, + 0.0026447661221027374, + 0.005151207558810711, + 0.0024997424334287643, + 0.002766047604382038, + 0.0017516401130706072, + 0.003888464532792568, + 0.0031421498861163855, + 0.0019623327534645796, + 0.002323754597455263, + 0.0026087446603924036, + 0.004263584036380053, + 0.0048456802032887936, + 0.007015560753643513, + 0.00419971439987421, + 0.005150700453668833, + 0.004277845844626427, + 0.004995237570255995, + 0.003627658123150468 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.016070714220404625 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_tensor([1307, 1307])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_tensor([1307, 1307])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e9fde2d6235b5b1a691cdc522565fca56593f35a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_tensor([1307, 1307])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Beginning with a high-angle shot rendered with stark, functional realism, the camera executes a slow, steady vertical descent, moving downward through the chilly night air. Below, a vast geothermal facility is revealed, its intricate network of pipes forming a maze of industrial-grade steel, crisscrossing the dark, frozen ground. Immense steam pressure surges upward from a large valve within this network, creating a powerful, sustained eruption that forms a thick, roiling cloud that billows and expands into the sky. The scene is dramatically illuminated by a spectacular aurora, its vibrant, dancing curtain of green and purple light sweeping across the facility. This celestial glow casts an otherworldly glow, causing the metal pipes to shimmer with shifting, cold points of light and causing the turbulent steam cloud itself to glow from within, creating a luminous, ethereal substance that contrasts starkly with the desolate, functional buildings and the vast, dark night.", + "seed": 4461, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_1307_seed_4461_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021964077651500703, + "vlm_logprob": -0.021964077651500703, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.021964077651500703, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08860336989164352 + ], + "opposite_logprob_tokens": [ + -2.5886032581329346 + ], + "logprob_reward": -0.08860336989164352, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.08860336989164352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3311053514480591, + 0.06360755860805511, + 0.09489234536886215, + 0.03861262649297714, + 0.07504508644342422, + 0.07258972525596619, + 0.03268831968307495, + 0.1336119920015335, + 0.09439204633235931, + 0.06345493346452713 + ], + "credit_map": [ + 0.04212116077542305, + 0.041026558727025986, + 0.04765167832374573, + 0.029218794777989388, + 0.03602731227874756, + 0.0562606006860733, + 0.05035524070262909, + 0.02844400145113468, + 0.005556837655603886, + 0.006804477423429489, + 0.00805710069835186, + 0.006879844702780247, + 0.008291929960250854, + 0.006199895869940519, + 0.013749489560723305, + 0.008067982271313667, + 0.010610785335302353, + 0.01286259014159441, + 0.011525939218699932, + 0.011645594611763954, + 0.009090244770050049, + 0.010986444540321827, + 0.017433155328035355, + 0.010737593285739422, + 0.003646413329988718, + 0.005356167443096638, + 0.00585391977801919, + 0.004904478322714567, + 0.0032992830965667963, + 0.0036365268751978874, + 0.006064271554350853, + 0.005851568188518286, + 0.007476518861949444, + 0.007397958543151617, + 0.012349143624305725, + 0.013434362597763538, + 0.006634081248193979, + 0.008661282248795033, + 0.011488571763038635, + 0.007603165693581104, + 0.006364010740071535, + 0.007841675542294979, + 0.01926887407898903, + 0.01271092426031828, + 0.003829395165666938, + 0.007210558280348778, + 0.008005158044397831, + 0.007359127514064312, + 0.0034151021391153336, + 0.006288011092692614, + 0.005495835095643997, + 0.004719607066363096, + 0.0029682861641049385, + 0.002980272052809596, + 0.003695684252306819, + 0.003125522518530488, + 0.01018914207816124, + 0.014992213808000088, + 0.021327929571270943, + 0.02711542136967182, + 0.009363583289086819, + 0.019287489354610443, + 0.016495967283844948, + 0.014840244315564632, + 0.007386112120002508, + 0.010322374291718006, + 0.016797000542283058, + 0.014753017574548721, + 0.010073966346681118, + 0.01219696830958128, + 0.012945303693413734, + 0.009917299263179302, + 0.011538561433553696, + 0.015114453621208668, + 0.005955420434474945, + 0.008418593555688858, + 0.006004732567816973, + 0.004636435769498348, + 0.006661706604063511, + 0.005125030875205994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.08860336989164352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012566667050123215 + ], + "opposite_logprob_tokens": [ + -4.387566566467285 + ], + "logprob_reward": -0.012566667050123215, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012566667050123215, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07909741252660751, + 0.051367245614528656, + 0.0463215708732605, + 0.03179140016436577, + 0.03812786936759949, + 0.06337438523769379, + 0.08577419817447662, + 0.10473807901144028, + 0.3229284882545471, + 0.17647932469844818 + ], + "credit_map": [ + 0.00856019463390112, + 0.017050743103027344, + 0.010437232442200184, + 0.007334407884627581, + 0.007796718273311853, + 0.010564394295215607, + 0.011448576115071774, + 0.005905152764171362, + 0.006903898902237415, + 0.007061193697154522, + 0.006484395358711481, + 0.0047506666742265224, + 0.005711996462196112, + 0.005980899557471275, + 0.008534207008779049, + 0.005939991679042578, + 0.0049584414809942245, + 0.00865784753113985, + 0.00628431374207139, + 0.0049710324965417385, + 0.005201372317969799, + 0.004967148415744305, + 0.006777115166187286, + 0.004504304379224777, + 0.004562524612993002, + 0.002399119548499584, + 0.009575078263878822, + 0.003414211329072714, + 0.0023263150360435247, + 0.0027059440035372972, + 0.00292757130227983, + 0.0038806372322142124, + 0.004223915282636881, + 0.0033784282859414816, + 0.006640944164246321, + 0.00715296994894743, + 0.003403967712074518, + 0.004346141591668129, + 0.005553406663239002, + 0.0034281020052731037, + 0.005967500153928995, + 0.009235328994691372, + 0.01473987940698862, + 0.009909319691359997, + 0.005200604908168316, + 0.007331657223403454, + 0.005522897467017174, + 0.005467202048748732, + 0.010405166074633598, + 0.011677060276269913, + 0.01061366405338049, + 0.013226638548076153, + 0.007248636335134506, + 0.014169983565807343, + 0.009074698202311993, + 0.009358360432088375, + 0.007908194325864315, + 0.009943373501300812, + 0.015166154131293297, + 0.021434295922517776, + 0.009162825532257557, + 0.016115901991724968, + 0.015422502532601357, + 0.00958484411239624, + 0.0331696979701519, + 0.04418748617172241, + 0.05637820437550545, + 0.047791216522455215, + 0.03444359451532364, + 0.048568811267614365, + 0.028440549969673157, + 0.029948946088552475, + 0.029191989451646805, + 0.04318641126155853, + 0.019692156463861465, + 0.02436812035739422, + 0.013646529987454414, + 0.017322875559329987, + 0.013671000488102436, + 0.015400251373648643 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.012566667050123215 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022078203037381172 + ], + "opposite_logprob_tokens": [ + -6.1272077560424805 + ], + "logprob_reward": -0.0022078203037381172, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022078203037381172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19082386791706085, + 0.046990711241960526, + 0.043939292430877686, + 0.04865798354148865, + 0.08346568793058395, + 0.05469215661287308, + 0.05697612091898918, + 0.144243061542511, + 0.1955159306526184, + 0.13469521701335907 + ], + "credit_map": [ + 0.02474968694150448, + 0.024388127028942108, + 0.01810595951974392, + 0.017425009980797768, + 0.021899113431572914, + 0.028860649093985558, + 0.03416847065091133, + 0.021226856857538223, + 0.0049852048978209496, + 0.007595119997859001, + 0.006037191953510046, + 0.004784928634762764, + 0.0046635461039841175, + 0.006065668072551489, + 0.007068841718137264, + 0.005790207535028458, + 0.0031040851026773453, + 0.0033217717427760363, + 0.006722305901348591, + 0.00428195483982563, + 0.0044644670560956, + 0.0069967275485396385, + 0.008645319379866123, + 0.006402658764272928, + 0.00451032305136323, + 0.005404527299106121, + 0.0058921221643686295, + 0.005957350134849548, + 0.006703833118081093, + 0.005992261692881584, + 0.009819602593779564, + 0.004377960227429867, + 0.007466228678822517, + 0.007539558690041304, + 0.0201196800917387, + 0.009208867326378822, + 0.005109146703034639, + 0.006358519196510315, + 0.008243194781243801, + 0.01942049153149128, + 0.004967695102095604, + 0.008642561733722687, + 0.015458245761692524, + 0.00761696370318532, + 0.0037052284460514784, + 0.005160566419363022, + 0.003969200421124697, + 0.005171689670532942, + 0.00626662140712142, + 0.006983044557273388, + 0.007332765497267246, + 0.008804747834801674, + 0.004742534831166267, + 0.009746716357767582, + 0.0062890914268791676, + 0.006810592953115702, + 0.011346633546054363, + 0.01441169623285532, + 0.022812495008111, + 0.028846492990851402, + 0.013906002976000309, + 0.021647965535521507, + 0.017874043434858322, + 0.01339771319180727, + 0.018809184432029724, + 0.022740881890058517, + 0.04413481429219246, + 0.026606248691678047, + 0.019181402400135994, + 0.02891751378774643, + 0.01698133535683155, + 0.018144527450203896, + 0.020036078989505768, + 0.01584075205028057, + 0.014574903063476086, + 0.02069871500134468, + 0.016012107953429222, + 0.0176150631159544, + 0.014637049287557602, + 0.015280541963875294 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0022078203037381172 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011820956133306026 + ], + "opposite_logprob_tokens": [ + -6.7511820793151855 + ], + "logprob_reward": -0.0011820956133306026, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011820956133306026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20295941829681396, + 0.17429949343204498, + 0.05556097254157066, + 0.05020836740732193, + 0.0996706560254097, + 0.033850736916065216, + 0.0561649352312088, + 0.04364052414894104, + 0.15635357797145844, + 0.12729127705097198 + ], + "credit_map": [ + 0.03195654973387718, + 0.027966011315584183, + 0.026559244841337204, + 0.014629559591412544, + 0.019948823377490044, + 0.03977960720658302, + 0.02302665449678898, + 0.019092978909611702, + 0.02708875946700573, + 0.021326936781406403, + 0.016595251858234406, + 0.01459096372127533, + 0.025056637823581696, + 0.016559511423110962, + 0.03939943015575409, + 0.01368201244622469, + 0.008368989452719688, + 0.01153328362852335, + 0.006858088541775942, + 0.0048054358921945095, + 0.004998470656573772, + 0.007560276426374912, + 0.006572362966835499, + 0.004864070564508438, + 0.0065729753114283085, + 0.004001757130026817, + 0.014314514584839344, + 0.004655598197132349, + 0.004337197169661522, + 0.0040816254913806915, + 0.00536154443398118, + 0.00688316160812974, + 0.010609407909214497, + 0.010514512658119202, + 0.018817253410816193, + 0.019357779994606972, + 0.007670846302062273, + 0.011290645226836205, + 0.0094820661470294, + 0.011928150430321693, + 0.0038535650819540024, + 0.005982857197523117, + 0.007147450931370258, + 0.0044671399518847466, + 0.002533884486183524, + 0.004439554177224636, + 0.0029582504648715258, + 0.002468035789206624, + 0.006147050764411688, + 0.007634451147168875, + 0.009400435723364353, + 0.006672928109765053, + 0.004485579207539558, + 0.008613532409071922, + 0.00668652867898345, + 0.006524432450532913, + 0.005080429837107658, + 0.00762020330876112, + 0.006251899991184473, + 0.006214630790054798, + 0.0047520785592496395, + 0.0055150664411485195, + 0.004068535286933184, + 0.004137681797146797, + 0.016533445566892624, + 0.0205334834754467, + 0.03391624987125397, + 0.02142525091767311, + 0.015255982056260109, + 0.02298525534570217, + 0.011864257045090199, + 0.01383967138826847, + 0.022488411515951157, + 0.022542864084243774, + 0.013311074115335941, + 0.01676955819129944, + 0.014591321349143982, + 0.01189576368778944, + 0.012756070122122765, + 0.012936224229633808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0011820956133306026 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005260435398668051 + ], + "opposite_logprob_tokens": [ + -5.255260467529297 + ], + "logprob_reward": -0.005260435398668051, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005260435398668051, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14240144193172455, + 0.07833313941955566, + 0.03706590086221695, + 0.04438794031739235, + 0.05251728743314743, + 0.05940847843885422, + 0.0700204074382782, + 0.2196807563304901, + 0.08941841125488281, + 0.20676617324352264 + ], + "credit_map": [ + 0.016419092193245888, + 0.02380545623600483, + 0.017548656091094017, + 0.016284629702568054, + 0.013576981611549854, + 0.020849552005529404, + 0.018319308757781982, + 0.01559776533395052, + 0.01160330418497324, + 0.013173992745578289, + 0.008880223147571087, + 0.008832604624330997, + 0.008089149370789528, + 0.009680709801614285, + 0.01098907645791769, + 0.007084074895828962, + 0.0048829615116119385, + 0.005531542934477329, + 0.0038323693443089724, + 0.0045528835617005825, + 0.0029314893763512373, + 0.003346286714076996, + 0.007993959821760654, + 0.0039944034069776535, + 0.005082692019641399, + 0.004736633040010929, + 0.010342802852392197, + 0.00588403083384037, + 0.003655885811895132, + 0.005777040030807257, + 0.004404983948916197, + 0.0045038736425340176, + 0.006358343176543713, + 0.0063604675233364105, + 0.008268186822533607, + 0.008817755617201328, + 0.004741029813885689, + 0.005668616853654385, + 0.007759918458759785, + 0.00454296451061964, + 0.005402474198490381, + 0.007625649217516184, + 0.012155030854046345, + 0.009938850998878479, + 0.0043057543225586414, + 0.006943541578948498, + 0.0070117381401360035, + 0.006025440525263548, + 0.006706170737743378, + 0.007578403688967228, + 0.009236011654138565, + 0.009413007646799088, + 0.005544794257730246, + 0.009867774322628975, + 0.008905394934117794, + 0.012768849730491638, + 0.02108701691031456, + 0.024179639294743538, + 0.03918307647109032, + 0.03525213897228241, + 0.015191754326224327, + 0.0352754145860672, + 0.0262435469776392, + 0.023268168792128563, + 0.009628658182919025, + 0.014798815362155437, + 0.014249197207391262, + 0.010929611511528492, + 0.01146162860095501, + 0.008028518408536911, + 0.012975719757378101, + 0.007346261292695999, + 0.03580699861049652, + 0.0399281345307827, + 0.024209922179579735, + 0.03143865615129471, + 0.020242612808942795, + 0.01683979108929634, + 0.019674833863973618, + 0.018625222146511078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.005260435398668051 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a high-angle shot rendered with stark, functional realism, the camera executes a slow, steady vertical descent, moving downward through the chilly night air. Below, a vast geothermal facility is revealed, its intricate network of pipes forming a maze of industrial-grade steel, crisscrossing the dark, frozen ground. Immense steam pressure surges upward from a large valve within this network, creating a powerful, sustained eruption that forms a thick, roiling cloud that billows and expands into the sky. The scene is dramatically illuminated by a spectacular aurora, its vibrant, dancing curtain of green and purple light sweeping across the facility. This celestial glow casts an otherworldly glow, causing the metal pipes to shimmer with shifting, cold points of light and causing the turbulent steam cloud itself to glow from within, creating a luminous, ethereal substance that contrasts starkly with the desolate, functional buildings and the vast, dark night.", + "seed": 4462, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_1307_seed_4462_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.024116956908255816, + "vlm_logprob": -0.024116956908255816, + "vlm_yes_no_margin": 4.7, + "vlm_token_credit": -0.024116956908255816, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08392070233821869 + ], + "opposite_logprob_tokens": [ + -2.70892071723938 + ], + "logprob_reward": -0.08392070233821869, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08392070233821869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19714492559432983, + 0.06838173419237137, + 0.0958024114370346, + 0.09921851009130478, + 0.06283217668533325, + 0.17702710628509521, + 0.13860762119293213, + 0.03246983513236046, + 0.08826780319213867, + 0.040247876197099686 + ], + "credit_map": [ + 0.029180722311139107, + 0.023998593911528587, + 0.030969860032200813, + 0.0298701673746109, + 0.024716531857848167, + 0.023482030257582664, + 0.016729174181818962, + 0.018197840079665184, + 0.00514702033251524, + 0.007224058732390404, + 0.008429921232163906, + 0.009786298498511314, + 0.008146489970386028, + 0.011178309097886086, + 0.007464718539267778, + 0.011004921048879623, + 0.006682282313704491, + 0.00805875938385725, + 0.008810347877442837, + 0.009359578602015972, + 0.013047980144619942, + 0.01864798180758953, + 0.01736425794661045, + 0.01383121870458126, + 0.010856150649487972, + 0.0165429525077343, + 0.00837788637727499, + 0.007844754494726658, + 0.009199061430990696, + 0.015460397116839886, + 0.01967848837375641, + 0.011258824728429317, + 0.006574020255357027, + 0.006133385468274355, + 0.010357935912907124, + 0.012582963332533836, + 0.0054318904876708984, + 0.007576672825962305, + 0.005760445259511471, + 0.008414863608777523, + 0.0190939512103796, + 0.021815896034240723, + 0.02311188541352749, + 0.03457018733024597, + 0.019753696396946907, + 0.0214371457695961, + 0.013963031582534313, + 0.023281317204236984, + 0.013555938377976418, + 0.018204525113105774, + 0.019521545618772507, + 0.018212351948022842, + 0.019383205100893974, + 0.015211598947644234, + 0.01477945689111948, + 0.01973901130259037, + 0.0032556988298892975, + 0.002881182124838233, + 0.004786815959960222, + 0.004868512507528067, + 0.004184321500360966, + 0.0034211277961730957, + 0.00507436553016305, + 0.00399780971929431, + 0.025779996067285538, + 0.008742484264075756, + 0.012283200398087502, + 0.006629555486142635, + 0.00661731930449605, + 0.012908837758004665, + 0.007513047195971012, + 0.007793360389769077, + 0.005035256966948509, + 0.0056291911751031876, + 0.005587458610534668, + 0.0049669258296489716, + 0.005427757743746042, + 0.00447692908346653, + 0.00516341719776392, + 0.0039609442465007305 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.08392070233821869 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026373712345957756 + ], + "opposite_logprob_tokens": [ + -3.6513736248016357 + ], + "logprob_reward": -0.026373712345957756, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026373712345957756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07195018231868744, + 0.04780982434749603, + 0.05603288114070892, + 0.22490832209587097, + 0.19540660083293915, + 0.11412319540977478, + 0.10385871678590775, + 0.048595547676086426, + 0.09814538061618805, + 0.03916940465569496 + ], + "credit_map": [ + 0.009652428328990936, + 0.010297710075974464, + 0.0117599256336689, + 0.01033126562833786, + 0.009376646019518375, + 0.0065152901224792, + 0.008631042204797268, + 0.005385881289839745, + 0.004084858577698469, + 0.007235875818878412, + 0.0034576570615172386, + 0.005764095112681389, + 0.006652428302913904, + 0.007361620664596558, + 0.00746951624751091, + 0.005783773493021727, + 0.0059025767259299755, + 0.004860537592321634, + 0.004766006488353014, + 0.007232784293591976, + 0.0075401924550533295, + 0.009018275886774063, + 0.006346187554299831, + 0.010366320610046387, + 0.02724582515656948, + 0.04740144684910774, + 0.020744474604725838, + 0.013694561086595058, + 0.023356720805168152, + 0.0328814797103405, + 0.03695109859108925, + 0.02263270504772663, + 0.022338181734085083, + 0.01680869609117508, + 0.030459221452474594, + 0.034440405666828156, + 0.026605477556586266, + 0.024448053911328316, + 0.016268080100417137, + 0.02403847686946392, + 0.013720905408263206, + 0.013114101253449917, + 0.027600795030593872, + 0.01688181795179844, + 0.01068277470767498, + 0.01210689079016447, + 0.007296841125935316, + 0.012719068676233292, + 0.010944606736302376, + 0.017819734290242195, + 0.01442214660346508, + 0.012572363018989563, + 0.014416136778891087, + 0.010481002740561962, + 0.009954788722097874, + 0.013247940689325333, + 0.003496226854622364, + 0.005797384772449732, + 0.007201544940471649, + 0.00788863841444254, + 0.005688908044248819, + 0.005237632431089878, + 0.006036258768290281, + 0.007248952053487301, + 0.023956555873155594, + 0.010815761983394623, + 0.013628445565700531, + 0.008377420715987682, + 0.007656829431653023, + 0.012911083176732063, + 0.010583531111478806, + 0.010215756483376026, + 0.003166814334690571, + 0.004423723556101322, + 0.004787725862115622, + 0.006507857237011194, + 0.005534475669264793, + 0.005230904556810856, + 0.004787154961377382, + 0.004730748478323221 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 5 + ], + "selected_reward": -0.026373712345957756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209723461419344 + ], + "opposite_logprob_tokens": [ + -6.127209663391113 + ], + "logprob_reward": -0.002209723461419344, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209723461419344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19663724303245544, + 0.11080677062273026, + 0.06898140162229538, + 0.11312944442033768, + 0.06831706315279007, + 0.12966293096542358, + 0.06306683272123337, + 0.06578417867422104, + 0.09816484153270721, + 0.08544928580522537 + ], + "credit_map": [ + 0.03070969507098198, + 0.021885601803660393, + 0.03324418142437935, + 0.0311532411724329, + 0.020398778840899467, + 0.023093663156032562, + 0.018739255145192146, + 0.01741281896829605, + 0.008778956718742847, + 0.014151720330119133, + 0.009276563301682472, + 0.014988013543188572, + 0.014981038868427277, + 0.01603199914097786, + 0.019294336438179016, + 0.013304133899509907, + 0.004721102770417929, + 0.006117689423263073, + 0.006504612974822521, + 0.010305332019925117, + 0.009118625894188881, + 0.012416177429258823, + 0.00826876237988472, + 0.011529102921485901, + 0.009184441529214382, + 0.021329786628484726, + 0.01406324002891779, + 0.01110164076089859, + 0.010642190463840961, + 0.016247650608420372, + 0.014727986417710781, + 0.01583249866962433, + 0.008652668446302414, + 0.006833534222096205, + 0.01019931212067604, + 0.008877298794686794, + 0.006701985839754343, + 0.009130529128015041, + 0.008554362691938877, + 0.00936736911535263, + 0.019022498279809952, + 0.011086848564445972, + 0.03557068109512329, + 0.015403909608721733, + 0.015148292295634747, + 0.012231064029037952, + 0.008215020410716534, + 0.012984616681933403, + 0.005096341948956251, + 0.009721163660287857, + 0.008390365168452263, + 0.008580374531447887, + 0.006419213023036718, + 0.007542774081230164, + 0.006919597741216421, + 0.010396999306976795, + 0.005443855654448271, + 0.011644475162029266, + 0.00943625159561634, + 0.009775472804903984, + 0.006738292519003153, + 0.006936316844075918, + 0.007442378904670477, + 0.008367134258151054, + 0.012561550363898277, + 0.011632513254880905, + 0.02002347633242607, + 0.009143722243607044, + 0.008522691205143929, + 0.018281206488609314, + 0.009253283031284809, + 0.008746392093598843, + 0.007598694413900375, + 0.009974018670618534, + 0.010782304219901562, + 0.009140565991401672, + 0.01187845878303051, + 0.01554642803966999, + 0.009386783465743065, + 0.011142024770379066 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.002209723461419344 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028219898231327534 + ], + "opposite_logprob_tokens": [ + -5.877821922302246 + ], + "logprob_reward": -0.0028219898231327534, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028219898231327534, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06133837252855301, + 0.04153811186552048, + 0.06487273424863815, + 0.0785742998123169, + 0.07569444179534912, + 0.14412733912467957, + 0.11267203837633133, + 0.113277867436409, + 0.16131460666656494, + 0.14659017324447632 + ], + "credit_map": [ + 0.009707742370665073, + 0.007047359831631184, + 0.010696941055357456, + 0.00939228292554617, + 0.005432507488876581, + 0.008795328438282013, + 0.00488510075956583, + 0.005381113849580288, + 0.003031390253454447, + 0.006067234091460705, + 0.006477206014096737, + 0.004493159707635641, + 0.005313670728355646, + 0.004730661399662495, + 0.0046943784691393375, + 0.006730412598699331, + 0.006833330728113651, + 0.0057879299856722355, + 0.008826769888401031, + 0.008332808502018452, + 0.0080428346991539, + 0.010977779515087605, + 0.008439388126134872, + 0.007631896063685417, + 0.011049807071685791, + 0.0180395245552063, + 0.00604568887501955, + 0.007691577542573214, + 0.008085612207651138, + 0.014280886389315128, + 0.005658065900206566, + 0.007723141461610794, + 0.007824680767953396, + 0.006272660568356514, + 0.013741978444159031, + 0.008136468939483166, + 0.009586558677256107, + 0.01083067525178194, + 0.008083109743893147, + 0.011218313127756119, + 0.014786205254495144, + 0.023052936419844627, + 0.024641023948788643, + 0.013708621263504028, + 0.01952999271452427, + 0.016651907935738564, + 0.012642485089600086, + 0.01911417953670025, + 0.009179345332086086, + 0.022323202341794968, + 0.014409997500479221, + 0.013231187127530575, + 0.013990622945129871, + 0.017414197325706482, + 0.00869728159159422, + 0.013426209799945354, + 0.008542243391275406, + 0.015865445137023926, + 0.015301806852221489, + 0.011629216372966766, + 0.020914312452077866, + 0.011775827966630459, + 0.01346239447593689, + 0.015786634758114815, + 0.02540770173072815, + 0.019177064299583435, + 0.04057604819536209, + 0.012770197354257107, + 0.009232443757355213, + 0.019687214866280556, + 0.013228531926870346, + 0.021235397085547447, + 0.015448260121047497, + 0.01672174781560898, + 0.015520782209932804, + 0.016896892338991165, + 0.01381079014390707, + 0.02754141390323639, + 0.017120255157351494, + 0.02353004179894924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.0028219898231327534 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09320623427629471, + 0.09103547036647797, + 0.10205859690904617, + 0.06975279748439789, + 0.1280631124973297, + 0.19033773243427277, + 0.07038547843694687, + 0.05885370820760727, + 0.1134733110666275, + 0.08283351361751556 + ], + "credit_map": [ + 0.012429243884980679, + 0.013971108011901379, + 0.013042101636528969, + 0.010405315086245537, + 0.01290920376777649, + 0.00932796485722065, + 0.011282939463853836, + 0.009838365949690342, + 0.005945298355072737, + 0.01551569439470768, + 0.009182922542095184, + 0.012881620787084103, + 0.01271664910018444, + 0.011473813094198704, + 0.010004645213484764, + 0.013314834795892239, + 0.006437614560127258, + 0.009199918247759342, + 0.009615713730454445, + 0.01289110817015171, + 0.014694560319185257, + 0.0157035943120718, + 0.014452066272497177, + 0.019064033403992653, + 0.006279349327087402, + 0.008571088314056396, + 0.009326368570327759, + 0.010717825964093208, + 0.005156185477972031, + 0.006132038310170174, + 0.010173746384680271, + 0.01339619979262352, + 0.013202022761106491, + 0.017742332071065903, + 0.01878146454691887, + 0.01840478926897049, + 0.009733536280691624, + 0.026928871870040894, + 0.009537518955767155, + 0.013732579536736012, + 0.022594690322875977, + 0.022702142596244812, + 0.0391482412815094, + 0.026099393144249916, + 0.019295569509267807, + 0.02063390053808689, + 0.011986815370619297, + 0.027876991778612137, + 0.0060758814215660095, + 0.010525278747081757, + 0.009816114790737629, + 0.009208687581121922, + 0.0103836040943861, + 0.007691361475735903, + 0.0066396803595125675, + 0.010044872760772705, + 0.004308886826038361, + 0.011223717592656612, + 0.00894026830792427, + 0.0070801894180476665, + 0.007532428950071335, + 0.0063436985947191715, + 0.0053936452604830265, + 0.008030872792005539, + 0.010482646524906158, + 0.015925303101539612, + 0.023403901606798172, + 0.008808969520032406, + 0.012373412027955055, + 0.013758311979472637, + 0.015303678810596466, + 0.013417097739875317, + 0.010486737824976444, + 0.007542220875620842, + 0.009476643055677414, + 0.009320421144366264, + 0.006280265282839537, + 0.017966194078326225, + 0.009441524744033813, + 0.012319514527916908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 8 + ], + "selected_reward": -0.005258656572550535 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a high-angle shot rendered with stark, functional realism, the camera executes a slow, steady vertical descent, moving downward through the chilly night air. Below, a vast geothermal facility is revealed, its intricate network of pipes forming a maze of industrial-grade steel, crisscrossing the dark, frozen ground. Immense steam pressure surges upward from a large valve within this network, creating a powerful, sustained eruption that forms a thick, roiling cloud that billows and expands into the sky. The scene is dramatically illuminated by a spectacular aurora, its vibrant, dancing curtain of green and purple light sweeping across the facility. This celestial glow casts an otherworldly glow, causing the metal pipes to shimmer with shifting, cold points of light and causing the turbulent steam cloud itself to glow from within, creating a luminous, ethereal substance that contrasts starkly with the desolate, functional buildings and the vast, dark night.", + "seed": 4463, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_1307_seed_4463_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.024989061651285737, + "vlm_logprob": -0.024989061651285737, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.024989061651285737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06417149305343628 + ], + "opposite_logprob_tokens": [ + -3.314171552658081 + ], + "logprob_reward": -0.06417149305343628, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.06417149305343628, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17287565767765045, + 0.07116139680147171, + 0.04407472908496857, + 0.128937229514122, + 0.028553888201713562, + 0.07132577896118164, + 0.07953312993049622, + 0.11975283920764923, + 0.19274362921714783, + 0.09104177355766296 + ], + "credit_map": [ + 0.022970424965023994, + 0.015706462785601616, + 0.02367543801665306, + 0.027948496863245964, + 0.020955020561814308, + 0.017492596060037613, + 0.02419188618659973, + 0.019935328513383865, + 0.007841208949685097, + 0.00558991776779294, + 0.006899840664118528, + 0.012147976085543633, + 0.010523371398448944, + 0.007616167888045311, + 0.011363035067915916, + 0.009179878048598766, + 0.0062223169952631, + 0.006357908248901367, + 0.005014197435230017, + 0.0029997793026268482, + 0.00779466237872839, + 0.006195782218128443, + 0.0041809785179793835, + 0.0053091044537723064, + 0.027258511632680893, + 0.014659448526799679, + 0.018170066177845, + 0.012404987588524818, + 0.009561499580740929, + 0.016517559066414833, + 0.014618037268519402, + 0.015747109428048134, + 0.004015509970486164, + 0.0030313958413898945, + 0.004302837885916233, + 0.002869286108762026, + 0.0025381704326719046, + 0.005720530170947313, + 0.003307133913040161, + 0.0027690213173627853, + 0.01082481350749731, + 0.01984376460313797, + 0.008615455590188503, + 0.006594747304916382, + 0.009381284937262535, + 0.005379728507250547, + 0.00626615434885025, + 0.0044198292307555676, + 0.010192174464464188, + 0.01169504038989544, + 0.016671117395162582, + 0.007433439139276743, + 0.005277760326862335, + 0.008239173330366611, + 0.011387530714273453, + 0.008636892773211002, + 0.015508822165429592, + 0.019404394552111626, + 0.01774701662361622, + 0.013163352385163307, + 0.012625754810869694, + 0.01489532832056284, + 0.01442989893257618, + 0.011978261172771454, + 0.023604080080986023, + 0.02764813043177128, + 0.026443418115377426, + 0.028572620823979378, + 0.0176334660500288, + 0.025956980884075165, + 0.01976177841424942, + 0.02312314137816429, + 0.01506354846060276, + 0.011688766069710255, + 0.014549520798027515, + 0.010174930095672607, + 0.0074286009185016155, + 0.01635703071951866, + 0.0072995079681277275, + 0.008479873649775982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.06417149305343628 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04869474843144417 + ], + "opposite_logprob_tokens": [ + -3.0486948490142822 + ], + "logprob_reward": -0.04869474843144417, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04869474843144417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07916878163814545, + 0.0644342303276062, + 0.028172850608825684, + 0.035175371915102005, + 0.021969739347696304, + 0.048510316759347916, + 0.025145959109067917, + 0.07350651174783707, + 0.5546720027923584, + 0.06924418359994888 + ], + "credit_map": [ + 0.009525644592940807, + 0.0071706916205585, + 0.013129358179867268, + 0.009808626025915146, + 0.00887206755578518, + 0.010380400344729424, + 0.009516654536128044, + 0.010765336453914642, + 0.015762850642204285, + 0.004364600405097008, + 0.006697791628539562, + 0.007216955069452524, + 0.008825909346342087, + 0.008957494981586933, + 0.006583615206182003, + 0.006025008391588926, + 0.004558507818728685, + 0.0029841165523976088, + 0.004025833215564489, + 0.0023023742251098156, + 0.0035627244506031275, + 0.003343653865158558, + 0.003379311878234148, + 0.004016329534351826, + 0.003843513084575534, + 0.0025841796305030584, + 0.009439965710043907, + 0.0018511057132855058, + 0.002580709522590041, + 0.00967012345790863, + 0.0024119634181261063, + 0.0027938124258071184, + 0.004128602799028158, + 0.003843549871817231, + 0.0032791048288345337, + 0.00172484724316746, + 0.0027294796891510487, + 0.002681182464584708, + 0.00166070566046983, + 0.0019222671398892999, + 0.0072593046352267265, + 0.013536237180233002, + 0.0059048570692539215, + 0.003592261578887701, + 0.006814685184508562, + 0.004822113551199436, + 0.003830480622127652, + 0.00275037856772542, + 0.0034677789080888033, + 0.003571821143850684, + 0.0034378175623714924, + 0.00222117081284523, + 0.0037000065203756094, + 0.0030706143006682396, + 0.0036072342190891504, + 0.0020695144776254892, + 0.008301743306219578, + 0.008445448242127895, + 0.011651840060949326, + 0.010890700854361057, + 0.008463622070848942, + 0.00963635928928852, + 0.010352927260100842, + 0.005763866938650608, + 0.06365109980106354, + 0.08288771659135818, + 0.065919429063797, + 0.0748855248093605, + 0.050046976655721664, + 0.069636769592762, + 0.05834741145372391, + 0.0892971009016037, + 0.0063875587657094, + 0.009250165894627571, + 0.00985887460410595, + 0.009522464126348495, + 0.009357585571706295, + 0.00746138347312808, + 0.00855604000389576, + 0.008850108832120895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.04869474843144417 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002507282653823495 + ], + "opposite_logprob_tokens": [ + -6.002507209777832 + ], + "logprob_reward": -0.002507282653823495, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002507282653823495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18414029479026794, + 0.07805214822292328, + 0.0546051524579525, + 0.1333553045988083, + 0.01849331334233284, + 0.039792250841856, + 0.047575775533914566, + 0.052755992859601974, + 0.27128931879997253, + 0.11994040757417679 + ], + "credit_map": [ + 0.01797705516219139, + 0.01743575930595398, + 0.018203791230916977, + 0.035141732543706894, + 0.04170504957437515, + 0.018627235665917397, + 0.019197629764676094, + 0.015852052718400955, + 0.009727250784635544, + 0.007262768689543009, + 0.008385497145354748, + 0.007313804235309362, + 0.015719322487711906, + 0.010988760739564896, + 0.008703261613845825, + 0.009951487183570862, + 0.009591813199222088, + 0.008849142119288445, + 0.007594147697091103, + 0.005079302471131086, + 0.005796588957309723, + 0.006305480841547251, + 0.005322503391653299, + 0.006066178437322378, + 0.013669122941792011, + 0.00976046733558178, + 0.038738053292036057, + 0.006019421387463808, + 0.009821109473705292, + 0.04357516020536423, + 0.005668880883604288, + 0.006103102117776871, + 0.0030552169773727655, + 0.0020290084648877382, + 0.002704913029447198, + 0.00236207596026361, + 0.001459212158806622, + 0.0025635112542659044, + 0.0019257152453064919, + 0.002393662929534912, + 0.00537986122071743, + 0.008823035284876823, + 0.004782569594681263, + 0.0038038159254938364, + 0.006378334481269121, + 0.002890776377171278, + 0.004056895151734352, + 0.0036769702564924955, + 0.0063413470052182674, + 0.010631337761878967, + 0.011747772805392742, + 0.0037868621293455362, + 0.005170570220798254, + 0.003422045847401023, + 0.003492063144221902, + 0.0029837805777788162, + 0.0045104562304914, + 0.00453548226505518, + 0.008122149854898453, + 0.008584301918745041, + 0.005034689791500568, + 0.004839781671762466, + 0.011904981918632984, + 0.005224151536822319, + 0.02839202992618084, + 0.03713323920965195, + 0.03246545046567917, + 0.03878936171531677, + 0.023547081276774406, + 0.033782560378313065, + 0.02998684160411358, + 0.047192808240652084, + 0.010839378461241722, + 0.016070228070020676, + 0.012479443103075027, + 0.014335030689835548, + 0.016286909580230713, + 0.011765795759856701, + 0.023297453299164772, + 0.014866173267364502 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 3 + ], + "selected_reward": -0.002507282653823495 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0009255892946384847 + ], + "opposite_logprob_tokens": [ + -7.000925540924072 + ], + "logprob_reward": -0.0009255892946384847, + "yes_no_margin_reward": 7.0, + "token_credit_reward": -0.0009255892946384847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07934509217739105, + 0.07856407016515732, + 0.054615363478660583, + 0.13081446290016174, + 0.03168519213795662, + 0.16034340858459473, + 0.06124250218272209, + 0.07458855956792831, + 0.2623104453086853, + 0.06649092584848404 + ], + "credit_map": [ + 0.00998957734555006, + 0.006019153166562319, + 0.0114546874538064, + 0.009193331003189087, + 0.008762610144913197, + 0.009902955032885075, + 0.016087939962744713, + 0.007934827357530594, + 0.013985449448227882, + 0.006075737066566944, + 0.009234968572854996, + 0.009138884022831917, + 0.013695182278752327, + 0.010630960576236248, + 0.007421534508466721, + 0.008381348103284836, + 0.011921371333301067, + 0.009582686237990856, + 0.006007323507219553, + 0.004050413612276316, + 0.006379582453519106, + 0.006652249023318291, + 0.0041203913278877735, + 0.005901338532567024, + 0.01672268658876419, + 0.012165941298007965, + 0.037147294729948044, + 0.00608043372631073, + 0.00877053290605545, + 0.03747282549738884, + 0.005287449341267347, + 0.007167282048612833, + 0.005894958507269621, + 0.005690888501703739, + 0.004126254003494978, + 0.0032681720331311226, + 0.004042681772261858, + 0.003182756481692195, + 0.002770861843600869, + 0.0027086175978183746, + 0.02141757681965828, + 0.04565912485122681, + 0.018875494599342346, + 0.014604643918573856, + 0.02441299520432949, + 0.012068336829543114, + 0.01362941600382328, + 0.00967580359429121, + 0.006971438881009817, + 0.010260848328471184, + 0.009520471096038818, + 0.0053098867647349834, + 0.0069587454199790955, + 0.0062557402998209, + 0.009660438634455204, + 0.006304929964244366, + 0.010788115672767162, + 0.010884251445531845, + 0.008626908995211124, + 0.01047260407358408, + 0.006269108969718218, + 0.009291893802583218, + 0.011010156013071537, + 0.007245519198477268, + 0.022751156240701675, + 0.02996164932847023, + 0.03710325062274933, + 0.03622046113014221, + 0.020865410566329956, + 0.037906497716903687, + 0.027035657316446304, + 0.050466347485780716, + 0.005254972260445356, + 0.005861019250005484, + 0.0076195430010557175, + 0.0085446797311306, + 0.008292041718959808, + 0.0078057656064629555, + 0.014437379315495491, + 0.008675524033606052 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.0009255892946384847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008646194823086262 + ], + "opposite_logprob_tokens": [ + -4.758646011352539 + ], + "logprob_reward": -0.008646194823086262, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008646194823086262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11727933585643768, + 0.06879542768001556, + 0.03193552792072296, + 0.06832420080900192, + 0.030473235994577408, + 0.021992700174450874, + 0.020675919950008392, + 0.08089939504861832, + 0.5052558779716492, + 0.05436833202838898 + ], + "credit_map": [ + 0.013015447184443474, + 0.0072725811041891575, + 0.0143229179084301, + 0.014015664346516132, + 0.01858607493340969, + 0.01387614943087101, + 0.022543152794241905, + 0.013647357001900673, + 0.009850681759417057, + 0.0060205720365047455, + 0.007182891946285963, + 0.006591949611902237, + 0.014729565940797329, + 0.009902260266244411, + 0.007054235320538282, + 0.007463270332664251, + 0.004273707512766123, + 0.004665502347052097, + 0.005776452831923962, + 0.0030551371164619923, + 0.0026854099705815315, + 0.003957896493375301, + 0.003861408680677414, + 0.0036600143648684025, + 0.007862004451453686, + 0.009788475930690765, + 0.012518463656306267, + 0.004606776405125856, + 0.007276891730725765, + 0.013433901593089104, + 0.005960570182651281, + 0.0068771266378462315, + 0.003773423144593835, + 0.0031735452357679605, + 0.004761260934174061, + 0.0031847364734858274, + 0.003522571874782443, + 0.004682533442974091, + 0.0050236014649271965, + 0.002351564820855856, + 0.0031139617785811424, + 0.003556416369974613, + 0.003385838819667697, + 0.0021621128544211388, + 0.002245503943413496, + 0.0031258047092705965, + 0.0023975048679858446, + 0.0020055577624589205, + 0.0029218413401395082, + 0.0017824230017140508, + 0.0031895304564386606, + 0.0021635170560330153, + 0.0018397303065285087, + 0.003444395726546645, + 0.0025975964963436127, + 0.0027368871960788965, + 0.007517598103731871, + 0.009068979881703854, + 0.011283050291240215, + 0.011138956062495708, + 0.008433233015239239, + 0.013889417983591557, + 0.011758270673453808, + 0.007809893693774939, + 0.05893464386463165, + 0.07472095638513565, + 0.05918213725090027, + 0.06904777884483337, + 0.04488413408398628, + 0.06161816045641899, + 0.055528461933135986, + 0.08133965730667114, + 0.007867386564612389, + 0.006279258988797665, + 0.011684011667966843, + 0.007788780611008406, + 0.005463757552206516, + 0.004846556577831507, + 0.005333476699888706, + 0.005105109885334969 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.008646194823086262 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Beginning with a high-angle shot rendered with stark, functional realism, the camera executes a slow, steady vertical descent, moving downward through the chilly night air. Below, a vast geothermal facility is revealed, its intricate network of pipes forming a maze of industrial-grade steel, crisscrossing the dark, frozen ground. Immense steam pressure surges upward from a large valve within this network, creating a powerful, sustained eruption that forms a thick, roiling cloud that billows and expands into the sky. The scene is dramatically illuminated by a spectacular aurora, its vibrant, dancing curtain of green and purple light sweeping across the facility. This celestial glow casts an otherworldly glow, causing the metal pipes to shimmer with shifting, cold points of light and causing the turbulent steam cloud itself to glow from within, creating a luminous, ethereal substance that contrasts starkly with the desolate, functional buildings and the vast, dark night.", + "seed": 4464, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_8/index_1307_seed_4464_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.020582368155010043, + "vlm_logprob": -0.020582368155010043, + "vlm_yes_no_margin": 4.75, + "vlm_token_credit": -0.020582368155010043, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.047797009348869324 + ], + "opposite_logprob_tokens": [ + -3.4227969646453857 + ], + "logprob_reward": -0.047797009348869324, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.047797009348869324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5472513437271118, + 0.06291503459215164, + 0.06421202421188354, + 0.022604040801525116, + 0.017179390415549278, + 0.021554624661803246, + 0.05280128866434097, + 0.03753330558538437, + 0.036184970289468765, + 0.13776391744613647 + ], + "credit_map": [ + 0.050502192229032516, + 0.034932941198349, + 0.01060738880187273, + 0.04684410244226456, + 0.04534485191106796, + 0.15721626579761505, + 0.019442953169345856, + 0.18236064910888672, + 0.006844051647931337, + 0.006884304340928793, + 0.004367009270936251, + 0.011466016992926598, + 0.007791650481522083, + 0.008124228566884995, + 0.009831099770963192, + 0.007606671191751957, + 0.007569991983473301, + 0.011928503401577473, + 0.0029978409875184298, + 0.011025574989616871, + 0.006707527209073305, + 0.01239143405109644, + 0.004365865606814623, + 0.007225281558930874, + 0.0031887434888631105, + 0.0026774366851896048, + 0.0017227601492777467, + 0.0035488512367010117, + 0.0022900865878909826, + 0.0025099024642258883, + 0.003504961496219039, + 0.003161296248435974, + 0.0027978529687970877, + 0.0017673983238637447, + 0.001037247246131301, + 0.0020774982403963804, + 0.0016065523959696293, + 0.003102943766862154, + 0.002780413720756769, + 0.0020094835199415684, + 0.0028922315686941147, + 0.003612749045714736, + 0.0022223235573619604, + 0.002436863724142313, + 0.001393193262629211, + 0.003728966461494565, + 0.0028961696662008762, + 0.0023721279576420784, + 0.008410491049289703, + 0.006540900561958551, + 0.0032627175096422434, + 0.005643327720463276, + 0.004411385394632816, + 0.011286580935120583, + 0.0076222410425543785, + 0.005623647011816502, + 0.0041876621544361115, + 0.003753263968974352, + 0.003175202989950776, + 0.004998993128538132, + 0.007345560472458601, + 0.0035179262049496174, + 0.003876984817907214, + 0.006677710916846991, + 0.004291632678359747, + 0.0027262037619948387, + 0.002098996192216873, + 0.0032442326191812754, + 0.008899793960154057, + 0.004018787760287523, + 0.002740717027336359, + 0.008164606988430023, + 0.01709366776049137, + 0.01808965764939785, + 0.010497350245714188, + 0.018511826172471046, + 0.01718118041753769, + 0.012029916979372501, + 0.009787035174667835, + 0.034573279321193695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.047797009348869324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043086692690849304 + ], + "opposite_logprob_tokens": [ + -3.1680867671966553 + ], + "logprob_reward": -0.043086692690849304, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043086692690849304, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28812745213508606, + 0.16078196465969086, + 0.06744129955768585, + 0.021951302886009216, + 0.019690634682774544, + 0.08925095200538635, + 0.028455587103962898, + 0.03916577249765396, + 0.12267949432134628, + 0.1624554991722107 + ], + "credit_map": [ + 0.024943752214312553, + 0.022403478622436523, + 0.007147942669689655, + 0.025364769622683525, + 0.022819628939032555, + 0.07961626350879669, + 0.011605761013925076, + 0.09422587603330612, + 0.018725009635090828, + 0.02821994386613369, + 0.01963689737021923, + 0.032698698341846466, + 0.01634526066482067, + 0.013457718305289745, + 0.013095207512378693, + 0.01860324665904045, + 0.008197249844670296, + 0.012179800309240818, + 0.003935772459954023, + 0.009898043237626553, + 0.011365111917257309, + 0.009017799980938435, + 0.006220933049917221, + 0.006626588758081198, + 0.0032448922283947468, + 0.002686112653464079, + 0.0024663496296852827, + 0.002600782085210085, + 0.003018452087417245, + 0.003093679901212454, + 0.0022465165238827467, + 0.0025945205707103014, + 0.003266219049692154, + 0.0019281457643955946, + 0.0018584360368549824, + 0.0026374636217951775, + 0.0018054349347949028, + 0.00325056747533381, + 0.002539643319323659, + 0.0024047261103987694, + 0.013010609894990921, + 0.01478930190205574, + 0.008849258534610271, + 0.00930064544081688, + 0.00580893037840724, + 0.014516061171889305, + 0.012952491641044617, + 0.010023660957813263, + 0.00378057942725718, + 0.003793755080550909, + 0.003611944615840912, + 0.0030705498065799475, + 0.0018886596662923694, + 0.005950788035988808, + 0.003773387521505356, + 0.002585923532024026, + 0.006989669986069202, + 0.003848574822768569, + 0.003702194429934025, + 0.005375021602958441, + 0.0029900327790528536, + 0.005193994380533695, + 0.00549317104741931, + 0.005573113914579153, + 0.01520564965903759, + 0.017240693792700768, + 0.007519499398767948, + 0.010479981079697609, + 0.025263730436563492, + 0.013410219922661781, + 0.008588156662881374, + 0.024971578270196915, + 0.021340908482670784, + 0.0143871596083045, + 0.010810132138431072, + 0.0185334924608469, + 0.02352367527782917, + 0.014038356021046638, + 0.014466051012277603, + 0.04535573720932007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.043086692690849304 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103692714124918 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.004103692714124918, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103692714124918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6438092589378357, + 0.035213716328144073, + 0.022155430167913437, + 0.020002581179142, + 0.0059227957390248775, + 0.04467691853642464, + 0.028968065977096558, + 0.058749377727508545, + 0.07259754091501236, + 0.06790432333946228 + ], + "credit_map": [ + 0.06291671097278595, + 0.03888051211833954, + 0.01260711345821619, + 0.05702490359544754, + 0.05033331364393234, + 0.18296325206756592, + 0.016745582222938538, + 0.22233785688877106, + 0.004625277128070593, + 0.003914691507816315, + 0.0020045724231749773, + 0.003934519365429878, + 0.008824696764349937, + 0.00643139099702239, + 0.0019968105480074883, + 0.0034817573614418507, + 0.0032114614732563496, + 0.0031824475154280663, + 0.0014411847805604339, + 0.002697103191167116, + 0.0029460014775395393, + 0.00450093811377883, + 0.0015097904251888394, + 0.0026665038894861937, + 0.0026302889455109835, + 0.0025602628011256456, + 0.0014252698747441173, + 0.002821782138198614, + 0.002070443006232381, + 0.0030759843066334724, + 0.003389206947758794, + 0.002029343042522669, + 0.0010501380311325192, + 0.0009132467675954103, + 0.0005257793236523867, + 0.0006465428741648793, + 0.0004891832359135151, + 0.0008753740112297237, + 0.0006314858328551054, + 0.0007910458953119814, + 0.006145814433693886, + 0.006776572670787573, + 0.004224343691021204, + 0.004906900227069855, + 0.0028877397999167442, + 0.007788192015141249, + 0.007059038151055574, + 0.004888318479061127, + 0.00465404475107789, + 0.0031097042374312878, + 0.0019561343360692263, + 0.003265703096985817, + 0.00280619109980762, + 0.006544991862028837, + 0.0031880380120128393, + 0.003443256951868534, + 0.008869508281350136, + 0.004667195957154036, + 0.004836431238800287, + 0.005905479658395052, + 0.009872052818536758, + 0.00645424984395504, + 0.007723852060735226, + 0.010420609265565872, + 0.009259160608053207, + 0.006700441706925631, + 0.0028017915319651365, + 0.005432582460343838, + 0.02229236252605915, + 0.006124039646238089, + 0.00486329011619091, + 0.015123873017728329, + 0.007102377247065306, + 0.012082865461707115, + 0.005860813893377781, + 0.007446888368576765, + 0.01163784321397543, + 0.01077469252049923, + 0.005870079156011343, + 0.0071287681348621845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.004103692714124918 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0011828100541606545 + ], + "opposite_logprob_tokens": [ + -6.751183032989502 + ], + "logprob_reward": -0.0011828100541606545, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.0011828100541606545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5900816321372986, + 0.0964571163058281, + 0.030533503741025925, + 0.01939776912331581, + 0.018491949886083603, + 0.011206213384866714, + 0.018710047006607056, + 0.040531788021326065, + 0.1426684707403183, + 0.03192149102687836 + ], + "credit_map": [ + 0.05569981411099434, + 0.03218172490596771, + 0.0127581637352705, + 0.0505521297454834, + 0.04822831228375435, + 0.1702163964509964, + 0.016154974699020386, + 0.20429015159606934, + 0.013585400767624378, + 0.010908166877925396, + 0.004266588017344475, + 0.007566850632429123, + 0.02845553122460842, + 0.01770160347223282, + 0.005855701398104429, + 0.008117265067994595, + 0.0037484902422875166, + 0.004598576109856367, + 0.0022130573634058237, + 0.0052451686933636665, + 0.002367996610701084, + 0.005928605329245329, + 0.002555946819484234, + 0.003875663038343191, + 0.0029709588270634413, + 0.002722637029364705, + 0.00109362811781466, + 0.0029859256464987993, + 0.0020880165975540876, + 0.0021540008019655943, + 0.0034201201051473618, + 0.0019624829292297363, + 0.002439298201352358, + 0.0021754801273345947, + 0.0014857881469652057, + 0.0023925083223730326, + 0.00147819088306278, + 0.0035852526780217886, + 0.0028979454655200243, + 0.002037484897300601, + 0.001345584518276155, + 0.0017776781460270286, + 0.0006780379335395992, + 0.001439077197574079, + 0.0007966490229591727, + 0.002168291946873069, + 0.0019502410432323813, + 0.0010506527032703161, + 0.003122734371572733, + 0.0021734514739364386, + 0.0019405974308028817, + 0.0021294078323990107, + 0.0011343088699504733, + 0.004364719148725271, + 0.0017590950010344386, + 0.0020857315976172686, + 0.005999772809445858, + 0.0037441826425492764, + 0.003741831984370947, + 0.0033769707661122084, + 0.007741210982203484, + 0.0037484143394976854, + 0.003941291011869907, + 0.008238110691308975, + 0.0179100651293993, + 0.01363697275519371, + 0.005432098172605038, + 0.010189798660576344, + 0.04307509958744049, + 0.010937850922346115, + 0.009007754735648632, + 0.03247883915901184, + 0.003482024185359478, + 0.0066967979073524475, + 0.0033744475804269314, + 0.005538771394640207, + 0.0042450884357094765, + 0.003259451361373067, + 0.002071936149150133, + 0.003252972150221467 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0011828100541606545 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Does the video begin with a high-angle shot of a geothermal facility?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a slow, steady vertical descent downward through the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is a large valve visible emitting a powerful, sustained eruption of steam that forms a thick, roiling cloud?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by a spectacular aurora displaying vibrant green and purple light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the metal pipes shimmer with shifting light and does the steam cloud glow from within due to the aurora?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741635967046022 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741635967046022, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741635967046022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7447177767753601, + 0.0668744295835495, + 0.02442755177617073, + 0.009802768006920815, + 0.024246757850050926, + 0.04103148356080055, + 0.008730118162930012, + 0.023210957646369934, + 0.037325188517570496, + 0.01963302679359913 + ], + "credit_map": [ + 0.06722905486822128, + 0.04435443505644798, + 0.014638162218034267, + 0.0623445026576519, + 0.06336696445941925, + 0.22183643281459808, + 0.019758056849241257, + 0.25119009613990784, + 0.00994445662945509, + 0.01009625568985939, + 0.004113955423235893, + 0.008057224564254284, + 0.012522301636636257, + 0.010124373249709606, + 0.00515521876513958, + 0.006860639899969101, + 0.0033214641734957695, + 0.004156792536377907, + 0.0018260215874761343, + 0.003847144776955247, + 0.0024883418809622526, + 0.004292663186788559, + 0.0016850437968969345, + 0.002810076344758272, + 0.0012502005556598306, + 0.0012231300352141261, + 0.0006228507845662534, + 0.0011930224718526006, + 0.0008935582009144127, + 0.001321776187978685, + 0.0023036084603518248, + 0.0009946201462298632, + 0.003426019102334976, + 0.003598104929551482, + 0.0021076230332255363, + 0.0029670658987015486, + 0.002091547241434455, + 0.0038280149456113577, + 0.0030648461543023586, + 0.003163535613566637, + 0.005820878315716982, + 0.009040871635079384, + 0.004360439255833626, + 0.004269855096936226, + 0.0029978433158248663, + 0.0053055197931826115, + 0.004572588950395584, + 0.004663484171032906, + 0.0012240001233294606, + 0.0011273929849267006, + 0.0008103176369331777, + 0.0008742472273297608, + 0.0005668849335052073, + 0.0016421526670455933, + 0.0015024308813735843, + 0.0009826909517869353, + 0.0036534906830638647, + 0.002002482768148184, + 0.0017312719719484448, + 0.0021161434706300497, + 0.004453369416296482, + 0.0021169392857700586, + 0.0018031809013336897, + 0.005334075074642897, + 0.00460009416565299, + 0.004449264612048864, + 0.0020657633431255817, + 0.003148999996483326, + 0.009503137320280075, + 0.0035685929469764233, + 0.002600293606519699, + 0.007389039266854525, + 0.0021634073927998543, + 0.0030052091460675, + 0.0020556910894811153, + 0.002795818028971553, + 0.0030279543716460466, + 0.002979442710056901, + 0.0015336298383772373, + 0.002071871655061841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.006741635967046022 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_tensor([3077, 3077])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_tensor([3077, 3077])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7d4ea639ea94a10b28d0cf20003e8a2a49050f33 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_tensor([3077, 3077])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a level angle, slowly pans right, following the graceful movements of a female swordsman dancing with a sword deep in a bamboo forest. She is situated in a natural circular clearing, surrounded by dense stands of towering, emerald green bamboo. Dappled sunlight filters through the dense foliage, casting flickering spots of light on the leaf-strewn ground. The female swordsman's long, cascading black hair, held high by a simple white headband, gently brushes against her delicate face as she moves. A heroic spirit etched between her brows, her gaze focused and resolute. She is clad in a light, almost translucent, pure white, cross-collared Hanfu. As she swivels swiftly and fluidly, the fine steel sword in her hand traces visible, fluid, and swift arcs of silver light through the air, its blade reflecting the faint light of the forest. At the same time, her wide, flowing white sleeves, with each graceful turn and decisive swing of her arm, fluttered, stretched, and waved in the air like outstretched wings, complementing the sword's radiance. The entire scene is dominated by the emerald green of the bamboo forest and the pure white of the chivalrous woman. The interweaving of light and shadow creates a serene, otherworldly, yet restrained and powerful martial arts atmosphere.", + "seed": 23940, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_3077_seed_23940_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03900245181284845, + "vlm_logprob": -0.03900245181284845, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03900245181284845, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01470724307000637 + ], + "opposite_logprob_tokens": [ + -4.514707088470459 + ], + "logprob_reward": -0.01470724307000637, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.01470724307000637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2209399938583374, + 0.09683175384998322, + 0.06211107224225998, + 0.07317906618118286, + 0.1392606794834137, + 0.11535374820232391, + 0.08158125728368759, + 0.08254584670066833, + 0.07030171900987625, + 0.05789481848478317 + ], + "credit_map": [ + 0.018477438017725945, + 0.045096784830093384, + 0.039917558431625366, + 0.027723802253603935, + 0.022714976221323013, + 0.03794850409030914, + 0.014236163347959518, + 0.014824772253632545, + 0.009191440418362617, + 0.025236502289772034, + 0.01574583910405636, + 0.010644076392054558, + 0.0063503822311758995, + 0.019660329446196556, + 0.004590747877955437, + 0.00541243189945817, + 0.01088513620197773, + 0.011420254595577717, + 0.011575527489185333, + 0.005104689858853817, + 0.005593033507466316, + 0.010021095164120197, + 0.004672075156122446, + 0.0028392598032951355, + 0.006609619595110416, + 0.011311313137412071, + 0.011377660557627678, + 0.004920285195112228, + 0.009185590781271458, + 0.012516801245510578, + 0.005235063843429089, + 0.012022733688354492, + 0.02084718458354473, + 0.020792359486222267, + 0.022164253517985344, + 0.013299853540956974, + 0.01869291439652443, + 0.017545174807310104, + 0.01322365365922451, + 0.012695289216935635, + 0.020503580570220947, + 0.017003946006298065, + 0.028789792209863663, + 0.008664024993777275, + 0.008301849476993084, + 0.018561188131570816, + 0.005572744179517031, + 0.007956627756357193, + 0.010922545567154884, + 0.011009780690073967, + 0.00924273207783699, + 0.009065394289791584, + 0.008675206452608109, + 0.010649812407791615, + 0.011702433228492737, + 0.010313357226550579, + 0.01278451457619667, + 0.021220967173576355, + 0.003896584501489997, + 0.005146456882357597, + 0.009986265562474728, + 0.01654377207159996, + 0.0062354435212910175, + 0.006731844507157803, + 0.01024146843701601, + 0.012594738975167274, + 0.005101421847939491, + 0.005774202756583691, + 0.007212488912045956, + 0.012473126873373985, + 0.009303620085120201, + 0.0076006497256457806, + 0.008141906931996346, + 0.01069872546941042, + 0.005286139901727438, + 0.004140499513596296, + 0.006705370731651783, + 0.008225060068070889, + 0.007425797637552023, + 0.007271319627761841 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.01470724307000637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262688733637333 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.005262688733637333, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262688733637333, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1337069272994995, + 0.04077454283833504, + 0.02858337014913559, + 0.04103957489132881, + 0.12068893760442734, + 0.057264648377895355, + 0.05320633202791214, + 0.06741797924041748, + 0.4142187237739563, + 0.04309890791773796 + ], + "credit_map": [ + 0.009870762936770916, + 0.01626131497323513, + 0.023233598098158836, + 0.014619706198573112, + 0.026712525635957718, + 0.019403675571084023, + 0.011432808823883533, + 0.012172525748610497, + 0.0021655268501490355, + 0.004719235002994537, + 0.005257786251604557, + 0.008091076277196407, + 0.005319873336702585, + 0.0055458699353039265, + 0.005248434375971556, + 0.004426742438226938, + 0.0044157314114272594, + 0.004036803264170885, + 0.005298785399645567, + 0.0024392467457801104, + 0.00496810395270586, + 0.0034968722611665726, + 0.0026547471061348915, + 0.0012730802409350872, + 0.004390568006783724, + 0.007319125812500715, + 0.006541473791003227, + 0.0048995064571499825, + 0.004094270057976246, + 0.005343361292034388, + 0.004266985692083836, + 0.004184282384812832, + 0.019457368180155754, + 0.016592029482126236, + 0.014443106018006802, + 0.01159927062690258, + 0.01956370659172535, + 0.014301500283181667, + 0.01550216507166624, + 0.009229794144630432, + 0.006144614424556494, + 0.006999490316957235, + 0.008765706792473793, + 0.007460470777004957, + 0.008525002747774124, + 0.007553245406597853, + 0.005952273495495319, + 0.00586384441703558, + 0.007075970526784658, + 0.006994110532104969, + 0.005963636562228203, + 0.006776542868465185, + 0.004736489616334438, + 0.006978183519095182, + 0.008018589578568935, + 0.006662811152637005, + 0.007225897163152695, + 0.0091903917491436, + 0.005609976593405008, + 0.007080825977027416, + 0.01107051782310009, + 0.008143256418406963, + 0.008487952873110771, + 0.010609159246087074, + 0.06814401596784592, + 0.06377164274454117, + 0.01900920830667019, + 0.03197569400072098, + 0.042505089193582535, + 0.05732089653611183, + 0.0698951855301857, + 0.06159699335694313, + 0.004136755596846342, + 0.005563328508287668, + 0.0031423508189618587, + 0.004252895247191191, + 0.006275682244449854, + 0.005341651849448681, + 0.007015320472419262, + 0.007370923180133104 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.005262688733637333 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012540768831968307 + ], + "opposite_logprob_tokens": [ + -4.387540817260742 + ], + "logprob_reward": -0.012540768831968307, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012540768831968307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.209352508187294, + 0.1156335100531578, + 0.04298608750104904, + 0.10381288081407547, + 0.10036493837833405, + 0.1312388777732849, + 0.09614024311304092, + 0.09396905452013016, + 0.05946224555373192, + 0.047039665281772614 + ], + "credit_map": [ + 0.013939620926976204, + 0.025967849418520927, + 0.05414760857820511, + 0.02875320240855217, + 0.026959478855133057, + 0.0274723619222641, + 0.015363985672593117, + 0.01674838177859783, + 0.006802596151828766, + 0.01134887058287859, + 0.015596117824316025, + 0.012180395424365997, + 0.022853145375847816, + 0.015889378264546394, + 0.014779896475374699, + 0.01618311181664467, + 0.00540862837806344, + 0.005011081229895353, + 0.007052046712487936, + 0.003821677528321743, + 0.007766688242554665, + 0.005359557922929525, + 0.005854127462953329, + 0.0027122790925204754, + 0.008767777122557163, + 0.015351329930126667, + 0.017922010272741318, + 0.010415252298116684, + 0.009950844570994377, + 0.011605004779994488, + 0.009065337479114532, + 0.020735329017043114, + 0.020530644804239273, + 0.017074307426810265, + 0.01877957582473755, + 0.007677875924855471, + 0.012167860753834248, + 0.011295763775706291, + 0.0056849075481295586, + 0.007154005113989115, + 0.022167999297380447, + 0.01628291606903076, + 0.020616309717297554, + 0.013601275160908699, + 0.018761489540338516, + 0.018612582236528397, + 0.009411205537617207, + 0.011785101145505905, + 0.012258983217179775, + 0.01431342400610447, + 0.007292682304978371, + 0.01121463906019926, + 0.011419257149100304, + 0.012388987466692924, + 0.015402790158987045, + 0.011849476955831051, + 0.010480811819434166, + 0.012321067973971367, + 0.012094110250473022, + 0.009937109425663948, + 0.012988243252038956, + 0.010467934422194958, + 0.01103016547858715, + 0.014649614691734314, + 0.009066244587302208, + 0.013330185785889626, + 0.003688776632770896, + 0.004171370528638363, + 0.00584221538156271, + 0.009053805842995644, + 0.008388891816139221, + 0.005920755676925182, + 0.006851287558674812, + 0.009491939097642899, + 0.003543949918821454, + 0.0035219865385442972, + 0.00508304126560688, + 0.007448470685631037, + 0.006323261186480522, + 0.0047757322899997234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.012540768831968307 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002196520334109664 + ], + "opposite_logprob_tokens": [ + -6.127196311950684 + ], + "logprob_reward": -0.002196520334109664, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002196520334109664, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1727159023284912, + 0.09669076651334763, + 0.04762673005461693, + 0.11251062899827957, + 0.12471670657396317, + 0.04981391504406929, + 0.0995870903134346, + 0.18390466272830963, + 0.06535080075263977, + 0.047082770615816116 + ], + "credit_map": [ + 0.013206485658884048, + 0.020116319879889488, + 0.042358335107564926, + 0.025524547323584557, + 0.02352876402437687, + 0.02366730011999607, + 0.01169331930577755, + 0.012620825320482254, + 0.005600991193205118, + 0.010385268367826939, + 0.011185934767127037, + 0.011276179924607277, + 0.0196146909147501, + 0.01486354973167181, + 0.01300668716430664, + 0.010757459327578545, + 0.004439996555447578, + 0.005841562524437904, + 0.007291437592357397, + 0.004668110981583595, + 0.007832867093384266, + 0.007028619758784771, + 0.007578276097774506, + 0.002945862477645278, + 0.00970938429236412, + 0.015081397257745266, + 0.017758507281541824, + 0.008529544807970524, + 0.015545866452157497, + 0.014366624876856804, + 0.009961828589439392, + 0.02155747078359127, + 0.0226641483604908, + 0.018926424905657768, + 0.01779850199818611, + 0.011160256341099739, + 0.01739073544740677, + 0.015977919101715088, + 0.012143111787736416, + 0.008655614219605923, + 0.007811431773006916, + 0.006032358389347792, + 0.004079362377524376, + 0.005369999911636114, + 0.009646134451031685, + 0.007482854183763266, + 0.003353889798745513, + 0.006037883926182985, + 0.012549794279038906, + 0.015184430405497551, + 0.010577076114714146, + 0.009650892578065395, + 0.013465089723467827, + 0.013110105879604816, + 0.013239415362477303, + 0.011810284107923508, + 0.023614060133695602, + 0.03056398592889309, + 0.014525268226861954, + 0.023500381037592888, + 0.025525731965899467, + 0.018357548862695694, + 0.022345446050167084, + 0.0254722461104393, + 0.009367798455059528, + 0.01415176223963499, + 0.004297662526369095, + 0.004594938363879919, + 0.006190519314259291, + 0.00966611783951521, + 0.009944996796548367, + 0.007137006614357233, + 0.005761654116213322, + 0.010291325859725475, + 0.0028328574262559414, + 0.004262187983840704, + 0.006524465978145599, + 0.0072239115834236145, + 0.004689385648816824, + 0.005496982950717211 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.002196520334109664 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16030503809452057 + ], + "opposite_logprob_tokens": [ + -1.9103050231933594 + ], + "logprob_reward": -0.16030503809452057, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16030503809452057, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2614469528198242, + 0.082339346408844, + 0.10564297437667847, + 0.08993276208639145, + 0.08360478281974792, + 0.06254609674215317, + 0.07116576284170151, + 0.1321948617696762, + 0.06827159970998764, + 0.04285483807325363 + ], + "credit_map": [ + 0.028546378016471863, + 0.029310496523976326, + 0.04618499055504799, + 0.03149142116308212, + 0.032454513013362885, + 0.035014670342206955, + 0.026934998109936714, + 0.03150947019457817, + 0.005843647755682468, + 0.01382326427847147, + 0.012125348672270775, + 0.011865186505019665, + 0.013080374337732792, + 0.011510004289448261, + 0.00766637921333313, + 0.006425137165933847, + 0.011080220341682434, + 0.012504847720265388, + 0.01916995272040367, + 0.00897700060158968, + 0.016394183039665222, + 0.014104834757745266, + 0.016310401260852814, + 0.007101533003151417, + 0.007808815222233534, + 0.016094518825411797, + 0.012728107161819935, + 0.009474298916757107, + 0.010364877060055733, + 0.011650297790765762, + 0.008859576657414436, + 0.012952280230820179, + 0.011102375574409962, + 0.012335424311459064, + 0.01083421427756548, + 0.007167994510382414, + 0.009211252443492413, + 0.01586139388382435, + 0.006258438341319561, + 0.01083369180560112, + 0.008648507297039032, + 0.010716364718973637, + 0.008683952502906322, + 0.005549024324864149, + 0.007725704461336136, + 0.008618163876235485, + 0.004636488854885101, + 0.007967889308929443, + 0.00978807546198368, + 0.0073643336072564125, + 0.009294532239437103, + 0.009294898249208927, + 0.007078840397298336, + 0.008622339926660061, + 0.010791061446070671, + 0.008931676857173443, + 0.01740046963095665, + 0.01927063800394535, + 0.010467329993844032, + 0.01246704999357462, + 0.020634755492210388, + 0.017928708344697952, + 0.014842537231743336, + 0.01918337494134903, + 0.012522779405117035, + 0.010890552774071693, + 0.00366270006634295, + 0.00603119982406497, + 0.00954605545848608, + 0.006714954506605864, + 0.009422428905963898, + 0.00948092620819807, + 0.005394275300204754, + 0.006457548122853041, + 0.00428866408765316, + 0.0031288480386137962, + 0.006313882768154144, + 0.005905910860747099, + 0.005481647793203592, + 0.005884059704840183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.16030503809452057 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans right, following the graceful movements of a female swordsman dancing with a sword deep in a bamboo forest. She is situated in a natural circular clearing, surrounded by dense stands of towering, emerald green bamboo. Dappled sunlight filters through the dense foliage, casting flickering spots of light on the leaf-strewn ground. The female swordsman's long, cascading black hair, held high by a simple white headband, gently brushes against her delicate face as she moves. A heroic spirit etched between her brows, her gaze focused and resolute. She is clad in a light, almost translucent, pure white, cross-collared Hanfu. As she swivels swiftly and fluidly, the fine steel sword in her hand traces visible, fluid, and swift arcs of silver light through the air, its blade reflecting the faint light of the forest. At the same time, her wide, flowing white sleeves, with each graceful turn and decisive swing of her arm, fluttered, stretched, and waved in the air like outstretched wings, complementing the sword's radiance. The entire scene is dominated by the emerald green of the bamboo forest and the pure white of the chivalrous woman. The interweaving of light and shadow creates a serene, otherworldly, yet restrained and powerful martial arts atmosphere.", + "seed": 23941, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_3077_seed_23941_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.01630460317246616, + "vlm_logprob": -0.01630460317246616, + "vlm_yes_no_margin": 4.8, + "vlm_token_credit": -0.01630460317246616, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009352798573672771 + ], + "opposite_logprob_tokens": [ + -5.259352684020996 + ], + "logprob_reward": -0.009352798573672771, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.009352798573672771, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06259165704250336, + 0.04045682027935982, + 0.018175872042775154, + 0.053438007831573486, + 0.04040297865867615, + 0.04212183132767677, + 0.028315303847193718, + 0.0508890375494957, + 0.07649332284927368, + 0.5871152281761169 + ], + "credit_map": [ + 0.006460601929575205, + 0.00799079891294241, + 0.006996005307883024, + 0.009046426974236965, + 0.007787202950567007, + 0.008789448998868465, + 0.005841294303536415, + 0.009679869748651981, + 0.005449599586427212, + 0.007453691214323044, + 0.007266470231115818, + 0.0026854798197746277, + 0.0035369449760764837, + 0.006392580922693014, + 0.002692404668778181, + 0.004979647230356932, + 0.0026067157741636038, + 0.0029118950478732586, + 0.002981745870783925, + 0.0011091773631051183, + 0.0019836777355521917, + 0.0037186003755778074, + 0.00136276101693511, + 0.00150129827670753, + 0.00789251271635294, + 0.003263564081862569, + 0.013030368834733963, + 0.00434962147846818, + 0.004270416218787432, + 0.005338698625564575, + 0.005773789249360561, + 0.009519031271338463, + 0.0057446518912911415, + 0.004086502827703953, + 0.003218340454623103, + 0.0030696780886501074, + 0.00703028729185462, + 0.008157090283930302, + 0.0045181442983448505, + 0.004578280728310347, + 0.006766162812709808, + 0.0030277578625828028, + 0.0034021716564893723, + 0.0023469782900065184, + 0.007318432908505201, + 0.011629038490355015, + 0.002380581106990576, + 0.005250703077763319, + 0.0041219498962163925, + 0.004531350452452898, + 0.002518740249797702, + 0.002256275387480855, + 0.0037208052817732096, + 0.005725592840462923, + 0.002223977819085121, + 0.0032166081946343184, + 0.007723324466496706, + 0.0024214822333306074, + 0.0025319268461316824, + 0.00281463167630136, + 0.009379778057336807, + 0.014308649115264416, + 0.0027266829274594784, + 0.008982556872069836, + 0.011843358166515827, + 0.004459945950657129, + 0.014722762629389763, + 0.0035206738393753767, + 0.012498350813984871, + 0.012046830728650093, + 0.004918486811220646, + 0.012482907623052597, + 0.046942975372076035, + 0.009871909394860268, + 0.22937573492527008, + 0.017313122749328613, + 0.23051728308200836, + 0.033773306757211685, + 0.007881914265453815, + 0.011438906192779541 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.009352798573672771 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004647050518542528 + ], + "opposite_logprob_tokens": [ + -5.379647254943848 + ], + "logprob_reward": -0.004647050518542528, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004647050518542528, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042316269129514694, + 0.02329491265118122, + 0.02243611589074135, + 0.01828751526772976, + 0.050349365919828415, + 0.017417266964912415, + 0.022867783904075623, + 0.026341654360294342, + 0.014669954776763916, + 0.7620192170143127 + ], + "credit_map": [ + 0.004112667869776487, + 0.0052454653196036816, + 0.004290533717721701, + 0.005735568702220917, + 0.007468692027032375, + 0.006080602761358023, + 0.004795899149030447, + 0.004586841445416212, + 0.0029204501770436764, + 0.0033938433043658733, + 0.0026171994395554066, + 0.0029035808984190226, + 0.0035004306118935347, + 0.0029293736442923546, + 0.0029517090879380703, + 0.002078324556350708, + 0.0026325611397624016, + 0.0032790438272058964, + 0.003141104942187667, + 0.0023726164363324642, + 0.005271365400403738, + 0.0018255430040881038, + 0.002150071319192648, + 0.0017638111021369696, + 0.002579665509983897, + 0.001228775130584836, + 0.003815843490883708, + 0.0020609593484550714, + 0.0017515080980956554, + 0.0019201519899070263, + 0.002331054536625743, + 0.0025995569303631783, + 0.007076404523104429, + 0.004434050060808659, + 0.004103867337107658, + 0.003289258573204279, + 0.00987762026488781, + 0.010632832534611225, + 0.005006453022360802, + 0.005928877741098404, + 0.0025396563578397036, + 0.0020069661550223827, + 0.0013581798411905766, + 0.001516450778581202, + 0.0032129681203514338, + 0.00435164337977767, + 0.001017607282847166, + 0.0014137965627014637, + 0.0031751818023622036, + 0.003292272100225091, + 0.002213659230619669, + 0.0021495288237929344, + 0.003315329784527421, + 0.004687417298555374, + 0.001359251793473959, + 0.002675142837688327, + 0.0038505124393850565, + 0.002099769888445735, + 0.0017853440949693322, + 0.0016177593497559428, + 0.005176087841391563, + 0.006600634660571814, + 0.001510641654022038, + 0.003700906177982688, + 0.0020931155886501074, + 0.0020063770934939384, + 0.0021343252155929804, + 0.0012182160280644894, + 0.0020057326182723045, + 0.0018521008314564824, + 0.0019253218779340386, + 0.0014347669202834368, + 0.06116791069507599, + 0.012273883447051048, + 0.29996275901794434, + 0.022787194699048996, + 0.3005450963973999, + 0.04408877342939377, + 0.010209627449512482, + 0.010983946733176708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.004647050518542528 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780817665159702 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.009780817665159702, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780817665159702, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11552304774522781, + 0.08229735493659973, + 0.07451623678207397, + 0.04252338036894798, + 0.34040579199790955, + 0.043437156826257706, + 0.052249372005462646, + 0.1022842526435852, + 0.07965053617954254, + 0.06711288541555405 + ], + "credit_map": [ + 0.01041125226765871, + 0.01940108835697174, + 0.01140550710260868, + 0.01890328899025917, + 0.014264854602515697, + 0.016999555751681328, + 0.010237317532300949, + 0.013900190591812134, + 0.008430887013673782, + 0.013985848054289818, + 0.015506476163864136, + 0.007193227764219046, + 0.009553234092891216, + 0.011614855378866196, + 0.005270305089652538, + 0.010742513462901115, + 0.009804558008909225, + 0.017398344352841377, + 0.013012235052883625, + 0.006618636194616556, + 0.006682593375444412, + 0.005134247709065676, + 0.006707134190946817, + 0.009158489294350147, + 0.00772131560370326, + 0.004512067884206772, + 0.007355114445090294, + 0.0039023030549287796, + 0.0046738749369978905, + 0.005195778328925371, + 0.006355249788612127, + 0.0028076726011931896, + 0.05048826336860657, + 0.02959410287439823, + 0.028437897562980652, + 0.025359613820910454, + 0.04882698506116867, + 0.07549947500228882, + 0.044725626707077026, + 0.03747382014989853, + 0.0066717336885631084, + 0.004214787855744362, + 0.004031553864479065, + 0.005709645804017782, + 0.006771867163479328, + 0.006336101330816746, + 0.004782161209732294, + 0.004919310100376606, + 0.005953628104180098, + 0.009325768798589706, + 0.004376319237053394, + 0.005654463078826666, + 0.007704620715230703, + 0.009217053651809692, + 0.00294799474067986, + 0.007069528102874756, + 0.01297490019351244, + 0.007722902577370405, + 0.009494256228208542, + 0.007768737617880106, + 0.023075595498085022, + 0.022872194647789, + 0.006800178438425064, + 0.011575493961572647, + 0.01017867960035801, + 0.011052863672375679, + 0.0072676315903663635, + 0.006171199027448893, + 0.011938286013901234, + 0.010339831002056599, + 0.009145320393145084, + 0.01355671975761652, + 0.00882711261510849, + 0.009007321670651436, + 0.011933904141187668, + 0.004010547883808613, + 0.013629446737468243, + 0.004699543118476868, + 0.007646725978702307, + 0.007358281873166561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.009780817665159702 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002815808402374387 + ], + "opposite_logprob_tokens": [ + -5.8778157234191895 + ], + "logprob_reward": -0.002815808402374387, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002815808402374387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06353416293859482, + 0.029926471412181854, + 0.025435373187065125, + 0.01757138967514038, + 0.11944667994976044, + 0.023212259635329247, + 0.03616219386458397, + 0.040827907621860504, + 0.034176427870988846, + 0.6097071766853333 + ], + "credit_map": [ + 0.005676054861396551, + 0.010340072214603424, + 0.00762137258425355, + 0.009259865619242191, + 0.007701063994318247, + 0.008840203285217285, + 0.006736726965755224, + 0.007358801085501909, + 0.002397073432803154, + 0.004789835773408413, + 0.004366706591099501, + 0.0035731568932533264, + 0.005008190404623747, + 0.003633656306192279, + 0.00352812884375453, + 0.0026297220028936863, + 0.003047812031581998, + 0.007082604803144932, + 0.004898448009043932, + 0.002031056210398674, + 0.00223619700409472, + 0.0016738937702029943, + 0.0020432414021342993, + 0.0024221192579716444, + 0.0020344387739896774, + 0.0010117345955222845, + 0.0031515266746282578, + 0.0020323237404227257, + 0.0018915771506726742, + 0.00222536688670516, + 0.0024141399189829826, + 0.002810281002894044, + 0.021186811849474907, + 0.006983176805078983, + 0.014131564646959305, + 0.008435129188001156, + 0.008451873436570168, + 0.031096268445253372, + 0.01113090943545103, + 0.018030934035778046, + 0.00345036992803216, + 0.0022619005758315325, + 0.0020585735328495502, + 0.00225889403373003, + 0.004035258665680885, + 0.0035546068102121353, + 0.0023535829968750477, + 0.003239073557779193, + 0.004105525556951761, + 0.0076305074617266655, + 0.0028678535018116236, + 0.0034085577353835106, + 0.005544919520616531, + 0.0058144996874034405, + 0.0023106916341930628, + 0.00447963597252965, + 0.006482379510998726, + 0.0028614948969334364, + 0.0031708572059869766, + 0.0029965632129460573, + 0.006216268986463547, + 0.010253190994262695, + 0.0024974134285002947, + 0.0063497330993413925, + 0.004941993393003941, + 0.00432499498128891, + 0.002855097409337759, + 0.0028523963410407305, + 0.004894567653536797, + 0.004587496165186167, + 0.004379086196422577, + 0.005340793170034885, + 0.04883373901247978, + 0.011825889348983765, + 0.23870308697223663, + 0.01812669448554516, + 0.24135927855968475, + 0.03467581048607826, + 0.008404672145843506, + 0.007777998223900795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.002815808402374387 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054926540702581406 + ], + "opposite_logprob_tokens": [ + -2.929926633834839 + ], + "logprob_reward": -0.054926540702581406, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054926540702581406, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0537789948284626, + 0.033334750682115555, + 0.01302025094628334, + 0.03778721019625664, + 0.013519746251404285, + 0.010147771798074245, + 0.015715651214122772, + 0.01370125263929367, + 0.05788629129528999, + 0.7511081099510193 + ], + "credit_map": [ + 0.00612462917342782, + 0.008455757051706314, + 0.0051533738151192665, + 0.007400128059089184, + 0.007034262642264366, + 0.007382148411124945, + 0.005999920889735222, + 0.006228778511285782, + 0.004372544586658478, + 0.004510087892413139, + 0.002542869420722127, + 0.003922669216990471, + 0.006188493687659502, + 0.0038426206447184086, + 0.00470033148303628, + 0.0032551309559494257, + 0.0018075669649988413, + 0.0015747580910101533, + 0.0019796318374574184, + 0.0013449041871353984, + 0.0023225927725434303, + 0.0010866953525692225, + 0.001473536598496139, + 0.0014305649092420936, + 0.004681058693677187, + 0.001949772471562028, + 0.006998949218541384, + 0.004352267365902662, + 0.003600775497034192, + 0.0037760050036013126, + 0.006762588396668434, + 0.0056657916866242886, + 0.002178120194002986, + 0.0013256792444735765, + 0.0013837917940691113, + 0.001373432227410376, + 0.0011032568290829659, + 0.002266475697979331, + 0.0017400920623913407, + 0.002148898085579276, + 0.0010721632279455662, + 0.0015238336054608226, + 0.0008285696385428309, + 0.00167121144477278, + 0.0016443772474303842, + 0.0018602906493470073, + 0.0006869033095426857, + 0.0008604222093708813, + 0.0016821338795125484, + 0.0025318411644548178, + 0.0016603978583589196, + 0.0018151620170101523, + 0.002114637987688184, + 0.002281651133671403, + 0.0014231755631044507, + 0.0022066510282456875, + 0.0019450082909315825, + 0.0013932253932580352, + 0.0011026313295587897, + 0.0009134708670899272, + 0.0020778500474989414, + 0.0032030614092946053, + 0.0009709791629575193, + 0.002095026895403862, + 0.007343859411776066, + 0.009393635205924511, + 0.005720025859773159, + 0.005512337200343609, + 0.007219723891466856, + 0.005967750679701567, + 0.008040486834943295, + 0.008688472211360931, + 0.06012836843729019, + 0.012530519627034664, + 0.2952720820903778, + 0.022409312427043915, + 0.29757675528526306, + 0.043458327651023865, + 0.009841504506766796, + 0.009891250170767307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.054926540702581406 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans right, following the graceful movements of a female swordsman dancing with a sword deep in a bamboo forest. She is situated in a natural circular clearing, surrounded by dense stands of towering, emerald green bamboo. Dappled sunlight filters through the dense foliage, casting flickering spots of light on the leaf-strewn ground. The female swordsman's long, cascading black hair, held high by a simple white headband, gently brushes against her delicate face as she moves. A heroic spirit etched between her brows, her gaze focused and resolute. She is clad in a light, almost translucent, pure white, cross-collared Hanfu. As she swivels swiftly and fluidly, the fine steel sword in her hand traces visible, fluid, and swift arcs of silver light through the air, its blade reflecting the faint light of the forest. At the same time, her wide, flowing white sleeves, with each graceful turn and decisive swing of her arm, fluttered, stretched, and waved in the air like outstretched wings, complementing the sword's radiance. The entire scene is dominated by the emerald green of the bamboo forest and the pure white of the chivalrous woman. The interweaving of light and shadow creates a serene, otherworldly, yet restrained and powerful martial arts atmosphere.", + "seed": 23942, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_3077_seed_23942_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.03668346479535103, + "vlm_logprob": -0.03668346479535103, + "vlm_yes_no_margin": 4.775, + "vlm_token_credit": -0.03668346479535103, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008197234943509102 + ], + "opposite_logprob_tokens": [ + -5.50819730758667 + ], + "logprob_reward": -0.008197234943509102, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.008197234943509102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13678057491779327, + 0.15276606380939484, + 0.07959482818841934, + 0.03575407713651657, + 0.09263288229703903, + 0.17062066495418549, + 0.07224933058023453, + 0.07420657575130463, + 0.13026413321495056, + 0.05513090640306473 + ], + "credit_map": [ + 0.017701230943202972, + 0.022221848368644714, + 0.023131106048822403, + 0.006962134037166834, + 0.017178580164909363, + 0.023923292756080627, + 0.00988476537168026, + 0.01577761024236679, + 0.01774180307984352, + 0.03648628666996956, + 0.01714569516479969, + 0.02044743485748768, + 0.020311100408434868, + 0.016426414251327515, + 0.011648338288068771, + 0.01255897805094719, + 0.01181007456034422, + 0.011266915127635002, + 0.010044415481388569, + 0.01071953121572733, + 0.008721052668988705, + 0.008413505740463734, + 0.009780929423868656, + 0.008838390931487083, + 0.005034102126955986, + 0.0038284205365926027, + 0.0059901331551373005, + 0.0033921501599252224, + 0.004271592944860458, + 0.003968488425016403, + 0.004820335190743208, + 0.004448853433132172, + 0.01634489931166172, + 0.006907128728926182, + 0.014121809974312782, + 0.003762201638892293, + 0.01205520797520876, + 0.014544131234288216, + 0.006204260513186455, + 0.018693244084715843, + 0.043072409927845, + 0.013763219118118286, + 0.014002951793372631, + 0.008224018849432468, + 0.036258574575185776, + 0.029188789427280426, + 0.010336996987462044, + 0.015773681923747063, + 0.010730489157140255, + 0.005388778634369373, + 0.009232217445969582, + 0.00433765072375536, + 0.016400059685111046, + 0.012318084947764874, + 0.004544280469417572, + 0.00929777231067419, + 0.008753789588809013, + 0.0093544851988554, + 0.010034347884356976, + 0.006485365331172943, + 0.011013262905180454, + 0.011956969276070595, + 0.0061784652061760426, + 0.010429887101054192, + 0.0174579881131649, + 0.010707980021834373, + 0.014677979052066803, + 0.009156563319265842, + 0.0343523807823658, + 0.020903583616018295, + 0.00992907676845789, + 0.013078571297228336, + 0.005786806810647249, + 0.004375977907329798, + 0.00464785797521472, + 0.0037787659093737602, + 0.015593228861689568, + 0.010818907991051674, + 0.0032680928707122803, + 0.006861267611384392 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.008197234943509102 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032024311367422342 + ], + "opposite_logprob_tokens": [ + -5.753202438354492 + ], + "logprob_reward": -0.0032024311367422342, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032024311367422342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11099103093147278, + 0.11117035150527954, + 0.08536531031131744, + 0.049838121980428696, + 0.1026998832821846, + 0.20429286360740662, + 0.061724159866571426, + 0.15354518592357635, + 0.0646064355969429, + 0.055766716599464417 + ], + "credit_map": [ + 0.01548038236796856, + 0.015870671719312668, + 0.017148775979876518, + 0.00899705570191145, + 0.017059696838259697, + 0.017336420714855194, + 0.009571526199579239, + 0.009526488371193409, + 0.012904530391097069, + 0.03048151172697544, + 0.012634864076972008, + 0.00757433520630002, + 0.024073995649814606, + 0.010163209401071072, + 0.005863121710717678, + 0.007474774960428476, + 0.018068306148052216, + 0.00819474458694458, + 0.013673638924956322, + 0.005401212722063065, + 0.008200989104807377, + 0.017908459529280663, + 0.005664369557052851, + 0.008253579959273338, + 0.005036236252635717, + 0.0052083879709243774, + 0.00797470472753048, + 0.005787082016468048, + 0.005713567603379488, + 0.0070496187545359135, + 0.007279869634658098, + 0.005788652691990137, + 0.024008935317397118, + 0.008444990031421185, + 0.012647702358663082, + 0.006480718497186899, + 0.011642134748399258, + 0.015607314184308052, + 0.0069908034056425095, + 0.016877276822924614, + 0.05162886530160904, + 0.01632748916745186, + 0.01823519356548786, + 0.010507102124392986, + 0.038459114730358124, + 0.03440200537443161, + 0.01654246263206005, + 0.018190618604421616, + 0.011194796301424503, + 0.006888578645884991, + 0.00848970003426075, + 0.004724094178527594, + 0.008410765789449215, + 0.009732404723763466, + 0.005142709240317345, + 0.007141101639717817, + 0.017155807465314865, + 0.019869813695549965, + 0.024806877598166466, + 0.014343422837555408, + 0.019920917227864265, + 0.02593761309981346, + 0.013464903458952904, + 0.018045814707875252, + 0.011861301958560944, + 0.007641180884093046, + 0.008750817738473415, + 0.0062956977635622025, + 0.008275636471807957, + 0.008440976031124592, + 0.006437703035771847, + 0.0069031137973070145, + 0.006150178611278534, + 0.003921029157936573, + 0.005323787685483694, + 0.0038332354743033648, + 0.015481420792639256, + 0.010841231793165207, + 0.0033860388211905956, + 0.006829786580055952 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 1 + ], + "selected_reward": -0.0032024311367422342 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009780462831258774 + ], + "opposite_logprob_tokens": [ + -4.634780406951904 + ], + "logprob_reward": -0.009780462831258774, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009780462831258774, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10582707822322845, + 0.15876799821853638, + 0.26261773705482483, + 0.10825061053037643, + 0.09050890058279037, + 0.07000632584095001, + 0.046849340200424194, + 0.09724684804677963, + 0.024950355291366577, + 0.03497478738427162 + ], + "credit_map": [ + 0.015170537866652012, + 0.012197267264127731, + 0.021288514137268066, + 0.007413238752633333, + 0.011665014550089836, + 0.01871076226234436, + 0.00850694254040718, + 0.010874796658754349, + 0.022509600967168808, + 0.04630981385707855, + 0.017931923270225525, + 0.011893865652382374, + 0.029324138537049294, + 0.01417300384491682, + 0.006657411810010672, + 0.009968237951397896, + 0.02800803817808628, + 0.037475455552339554, + 0.03389723598957062, + 0.03668138384819031, + 0.022021330893039703, + 0.03833276033401489, + 0.035095639526844025, + 0.03110591135919094, + 0.011788396164774895, + 0.009584981016814709, + 0.01688879355788231, + 0.01019514910876751, + 0.011567804962396622, + 0.017685357481241226, + 0.015622451901435852, + 0.014917674474418163, + 0.022249501198530197, + 0.007467097602784634, + 0.010673973709344864, + 0.005766452755779028, + 0.011304832994937897, + 0.0140198590233922, + 0.007027632091194391, + 0.011999549344182014, + 0.009719233959913254, + 0.007426409516483545, + 0.005850058514624834, + 0.005559828132390976, + 0.010424727573990822, + 0.009785737842321396, + 0.013261270709335804, + 0.007979062385857105, + 0.008227159269154072, + 0.0035311742685735226, + 0.00636750552803278, + 0.002697731601074338, + 0.008027801290154457, + 0.007701106369495392, + 0.003058711299672723, + 0.007238153368234634, + 0.011418746784329414, + 0.011536149308085442, + 0.014178804121911526, + 0.008709241636097431, + 0.014902947470545769, + 0.016038617119193077, + 0.008688898757100105, + 0.01177343912422657, + 0.002714587142691016, + 0.0033312691375613213, + 0.0033529698848724365, + 0.002508939476683736, + 0.004179221112281084, + 0.0038627053145319223, + 0.0019607520662248135, + 0.003039912786334753, + 0.00478342454880476, + 0.00301571236923337, + 0.0034484376665204763, + 0.00283258780837059, + 0.007355748675763607, + 0.006165012717247009, + 0.002447286620736122, + 0.004926576744765043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.009780462831258774 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019382515456527472 + ], + "opposite_logprob_tokens": [ + -6.251938343048096 + ], + "logprob_reward": -0.0019382515456527472, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019382515456527472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10175411403179169, + 0.11270914226770401, + 0.23669438064098358, + 0.08260628581047058, + 0.09419746696949005, + 0.10966262966394424, + 0.049734365195035934, + 0.16494576632976532, + 0.019351810216903687, + 0.028344040736556053 + ], + "credit_map": [ + 0.011676475405693054, + 0.016028666868805885, + 0.014933965168893337, + 0.00946310255676508, + 0.015789180994033813, + 0.014492989517748356, + 0.010162686929106712, + 0.009207041002810001, + 0.020303627476096153, + 0.021054720506072044, + 0.016071857884526253, + 0.009200314991176128, + 0.012860183604061604, + 0.015604743734002113, + 0.008129716850817204, + 0.009483983740210533, + 0.025489456951618195, + 0.03074471279978752, + 0.03045083023607731, + 0.03360191360116005, + 0.020329469814896584, + 0.034754715859889984, + 0.03141939640045166, + 0.029903870075941086, + 0.00901846680790186, + 0.008600739762187004, + 0.011542055755853653, + 0.00901286955922842, + 0.010759050026535988, + 0.012143189087510109, + 0.012292040511965752, + 0.009237876161932945, + 0.0217896681278944, + 0.008048699237406254, + 0.011858937330543995, + 0.006003329064697027, + 0.010250844061374664, + 0.01425156556069851, + 0.006457317620515823, + 0.015537104569375515, + 0.02825409173965454, + 0.01029068510979414, + 0.009269203059375286, + 0.00536770885810256, + 0.020526694133877754, + 0.018489737063646317, + 0.008018518798053265, + 0.009445996023714542, + 0.009816614910960197, + 0.0042807795107364655, + 0.0063013313338160515, + 0.0025329608470201492, + 0.011056850664317608, + 0.00795113667845726, + 0.0030291492585092783, + 0.004765540827065706, + 0.021695327013731003, + 0.015832336619496346, + 0.03861963376402855, + 0.01376779843121767, + 0.016061309725046158, + 0.02673647180199623, + 0.01089618168771267, + 0.02133670449256897, + 0.0025505125522613525, + 0.0025246315635740757, + 0.002338563557714224, + 0.001629528822377324, + 0.003122578840702772, + 0.0031342722941190004, + 0.0016728013288229704, + 0.002378921490162611, + 0.003026747377589345, + 0.0036747215781360865, + 0.004376361146569252, + 0.0018676939653232694, + 0.005236229859292507, + 0.004728628788143396, + 0.0020733592100441456, + 0.0033603005576878786 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 7, + 1 + ], + "selected_reward": -0.0019382515456527472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16029894351959229 + ], + "opposite_logprob_tokens": [ + -1.9102989435195923 + ], + "logprob_reward": -0.16029894351959229, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16029894351959229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053714320063591, + 0.17772658169269562, + 0.259001225233078, + 0.07693638652563095, + 0.06881221383810043, + 0.25135356187820435, + 0.043308671563863754, + 0.023074917495250702, + 0.024195892736315727, + 0.021876171231269836 + ], + "credit_map": [ + 0.007388809695839882, + 0.007694307714700699, + 0.006444230210036039, + 0.005730991251766682, + 0.0065598757937550545, + 0.009059420786798, + 0.005631742998957634, + 0.005204946268349886, + 0.028609121218323708, + 0.048215921968221664, + 0.023293977603316307, + 0.015351925045251846, + 0.02361372858285904, + 0.018231399357318878, + 0.008905947208404541, + 0.011504573747515678, + 0.028087658807635307, + 0.033336929976940155, + 0.03354799374938011, + 0.0354112833738327, + 0.021771185100078583, + 0.0402425080537796, + 0.033256229013204575, + 0.03334745764732361, + 0.007097700145095587, + 0.0058867004700005054, + 0.01183058600872755, + 0.008393573574721813, + 0.008563079871237278, + 0.011616546660661697, + 0.013362567871809006, + 0.010185640305280685, + 0.01552355196326971, + 0.004857736639678478, + 0.008576332591474056, + 0.004036450758576393, + 0.007444353308528662, + 0.010097336024045944, + 0.005686651449650526, + 0.012589803896844387, + 0.06625885516405106, + 0.02001153863966465, + 0.02268427610397339, + 0.010462494567036629, + 0.05229946970939636, + 0.044142987579107285, + 0.012739473953843117, + 0.022754494100809097, + 0.0063045150600373745, + 0.003602195531129837, + 0.006242025643587112, + 0.003920023329555988, + 0.0044006966054439545, + 0.008184981532394886, + 0.003509090980514884, + 0.007145147770643234, + 0.0028128086123615503, + 0.0021422330755740404, + 0.0041087293066084385, + 0.0020554731599986553, + 0.003618731861934066, + 0.0033738475758582354, + 0.0018565519712865353, + 0.003106541931629181, + 0.0037332954816520214, + 0.002060116035863757, + 0.003811146365478635, + 0.0019233982311561704, + 0.004231254104524851, + 0.0037317476235330105, + 0.001988777657970786, + 0.0027161575853824615, + 0.0041800811886787415, + 0.0018991321558132768, + 0.002981322817504406, + 0.0010637902887538075, + 0.0038054296746850014, + 0.003636863548308611, + 0.0010665771551430225, + 0.0032429767306894064 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.16029894351959229 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a level angle, slowly pans right, following the graceful movements of a female swordsman dancing with a sword deep in a bamboo forest. She is situated in a natural circular clearing, surrounded by dense stands of towering, emerald green bamboo. Dappled sunlight filters through the dense foliage, casting flickering spots of light on the leaf-strewn ground. The female swordsman's long, cascading black hair, held high by a simple white headband, gently brushes against her delicate face as she moves. A heroic spirit etched between her brows, her gaze focused and resolute. She is clad in a light, almost translucent, pure white, cross-collared Hanfu. As she swivels swiftly and fluidly, the fine steel sword in her hand traces visible, fluid, and swift arcs of silver light through the air, its blade reflecting the faint light of the forest. At the same time, her wide, flowing white sleeves, with each graceful turn and decisive swing of her arm, fluttered, stretched, and waved in the air like outstretched wings, complementing the sword's radiance. The entire scene is dominated by the emerald green of the bamboo forest and the pure white of the chivalrous woman. The interweaving of light and shadow creates a serene, otherworldly, yet restrained and powerful martial arts atmosphere.", + "seed": 23943, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000094/group_9/index_3077_seed_23943_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.018345325021073222, + "vlm_logprob": -0.018345325021073222, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.018345325021073222, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009354096837341785 + ], + "opposite_logprob_tokens": [ + -5.509354114532471 + ], + "logprob_reward": -0.009354096837341785, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.009354096837341785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23771026730537415, + 0.08543925732374191, + 0.07326628267765045, + 0.10917310416698456, + 0.02954242378473282, + 0.09216556698083878, + 0.1006573811173439, + 0.05977112427353859, + 0.040540724992752075, + 0.17173388600349426 + ], + "credit_map": [ + 0.028336908668279648, + 0.015729503706097603, + 0.05501307547092438, + 0.01821182295680046, + 0.011868190951645374, + 0.04694899171590805, + 0.011897235177457333, + 0.049704525619745255, + 0.013721613213419914, + 0.016439709812402725, + 0.011475474573671818, + 0.008423250168561935, + 0.008548830635845661, + 0.010619666427373886, + 0.00668644392862916, + 0.009524262510240078, + 0.007179536856710911, + 0.009201818145811558, + 0.012564324773848057, + 0.018258435651659966, + 0.009150184690952301, + 0.005905176512897015, + 0.0055452450178563595, + 0.005461555905640125, + 0.013195441104471684, + 0.026573259383440018, + 0.01660049706697464, + 0.019020209088921547, + 0.008444231934845448, + 0.015218516811728477, + 0.00499730883166194, + 0.005123625509440899, + 0.003296605311334133, + 0.0034523375798016787, + 0.004362971987575293, + 0.0032823665533214808, + 0.005179317202419043, + 0.0032468356657773256, + 0.0025873640552163124, + 0.004134622402489185, + 0.018732955679297447, + 0.021168669685721397, + 0.010187072679400444, + 0.005877833813428879, + 0.012940572574734688, + 0.013287600129842758, + 0.004789487458765507, + 0.0051813675090670586, + 0.009857726283371449, + 0.009138637222349644, + 0.014660040847957134, + 0.013708374463021755, + 0.013889681547880173, + 0.01396976225078106, + 0.010450066067278385, + 0.014983079396188259, + 0.004468242637813091, + 0.0066417814232409, + 0.009439260698854923, + 0.004891660995781422, + 0.008476495742797852, + 0.01223793439567089, + 0.00705529423430562, + 0.006560449488461018, + 0.004309090785682201, + 0.005182913038879633, + 0.004944093991070986, + 0.00365970260463655, + 0.004743118304759264, + 0.004677915014326572, + 0.008168460801243782, + 0.0048554278910160065, + 0.04050242528319359, + 0.025775467976927757, + 0.025419507175683975, + 0.013397607021033764, + 0.01123986765742302, + 0.022453615441918373, + 0.010370633564889431, + 0.022574737668037415 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.009354096837341785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028269824106246233 + ], + "opposite_logprob_tokens": [ + -5.877827167510986 + ], + "logprob_reward": -0.0028269824106246233, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028269824106246233, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22556203603744507, + 0.1886831670999527, + 0.09062608331441879, + 0.05189137160778046, + 0.05432203412055969, + 0.05683131143450737, + 0.06627308577299118, + 0.08041837066411972, + 0.06265266984701157, + 0.12273981422185898 + ], + "credit_map": [ + 0.02306613326072693, + 0.01003407221287489, + 0.027680471539497375, + 0.01672709360718727, + 0.02079232968389988, + 0.0470627062022686, + 0.019252941012382507, + 0.06094631925225258, + 0.01582525111734867, + 0.017611224204301834, + 0.02703707106411457, + 0.04785805195569992, + 0.019530514255166054, + 0.019595367833971977, + 0.01769251935184002, + 0.02353319525718689, + 0.008782444521784782, + 0.008763917721807957, + 0.01419670507311821, + 0.010492950677871704, + 0.01486614253371954, + 0.009498538449406624, + 0.010415604338049889, + 0.013609785586595535, + 0.00620204908773303, + 0.010346964001655579, + 0.007566077169030905, + 0.008035258390009403, + 0.0056002442725002766, + 0.006270804442465305, + 0.004195268731564283, + 0.0036747094709426165, + 0.008250756189227104, + 0.00794615875929594, + 0.007842393592000008, + 0.004768368788063526, + 0.005882455501705408, + 0.006271037273108959, + 0.004717316012829542, + 0.008643555454909801, + 0.0057212552055716515, + 0.006180437747389078, + 0.007062749471515417, + 0.005383880343288183, + 0.008856943808495998, + 0.009815706871449947, + 0.004111540038138628, + 0.009698797017335892, + 0.0058403038419783115, + 0.007139245513826609, + 0.009076003916561604, + 0.010783538222312927, + 0.012189292348921299, + 0.007451155688613653, + 0.007090659812092781, + 0.006702891085296869, + 0.005901971831917763, + 0.006037823855876923, + 0.014769642613828182, + 0.0070695835165679455, + 0.008934597484767437, + 0.02057645097374916, + 0.008773261681199074, + 0.008355040103197098, + 0.00597753981128335, + 0.00580581184476614, + 0.005039970390498638, + 0.005488438997417688, + 0.008338438346982002, + 0.007336985785514116, + 0.01325281523168087, + 0.01141267828643322, + 0.025519665330648422, + 0.014311280101537704, + 0.01643984578549862, + 0.008869466371834278, + 0.011866724118590355, + 0.01653372123837471, + 0.007435915991663933, + 0.021763207390904427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0028269824106246233 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00762575538828969 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.00762575538828969, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.00762575538828969, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1593172550201416, + 0.26465293765068054, + 0.044282056391239166, + 0.05758253112435341, + 0.02959235943853855, + 0.04927228018641472, + 0.09724096953868866, + 0.13285529613494873, + 0.05295157805085182, + 0.11225276440382004 + ], + "credit_map": [ + 0.021165747195482254, + 0.01217915304005146, + 0.03722899779677391, + 0.013091614469885826, + 0.010884018614888191, + 0.02510100230574608, + 0.01332875806838274, + 0.02633794955909252, + 0.02800939604640007, + 0.04408895969390869, + 0.0413329154253006, + 0.04944591224193573, + 0.02491694688796997, + 0.02771821618080139, + 0.019664781168103218, + 0.02947578951716423, + 0.00399761414155364, + 0.00432143360376358, + 0.007162987254559994, + 0.005052491556853056, + 0.00574623653665185, + 0.007987499237060547, + 0.0038594056386500597, + 0.00615439098328352, + 0.006371248047798872, + 0.011516314931213856, + 0.007885039784014225, + 0.0076876976527273655, + 0.0054965512827038765, + 0.007502053864300251, + 0.0058462838642299175, + 0.005277344956994057, + 0.00412743678316474, + 0.003652039682492614, + 0.004327865783125162, + 0.0029887855052948, + 0.003242762526497245, + 0.003548041684553027, + 0.0027359637897461653, + 0.004969466477632523, + 0.008638695813715458, + 0.006606067065149546, + 0.005928792525082827, + 0.0028882708866149187, + 0.0055493321269750595, + 0.007206040900200605, + 0.003661257680505514, + 0.00879382248967886, + 0.010336280800402164, + 0.0077150361612439156, + 0.0162807609885931, + 0.011797744780778885, + 0.012815840542316437, + 0.015044823288917542, + 0.007873968221247196, + 0.015376511961221695, + 0.008223461918532848, + 0.011720627546310425, + 0.0236930213868618, + 0.009713943116366863, + 0.020136255770921707, + 0.030723869800567627, + 0.015797453001141548, + 0.01284666545689106, + 0.006858717184513807, + 0.007303648628294468, + 0.006009573116898537, + 0.0044481451623141766, + 0.008042732253670692, + 0.00620876532047987, + 0.005640116985887289, + 0.008439880795776844, + 0.024673426523804665, + 0.013412869535386562, + 0.01611681468784809, + 0.010204880498349667, + 0.008758367970585823, + 0.014636605978012085, + 0.007226640824228525, + 0.017223164439201355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.00762575538828969 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019376566633582115 + ], + "opposite_logprob_tokens": [ + -6.2519378662109375 + ], + "logprob_reward": -0.0019376566633582115, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019376566633582115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07155027985572815, + 0.07120784372091293, + 0.0661417692899704, + 0.054823122918605804, + 0.06342995166778564, + 0.09667567908763885, + 0.06980540603399277, + 0.2666182816028595, + 0.05337478592991829, + 0.18637286126613617 + ], + "credit_map": [ + 0.011500222608447075, + 0.004509858321398497, + 0.010791584849357605, + 0.005337010603398085, + 0.0063484166748821735, + 0.011342955753207207, + 0.0051223523914813995, + 0.01659788005053997, + 0.006337830331176519, + 0.008986839093267918, + 0.010648272931575775, + 0.01415692176669836, + 0.0070212651044130325, + 0.007771892938762903, + 0.0067424350418150425, + 0.009542388841509819, + 0.006097231525927782, + 0.0058384123258292675, + 0.010924667119979858, + 0.007731282617896795, + 0.012517351657152176, + 0.006489638704806566, + 0.007159174885600805, + 0.009384012781083584, + 0.005622021853923798, + 0.01011989451944828, + 0.0068272133357822895, + 0.006663480307906866, + 0.006927001755684614, + 0.006606580223888159, + 0.006323081906884909, + 0.0057338522747159, + 0.009225139394402504, + 0.009396355599164963, + 0.008036265149712563, + 0.0065274350345134735, + 0.007980707101523876, + 0.0068146525882184505, + 0.00626544700935483, + 0.00918395072221756, + 0.023307764902710915, + 0.020073017105460167, + 0.009046281687915325, + 0.003810135880485177, + 0.010523115284740925, + 0.013041437603533268, + 0.006032049655914307, + 0.010841879062354565, + 0.006766350008547306, + 0.007064887788146734, + 0.012486162595450878, + 0.009490866214036942, + 0.010064070113003254, + 0.0096500925719738, + 0.006576082669198513, + 0.007706888020038605, + 0.013754128478467464, + 0.015448149293661118, + 0.04941719397902489, + 0.021446676924824715, + 0.02210388332605362, + 0.0726519450545311, + 0.03909575939178467, + 0.032700538635253906, + 0.005840641912072897, + 0.004782921634614468, + 0.0066605969332158566, + 0.004351472947746515, + 0.006922772619873285, + 0.007376211695373058, + 0.008375583216547966, + 0.009064585901796818, + 0.036984290927648544, + 0.023608336225152016, + 0.02543029747903347, + 0.022286787629127502, + 0.02024013176560402, + 0.02358986623585224, + 0.015083547681570053, + 0.019149592146277428 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 5 + ], + "selected_reward": -0.0019376566633582115 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a natural circular clearing surrounded by dense stands of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female swordsman wearing a light, translucent, pure white, cross-collared Hanfu with long black hair held by a white headband?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the female swordsman performing graceful, swift, and fluid martial arts movements with a steel sword that traces arcs of silver light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the female swordsman's wide, flowing white sleeves flutter and wave in the air like outstretched wings during her movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot from a level angle in a medium shot that slowly pans right to follow the subject?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0699821338057518 + ], + "opposite_logprob_tokens": [ + -2.6949820518493652 + ], + "logprob_reward": -0.0699821338057518, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0699821338057518, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15754707157611847, + 0.07504092901945114, + 0.18737448751926422, + 0.12629546225070953, + 0.05410172790288925, + 0.1508355438709259, + 0.11032330989837646, + 0.053008224815130234, + 0.047248415648937225, + 0.0382249541580677 + ], + "credit_map": [ + 0.03051764704287052, + 0.00964822992682457, + 0.026314569637179375, + 0.015812715515494347, + 0.017395315691828728, + 0.02027837745845318, + 0.012469910085201263, + 0.025110304355621338, + 0.006902559660375118, + 0.013244329951703548, + 0.014459256082773209, + 0.011534336023032665, + 0.007699070498347282, + 0.0075920927338302135, + 0.007008090615272522, + 0.006601191591471434, + 0.015908870846033096, + 0.016271008178591728, + 0.02642677165567875, + 0.01720268279314041, + 0.04321833327412605, + 0.025580186396837234, + 0.01660071313381195, + 0.02616591937839985, + 0.008723067119717598, + 0.009465709328651428, + 0.017323488369584084, + 0.018324922770261765, + 0.02226416766643524, + 0.012210329994559288, + 0.019866935908794403, + 0.018116841092705727, + 0.008587258867919445, + 0.0075155142694711685, + 0.007479094434529543, + 0.004630592651665211, + 0.006796507630497217, + 0.006292411591857672, + 0.00456681614741683, + 0.008233532309532166, + 0.03469724580645561, + 0.03301829844713211, + 0.014595857821404934, + 0.007646927610039711, + 0.01753602735698223, + 0.019720112904906273, + 0.0095956576988101, + 0.014025406911969185, + 0.00936032272875309, + 0.01452780794352293, + 0.014157358556985855, + 0.008335163816809654, + 0.01829606667160988, + 0.01786663942039013, + 0.01072527002543211, + 0.01705467700958252, + 0.005136540625244379, + 0.0028667377773672342, + 0.013172384351491928, + 0.0027066939510405064, + 0.005454318132251501, + 0.01261847373098135, + 0.0042474460788071156, + 0.006805627606809139, + 0.003938173409551382, + 0.006286237854510546, + 0.003565495600923896, + 0.004616141319274902, + 0.0044678025878965855, + 0.005578887183219194, + 0.011850904673337936, + 0.006944774184376001, + 0.0031517937313765287, + 0.0032778275199234486, + 0.00575824361294508, + 0.003117404645308852, + 0.007996861822903156, + 0.006879383232444525, + 0.003458216320723295, + 0.004585218615829945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.0699821338057518 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_tensor([2391, 2391])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_tensor([2391, 2391])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..33f460aba3ed379c491c6a0b754f1e7f49ae1687 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_tensor([2391, 2391])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, pulling-back tracking shot follows a modern sedan as it steadily advances through the pitch-black night. The car's sleek metallic black paint reflects the faint ambient light from the humid air. Its glaring white LED headlights cast two long, water-distorted beams across the deserted two-lane asphalt. The car drives straight ahead, where a torrential downpour pours down like a cascade, forming a nearly opaque wall of water. The raindrops are so dense that they create a dynamic, grainy texture in the headlights. As the car plunges unhesitatingly into the rain, countless water droplets are instantly illuminated by the headlights, creating a chaotic, shimmering cloud of particles. Then, a massive spray of water explodes upward and to the sides from the front of the car, instantly engulfing the windshield and rendering the frantic thrashing of the wipers futile. The entire scene is shrouded in cold tones, with only the bright light of the car lights contrasting sharply with the endless darkness, creating a strong dramatic tension and tense atmosphere of being isolated from the world and resolutely rushing towards unknown challenges.", + "seed": 20199, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_2391_seed_20199_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.10150414342060685, + "vlm_logprob": -0.10150414342060685, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.10150414342060685, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12130668759346008 + ], + "opposite_logprob_tokens": [ + -2.2463066577911377 + ], + "logprob_reward": -0.12130668759346008, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12130668759346008, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07501168549060822, + 0.029527908191084862, + 0.017701121047139168, + 0.03561278432607651, + 0.07217276841402054, + 0.27420544624328613, + 0.2279725819826126, + 0.11703883856534958, + 0.05687851831316948, + 0.0938783660531044 + ], + "credit_map": [ + 0.0071708024479448795, + 0.008502092212438583, + 0.016892526298761368, + 0.008350961841642857, + 0.008251365274190903, + 0.008968973532319069, + 0.007959052920341492, + 0.00891591515392065, + 0.0029137791134417057, + 0.0032336097210645676, + 0.006132207345217466, + 0.004422412719577551, + 0.0027904382441192865, + 0.0036061436403542757, + 0.0026217573322355747, + 0.0038075570482760668, + 0.0018040998838841915, + 0.0014753846917301416, + 0.0031074373982846737, + 0.002022065222263336, + 0.002021541353315115, + 0.002438595285639167, + 0.0025169826112687588, + 0.0023150132037699223, + 0.003142748726531863, + 0.0028439119923859835, + 0.004078215453773737, + 0.0025961054489016533, + 0.005661100149154663, + 0.005537430290132761, + 0.006142789963632822, + 0.005610479973256588, + 0.01021910086274147, + 0.004274028353393078, + 0.009213619865477085, + 0.004950832575559616, + 0.0087658129632473, + 0.013048825785517693, + 0.00626659719273448, + 0.01543395034968853, + 0.03418794646859169, + 0.024888725951313972, + 0.02439827471971512, + 0.025313086807727814, + 0.04694129154086113, + 0.06065308675169945, + 0.02119932882487774, + 0.03662367910146713, + 0.020327163860201836, + 0.027334963902831078, + 0.029778601601719856, + 0.028989017009735107, + 0.026909371837973595, + 0.03465977683663368, + 0.02596784010529518, + 0.03400583565235138, + 0.011949102394282818, + 0.01448281854391098, + 0.018000544980168343, + 0.011998829431831837, + 0.008867439813911915, + 0.015326093882322311, + 0.02155991643667221, + 0.014854084700345993, + 0.007624393329024315, + 0.012737042270600796, + 0.009936180897057056, + 0.006429661065340042, + 0.00416063005104661, + 0.0034768078476190567, + 0.00633692741394043, + 0.006176872178912163, + 0.010596039704978466, + 0.018088897690176964, + 0.010764564387500286, + 0.015432223677635193, + 0.008019518107175827, + 0.012583764269948006, + 0.008928310126066208, + 0.009465038776397705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 7 + ], + "selected_reward": -0.12130668759346008 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07006850093603134 + ], + "opposite_logprob_tokens": [ + -2.695068597793579 + ], + "logprob_reward": -0.07006850093603134, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07006850093603134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32670214772224426, + 0.13520845770835876, + 0.095998115837574, + 0.1685667783021927, + 0.038508590310811996, + 0.047643329948186874, + 0.02520401030778885, + 0.032701004296541214, + 0.06713833659887314, + 0.06232921779155731 + ], + "credit_map": [ + 0.03882833197712898, + 0.024987362325191498, + 0.057502519339323044, + 0.10275837033987045, + 0.021111024543642998, + 0.024293547496199608, + 0.035380225628614426, + 0.02184073068201542, + 0.018412996083498, + 0.009372488595545292, + 0.028452955186367035, + 0.037134893238544464, + 0.007408329285681248, + 0.0120009770616889, + 0.012787683866918087, + 0.009638136252760887, + 0.01039206888526678, + 0.008600755594670773, + 0.018073854967951775, + 0.00785085093230009, + 0.010688050650060177, + 0.014679493382573128, + 0.007901208475232124, + 0.01781182736158371, + 0.01752135530114174, + 0.013841215521097183, + 0.02464710921049118, + 0.016197172924876213, + 0.016443809494376183, + 0.030156679451465607, + 0.018098978325724602, + 0.031660471111536026, + 0.004751583561301231, + 0.0024849080946296453, + 0.004317702725529671, + 0.0037602230440825224, + 0.0035166062880307436, + 0.009303577244281769, + 0.0033293042797595263, + 0.007044682744890451, + 0.007002048194408417, + 0.004183642566204071, + 0.0037669679149985313, + 0.003573126159608364, + 0.005484390538185835, + 0.013649232685565948, + 0.002540444489568472, + 0.0074434722773730755, + 0.0025367317721247673, + 0.002661469392478466, + 0.0023770041298121214, + 0.002654368756338954, + 0.00450163334608078, + 0.003600784344598651, + 0.002692936221137643, + 0.004179081879556179, + 0.0028952250722795725, + 0.004817886743694544, + 0.0025382530875504017, + 0.002852936275303364, + 0.002447641221806407, + 0.0028254901990294456, + 0.009896859526634216, + 0.0044267079792916775, + 0.00653800368309021, + 0.015744607895612717, + 0.008871704339981079, + 0.005939403548836708, + 0.006358543876558542, + 0.006405544001609087, + 0.009719937108457088, + 0.007560588885098696, + 0.007428422104567289, + 0.006195572670549154, + 0.007775362581014633, + 0.012304224073886871, + 0.004524058196693659, + 0.012613505125045776, + 0.005562230013310909, + 0.005925843026489019 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.07006850093603134 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559206187725067 + ], + "opposite_logprob_tokens": [ + -1.600592017173767 + ], + "logprob_reward": -0.22559206187725067, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559206187725067, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2347571849822998, + 0.0813065618276596, + 0.08473768085241318, + 0.08455538004636765, + 0.11534982919692993, + 0.1265917271375656, + 0.06523136049509048, + 0.057917993515729904, + 0.04370911419391632, + 0.10584314167499542 + ], + "credit_map": [ + 0.0294131338596344, + 0.02017308957874775, + 0.05637793615460396, + 0.04818376153707504, + 0.010053365491330624, + 0.026901694014668465, + 0.0171340461820364, + 0.026520147919654846, + 0.009294089861214161, + 0.00627939822152257, + 0.02172461524605751, + 0.014373627491295338, + 0.005835168529301882, + 0.008663074113428593, + 0.008081603795289993, + 0.007054989226162434, + 0.010144627653062344, + 0.008090920746326447, + 0.015033558011054993, + 0.00694451155140996, + 0.007216500584036112, + 0.013026670552790165, + 0.0074197095818817616, + 0.01686118170619011, + 0.010048291645944118, + 0.005921549629420042, + 0.014104131609201431, + 0.010164034552872181, + 0.005520014092326164, + 0.013504927977919579, + 0.009233496151864529, + 0.016058936715126038, + 0.01091222558170557, + 0.008334644138813019, + 0.016033492982387543, + 0.01279485784471035, + 0.010531137697398663, + 0.020054170861840248, + 0.015578662976622581, + 0.02111063338816166, + 0.017778223380446434, + 0.010487648658454418, + 0.009800180792808533, + 0.010686848312616348, + 0.016868602484464645, + 0.0321640782058239, + 0.007887888699769974, + 0.02091825008392334, + 0.009343326091766357, + 0.0057968636974692345, + 0.007526973262429237, + 0.00834421906620264, + 0.004488520789891481, + 0.011479631066322327, + 0.005608870182186365, + 0.012642959132790565, + 0.005916970316320658, + 0.007269614841789007, + 0.006269198842346668, + 0.005660475697368383, + 0.00518557196483016, + 0.007898133248090744, + 0.011876532807946205, + 0.007841500453650951, + 0.004916390869766474, + 0.00572751322761178, + 0.007187939714640379, + 0.004738366696983576, + 0.005498417187482119, + 0.0036894057411700487, + 0.006294448394328356, + 0.0056566293351352215, + 0.011559988372027874, + 0.013146936893463135, + 0.018538188189268112, + 0.01820356212556362, + 0.011299865320324898, + 0.011370431631803513, + 0.010860571637749672, + 0.010863601230084896 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.22559206187725067 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015166220255196095 + ], + "opposite_logprob_tokens": [ + -6.501516819000244 + ], + "logprob_reward": -0.0015166220255196095, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015166220255196095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10548362135887146, + 0.17952942848205566, + 0.05483027920126915, + 0.034874964505434036, + 0.09207160025835037, + 0.10575801879167557, + 0.2660824656486511, + 0.04265660420060158, + 0.04003939777612686, + 0.07867363095283508 + ], + "credit_map": [ + 0.011022455058991909, + 0.00929947104305029, + 0.02518383599817753, + 0.013621922582387924, + 0.00974934920668602, + 0.010806032456457615, + 0.008933854289352894, + 0.01686670258641243, + 0.023735646158456802, + 0.012222486548125744, + 0.03357980400323868, + 0.052237629890441895, + 0.010258457623422146, + 0.017314521595835686, + 0.017764324322342873, + 0.012416550889611244, + 0.00457373820245266, + 0.005174966994673014, + 0.01133748609572649, + 0.004928627051413059, + 0.0054504224099218845, + 0.008691021241247654, + 0.006263113114982843, + 0.00841089989989996, + 0.0029251729138195515, + 0.003688219003379345, + 0.005052517633885145, + 0.002404842060059309, + 0.004401734564453363, + 0.006155985873192549, + 0.0035542496480047703, + 0.006692240480333567, + 0.011775003746151924, + 0.005397069733589888, + 0.011762621812522411, + 0.009766663424670696, + 0.007477763574570417, + 0.019087763503193855, + 0.008080312982201576, + 0.01872440241277218, + 0.012319535948336124, + 0.007752109784632921, + 0.010060991160571575, + 0.009536637924611568, + 0.015671536326408386, + 0.026176422834396362, + 0.008336635306477547, + 0.0159041415899992, + 0.026697559282183647, + 0.031656038016080856, + 0.03863418102264404, + 0.039221201092004776, + 0.028751421719789505, + 0.03560485690832138, + 0.03165672346949577, + 0.03386050462722778, + 0.0038271965458989143, + 0.004775373265147209, + 0.007676179986447096, + 0.0028505909722298384, + 0.002482644049450755, + 0.003532702336087823, + 0.012701495550572872, + 0.004810419399291277, + 0.003988826181739569, + 0.005636061541736126, + 0.005627800710499287, + 0.00531697366386652, + 0.004187326412647963, + 0.0043495954014360905, + 0.007449088152498007, + 0.0034837264101952314, + 0.009812446311116219, + 0.013000404462218285, + 0.010243255645036697, + 0.013151773251593113, + 0.008533255197107792, + 0.007422673981636763, + 0.0101886922493577, + 0.006321127526462078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 5 + ], + "selected_reward": -0.0015166220255196095 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08903684467077255 + ], + "opposite_logprob_tokens": [ + -2.4640369415283203 + ], + "logprob_reward": -0.08903684467077255, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08903684467077255, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3072813153266907, + 0.08347564190626144, + 0.028256051242351532, + 0.039682745933532715, + 0.15287354588508606, + 0.06693272292613983, + 0.20565353333950043, + 0.022295650094747543, + 0.04797343909740448, + 0.04557536914944649 + ], + "credit_map": [ + 0.033530302345752716, + 0.023655986413359642, + 0.05617636442184448, + 0.0964365005493164, + 0.019918227568268776, + 0.022494103759527206, + 0.033883024007081985, + 0.021186813712120056, + 0.009738174267113209, + 0.005919856019318104, + 0.015487654134631157, + 0.02353338897228241, + 0.005865284241735935, + 0.008239378221333027, + 0.00838925875723362, + 0.00630264263600111, + 0.0035198775585740805, + 0.0030399865936487913, + 0.00401692371815443, + 0.0023069775197654963, + 0.002942495746538043, + 0.0037058289162814617, + 0.003289117245003581, + 0.0054348427802324295, + 0.004423429723829031, + 0.0034019018057733774, + 0.0041809044778347015, + 0.0028475977014750242, + 0.0060138944536447525, + 0.006750144995748997, + 0.005085499957203865, + 0.006979372818022966, + 0.016199445351958275, + 0.00898631289601326, + 0.02136560156941414, + 0.017425796017050743, + 0.01402595266699791, + 0.024219131097197533, + 0.017131080850958824, + 0.033520229160785675, + 0.006884787231683731, + 0.00521710654720664, + 0.006594174541532993, + 0.00563469622284174, + 0.0107593247666955, + 0.011597252450883389, + 0.005338737741112709, + 0.014906642958521843, + 0.019143467769026756, + 0.026209168136119843, + 0.030219582840800285, + 0.033414632081985474, + 0.022439925000071526, + 0.026628194376826286, + 0.023052269592881203, + 0.024546289816498756, + 0.0023217706475406885, + 0.0022918512113392353, + 0.003812319366261363, + 0.0027583155315369368, + 0.0013467336539179087, + 0.002574679907411337, + 0.005233454518020153, + 0.001956523396074772, + 0.006370242685079575, + 0.013311883434653282, + 0.007535379379987717, + 0.0033833247143775225, + 0.003432827303186059, + 0.002971807261928916, + 0.005610701162368059, + 0.005357272457331419, + 0.004751607310026884, + 0.006339399144053459, + 0.007230364717543125, + 0.007659731898456812, + 0.0042088571935892105, + 0.007004737854003906, + 0.00440472224727273, + 0.0039759487845003605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.08903684467077255 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, pulling-back tracking shot follows a modern sedan as it steadily advances through the pitch-black night. The car's sleek metallic black paint reflects the faint ambient light from the humid air. Its glaring white LED headlights cast two long, water-distorted beams across the deserted two-lane asphalt. The car drives straight ahead, where a torrential downpour pours down like a cascade, forming a nearly opaque wall of water. The raindrops are so dense that they create a dynamic, grainy texture in the headlights. As the car plunges unhesitatingly into the rain, countless water droplets are instantly illuminated by the headlights, creating a chaotic, shimmering cloud of particles. Then, a massive spray of water explodes upward and to the sides from the front of the car, instantly engulfing the windshield and rendering the frantic thrashing of the wipers futile. The entire scene is shrouded in cold tones, with only the bright light of the car lights contrasting sharply with the endless darkness, creating a strong dramatic tension and tense atmosphere of being isolated from the world and resolutely rushing towards unknown challenges.", + "seed": 20200, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_2391_seed_20200_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.050057011959142986, + "vlm_logprob": -0.050057011959142986, + "vlm_yes_no_margin": 3.825, + "vlm_token_credit": -0.050057011959142986, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.028553439304232597 + ], + "opposite_logprob_tokens": [ + -3.7785534858703613 + ], + "logprob_reward": -0.028553439304232597, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.028553439304232597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12115449458360672, + 0.02195669896900654, + 0.031995728611946106, + 0.028498278930783272, + 0.06606924533843994, + 0.01958511210978031, + 0.12315163016319275, + 0.16285395622253418, + 0.24153712391853333, + 0.1831977665424347 + ], + "credit_map": [ + 0.011002328246831894, + 0.030207505449652672, + 0.011975022964179516, + 0.009468085132539272, + 0.017964232712984085, + 0.02253301441669464, + 0.007953435182571411, + 0.010050863958895206, + 0.0017080431571230292, + 0.004855731502175331, + 0.002543144393712282, + 0.002417123643681407, + 0.00326780928298831, + 0.0032524182461202145, + 0.0022052964195609093, + 0.0017071312759071589, + 0.002950118388980627, + 0.004106500186026096, + 0.003641669172793627, + 0.0034789463970810175, + 0.006236438639461994, + 0.0045485105365514755, + 0.0034999048803001642, + 0.0035336387809365988, + 0.0031177878845483065, + 0.005854692775756121, + 0.0024871379137039185, + 0.002670504152774811, + 0.004398631863296032, + 0.0048060668632388115, + 0.0026757814921438694, + 0.002487672260031104, + 0.006079722661525011, + 0.006125848740339279, + 0.007913895882666111, + 0.008805746212601662, + 0.008046976290643215, + 0.009004470892250538, + 0.009789090603590012, + 0.010303493589162827, + 0.0020304154604673386, + 0.002019074745476246, + 0.002416382310912013, + 0.00333599210716784, + 0.0029950460884720087, + 0.0028703531716018915, + 0.0018208064138889313, + 0.0020970406476408243, + 0.012163168750703335, + 0.00905651319772005, + 0.022050971165299416, + 0.013561717234551907, + 0.013776697218418121, + 0.013600836507976055, + 0.021270237863063812, + 0.017671480774879456, + 0.015535799786448479, + 0.014005630277097225, + 0.03334415331482887, + 0.02046012505888939, + 0.01924172416329384, + 0.01603403314948082, + 0.027635673061013222, + 0.016596822068095207, + 0.032953884452581406, + 0.025601452216506004, + 0.026424294337630272, + 0.029722144827246666, + 0.037769466638565063, + 0.03581654280424118, + 0.029991505667567253, + 0.023257823660969734, + 0.017322860658168793, + 0.011640499345958233, + 0.014523443765938282, + 0.025285892188549042, + 0.019870202988386154, + 0.017515091225504875, + 0.04692067578434944, + 0.030119098722934723 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.028553439304232597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08907785266637802 + ], + "opposite_logprob_tokens": [ + -2.464077949523926 + ], + "logprob_reward": -0.08907785266637802, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08907785266637802, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10622570663690567, + 0.05433826521039009, + 0.04123680666089058, + 0.04028642550110817, + 0.054368533194065094, + 0.013506465591490269, + 0.14771029353141785, + 0.17273792624473572, + 0.03681908920407295, + 0.33277055621147156 + ], + "credit_map": [ + 0.009185045026242733, + 0.01850670389831066, + 0.014854107983410358, + 0.011900766752660275, + 0.013977128081023693, + 0.01529020443558693, + 0.00907210260629654, + 0.013439644128084183, + 0.005258182063698769, + 0.009846535511314869, + 0.007620649412274361, + 0.0062240310944616795, + 0.005632552783936262, + 0.007280593737959862, + 0.005368324927985668, + 0.007107400335371494, + 0.004215361550450325, + 0.009263555519282818, + 0.0032745732460170984, + 0.003123957198113203, + 0.007991835474967957, + 0.005835572257637978, + 0.004073478281497955, + 0.0034584752283990383, + 0.005018122959882021, + 0.007779691834002733, + 0.003516971366479993, + 0.002193949418142438, + 0.009622114710509777, + 0.005868069361895323, + 0.003466811729595065, + 0.002820695051923394, + 0.005198822356760502, + 0.005765276495367289, + 0.005956943146884441, + 0.008133872412145138, + 0.005660507827997208, + 0.007479368709027767, + 0.007766164839267731, + 0.008407575078308582, + 0.0018242041114717722, + 0.0012177302269265056, + 0.001356329768896103, + 0.001645333948545158, + 0.0024486619513481855, + 0.002185949357226491, + 0.0013365385821089149, + 0.0014917183434590697, + 0.015874581411480904, + 0.009212363511323929, + 0.03842047601938248, + 0.010706055909395218, + 0.01509630773216486, + 0.01763790100812912, + 0.016059337183833122, + 0.02470327354967594, + 0.018444573506712914, + 0.016887445002794266, + 0.029249152168631554, + 0.01996472477912903, + 0.024520736187696457, + 0.022288890555500984, + 0.021733971312642097, + 0.01964843086898327, + 0.005214126780629158, + 0.0049089668318629265, + 0.003419993445277214, + 0.0035227106418460608, + 0.005910246167331934, + 0.007126593496650457, + 0.003537197830155492, + 0.003179253777489066, + 0.028678767383098602, + 0.018865693360567093, + 0.030448932200670242, + 0.04661398008465767, + 0.03362170606851578, + 0.03182154893875122, + 0.0884169340133667, + 0.054302994161844254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 6 + ], + "selected_reward": -0.08907785266637802 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04306888207793236 + ], + "opposite_logprob_tokens": [ + -3.1680688858032227 + ], + "logprob_reward": -0.04306888207793236, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04306888207793236, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17080584168434143, + 0.11634868383407593, + 0.07169809192419052, + 0.07452190667390823, + 0.12530826032161713, + 0.05082768201828003, + 0.07867606729269028, + 0.11222223192453384, + 0.0811513289809227, + 0.1184399276971817 + ], + "credit_map": [ + 0.012858809903264046, + 0.030932482331991196, + 0.02181217260658741, + 0.026117216795682907, + 0.017841096967458725, + 0.024228980764746666, + 0.020604601129889488, + 0.016410471871495247, + 0.010440056212246418, + 0.025577092543244362, + 0.014928492717444897, + 0.011550032533705235, + 0.015461464412510395, + 0.019437018781900406, + 0.007714500650763512, + 0.011240025982260704, + 0.006893414072692394, + 0.014335374347865582, + 0.00629167165607214, + 0.006006081122905016, + 0.01591310277581215, + 0.00930001586675644, + 0.007391920778900385, + 0.005566504783928394, + 0.007482993416488171, + 0.010791582986712456, + 0.00967326294630766, + 0.004640086088329554, + 0.016220418736338615, + 0.011935112997889519, + 0.005837769713252783, + 0.007940675131976604, + 0.012021070346236229, + 0.012277187779545784, + 0.011093082837760448, + 0.016229329630732536, + 0.014877351932227612, + 0.014501836150884628, + 0.021872272714972496, + 0.022436117753386497, + 0.005972571205347776, + 0.004705785308033228, + 0.00656037637963891, + 0.005737303290516138, + 0.013073726557195187, + 0.005868289154022932, + 0.005485950969159603, + 0.003423672402277589, + 0.008432658389210701, + 0.007956868037581444, + 0.008788306266069412, + 0.010195180773735046, + 0.010099616833031178, + 0.01044587790966034, + 0.012654435820877552, + 0.010103119537234306, + 0.009257329627871513, + 0.012442907318472862, + 0.020403897389769554, + 0.020106442272663116, + 0.011762449517846107, + 0.009099042043089867, + 0.017050612717866898, + 0.012099547311663628, + 0.01240517944097519, + 0.009876389056444168, + 0.007358239963650703, + 0.006340093910694122, + 0.013255460187792778, + 0.015129993669688702, + 0.005934754852205515, + 0.010851213708519936, + 0.01324810553342104, + 0.009407482109963894, + 0.01306515745818615, + 0.01737656258046627, + 0.00989014096558094, + 0.012230169028043747, + 0.023846222087740898, + 0.019376074895262718 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.04306888207793236 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0005616756388917565 + ], + "opposite_logprob_tokens": [ + -7.500561714172363 + ], + "logprob_reward": -0.0005616756388917565, + "yes_no_margin_reward": 7.5, + "token_credit_reward": -0.0005616756388917565, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1352945864200592, + 0.03393851965665817, + 0.04952135682106018, + 0.03420182317495346, + 0.1077064722776413, + 0.023824447765946388, + 0.102558434009552, + 0.05662264674901962, + 0.03865370899438858, + 0.4176780879497528 + ], + "credit_map": [ + 0.013285701163113117, + 0.028129516169428825, + 0.01670907437801361, + 0.011706840246915817, + 0.019411573186516762, + 0.02035284973680973, + 0.010364548303186893, + 0.01533447578549385, + 0.0026741584297269583, + 0.005402424372732639, + 0.005264127627015114, + 0.0042391205206513405, + 0.0041978186927735806, + 0.0039649344980716705, + 0.003821933176368475, + 0.004374001175165176, + 0.005157681182026863, + 0.005359848029911518, + 0.004979574121534824, + 0.005431219935417175, + 0.007839448750019073, + 0.008285985328257084, + 0.006979976315051317, + 0.005487620830535889, + 0.0032660665456205606, + 0.004097578581422567, + 0.00421116454526782, + 0.0028159895446151495, + 0.008042951114475727, + 0.004854960832744837, + 0.0035700213629752398, + 0.003343088086694479, + 0.010464586317539215, + 0.010245544835925102, + 0.010342265479266644, + 0.014596574008464813, + 0.012090496718883514, + 0.014156116172671318, + 0.016694752499461174, + 0.019116127863526344, + 0.002877449383959174, + 0.002329600974917412, + 0.0025510762352496386, + 0.002777323592454195, + 0.006000536028295755, + 0.0025291810743510723, + 0.003300410695374012, + 0.0014588688500225544, + 0.010745082050561905, + 0.006221792660653591, + 0.02897942252457142, + 0.00959025789052248, + 0.00974258966743946, + 0.009302442893385887, + 0.009887813590466976, + 0.01808902807533741, + 0.006124775856733322, + 0.006081984844058752, + 0.007991177029907703, + 0.008386236615478992, + 0.007727755233645439, + 0.00765993632376194, + 0.006357627455145121, + 0.00629314873367548, + 0.005448433570563793, + 0.0038167438469827175, + 0.004837007261812687, + 0.0044007510878145695, + 0.005842588376253843, + 0.00760685233399272, + 0.0028063119389116764, + 0.0038950166199356318, + 0.035053472965955734, + 0.024429326876997948, + 0.03450077772140503, + 0.05918915569782257, + 0.044833239167928696, + 0.037258658558130264, + 0.11505336314439774, + 0.0673600360751152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.0005616756388917565 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08902321010828018 + ], + "opposite_logprob_tokens": [ + -2.4640231132507324 + ], + "logprob_reward": -0.08902321010828018, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08902321010828018, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07086383551359177, + 0.02948513999581337, + 0.02722746692597866, + 0.02117617428302765, + 0.05559343099594116, + 0.08458075672388077, + 0.23166005313396454, + 0.08470065146684647, + 0.026582958176732063, + 0.3681294918060303 + ], + "credit_map": [ + 0.006330987438559532, + 0.014310424216091633, + 0.010561419650912285, + 0.005053362343460321, + 0.009934418834745884, + 0.010556682012975216, + 0.0055695511400699615, + 0.008546999655663967, + 0.002355253091081977, + 0.005611300002783537, + 0.004008774179965258, + 0.0042344327084720135, + 0.003879691008478403, + 0.0034789759665727615, + 0.002766659250482917, + 0.003150058211758733, + 0.002608196111395955, + 0.0035340548492968082, + 0.004036754369735718, + 0.003120778826996684, + 0.003807850880548358, + 0.0032229116186499596, + 0.003285404061898589, + 0.0036115196999162436, + 0.0020390197169035673, + 0.002579605206847191, + 0.002124773571267724, + 0.002924674190580845, + 0.0035330888349562883, + 0.0022662864066660404, + 0.0036532296799123287, + 0.0020554971415549517, + 0.005731933284550905, + 0.004561719484627247, + 0.00694294786080718, + 0.007306843064725399, + 0.005352543201297522, + 0.007321922108530998, + 0.007442817557603121, + 0.010932710953056812, + 0.01192538347095251, + 0.005914201959967613, + 0.010784893296658993, + 0.007741847075521946, + 0.015265464782714844, + 0.017235465347766876, + 0.008643757551908493, + 0.007069747429341078, + 0.027289897203445435, + 0.017354663461446762, + 0.06488605588674545, + 0.017537588253617287, + 0.023308873176574707, + 0.026752356439828873, + 0.017834454774856567, + 0.0366961844265461, + 0.007294123061001301, + 0.008349502459168434, + 0.011597895994782448, + 0.016086941584944725, + 0.010074710473418236, + 0.007591960020363331, + 0.014807517640292645, + 0.008898002095520496, + 0.0034975926391780376, + 0.0029863615054637194, + 0.0026935182977467775, + 0.0027726441621780396, + 0.004315446596592665, + 0.0039537823759019375, + 0.00340422336012125, + 0.002959392499178648, + 0.03192273899912834, + 0.02128451317548752, + 0.03083100914955139, + 0.051473625004291534, + 0.03801422566175461, + 0.03433002531528473, + 0.10072971880435944, + 0.05954364687204361 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 7 + ], + "selected_reward": -0.08902321010828018 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, pulling-back tracking shot follows a modern sedan as it steadily advances through the pitch-black night. The car's sleek metallic black paint reflects the faint ambient light from the humid air. Its glaring white LED headlights cast two long, water-distorted beams across the deserted two-lane asphalt. The car drives straight ahead, where a torrential downpour pours down like a cascade, forming a nearly opaque wall of water. The raindrops are so dense that they create a dynamic, grainy texture in the headlights. As the car plunges unhesitatingly into the rain, countless water droplets are instantly illuminated by the headlights, creating a chaotic, shimmering cloud of particles. Then, a massive spray of water explodes upward and to the sides from the front of the car, instantly engulfing the windshield and rendering the frantic thrashing of the wipers futile. The entire scene is shrouded in cold tones, with only the bright light of the car lights contrasting sharply with the endless darkness, creating a strong dramatic tension and tense atmosphere of being isolated from the world and resolutely rushing towards unknown challenges.", + "seed": 20201, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_2391_seed_20201_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.13055754997767507, + "vlm_logprob": -0.13055754997767507, + "vlm_yes_no_margin": 2.65, + "vlm_token_credit": -0.13055754997767507, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15614432096481323 + ], + "opposite_logprob_tokens": [ + -2.031144380569458 + ], + "logprob_reward": -0.15614432096481323, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15614432096481323, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11486885696649551, + 0.0705844983458519, + 0.05234343186020851, + 0.14721588790416718, + 0.17687620222568512, + 0.11466345191001892, + 0.10621815919876099, + 0.062461044639348984, + 0.0842117890715599, + 0.07055667042732239 + ], + "credit_map": [ + 0.010745971463620663, + 0.025585675612092018, + 0.013333767652511597, + 0.00608422327786684, + 0.012492196634411812, + 0.01731139048933983, + 0.00884426012635231, + 0.020471377298235893, + 0.006255012005567551, + 0.013483909890055656, + 0.00676512997597456, + 0.004959512501955032, + 0.008845597505569458, + 0.0160662941634655, + 0.0051721991039812565, + 0.009036845527589321, + 0.005031108856201172, + 0.00820990465581417, + 0.0038113384507596493, + 0.006332131568342447, + 0.007489494979381561, + 0.007818331941962242, + 0.005343258380889893, + 0.008307861164212227, + 0.013102645054459572, + 0.0236783716827631, + 0.012217310257256031, + 0.015342443250119686, + 0.024441484361886978, + 0.02173473685979843, + 0.020304331555962563, + 0.016394559293985367, + 0.015689566731452942, + 0.03302203491330147, + 0.015615520067512989, + 0.019072972238063812, + 0.03726191446185112, + 0.026950301602482796, + 0.011827587150037289, + 0.017436306923627853, + 0.01805066503584385, + 0.013607471249997616, + 0.01854894310235977, + 0.00588261941447854, + 0.01968569867312908, + 0.020069098100066185, + 0.008708625100553036, + 0.010110327042639256, + 0.010820494033396244, + 0.013648680411279202, + 0.009409903548657894, + 0.010012621991336346, + 0.022725339978933334, + 0.016762254759669304, + 0.015148195438086987, + 0.007690667640417814, + 0.006758987903594971, + 0.008973789401352406, + 0.004926577210426331, + 0.006909308955073357, + 0.01165312621742487, + 0.00896388664841652, + 0.0065818713046610355, + 0.0076934946700930595, + 0.006524194963276386, + 0.010392184369266033, + 0.006757648661732674, + 0.013227029703557491, + 0.014017093926668167, + 0.015165251679718494, + 0.008283628150820732, + 0.009844758547842503, + 0.00646364688873291, + 0.008979827165603638, + 0.003986246418207884, + 0.01578441821038723, + 0.013773311860859394, + 0.011958098970353603, + 0.004212244879454374, + 0.005398876499384642 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.15614432096481323 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11297738552093506 + ], + "opposite_logprob_tokens": [ + -2.2379775047302246 + ], + "logprob_reward": -0.11297738552093506, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11297738552093506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36736732721328735, + 0.12324225157499313, + 0.04871411994099617, + 0.05270080268383026, + 0.052364446222782135, + 0.21061982214450836, + 0.06237392500042915, + 0.047149788588285446, + 0.016867930069565773, + 0.01859959401190281 + ], + "credit_map": [ + 0.032768260687589645, + 0.12733423709869385, + 0.02332766354084015, + 0.013873293064534664, + 0.034720245748758316, + 0.08651679009199142, + 0.01557200588285923, + 0.03325485438108444, + 0.009675302542746067, + 0.026167558506131172, + 0.01075703464448452, + 0.00699873361736536, + 0.01629374735057354, + 0.03094618394970894, + 0.006252765189856291, + 0.016150930896401405, + 0.004248359240591526, + 0.006613588891923428, + 0.005326460115611553, + 0.006111151073127985, + 0.005440842360258102, + 0.006847835145890713, + 0.004861103370785713, + 0.009264783002436161, + 0.006649140268564224, + 0.008230033330619335, + 0.0036126484628766775, + 0.004596003331243992, + 0.00848955474793911, + 0.009215028956532478, + 0.0073455520905554295, + 0.004562841262668371, + 0.0052168359979987144, + 0.007510711904615164, + 0.0056579685769975185, + 0.0050768181681632996, + 0.011187247931957245, + 0.007939727045595646, + 0.0037250586319714785, + 0.00605008052662015, + 0.029255816712975502, + 0.028196273371577263, + 0.029651695862412453, + 0.015481632202863693, + 0.02783355489373207, + 0.03050246462225914, + 0.02147662453353405, + 0.028221774846315384, + 0.00625488068908453, + 0.007167984265834093, + 0.01226673275232315, + 0.004106545355170965, + 0.013750213198363781, + 0.009035981260240078, + 0.006137384567409754, + 0.00365420151501894, + 0.0097737992182374, + 0.004607390612363815, + 0.003344095079228282, + 0.005040949210524559, + 0.006025153677910566, + 0.00826941803097725, + 0.004382137209177017, + 0.005706843454390764, + 0.0016023654025048018, + 0.0024213537108153105, + 0.0014744577929377556, + 0.0026011341251432896, + 0.002767367986962199, + 0.0024728192947804928, + 0.001594499684870243, + 0.0019339321879670024, + 0.0017112904461100698, + 0.0026992883067578077, + 0.0016337555134668946, + 0.0027021856512874365, + 0.0033671895507723093, + 0.0029622563160955906, + 0.0017626056214794517, + 0.0017610227223485708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.11297738552093506 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1799009144306183 + ], + "opposite_logprob_tokens": [ + -1.804900884628296 + ], + "logprob_reward": -0.1799009144306183, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1799009144306183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49360552430152893, + 0.08448053896427155, + 0.03555666655302048, + 0.0540718138217926, + 0.05250892415642738, + 0.15662063658237457, + 0.033727459609508514, + 0.036970049142837524, + 0.02182355336844921, + 0.030634805560112 + ], + "credit_map": [ + 0.03886638209223747, + 0.21484217047691345, + 0.014041811227798462, + 0.009555958211421967, + 0.048857755959033966, + 0.1470377892255783, + 0.007070460356771946, + 0.013333234935998917, + 0.007655290886759758, + 0.014036311767995358, + 0.006055215839296579, + 0.0054059866815805435, + 0.012065894901752472, + 0.01718304492533207, + 0.006355781573802233, + 0.01572301797568798, + 0.004395575262606144, + 0.005254305433481932, + 0.002533805090934038, + 0.002887977287173271, + 0.0049929264932870865, + 0.008467346429824829, + 0.0027888119220733643, + 0.004235920496284962, + 0.0047940039075911045, + 0.009796775877475739, + 0.004729752894490957, + 0.004969767760485411, + 0.008838996291160583, + 0.007757037412375212, + 0.004836815409362316, + 0.008348667994141579, + 0.006592565681785345, + 0.0051999506540596485, + 0.004488978069275618, + 0.004535251762717962, + 0.010620948858559132, + 0.007960580289363861, + 0.0071509298868477345, + 0.005959721747785807, + 0.02418479323387146, + 0.021321479231119156, + 0.02493441291153431, + 0.009406629018485546, + 0.020498203113675117, + 0.022858940064907074, + 0.013323133811354637, + 0.020093059167265892, + 0.005928636062890291, + 0.002597917104139924, + 0.004534978419542313, + 0.004750208929181099, + 0.0035274699330329895, + 0.0042136977426707745, + 0.0033520006109029055, + 0.004822554066777229, + 0.004354479722678661, + 0.003697680775076151, + 0.003912958316504955, + 0.005521820858120918, + 0.005437337793409824, + 0.005263513885438442, + 0.004549759905785322, + 0.004232500679790974, + 0.0016119936481118202, + 0.002743300748988986, + 0.0016481331549584866, + 0.003338891314342618, + 0.0037438133731484413, + 0.0033876942470669746, + 0.0028601749800145626, + 0.002489553764462471, + 0.0026706336066126823, + 0.0036783136893063784, + 0.002267218194901943, + 0.005792881827801466, + 0.006279968656599522, + 0.005152693018317223, + 0.002380114048719406, + 0.002412984613329172 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.1799009144306183 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00220104050822556 + ], + "opposite_logprob_tokens": [ + -6.127201080322266 + ], + "logprob_reward": -0.00220104050822556, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.00220104050822556, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4386765658855438, + 0.09164807200431824, + 0.05309038981795311, + 0.0660528615117073, + 0.08852700144052505, + 0.07485730201005936, + 0.05283261835575104, + 0.054124865680933, + 0.03630669042468071, + 0.04388362914323807 + ], + "credit_map": [ + 0.03094787523150444, + 0.1358318030834198, + 0.025112541392445564, + 0.01545112207531929, + 0.03605175018310547, + 0.10070572048425674, + 0.021691106259822845, + 0.07288463413715363, + 0.008025345392525196, + 0.013355677016079426, + 0.013721478171646595, + 0.006455729715526104, + 0.009963133372366428, + 0.018344758078455925, + 0.0064276522025465965, + 0.015354296192526817, + 0.006244009826332331, + 0.007505897898226976, + 0.004035084042698145, + 0.005602429620921612, + 0.006846834905445576, + 0.010859631933271885, + 0.0054518380202353, + 0.00654465751722455, + 0.005978572182357311, + 0.009998993948101997, + 0.008224552497267723, + 0.007701015565544367, + 0.007080484181642532, + 0.009152661077678204, + 0.010106800124049187, + 0.007809778675436974, + 0.0090049272403121, + 0.013961616903543472, + 0.008386706002056599, + 0.008813058957457542, + 0.016226530075073242, + 0.012749910354614258, + 0.008594159036874771, + 0.010790091939270496, + 0.010727942921221256, + 0.011059649288654327, + 0.009616496972739697, + 0.004944863263517618, + 0.010804466903209686, + 0.010008267126977444, + 0.007304620463401079, + 0.010390986688435078, + 0.006719456985592842, + 0.005083369091153145, + 0.007632849272340536, + 0.004672931507229805, + 0.010176621377468109, + 0.007945605553686619, + 0.005800721235573292, + 0.004801063798367977, + 0.006301437970250845, + 0.00642796466127038, + 0.005655386485159397, + 0.00635142158716917, + 0.00877272617071867, + 0.00803388562053442, + 0.006626661866903305, + 0.005955378990620375, + 0.002552760997787118, + 0.003908619750291109, + 0.003761664731428027, + 0.006502971984446049, + 0.005561038851737976, + 0.005345532204955816, + 0.005017453338950872, + 0.0036566476337611675, + 0.002911729970946908, + 0.004511727020144463, + 0.003042415948584676, + 0.008264336735010147, + 0.0077394768595695496, + 0.007750834338366985, + 0.005010205786675215, + 0.0046528978273272514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00220104050822556 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156408846378326 + ], + "opposite_logprob_tokens": [ + -1.701564073562622 + ], + "logprob_reward": -0.20156408846378326, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156408846378326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37903690338134766, + 0.07939858734607697, + 0.03629257157444954, + 0.07804204523563385, + 0.14176972210407257, + 0.046338606625795364, + 0.07191606611013412, + 0.06102930009365082, + 0.04247293993830681, + 0.06370329856872559 + ], + "credit_map": [ + 0.02410997450351715, + 0.1511516571044922, + 0.011958681978285313, + 0.007638165261596441, + 0.033535782247781754, + 0.11396133899688721, + 0.00967035535722971, + 0.027010919526219368, + 0.006470549385994673, + 0.01369012612849474, + 0.008954270742833614, + 0.004390506539493799, + 0.008800174109637737, + 0.016339896246790886, + 0.005135545041412115, + 0.015617518685758114, + 0.004066954366862774, + 0.005585141479969025, + 0.00237071025185287, + 0.0028437264263629913, + 0.0056794174015522, + 0.008096305653452873, + 0.002303843619301915, + 0.005346468649804592, + 0.00645132502540946, + 0.016777455806732178, + 0.006517819128930569, + 0.006304297596216202, + 0.012209584936499596, + 0.014767642132937908, + 0.005475501995533705, + 0.009538420476019382, + 0.012679992243647575, + 0.02507064677774906, + 0.012865555472671986, + 0.014231067150831223, + 0.03258363902568817, + 0.022022463381290436, + 0.008976004086434841, + 0.01334035862237215, + 0.006267176475375891, + 0.005718529690057039, + 0.008680547587573528, + 0.004080116283148527, + 0.00566693814471364, + 0.006111428141593933, + 0.0038177950773388147, + 0.005996072664856911, + 0.006679007317870855, + 0.006686939392238855, + 0.005974297411739826, + 0.004433697555214167, + 0.020845260471105576, + 0.012762968428432941, + 0.009686230681836605, + 0.0048476592637598515, + 0.0057488661259412766, + 0.005837157368659973, + 0.0058449856005609035, + 0.011000590398907661, + 0.00832943245768547, + 0.007593912072479725, + 0.009011918678879738, + 0.007662429008632898, + 0.003415906336158514, + 0.004286718089133501, + 0.0043407101184129715, + 0.005749461706727743, + 0.006757485680282116, + 0.005872388835996389, + 0.006113887764513493, + 0.0059363809414207935, + 0.006180838216096163, + 0.006795521825551987, + 0.005245777312666178, + 0.01394808292388916, + 0.007578173186630011, + 0.010819402523338795, + 0.006228094454854727, + 0.006907399743795395 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.20156408846378326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, pulling-back tracking shot follows a modern sedan as it steadily advances through the pitch-black night. The car's sleek metallic black paint reflects the faint ambient light from the humid air. Its glaring white LED headlights cast two long, water-distorted beams across the deserted two-lane asphalt. The car drives straight ahead, where a torrential downpour pours down like a cascade, forming a nearly opaque wall of water. The raindrops are so dense that they create a dynamic, grainy texture in the headlights. As the car plunges unhesitatingly into the rain, countless water droplets are instantly illuminated by the headlights, creating a chaotic, shimmering cloud of particles. Then, a massive spray of water explodes upward and to the sides from the front of the car, instantly engulfing the windshield and rendering the frantic thrashing of the wipers futile. The entire scene is shrouded in cold tones, with only the bright light of the car lights contrasting sharply with the endless darkness, creating a strong dramatic tension and tense atmosphere of being isolated from the world and resolutely rushing towards unknown challenges.", + "seed": 20202, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_0/index_2391_seed_20202_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.20762284446973353, + "vlm_logprob": -0.20762284446973353, + "vlm_yes_no_margin": 2.449999976158142, + "vlm_token_credit": -0.20762284446973353, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6391701698303223 + ], + "opposite_logprob_tokens": [ + -0.7641701698303223 + ], + "logprob_reward": -0.6391701698303223, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6391701698303223, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12042354047298431, + 0.16261400282382965, + 0.13622967898845673, + 0.052665531635284424, + 0.034152936190366745, + 0.03560037910938263, + 0.14716953039169312, + 0.034015461802482605, + 0.06015874817967415, + 0.21697022020816803 + ], + "credit_map": [ + 0.009781680069863796, + 0.011690095067024231, + 0.01705561764538288, + 0.008464721962809563, + 0.024016013368964195, + 0.01954331248998642, + 0.007785737980157137, + 0.022086357697844505, + 0.012478540651500225, + 0.011098570190370083, + 0.005895249079912901, + 0.003832352813333273, + 0.051553018391132355, + 0.06366217881441116, + 0.0035746474750339985, + 0.010519439354538918, + 0.009558817371726036, + 0.014167635701596737, + 0.016136178746819496, + 0.016414722427725792, + 0.029361655935645103, + 0.022625930607318878, + 0.013457205146551132, + 0.014507521875202656, + 0.004573899321258068, + 0.009137596003711224, + 0.006602139677852392, + 0.004333506338298321, + 0.006041867658495903, + 0.006696419790387154, + 0.004016600549221039, + 0.011263500899076462, + 0.002340006409212947, + 0.005225705448538065, + 0.002653521951287985, + 0.0027728541754186153, + 0.004369559232145548, + 0.005546831991523504, + 0.003136888612061739, + 0.008107571862637997, + 0.006458789575845003, + 0.004746968392282724, + 0.0028035969007760286, + 0.002064740052446723, + 0.004281121771782637, + 0.004536089487373829, + 0.0035181730054318905, + 0.007190901320427656, + 0.02031666785478592, + 0.023256659507751465, + 0.010287665762007236, + 0.00817082915455103, + 0.021214384585618973, + 0.017293862998485565, + 0.020486580207943916, + 0.026142867282032967, + 0.004524950869381428, + 0.004285127390176058, + 0.002426965394988656, + 0.002200174378231168, + 0.0042707910761237144, + 0.004930492490530014, + 0.0029727425426244736, + 0.008404218591749668, + 0.012888944707810879, + 0.011729973368346691, + 0.005509990733116865, + 0.005059930495917797, + 0.006046032998710871, + 0.008664351888000965, + 0.0034338929690420628, + 0.0068256305530667305, + 0.01749916933476925, + 0.03638656064867973, + 0.018664749339222908, + 0.03957955539226532, + 0.02712039276957512, + 0.042078278958797455, + 0.01479019783437252, + 0.02085132524371147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 6 + ], + "selected_reward": -0.6391701698303223 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12711314857006073 + ], + "opposite_logprob_tokens": [ + -2.127113103866577 + ], + "logprob_reward": -0.12711314857006073, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12711314857006073, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.534756600856781, + 0.147626131772995, + 0.09081688523292542, + 0.01840890571475029, + 0.02729869820177555, + 0.035281576216220856, + 0.031595319509506226, + 0.017157653346657753, + 0.03803345933556557, + 0.05902482196688652 + ], + "credit_map": [ + 0.037446700036525726, + 0.026199664920568466, + 0.0142745953053236, + 0.009964720346033573, + 0.18983282148838043, + 0.21308349072933197, + 0.0097444336861372, + 0.034210141748189926, + 0.010533672757446766, + 0.008575822226703167, + 0.003857079427689314, + 0.00249025272205472, + 0.04888097196817398, + 0.060598794370889664, + 0.0022262870334088802, + 0.01046324335038662, + 0.00942959450185299, + 0.008210107684135437, + 0.00815479177981615, + 0.008296091109514236, + 0.017589669674634933, + 0.02614067867398262, + 0.005967994686216116, + 0.007027947343885899, + 0.001863223616965115, + 0.0030931858345866203, + 0.002215260174125433, + 0.0009803822031244636, + 0.002712986897677183, + 0.0034821140579879284, + 0.0008962739957496524, + 0.0031654771883040667, + 0.0023138776887208223, + 0.00450735492631793, + 0.0021969133522361517, + 0.0028711401391774416, + 0.002933569485321641, + 0.0032870210707187653, + 0.0032038011122494936, + 0.005985018331557512, + 0.0035036634653806686, + 0.00838460773229599, + 0.0019402526086196303, + 0.0013135155895724893, + 0.004231720231473446, + 0.010699289850890636, + 0.0019472627900540829, + 0.0032612630166113377, + 0.004537731409072876, + 0.004038133658468723, + 0.0022675693035125732, + 0.0018754913471639156, + 0.0061800614930689335, + 0.006028495263308287, + 0.0030891222413629293, + 0.0035787110682576895, + 0.0019782236777246, + 0.002260761335492134, + 0.0008005177369341254, + 0.0007192916818894446, + 0.002522665774449706, + 0.003348890459164977, + 0.0011327156098559499, + 0.004394585266709328, + 0.007523624692112207, + 0.007720019668340683, + 0.003399468958377838, + 0.0029256080742925406, + 0.0042549786157906055, + 0.006078353151679039, + 0.002368777524679899, + 0.00376262329518795, + 0.00452454574406147, + 0.01670992374420166, + 0.00333856581710279, + 0.0039977035485208035, + 0.005689116660505533, + 0.019869813695549965, + 0.001997543266043067, + 0.0028975962195545435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.12711314857006073 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270829141139984 + ], + "opposite_logprob_tokens": [ + -2.1270828247070312 + ], + "logprob_reward": -0.1270829141139984, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.1270829141139984, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.49113741517066956, + 0.2193020135164261, + 0.06263121217489243, + 0.03839097172021866, + 0.02459501102566719, + 0.020546985790133476, + 0.03648548200726509, + 0.0269952230155468, + 0.029962733387947083, + 0.04995294660329819 + ], + "credit_map": [ + 0.059884294867515564, + 0.04287540912628174, + 0.025505218654870987, + 0.020074112340807915, + 0.10368045419454575, + 0.15791040658950806, + 0.02375865913927555, + 0.05744883790612221, + 0.018026186153292656, + 0.016216594725847244, + 0.006138126831501722, + 0.003785402048379183, + 0.06858990341424942, + 0.09088683873414993, + 0.0036777928471565247, + 0.01198117807507515, + 0.0045499494299292564, + 0.007981396280229092, + 0.009769218973815441, + 0.005023821722716093, + 0.012529595755040646, + 0.008618971332907677, + 0.004397042561322451, + 0.009761217050254345, + 0.0034126085229218006, + 0.004683285020291805, + 0.004740388598293066, + 0.0026683902833610773, + 0.004901032894849777, + 0.008822173811495304, + 0.002333842683583498, + 0.006829249672591686, + 0.001964069437235594, + 0.0032783597707748413, + 0.0019414094276726246, + 0.0014939089305698872, + 0.0033117737621068954, + 0.005570361856371164, + 0.0016792279202491045, + 0.005355900153517723, + 0.0026133174542337656, + 0.0019103731028735638, + 0.0018338164081797004, + 0.0011917572701349854, + 0.0030629525426775217, + 0.0037022046744823456, + 0.0019077700562775135, + 0.004324792884290218, + 0.003837071591988206, + 0.004044422414153814, + 0.0015477797714993358, + 0.0013578799553215504, + 0.008928398601710796, + 0.009498306550085545, + 0.002913983538746834, + 0.004357639700174332, + 0.0033361867535859346, + 0.0023956343065947294, + 0.0016254406655207276, + 0.0014699967578053474, + 0.005718786269426346, + 0.0068266550078988075, + 0.0016301337163895369, + 0.003992388490587473, + 0.0033926439937204123, + 0.004895188380032778, + 0.002871175529435277, + 0.004341140855103731, + 0.003354673972353339, + 0.004452850669622421, + 0.0029428680427372456, + 0.0037121919449418783, + 0.00441302265971899, + 0.007449374999850988, + 0.0049414969980716705, + 0.010623278096318245, + 0.006539817899465561, + 0.006636114791035652, + 0.0038675381802022457, + 0.005482305306941271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.1270829141139984 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019436056027188897 + ], + "opposite_logprob_tokens": [ + -6.251943588256836 + ], + "logprob_reward": -0.0019436056027188897, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019436056027188897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38831374049186707, + 0.3152090013027191, + 0.07450103759765625, + 0.023638691753149033, + 0.020806465297937393, + 0.017632663249969482, + 0.0173108521848917, + 0.03110894188284874, + 0.043857935816049576, + 0.06762068718671799 + ], + "credit_map": [ + 0.03912802040576935, + 0.0186528991907835, + 0.016541389748454094, + 0.009216777049005032, + 0.11978089064359665, + 0.15528549253940582, + 0.0073130568489432335, + 0.022395189851522446, + 0.028081413358449936, + 0.02181665040552616, + 0.008476939052343369, + 0.005410587880760431, + 0.09849737584590912, + 0.13410134613513947, + 0.004324756562709808, + 0.014499906450510025, + 0.004140967968851328, + 0.007264719810336828, + 0.007512993179261684, + 0.005739201325923204, + 0.019580524414777756, + 0.018798111006617546, + 0.005676280241459608, + 0.005788230802863836, + 0.0017019358929246664, + 0.0050233155488967896, + 0.0030784793198108673, + 0.001048819744028151, + 0.004066118039190769, + 0.0033080128487199545, + 0.0012830811319872737, + 0.004128927364945412, + 0.001576435286551714, + 0.003507687710225582, + 0.0016334881074726582, + 0.0014744088985025883, + 0.002299634972587228, + 0.0040273768827319145, + 0.0015409139450639486, + 0.004746518097817898, + 0.0022524199448525906, + 0.0032302795443683863, + 0.0008952195639722049, + 0.000921873317565769, + 0.0021481537260115147, + 0.0044102743268013, + 0.0012148019159212708, + 0.0025596378836780787, + 0.002435632050037384, + 0.0028342651203274727, + 0.001052492531016469, + 0.0009647904662415385, + 0.002807743614539504, + 0.002581763081252575, + 0.001868645427748561, + 0.0027655181474983692, + 0.004526960197836161, + 0.0029204420279711485, + 0.0012326312717050314, + 0.0015367618761956692, + 0.00578604731708765, + 0.007024721242487431, + 0.0015236131148412824, + 0.006557763554155827, + 0.006168421823531389, + 0.009796952828764915, + 0.0033603247720748186, + 0.0028098844923079014, + 0.002696460811421275, + 0.011157823726534843, + 0.003803964238613844, + 0.004064102191478014, + 0.005721419118344784, + 0.019673319533467293, + 0.003509833011776209, + 0.004562684800475836, + 0.007659635506570339, + 0.019750598818063736, + 0.0030698131304234266, + 0.0036733748856931925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.0019436056027188897 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a modern sedan with sleek metallic black paint?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the scene set at night with a pitch-black background and cold tones?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the car driving straight ahead through a torrential downpour that forms a nearly opaque wall of water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do the car's white LED headlights cast long beams that are distorted by water and illuminate a chaotic cloud of rain particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle pulling-back tracking shot following the car?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280438423156738 + ], + "opposite_logprob_tokens": [ + -2.0178043842315674 + ], + "logprob_reward": -0.14280438423156738, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280438423156738, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06025884672999382, + 0.19798289239406586, + 0.1442916989326477, + 0.056711554527282715, + 0.06994924694299698, + 0.19743697345256805, + 0.13891500234603882, + 0.052287228405475616, + 0.048279132694005966, + 0.03388744965195656 + ], + "credit_map": [ + 0.005051952321082354, + 0.006041846238076687, + 0.007355339825153351, + 0.004401023965328932, + 0.01260159071534872, + 0.01219955738633871, + 0.0039285351522266865, + 0.008678999729454517, + 0.014573405496776104, + 0.013951465487480164, + 0.01095262635499239, + 0.00561936991289258, + 0.051368944346904755, + 0.06407836824655533, + 0.006181210745126009, + 0.031257513910532, + 0.008379312232136726, + 0.01429536659270525, + 0.014372507110238075, + 0.01850544847548008, + 0.03247975558042526, + 0.02815346233546734, + 0.012554868124425411, + 0.015550980344414711, + 0.004553600680083036, + 0.010361098684370518, + 0.006937062367796898, + 0.0045790402218699455, + 0.006308126263320446, + 0.006819305010139942, + 0.0041642989963293076, + 0.012989022769033909, + 0.006258928682655096, + 0.009656047448515892, + 0.005079331342130899, + 0.005254982504993677, + 0.0073826927691698074, + 0.013238398358225822, + 0.006059807259589434, + 0.017019059509038925, + 0.02576160617172718, + 0.034165360033512115, + 0.012586143799126148, + 0.010831962339580059, + 0.027539128437638283, + 0.04110301285982132, + 0.018356679007411003, + 0.027093078941106796, + 0.015773862600326538, + 0.014522857032716274, + 0.010832665488123894, + 0.0073890965431928635, + 0.0234743170440197, + 0.020640946924686432, + 0.023529957979917526, + 0.02275129407644272, + 0.005909498315304518, + 0.007940120063722134, + 0.00442484999075532, + 0.0037508588284254074, + 0.00832362286746502, + 0.010249542072415352, + 0.003956833388656378, + 0.007731902413070202, + 0.006679887883365154, + 0.009939830750226974, + 0.004165775142610073, + 0.00440694484859705, + 0.0034868111833930016, + 0.010124849155545235, + 0.005036505870521069, + 0.004438526462763548, + 0.0033203631173819304, + 0.004758034832775593, + 0.0028729974292218685, + 0.007162503898143768, + 0.004715213552117348, + 0.00483655696734786, + 0.0028664912097156048, + 0.003355288878083229 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.14280438423156738 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_tensor([4782, 4782])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_tensor([4782, 4782])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9f76307947859265dcc84c4b2e33cb8e4eaea42f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_tensor([4782, 4782])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static, medium shot shows an ornate longsword lying in a shallow pool of blood on a cold, cracked stone floor. The sword's hilt, bound in worn black leather, is topped with a snarling wolf's head, adorned with small, gleaming ruby eyes. Its cross is an intricate grid of faded, thorn-like metal. The black, non-reflective steel of the blade is covered in deeply etched ancient runes. The focal point of the scene is a slow, mesmerizing horror: viscous crimson blood is deliberately absorbed into the blade. Thin streams of blood flow unnaturally from the puddle, as if drawn toward the etched symbols by an unseen force. As the liquid contacts each rune, the symbol comes alive, emitting a pulsating crimson glow that grows brighter with each sip. The blood vanished into the steel, leaving no residue behind, as if the weapon itself were drinking the water. These malevolent, rhythmic rays from the runes served as the primary light source, casting deep, flickering shadows and bathing the surrounding area in a hellish red glow, creating a tense and deeply sinister atmosphere as the dark ritual neared its conclusion.", + "seed": 4754, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4754_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.07661922136321664, + "vlm_logprob": -0.07661922136321664, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.07661922136321664, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05276188254356384 + ], + "opposite_logprob_tokens": [ + -3.1777617931365967 + ], + "logprob_reward": -0.05276188254356384, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.05276188254356384, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2936367988586426, + 0.07414078712463379, + 0.054845619946718216, + 0.08969257026910782, + 0.053825587034225464, + 0.06851226836442947, + 0.09663694351911545, + 0.15738767385482788, + 0.0352652482688427, + 0.07605645805597305 + ], + "credit_map": [ + 0.06459108740091324, + 0.04403461888432503, + 0.0967278778553009, + 0.019054101780056953, + 0.020951004698872566, + 0.020224573090672493, + 0.01377434004098177, + 0.014279224909842014, + 0.01363375410437584, + 0.01206167507916689, + 0.011755810119211674, + 0.004134602844715118, + 0.006569727789610624, + 0.0059530483558773994, + 0.006954098120331764, + 0.013078080490231514, + 0.01015869714319706, + 0.009124324657022953, + 0.013922440819442272, + 0.003937835805118084, + 0.003985752817243338, + 0.005603438708931208, + 0.003471778705716133, + 0.004641358274966478, + 0.01723351888358593, + 0.013443194329738617, + 0.010678454302251339, + 0.006091642659157515, + 0.016810718923807144, + 0.008605497889220715, + 0.008257731795310974, + 0.008571812883019447, + 0.009346836246550083, + 0.010017496533691883, + 0.006287633907049894, + 0.0038316433783620596, + 0.005589384585618973, + 0.006403705105185509, + 0.005571260116994381, + 0.006777628790587187, + 0.012791394256055355, + 0.013541869819164276, + 0.009323122911155224, + 0.004003727342933416, + 0.010541758500039577, + 0.00824069231748581, + 0.00497056357562542, + 0.005099141504615545, + 0.015939861536026, + 0.011952206492424011, + 0.010393255390226841, + 0.00648851040750742, + 0.01998896710574627, + 0.013803706504404545, + 0.009167981334030628, + 0.008902464061975479, + 0.019220205023884773, + 0.018707361072301865, + 0.014320097863674164, + 0.024583203718066216, + 0.017787417396903038, + 0.010153883136808872, + 0.032100312411785126, + 0.020515216514468193, + 0.004970422014594078, + 0.006310532800853252, + 0.005135448183864355, + 0.002832662547007203, + 0.005184960085898638, + 0.004028387367725372, + 0.003144447924569249, + 0.0036583873443305492, + 0.014156036078929901, + 0.010215210728347301, + 0.013326387852430344, + 0.0069597638212144375, + 0.005925227887928486, + 0.006083805114030838, + 0.00805592816323042, + 0.01133410632610321 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.05276188254356384 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04311215505003929 + ], + "opposite_logprob_tokens": [ + -3.16811203956604 + ], + "logprob_reward": -0.04311215505003929, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04311215505003929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30792099237442017, + 0.02045336924493313, + 0.04753590375185013, + 0.07324589043855667, + 0.05366579815745354, + 0.092771515250206, + 0.13369032740592957, + 0.18444521725177765, + 0.017056921496987343, + 0.0692141130566597 + ], + "credit_map": [ + 0.07195097208023071, + 0.047300610691308975, + 0.1078220084309578, + 0.016368204727768898, + 0.01951424777507782, + 0.01799115352332592, + 0.01360451988875866, + 0.013369236141443253, + 0.0027067118790000677, + 0.0029225063044577837, + 0.004059520550072193, + 0.0010558452922850847, + 0.0025465497747063637, + 0.0027111838571727276, + 0.0019006027141585946, + 0.002550446195527911, + 0.00784700084477663, + 0.008589404635131359, + 0.008083529770374298, + 0.0022998114582151175, + 0.005818687845021486, + 0.004417930729687214, + 0.004225606564432383, + 0.006253923289477825, + 0.01256273128092289, + 0.010443580336868763, + 0.011301958933472633, + 0.004718318115919828, + 0.010253572836518288, + 0.0070655327290296555, + 0.0069748591631650925, + 0.009925334714353085, + 0.008954940363764763, + 0.008911415003240108, + 0.007572174072265625, + 0.003913562279194593, + 0.005707256495952606, + 0.005783941131085157, + 0.005396510940045118, + 0.007425994146615267, + 0.03141392767429352, + 0.03352954983711243, + 0.004907702561467886, + 0.0031872126273810863, + 0.006168130785226822, + 0.005057486239820719, + 0.004248964600265026, + 0.004258531611412764, + 0.023073039948940277, + 0.021794941276311874, + 0.013846876099705696, + 0.007645537611097097, + 0.028642352670431137, + 0.019090628251433372, + 0.01082656066864729, + 0.008770393207669258, + 0.020878136157989502, + 0.02241593226790428, + 0.01798286661505699, + 0.030432362109422684, + 0.01858050562441349, + 0.012232920154929161, + 0.03808481618762016, + 0.023837659507989883, + 0.0026388142723590136, + 0.0020098406821489334, + 0.002113538794219494, + 0.0018626145320013165, + 0.002130758948624134, + 0.002078777877613902, + 0.0021573975682258606, + 0.0020651777740567923, + 0.013689464889466763, + 0.011567384004592896, + 0.009424682706594467, + 0.0061624725349247456, + 0.004867296200245619, + 0.005938603542745113, + 0.006797286681830883, + 0.010766918770968914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.04311215505003929 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521505653858185 + ], + "opposite_logprob_tokens": [ + -1.502150535583496 + ], + "logprob_reward": -0.2521505653858185, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521505653858185, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3282169699668884, + 0.034813061356544495, + 0.04420540854334831, + 0.067515067756176, + 0.04802275821566582, + 0.09574505686759949, + 0.05211052671074867, + 0.22391138970851898, + 0.028381243348121643, + 0.07707849144935608 + ], + "credit_map": [ + 0.07072879374027252, + 0.05467475950717926, + 0.1162622720003128, + 0.021429527550935745, + 0.018849024549126625, + 0.018993012607097626, + 0.013730076141655445, + 0.013549498282372952, + 0.006734772585332394, + 0.0045902119018137455, + 0.008547010831534863, + 0.001984742470085621, + 0.002257382031530142, + 0.003269877750426531, + 0.002678581280633807, + 0.004750481806695461, + 0.007769078481942415, + 0.006836633197963238, + 0.007732616737484932, + 0.002320610685274005, + 0.004624504130333662, + 0.0048087649047374725, + 0.003970626275986433, + 0.006142576225101948, + 0.01225690171122551, + 0.008890309371054173, + 0.007116010412573814, + 0.005833418108522892, + 0.009629516862332821, + 0.009472809731960297, + 0.006621216889470816, + 0.007694880478084087, + 0.008454936556518078, + 0.007053959649056196, + 0.007840386591851711, + 0.0034087677486240864, + 0.005690698977559805, + 0.004952925723046064, + 0.004913180600851774, + 0.005707903299480677, + 0.03456878289580345, + 0.035452209413051605, + 0.004233403131365776, + 0.0035283444449305534, + 0.004653161391615868, + 0.003908940125256777, + 0.004409419372677803, + 0.004990804009139538, + 0.007952678017318249, + 0.00903263408690691, + 0.006022034678608179, + 0.003394691040739417, + 0.008822127245366573, + 0.007889614440500736, + 0.00486810551956296, + 0.00412864051759243, + 0.026893749833106995, + 0.026833662763237953, + 0.02231457456946373, + 0.03934135288000107, + 0.019044501706957817, + 0.011870275251567364, + 0.047586072236299515, + 0.030027205124497414, + 0.0038328918162733316, + 0.0038140504620969296, + 0.004301588051021099, + 0.0029060030356049538, + 0.004092132207006216, + 0.0031249274034053087, + 0.003047317499294877, + 0.003262331010773778, + 0.010319910012185574, + 0.01249610260128975, + 0.014113028533756733, + 0.009013136848807335, + 0.006103514693677425, + 0.006116355303674936, + 0.009932338260114193, + 0.008984106592833996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.2521505653858185 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00525640370324254 + ], + "opposite_logprob_tokens": [ + -5.255256175994873 + ], + "logprob_reward": -0.00525640370324254, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00525640370324254, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09107375890016556, + 0.014220311306416988, + 0.00875943899154663, + 0.029782993718981743, + 0.018343903124332428, + 0.020779302343726158, + 0.035208411514759064, + 0.5577982664108276, + 0.03110046125948429, + 0.19293317198753357 + ], + "credit_map": [ + 0.0198379997164011, + 0.014559990726411343, + 0.032985344529151917, + 0.004175052512437105, + 0.00469659548252821, + 0.004570745397359133, + 0.00500383460894227, + 0.0052441940642893314, + 0.001844933838583529, + 0.0019687023013830185, + 0.0027507718186825514, + 0.0011211964301764965, + 0.0015286211855709553, + 0.001558887423016131, + 0.001332114334218204, + 0.002115082461386919, + 0.0014078030362725258, + 0.0014330229023471475, + 0.0021890727803111076, + 0.0005451762699522078, + 0.00068483134964481, + 0.0011880692327395082, + 0.0006266784039326012, + 0.0006847847835160792, + 0.006521331612020731, + 0.00507521815598011, + 0.004105553496629, + 0.0018139883177354932, + 0.0026270521339029074, + 0.002925676992163062, + 0.0028768223710358143, + 0.003837346797809005, + 0.0030830579344183207, + 0.0023876121267676353, + 0.0025767763145267963, + 0.0008866076823323965, + 0.002598605351522565, + 0.0022210730239748955, + 0.0020432702731341124, + 0.002546899253502488, + 0.004073835909366608, + 0.004087632056325674, + 0.0029900348745286465, + 0.001227838103659451, + 0.0027028946205973625, + 0.002076094038784504, + 0.00182460923679173, + 0.0017963615246117115, + 0.0067392634227871895, + 0.005075694061815739, + 0.004605691879987717, + 0.0022780310828238726, + 0.005766855087131262, + 0.004334283526986837, + 0.0031245737336575985, + 0.003284017788246274, + 0.06658849120140076, + 0.06542983651161194, + 0.05423770472407341, + 0.09766804426908493, + 0.05413425713777542, + 0.028230953961610794, + 0.11783938109874725, + 0.07366961240768433, + 0.004887365270406008, + 0.0046868338249623775, + 0.0037268088199198246, + 0.0034653854090720415, + 0.0035668672062456608, + 0.004024225287139416, + 0.002881795633584261, + 0.003861177945509553, + 0.05053793266415596, + 0.01756368950009346, + 0.03143388032913208, + 0.021209048107266426, + 0.014592725783586502, + 0.014801165089011192, + 0.02165352553129196, + 0.021141184493899345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00525640370324254 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029815100133419037 + ], + "opposite_logprob_tokens": [ + -3.529815196990967 + ], + "logprob_reward": -0.029815100133419037, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029815100133419037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24549046158790588, + 0.06322906166315079, + 0.022087380290031433, + 0.06635750830173492, + 0.05585090070962906, + 0.05513767898082733, + 0.09158780425786972, + 0.09158581495285034, + 0.07064057141542435, + 0.2380327433347702 + ], + "credit_map": [ + 0.04085344821214676, + 0.039199165999889374, + 0.07280838489532471, + 0.02434777095913887, + 0.016318928450345993, + 0.02013658732175827, + 0.01548722106963396, + 0.01633896864950657, + 0.02021946758031845, + 0.006426298059523106, + 0.013242527842521667, + 0.002255331026390195, + 0.0038452327717095613, + 0.004300351720303297, + 0.005099416244775057, + 0.007840439677238464, + 0.0031780533026903868, + 0.0035192673094570637, + 0.004297537263482809, + 0.0012298779329285026, + 0.0032633556984364986, + 0.0025595000479370356, + 0.001810860470868647, + 0.00222893082536757, + 0.01356926467269659, + 0.010876483283936977, + 0.010275793261826038, + 0.004046296700835228, + 0.005520847160369158, + 0.006204878445714712, + 0.00677808653563261, + 0.009085857309401035, + 0.008797106333076954, + 0.009531436488032341, + 0.006208064500242472, + 0.00358018116094172, + 0.005564041435718536, + 0.006643129047006369, + 0.006420718505978584, + 0.00910622626543045, + 0.016157114878296852, + 0.016929784789681435, + 0.004392990842461586, + 0.0022811859380453825, + 0.0046567413955926895, + 0.004261204041540623, + 0.0030525263864547014, + 0.0034061302430927753, + 0.015994446352124214, + 0.014088550582528114, + 0.01112330798059702, + 0.005980323068797588, + 0.017223848029971123, + 0.010547773912549019, + 0.008494284935295582, + 0.008135275915265083, + 0.013481284491717815, + 0.010855307802557945, + 0.011078573763370514, + 0.0067441947758197784, + 0.015720274299383163, + 0.007221988867968321, + 0.013441842049360275, + 0.013042351230978966, + 0.011196478269994259, + 0.007382345851510763, + 0.012867839075624943, + 0.007308317814022303, + 0.00931788980960846, + 0.008190938271582127, + 0.007419591769576073, + 0.006957175210118294, + 0.029762456193566322, + 0.03471821919083595, + 0.04380102828145027, + 0.028629889711737633, + 0.02563468925654888, + 0.02299821563065052, + 0.029835250228643417, + 0.022653013467788696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.029815100133419037 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, medium shot shows an ornate longsword lying in a shallow pool of blood on a cold, cracked stone floor. The sword's hilt, bound in worn black leather, is topped with a snarling wolf's head, adorned with small, gleaming ruby eyes. Its cross is an intricate grid of faded, thorn-like metal. The black, non-reflective steel of the blade is covered in deeply etched ancient runes. The focal point of the scene is a slow, mesmerizing horror: viscous crimson blood is deliberately absorbed into the blade. Thin streams of blood flow unnaturally from the puddle, as if drawn toward the etched symbols by an unseen force. As the liquid contacts each rune, the symbol comes alive, emitting a pulsating crimson glow that grows brighter with each sip. The blood vanished into the steel, leaving no residue behind, as if the weapon itself were drinking the water. These malevolent, rhythmic rays from the runes served as the primary light source, casting deep, flickering shadows and bathing the surrounding area in a hellish red glow, creating a tense and deeply sinister atmosphere as the dark ritual neared its conclusion.", + "seed": 4755, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4755_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.07271162793040276, + "vlm_logprob": -0.07271162793040276, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.07271162793040276, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04468749836087227 + ], + "opposite_logprob_tokens": [ + -3.419687509536743 + ], + "logprob_reward": -0.04468749836087227, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04468749836087227, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33244264125823975, + 0.06118667498230934, + 0.053049251437187195, + 0.039976075291633606, + 0.024474846199154854, + 0.05588330328464508, + 0.050848666578531265, + 0.03135787695646286, + 0.2004723697900772, + 0.1503082662820816 + ], + "credit_map": [ + 0.03842183202505112, + 0.04131481423974037, + 0.03903133049607277, + 0.03351830318570137, + 0.07342907786369324, + 0.045226890593767166, + 0.03576623275876045, + 0.02573416568338871, + 0.007204228546470404, + 0.006320171989500523, + 0.007132710423320532, + 0.005546989850699902, + 0.01577739603817463, + 0.0086442269384861, + 0.006216451525688171, + 0.004344498738646507, + 0.005925151519477367, + 0.0051026493310928345, + 0.006898175459355116, + 0.0034176884219050407, + 0.012391758151352406, + 0.009305152110755444, + 0.005982113070785999, + 0.004026563838124275, + 0.005081285256892443, + 0.00531466631218791, + 0.0070083606988191605, + 0.003124572103843093, + 0.004856903105974197, + 0.006622401066124439, + 0.0037651918828487396, + 0.004202697891741991, + 0.00311480020172894, + 0.0033468645997345448, + 0.0033898476976901293, + 0.002148224273696542, + 0.0038840468041598797, + 0.0029926036950200796, + 0.00325861107558012, + 0.002339849481359124, + 0.00682347035035491, + 0.008435605093836784, + 0.0073244269005954266, + 0.004649489652365446, + 0.00825955905020237, + 0.008972927927970886, + 0.006293367128819227, + 0.005124460905790329, + 0.004488905891776085, + 0.007149714045226574, + 0.00454051373526454, + 0.004039863124489784, + 0.008399228565394878, + 0.008985169231891632, + 0.009271892718970776, + 0.003973378334194422, + 0.0033397956285625696, + 0.005259267054498196, + 0.0023412008304148912, + 0.0033013036008924246, + 0.004893103148788214, + 0.004275238141417503, + 0.005247732158750296, + 0.0027002335991710424, + 0.02076508291065693, + 0.014158068224787712, + 0.016344372183084488, + 0.012227393686771393, + 0.019466448575258255, + 0.07189580798149109, + 0.01453429739922285, + 0.031080903485417366, + 0.0206197127699852, + 0.024461276829242706, + 0.013714001514017582, + 0.007543641608208418, + 0.014033004641532898, + 0.033465947955846786, + 0.009207677096128464, + 0.027263006195425987 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.04468749836087227 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04871223121881485 + ], + "opposite_logprob_tokens": [ + -3.0487122535705566 + ], + "logprob_reward": -0.04871223121881485, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04871223121881485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19408783316612244, + 0.06791328638792038, + 0.023794401437044144, + 0.015280376188457012, + 0.02528020367026329, + 0.021559786051511765, + 0.08818937093019485, + 0.029384726658463478, + 0.21737340092658997, + 0.3171366751194 + ], + "credit_map": [ + 0.024809254333376884, + 0.028285222128033638, + 0.022418618202209473, + 0.01823192648589611, + 0.028142323717474937, + 0.03233081102371216, + 0.02219434268772602, + 0.017675328999757767, + 0.007892899215221405, + 0.009002533741295338, + 0.008068026043474674, + 0.006416257470846176, + 0.015721842646598816, + 0.009023440070450306, + 0.006873603910207748, + 0.004914682824164629, + 0.0032471318263560534, + 0.0026338426396250725, + 0.0030934056267142296, + 0.002004181267693639, + 0.004243315663188696, + 0.0032624490559101105, + 0.0030107744969427586, + 0.0022992996964603662, + 0.001745061483234167, + 0.0016274608206003904, + 0.0031130556017160416, + 0.001139978296123445, + 0.0025302644353359938, + 0.0019100605277344584, + 0.001817570417188108, + 0.0013969235587865114, + 0.002727048471570015, + 0.0025142766535282135, + 0.005350401159375906, + 0.0023458446376025677, + 0.0037905287463217974, + 0.0030110094230622053, + 0.003350136335939169, + 0.0021909582428634167, + 0.0027155489660799503, + 0.0029627755284309387, + 0.003014081623405218, + 0.001614202163182199, + 0.0034710729960352182, + 0.00356726604513824, + 0.002550958888605237, + 0.0016638791421428323, + 0.007002092897891998, + 0.010169832035899162, + 0.006914456374943256, + 0.005518685560673475, + 0.01722927950322628, + 0.017507174983620644, + 0.016170434653759003, + 0.007677410263568163, + 0.003170209703966975, + 0.004109944682568312, + 0.003423216287046671, + 0.0017783642979338765, + 0.00783496629446745, + 0.00456249387934804, + 0.002587753813713789, + 0.0019177749054506421, + 0.030823783949017525, + 0.020465277135372162, + 0.01853492297232151, + 0.012018918059766293, + 0.021586501970887184, + 0.06925440579652786, + 0.015658807009458542, + 0.029030755162239075, + 0.06178922578692436, + 0.07435979694128036, + 0.020879575982689857, + 0.013502828776836395, + 0.02743358723819256, + 0.056752242147922516, + 0.015966836363077164, + 0.046452563256025314 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.04871223121881485 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22560766339302063 + ], + "opposite_logprob_tokens": [ + -1.6006076335906982 + ], + "logprob_reward": -0.22560766339302063, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22560766339302063, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15528522431850433, + 0.09145431965589523, + 0.035153038799762726, + 0.019542083144187927, + 0.015056225471198559, + 0.00995996966958046, + 0.046705491840839386, + 0.047786515206098557, + 0.2595541179180145, + 0.3195030391216278 + ], + "credit_map": [ + 0.01611931249499321, + 0.021092617884278297, + 0.018327893689274788, + 0.016210950911045074, + 0.030446762219071388, + 0.021903902292251587, + 0.014515836723148823, + 0.016667932271957397, + 0.007178409956395626, + 0.009819850325584412, + 0.01225745864212513, + 0.00767820980399847, + 0.021370772272348404, + 0.015700234100222588, + 0.008227976970374584, + 0.009221396408975124, + 0.0025136449839919806, + 0.003709430107846856, + 0.004494938533753157, + 0.002231958322227001, + 0.010629775933921337, + 0.005440841428935528, + 0.0031952527351677418, + 0.002937193028628826, + 0.002293671714141965, + 0.003217288525775075, + 0.0021230466663837433, + 0.0018637941684573889, + 0.003432669211179018, + 0.002697265474125743, + 0.0017993361689150333, + 0.002115009119734168, + 0.0011467360891401768, + 0.0014837636845186353, + 0.002790832659229636, + 0.0015515144914388657, + 0.0024683764204382896, + 0.002076096134260297, + 0.001834447612054646, + 0.0017044574487954378, + 0.0015682002995163202, + 0.0022811912931501865, + 0.0009132975828833878, + 0.0007153066690079868, + 0.0013449953403323889, + 0.0012275753542780876, + 0.001149575226008892, + 0.0007598277879878879, + 0.004879784304648638, + 0.005034395027905703, + 0.0047807167284190655, + 0.0035612834617495537, + 0.009682411327958107, + 0.008970791473984718, + 0.005140081979334354, + 0.004656024742871523, + 0.005109832622110844, + 0.006973710376769304, + 0.005256684496998787, + 0.002914789831265807, + 0.010747977532446384, + 0.0069145034067332745, + 0.006335441023111343, + 0.0035335770808160305, + 0.06551897525787354, + 0.08005797117948532, + 0.013963796198368073, + 0.010646275244653225, + 0.022223595529794693, + 0.034494321793317795, + 0.011510059237480164, + 0.0211391132324934, + 0.06425388902425766, + 0.11133377999067307, + 0.011059416458010674, + 0.01550027634948492, + 0.02027866616845131, + 0.06850079447031021, + 0.013095983304083347, + 0.015480208210647106 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.22560766339302063 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194913864135742 + ], + "opposite_logprob_tokens": [ + -4.018195152282715 + ], + "logprob_reward": -0.018194913864135742, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194913864135742, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06907251477241516, + 0.026111628860235214, + 0.02737387642264366, + 0.021068787202239037, + 0.022683652117848396, + 0.024179771542549133, + 0.03771095722913742, + 0.0403083972632885, + 0.33501407504081726, + 0.39647629857063293 + ], + "credit_map": [ + 0.009491556324064732, + 0.009379709139466286, + 0.007546843495219946, + 0.006641309708356857, + 0.011728094890713692, + 0.012461775913834572, + 0.007227700669318438, + 0.004595525562763214, + 0.003696171334013343, + 0.0025073953438550234, + 0.0030873508658260107, + 0.0016574448673054576, + 0.0057733310386538506, + 0.005636744201183319, + 0.0020396269392222166, + 0.0017135649686679244, + 0.0041230181232094765, + 0.003719890024513006, + 0.002886540722101927, + 0.0028396642301231623, + 0.005736573599278927, + 0.00398351950570941, + 0.0023848158307373524, + 0.0016998552018776536, + 0.0031317814718931913, + 0.0038960007950663567, + 0.0027586822398006916, + 0.0021789574529975653, + 0.0029664805624634027, + 0.0023751258850097656, + 0.0019503971561789513, + 0.0018113611731678247, + 0.00241553271189332, + 0.003155283397063613, + 0.0037758720573037863, + 0.0020721254404634237, + 0.0036521409638226032, + 0.0026137246750295162, + 0.0030373872723430395, + 0.001961585832759738, + 0.002693799091503024, + 0.0034027588553726673, + 0.0035481879021972418, + 0.002285386435687542, + 0.0036841656547039747, + 0.003246332285925746, + 0.003432196332141757, + 0.001886943937279284, + 0.0036748461425304413, + 0.004828345496207476, + 0.003818415803834796, + 0.00246788514778018, + 0.005558174103498459, + 0.006351000629365444, + 0.007533346302807331, + 0.0034789445344358683, + 0.006722393445670605, + 0.009388288483023643, + 0.0037187226116657257, + 0.002928276313468814, + 0.006581462919712067, + 0.004305263515561819, + 0.0035860566422343254, + 0.0030779330991208553, + 0.07089149951934814, + 0.07559147477149963, + 0.023582175374031067, + 0.016092529520392418, + 0.030550025403499603, + 0.0639461949467659, + 0.018927043303847313, + 0.03543313965201378, + 0.09119507670402527, + 0.14372633397579193, + 0.02121494710445404, + 0.018979720771312714, + 0.025041822344064713, + 0.05598835274577141, + 0.01601707935333252, + 0.024312986060976982 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 0 + ], + "selected_reward": -0.018194913864135742 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026355832815170288 + ], + "opposite_logprob_tokens": [ + -3.651355743408203 + ], + "logprob_reward": -0.026355832815170288, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026355832815170288, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20139510929584503, + 0.052353039383888245, + 0.027760865166783333, + 0.025403589010238647, + 0.027509259060025215, + 0.02966953255236149, + 0.07795052230358124, + 0.060823362320661545, + 0.1827331781387329, + 0.3144015073776245 + ], + "credit_map": [ + 0.031557053327560425, + 0.021200288087129593, + 0.024218952283263206, + 0.015484679490327835, + 0.03040122799575329, + 0.035862889140844345, + 0.024911319836974144, + 0.017758691683411598, + 0.007135217543691397, + 0.006575895939022303, + 0.007283282000571489, + 0.004426386207342148, + 0.009396255947649479, + 0.006533384323120117, + 0.0060135736130177975, + 0.0049890452064573765, + 0.0038498586509376764, + 0.003187824273481965, + 0.003401912050321698, + 0.002290544332936406, + 0.005148754920810461, + 0.004052308388054371, + 0.002519953064620495, + 0.003309707622975111, + 0.0026302176993340254, + 0.0028103061486035585, + 0.0036990302614867687, + 0.0021810410544276237, + 0.0032164526637643576, + 0.004106637090444565, + 0.0031871343962848186, + 0.0035727699287235737, + 0.0027455331292003393, + 0.00354700512252748, + 0.0031734989024698734, + 0.0023316701408475637, + 0.0049049463123083115, + 0.0035992690827697515, + 0.004230200313031673, + 0.002977136056870222, + 0.00421792920678854, + 0.005682544317096472, + 0.0030737905763089657, + 0.0015797846717759967, + 0.005532088223844767, + 0.004965759348124266, + 0.0026222404558211565, + 0.0019953951705247164, + 0.007935122586786747, + 0.010285502299666405, + 0.006661758758127689, + 0.005016697105020285, + 0.014500001445412636, + 0.015823116526007652, + 0.01118050329387188, + 0.006547818426042795, + 0.006569474004209042, + 0.007756389211863279, + 0.008132017217576504, + 0.004887024872004986, + 0.012643866240978241, + 0.009260056540369987, + 0.0063300784677267075, + 0.005244455300271511, + 0.02075258269906044, + 0.018014414235949516, + 0.019824344664812088, + 0.01391181442886591, + 0.018155749887228012, + 0.052626822143793106, + 0.020779583603143692, + 0.018667882308363914, + 0.05279579758644104, + 0.09855629503726959, + 0.026918333023786545, + 0.01453055627644062, + 0.02046682871878147, + 0.06523922085762024, + 0.017098961398005486, + 0.018795516341924667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 8 + ], + "selected_reward": -0.026355832815170288 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, medium shot shows an ornate longsword lying in a shallow pool of blood on a cold, cracked stone floor. The sword's hilt, bound in worn black leather, is topped with a snarling wolf's head, adorned with small, gleaming ruby eyes. Its cross is an intricate grid of faded, thorn-like metal. The black, non-reflective steel of the blade is covered in deeply etched ancient runes. The focal point of the scene is a slow, mesmerizing horror: viscous crimson blood is deliberately absorbed into the blade. Thin streams of blood flow unnaturally from the puddle, as if drawn toward the etched symbols by an unseen force. As the liquid contacts each rune, the symbol comes alive, emitting a pulsating crimson glow that grows brighter with each sip. The blood vanished into the steel, leaving no residue behind, as if the weapon itself were drinking the water. These malevolent, rhythmic rays from the runes served as the primary light source, casting deep, flickering shadows and bathing the surrounding area in a hellish red glow, creating a tense and deeply sinister atmosphere as the dark ritual neared its conclusion.", + "seed": 4756, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4756_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.08524208753369748, + "vlm_logprob": -0.08524208753369748, + "vlm_yes_no_margin": 3.375, + "vlm_token_credit": -0.08524208753369748, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.066508948802948 + ], + "opposite_logprob_tokens": [ + -3.0665090084075928 + ], + "logprob_reward": -0.066508948802948, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.066508948802948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5119892954826355, + 0.055766236037015915, + 0.021237310022115707, + 0.023823663592338562, + 0.02491910569369793, + 0.025050094351172447, + 0.014422762207686901, + 0.025824470445513725, + 0.03847580403089523, + 0.2584911584854126 + ], + "credit_map": [ + 0.03658353537321091, + 0.05560757964849472, + 0.021208519116044044, + 0.018024373799562454, + 0.21353480219841003, + 0.031744975596666336, + 0.01146820466965437, + 0.1238173395395279, + 0.00423578592017293, + 0.004793210886418819, + 0.005658989306539297, + 0.0032527316361665726, + 0.015430369414389133, + 0.005325700622051954, + 0.0045435260981321335, + 0.012525922618806362, + 0.0027792376931756735, + 0.0021685336250811815, + 0.0033216734882444143, + 0.0013323461171239614, + 0.002882793778553605, + 0.003625561948865652, + 0.0018090958474203944, + 0.0033180685713887215, + 0.0038679679855704308, + 0.0023011930752545595, + 0.004712752532213926, + 0.001242459868080914, + 0.002429224783554673, + 0.0034254041966050863, + 0.0015066824853420258, + 0.004337979480624199, + 0.004266532603651285, + 0.0030423440039157867, + 0.003296511946246028, + 0.0025288949254900217, + 0.0031309877522289753, + 0.003810826689004898, + 0.0017020059749484062, + 0.003141002729535103, + 0.002940315054729581, + 0.0031789159402251244, + 0.002064700936898589, + 0.002793793100863695, + 0.0025390421506017447, + 0.0036287899129092693, + 0.0019073269795626402, + 0.005997208878397942, + 0.002768712118268013, + 0.0023403072264045477, + 0.0015221023932099342, + 0.0009729607845656574, + 0.0012318812077865005, + 0.0023647150956094265, + 0.001141440705396235, + 0.002080643316730857, + 0.003034718567505479, + 0.0038177778478711843, + 0.0018257489427924156, + 0.0012683637905865908, + 0.002853675512596965, + 0.0076714977622032166, + 0.0017825678223744035, + 0.003570119384676218, + 0.0041541773825883865, + 0.0037986477836966515, + 0.002263573696836829, + 0.0018936411943286657, + 0.002322980435565114, + 0.012856927700340748, + 0.00202212599106133, + 0.009163731709122658, + 0.017603233456611633, + 0.0897894874215126, + 0.009370333515107632, + 0.007679619826376438, + 0.01662568375468254, + 0.017891814932227135, + 0.013674607500433922, + 0.0858563780784607 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.066508948802948 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062118642032146454 + ], + "opposite_logprob_tokens": [ + -2.8121185302734375 + ], + "logprob_reward": -0.062118642032146454, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062118642032146454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6564262509346008, + 0.06514587998390198, + 0.04180055111646652, + 0.01931367628276348, + 0.028481706976890564, + 0.03637431189417839, + 0.0198063962161541, + 0.026564519852399826, + 0.029464635998010635, + 0.07662215083837509 + ], + "credit_map": [ + 0.043267153203487396, + 0.07228941470384598, + 0.024929581210017204, + 0.023212328553199768, + 0.2761665880680084, + 0.040876638144254684, + 0.014984128065407276, + 0.16070035099983215, + 0.005991509184241295, + 0.005178859457373619, + 0.004497853573411703, + 0.002451099455356598, + 0.019491039216518402, + 0.006921207532286644, + 0.003617675043642521, + 0.016996631398797035, + 0.005285503808408976, + 0.004539854358881712, + 0.005190991330891848, + 0.0034435500856488943, + 0.005859345197677612, + 0.008627170696854591, + 0.002905720379203558, + 0.005948410835117102, + 0.002803265117108822, + 0.0019142706878483295, + 0.0025745495222508907, + 0.0013764559989795089, + 0.002223411574959755, + 0.0033690629061311483, + 0.0019126411061733961, + 0.0031400176230818033, + 0.004367594141513109, + 0.003478929866105318, + 0.003214177442714572, + 0.002417551353573799, + 0.0037178555503487587, + 0.003911916632205248, + 0.002384708495810628, + 0.00498897023499012, + 0.004716104362159967, + 0.004609991330653429, + 0.003195237834006548, + 0.004049246199429035, + 0.00399848772212863, + 0.004904304631054401, + 0.0029272467363625765, + 0.007973687723279, + 0.0035220743156969547, + 0.0026665471959859133, + 0.0022594488691538572, + 0.0019966834224760532, + 0.0018078177236020565, + 0.002859363565221429, + 0.001957568107172847, + 0.0027368911541998386, + 0.0033597543369978666, + 0.003042670665308833, + 0.0017679319716989994, + 0.0015408310573548079, + 0.002600329229608178, + 0.007770252879709005, + 0.0016631269827485085, + 0.0048196204006671906, + 0.0036384023260325193, + 0.00299647799693048, + 0.002083961619064212, + 0.0017276655416935682, + 0.002494796644896269, + 0.009396384470164776, + 0.0022565829567611217, + 0.004870362114161253, + 0.008750488981604576, + 0.013115378096699715, + 0.005403689108788967, + 0.0027430891059339046, + 0.00786843616515398, + 0.007638839073479176, + 0.004559071734547615, + 0.026543155312538147 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.062118642032146454 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2814101576805115 + ], + "opposite_logprob_tokens": [ + -1.4064102172851562 + ], + "logprob_reward": -0.2814101576805115, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.2814101576805115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38208091259002686, + 0.23805762827396393, + 0.041087351739406586, + 0.054913230240345, + 0.07430846244096756, + 0.04225552827119827, + 0.01950540952384472, + 0.027087217196822166, + 0.06227776035666466, + 0.058426499366760254 + ], + "credit_map": [ + 0.026378203183412552, + 0.04219779744744301, + 0.01451838482171297, + 0.013483064249157906, + 0.1600758582353592, + 0.024619320407509804, + 0.009610314853489399, + 0.09119799733161926, + 0.016352279111742973, + 0.025301968678832054, + 0.009606026113033295, + 0.007660903502255678, + 0.0987897738814354, + 0.014471784234046936, + 0.006340262480080128, + 0.059534672647714615, + 0.006669754162430763, + 0.0052892714738845825, + 0.0050231958739459515, + 0.0027636794839054346, + 0.004317163489758968, + 0.005715581122785807, + 0.003182959044352174, + 0.008125749416649342, + 0.008445581421256065, + 0.005900072865188122, + 0.011819993145763874, + 0.0028859188314527273, + 0.005789464805275202, + 0.006173301488161087, + 0.0039424286223948, + 0.009956471621990204, + 0.012313800863921642, + 0.009299637749791145, + 0.008591484278440475, + 0.006076749414205551, + 0.008031590841710567, + 0.012960712425410748, + 0.006671363487839699, + 0.010363128036260605, + 0.004817490465939045, + 0.004884745925664902, + 0.0036233130376785994, + 0.0054900189861655235, + 0.0045171333476901054, + 0.0057013146579265594, + 0.00413353368639946, + 0.009087984450161457, + 0.0036425935104489326, + 0.0025171423330903053, + 0.002579492749646306, + 0.0018419265979900956, + 0.0019517289474606514, + 0.002732561668381095, + 0.0017587917391210794, + 0.002481174422428012, + 0.0032733886037021875, + 0.003943360410630703, + 0.0023332922719419003, + 0.001920362119562924, + 0.0037952738348394632, + 0.006728651002049446, + 0.0022743013687431812, + 0.002818591194227338, + 0.007732170168310404, + 0.00844248291105032, + 0.00475397938862443, + 0.004437868017703295, + 0.00566283892840147, + 0.016601741313934326, + 0.00528109772130847, + 0.009365583769977093, + 0.008964044973254204, + 0.006513160653412342, + 0.0073136864230036736, + 0.0033097986597567797, + 0.005286287050694227, + 0.006556548643857241, + 0.0054261088371276855, + 0.015056869946420193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.2814101576805115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036206187214702368 + ], + "opposite_logprob_tokens": [ + -5.628620624542236 + ], + "logprob_reward": -0.0036206187214702368, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036206187214702368, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3272612690925598, + 0.061923496425151825, + 0.07807628065347672, + 0.03326026350259781, + 0.04883401840925217, + 0.03220313414931297, + 0.01750057190656662, + 0.05018142610788345, + 0.049969688057899475, + 0.30078980326652527 + ], + "credit_map": [ + 0.02421468496322632, + 0.027507515624165535, + 0.014709056355059147, + 0.011297596618533134, + 0.1272130310535431, + 0.025792820379137993, + 0.00904942024499178, + 0.08747713267803192, + 0.0069962856359779835, + 0.007312086410820484, + 0.009531570598483086, + 0.0036579708103090525, + 0.007923654280602932, + 0.006257288157939911, + 0.006933376658707857, + 0.013311265967786312, + 0.009704031981527805, + 0.008512301370501518, + 0.014516697265207767, + 0.006725427228957415, + 0.008779059164226055, + 0.01092367060482502, + 0.006752558518201113, + 0.012162531726062298, + 0.005156959407031536, + 0.003267919644713402, + 0.0075523522682487965, + 0.0017343212384730577, + 0.0035211758222430944, + 0.004639456979930401, + 0.0033789861481636763, + 0.004009094554930925, + 0.00766028044745326, + 0.005859272088855505, + 0.006499025039374828, + 0.00282620289362967, + 0.005978976376354694, + 0.010277344845235348, + 0.003074733540415764, + 0.006658185739070177, + 0.005104260519146919, + 0.004622464068233967, + 0.003717998508363962, + 0.002697370480746031, + 0.0040982430800795555, + 0.005343703553080559, + 0.002406070241704583, + 0.004213028121739626, + 0.002014326862990856, + 0.0020386921241879463, + 0.0016993202734738588, + 0.002291678683832288, + 0.0018319317605346441, + 0.0030255226884037256, + 0.0018622096395120025, + 0.0027368911541998386, + 0.00586307467892766, + 0.0031283374410122633, + 0.00302962027490139, + 0.0021712137386202812, + 0.004702270962297916, + 0.012052700854837894, + 0.0028505849186331034, + 0.01638362556695938, + 0.0062995594926178455, + 0.006188498809933662, + 0.004224593751132488, + 0.005518733989447355, + 0.0047062961384654045, + 0.012903253547847271, + 0.0053022317588329315, + 0.004826523829251528, + 0.02249566838145256, + 0.12439773976802826, + 0.0076538631692528725, + 0.008267062716186047, + 0.02666325867176056, + 0.023861493915319443, + 0.015154707245528698, + 0.07229602336883545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.0036206187214702368 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012552070431411266 + ], + "opposite_logprob_tokens": [ + -4.387552261352539 + ], + "logprob_reward": -0.012552070431411266, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012552070431411266, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4950074553489685, + 0.07883810251951218, + 0.03145000711083412, + 0.025651969015598297, + 0.023191187530755997, + 0.026463888585567474, + 0.010928060859441757, + 0.04313068464398384, + 0.08086144179105759, + 0.18447716534137726 + ], + "credit_map": [ + 0.03268631175160408, + 0.05361789092421532, + 0.016347384080290794, + 0.017676150426268578, + 0.21128880977630615, + 0.029358558356761932, + 0.01055589783936739, + 0.12347649037837982, + 0.005430672317743301, + 0.008618353866040707, + 0.005814418662339449, + 0.002986747305840254, + 0.03038710542023182, + 0.005649171769618988, + 0.003221680410206318, + 0.016729960218071938, + 0.003298016032204032, + 0.004045965149998665, + 0.00463314913213253, + 0.0037490285467356443, + 0.0038558454252779484, + 0.0028536489699035883, + 0.00369994412176311, + 0.005314414389431477, + 0.003877905895933509, + 0.002139423508197069, + 0.004253364633768797, + 0.0021829763427376747, + 0.002437898889183998, + 0.003945224918425083, + 0.0026796262245625257, + 0.004135549534112215, + 0.003931297454982996, + 0.0027992487885057926, + 0.0033351401798427105, + 0.001970418496057391, + 0.0028956285677850246, + 0.003627354744821787, + 0.0019083158113062382, + 0.0027237844187766314, + 0.004473545588552952, + 0.0032469499856233597, + 0.0029784489888697863, + 0.0022856607101857662, + 0.0031744553707540035, + 0.00450020981952548, + 0.0021212405990809202, + 0.00368337775580585, + 0.0019841992761939764, + 0.0012909749057143927, + 0.0009389017359353602, + 0.0008380202343687415, + 0.0010349461808800697, + 0.0017286391230300069, + 0.0014114428777247667, + 0.0017009371658787131, + 0.0068161385133862495, + 0.009191969409584999, + 0.0026569939218461514, + 0.0018833652138710022, + 0.0031426777131855488, + 0.00471478933468461, + 0.002683382946997881, + 0.012041371315717697, + 0.0119891706854105, + 0.018757611513137817, + 0.006197680253535509, + 0.006267834454774857, + 0.006353937089443207, + 0.015895986929535866, + 0.0066111101768910885, + 0.008788113482296467, + 0.01784643530845642, + 0.0800141841173172, + 0.004831175785511732, + 0.0054903412237763405, + 0.02117036283016205, + 0.017312360927462578, + 0.00986661296337843, + 0.027945714071393013 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 8 + ], + "selected_reward": -0.012552070431411266 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static, medium shot shows an ornate longsword lying in a shallow pool of blood on a cold, cracked stone floor. The sword's hilt, bound in worn black leather, is topped with a snarling wolf's head, adorned with small, gleaming ruby eyes. Its cross is an intricate grid of faded, thorn-like metal. The black, non-reflective steel of the blade is covered in deeply etched ancient runes. The focal point of the scene is a slow, mesmerizing horror: viscous crimson blood is deliberately absorbed into the blade. Thin streams of blood flow unnaturally from the puddle, as if drawn toward the etched symbols by an unseen force. As the liquid contacts each rune, the symbol comes alive, emitting a pulsating crimson glow that grows brighter with each sip. The blood vanished into the steel, leaving no residue behind, as if the weapon itself were drinking the water. These malevolent, rhythmic rays from the runes served as the primary light source, casting deep, flickering shadows and bathing the surrounding area in a hellish red glow, creating a tense and deeply sinister atmosphere as the dark ritual neared its conclusion.", + "seed": 4757, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_1/index_4782_seed_4757_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.11158546730875969, + "vlm_logprob": -0.11158546730875969, + "vlm_yes_no_margin": 2.250000023841858, + "vlm_token_credit": -0.11158546730875969, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.13753822445869446 + ], + "opposite_logprob_tokens": [ + -2.137538194656372 + ], + "logprob_reward": -0.13753822445869446, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.13753822445869446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5384557843208313, + 0.22065559029579163, + 0.02639181725680828, + 0.02728983573615551, + 0.016674820333719254, + 0.017404191195964813, + 0.0810268223285675, + 0.028521478176116943, + 0.023389307782053947, + 0.02019030600786209 + ], + "credit_map": [ + 0.06400640308856964, + 0.06373660266399384, + 0.05755072459578514, + 0.046411313116550446, + 0.06537386775016785, + 0.08688872307538986, + 0.08571168780326843, + 0.06877647340297699, + 0.027785524725914, + 0.027772467583417892, + 0.021768102422356606, + 0.018555521965026855, + 0.027791233733296394, + 0.04052562266588211, + 0.02977764420211315, + 0.026679476723074913, + 0.0024219676852226257, + 0.0032652004156261683, + 0.0025832615792751312, + 0.0024917838163673878, + 0.0036349836736917496, + 0.004703556653112173, + 0.0038445587269961834, + 0.003446505405008793, + 0.0028963310178369284, + 0.00383641361258924, + 0.0024085030891001225, + 0.0025266625452786684, + 0.004125190898776054, + 0.004038422368466854, + 0.004396598320454359, + 0.0030617143493145704, + 0.0016649741446599364, + 0.0020904792472720146, + 0.0027079498395323753, + 0.0015556096332147717, + 0.001947146374732256, + 0.0025886590592563152, + 0.0022386203054338694, + 0.0018813820788636804, + 0.0023297143634408712, + 0.0022710503544658422, + 0.0018925105687230825, + 0.0016355827683582902, + 0.0021467360202223063, + 0.0023473002947866917, + 0.0029162615537643433, + 0.001865035155788064, + 0.00862629059702158, + 0.010334030725061893, + 0.0071326750330626965, + 0.009447754360735416, + 0.011492100544273853, + 0.01233018096536398, + 0.012605421245098114, + 0.009058373980224133, + 0.0035637924447655678, + 0.0036611096002161503, + 0.0036501812282949686, + 0.0031000033486634493, + 0.003174019046127796, + 0.004169253632426262, + 0.0035820070188492537, + 0.0036211099941283464, + 0.003433362813666463, + 0.002890421077609062, + 0.0022501195780932903, + 0.002012233017012477, + 0.002465465571731329, + 0.003171822987496853, + 0.0045892405323684216, + 0.002576642669737339, + 0.0035729436203837395, + 0.0028890655376017094, + 0.001908646896481514, + 0.001869188155978918, + 0.001747778500430286, + 0.0024424216244369745, + 0.0034297544043511152, + 0.0023305076174438 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.13753822445869446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0621364526450634 + ], + "opposite_logprob_tokens": [ + -2.81213641166687 + ], + "logprob_reward": -0.0621364526450634, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0621364526450634, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24174050986766815, + 0.0856621116399765, + 0.06452841311693192, + 0.08007983863353729, + 0.04081317409873009, + 0.11584242433309555, + 0.08307256549596786, + 0.13192696869373322, + 0.08645331114530563, + 0.0698806494474411 + ], + "credit_map": [ + 0.031553834676742554, + 0.03057768940925598, + 0.023040493950247765, + 0.01991238258779049, + 0.02650108002126217, + 0.035710833966732025, + 0.038504354655742645, + 0.03593984991312027, + 0.012127786874771118, + 0.010287594050168991, + 0.008661692962050438, + 0.009272617287933826, + 0.01059689186513424, + 0.013139083981513977, + 0.012617475353181362, + 0.008958977647125721, + 0.005894350353628397, + 0.006154583301395178, + 0.007654077839106321, + 0.007517621386796236, + 0.009684830904006958, + 0.011346879415214062, + 0.007953433319926262, + 0.008322643116116524, + 0.00936118047684431, + 0.00999742466956377, + 0.0072305575013160706, + 0.007744212634861469, + 0.009645281359553337, + 0.010956616140902042, + 0.015980714932084084, + 0.009163854643702507, + 0.004432925023138523, + 0.005273640155792236, + 0.006499812006950378, + 0.003857191652059555, + 0.004112593363970518, + 0.005823215004056692, + 0.005475128069519997, + 0.005338673479855061, + 0.01732337847352028, + 0.015854567289352417, + 0.007267385255545378, + 0.008448237553238869, + 0.013338391669094563, + 0.013782566413283348, + 0.02471952699124813, + 0.015108373947441578, + 0.007790947332978249, + 0.010298214852809906, + 0.010659086517989635, + 0.010369472205638885, + 0.008330745622515678, + 0.014263812452554703, + 0.011789862997829914, + 0.009570433758199215, + 0.018665803596377373, + 0.01432022824883461, + 0.013175169005990028, + 0.013889659196138382, + 0.01663776859641075, + 0.01852824166417122, + 0.01871314086019993, + 0.017996961250901222, + 0.011360647156834602, + 0.011195378378033638, + 0.00642202515155077, + 0.00895810779184103, + 0.0086427116766572, + 0.01125367358326912, + 0.017317933961749077, + 0.01130283996462822, + 0.010050761513411999, + 0.010373723693192005, + 0.006111562252044678, + 0.007317155133932829, + 0.006118094082921743, + 0.009561766870319843, + 0.01197519339621067, + 0.008372393436729908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.0621364526450634 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14285863935947418 + ], + "opposite_logprob_tokens": [ + -2.0178587436676025 + ], + "logprob_reward": -0.14285863935947418, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14285863935947418, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2755778729915619, + 0.25315535068511963, + 0.07243139296770096, + 0.06291546672582626, + 0.023541931062936783, + 0.0538545586168766, + 0.0287177637219429, + 0.11983072012662888, + 0.04064616560935974, + 0.06932873278856277 + ], + "credit_map": [ + 0.03789873793721199, + 0.03386620059609413, + 0.03115137480199337, + 0.025079458951950073, + 0.027893276885151863, + 0.040467653423547745, + 0.045596327632665634, + 0.03362485393881798, + 0.030520565807819366, + 0.02591395005583763, + 0.02473263628780842, + 0.023350263014435768, + 0.03291342779994011, + 0.04284154623746872, + 0.038004398345947266, + 0.03487854078412056, + 0.008006694726645947, + 0.008276693522930145, + 0.005976427346467972, + 0.006381532642990351, + 0.006734311580657959, + 0.013439679518342018, + 0.013597176410257816, + 0.010018878616392612, + 0.006473701447248459, + 0.007968608289957047, + 0.0048257349990308285, + 0.005953473038971424, + 0.009108561091125011, + 0.00919981300830841, + 0.012614683248102665, + 0.006770889274775982, + 0.002955613425001502, + 0.0034593248274177313, + 0.0025638602674007416, + 0.0021273086313158274, + 0.0029644460882991552, + 0.0033403264824301004, + 0.0034284526482224464, + 0.002702601021155715, + 0.004108683671802282, + 0.00613012071698904, + 0.006858163047581911, + 0.005740106105804443, + 0.006437715608626604, + 0.010816511698067188, + 0.007856038399040699, + 0.005907220300287008, + 0.0035427568946033716, + 0.0030732506420463324, + 0.0021974954288452864, + 0.003318675560876727, + 0.003927839454263449, + 0.005081005394458771, + 0.003936602268368006, + 0.003640137379989028, + 0.017357250675559044, + 0.012630583718419075, + 0.01236080564558506, + 0.012850264087319374, + 0.011560841463506222, + 0.01681768149137497, + 0.01917448081076145, + 0.017078805714845657, + 0.005039914045482874, + 0.004410421010106802, + 0.004336874932050705, + 0.004552973434329033, + 0.0038349092938005924, + 0.006303946953266859, + 0.007065036799758673, + 0.005102087743580341, + 0.006901117041707039, + 0.007973970845341682, + 0.007187629118561745, + 0.008540877141058445, + 0.007020480930805206, + 0.008105621673166752, + 0.015019610524177551, + 0.008579428307712078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.14285863935947418 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16040688753128052 + ], + "opposite_logprob_tokens": [ + -1.9104068279266357 + ], + "logprob_reward": -0.16040688753128052, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16040688753128052, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18343210220336914, + 0.06128197908401489, + 0.04300906881690025, + 0.18322163820266724, + 0.0602264329791069, + 0.04801829531788826, + 0.06009306013584137, + 0.13787591457366943, + 0.12596745789051056, + 0.09687399119138718 + ], + "credit_map": [ + 0.02859346941113472, + 0.02184736542403698, + 0.02253204584121704, + 0.018310047686100006, + 0.017611924558877945, + 0.025221969932317734, + 0.02661527507007122, + 0.02270001918077469, + 0.010535022243857384, + 0.009315998293459415, + 0.0068293726071715355, + 0.0044630891643464565, + 0.007617645431309938, + 0.00939358863979578, + 0.006574693601578474, + 0.006552574224770069, + 0.004667660687118769, + 0.004179001320153475, + 0.005227458197623491, + 0.003543898230418563, + 0.005838403478264809, + 0.01008014939725399, + 0.00452199624851346, + 0.004950504284352064, + 0.018698176369071007, + 0.02403617836534977, + 0.016080079600214958, + 0.01773586869239807, + 0.02434745989739895, + 0.02430756576359272, + 0.037948913872241974, + 0.020067395642399788, + 0.006268172990530729, + 0.0075343698263168335, + 0.00984280463308096, + 0.005463895853608847, + 0.008241686969995499, + 0.008100783452391624, + 0.0072866762056946754, + 0.007488040253520012, + 0.005117327906191349, + 0.004332592710852623, + 0.005669332109391689, + 0.004265124443918467, + 0.00558069022372365, + 0.007036122493445873, + 0.01102191861718893, + 0.004995185881853104, + 0.007565941195935011, + 0.006963774096220732, + 0.005452274344861507, + 0.0075485981069505215, + 0.00877821259200573, + 0.008700139820575714, + 0.007988527417182922, + 0.007095590699464083, + 0.02134046144783497, + 0.014462714083492756, + 0.012031182646751404, + 0.013069739565253258, + 0.015476887114346027, + 0.021394135430455208, + 0.02261551097035408, + 0.017485279589891434, + 0.01703452691435814, + 0.015854021534323692, + 0.008892577141523361, + 0.013711840845644474, + 0.013280784711241722, + 0.01622353307902813, + 0.024529818445444107, + 0.01644035615026951, + 0.013616742566227913, + 0.013793566264212132, + 0.009885035455226898, + 0.008734559640288353, + 0.00808640755712986, + 0.013486051931977272, + 0.018027445301413536, + 0.011244182474911213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.16040688753128052 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a low-angle, static, medium shot of an ornate longsword lying in a shallow pool of blood on a cracked stone floor?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the sword's hilt feature worn black leather bound with a snarling wolf's head topped with small gleaming ruby eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the blade made of black non-reflective steel covered in deeply etched ancient runes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does viscous crimson blood flow from the puddle into the blade, causing the runes to emit a pulsating crimson glow that grows brighter?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Do the glowing runes serve as the primary light source, casting deep flickering shadows and bathing the scene in a hellish red glow?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05498713254928589 + ], + "opposite_logprob_tokens": [ + -2.9299871921539307 + ], + "logprob_reward": -0.05498713254928589, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05498713254928589, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13053244352340698, + 0.07776568084955215, + 0.0664130225777626, + 0.12975940108299255, + 0.03220720589160919, + 0.17425653338432312, + 0.10534266382455826, + 0.1052917018532753, + 0.08132678270339966, + 0.0971045270562172 + ], + "credit_map": [ + 0.020094484090805054, + 0.018863698467612267, + 0.012620505876839161, + 0.010642033070325851, + 0.01619974710047245, + 0.021943535655736923, + 0.013665856793522835, + 0.016502585262060165, + 0.008704663254320621, + 0.009914522059261799, + 0.005878747906535864, + 0.006343076471239328, + 0.007635015062987804, + 0.014247354120016098, + 0.01413424126803875, + 0.010908063501119614, + 0.00814240612089634, + 0.009426413103938103, + 0.005124418996274471, + 0.004900025669485331, + 0.008392434567213058, + 0.013544113375246525, + 0.0075967744924128056, + 0.00928643997758627, + 0.013542557135224342, + 0.017600663006305695, + 0.010890475474298, + 0.011941933073103428, + 0.0173533596098423, + 0.018784483894705772, + 0.023450544103980064, + 0.01619539223611355, + 0.003518348792567849, + 0.0040082684718072414, + 0.004211248829960823, + 0.0027481629513204098, + 0.0031258983071893454, + 0.003972368780523539, + 0.006589117459952831, + 0.004033789969980717, + 0.027017122134566307, + 0.02553226239979267, + 0.014381674118340015, + 0.011884985491633415, + 0.019583487883210182, + 0.023421015590429306, + 0.031876519322395325, + 0.020559480413794518, + 0.011360343545675278, + 0.009769192896783352, + 0.012077870778739452, + 0.017517361789941788, + 0.010992536321282387, + 0.011710744351148605, + 0.019142957404255867, + 0.012771662324666977, + 0.011291902512311935, + 0.011162471026182175, + 0.008954750373959541, + 0.010503997094929218, + 0.011959359981119633, + 0.020485732704401016, + 0.018802620470523834, + 0.012130863033235073, + 0.011206562630832195, + 0.01003041211515665, + 0.005986360367387533, + 0.009149848483502865, + 0.007551901508122683, + 0.009469625540077686, + 0.018803313374519348, + 0.009128764271736145, + 0.013158967718482018, + 0.01624334789812565, + 0.010209151543676853, + 0.008012136444449425, + 0.010481096804141998, + 0.012666042894124985, + 0.015626635402441025, + 0.010707154870033264 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.05498713254928589 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_tensor([86, 86])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_tensor([86, 86])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e177f9c1d5f8d57557f41b10aa1784ebc468fc7f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_tensor([86, 86])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, eye-level wide shot captures the atmosphere of a crowded wrestling ring, where the camera pans rapidly to track the powerful, muscular wrestler. He wears a vibrant pink and gold lucha libre-style mask adorned with intricate gold filigree, perfectly matching his pink leggings and knee-high boots, which gleam under the intense overhead spotlights. He powerfully grasps the top rope with both hands, harnessing its elastic tension to propel himself across the blue canvas mat with explosive speed, charging towards his opponent, who is dressed in simple black shorts. In one fluid movement, the masked wrestler bends down, grabs his opponent, and with a burst of raw power, hoisting him onto his broad shoulders. He firmly plants his feet and spins with a swift, controlled movement, creating a blur of pink and gold. The move culminates in a decisive arc as he flips his opponent onto his back, sending him crashing to the mat with a thunderous impact that reverberates throughout the ring. Without hesitation, the masked wrestler drops down, hooking his opponent's leg. The referee, dressed in a classic black and white striped shirt, immediately dropped to one knee beside them, his expression focused, and decisively struck the mat with his palm three times, sharp, rhythmic slaps that interrupted the roar of the cheering, out-of-focus crowd and sealed the victory in a piece of energetic, triumphant drama.", + "seed": 8997, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_8997_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.09477403052151204, + "vlm_logprob": -0.09477403052151204, + "vlm_yes_no_margin": 2.624999976158142, + "vlm_token_credit": -0.09477403052151204, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14424119889736176 + ], + "opposite_logprob_tokens": [ + -2.1442410945892334 + ], + "logprob_reward": -0.14424119889736176, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.14424119889736176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1338660567998886, + 0.15871110558509827, + 0.06174944341182709, + 0.07854317873716354, + 0.07934820652008057, + 0.17996452748775482, + 0.04313821718096733, + 0.15798507630825043, + 0.07980381697416306, + 0.02689046412706375 + ], + "credit_map": [ + 0.01947796158492565, + 0.013270119205117226, + 0.02409466914832592, + 0.01753651723265648, + 0.015487362630665302, + 0.012924524024128914, + 0.01498525869101286, + 0.016089633107185364, + 0.03056960552930832, + 0.01427460927516222, + 0.01646602526307106, + 0.017301009967923164, + 0.013611333444714546, + 0.02190590836107731, + 0.01818370446562767, + 0.026398897171020508, + 0.008914143778383732, + 0.006239346694201231, + 0.009130207821726799, + 0.006148046348243952, + 0.009133812040090561, + 0.007267843931913376, + 0.006934509612619877, + 0.007981527596712112, + 0.010801003314554691, + 0.012944133952260017, + 0.010533972643315792, + 0.010952374897897243, + 0.008789081126451492, + 0.011460382491350174, + 0.007696101907640696, + 0.005366121418774128, + 0.01115415245294571, + 0.015544790774583817, + 0.006043402943760157, + 0.009529258124530315, + 0.008961520157754421, + 0.009009943343698978, + 0.012684664689004421, + 0.006420467048883438, + 0.017769968137145042, + 0.012683892622590065, + 0.05287633463740349, + 0.026859572157263756, + 0.010755130089819431, + 0.018098099157214165, + 0.016043802723288536, + 0.024877725169062614, + 0.0038446851540356874, + 0.005287014879286289, + 0.0047557842917740345, + 0.00531478226184845, + 0.005172188393771648, + 0.010061657056212425, + 0.004888170398771763, + 0.0038139319512993097, + 0.019103698432445526, + 0.01195608638226986, + 0.014898543246090412, + 0.026869429275393486, + 0.014081694185733795, + 0.0269829910248518, + 0.023125268518924713, + 0.02096734009683132, + 0.00984496995806694, + 0.030806196853518486, + 0.0036648684181272984, + 0.010836219415068626, + 0.0067015052773058414, + 0.003949237987399101, + 0.005552468355745077, + 0.008448348380625248, + 0.003124173730611801, + 0.0017542669083923101, + 0.0024107745848596096, + 0.0025723979342728853, + 0.0031966473907232285, + 0.0036144584883004427, + 0.004400447942316532, + 0.005817293655127287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 7 + ], + "selected_reward": -0.14424119889736176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295917630195618 + ], + "opposite_logprob_tokens": [ + -2.237959146499634 + ], + "logprob_reward": -0.11295917630195618, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295917630195618, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06655103713274002, + 0.07814381271600723, + 0.0949031263589859, + 0.08818341791629791, + 0.20836105942726135, + 0.16214407980442047, + 0.056349460035562515, + 0.03802914917469025, + 0.17304912209510803, + 0.03428566828370094 + ], + "credit_map": [ + 0.003956212196499109, + 0.005894087255001068, + 0.009955954737961292, + 0.013981514610350132, + 0.009808456525206566, + 0.0037592663429677486, + 0.011037876829504967, + 0.008157677948474884, + 0.010919316671788692, + 0.005242335144430399, + 0.010594518855214119, + 0.013034294359385967, + 0.007120461203157902, + 0.008568896912038326, + 0.009094398468732834, + 0.013569600880146027, + 0.01247014943510294, + 0.010517960414290428, + 0.013077066279947758, + 0.009864597581326962, + 0.015085263177752495, + 0.011834919452667236, + 0.010819331742823124, + 0.011233843863010406, + 0.017366304993629456, + 0.015058053657412529, + 0.012694267556071281, + 0.012088288553059101, + 0.007319225929677486, + 0.008287126198410988, + 0.00900482852011919, + 0.006365334149450064, + 0.018283318728208542, + 0.02125857211649418, + 0.028684942051768303, + 0.02829609625041485, + 0.018682075664401054, + 0.016558535397052765, + 0.04396538808941841, + 0.03263213858008385, + 0.01473868079483509, + 0.01202473882585764, + 0.025022095069289207, + 0.018915368244051933, + 0.01810011826455593, + 0.028459642082452774, + 0.017782622948288918, + 0.027100831270217896, + 0.00415392592549324, + 0.0059051853604614735, + 0.006165303289890289, + 0.010100421495735645, + 0.0077561247162520885, + 0.00547264376655221, + 0.009939761832356453, + 0.006856098305433989, + 0.0034734485670924187, + 0.002249200129881501, + 0.00274768378585577, + 0.007023426704108715, + 0.0027236940804868937, + 0.005227867979556322, + 0.008188503794372082, + 0.006395325064659119, + 0.019560672342777252, + 0.055120326578617096, + 0.012325170449912548, + 0.028831928968429565, + 0.01243804395198822, + 0.009730597026646137, + 0.02017032727599144, + 0.014872075989842415, + 0.0038914461620151997, + 0.0024346583522856236, + 0.0022414790000766516, + 0.005231161136180162, + 0.005067554302513599, + 0.0033571168314665556, + 0.007204069755971432, + 0.004858187399804592 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 5 + ], + "selected_reward": -0.11295917630195618 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038107167929410934 + ], + "opposite_logprob_tokens": [ + -3.28810715675354 + ], + "logprob_reward": -0.038107167929410934, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038107167929410934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048595793545246124, + 0.039127420634031296, + 0.03866790235042572, + 0.03743797913193703, + 0.059350740164518356, + 0.05447426065802574, + 0.4833771288394928, + 0.04938438907265663, + 0.07601393759250641, + 0.1135704442858696 + ], + "credit_map": [ + 0.0047522131353616714, + 0.005216154735535383, + 0.007049809210002422, + 0.006940871011465788, + 0.008315104059875011, + 0.0033740543294698, + 0.00622602179646492, + 0.006721567362546921, + 0.005151758436113596, + 0.003767425427213311, + 0.004555113613605499, + 0.005345379002392292, + 0.005480232648551464, + 0.004441029857844114, + 0.005900655407458544, + 0.004485825542360544, + 0.00507857371121645, + 0.0034741763956844807, + 0.005160144995898008, + 0.004078196361660957, + 0.006364330183714628, + 0.005875019822269678, + 0.004096942022442818, + 0.0045405165292322636, + 0.008120657876133919, + 0.005495850928127766, + 0.0036097969859838486, + 0.0025439755991101265, + 0.005190988536924124, + 0.005226156674325466, + 0.0037607778795063496, + 0.0034897769801318645, + 0.005649789702147245, + 0.006407438311725855, + 0.006079609040170908, + 0.0067898305132985115, + 0.005944326985627413, + 0.005225013941526413, + 0.012588579207658768, + 0.010666157118976116, + 0.004968032706528902, + 0.0041405213996768, + 0.012330468744039536, + 0.008413340896368027, + 0.005202183499932289, + 0.005527103319764137, + 0.0058899843133986, + 0.00800262950360775, + 0.03525291010737419, + 0.048109836876392365, + 0.054159101098775864, + 0.10881910473108292, + 0.0482184998691082, + 0.03464839607477188, + 0.09551721811294556, + 0.058652088046073914, + 0.003988140728324652, + 0.00525294803082943, + 0.004670416470617056, + 0.01159573718905449, + 0.003281625220552087, + 0.004454443231225014, + 0.010513963177800179, + 0.0056271180510520935, + 0.008569509722292423, + 0.022873586043715477, + 0.004144259262830019, + 0.010610776953399181, + 0.006149681285023689, + 0.005386906210333109, + 0.005020163953304291, + 0.01325905229896307, + 0.016515590250492096, + 0.007869013585150242, + 0.011275405995547771, + 0.021855931729078293, + 0.011287976987659931, + 0.010817798785865307, + 0.014000164344906807, + 0.01994856260716915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 8 + ], + "selected_reward": -0.038107167929410934 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16037003695964813 + ], + "opposite_logprob_tokens": [ + -1.9103699922561646 + ], + "logprob_reward": -0.16037003695964813, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16037003695964813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15904738008975983, + 0.06609708815813065, + 0.04549166187644005, + 0.06403624266386032, + 0.06894383579492569, + 0.1112888902425766, + 0.19092795252799988, + 0.08298734575510025, + 0.11580196022987366, + 0.09537768363952637 + ], + "credit_map": [ + 0.01477607898414135, + 0.011102304793894291, + 0.022865520790219307, + 0.02621896006166935, + 0.022309569641947746, + 0.009918265044689178, + 0.021966392174363136, + 0.02989029325544834, + 0.005552150774747133, + 0.005968710873275995, + 0.00785969290882349, + 0.012018316425383091, + 0.011759697459638119, + 0.0051396372728049755, + 0.007614762522280216, + 0.010184122249484062, + 0.0059987623244524, + 0.004286383278667927, + 0.006051407661288977, + 0.005906241480261087, + 0.006461724638938904, + 0.004872748162597418, + 0.005203124135732651, + 0.006711270194500685, + 0.010781862773001194, + 0.010779816657304764, + 0.006271666381508112, + 0.005934913642704487, + 0.00778987118974328, + 0.009889542125165462, + 0.007087402511388063, + 0.005501167383044958, + 0.007961980067193508, + 0.007853393442928791, + 0.009133185259997845, + 0.009640623815357685, + 0.008324871771037579, + 0.005598684307187796, + 0.012595510110259056, + 0.007835585623979568, + 0.01319923996925354, + 0.008419922553002834, + 0.027846219018101692, + 0.01145893894135952, + 0.009996325708925724, + 0.01452604029327631, + 0.01227454375475645, + 0.013567665591835976, + 0.014484662562608719, + 0.019025355577468872, + 0.022460805252194405, + 0.04304653778672218, + 0.018775353208184242, + 0.013277726247906685, + 0.03768093138933182, + 0.022176571190357208, + 0.010907399468123913, + 0.006979604717344046, + 0.008194472640752792, + 0.012628698721528053, + 0.005487785208970308, + 0.009404142387211323, + 0.014291136525571346, + 0.015094111673533916, + 0.018341034650802612, + 0.022839201614260674, + 0.00876932218670845, + 0.015095300041139126, + 0.015462561510503292, + 0.008693787269294262, + 0.011147907935082912, + 0.015452844090759754, + 0.010636558756232262, + 0.008536606095731258, + 0.011658395640552044, + 0.009806754998862743, + 0.013293632306158543, + 0.019668035209178925, + 0.013919441029429436, + 0.007858258672058582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.16037003695964813 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01819257251918316 + ], + "opposite_logprob_tokens": [ + -4.018192768096924 + ], + "logprob_reward": -0.01819257251918316, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01819257251918316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11050733178853989, + 0.12830804288387299, + 0.06453189998865128, + 0.029093822464346886, + 0.0511888712644577, + 0.1070384532213211, + 0.24615494906902313, + 0.12410837411880493, + 0.09576385468244553, + 0.04330442473292351 + ], + "credit_map": [ + 0.014316295273602009, + 0.010878358036279678, + 0.015010617673397064, + 0.013665101490914822, + 0.01595352590084076, + 0.011061959899961948, + 0.01276776846498251, + 0.016853705048561096, + 0.023381853476166725, + 0.009357953444123268, + 0.02092875726521015, + 0.018559549003839493, + 0.00967059750109911, + 0.013544868677854538, + 0.014844837598502636, + 0.018019622191786766, + 0.01098975632339716, + 0.006922197062522173, + 0.007386534009128809, + 0.005977493245154619, + 0.006495116278529167, + 0.010564355179667473, + 0.006035926751792431, + 0.01016052532941103, + 0.00652682688087225, + 0.005870534107089043, + 0.002370333531871438, + 0.0021878499537706375, + 0.003165655070915818, + 0.00467847241088748, + 0.0022459409665316343, + 0.0020482121035456657, + 0.006426421459764242, + 0.008848534896969795, + 0.005608460400253534, + 0.0067795319482684135, + 0.004142628982663155, + 0.005162586458027363, + 0.007903087884187698, + 0.006317620165646076, + 0.010376794263720512, + 0.007943195290863514, + 0.028419476002454758, + 0.013211690820753574, + 0.007506841327995062, + 0.013623539358377457, + 0.010840706527233124, + 0.015116208232939243, + 0.01793859899044037, + 0.025333914905786514, + 0.03483681380748749, + 0.048204828053712845, + 0.023607827723026276, + 0.018260614946484566, + 0.04597585275769234, + 0.03199649229645729, + 0.010890264064073563, + 0.010614892467856407, + 0.011806418187916279, + 0.02517036348581314, + 0.007527740206569433, + 0.01908825896680355, + 0.026263795793056488, + 0.012746642343699932, + 0.010021467693150043, + 0.02417678013443947, + 0.005389622412621975, + 0.013151773251593113, + 0.007188467308878899, + 0.006463181227445602, + 0.008537853136658669, + 0.020834708586335182, + 0.007824542000889778, + 0.0024330741725862026, + 0.0029172785580158234, + 0.004208574537187815, + 0.007722201757133007, + 0.008184831589460373, + 0.004795062821358442, + 0.00521885696798563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 1, + 7 + ], + "selected_reward": -0.01819257251918316 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, eye-level wide shot captures the atmosphere of a crowded wrestling ring, where the camera pans rapidly to track the powerful, muscular wrestler. He wears a vibrant pink and gold lucha libre-style mask adorned with intricate gold filigree, perfectly matching his pink leggings and knee-high boots, which gleam under the intense overhead spotlights. He powerfully grasps the top rope with both hands, harnessing its elastic tension to propel himself across the blue canvas mat with explosive speed, charging towards his opponent, who is dressed in simple black shorts. In one fluid movement, the masked wrestler bends down, grabs his opponent, and with a burst of raw power, hoisting him onto his broad shoulders. He firmly plants his feet and spins with a swift, controlled movement, creating a blur of pink and gold. The move culminates in a decisive arc as he flips his opponent onto his back, sending him crashing to the mat with a thunderous impact that reverberates throughout the ring. Without hesitation, the masked wrestler drops down, hooking his opponent's leg. The referee, dressed in a classic black and white striped shirt, immediately dropped to one knee beside them, his expression focused, and decisively struck the mat with his palm three times, sharp, rhythmic slaps that interrupted the roar of the cheering, out-of-focus crowd and sealed the victory in a piece of energetic, triumphant drama.", + "seed": 8998, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_8998_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10298004187643528, + "vlm_logprob": -0.10298004187643528, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.10298004187643528, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2439797967672348 + ], + "opposite_logprob_tokens": [ + -1.618979811668396 + ], + "logprob_reward": -0.2439797967672348, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2439797967672348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08424163609743118, + 0.5702048540115356, + 0.048880089074373245, + 0.03324417769908905, + 0.06937476247549057, + 0.02845025807619095, + 0.014421643689274788, + 0.017283417284488678, + 0.11219604313373566, + 0.02170308120548725 + ], + "credit_map": [ + 0.01164102554321289, + 0.014996482990682125, + 0.00492847478017211, + 0.005678972695022821, + 0.02568831853568554, + 0.0075749545358121395, + 0.008933468721807003, + 0.004799936898052692, + 0.013096384704113007, + 0.05242963880300522, + 0.011460334062576294, + 0.01781521923840046, + 0.24031436443328857, + 0.19797928631305695, + 0.02462439052760601, + 0.012485229410231113, + 0.006175227928906679, + 0.006414229515939951, + 0.00499714957550168, + 0.013131811283528805, + 0.005184961948543787, + 0.003926191944628954, + 0.0047434489242732525, + 0.004307068884372711, + 0.006502960808575153, + 0.00711953267455101, + 0.0026690836530178785, + 0.0025940698105841875, + 0.004879619926214218, + 0.0036506704054772854, + 0.0028586203698068857, + 0.0029696188867092133, + 0.009741129353642464, + 0.024831846356391907, + 0.0050883241929113865, + 0.005023668520152569, + 0.005518355872482061, + 0.011480302549898624, + 0.0035236645489931107, + 0.004167472943663597, + 0.00490256492048502, + 0.002912292256951332, + 0.004050306510180235, + 0.003884380916133523, + 0.0029551065526902676, + 0.004075057804584503, + 0.0028210070449858904, + 0.0028495416045188904, + 0.0018263760721310973, + 0.001421708962880075, + 0.0022411521058529615, + 0.0014714575372636318, + 0.0012250067666172981, + 0.0018746325513347983, + 0.0018852923531085253, + 0.002476017689332366, + 0.002328367903828621, + 0.00145864172372967, + 0.002528541721403599, + 0.00231582741253078, + 0.0019870304968208075, + 0.0021068560890853405, + 0.0023473533801734447, + 0.002210797043517232, + 0.0026546556036919355, + 0.04707165062427521, + 0.001260210876353085, + 0.0018236159812659025, + 0.005795870907604694, + 0.049619950354099274, + 0.0021932863164693117, + 0.0017768017714843154, + 0.003623552620410919, + 0.0024298762436956167, + 0.0024823916610330343, + 0.0030602384358644485, + 0.003279227064922452, + 0.003643918549641967, + 0.001927497098222375, + 0.001256380695849657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.2439797967672348 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430944561958313 + ], + "opposite_logprob_tokens": [ + -3.1680943965911865 + ], + "logprob_reward": -0.0430944561958313, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430944561958313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.023126132786273956, + 0.8338074088096619, + 0.01868029497563839, + 0.008770660497248173, + 0.009388181380927563, + 0.011065419763326645, + 0.01538011059165001, + 0.0050985608249902725, + 0.06483621895313263, + 0.009846989996731281 + ], + "credit_map": [ + 0.0027075165417045355, + 0.003531284863129258, + 0.0023660531733185053, + 0.002334354445338249, + 0.003369296668097377, + 0.0030839606188237667, + 0.003983700182288885, + 0.0017499671084806323, + 0.018087875097990036, + 0.07655195891857147, + 0.016331646591424942, + 0.022729534655809402, + 0.3576294183731079, + 0.2928577959537506, + 0.033301983028650284, + 0.016317153349518776, + 0.001752750831656158, + 0.001998153980821371, + 0.0017577748512849212, + 0.002442972268909216, + 0.0034495696891099215, + 0.0027301376685500145, + 0.0024895945098251104, + 0.00205934327095747, + 0.0014202028978616, + 0.0014071711339056492, + 0.0007485421956516802, + 0.0008811768493615091, + 0.0012644949601963162, + 0.00130936480127275, + 0.0008286989177577198, + 0.000911009032279253, + 0.0010110285365954041, + 0.002082431921735406, + 0.0014105763984844089, + 0.0009384859586134553, + 0.0007796887075528502, + 0.0013260599225759506, + 0.0009988152887672186, + 0.0008410952286794782, + 0.002532962476834655, + 0.0010042053181678057, + 0.0016637170920148492, + 0.0009622556390240788, + 0.001045778626576066, + 0.0012412116630002856, + 0.0013662386918440461, + 0.0012490509543567896, + 0.0012717775534838438, + 0.00117868569213897, + 0.0025280837435275316, + 0.0018228882690891623, + 0.001194274635054171, + 0.001548868021927774, + 0.0026067784056067467, + 0.0032287552021443844, + 0.00046606126124970615, + 0.0004543703398667276, + 0.0006101374165154994, + 0.0006218701601028442, + 0.0008883057162165642, + 0.0007125664269551635, + 0.000734842149540782, + 0.0006104076164774597, + 0.0016509690321981907, + 0.02639913000166416, + 0.0009245559340342879, + 0.0010792241664603353, + 0.0034540577325969934, + 0.02873278222978115, + 0.001470424933359027, + 0.0011250844690948725, + 0.0011468478478491306, + 0.0011228887597098947, + 0.001041124458424747, + 0.0012934289406985044, + 0.0014553252840414643, + 0.0015651878202334046, + 0.0015300310915336013, + 0.0006921567837707698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 0 + ], + "selected_reward": -0.0430944561958313 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370063751935959 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370063751935959, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370063751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04550172761082649, + 0.5436251163482666, + 0.01692638173699379, + 0.023072483018040657, + 0.020806528627872467, + 0.02420768514275551, + 0.02052975632250309, + 0.03317015618085861, + 0.20877504348754883, + 0.06338514387607574 + ], + "credit_map": [ + 0.004837594460695982, + 0.006009772885590792, + 0.00498499721288681, + 0.00473381532356143, + 0.010233588516712189, + 0.004619694780558348, + 0.007306965067982674, + 0.0027752965688705444, + 0.01318153366446495, + 0.049797121435403824, + 0.011018350720405579, + 0.01478959433734417, + 0.23311208188533783, + 0.19006934762001038, + 0.021389594301581383, + 0.010267469100654125, + 0.0017904478590935469, + 0.0018525319173932076, + 0.0017545241862535477, + 0.003980060573667288, + 0.0020252603571861982, + 0.002007732167840004, + 0.0022397313732653856, + 0.00127609190531075, + 0.0034575490280985832, + 0.004498927854001522, + 0.0023150381166487932, + 0.002092607319355011, + 0.002859759144484997, + 0.0028933254070580006, + 0.0021285638213157654, + 0.0028267106972634792, + 0.0025786347687244415, + 0.0062842289917171, + 0.0024965086486190557, + 0.0011113999644294381, + 0.0016363030299544334, + 0.003121598158031702, + 0.0012411045609042048, + 0.0023367470130324364, + 0.0033425043802708387, + 0.0025948553811758757, + 0.003212364623323083, + 0.002907049609348178, + 0.0031464549247175455, + 0.0032148745376616716, + 0.0033395295031368732, + 0.002450051950290799, + 0.00309395813383162, + 0.0019251112826168537, + 0.0032325906213372946, + 0.0021436905954033136, + 0.002142903860658407, + 0.001781319035217166, + 0.0027767405845224857, + 0.0034334424417465925, + 0.004172844812273979, + 0.0030118331778794527, + 0.005922280717641115, + 0.0038900983054190874, + 0.0045846858993172646, + 0.004181270021945238, + 0.004384520463645458, + 0.0030226223170757294, + 0.006791302002966404, + 0.08306316286325455, + 0.0028388528153300285, + 0.00512600177899003, + 0.011148500256240368, + 0.09005416184663773, + 0.005916350055485964, + 0.0038366944063454866, + 0.007530453614890575, + 0.007483831141144037, + 0.010177829302847385, + 0.008627966977655888, + 0.008700069040060043, + 0.009810228832066059, + 0.007639836519956589, + 0.0034149170387536287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.03370063751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17991873621940613 + ], + "opposite_logprob_tokens": [ + -1.8049187660217285 + ], + "logprob_reward": -0.17991873621940613, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17991873621940613, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07382302731275558, + 0.5427160263061523, + 0.016691727563738823, + 0.024130916222929955, + 0.03278861939907074, + 0.020343119278550148, + 0.028737301006913185, + 0.0372709184885025, + 0.09857875108718872, + 0.12491962313652039 + ], + "credit_map": [ + 0.006377574522048235, + 0.009274078533053398, + 0.011118714697659016, + 0.008914049714803696, + 0.008098803460597992, + 0.008627324365079403, + 0.015492516569793224, + 0.005919969640672207, + 0.011385559104382992, + 0.0493675097823143, + 0.010951081290841103, + 0.015232011675834656, + 0.23213456571102142, + 0.1908946931362152, + 0.022167136892676353, + 0.01058341097086668, + 0.001913438900373876, + 0.0016707493923604488, + 0.002005815738812089, + 0.002892202464863658, + 0.0019492276478558779, + 0.002532364334911108, + 0.0016386252827942371, + 0.002089305315166712, + 0.003886695019900799, + 0.004934133496135473, + 0.0024888929910957813, + 0.00234550004824996, + 0.0027319462969899178, + 0.0027248982805758715, + 0.0023150278721004725, + 0.002703822450712323, + 0.003459554398432374, + 0.007311934605240822, + 0.005287951324135065, + 0.002528212498873472, + 0.0028338676784187555, + 0.004601645283401012, + 0.00387533288449049, + 0.002890123287215829, + 0.0024814060889184475, + 0.0016519775381311774, + 0.0029281049501150846, + 0.002852430334314704, + 0.0023160395212471485, + 0.0022091481368988752, + 0.003169008530676365, + 0.002735002664849162, + 0.004148698411881924, + 0.0025812415406107903, + 0.004745245911180973, + 0.003886304097250104, + 0.0019178378861397505, + 0.0019596207421272993, + 0.0038047439884394407, + 0.005693607032299042, + 0.0040779151022434235, + 0.0029380267951637506, + 0.004304270260035992, + 0.003972518723458052, + 0.005281495861709118, + 0.005019580479711294, + 0.006408026907593012, + 0.005269086919724941, + 0.008114524185657501, + 0.02634459361433983, + 0.004200398921966553, + 0.009717084467411041, + 0.008723320439457893, + 0.027286654338240623, + 0.009487560018897057, + 0.004704620689153671, + 0.019688047468662262, + 0.013141119852662086, + 0.013788333162665367, + 0.015087073668837547, + 0.020627029240131378, + 0.02151242457330227, + 0.012926580384373665, + 0.008149009197950363 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.17991873621940613 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014206582680344582 + ], + "opposite_logprob_tokens": [ + -4.264206409454346 + ], + "logprob_reward": -0.014206582680344582, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014206582680344582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13752777874469757, + 0.294577032327652, + 0.009472809731960297, + 0.03129834681749344, + 0.029398983344435692, + 0.015249121934175491, + 0.015845030546188354, + 0.016923213377594948, + 0.43554043769836426, + 0.01416720449924469 + ], + "credit_map": [ + 0.014965418726205826, + 0.022277509793639183, + 0.008219555020332336, + 0.01107484009116888, + 0.048107318580150604, + 0.012293226085603237, + 0.013121341355144978, + 0.007468573749065399, + 0.006336993537843227, + 0.026822246611118317, + 0.005954909138381481, + 0.008460761979222298, + 0.1254223734140396, + 0.10382143408060074, + 0.011956111527979374, + 0.005802231375128031, + 0.0012794240610674024, + 0.000986132537946105, + 0.001185700879432261, + 0.0018595875008031726, + 0.0009173620492219925, + 0.0011690566316246986, + 0.00109478656668216, + 0.0009807611349970102, + 0.004890786949545145, + 0.006731980945914984, + 0.0035239015705883503, + 0.0025204189587384462, + 0.0033915790263563395, + 0.0036014937795698643, + 0.002796469023451209, + 0.0038417172618210316, + 0.003770710900425911, + 0.009344996884465218, + 0.004730678163468838, + 0.0017883701948449016, + 0.0017745733493939042, + 0.004528268706053495, + 0.0017701378092169762, + 0.0016912496648728848, + 0.0025583894457668066, + 0.002035465557128191, + 0.0021988109219819307, + 0.0017574273515492678, + 0.0018210895359516144, + 0.0013787400675937533, + 0.0019697346724569798, + 0.001529466244392097, + 0.002280081622302532, + 0.0013489588163793087, + 0.002363201230764389, + 0.0018918070709332824, + 0.0013437860179692507, + 0.0015956283314153552, + 0.002624392043799162, + 0.002397174946963787, + 0.0012349337339401245, + 0.0012061633169651031, + 0.0024599707685410976, + 0.0017916308715939522, + 0.0017893005860969424, + 0.002406733576208353, + 0.0030524551402777433, + 0.0029820275958627462, + 0.008168574422597885, + 0.18535244464874268, + 0.0028823495376855135, + 0.006249561440199614, + 0.02398489974439144, + 0.19687975943088531, + 0.006064352113753557, + 0.005958510097116232, + 0.0017233082326129079, + 0.0015443108277395368, + 0.0017736883601173759, + 0.001835868926718831, + 0.0021345054265111685, + 0.002247757278382778, + 0.001965094590559602, + 0.0009426715550944209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.014206582680344582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, eye-level wide shot captures the atmosphere of a crowded wrestling ring, where the camera pans rapidly to track the powerful, muscular wrestler. He wears a vibrant pink and gold lucha libre-style mask adorned with intricate gold filigree, perfectly matching his pink leggings and knee-high boots, which gleam under the intense overhead spotlights. He powerfully grasps the top rope with both hands, harnessing its elastic tension to propel himself across the blue canvas mat with explosive speed, charging towards his opponent, who is dressed in simple black shorts. In one fluid movement, the masked wrestler bends down, grabs his opponent, and with a burst of raw power, hoisting him onto his broad shoulders. He firmly plants his feet and spins with a swift, controlled movement, creating a blur of pink and gold. The move culminates in a decisive arc as he flips his opponent onto his back, sending him crashing to the mat with a thunderous impact that reverberates throughout the ring. Without hesitation, the masked wrestler drops down, hooking his opponent's leg. The referee, dressed in a classic black and white striped shirt, immediately dropped to one knee beside them, his expression focused, and decisively struck the mat with his palm three times, sharp, rhythmic slaps that interrupted the roar of the cheering, out-of-focus crowd and sealed the victory in a piece of energetic, triumphant drama.", + "seed": 8999, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_8999_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1711096443235874, + "vlm_logprob": -0.1711096443235874, + "vlm_yes_no_margin": 2.1, + "vlm_token_credit": -0.1711096443235874, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3738487958908081 + ], + "opposite_logprob_tokens": [ + -1.248848795890808 + ], + "logprob_reward": -0.3738487958908081, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3738487958908081, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12321135401725769, + 0.054526638239622116, + 0.21778172254562378, + 0.24833494424819946, + 0.09799934178590775, + 0.08779773116111755, + 0.05692962557077408, + 0.05050148814916611, + 0.031046662479639053, + 0.03187048062682152 + ], + "credit_map": [ + 0.014855089597404003, + 0.016280455514788628, + 0.013793742284178734, + 0.010902401991188526, + 0.010127978399395943, + 0.025854313746094704, + 0.0075883446261286736, + 0.023809028789401054, + 0.009351104497909546, + 0.010698722675442696, + 0.004926871974021196, + 0.004306284710764885, + 0.0090186707675457, + 0.006745600141584873, + 0.005072740837931633, + 0.0044066403061151505, + 0.060463376343250275, + 0.01227027177810669, + 0.025241106748580933, + 0.013346563093364239, + 0.019918616861104965, + 0.05945258587598801, + 0.012291258201003075, + 0.014797934330999851, + 0.023828361183404922, + 0.023050488904118538, + 0.05103451386094093, + 0.04167930781841278, + 0.014547440223395824, + 0.02881419099867344, + 0.0312076173722744, + 0.03417303413152695, + 0.010357278399169445, + 0.015488220378756523, + 0.007516141049563885, + 0.009329404681921005, + 0.02111712656915188, + 0.013034342788159847, + 0.013922622427344322, + 0.00723420362919569, + 0.011951569467782974, + 0.01422654464840889, + 0.01714751124382019, + 0.013417289592325687, + 0.005425443407148123, + 0.013302545063197613, + 0.005798182915896177, + 0.006528644822537899, + 0.007136065512895584, + 0.00752972811460495, + 0.00872307363897562, + 0.009683662094175816, + 0.005804738495498896, + 0.007216118276119232, + 0.005250647664070129, + 0.00558558851480484, + 0.0064560407772660255, + 0.006326073314994574, + 0.008557199500501156, + 0.006880417466163635, + 0.004346847999840975, + 0.007304596249014139, + 0.005063824355602264, + 0.005566487088799477, + 0.00560114486142993, + 0.004222417715936899, + 0.002705031307414174, + 0.0019937085453420877, + 0.005800236016511917, + 0.005546350497752428, + 0.0030165540520101786, + 0.0021612215787172318, + 0.002854069462046027, + 0.004074444063007832, + 0.002572495024651289, + 0.002245505340397358, + 0.0048749325796961784, + 0.007139616645872593, + 0.004185816738754511, + 0.0039235991425812244 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.3738487958908081 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11296162754297256 + ], + "opposite_logprob_tokens": [ + -2.237961530685425 + ], + "logprob_reward": -0.11296162754297256, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11296162754297256, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10614202171564102, + 0.06731285899877548, + 0.15649336576461792, + 0.27118363976478577, + 0.0927673727273941, + 0.09346796572208405, + 0.05655490234494209, + 0.043115850538015366, + 0.07428260892629623, + 0.038679398596286774 + ], + "credit_map": [ + 0.010613463819026947, + 0.011710462160408497, + 0.012723083607852459, + 0.006088451482355595, + 0.012383638881146908, + 0.026660801842808723, + 0.006787319667637348, + 0.019174806773662567, + 0.009561214596033096, + 0.012197131291031837, + 0.006569060496985912, + 0.006171908229589462, + 0.010062530636787415, + 0.007080279756337404, + 0.0072583700530231, + 0.008412369526922703, + 0.03805707395076752, + 0.012876927852630615, + 0.016731973737478256, + 0.011660889722406864, + 0.016024364158511162, + 0.03325612097978592, + 0.014172105118632317, + 0.013713912107050419, + 0.03388916701078415, + 0.038259610533714294, + 0.05546038970351219, + 0.02578749693930149, + 0.02346954122185707, + 0.03607643023133278, + 0.020675860345363617, + 0.03756515681743622, + 0.00822596438229084, + 0.010468744672834873, + 0.006814411375671625, + 0.008899430744349957, + 0.022096697241067886, + 0.013885504566133022, + 0.015550916083157063, + 0.006825712975114584, + 0.010558195412158966, + 0.012190144509077072, + 0.015239477157592773, + 0.01393676083534956, + 0.00745950173586607, + 0.011989515274763107, + 0.009668175131082535, + 0.012426196597516537, + 0.00679584126919508, + 0.007304409518837929, + 0.007035119459033012, + 0.008562849834561348, + 0.007921009324491024, + 0.007344848942011595, + 0.005619092844426632, + 0.005971732549369335, + 0.007519111968576908, + 0.005458928178995848, + 0.004581400193274021, + 0.003454620949923992, + 0.007414325140416622, + 0.007477947510778904, + 0.003658393630757928, + 0.003551125293597579, + 0.008660412393510342, + 0.009164701215922832, + 0.01912657544016838, + 0.007279711309820414, + 0.006435404531657696, + 0.005824096500873566, + 0.008596371859312057, + 0.009195348247885704, + 0.0034797354601323605, + 0.004189684055745602, + 0.004723336547613144, + 0.0030694049783051014, + 0.004946728702634573, + 0.006713277194648981, + 0.007073864806443453, + 0.004483368713408709 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.11296162754297256 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270623356103897 + ], + "opposite_logprob_tokens": [ + -2.1270623207092285 + ], + "logprob_reward": -0.1270623356103897, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270623356103897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051235131919384, + 0.0441177599132061, + 0.2460673302412033, + 0.16057506203651428, + 0.11310488730669022, + 0.08205702155828476, + 0.035792481154203415, + 0.03854900598526001, + 0.1530744433403015, + 0.07542689889669418 + ], + "credit_map": [ + 0.005508820526301861, + 0.00780186103656888, + 0.007643138058483601, + 0.003590088803321123, + 0.005963529925793409, + 0.00942307710647583, + 0.0049925209023058414, + 0.006312092766165733, + 0.0059609971940517426, + 0.009402590803802013, + 0.0037808939814567566, + 0.004231269937008619, + 0.005511424969881773, + 0.0050909132696688175, + 0.006157805677503347, + 0.003981860354542732, + 0.09276929497718811, + 0.016322875395417213, + 0.01981154829263687, + 0.009639785625040531, + 0.014823492616415024, + 0.0693930834531784, + 0.012429418042302132, + 0.010877816006541252, + 0.019924571737647057, + 0.026035405695438385, + 0.03122224472463131, + 0.015007763169705868, + 0.015991857275366783, + 0.018214473500847816, + 0.014846855774521828, + 0.019331878051161766, + 0.008296074345707893, + 0.015265780501067638, + 0.009129160083830357, + 0.007044861558824778, + 0.026779288426041603, + 0.022746305912733078, + 0.015139829367399216, + 0.008703579194843769, + 0.011230377480387688, + 0.010480236262083054, + 0.019495712593197823, + 0.013569828122854233, + 0.005095054395496845, + 0.010382979176938534, + 0.005188960116356611, + 0.006613872479647398, + 0.004245501011610031, + 0.005301820579916239, + 0.004126886837184429, + 0.006311841309070587, + 0.004088966175913811, + 0.004247120115906, + 0.0038006994873285294, + 0.0036696463357657194, + 0.004339283332228661, + 0.004342556931078434, + 0.003645497839897871, + 0.005579834803938866, + 0.005782900378108025, + 0.006402580067515373, + 0.004287474788725376, + 0.004168872255831957, + 0.020275620743632317, + 0.02002628520131111, + 0.028187867254018784, + 0.014974970370531082, + 0.02210155874490738, + 0.01296095922589302, + 0.018862469121813774, + 0.01568470522761345, + 0.006452234461903572, + 0.008530802093446255, + 0.006242379080504179, + 0.005874814931303263, + 0.010212810710072517, + 0.01619764044880867, + 0.012517468072474003, + 0.009398744441568851 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 8 + ], + "selected_reward": -0.1270623356103897 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22559958696365356 + ], + "opposite_logprob_tokens": [ + -1.6005995273590088 + ], + "logprob_reward": -0.22559958696365356, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22559958696365356, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08375599980354309, + 0.04795904457569122, + 0.2510667145252228, + 0.1195688396692276, + 0.05411224067211151, + 0.1978490650653839, + 0.050713080912828445, + 0.058680836111307144, + 0.06882356852293015, + 0.06747061759233475 + ], + "credit_map": [ + 0.008189188316464424, + 0.012872803956270218, + 0.008880142122507095, + 0.006234500557184219, + 0.010854778811335564, + 0.015687910839915276, + 0.010489165782928467, + 0.010547501035034657, + 0.007002083584666252, + 0.006290972698479891, + 0.00805946346372366, + 0.004739934112876654, + 0.006123923230916262, + 0.0059410990215837955, + 0.004233052954077721, + 0.0055685085244476795, + 0.07667798548936844, + 0.0170342568308115, + 0.020036978647112846, + 0.010569094680249691, + 0.021497685462236404, + 0.08058304339647293, + 0.010921877808868885, + 0.013745786622166634, + 0.014126397669315338, + 0.022624652832746506, + 0.019021762534976006, + 0.011534638702869415, + 0.012513572350144386, + 0.016203774139285088, + 0.011278390884399414, + 0.012265635654330254, + 0.0054327561520040035, + 0.006536491215229034, + 0.006435634568333626, + 0.0035144039429724216, + 0.012664142064750195, + 0.007009944412857294, + 0.007876361720263958, + 0.004642500542104244, + 0.0247573833912611, + 0.025041766464710236, + 0.05508730188012123, + 0.031861405819654465, + 0.012124271132051945, + 0.02415132336318493, + 0.010269305668771267, + 0.014556284062564373, + 0.006023947149515152, + 0.007708986289799213, + 0.006312251556664705, + 0.009133338928222656, + 0.00572830019518733, + 0.006159220822155476, + 0.0037725556176155806, + 0.0058744740672409534, + 0.008169879205524921, + 0.007432825397700071, + 0.006349567323923111, + 0.00637883460149169, + 0.009905521757900715, + 0.008320258930325508, + 0.006578636355698109, + 0.005545308347791433, + 0.007754565682262182, + 0.006194308400154114, + 0.0132548613473773, + 0.011338688433170319, + 0.008335387334227562, + 0.0064866929315030575, + 0.008968651294708252, + 0.006490405648946762, + 0.011584805324673653, + 0.00866530928760767, + 0.008705697022378445, + 0.0055423821322619915, + 0.010558194480836391, + 0.00930766575038433, + 0.005399329587817192, + 0.007707230746746063 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 3 + ], + "selected_reward": -0.22559958696365356 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016075875610113144 + ], + "opposite_logprob_tokens": [ + -4.14107608795166 + ], + "logprob_reward": -0.016075875610113144, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016075875610113144, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25830039381980896, + 0.03950286656618118, + 0.14485977590084076, + 0.0447046235203743, + 0.04454639181494713, + 0.12417411804199219, + 0.11136417835950851, + 0.16863895952701569, + 0.043785154819488525, + 0.02012355625629425 + ], + "credit_map": [ + 0.02665541134774685, + 0.01304609701037407, + 0.027609707787632942, + 0.014713307842612267, + 0.026314539834856987, + 0.08057937771081924, + 0.012354888953268528, + 0.05702706053853035, + 0.008575174026191235, + 0.0041794199496507645, + 0.007244337350130081, + 0.002415794413536787, + 0.003121117129921913, + 0.004961997736245394, + 0.002770449034869671, + 0.006234578788280487, + 0.02355913631618023, + 0.0065323784947395325, + 0.034685783088207245, + 0.0071525126695632935, + 0.010554675944149494, + 0.047306668013334274, + 0.006675345823168755, + 0.008393282070755959, + 0.005701335612684488, + 0.0046666949056088924, + 0.006571480073034763, + 0.00676595838740468, + 0.003312893444672227, + 0.00702678132802248, + 0.005116280168294907, + 0.005543198902159929, + 0.005164166912436485, + 0.009172219783067703, + 0.004325976595282555, + 0.004478853195905685, + 0.00704475911334157, + 0.007554009091109037, + 0.003773608012124896, + 0.003032797947525978, + 0.0147933978587389, + 0.01642371341586113, + 0.021074336022138596, + 0.023574525490403175, + 0.010031148791313171, + 0.016645057126879692, + 0.011520675383508205, + 0.010111257433891296, + 0.015713931992650032, + 0.01656334660947323, + 0.01812838949263096, + 0.019932745024561882, + 0.008798348717391491, + 0.012556944973766804, + 0.0100958701223135, + 0.009574607945978642, + 0.021154388785362244, + 0.024354394525289536, + 0.02681337110698223, + 0.015567094087600708, + 0.023246806114912033, + 0.02604900486767292, + 0.015865007415413857, + 0.015588891692459583, + 0.005108692217618227, + 0.005997578147798777, + 0.007771574426442385, + 0.005539518781006336, + 0.004879934713244438, + 0.0048513333313167095, + 0.00569350179284811, + 0.003943020477890968, + 0.001616207999177277, + 0.002438499592244625, + 0.003472353331744671, + 0.0018422359135001898, + 0.0026616959366947412, + 0.0037715379148721695, + 0.002695196308195591, + 0.0016258289106190205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.016075875610113144 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, eye-level wide shot captures the atmosphere of a crowded wrestling ring, where the camera pans rapidly to track the powerful, muscular wrestler. He wears a vibrant pink and gold lucha libre-style mask adorned with intricate gold filigree, perfectly matching his pink leggings and knee-high boots, which gleam under the intense overhead spotlights. He powerfully grasps the top rope with both hands, harnessing its elastic tension to propel himself across the blue canvas mat with explosive speed, charging towards his opponent, who is dressed in simple black shorts. In one fluid movement, the masked wrestler bends down, grabs his opponent, and with a burst of raw power, hoisting him onto his broad shoulders. He firmly plants his feet and spins with a swift, controlled movement, creating a blur of pink and gold. The move culminates in a decisive arc as he flips his opponent onto his back, sending him crashing to the mat with a thunderous impact that reverberates throughout the ring. Without hesitation, the masked wrestler drops down, hooking his opponent's leg. The referee, dressed in a classic black and white striped shirt, immediately dropped to one knee beside them, his expression focused, and decisively struck the mat with his palm three times, sharp, rhythmic slaps that interrupted the roar of the cheering, out-of-focus crowd and sealed the victory in a piece of energetic, triumphant drama.", + "seed": 9000, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_10/index_86_seed_9000_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06443592961877584, + "vlm_logprob": -0.06443592961877584, + "vlm_yes_no_margin": 2.975, + "vlm_token_credit": -0.06443592961877584, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10447235405445099 + ], + "opposite_logprob_tokens": [ + -2.4794723987579346 + ], + "logprob_reward": -0.10447235405445099, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10447235405445099, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1878875494003296, + 0.20523999631404877, + 0.07921529561281204, + 0.06410132348537445, + 0.07031015306711197, + 0.17247872054576874, + 0.08922437578439713, + 0.049093570560216904, + 0.0628180131316185, + 0.019630976021289825 + ], + "credit_map": [ + 0.00864800252020359, + 0.011638517491519451, + 0.06224560737609863, + 0.07171203196048737, + 0.007424334529787302, + 0.010297714732587337, + 0.007328527979552746, + 0.008592813275754452, + 0.019044838845729828, + 0.014322715811431408, + 0.06611905992031097, + 0.02909361757338047, + 0.043706778436899185, + 0.01587693765759468, + 0.009645857848227024, + 0.0074302139692008495, + 0.009489932097494602, + 0.011292725801467896, + 0.005161619745194912, + 0.006958693731576204, + 0.02387487143278122, + 0.0077876681461930275, + 0.007869421504437923, + 0.0067803687416017056, + 0.01763264089822769, + 0.0060860393568873405, + 0.004978574812412262, + 0.00784272886812687, + 0.006748639978468418, + 0.010623195208609104, + 0.004742225632071495, + 0.005447286646813154, + 0.02066813036799431, + 0.023008497431874275, + 0.003446124028414488, + 0.008500495925545692, + 0.0037312067579478025, + 0.0028372083324939013, + 0.0049094706773757935, + 0.003209026763215661, + 0.025469981133937836, + 0.03417573124170303, + 0.014680334366858006, + 0.017475675791502, + 0.01347518153488636, + 0.035846710205078125, + 0.011935746297240257, + 0.019419372081756592, + 0.011273214593529701, + 0.01016153022646904, + 0.021349193528294563, + 0.011801387183368206, + 0.011498495005071163, + 0.006153278052806854, + 0.008832977153360844, + 0.0081543093547225, + 0.008240441791713238, + 0.005380615592002869, + 0.0038986271247267723, + 0.0047346316277980804, + 0.00854471605271101, + 0.009055788628757, + 0.005195945501327515, + 0.004042807966470718, + 0.007618294563144445, + 0.0075503126718103886, + 0.005575863644480705, + 0.007233850657939911, + 0.010656988248229027, + 0.013253036886453629, + 0.005542516242712736, + 0.005387153942137957, + 0.0025257489178329706, + 0.004872581921517849, + 0.0014663971960544586, + 0.002244796836748719, + 0.0028349305503070354, + 0.0023709172382950783, + 0.001811139634810388, + 0.0015044667525216937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.10447235405445099 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899352163076401 + ], + "opposite_logprob_tokens": [ + -2.578993558883667 + ], + "logprob_reward": -0.07899352163076401, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899352163076401, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03999614343047142, + 0.5187329053878784, + 0.07146260887384415, + 0.05369128659367561, + 0.05650627613067627, + 0.1014559268951416, + 0.06710667908191681, + 0.03310297429561615, + 0.04100098833441734, + 0.01694418303668499 + ], + "credit_map": [ + 0.00583010446280241, + 0.006455227732658386, + 0.0030909846536815166, + 0.003056582296267152, + 0.005167058669030666, + 0.007050242740660906, + 0.003960881847888231, + 0.005385060328990221, + 0.04924951493740082, + 0.06414636969566345, + 0.11804673820734024, + 0.04416671395301819, + 0.1544424295425415, + 0.04972444102168083, + 0.022038819268345833, + 0.016917914152145386, + 0.008461765944957733, + 0.005864104721695185, + 0.015272011049091816, + 0.006715758703649044, + 0.015066484920680523, + 0.007044085767120123, + 0.0074930754490196705, + 0.005545326042920351, + 0.010859834030270576, + 0.006898357532918453, + 0.004940827842801809, + 0.003630024380981922, + 0.006501738913357258, + 0.011205156333744526, + 0.00457443343475461, + 0.005080918315798044, + 0.01578078791499138, + 0.014544048346579075, + 0.0045312028378248215, + 0.004046289715915918, + 0.0031444947235286236, + 0.0037223880644887686, + 0.0039143068715929985, + 0.006822759285569191, + 0.015196630731225014, + 0.023949863389134407, + 0.007510246708989143, + 0.009172216057777405, + 0.00697445310652256, + 0.016643479466438293, + 0.007315813563764095, + 0.014693229459226131, + 0.009772795252501965, + 0.009647213853895664, + 0.005806319881230593, + 0.00763031467795372, + 0.010429124347865582, + 0.007647990714758635, + 0.0064758919179439545, + 0.009697034023702145, + 0.004437513183802366, + 0.004987181164324284, + 0.003116225590929389, + 0.003706294111907482, + 0.005612140987068415, + 0.004072398412972689, + 0.004029123578220606, + 0.003142100526019931, + 0.004820044618099928, + 0.009126975201070309, + 0.004206601995974779, + 0.004269972909241915, + 0.005774537567049265, + 0.0067492383532226086, + 0.00288867368362844, + 0.003164941444993019, + 0.0015805892180651426, + 0.002498233923688531, + 0.0015155067667365074, + 0.0020992460194975138, + 0.002863344270735979, + 0.002138706622645259, + 0.0026378033217042685, + 0.0016107553383335471 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 2 + ], + "selected_reward": -0.07899352163076401 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048654552549123764 + ], + "opposite_logprob_tokens": [ + -3.048654556274414 + ], + "logprob_reward": -0.048654552549123764, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048654552549123764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051405876874923706, + 0.5855305790901184, + 0.0962209552526474, + 0.08961262553930283, + 0.01658977009356022, + 0.05047941952943802, + 0.059586867690086365, + 0.011516190133988857, + 0.027187012135982513, + 0.011870641261339188 + ], + "credit_map": [ + 0.005439997185021639, + 0.005802469793707132, + 0.00927458144724369, + 0.012949036434292793, + 0.0039841593243181705, + 0.006862396374344826, + 0.0028472926933318377, + 0.004245948977768421, + 0.053920961916446686, + 0.05976313352584839, + 0.15534433722496033, + 0.06383958458900452, + 0.16004137694835663, + 0.05213860794901848, + 0.023313766345381737, + 0.017168885096907616, + 0.007004886865615845, + 0.008771845139563084, + 0.015493329614400864, + 0.005813978146761656, + 0.03947780281305313, + 0.010125021450221539, + 0.006436972878873348, + 0.003097132546827197, + 0.012164008803665638, + 0.00841526873409748, + 0.006012528669089079, + 0.008088093250989914, + 0.013181515969336033, + 0.017979424446821213, + 0.015950003638863564, + 0.00782178808003664, + 0.0018771557370200753, + 0.002292933641001582, + 0.0016673763748258352, + 0.0036347119603306055, + 0.0013089252170175314, + 0.001395408296957612, + 0.002847429132089019, + 0.001565831247717142, + 0.007068257313221693, + 0.007575119845569134, + 0.005484148394316435, + 0.004326528403908014, + 0.00551211042329669, + 0.012590905651450157, + 0.003829328343272209, + 0.004093024414032698, + 0.008326171897351742, + 0.007020806893706322, + 0.008480925112962723, + 0.00960399117320776, + 0.00776556134223938, + 0.004072792828083038, + 0.006009131669998169, + 0.008307491429150105, + 0.0017567327013239264, + 0.0013981035444885492, + 0.0013461570488288999, + 0.0016216340009123087, + 0.001528397435322404, + 0.0018068323843181133, + 0.0009915397968143225, + 0.0010667945025488734, + 0.00316622038371861, + 0.007800386752933264, + 0.0024044979363679886, + 0.003207716392353177, + 0.0023066720459610224, + 0.003851095447316766, + 0.0025945683009922504, + 0.0018558576703071594, + 0.0011392362648621202, + 0.0030753023456782103, + 0.0009680343209765851, + 0.0012808430474251509, + 0.0014024663250893354, + 0.0016147064743563533, + 0.0013194752391427755, + 0.001070577185600996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 3 + ], + "selected_reward": -0.048654552549123764 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898129522800446 + ], + "opposite_logprob_tokens": [ + -2.578981399536133 + ], + "logprob_reward": -0.07898129522800446, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898129522800446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06965722888708115, + 0.07988885045051575, + 0.01611984707415104, + 0.036773547530174255, + 0.022679081186652184, + 0.6248286962509155, + 0.06642759591341019, + 0.034099288284778595, + 0.022702105343341827, + 0.026823733001947403 + ], + "credit_map": [ + 0.005519732367247343, + 0.00560328084975481, + 0.022125987336039543, + 0.01847921684384346, + 0.003917619585990906, + 0.005052622873336077, + 0.00426690885797143, + 0.004691869951784611, + 0.0072099920362234116, + 0.00835566595196724, + 0.020302923396229744, + 0.009867328219115734, + 0.018781201913952827, + 0.006434183102101088, + 0.006196543108671904, + 0.002741016447544098, + 0.001839752309024334, + 0.002118620090186596, + 0.001143416971899569, + 0.0025014649145305157, + 0.0036540720611810684, + 0.0014594212407246232, + 0.002011100761592388, + 0.001391999190673232, + 0.006699154619127512, + 0.004211218096315861, + 0.003574950620532036, + 0.0017554875230416656, + 0.0056547317653894424, + 0.00951758585870266, + 0.0030843429267406464, + 0.002276078797876835, + 0.007116072345525026, + 0.0064080990850925446, + 0.0012326734140515327, + 0.001872286549769342, + 0.0012748059816658497, + 0.001340076676569879, + 0.0019835918210446835, + 0.0014514765935018659, + 0.0977286696434021, + 0.13320285081863403, + 0.05566837266087532, + 0.04802922531962395, + 0.050031181424856186, + 0.1309986114501953, + 0.039709270000457764, + 0.06946053355932236, + 0.00946938805282116, + 0.006982137914747, + 0.012345673516392708, + 0.009939699433743954, + 0.008224810473620892, + 0.004193160217255354, + 0.008406909182667732, + 0.006865821313112974, + 0.0052129123359918594, + 0.003068485762923956, + 0.004136416595429182, + 0.003509407164528966, + 0.005018325522542, + 0.006624871399253607, + 0.002226619515568018, + 0.004302252549678087, + 0.0026928004808723927, + 0.005522598512470722, + 0.002047751797363162, + 0.0026227247435599566, + 0.0022338784765452147, + 0.0038570661563426256, + 0.0021558981388807297, + 0.0015693898312747478, + 0.003390613943338394, + 0.003971205558627844, + 0.0031479562167078257, + 0.0034762395080178976, + 0.003499525599181652, + 0.003450465854257345, + 0.0035093065816909075, + 0.0023784232325851917 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.07898129522800446 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level wide shot of a crowded wrestling ring?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera pan rapidly to track a muscular wrestler wearing a pink and gold lucha libre mask?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the wrestler perform a move where he hoists his opponent onto his shoulders, spins, and flips the opponent onto their back?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the referee visible wearing a black and white striped shirt, dropping to one knee, and striking the mat three times?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background crowd visible as out-of-focus and cheering?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011077924631536007 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011077924631536007, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011077924631536007, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19715282320976257, + 0.20961974561214447, + 0.13382256031036377, + 0.04611610248684883, + 0.019125480204820633, + 0.21713818609714508, + 0.03657466173171997, + 0.06363718211650848, + 0.059131406247615814, + 0.017681853845715523 + ], + "credit_map": [ + 0.010675120167434216, + 0.015060707926750183, + 0.05846862122416496, + 0.07473158836364746, + 0.008646630682051182, + 0.011499777436256409, + 0.008875391446053982, + 0.009194985963404179, + 0.021348273381590843, + 0.02740122750401497, + 0.057431410998106, + 0.021747859194874763, + 0.05113759636878967, + 0.014001067727804184, + 0.009395853616297245, + 0.007156468462198973, + 0.018499180674552917, + 0.016876786947250366, + 0.01926473155617714, + 0.01036107912659645, + 0.027109520509839058, + 0.013688898645341396, + 0.015036472119390965, + 0.012985903769731522, + 0.009518950246274471, + 0.005130395293235779, + 0.0021896185353398323, + 0.0032545016147196293, + 0.007192898541688919, + 0.013590708374977112, + 0.0025509102270007133, + 0.0026881261728703976, + 0.002945538843050599, + 0.0027034934610128403, + 0.0019342985469847918, + 0.003682223381474614, + 0.0014596377732232213, + 0.0014862859388813376, + 0.003250835696235299, + 0.0016631678445264697, + 0.03359483182430267, + 0.04538688063621521, + 0.01949913427233696, + 0.017957914620637894, + 0.01605934463441372, + 0.04657036438584328, + 0.01370861753821373, + 0.024361107498407364, + 0.006300256587564945, + 0.003347112564370036, + 0.006389243062585592, + 0.003887747647240758, + 0.004549542907625437, + 0.0032773602288216352, + 0.004467009101063013, + 0.004356394521892071, + 0.010125203058123589, + 0.006653898861259222, + 0.007350573781877756, + 0.0054684653878211975, + 0.01193485502153635, + 0.009006119333207607, + 0.005703463684767485, + 0.007394609972834587, + 0.007553847040981054, + 0.009254773147404194, + 0.005176203791052103, + 0.0063461726531386375, + 0.0073170773684978485, + 0.011907839216291904, + 0.006174132227897644, + 0.005401367321610451, + 0.002291998825967312, + 0.004232405684888363, + 0.001316584413871169, + 0.0015487706987187266, + 0.002195789711549878, + 0.0024080139119178057, + 0.002083750441670418, + 0.0016045408556237817 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.011077924631536007 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_tensor([9585, 9585])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_tensor([9585, 9585])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..063cb60c70a5ddaf608dcf1a0667159e65ec59ff --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_tensor([9585, 9585])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "The film begins with a low-angle close-up, focused on a dark, shadowy crevice within a massive artificial rock structure within a vast aquarium. In the dim light, the moray eel's mottled green-brown head is barely visible, its leathery skin coated in mucus, forming an intricate, maze-like pattern. Its beady, black eyes stare at its hiding place with a menacing stillness, its slightly open mouth offering a glimpse of sharp, translucent teeth. The surrounding rock texture mimics a porous volcanic surface, dotted with black algae, while in the distant, out-of-focus depths, the blurry shapes of smaller, colorful fish drift lazily by. As the camera slowly and deliberately tilts upward, the perspective shifts away from the eel and toward the shimmering water surface above. This movement causes dancing caustic light patterns to ripple downward through the crystal-clear water, illuminating drifting particles and casting fleeting highlights on the rock walls. Bathed in cool, aquatic colors of deep blue and hazy gray, the scene creates an atmosphere of tranquility and pristine mystery, capturing a hidden, watchful moment in the vast underwater world.", + "seed": 14338, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_9585_seed_14338_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.009060131059959531, + "vlm_logprob": -0.009060131059959531, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.009060131059959531, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011718484573066235 + ], + "opposite_logprob_tokens": [ + -4.886718273162842 + ], + "logprob_reward": -0.011718484573066235, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.011718484573066235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08914276957511902, + 0.0717630386352539, + 0.07081378996372223, + 0.048860225826501846, + 0.04239163547754288, + 0.061724115163087845, + 0.11303079128265381, + 0.1314428150653839, + 0.33513516187667847, + 0.035695694386959076 + ], + "credit_map": [ + 0.012108036316931248, + 0.012462425976991653, + 0.018319901078939438, + 0.009701467119157314, + 0.007145666982978582, + 0.009042729623615742, + 0.005400450434535742, + 0.014962094835937023, + 0.011752475053071976, + 0.01077989861369133, + 0.013305297121405602, + 0.009646599180996418, + 0.005747785326093435, + 0.008077716454863548, + 0.006080825813114643, + 0.006372442003339529, + 0.005153856240212917, + 0.005286685191094875, + 0.007143952418118715, + 0.005841388367116451, + 0.0061301011592149734, + 0.019315624609589577, + 0.01553803589195013, + 0.006404141429811716, + 0.0050369445234537125, + 0.004905349109321833, + 0.008008434437215328, + 0.0069641671143472195, + 0.004380026366561651, + 0.007190084550529718, + 0.007238495163619518, + 0.00513672549277544, + 0.0051353732123970985, + 0.003931369166821241, + 0.004685853607952595, + 0.0042369915172457695, + 0.0031874103005975485, + 0.0030177272856235504, + 0.007836001925170422, + 0.01036090962588787, + 0.008098011836409569, + 0.0056870728731155396, + 0.01110356580466032, + 0.008658051490783691, + 0.0055603026412427425, + 0.007710399106144905, + 0.007854579947888851, + 0.007052132394164801, + 0.019474497064948082, + 0.01939297281205654, + 0.017388829961419106, + 0.00846183206886053, + 0.01690542884171009, + 0.017535299062728882, + 0.006224213633686304, + 0.007647714111953974, + 0.020261555910110474, + 0.010654094628989697, + 0.025126054883003235, + 0.010023239068686962, + 0.015728307887911797, + 0.021702811121940613, + 0.014403469860553741, + 0.013543287292122841, + 0.06544382125139236, + 0.015729444101452827, + 0.0289569441229105, + 0.010359566658735275, + 0.12745819985866547, + 0.0654737576842308, + 0.011744909919798374, + 0.00996850524097681, + 0.0049070981331169605, + 0.0044052028097212315, + 0.00435148598626256, + 0.002865298418328166, + 0.004807634279131889, + 0.005493208300322294, + 0.0037919541355222464, + 0.005073814652860165 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.011718484573066235 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036256073508411646 + ], + "opposite_logprob_tokens": [ + -5.628625392913818 + ], + "logprob_reward": -0.0036256073508411646, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036256073508411646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053309179842472076, + 0.06105818599462509, + 0.01533066388219595, + 0.032793156802654266, + 0.028482511639595032, + 0.023631561547517776, + 0.021752897650003433, + 0.032290492206811905, + 0.7218897342681885, + 0.009461634792387486 + ], + "credit_map": [ + 0.005250211339443922, + 0.0070149595849215984, + 0.005911550484597683, + 0.0067441463470458984, + 0.007839404046535492, + 0.007611511740833521, + 0.005720809567719698, + 0.007216581609100103, + 0.00432585459202528, + 0.007410957012325525, + 0.007332996930927038, + 0.0064734164625406265, + 0.008304283954203129, + 0.008858258835971355, + 0.007351305335760117, + 0.011001113802194595, + 0.0017102365382015705, + 0.001679397071711719, + 0.0018130637472495437, + 0.0014168183552101254, + 0.0018019373528659344, + 0.0031889425590634346, + 0.0018960533197969198, + 0.0018242159858345985, + 0.00271948822773993, + 0.0034858486615121365, + 0.003823671257123351, + 0.004896723199635744, + 0.003927389159798622, + 0.004062416963279247, + 0.006471023429185152, + 0.0034065968357026577, + 0.0026781472843140364, + 0.0036259430926293135, + 0.004227524623274803, + 0.00504347076639533, + 0.0021554597187787294, + 0.00232111313380301, + 0.0034288663882762194, + 0.0050019873306155205, + 0.0029023599345237017, + 0.002132933586835861, + 0.003307939739897847, + 0.002937605604529381, + 0.003385363845154643, + 0.003222703468054533, + 0.0029511889442801476, + 0.0027914668899029493, + 0.0027752318419516087, + 0.004034915007650852, + 0.002653565024957061, + 0.0024334837216883898, + 0.0019229411846026778, + 0.002350540366023779, + 0.002762178424745798, + 0.002820041496306658, + 0.005001296754926443, + 0.0022572148591279984, + 0.0035038634669035673, + 0.004048381000757217, + 0.0038709265645593405, + 0.003348537255078554, + 0.005580715369433165, + 0.004679559264332056, + 0.10171937197446823, + 0.03244967758655548, + 0.21397249400615692, + 0.027351122349500656, + 0.22499980032444, + 0.08026563376188278, + 0.021672440692782402, + 0.01945924013853073, + 0.0012947822688147426, + 0.0012091218959540129, + 0.001229601912200451, + 0.001097573316656053, + 0.0009751932229846716, + 0.00161371985450387, + 0.0010462417267262936, + 0.00099539989605546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.0036256073508411646 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640168234705925 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640168234705925, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640168234705925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.029789431020617485, + 0.0672362744808197, + 0.058089565485715866, + 0.030693793669342995, + 0.040916506201028824, + 0.10700428485870361, + 0.06371579319238663, + 0.05425956845283508, + 0.5226054191589355, + 0.025689370930194855 + ], + "credit_map": [ + 0.005115770734846592, + 0.004955730866640806, + 0.004362822510302067, + 0.003775977296754718, + 0.00205647642724216, + 0.0034037467557936907, + 0.003633915912359953, + 0.002484990283846855, + 0.009390478953719139, + 0.008582666516304016, + 0.010450094006955624, + 0.005067234393209219, + 0.005550561472773552, + 0.011080075986683369, + 0.005544265732169151, + 0.011570900678634644, + 0.004993199836462736, + 0.003782241838052869, + 0.0062283980660140514, + 0.004327836912125349, + 0.004711446352303028, + 0.011705493554472923, + 0.01581471785902977, + 0.0065262289717793465, + 0.003007889725267887, + 0.004337602760642767, + 0.0047954474575817585, + 0.004827738739550114, + 0.002358304802328348, + 0.003544145729392767, + 0.003959265071898699, + 0.0038633989170193672, + 0.003257593372836709, + 0.0032741164322942495, + 0.006009490229189396, + 0.00809139758348465, + 0.0030070149805396795, + 0.0031667000148445368, + 0.007184007205069065, + 0.006926184985786676, + 0.014948348514735699, + 0.008671038784086704, + 0.018364008516073227, + 0.013564087450504303, + 0.010239152237772942, + 0.014367354102432728, + 0.013516654260456562, + 0.0133336391299963, + 0.007062889169901609, + 0.007125135511159897, + 0.010363627225160599, + 0.010838652029633522, + 0.005325463134795427, + 0.005788288079202175, + 0.009078560397028923, + 0.008133177645504475, + 0.010005710646510124, + 0.004860245622694492, + 0.006636481266468763, + 0.00729186600074172, + 0.00409515667706728, + 0.005080238915979862, + 0.0074773854576051235, + 0.008812484331429005, + 0.076213538646698, + 0.024336671456694603, + 0.14811769127845764, + 0.020561762154102325, + 0.1571924388408661, + 0.060648828744888306, + 0.020381024107336998, + 0.015153461135923862, + 0.0033469160553067923, + 0.0026044845581054688, + 0.0037014188710600138, + 0.002738113049417734, + 0.002711167326197028, + 0.004102489911019802, + 0.0032700668089091778, + 0.003214712953194976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 1 + ], + "selected_reward": -0.008640168234705925 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606437936425209 + ], + "opposite_logprob_tokens": [ + -4.141064167022705 + ], + "logprob_reward": -0.01606437936425209, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606437936425209, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.020748423412442207, + 0.015490527264773846, + 0.01654893346130848, + 0.016392750665545464, + 0.007295737974345684, + 0.021955376490950584, + 0.022834505885839462, + 0.02623574621975422, + 0.8435394763946533, + 0.008958465419709682 + ], + "credit_map": [ + 0.0023156071547418833, + 0.002740711672231555, + 0.004918197635561228, + 0.0012658904306590557, + 0.002372226445004344, + 0.0023043614346534014, + 0.0016839075833559036, + 0.003147520823404193, + 0.0026974622160196304, + 0.0026300731115043163, + 0.002549100900068879, + 0.001308162696659565, + 0.0015330591704696417, + 0.0019693232607096434, + 0.0012367464369162917, + 0.001566599472425878, + 0.0010281482245773077, + 0.0012237580958753824, + 0.0017577267717570066, + 0.001452825148589909, + 0.0012544036144390702, + 0.004150958266109228, + 0.004160765092819929, + 0.0015203488292172551, + 0.0015659775817766786, + 0.0015327013097703457, + 0.0025047759991139174, + 0.002105527091771364, + 0.0014419682556763291, + 0.003280098782852292, + 0.0021699939388781786, + 0.0017917080549523234, + 0.0008677338482812047, + 0.0007539892103523016, + 0.0010739871067926288, + 0.001059238682501018, + 0.0008467490551993251, + 0.0007358286529779434, + 0.0009955427376553416, + 0.0009626691462472081, + 0.0029405958484858274, + 0.0017291888361796737, + 0.0038179573602974415, + 0.0019008051604032516, + 0.0023282563779503107, + 0.0026067583821713924, + 0.002962440252304077, + 0.0036693757865577936, + 0.003023199737071991, + 0.004097204189747572, + 0.0024312094319611788, + 0.003159566083922982, + 0.0026177014224231243, + 0.0027227597311139107, + 0.0027952389791607857, + 0.001987625379115343, + 0.004385878797620535, + 0.0018217646284028888, + 0.0027541802264750004, + 0.0024401608388870955, + 0.004687658045440912, + 0.003555821720510721, + 0.002577287843450904, + 0.004012995399534702, + 0.12331092357635498, + 0.037938158959150314, + 0.2536710798740387, + 0.03141293302178383, + 0.2591676712036133, + 0.09097248315811157, + 0.025095077231526375, + 0.021971171721816063, + 0.0011419509537518024, + 0.0010458484757691622, + 0.001036380184814334, + 0.0009988099336624146, + 0.0010364553891122341, + 0.0013412154512479901, + 0.0011368583654984832, + 0.0012209468986839056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.01606437936425209 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0526869036257267, + 0.012298381887376308, + 0.014320260845124722, + 0.01651824451982975, + 0.019229551777243614, + 0.008572830818593502, + 0.014349494129419327, + 0.03390970826148987, + 0.8121700286865234, + 0.01594454236328602 + ], + "credit_map": [ + 0.005243338644504547, + 0.007947453297674656, + 0.014417443424463272, + 0.003656260436400771, + 0.0032465907279402018, + 0.005392828956246376, + 0.0026329895481467247, + 0.010150004178285599, + 0.001969063188880682, + 0.0015204964438453317, + 0.00219877902418375, + 0.0014701419277116656, + 0.0014918326633051038, + 0.0014676603022962809, + 0.0009367004386149347, + 0.0012437091208994389, + 0.0016345104668289423, + 0.0018418175168335438, + 0.0014308267273008823, + 0.0013676047092303634, + 0.001218468532897532, + 0.003209932940080762, + 0.0018519264413043857, + 0.0017651745583862066, + 0.0016515252646058798, + 0.001801209058612585, + 0.0022889638785272837, + 0.002376220189034939, + 0.0012880648719146848, + 0.0025400149170309305, + 0.0024527960922569036, + 0.0021194538567215204, + 0.0019016150617972016, + 0.0017195014515891671, + 0.0034739235416054726, + 0.0040503013879060745, + 0.0013787512434646487, + 0.00162682740483433, + 0.002445876831188798, + 0.0026327583473175764, + 0.0011662461329251528, + 0.000781892507802695, + 0.0013492244761437178, + 0.000992699759081006, + 0.0008674098644405603, + 0.001040183356963098, + 0.0011464570416137576, + 0.0012287184363231063, + 0.0019596375059336424, + 0.002676306990906596, + 0.002216156106442213, + 0.0015004394808784127, + 0.001653433428145945, + 0.0017191602382808924, + 0.0013168546138331294, + 0.0013075083261355758, + 0.0039276485331356525, + 0.002345076994970441, + 0.0031901143956929445, + 0.002053805161267519, + 0.008744612336158752, + 0.005172899924218655, + 0.004261572379618883, + 0.004213986452668905, + 0.09335269778966904, + 0.038669563829898834, + 0.21924272179603577, + 0.029092609882354736, + 0.276870995759964, + 0.1053173765540123, + 0.026268795132637024, + 0.0233553908765316, + 0.001978325191885233, + 0.0023649598006159067, + 0.0016137735219672322, + 0.0012973499251529574, + 0.002557940548285842, + 0.0028907356318086386, + 0.0014668999938294291, + 0.0017745617078617215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.0052520157769322395 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle close-up, focused on a dark, shadowy crevice within a massive artificial rock structure within a vast aquarium. In the dim light, the moray eel's mottled green-brown head is barely visible, its leathery skin coated in mucus, forming an intricate, maze-like pattern. Its beady, black eyes stare at its hiding place with a menacing stillness, its slightly open mouth offering a glimpse of sharp, translucent teeth. The surrounding rock texture mimics a porous volcanic surface, dotted with black algae, while in the distant, out-of-focus depths, the blurry shapes of smaller, colorful fish drift lazily by. As the camera slowly and deliberately tilts upward, the perspective shifts away from the eel and toward the shimmering water surface above. This movement causes dancing caustic light patterns to ripple downward through the crystal-clear water, illuminating drifting particles and casting fleeting highlights on the rock walls. Bathed in cool, aquatic colors of deep blue and hazy gray, the scene creates an atmosphere of tranquility and pristine mystery, capturing a hidden, watchful moment in the vast underwater world.", + "seed": 14339, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_9585_seed_14339_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.008514226367697119, + "vlm_logprob": -0.008514226367697119, + "vlm_yes_no_margin": 5.0, + "vlm_token_credit": -0.008514226367697119, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011125553399324417 + ], + "opposite_logprob_tokens": [ + -4.761125564575195 + ], + "logprob_reward": -0.011125553399324417, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011125553399324417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06541211158037186, + 0.23962517082691193, + 0.0557132288813591, + 0.0499483123421669, + 0.11147152632474899, + 0.056014783680438995, + 0.06824375689029694, + 0.06922172009944916, + 0.15897466242313385, + 0.1253747045993805 + ], + "credit_map": [ + 0.00753364572301507, + 0.011134011670947075, + 0.006102410610765219, + 0.006775797810405493, + 0.00831806194037199, + 0.010662484914064407, + 0.00580508541315794, + 0.009080607444047928, + 0.03147295117378235, + 0.018781084567308426, + 0.03186967596411705, + 0.033330973237752914, + 0.03386224806308746, + 0.03775424137711525, + 0.023057857528328896, + 0.02949610911309719, + 0.0057275984436273575, + 0.008627716451883316, + 0.0075186328031122684, + 0.006144986022263765, + 0.005763942375779152, + 0.009475728496909142, + 0.00484866788610816, + 0.007605951279401779, + 0.0038624198641628027, + 0.0038999775424599648, + 0.009179445914924145, + 0.008786531165242195, + 0.004102801438421011, + 0.005081032402813435, + 0.004384130239486694, + 0.010651973076164722, + 0.010881002992391586, + 0.015196667052805424, + 0.01655895635485649, + 0.01529852394014597, + 0.010340605862438679, + 0.013263873755931854, + 0.01128032524138689, + 0.018651554360985756, + 0.006328255403786898, + 0.0065110186114907265, + 0.010443948209285736, + 0.00811352115124464, + 0.006914759054780006, + 0.005067465826869011, + 0.006218027323484421, + 0.006417781580239534, + 0.005391136277467012, + 0.005568839143961668, + 0.01015403401106596, + 0.007175421807914972, + 0.004563528578728437, + 0.004469736013561487, + 0.01603226363658905, + 0.014888790436089039, + 0.008807649835944176, + 0.01507395040243864, + 0.009652948006987572, + 0.009002386592328548, + 0.00734338816255331, + 0.005059372633695602, + 0.007345121819525957, + 0.0069368924014270306, + 0.02365480177104473, + 0.0194777213037014, + 0.016124896705150604, + 0.01445344090461731, + 0.029001662507653236, + 0.030414298176765442, + 0.012087074108421803, + 0.013760766014456749, + 0.019330214709043503, + 0.0144429262727499, + 0.02291210927069187, + 0.017333798110485077, + 0.014613071456551552, + 0.014536566101014614, + 0.01309962198138237, + 0.009106378071010113 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 9 + ], + "selected_reward": -0.011125553399324417 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005265653133392334 + ], + "opposite_logprob_tokens": [ + -5.255265712738037 + ], + "logprob_reward": -0.005265653133392334, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005265653133392334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10783372074365616, + 0.2317466139793396, + 0.27942967414855957, + 0.0956869050860405, + 0.040623314678668976, + 0.0420772023499012, + 0.04537303000688553, + 0.04992533475160599, + 0.06138095632195473, + 0.045923229306936264 + ], + "credit_map": [ + 0.015237157233059406, + 0.011614308692514896, + 0.017336847260594368, + 0.009601938538253307, + 0.008805457502603531, + 0.016493670642375946, + 0.015161877498030663, + 0.013582460582256317, + 0.037739578634500504, + 0.021328434348106384, + 0.04042226821184158, + 0.0309641994535923, + 0.02315046451985836, + 0.030237935483455658, + 0.019128989428281784, + 0.028774745762348175, + 0.025143738836050034, + 0.02998138777911663, + 0.02927260845899582, + 0.030595457181334496, + 0.045160166919231415, + 0.044809747487306595, + 0.026037082076072693, + 0.04842948168516159, + 0.007253977004438639, + 0.0071495757438242435, + 0.01372282300144434, + 0.023523541167378426, + 0.009177041240036488, + 0.008545617572963238, + 0.013817079365253448, + 0.012497246265411377, + 0.003989909775555134, + 0.00614921934902668, + 0.005275738891214132, + 0.0056002093479037285, + 0.004359613172709942, + 0.0054754638113081455, + 0.004133269190788269, + 0.00563989020884037, + 0.0045429435558617115, + 0.004800589755177498, + 0.009106083773076534, + 0.005888528656214476, + 0.003724790411069989, + 0.002752298954874277, + 0.005423886701464653, + 0.005838083103299141, + 0.0030365039128810167, + 0.00379316252656281, + 0.0060038319788873196, + 0.0073013524524867535, + 0.004013040568679571, + 0.0026999665424227715, + 0.009504543617367744, + 0.009020627476274967, + 0.005985551048070192, + 0.006376519799232483, + 0.006156331393867731, + 0.007013991009443998, + 0.006776778493076563, + 0.00500023178756237, + 0.006003659218549728, + 0.006612271070480347, + 0.010123643092811108, + 0.0067924754694104195, + 0.00970594584941864, + 0.005863797385245562, + 0.008331166580319405, + 0.008865948766469955, + 0.00550549291074276, + 0.006192489061504602, + 0.004880587570369244, + 0.004647149238735437, + 0.007191943004727364, + 0.011140025220811367, + 0.0035355736035853624, + 0.005254323594272137, + 0.004967335145920515, + 0.004306289367377758 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.005265653133392334 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625637110322714 + ], + "opposite_logprob_tokens": [ + -4.882625579833984 + ], + "logprob_reward": -0.007625637110322714, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625637110322714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10896338522434235, + 0.3474494516849518, + 0.09707675874233246, + 0.0544346421957016, + 0.07064801454544067, + 0.07496753334999084, + 0.03421685844659805, + 0.05038207769393921, + 0.08970031887292862, + 0.07216089218854904 + ], + "credit_map": [ + 0.0114169642329216, + 0.014976800419390202, + 0.015371798537671566, + 0.014711443334817886, + 0.010225687175989151, + 0.01576254703104496, + 0.011421817354857922, + 0.015076330862939358, + 0.04606126993894577, + 0.026293640956282616, + 0.046616606414318085, + 0.048277150839567184, + 0.04682145640254021, + 0.0555131770670414, + 0.03345540910959244, + 0.04441076144576073, + 0.010665548965334892, + 0.009422247298061848, + 0.014452754519879818, + 0.012739677913486958, + 0.009492742829024792, + 0.013260728679597378, + 0.009810445830225945, + 0.017232611775398254, + 0.0039417739026248455, + 0.004904532805085182, + 0.0090874582529068, + 0.00833594985306263, + 0.005635468289256096, + 0.006232193671166897, + 0.005144372116774321, + 0.011152895167469978, + 0.004851579666137695, + 0.00705741997808218, + 0.009221997112035751, + 0.011330753564834595, + 0.007268931251019239, + 0.0065268841572105885, + 0.011354651302099228, + 0.013035794720053673, + 0.007219976745545864, + 0.007498735561966896, + 0.011217527091503143, + 0.010303945280611515, + 0.009301505982875824, + 0.010431423783302307, + 0.009870562702417374, + 0.009123854339122772, + 0.003543758299201727, + 0.003360120812430978, + 0.0050747874192893505, + 0.004721249453723431, + 0.003971942234784365, + 0.0024831166956573725, + 0.00574604794383049, + 0.00531583558768034, + 0.006599143147468567, + 0.006886955350637436, + 0.006656323093920946, + 0.005691824480891228, + 0.005700848065316677, + 0.005060796160250902, + 0.00752722704783082, + 0.006258957553654909, + 0.014371100813150406, + 0.01107724942266941, + 0.008423520252108574, + 0.01005414966493845, + 0.014518480747938156, + 0.016474176198244095, + 0.007158545311540365, + 0.007623097859323025, + 0.006266324315220118, + 0.010757065378129482, + 0.009776768274605274, + 0.013462812639772892, + 0.009906630031764507, + 0.008061433210968971, + 0.007658033631742001, + 0.0062718172557652 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007625637110322714 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016065318137407303 + ], + "opposite_logprob_tokens": [ + -4.1410651206970215 + ], + "logprob_reward": -0.016065318137407303, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016065318137407303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07916732132434845, + 0.20892147719860077, + 0.09523431956768036, + 0.09757193922996521, + 0.12774178385734558, + 0.10853561013936996, + 0.04803002253174782, + 0.029585005715489388, + 0.0909242033958435, + 0.11428837478160858 + ], + "credit_map": [ + 0.007672278676182032, + 0.010019276291131973, + 0.009776188991963863, + 0.011694801971316338, + 0.007197459693998098, + 0.013517717830836773, + 0.007386703044176102, + 0.011902890168130398, + 0.030288606882095337, + 0.016170497983694077, + 0.03144388645887375, + 0.029723120853304863, + 0.024897921830415726, + 0.030900001525878906, + 0.019125457853078842, + 0.02637198008596897, + 0.014909794554114342, + 0.015369845554232597, + 0.010580699890851974, + 0.011587892659008503, + 0.011070098727941513, + 0.010514547117054462, + 0.007929033599793911, + 0.013272405602037907, + 0.004812002647668123, + 0.005030449014157057, + 0.027072763070464134, + 0.007851040922105312, + 0.00869564339518547, + 0.007877127267420292, + 0.007334888447076082, + 0.028898024931550026, + 0.00846474152058363, + 0.010349122807383537, + 0.01818673685193062, + 0.01939123123884201, + 0.011387964710593224, + 0.010553548112511635, + 0.020412718877196312, + 0.028995707631111145, + 0.012015906162559986, + 0.008971165865659714, + 0.016959184780716896, + 0.014357557520270348, + 0.011938342824578285, + 0.015890460461378098, + 0.015575203113257885, + 0.01282778661698103, + 0.004038433078676462, + 0.004482367541640997, + 0.005439681466668844, + 0.006066993344575167, + 0.005299507174640894, + 0.005649870727211237, + 0.010517324320971966, + 0.006535844411700964, + 0.003876802511513233, + 0.0033497721888124943, + 0.0035084648989140987, + 0.003632423235103488, + 0.004208275116980076, + 0.002345219487324357, + 0.00592639297246933, + 0.0027376553043723106, + 0.01376948133111, + 0.010062990710139275, + 0.014441718347370625, + 0.009428917430341244, + 0.01494582463055849, + 0.012542994692921638, + 0.008235728368163109, + 0.007496544625610113, + 0.01359055656939745, + 0.012963089160621166, + 0.02241659350693226, + 0.014692052267491817, + 0.015460058115422726, + 0.013829057104885578, + 0.012162134051322937, + 0.009174828417599201 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 9 + ], + "selected_reward": -0.016065318137407303 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024889700580388308 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.0024889700580388308, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024889700580388308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16560259461402893, + 0.10625249892473221, + 0.10170687735080719, + 0.07655941694974899, + 0.04455149546265602, + 0.12725237011909485, + 0.07403959333896637, + 0.07829611748456955, + 0.09964759647846222, + 0.12609145045280457 + ], + "credit_map": [ + 0.01826275698840618, + 0.017512472346425056, + 0.01887720264494419, + 0.01827879063785076, + 0.01725631393492222, + 0.04036213457584381, + 0.014159156009554863, + 0.020893767476081848, + 0.013067287392914295, + 0.008247299119830132, + 0.016488423570990562, + 0.010364794172346592, + 0.01568613387644291, + 0.018967485055327415, + 0.01047748513519764, + 0.012953591533005238, + 0.010782678611576557, + 0.012299581430852413, + 0.011126523837447166, + 0.009711237624287605, + 0.012706826440989971, + 0.0179832112044096, + 0.007964144460856915, + 0.019132666289806366, + 0.00576388742774725, + 0.006145705934613943, + 0.016461264342069626, + 0.00906459242105484, + 0.006994420662522316, + 0.009659828618168831, + 0.007913521490991116, + 0.014556190930306911, + 0.0035787990782409906, + 0.003226278116926551, + 0.005370482802391052, + 0.009126558899879456, + 0.005361694376915693, + 0.0045073144137859344, + 0.00722920335829258, + 0.006151165813207626, + 0.015068942680954933, + 0.011789090931415558, + 0.019568417221307755, + 0.020166192203760147, + 0.009417802095413208, + 0.009750195778906345, + 0.022492609918117523, + 0.018999123945832253, + 0.007510388270020485, + 0.007689810357987881, + 0.010008375160396099, + 0.008393894881010056, + 0.008051825687289238, + 0.005739199463278055, + 0.01117473840713501, + 0.01547135692089796, + 0.010723530314862728, + 0.017326243221759796, + 0.010192074812948704, + 0.004636245779693127, + 0.01080773863941431, + 0.006741900462657213, + 0.008362223394215107, + 0.009506162256002426, + 0.016875382512807846, + 0.011989940889179707, + 0.012194349430501461, + 0.012907023541629314, + 0.015287411399185658, + 0.013836370781064034, + 0.00867527537047863, + 0.00788184441626072, + 0.014619970694184303, + 0.013054178096354008, + 0.02664404921233654, + 0.0148379011079669, + 0.01854042522609234, + 0.01666930317878723, + 0.0119788758456707, + 0.009746738709509373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 9 + ], + "selected_reward": -0.0024889700580388308 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle close-up, focused on a dark, shadowy crevice within a massive artificial rock structure within a vast aquarium. In the dim light, the moray eel's mottled green-brown head is barely visible, its leathery skin coated in mucus, forming an intricate, maze-like pattern. Its beady, black eyes stare at its hiding place with a menacing stillness, its slightly open mouth offering a glimpse of sharp, translucent teeth. The surrounding rock texture mimics a porous volcanic surface, dotted with black algae, while in the distant, out-of-focus depths, the blurry shapes of smaller, colorful fish drift lazily by. As the camera slowly and deliberately tilts upward, the perspective shifts away from the eel and toward the shimmering water surface above. This movement causes dancing caustic light patterns to ripple downward through the crystal-clear water, illuminating drifting particles and casting fleeting highlights on the rock walls. Bathed in cool, aquatic colors of deep blue and hazy gray, the scene creates an atmosphere of tranquility and pristine mystery, capturing a hidden, watchful moment in the vast underwater world.", + "seed": 14340, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_9585_seed_14340_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.010026424331590534, + "vlm_logprob": -0.010026424331590534, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.010026424331590534, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009551186114549637 + ], + "opposite_logprob_tokens": [ + -5.009551048278809 + ], + "logprob_reward": -0.009551186114549637, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.009551186114549637, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7020795941352844, + 0.03423911705613136, + 0.03465281054377556, + 0.047644682228565216, + 0.020047109574079514, + 0.04541917145252228, + 0.060869310051202774, + 0.02096940204501152, + 0.017223354429006577, + 0.01685541495680809 + ], + "credit_map": [ + 0.06992689520120621, + 0.080339215695858, + 0.05457687750458717, + 0.06132141873240471, + 0.13485337793827057, + 0.1929062008857727, + 0.04598541930317879, + 0.062170159071683884, + 0.004313911311328411, + 0.0055895838886499405, + 0.004545486066490412, + 0.0022104799281805754, + 0.00725509412586689, + 0.004278147127479315, + 0.0033940847497433424, + 0.0026523256674408913, + 0.0027569814119488, + 0.0038799543399363756, + 0.0037910433020442724, + 0.0036508168559521437, + 0.005534790921956301, + 0.008287300355732441, + 0.0029743509367108345, + 0.003777572186663747, + 0.003864109516143799, + 0.007342074997723103, + 0.007493099197745323, + 0.006886407732963562, + 0.005205081310123205, + 0.00530092790722847, + 0.0052505903877317905, + 0.0063023921102285385, + 0.002602760447189212, + 0.0023922238033264875, + 0.0028884916100651026, + 0.0031115792226046324, + 0.002372878137975931, + 0.002043244894593954, + 0.0022182329557836056, + 0.0024176982697099447, + 0.0031213550828397274, + 0.0037678126245737076, + 0.006344596855342388, + 0.006770285312086344, + 0.005528699140995741, + 0.004166153725236654, + 0.007611114997416735, + 0.008109156042337418, + 0.004836894106119871, + 0.012086202390491962, + 0.0044036926701664925, + 0.007685631047934294, + 0.014966804534196854, + 0.0059270719066262245, + 0.0056996820494532585, + 0.005263327620923519, + 0.002215776825323701, + 0.002612703014165163, + 0.0030717039480805397, + 0.0031682103872299194, + 0.0022764166351407766, + 0.0027750604785978794, + 0.002639354905113578, + 0.0022101756185293198, + 0.0013719061389565468, + 0.0030388731975108385, + 0.0020282657351344824, + 0.002011546865105629, + 0.001960833091288805, + 0.0026511738542467356, + 0.002491637133061886, + 0.0016691186465322971, + 0.0022792189847677946, + 0.004152912180870771, + 0.0013035617303103209, + 0.0014449618756771088, + 0.002988258609548211, + 0.001905081095173955, + 0.0016037722816690803, + 0.00117764831520617 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.009551186114549637 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0046457452699542046 + ], + "opposite_logprob_tokens": [ + -5.379645824432373 + ], + "logprob_reward": -0.0046457452699542046, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.0046457452699542046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5810737013816833, + 0.0471990667283535, + 0.04547005519270897, + 0.10779084265232086, + 0.05812625214457512, + 0.09490945190191269, + 0.009279677644371986, + 0.02143036015331745, + 0.017274554818868637, + 0.017446016892790794 + ], + "credit_map": [ + 0.05833179131150246, + 0.06852410733699799, + 0.04134277626872063, + 0.05034886673092842, + 0.11367465555667877, + 0.16289494931697845, + 0.035197388380765915, + 0.050759170204401016, + 0.0035932117607444525, + 0.005967882927507162, + 0.005081051029264927, + 0.003336291993036866, + 0.008902312256395817, + 0.01253685262054205, + 0.004448814783245325, + 0.0033326479606330395, + 0.003492082003504038, + 0.005986275151371956, + 0.004220390692353249, + 0.006176067050546408, + 0.006259281188249588, + 0.008479233831167221, + 0.004359386395663023, + 0.006497336085885763, + 0.011235618032515049, + 0.015001087449491024, + 0.014511228539049625, + 0.011951534077525139, + 0.012078391388058662, + 0.01232348382472992, + 0.01715605892241001, + 0.013533437624573708, + 0.005245603155344725, + 0.005782255902886391, + 0.009573026560246944, + 0.009241004474461079, + 0.0049078515730798244, + 0.007074497640132904, + 0.008161534555256367, + 0.008140472695231438, + 0.0045886291190981865, + 0.009888277389109135, + 0.01518517266958952, + 0.01787514239549637, + 0.007491590920835733, + 0.006815922912210226, + 0.016881676390767097, + 0.016183029860258102, + 0.0010925254318863153, + 0.0013423697091639042, + 0.000976610230281949, + 0.0012353332713246346, + 0.0012205499224364758, + 0.0009484312613494694, + 0.0012919137952849269, + 0.0011719439644366503, + 0.0021083978936076164, + 0.003222711617127061, + 0.002374427393078804, + 0.0036934323143213987, + 0.002747467253357172, + 0.00300825503654778, + 0.002268977463245392, + 0.00200669071637094, + 0.0012013681698590517, + 0.003189608920365572, + 0.001961736474186182, + 0.0028879959136247635, + 0.002139681251719594, + 0.002014726400375366, + 0.0020090125035494566, + 0.0018704256508499384, + 0.0018425205489620566, + 0.004391815979033709, + 0.0017224233597517014, + 0.0017470482271164656, + 0.002842307323589921, + 0.002026145113632083, + 0.0014149259077385068, + 0.0014588292688131332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.0046457452699542046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057752013206482 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.02057752013206482, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057752013206482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15243394672870636, + 0.07947281002998352, + 0.06588030606508255, + 0.05811067670583725, + 0.055127933621406555, + 0.13213278353214264, + 0.09698279201984406, + 0.06324003636837006, + 0.15836922824382782, + 0.1382495015859604 + ], + "credit_map": [ + 0.02289024367928505, + 0.024222712963819504, + 0.022080598399043083, + 0.015030533075332642, + 0.02378472127020359, + 0.02026713825762272, + 0.01203729584813118, + 0.012120706029236317, + 0.007003618869930506, + 0.01445934921503067, + 0.005452585406601429, + 0.0051044318825006485, + 0.024008937180042267, + 0.013588132336735725, + 0.006460578180849552, + 0.003395184176042676, + 0.0049871886149048805, + 0.009062119759619236, + 0.006363559514284134, + 0.008701973594725132, + 0.009551220573484898, + 0.01477206964045763, + 0.006306590046733618, + 0.0061355833895504475, + 0.005140411667525768, + 0.009378129616379738, + 0.00932251289486885, + 0.00786658376455307, + 0.0067516411654651165, + 0.006652898155152798, + 0.006902648601680994, + 0.006095853168517351, + 0.006219716742634773, + 0.006022030487656593, + 0.010916251689195633, + 0.008854501880705357, + 0.006358240731060505, + 0.005061337258666754, + 0.005798035301268101, + 0.005897815339267254, + 0.006767625920474529, + 0.00934904906898737, + 0.022736400365829468, + 0.02441864088177681, + 0.0105678616091609, + 0.010956140235066414, + 0.02374505065381527, + 0.023592019453644753, + 0.010272384621202946, + 0.018661070615053177, + 0.008907254785299301, + 0.008586452342569828, + 0.021448003128170967, + 0.008391287177801132, + 0.011408289894461632, + 0.00930805318057537, + 0.0053867013193666935, + 0.006479720585048199, + 0.008822633884847164, + 0.012718530371785164, + 0.006582737434655428, + 0.00710432231426239, + 0.008983904495835304, + 0.007161492947489023, + 0.011486716568470001, + 0.033282533288002014, + 0.016780991107225418, + 0.02710598148405552, + 0.022816291078925133, + 0.013977797701954842, + 0.017545394599437714, + 0.015373524278402328, + 0.016918765380978584, + 0.024841563776135445, + 0.010706108994781971, + 0.012258993461728096, + 0.032897401601076126, + 0.018660062924027443, + 0.012365785427391529, + 0.00960083119571209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.02057752013206482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01253841444849968 + ], + "opposite_logprob_tokens": [ + -4.387538433074951 + ], + "logprob_reward": -0.01253841444849968, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01253841444849968, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.275317519903183, + 0.027759777382016182, + 0.06011328846216202, + 0.04791296645998955, + 0.03362339362502098, + 0.11658437550067902, + 0.054944392293691635, + 0.07153142243623734, + 0.2514798939228058, + 0.06073301285505295 + ], + "credit_map": [ + 0.031723521649837494, + 0.03134418651461601, + 0.02363399788737297, + 0.024945829063653946, + 0.04684668034315109, + 0.06743969023227692, + 0.020804474130272865, + 0.028579121455550194, + 0.003131493227556348, + 0.003984002862125635, + 0.00243568466976285, + 0.0017386191757395864, + 0.005724417045712471, + 0.005887956824153662, + 0.0026663600001484156, + 0.0021912422962486744, + 0.004821357317268848, + 0.00808153860270977, + 0.005257102660834789, + 0.008287250064313412, + 0.009961666539311409, + 0.014436570927500725, + 0.004008143674582243, + 0.005259656347334385, + 0.00420150114223361, + 0.007822084240615368, + 0.007790905423462391, + 0.007710292004048824, + 0.006246109958738089, + 0.004086395259946585, + 0.006299148313701153, + 0.0037565259262919426, + 0.003804996144026518, + 0.0039536901749670506, + 0.0058268397115170956, + 0.00569954002276063, + 0.0029115646611899137, + 0.0030851552728563547, + 0.003802231978625059, + 0.004539371933788061, + 0.00945584662258625, + 0.012466808781027794, + 0.02038360759615898, + 0.01800696738064289, + 0.013754877261817455, + 0.012071906588971615, + 0.01579468324780464, + 0.014649668708443642, + 0.006472282111644745, + 0.010289323516190052, + 0.005490799900144339, + 0.0048021189868450165, + 0.012600989080965519, + 0.004734102636575699, + 0.006068298127502203, + 0.004486470017582178, + 0.007428288459777832, + 0.010894659906625748, + 0.007690735161304474, + 0.01281142607331276, + 0.007854029536247253, + 0.008794342167675495, + 0.009211764670908451, + 0.006846164353191853, + 0.017451804131269455, + 0.029581239446997643, + 0.03648977726697922, + 0.028897928074002266, + 0.02740548923611641, + 0.035409729927778244, + 0.04407469183206558, + 0.032169219106435776, + 0.008384808897972107, + 0.008515960536897182, + 0.004025961738079786, + 0.004857681691646576, + 0.01613551750779152, + 0.0070298295468091965, + 0.006327234208583832, + 0.0054560168646276 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.01253841444849968 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002819255692884326 + ], + "opposite_logprob_tokens": [ + -5.877819061279297 + ], + "logprob_reward": -0.002819255692884326, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002819255692884326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3551068603992462, + 0.10972031950950623, + 0.05060852691531181, + 0.05159996077418327, + 0.03997800871729851, + 0.06613761931657791, + 0.12275803834199905, + 0.04789814352989197, + 0.04392026364803314, + 0.11227226257324219 + ], + "credit_map": [ + 0.04107910767197609, + 0.0407387837767601, + 0.03570246696472168, + 0.03384670987725258, + 0.05897822603583336, + 0.08346381783485413, + 0.023930078372359276, + 0.037367671728134155, + 0.009026361629366875, + 0.021837692707777023, + 0.009063338860869408, + 0.007601600140333176, + 0.03008028119802475, + 0.018178988248109818, + 0.007767157629132271, + 0.006164904218167067, + 0.005106151569634676, + 0.005481983534991741, + 0.00637329276651144, + 0.005414241459220648, + 0.0076311733573675156, + 0.010020803660154343, + 0.004754745401442051, + 0.005826138891279697, + 0.004287520423531532, + 0.009689952246844769, + 0.008069862611591816, + 0.007569343317300081, + 0.005777187645435333, + 0.005025872029364109, + 0.006746251601725817, + 0.004433969967067242, + 0.005158553831279278, + 0.00524454191327095, + 0.005774283781647682, + 0.005508099216967821, + 0.005732942372560501, + 0.0036527374759316444, + 0.004499160218983889, + 0.004407691769301891, + 0.004899695049971342, + 0.006612254772335291, + 0.012293165549635887, + 0.011380844749510288, + 0.006286928430199623, + 0.00566577585414052, + 0.00963294506072998, + 0.009366010315716267, + 0.01187245361506939, + 0.012657738290727139, + 0.0188174806535244, + 0.020893361419439316, + 0.009326672181487083, + 0.01229647547006607, + 0.023605026304721832, + 0.013288838788866997, + 0.005384530872106552, + 0.007383155170828104, + 0.006274257320910692, + 0.006079912185668945, + 0.005758383311331272, + 0.007344275247305632, + 0.005109928082674742, + 0.004563703667372465, + 0.004380994010716677, + 0.0059179202653467655, + 0.0038792656268924475, + 0.0050444104708731174, + 0.005443522706627846, + 0.006963821593672037, + 0.007116798777133226, + 0.005173529963940382, + 0.014125724323093891, + 0.025561578571796417, + 0.009467270225286484, + 0.010380398482084274, + 0.012102700769901276, + 0.017257442697882652, + 0.013180174864828587, + 0.010196979157626629 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.002819255692884326 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "The film begins with a low-angle close-up, focused on a dark, shadowy crevice within a massive artificial rock structure within a vast aquarium. In the dim light, the moray eel's mottled green-brown head is barely visible, its leathery skin coated in mucus, forming an intricate, maze-like pattern. Its beady, black eyes stare at its hiding place with a menacing stillness, its slightly open mouth offering a glimpse of sharp, translucent teeth. The surrounding rock texture mimics a porous volcanic surface, dotted with black algae, while in the distant, out-of-focus depths, the blurry shapes of smaller, colorful fish drift lazily by. As the camera slowly and deliberately tilts upward, the perspective shifts away from the eel and toward the shimmering water surface above. This movement causes dancing caustic light patterns to ripple downward through the crystal-clear water, illuminating drifting particles and casting fleeting highlights on the rock walls. Bathed in cool, aquatic colors of deep blue and hazy gray, the scene creates an atmosphere of tranquility and pristine mystery, capturing a hidden, watchful moment in the vast underwater world.", + "seed": 14341, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_11/index_9585_seed_14341_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.009105381416156888, + "vlm_logprob": -0.009105381416156888, + "vlm_yes_no_margin": 4.925, + "vlm_token_credit": -0.009105381416156888, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017800109460949898 + ], + "opposite_logprob_tokens": [ + -4.267800331115723 + ], + "logprob_reward": -0.017800109460949898, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.017800109460949898, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09935778379440308, + 0.04372716695070267, + 0.012950503267347813, + 0.05952510982751846, + 0.08708250522613525, + 0.02290688268840313, + 0.04063945636153221, + 0.39203038811683655, + 0.16587507724761963, + 0.07590506225824356 + ], + "credit_map": [ + 0.016695883125066757, + 0.012062658555805683, + 0.01772928237915039, + 0.009861578233540058, + 0.009657047688961029, + 0.012237380258738995, + 0.010109249502420425, + 0.011004703119397163, + 0.003958952613174915, + 0.007021415513008833, + 0.005900076124817133, + 0.00451299175620079, + 0.004508072044700384, + 0.005120310466736555, + 0.005857569631189108, + 0.006847780663520098, + 0.0016053634462878108, + 0.0009924119804054499, + 0.0015867255860939622, + 0.0017023435793817043, + 0.0017091697081923485, + 0.001547753345221281, + 0.0020886193960905075, + 0.0017181159928441048, + 0.00561347184702754, + 0.004601041320711374, + 0.004489060491323471, + 0.016228260472416878, + 0.004366728011518717, + 0.004314749967306852, + 0.007439890410751104, + 0.012471906840801239, + 0.00896378979086876, + 0.010228992439806461, + 0.014590715058147907, + 0.010704992339015007, + 0.011188963428139687, + 0.008526797406375408, + 0.011754230596125126, + 0.011124027892947197, + 0.003840074874460697, + 0.0028739292174577713, + 0.0019230433972552419, + 0.00238698604516685, + 0.002836815780028701, + 0.003188996808603406, + 0.0025263165589421988, + 0.003330719657242298, + 0.004777335561811924, + 0.005394332110881805, + 0.003927507903426886, + 0.0048829540610313416, + 0.0058662788942456245, + 0.005184266250580549, + 0.004180787596851587, + 0.006425993051379919, + 0.051225997507572174, + 0.05071428418159485, + 0.04473697766661644, + 0.056859660893678665, + 0.044802408665418625, + 0.03943462297320366, + 0.05170426145195961, + 0.05255216732621193, + 0.017927896231412888, + 0.011124970391392708, + 0.05575457587838173, + 0.010718920268118382, + 0.01413749624043703, + 0.007494668010622263, + 0.010854165069758892, + 0.037862394005060196, + 0.009298604913055897, + 0.010426792316138744, + 0.009161250665783882, + 0.007402028422802687, + 0.009881929494440556, + 0.012634861283004284, + 0.006403688807040453, + 0.01069590449333191 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.017800109460949898 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005262807011604309 + ], + "opposite_logprob_tokens": [ + -5.255262851715088 + ], + "logprob_reward": -0.005262807011604309, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005262807011604309, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1286005824804306, + 0.1315782517194748, + 0.044996198266744614, + 0.08019328862428665, + 0.038358937948942184, + 0.015127957798540592, + 0.045456912368535995, + 0.08530011773109436, + 0.4007214605808258, + 0.02966623194515705 + ], + "credit_map": [ + 0.01026913896203041, + 0.0194168109446764, + 0.018718868494033813, + 0.013803385198116302, + 0.014730454422533512, + 0.017629681155085564, + 0.02464710921049118, + 0.009385150857269764, + 0.011098736897110939, + 0.02782195247709751, + 0.011511184275150299, + 0.012413726188242435, + 0.015165362507104874, + 0.01997075043618679, + 0.01907997578382492, + 0.014516561292111874, + 0.006272375117987394, + 0.0028579391073435545, + 0.004823293536901474, + 0.005042550154030323, + 0.007069421466439962, + 0.006137525662779808, + 0.008660164661705494, + 0.004132935311645269, + 0.005245534237474203, + 0.004173770546913147, + 0.017069637775421143, + 0.014212265610694885, + 0.004340247251093388, + 0.0036634360440075397, + 0.005744876340031624, + 0.025743529200553894, + 0.004684012848883867, + 0.004707095213234425, + 0.004028571769595146, + 0.004863961599767208, + 0.005008609965443611, + 0.0034436362329870462, + 0.007148765958845615, + 0.004474287386983633, + 0.0019547336269170046, + 0.001562944264151156, + 0.0019150960724800825, + 0.0018729795701801777, + 0.0019300549756735563, + 0.0022993457969278097, + 0.001976838568225503, + 0.0016159664373844862, + 0.005726569332182407, + 0.004896881058812141, + 0.003939360845834017, + 0.005022931378334761, + 0.008230702020227909, + 0.0064121875911951065, + 0.005921603180468082, + 0.005306682549417019, + 0.007609613239765167, + 0.0076308404095470905, + 0.012775301933288574, + 0.011974070221185684, + 0.00843840092420578, + 0.01228999998420477, + 0.013449658639729023, + 0.011132240295410156, + 0.023290477693080902, + 0.012696007266640663, + 0.15653583407402039, + 0.023948654532432556, + 0.07281950861215591, + 0.014150498434901237, + 0.020957836881279945, + 0.07632273435592651, + 0.0038057335186749697, + 0.0034831229131668806, + 0.0038872871082276106, + 0.002298698527738452, + 0.005830551031976938, + 0.004349902272224426, + 0.002733496017754078, + 0.0032774426508694887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.005262807011604309 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031922117341309786 + ], + "opposite_logprob_tokens": [ + -5.75319242477417 + ], + "logprob_reward": -0.0031922117341309786, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031922117341309786, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053010862320661545, + 0.05130520835518837, + 0.040531642735004425, + 0.08647283911705017, + 0.057686589658260345, + 0.019018815830349922, + 0.04722503200173378, + 0.06271801143884659, + 0.5551025867462158, + 0.026928409934043884 + ], + "credit_map": [ + 0.0059885685332119465, + 0.007341832388192415, + 0.009525233879685402, + 0.00579206133261323, + 0.005366711877286434, + 0.006634692661464214, + 0.007168229669332504, + 0.005193531978875399, + 0.003817988559603691, + 0.014166416600346565, + 0.0036669564433395863, + 0.004179155454039574, + 0.003709023352712393, + 0.010128021240234375, + 0.004507082514464855, + 0.007130561862140894, + 0.005699341185390949, + 0.0029884078539907932, + 0.00554606132209301, + 0.005460778251290321, + 0.00496156420558691, + 0.005038691684603691, + 0.006547666620463133, + 0.004289126954972744, + 0.004297980573028326, + 0.005345657002180815, + 0.01093370746821165, + 0.02215258963406086, + 0.0038145023863762617, + 0.007526872213929892, + 0.003422905458137393, + 0.0289786197245121, + 0.00554850697517395, + 0.007339400704950094, + 0.006293473765254021, + 0.008253665640950203, + 0.005259058903902769, + 0.004727926570922136, + 0.007813395000994205, + 0.012451160699129105, + 0.0022010034881532192, + 0.001653888844884932, + 0.003142115194350481, + 0.002831705380231142, + 0.0020229145884513855, + 0.0024675384629517794, + 0.0025856555439531803, + 0.002113993978127837, + 0.005224829073995352, + 0.006660023704171181, + 0.00475468672811985, + 0.0046011111699044704, + 0.008166608400642872, + 0.006623754743486643, + 0.005481233820319176, + 0.005712783895432949, + 0.00920584611594677, + 0.008507373742759228, + 0.006576624233275652, + 0.007618822157382965, + 0.007743376307189465, + 0.006916918326169252, + 0.007667918223887682, + 0.008481133729219437, + 0.032339952886104584, + 0.015779031440615654, + 0.20863531529903412, + 0.03295878693461418, + 0.0853191614151001, + 0.020552773028612137, + 0.034044165164232254, + 0.12547339498996735, + 0.0034492909908294678, + 0.003453608136624098, + 0.004181557334959507, + 0.0029097292572259903, + 0.003824736922979355, + 0.004629694391041994, + 0.002347748028114438, + 0.0021320439409464598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 3, + 7 + ], + "selected_reward": -0.0031922117341309786 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535471469163895 + ], + "opposite_logprob_tokens": [ + -4.387535572052002 + ], + "logprob_reward": -0.012535471469163895, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535471469163895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05256000906229019, + 0.021107640117406845, + 0.024679463356733322, + 0.015056242235004902, + 0.02516951784491539, + 0.02044624090194702, + 0.050458621233701706, + 0.5357264280319214, + 0.20890533924102783, + 0.045890480279922485 + ], + "credit_map": [ + 0.007422858849167824, + 0.006947963032871485, + 0.007616263348609209, + 0.006586933042854071, + 0.005912939552217722, + 0.007030115928500891, + 0.00512033561244607, + 0.0059225973673164845, + 0.0017705747159197927, + 0.0028699457179754972, + 0.002947085304185748, + 0.00239584781229496, + 0.002589291660115123, + 0.0036078242119401693, + 0.0025304590817540884, + 0.002396610099822283, + 0.004462084732949734, + 0.0017756144516170025, + 0.0035304317716509104, + 0.00388889922760427, + 0.0026242684107273817, + 0.0031573681626468897, + 0.0023386524990200996, + 0.0029021447990089655, + 0.001633622101508081, + 0.001505350461229682, + 0.001832165289670229, + 0.0023660690058022738, + 0.0014616275439038873, + 0.0015510611701756716, + 0.0013579968363046646, + 0.0033483502920717, + 0.0031219073571264744, + 0.0030248193070292473, + 0.0038170793559402227, + 0.0039423927664756775, + 0.0028094768058508635, + 0.0020641067530959845, + 0.0033146918285638094, + 0.003075043670833111, + 0.002999203745275736, + 0.0022693427745252848, + 0.0025563486851751804, + 0.0038824251387268305, + 0.001775320153683424, + 0.0024378777015954256, + 0.0023801818024367094, + 0.0021455390378832817, + 0.005670810583978891, + 0.006100643891841173, + 0.0037874476984143257, + 0.003980807960033417, + 0.010459830053150654, + 0.006562171503901482, + 0.005979468580335379, + 0.007917444221675396, + 0.06727096438407898, + 0.06837057322263718, + 0.06109767034649849, + 0.08061624318361282, + 0.059892576187849045, + 0.05659351125359535, + 0.07037035375833511, + 0.0715145692229271, + 0.014260884374380112, + 0.008436505682766438, + 0.08683713525533676, + 0.013878262601792812, + 0.032830264419317245, + 0.008023479953408241, + 0.009842622093856335, + 0.03479619696736336, + 0.006523569580167532, + 0.007992934435606003, + 0.005646693054586649, + 0.0035583835560828447, + 0.006207269616425037, + 0.007714366540312767, + 0.003859436372295022, + 0.004387828521430492 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.012535471469163895 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a vast aquarium featuring a massive artificial rock structure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a moray eel with mottled green-brown, leathery skin and sharp translucent teeth visible in a dark crevice?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the camera start with a low-angle close-up and slowly tilt upward toward the water surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are dancing caustic light patterns visible rippling downward through the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in cool aquatic colors of deep blue and hazy gray?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00673630740493536 + ], + "opposite_logprob_tokens": [ + -5.0067362785339355 + ], + "logprob_reward": -0.00673630740493536, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00673630740493536, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040263280272483826, + 0.04076632857322693, + 0.05178103968501091, + 0.03503410518169403, + 0.03185387700796127, + 0.017206059768795967, + 0.05704797804355621, + 0.5327902436256409, + 0.12616437673568726, + 0.06709273904561996 + ], + "credit_map": [ + 0.0068742320872843266, + 0.004924178123474121, + 0.008209379389882088, + 0.0034796043764799833, + 0.00389806623570621, + 0.004566896241158247, + 0.004354767967015505, + 0.003956158179789782, + 0.0029931478202342987, + 0.005396097898483276, + 0.005564714781939983, + 0.005607606377452612, + 0.005011693574488163, + 0.006401645950973034, + 0.004298666026443243, + 0.005492753814905882, + 0.00736901955679059, + 0.004092066083103418, + 0.007756360340863466, + 0.0073970588855445385, + 0.00713997520506382, + 0.005579142365604639, + 0.0076248240657150745, + 0.004822591319680214, + 0.001882202923297882, + 0.0027450150810182095, + 0.0047265226021409035, + 0.005213893949985504, + 0.003482005326077342, + 0.005229828413575888, + 0.003831302747130394, + 0.00792333297431469, + 0.0032488922588527203, + 0.0030886444728821516, + 0.005267136264592409, + 0.004538614768534899, + 0.003171490738168359, + 0.003342872951179743, + 0.005105364602059126, + 0.004090860951691866, + 0.0021092931274324656, + 0.002161617623642087, + 0.0018309017177671194, + 0.0023382108192890882, + 0.0022716776002198458, + 0.0020438218489289284, + 0.0021094761323183775, + 0.002341061132028699, + 0.007259824778884649, + 0.012555413879454136, + 0.005240079015493393, + 0.0058593531139194965, + 0.01013218518346548, + 0.005504960659891367, + 0.005556208547204733, + 0.004939955659210682, + 0.06278848648071289, + 0.06722727417945862, + 0.05949496477842331, + 0.08309667557477951, + 0.05840745568275452, + 0.05927519500255585, + 0.07125450670719147, + 0.07124567031860352, + 0.02236889861524105, + 0.011612690053880215, + 0.011952962726354599, + 0.007735629566013813, + 0.026600930839776993, + 0.008109941147267818, + 0.011644229292869568, + 0.02613908052444458, + 0.008745316416025162, + 0.010845848359167576, + 0.007778805680572987, + 0.005168679635971785, + 0.010793115012347698, + 0.012937293387949467, + 0.0051788571290671825, + 0.005644822958856821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 9 + ], + "selected_reward": -0.00673630740493536 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_tensor([6523, 6523])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_tensor([6523, 6523])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..92393a1fa45b47bbcb0e480fd630570f2cdc24b0 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_tensor([6523, 6523])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "Through a static, low-angle close-up, the camera captures the peak of athletic prowess as a balding, undersized, adult East Asian male athlete executes a flawless hurdle jump. His body, a compact engine of muscle, defined by the intense glare of the overhead stadium lights, is clad in a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts that cling to his powerful frame. Sweat gleams on his scalp and wrinkled forehead, his face a mask of pure, unwavering concentration. Trapped in mid-air, suspended for a moment, his form is a study in dynamic precision; his front leg, stretched straight and high, foot bent, just clear of the hurdle's white top bar, while his back leg is tucked tightly beside his hip. His arms swing in perfect, synchronized counter-swings, maintaining balance as he advances. The vibrant Simpson red synthetic track stretches beneath him, its grainy texture and crisp white lane lines providing a stark, high-contrast backdrop against the softly blurred stadium stands. The entire scene is imbued with an atmosphere of intense focus and explosive energy, immortalizing pure, unadulterated athletic grace and power.", + "seed": 10949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_6523_seed_10949_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.075933408876881, + "vlm_logprob": -0.075933408876881, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.075933408876881, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05084725841879845 + ], + "opposite_logprob_tokens": [ + -3.050847291946411 + ], + "logprob_reward": -0.05084725841879845, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05084725841879845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37724563479423523, + 0.04756270721554756, + 0.021290989592671394, + 0.08491487801074982, + 0.1430133879184723, + 0.060181405395269394, + 0.06761783361434937, + 0.11483746767044067, + 0.026855478063225746, + 0.05648011714220047 + ], + "credit_map": [ + 0.03585892170667648, + 0.035278793424367905, + 0.06913940608501434, + 0.07178029417991638, + 0.025424541905522346, + 0.030785832554101944, + 0.019492026418447495, + 0.08948589116334915, + 0.004740454722195864, + 0.005263808649033308, + 0.007168363779783249, + 0.007122251205146313, + 0.003529037581756711, + 0.005050204694271088, + 0.004344506189227104, + 0.01034408900886774, + 0.002364155137911439, + 0.0027695405296981335, + 0.0038574521895498037, + 0.003991146106272936, + 0.002007216913625598, + 0.0027014431543648243, + 0.0018752613104879856, + 0.001724777976050973, + 0.005343030206859112, + 0.007500621490180492, + 0.025140484794974327, + 0.024416768923401833, + 0.003655829466879368, + 0.0066324942745268345, + 0.004022108856588602, + 0.008203559555113316, + 0.01344733964651823, + 0.02572055347263813, + 0.027684587985277176, + 0.029037104919552803, + 0.008498295210301876, + 0.012061557732522488, + 0.012024118565022945, + 0.014539864845573902, + 0.004463459365069866, + 0.005504158791154623, + 0.01289331167936325, + 0.012747898697853088, + 0.004189038649201393, + 0.009070977568626404, + 0.0042169466614723206, + 0.007095626089721918, + 0.00967475213110447, + 0.018364807590842247, + 0.015227960422635078, + 0.010667691938579082, + 0.004492281004786491, + 0.0034299471881240606, + 0.0024711168371140957, + 0.003289288841187954, + 0.014770755544304848, + 0.032761652022600174, + 0.031169749796390533, + 0.017424747347831726, + 0.004638598766177893, + 0.006007279735058546, + 0.003484189510345459, + 0.0045805131085217, + 0.0033550309017300606, + 0.00505735632032156, + 0.004270393867045641, + 0.006478030700236559, + 0.0016973036108538508, + 0.001915551838465035, + 0.0018925181357190013, + 0.0021892969962209463, + 0.00514416815713048, + 0.0067269466817379, + 0.00830064620822668, + 0.01126562524586916, + 0.00531987426802516, + 0.0078054494224488735, + 0.006390731316059828, + 0.0055266921408474445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.05084725841879845 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07901059091091156 + ], + "opposite_logprob_tokens": [ + -2.579010486602783 + ], + "logprob_reward": -0.07901059091091156, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07901059091091156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20699289441108704, + 0.03973430395126343, + 0.01896282471716404, + 0.047487299889326096, + 0.2816108167171478, + 0.07228943705558777, + 0.03666329011321068, + 0.19299666583538055, + 0.05278163403272629, + 0.050480812788009644 + ], + "credit_map": [ + 0.020403413102030754, + 0.019143830984830856, + 0.04838566854596138, + 0.047879237681627274, + 0.013073270209133625, + 0.013454761356115341, + 0.009682808071374893, + 0.03496989607810974, + 0.003675188636407256, + 0.005290172528475523, + 0.007503313012421131, + 0.006627019960433245, + 0.003472824115306139, + 0.0035942713730037212, + 0.003648658748716116, + 0.005922854878008366, + 0.0022564169485121965, + 0.0023154865484684706, + 0.003454632591456175, + 0.003557346062734723, + 0.0015753619372844696, + 0.002092004520818591, + 0.0016901142662391067, + 0.0020214607939124107, + 0.0026745733339339495, + 0.0034346203319728374, + 0.01553276740014553, + 0.015253203921020031, + 0.0016776956617832184, + 0.002921320730820298, + 0.001979254186153412, + 0.004013862460851669, + 0.026374008506536484, + 0.05453268066048622, + 0.051717765629291534, + 0.061903756111860275, + 0.01575472019612789, + 0.021236253902316093, + 0.020805856212973595, + 0.029285795986652374, + 0.005798849742859602, + 0.006443576421588659, + 0.016080019995570183, + 0.016742974519729614, + 0.003655793610960245, + 0.011321625672280788, + 0.005156784784048796, + 0.007089811377227306, + 0.004766165744513273, + 0.00973488762974739, + 0.008230713196098804, + 0.005873695015907288, + 0.0024743410758674145, + 0.002189408987760544, + 0.0015138847520574927, + 0.0018801948754116893, + 0.025764193385839462, + 0.05738062039017677, + 0.0531209297478199, + 0.027663009241223335, + 0.008104884997010231, + 0.00847155973315239, + 0.006090487353503704, + 0.006400987505912781, + 0.0049498132430016994, + 0.007703409530222416, + 0.008110570721328259, + 0.009691009297966957, + 0.005639124661684036, + 0.005927384831011295, + 0.006127876695245504, + 0.004632444120943546, + 0.004666916560381651, + 0.006233728025108576, + 0.007418818771839142, + 0.00997377559542656, + 0.004735306371003389, + 0.006473473273217678, + 0.005681314505636692, + 0.005297479219734669 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.07901059091091156 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256108969449997 + ], + "opposite_logprob_tokens": [ + -1.6006108522415161 + ], + "logprob_reward": -0.2256108969449997, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256108969449997, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31472858786582947, + 0.053085606545209885, + 0.019564056769013405, + 0.09793054312467575, + 0.2760593295097351, + 0.03342806547880173, + 0.037966106086969376, + 0.08153829723596573, + 0.025212779641151428, + 0.06048652529716492 + ], + "credit_map": [ + 0.026527512818574905, + 0.02666875720024109, + 0.058555182069540024, + 0.06208420544862747, + 0.019665878266096115, + 0.026858340948820114, + 0.01669939234852791, + 0.07766934484243393, + 0.004433833993971348, + 0.005198624450713396, + 0.008359924890100956, + 0.008468042127788067, + 0.0038856733590364456, + 0.005915703251957893, + 0.005288452375680208, + 0.01153535395860672, + 0.00195033208001405, + 0.003070482751354575, + 0.004058732185512781, + 0.0035009540151804686, + 0.0012958607403561473, + 0.0018771222094073892, + 0.0018978214357048273, + 0.0019127525156363845, + 0.006232455372810364, + 0.00916489027440548, + 0.027930855751037598, + 0.027196004986763, + 0.004360295366495848, + 0.007559462916105986, + 0.005061048083007336, + 0.010425537824630737, + 0.025567760691046715, + 0.05314432457089424, + 0.052491724491119385, + 0.06071768328547478, + 0.015198971144855022, + 0.021153587847948074, + 0.020838510245084763, + 0.026946814730763435, + 0.002916257129982114, + 0.003501886734738946, + 0.005647588055580854, + 0.00621245801448822, + 0.0031005830969661474, + 0.004844257142394781, + 0.0032366495579481125, + 0.003968391567468643, + 0.004697720520198345, + 0.008876161649823189, + 0.009224113076925278, + 0.0061494335532188416, + 0.001905927900224924, + 0.0013844266068190336, + 0.0014483542181551456, + 0.004279970657080412, + 0.009845678694546223, + 0.021892322227358818, + 0.021060578525066376, + 0.01246787142008543, + 0.003420846536755562, + 0.0049780672416090965, + 0.003347810823470354, + 0.004525129217654467, + 0.003079476533457637, + 0.004585119895637035, + 0.004318902734667063, + 0.00497756851837039, + 0.0020381745416671038, + 0.0019421055912971497, + 0.002439083531498909, + 0.0018323505064472556, + 0.00568514596670866, + 0.006684737280011177, + 0.008530800230801105, + 0.011081619188189507, + 0.005891102831810713, + 0.008651173673570156, + 0.007234347518533468, + 0.00672760559245944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.2256108969449997 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036162238102406263 + ], + "opposite_logprob_tokens": [ + -5.6286163330078125 + ], + "logprob_reward": -0.0036162238102406263, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036162238102406263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3797064423561096, + 0.11367751657962799, + 0.0321616567671299, + 0.06631183624267578, + 0.1304784119129181, + 0.03901135176420212, + 0.03156837448477745, + 0.12212780117988586, + 0.04956428334116936, + 0.03539234772324562 + ], + "credit_map": [ + 0.04176072031259537, + 0.04452752321958542, + 0.06502052396535873, + 0.0667983666062355, + 0.02985183149576187, + 0.032476119697093964, + 0.020502666011452675, + 0.07876867055892944, + 0.012137068435549736, + 0.011651461012661457, + 0.02306351810693741, + 0.020658865571022034, + 0.007079609204083681, + 0.011822590604424477, + 0.007413170766085386, + 0.019851241260766983, + 0.004019217099994421, + 0.004770971834659576, + 0.004868319723755121, + 0.006444144528359175, + 0.002809933852404356, + 0.003205088898539543, + 0.0030134678818285465, + 0.0030305099207907915, + 0.005114343483000994, + 0.008648616261780262, + 0.01768713817000389, + 0.01765606738626957, + 0.003118158085271716, + 0.004559548571705818, + 0.003979820758104324, + 0.005548145156353712, + 0.011863064020872116, + 0.025270527228713036, + 0.027782142162322998, + 0.028119640424847603, + 0.007357300259172916, + 0.009034330025315285, + 0.0094137042760849, + 0.011637703515589237, + 0.004444296471774578, + 0.006036707665771246, + 0.006459137890487909, + 0.008715696632862091, + 0.002665750216692686, + 0.004302517045289278, + 0.003149139927700162, + 0.0032381045166403055, + 0.003679252928122878, + 0.006513865198940039, + 0.006466302089393139, + 0.0062665487639606, + 0.0019428705563768744, + 0.0023927746806293726, + 0.0018997046863660216, + 0.0024070520885288715, + 0.016036594286561012, + 0.03554551303386688, + 0.03317307308316231, + 0.018607528880238533, + 0.005041697062551975, + 0.005394360050559044, + 0.003865292761474848, + 0.0044637406244874, + 0.006209160666912794, + 0.008349443785846233, + 0.009161137975752354, + 0.015246494673192501, + 0.0023279734887182713, + 0.002445906400680542, + 0.002473840955644846, + 0.0033503249287605286, + 0.004496367648243904, + 0.003474860917776823, + 0.006128678098320961, + 0.005372434854507446, + 0.002292502438649535, + 0.004127373453229666, + 0.0029423998203128576, + 0.006557731423527002 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.0036162238102406263 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02058207429945469 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.02058207429945469, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02058207429945469, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2681836485862732, + 0.1320890337228775, + 0.07081876695156097, + 0.07308852672576904, + 0.07162546366453171, + 0.05565207451581955, + 0.08013707399368286, + 0.16180966794490814, + 0.04242214933037758, + 0.044173624366521835 + ], + "credit_map": [ + 0.04447000101208687, + 0.041218314319849014, + 0.05151163786649704, + 0.04693949222564697, + 0.02221635729074478, + 0.021512867882847786, + 0.015665678307414055, + 0.02464929036796093, + 0.01151233445852995, + 0.015084837563335896, + 0.023988641798496246, + 0.021778183057904243, + 0.009692724794149399, + 0.014071479439735413, + 0.011590959504246712, + 0.024369869381189346, + 0.007993689738214016, + 0.010618738830089569, + 0.011592168360948563, + 0.012561498209834099, + 0.006706247106194496, + 0.008461159653961658, + 0.0065438770689070225, + 0.006341388914734125, + 0.005102000664919615, + 0.007737281732261181, + 0.02005932480096817, + 0.019323699176311493, + 0.0036061701830476522, + 0.006543688476085663, + 0.003966215997934341, + 0.006750146392732859, + 0.006292905192822218, + 0.012039002031087875, + 0.015481575392186642, + 0.014550785534083843, + 0.004478514194488525, + 0.00581227196380496, + 0.005831392016261816, + 0.0071390243247151375, + 0.004669040907174349, + 0.005402314011007547, + 0.010848167352378368, + 0.012240155600011349, + 0.004025720525532961, + 0.0075919730588793755, + 0.00490752886980772, + 0.0059671783819794655, + 0.011672121472656727, + 0.019528137519955635, + 0.016482846811413765, + 0.012717701494693756, + 0.0050799488089978695, + 0.005201615858823061, + 0.004383603576570749, + 0.0050711012445390224, + 0.021081766113638878, + 0.04185783863067627, + 0.03909046947956085, + 0.023737598210573196, + 0.008225752972066402, + 0.012257619760930538, + 0.007283192127943039, + 0.008275418542325497, + 0.0054877945221960545, + 0.0060043311677873135, + 0.0076063526794314384, + 0.011813940480351448, + 0.0030269220005720854, + 0.0019919276237487793, + 0.002711269073188305, + 0.0037796134129166603, + 0.005188551731407642, + 0.0055729905143380165, + 0.006115212570875883, + 0.005299286916851997, + 0.0031212421599775553, + 0.005100471433252096, + 0.004476901143789291, + 0.009298969991505146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.02058207429945469 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a static, low-angle close-up, the camera captures the peak of athletic prowess as a balding, undersized, adult East Asian male athlete executes a flawless hurdle jump. His body, a compact engine of muscle, defined by the intense glare of the overhead stadium lights, is clad in a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts that cling to his powerful frame. Sweat gleams on his scalp and wrinkled forehead, his face a mask of pure, unwavering concentration. Trapped in mid-air, suspended for a moment, his form is a study in dynamic precision; his front leg, stretched straight and high, foot bent, just clear of the hurdle's white top bar, while his back leg is tucked tightly beside his hip. His arms swing in perfect, synchronized counter-swings, maintaining balance as he advances. The vibrant Simpson red synthetic track stretches beneath him, its grainy texture and crisp white lane lines providing a stark, high-contrast backdrop against the softly blurred stadium stands. The entire scene is imbued with an atmosphere of intense focus and explosive energy, immortalizing pure, unadulterated athletic grace and power.", + "seed": 10950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_6523_seed_10950_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.10770481890067458, + "vlm_logprob": -0.10770481890067458, + "vlm_yes_no_margin": 3.0, + "vlm_token_credit": -0.10770481890067458, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.388907790184021 + ], + "opposite_logprob_tokens": [ + -1.138907790184021 + ], + "logprob_reward": -0.388907790184021, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.388907790184021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2952796220779419, + 0.08062395453453064, + 0.05708696320652962, + 0.0620507188141346, + 0.09124402701854706, + 0.08328545838594437, + 0.23308853805065155, + 0.02930968441069126, + 0.045370928943157196, + 0.02266010455787182 + ], + "credit_map": [ + 0.026735736057162285, + 0.02890724129974842, + 0.0367862768471241, + 0.04420984908938408, + 0.06751959025859833, + 0.025915171951055527, + 0.030632924288511276, + 0.03457280248403549, + 0.0075477855280041695, + 0.009812440723180771, + 0.008839436806738377, + 0.008521847426891327, + 0.022014040499925613, + 0.00787162035703659, + 0.007378230337053537, + 0.008638545870780945, + 0.005260716192424297, + 0.006866846699267626, + 0.006726797204464674, + 0.009117163717746735, + 0.007647717837244272, + 0.008747878484427929, + 0.005188499111682177, + 0.007531336508691311, + 0.006018635351210833, + 0.005585685838013887, + 0.008892192505300045, + 0.010720822960138321, + 0.008004237897694111, + 0.00805285107344389, + 0.00591849023476243, + 0.008857800625264645, + 0.018611757084727287, + 0.008178514428436756, + 0.014932100661098957, + 0.0038054268807172775, + 0.0099040437489748, + 0.019653303548693657, + 0.006507922895252705, + 0.009650948457419872, + 0.010234124027192593, + 0.009828933514654636, + 0.011630008928477764, + 0.006609686184674501, + 0.012342100031673908, + 0.010842478834092617, + 0.00843950267881155, + 0.013358610682189465, + 0.01923239789903164, + 0.025290770456194878, + 0.0363052673637867, + 0.015272829681634903, + 0.042837705463171005, + 0.03394059091806412, + 0.017493480816483498, + 0.04271549731492996, + 0.0029327471274882555, + 0.006227695383131504, + 0.004089110065251589, + 0.00241615017876029, + 0.003936197143048048, + 0.0034246277064085007, + 0.0014082173584029078, + 0.004874936304986477, + 0.004010182339698076, + 0.005718569736927748, + 0.0059613934718072414, + 0.004625266883522272, + 0.006133233662694693, + 0.007435687817633152, + 0.004149114713072777, + 0.007337478920817375, + 0.0032274744007736444, + 0.00276493513956666, + 0.002998248441144824, + 0.0023605667520314455, + 0.0025672554038465023, + 0.0027201101183891296, + 0.002174900146201253, + 0.0038466115947812796 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.388907790184021 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05499108135700226 + ], + "opposite_logprob_tokens": [ + -2.9299910068511963 + ], + "logprob_reward": -0.05499108135700226, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05499108135700226, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09715010225772858, + 0.02256029285490513, + 0.024653276428580284, + 0.6884482502937317, + 0.021854421123862267, + 0.016927607357501984, + 0.02406228706240654, + 0.017509909346699715, + 0.05970663204789162, + 0.027127254754304886 + ], + "credit_map": [ + 0.014212407171726227, + 0.009772402234375477, + 0.015202420763671398, + 0.01105944998562336, + 0.01835515908896923, + 0.008384793996810913, + 0.006248190999031067, + 0.01391527708619833, + 0.0029964155983179808, + 0.0034771468490362167, + 0.002641563769429922, + 0.0018176912562921643, + 0.004123407881706953, + 0.0026016251649707556, + 0.0020331686828285456, + 0.0028692723717540503, + 0.0031821904703974724, + 0.0021704176906496286, + 0.0031315493397414684, + 0.002814160892739892, + 0.0028463646303862333, + 0.004966726992279291, + 0.0023301003966480494, + 0.0032117643859237432, + 0.06413181126117706, + 0.052196204662323, + 0.10325326025485992, + 0.1431902050971985, + 0.05927405133843422, + 0.0789736807346344, + 0.049345750361680984, + 0.13808320462703705, + 0.0032852236181497574, + 0.001831349916756153, + 0.003217178862541914, + 0.0011870302259922028, + 0.0024935558903962374, + 0.0038962357211858034, + 0.0018499320140108466, + 0.004093911498785019, + 0.0016232701018452644, + 0.0017076685326173902, + 0.0028699089307338, + 0.0014282232150435448, + 0.002334881341084838, + 0.0023493256885558367, + 0.0012756084324792027, + 0.0033387206494808197, + 0.0027553304098546505, + 0.002923649037256837, + 0.0037622868549078703, + 0.0018848759355023503, + 0.0026659129653126, + 0.0042599900625646114, + 0.0015284441178664565, + 0.004281796980649233, + 0.003912825137376785, + 0.0020912184845656157, + 0.002429375657811761, + 0.0011476028012111783, + 0.0021915773395448923, + 0.002006140071898699, + 0.0010690842755138874, + 0.002662084763869643, + 0.008339454419910908, + 0.003562599653378129, + 0.009042133577167988, + 0.0038920589722692966, + 0.003593540983274579, + 0.010011228732764721, + 0.0029278213623911142, + 0.018337789922952652, + 0.0033555165864527225, + 0.0021562371402978897, + 0.004650984425097704, + 0.0022617692593485117, + 0.0019207012373954058, + 0.0036686582025140524, + 0.0017550860065966845, + 0.0073582991026341915 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 8 + ], + "selected_reward": -0.05499108135700226 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054977767169475555 + ], + "opposite_logprob_tokens": [ + -2.9299776554107666 + ], + "logprob_reward": -0.054977767169475555, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054977767169475555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13844606280326843, + 0.04524797946214676, + 0.09860029816627502, + 0.23568564653396606, + 0.14447985589504242, + 0.03964107111096382, + 0.04349130392074585, + 0.0980396717786789, + 0.06987955421209335, + 0.08648848533630371 + ], + "credit_map": [ + 0.021524062380194664, + 0.016149042174220085, + 0.015906061977148056, + 0.021462293341755867, + 0.020299550145864487, + 0.011482108384370804, + 0.009801427833735943, + 0.021821536123752594, + 0.007119182962924242, + 0.006056966260075569, + 0.0056693353690207005, + 0.004019429441541433, + 0.0077169789001345634, + 0.0064695654436945915, + 0.002439866540953517, + 0.005756659433245659, + 0.011435816995799541, + 0.00952976942062378, + 0.013481780886650085, + 0.010615408420562744, + 0.007874274626374245, + 0.020273568108677864, + 0.006124031264334917, + 0.019265657290816307, + 0.02042585238814354, + 0.01982317678630352, + 0.037109967321157455, + 0.04012008383870125, + 0.022254526615142822, + 0.030109215527772903, + 0.01524425484240055, + 0.05059857293963432, + 0.013983293436467648, + 0.011026795022189617, + 0.02352638728916645, + 0.007826371118426323, + 0.015898652374744415, + 0.027077771723270416, + 0.014454579912126064, + 0.03068600967526436, + 0.0067695127800107, + 0.005064093042165041, + 0.006547618191689253, + 0.002492178464308381, + 0.003315581241622567, + 0.00566547317430377, + 0.0020724828355014324, + 0.007714134640991688, + 0.00423424644395709, + 0.005181755870580673, + 0.006069723051041365, + 0.003939296118915081, + 0.007628323044627905, + 0.008221662603318691, + 0.003852847497910261, + 0.004363454878330231, + 0.02003030851483345, + 0.019537974148988724, + 0.015742227435112, + 0.006344478111714125, + 0.012768149375915527, + 0.010716186836361885, + 0.005241569131612778, + 0.007658789865672588, + 0.008328423835337162, + 0.0046301637776196, + 0.01281590387225151, + 0.005608930718153715, + 0.005742336623370647, + 0.012177610769867897, + 0.002807710086926818, + 0.017768481746315956, + 0.009189891628921032, + 0.010089182294905186, + 0.015829401090741158, + 0.007886816747486591, + 0.008381133899092674, + 0.008687162771821022, + 0.0058205281384289265, + 0.02060438133776188 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 0 + ], + "selected_reward": -0.054977767169475555 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947511177510023 + ], + "opposite_logprob_tokens": [ + -5.130947589874268 + ], + "logprob_reward": -0.005947511177510023, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947511177510023, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09959784150123596, + 0.04222622141242027, + 0.08480861783027649, + 0.4978143870830536, + 0.05189615860581398, + 0.02711631916463375, + 0.09486246854066849, + 0.023615051060914993, + 0.04627816379070282, + 0.031784724444150925 + ], + "credit_map": [ + 0.010805848054587841, + 0.009778247214853764, + 0.013827784918248653, + 0.01412755623459816, + 0.018648186698555946, + 0.008467491716146469, + 0.008866621181368828, + 0.015076108276844025, + 0.006040414795279503, + 0.0058274175971746445, + 0.004876290913671255, + 0.004754063673317432, + 0.006147770211100578, + 0.005227821413427591, + 0.002593196230009198, + 0.006759248208254576, + 0.008906626142561436, + 0.006299351807683706, + 0.01087490189820528, + 0.009845559485256672, + 0.008498280309140682, + 0.015600153245031834, + 0.00474682729691267, + 0.020036915317177773, + 0.04748039320111275, + 0.039044756442308426, + 0.07414092123508453, + 0.10137051343917847, + 0.045406073331832886, + 0.055746693164110184, + 0.03812685236334801, + 0.09649816155433655, + 0.00947201531380415, + 0.0043640220537781715, + 0.007868483662605286, + 0.0024440307170152664, + 0.005423327442258596, + 0.010079039260745049, + 0.004879889078438282, + 0.007365346886217594, + 0.003266416722908616, + 0.0021765953861176968, + 0.004827835131436586, + 0.001997252693399787, + 0.003633348038420081, + 0.0048606921918690205, + 0.002245592186227441, + 0.004108586814254522, + 0.00956034567207098, + 0.008904402144253254, + 0.013123832643032074, + 0.007179599720984697, + 0.01536700688302517, + 0.018566276878118515, + 0.0076224301010370255, + 0.01453857496380806, + 0.0043858131393790245, + 0.0030552526004612446, + 0.003413757309317589, + 0.0017786708194762468, + 0.0036072467919439077, + 0.002586607588455081, + 0.0012240272480994463, + 0.0035636762622743845, + 0.005700585898011923, + 0.003063797950744629, + 0.007187464274466038, + 0.0039824871346354485, + 0.00473089748993516, + 0.00783469993621111, + 0.0030655006412416697, + 0.010712727904319763, + 0.0040760766714811325, + 0.0029193523805588484, + 0.0042965104803442955, + 0.0024868801701813936, + 0.003128217998892069, + 0.005957561545073986, + 0.0032050651498138905, + 0.005715059116482735 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.005947511177510023 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033699944615364075 + ], + "opposite_logprob_tokens": [ + -3.4086999893188477 + ], + "logprob_reward": -0.033699944615364075, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033699944615364075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13205650448799133, + 0.11934539675712585, + 0.11583948135375977, + 0.19187231361865997, + 0.0817255973815918, + 0.04909679666161537, + 0.0765758752822876, + 0.04309232905507088, + 0.11054988950490952, + 0.07984577119350433 + ], + "credit_map": [ + 0.018419770523905754, + 0.015003539621829987, + 0.015470007434487343, + 0.01784520037472248, + 0.01864173263311386, + 0.015353609807789326, + 0.012783230282366276, + 0.018539419397711754, + 0.01857464574277401, + 0.013811448588967323, + 0.013849505223333836, + 0.010526446625590324, + 0.02192062884569168, + 0.011893114075064659, + 0.011121307499706745, + 0.01764831133186817, + 0.01305060088634491, + 0.011574259027838707, + 0.016631530597805977, + 0.011270974762737751, + 0.01059738826006651, + 0.018345579504966736, + 0.006015754770487547, + 0.02835340052843094, + 0.023334242403507233, + 0.024705680087208748, + 0.02441520430147648, + 0.023962577804923058, + 0.025117626413702965, + 0.025227859616279602, + 0.018663622438907623, + 0.026445487514138222, + 0.014688227325677872, + 0.006150269880890846, + 0.014226306229829788, + 0.003953930921852589, + 0.008057668805122375, + 0.015055538155138493, + 0.0051161153241992, + 0.014477543532848358, + 0.007780082523822784, + 0.003887650091201067, + 0.008139452897012234, + 0.0031012578401714563, + 0.006242774892598391, + 0.009083247743546963, + 0.003546758322045207, + 0.007315571419894695, + 0.005405481439083815, + 0.007151328958570957, + 0.012982027605175972, + 0.007504331413656473, + 0.009113778360188007, + 0.01483418233692646, + 0.004187935497611761, + 0.015396800823509693, + 0.003930822946131229, + 0.0037050279788672924, + 0.006533102132380009, + 0.0037254851777106524, + 0.006549663841724396, + 0.006083476822823286, + 0.003203222993761301, + 0.009361527860164642, + 0.014507351443171501, + 0.00946502760052681, + 0.01624828763306141, + 0.009182961657643318, + 0.009456753730773926, + 0.020655564963817596, + 0.004908549599349499, + 0.02612539567053318, + 0.009824085049331188, + 0.004877889063209295, + 0.014498619362711906, + 0.006626930553466082, + 0.006027746479958296, + 0.01104721799492836, + 0.006166249979287386, + 0.020777041092514992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.033699944615364075 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a static, low-angle close-up, the camera captures the peak of athletic prowess as a balding, undersized, adult East Asian male athlete executes a flawless hurdle jump. His body, a compact engine of muscle, defined by the intense glare of the overhead stadium lights, is clad in a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts that cling to his powerful frame. Sweat gleams on his scalp and wrinkled forehead, his face a mask of pure, unwavering concentration. Trapped in mid-air, suspended for a moment, his form is a study in dynamic precision; his front leg, stretched straight and high, foot bent, just clear of the hurdle's white top bar, while his back leg is tucked tightly beside his hip. His arms swing in perfect, synchronized counter-swings, maintaining balance as he advances. The vibrant Simpson red synthetic track stretches beneath him, its grainy texture and crisp white lane lines providing a stark, high-contrast backdrop against the softly blurred stadium stands. The entire scene is imbued with an atmosphere of intense focus and explosive energy, immortalizing pure, unadulterated athletic grace and power.", + "seed": 10951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_6523_seed_10951_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10947742210701109, + "vlm_logprob": -0.10947742210701109, + "vlm_yes_no_margin": 3.024999976158142, + "vlm_token_credit": -0.10947742210701109, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.351410835981369 + ], + "opposite_logprob_tokens": [ + -1.2264108657836914 + ], + "logprob_reward": -0.351410835981369, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.351410835981369, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2462952584028244, + 0.07345619052648544, + 0.07489142566919327, + 0.03635355457663536, + 0.04659714177250862, + 0.14544786512851715, + 0.1594584584236145, + 0.10241595655679703, + 0.04877036064863205, + 0.0663137286901474 + ], + "credit_map": [ + 0.032197512686252594, + 0.06271466612815857, + 0.01758633181452751, + 0.036770399659872055, + 0.029781531542539597, + 0.020594581961631775, + 0.02456653118133545, + 0.022083723917603493, + 0.006319974083453417, + 0.005638333968818188, + 0.007139938417822123, + 0.011401688680052757, + 0.0075979409739375114, + 0.007464027497917414, + 0.012597797438502312, + 0.015296489000320435, + 0.010644644498825073, + 0.011877359822392464, + 0.007960161194205284, + 0.008262279443442822, + 0.00656037125736475, + 0.00988225731998682, + 0.00783616118133068, + 0.011868191882967949, + 0.004766819532960653, + 0.003835394512861967, + 0.0036439108662307262, + 0.004571614786982536, + 0.005098165012896061, + 0.002493741689249873, + 0.005313168745487928, + 0.006630740128457546, + 0.009864536114037037, + 0.005653440486639738, + 0.004352096933871508, + 0.0038360808975994587, + 0.00415278784930706, + 0.008337280713021755, + 0.004073196090757847, + 0.006327723152935505, + 0.016864998266100883, + 0.012686614878475666, + 0.01561149675399065, + 0.007723870221525431, + 0.016454068943858147, + 0.037699803709983826, + 0.01145243737846613, + 0.026954567059874535, + 0.03982969745993614, + 0.020125512033700943, + 0.019247455522418022, + 0.011686904355883598, + 0.013218803331255913, + 0.020579183474183083, + 0.011975985951721668, + 0.022794928401708603, + 0.011581267230212688, + 0.015494097955524921, + 0.012644192203879356, + 0.009384210221469402, + 0.014601992443203926, + 0.014823775738477707, + 0.010553725995123386, + 0.013332691043615341, + 0.007472655735909939, + 0.006234473083168268, + 0.0074352724477648735, + 0.00494385464116931, + 0.005883794277906418, + 0.004805657081305981, + 0.005608764011412859, + 0.006385891232639551, + 0.007825331762433052, + 0.016754304990172386, + 0.0051250397227704525, + 0.008777135983109474, + 0.010400375351309776, + 0.006425567902624607, + 0.00415427889674902, + 0.006851694546639919 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.351410835981369 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043077901005744934 + ], + "opposite_logprob_tokens": [ + -3.1680779457092285 + ], + "logprob_reward": -0.043077901005744934, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043077901005744934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12154717743396759, + 0.10481737554073334, + 0.04677343741059303, + 0.025071805343031883, + 0.04722055792808533, + 0.3012082576751709, + 0.052449870854616165, + 0.207545667886734, + 0.02885882370173931, + 0.06450698524713516 + ], + "credit_map": [ + 0.02309245988726616, + 0.011576286517083645, + 0.014945711940526962, + 0.017199181020259857, + 0.012208541855216026, + 0.012362092733383179, + 0.014278237707912922, + 0.015884682536125183, + 0.008532091975212097, + 0.011615738272666931, + 0.008316939696669579, + 0.016484912484884262, + 0.016973000019788742, + 0.007949861697852612, + 0.013924055732786655, + 0.021020779386162758, + 0.007207856513559818, + 0.00492529571056366, + 0.005464939400553703, + 0.006339613813906908, + 0.006713955197483301, + 0.004317828919738531, + 0.0030986464116722345, + 0.00870530679821968, + 0.0028820899315178394, + 0.003022151067852974, + 0.002433258341625333, + 0.0029226248152554035, + 0.0038226137403398752, + 0.00228721066378057, + 0.0037771814968436956, + 0.00392467575147748, + 0.012568642385303974, + 0.004617275204509497, + 0.005200603045523167, + 0.0035469643771648407, + 0.0035275935661047697, + 0.006453429814428091, + 0.004162255208939314, + 0.007143794093281031, + 0.03339353948831558, + 0.026843370869755745, + 0.032372888177633286, + 0.015340124256908894, + 0.0331963412463665, + 0.07802961766719818, + 0.023896681144833565, + 0.058135710656642914, + 0.013983286917209625, + 0.006104740314185619, + 0.007075476460158825, + 0.0040455106645822525, + 0.003760423045605421, + 0.005337617360055447, + 0.0036636979784816504, + 0.008479123003780842, + 0.02136581391096115, + 0.03148680552840233, + 0.022694271057844162, + 0.019319918006658554, + 0.033922843635082245, + 0.030695678666234016, + 0.025181984528899193, + 0.02287834882736206, + 0.004894412588328123, + 0.0030072215013206005, + 0.004800321068614721, + 0.002589423442259431, + 0.004024791996926069, + 0.00309893861413002, + 0.002888328395783901, + 0.003555387258529663, + 0.009684565477073193, + 0.0071258838288486, + 0.0076120588928461075, + 0.007083505857735872, + 0.010359402745962143, + 0.006516716443002224, + 0.006467483006417751, + 0.00965737085789442 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 0 + ], + "selected_reward": -0.043077901005744934 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706978619098663 + ], + "opposite_logprob_tokens": [ + -2.1270697116851807 + ], + "logprob_reward": -0.12706978619098663, + "yes_no_margin_reward": 1.9999998807907104, + "token_credit_reward": -0.12706978619098663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09738828986883163, + 0.036760516464710236, + 0.039377421140670776, + 0.0682273879647255, + 0.08003808557987213, + 0.44408854842185974, + 0.0711667537689209, + 0.03457389771938324, + 0.1022966057062149, + 0.026082482188940048 + ], + "credit_map": [ + 0.02070847526192665, + 0.01191392820328474, + 0.011408676393330097, + 0.013377017341554165, + 0.00852600485086441, + 0.008726709522306919, + 0.010808918625116348, + 0.01191855687648058, + 0.003353662556037307, + 0.004771091975271702, + 0.0030429826583713293, + 0.007724379189312458, + 0.004114019218832254, + 0.002583159366622567, + 0.005360011477023363, + 0.005811211187392473, + 0.005226007662713528, + 0.007201146800071001, + 0.00406281091272831, + 0.004894782789051533, + 0.0032117413356900215, + 0.004199557937681675, + 0.0047202929854393005, + 0.005861079320311546, + 0.012529239058494568, + 0.00672975042834878, + 0.007549199741333723, + 0.006708802655339241, + 0.008657919242978096, + 0.00725216930732131, + 0.0070355720818042755, + 0.011764729395508766, + 0.017634043470025063, + 0.0069046178832650185, + 0.007681193295866251, + 0.007433516904711723, + 0.004450765438377857, + 0.015461673960089684, + 0.009846040979027748, + 0.010626229457557201, + 0.04929760470986366, + 0.046818118542432785, + 0.04344905912876129, + 0.025188321247696877, + 0.04849480465054512, + 0.11782943457365036, + 0.0361214354634285, + 0.0768897607922554, + 0.01631867326796055, + 0.010702291503548622, + 0.008488759398460388, + 0.00589695805683732, + 0.005704296287149191, + 0.008876481093466282, + 0.005247931927442551, + 0.009931362234055996, + 0.0061844391748309135, + 0.003627658123150468, + 0.007391819730401039, + 0.0025491095148026943, + 0.003344279946759343, + 0.00432591512799263, + 0.00323122413828969, + 0.0039194505661726, + 0.015913885086774826, + 0.009341087192296982, + 0.024136126041412354, + 0.010185407474637032, + 0.00826577004045248, + 0.01169328298419714, + 0.005227494053542614, + 0.017533551901578903, + 0.004059589933604002, + 0.005102412309497595, + 0.00235390430316329, + 0.0033934558741748333, + 0.004019807558506727, + 0.0022867023944854736, + 0.0021211395505815744, + 0.0027454711962491274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 0 + ], + "selected_reward": -0.12706978619098663 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005251067224889994 + ], + "opposite_logprob_tokens": [ + -5.255250930786133 + ], + "logprob_reward": -0.005251067224889994, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005251067224889994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12392360717058182, + 0.06309718638658524, + 0.10300005227327347, + 0.032348670065402985, + 0.143898606300354, + 0.1921629160642624, + 0.04628097265958786, + 0.07123631983995438, + 0.18417826294898987, + 0.03987341374158859 + ], + "credit_map": [ + 0.01853216625750065, + 0.019923636689782143, + 0.010787801817059517, + 0.022923773154616356, + 0.01009331177920103, + 0.016094326972961426, + 0.00976122822612524, + 0.01580735854804516, + 0.012265028432011604, + 0.008156565017998219, + 0.007846291176974773, + 0.005924254655838013, + 0.005053667351603508, + 0.010523386299610138, + 0.00467588659375906, + 0.008652101270854473, + 0.02054201252758503, + 0.014938846230506897, + 0.010099753737449646, + 0.011415112763643265, + 0.010397338308393955, + 0.012810527347028255, + 0.007031633052974939, + 0.015764819458127022, + 0.00337964971549809, + 0.007335165981203318, + 0.003196652280166745, + 0.0034021560568362474, + 0.0030394638888537884, + 0.0038015537429600954, + 0.003291225992143154, + 0.004902800545096397, + 0.03825600817799568, + 0.01327954325824976, + 0.015319496393203735, + 0.011612968519330025, + 0.010297683998942375, + 0.020041095092892647, + 0.011872198432683945, + 0.023219598457217216, + 0.024522045627236366, + 0.020406058058142662, + 0.01929272897541523, + 0.010142291896045208, + 0.019185153767466545, + 0.05090917646884918, + 0.01345212385058403, + 0.034253332763910294, + 0.007194537669420242, + 0.005660620052367449, + 0.006498002912849188, + 0.0044890400022268295, + 0.005811150185763836, + 0.004926104564219713, + 0.004940946586430073, + 0.0067605674266815186, + 0.009820418432354927, + 0.009208745323121548, + 0.009901161305606365, + 0.008056245744228363, + 0.009843623265624046, + 0.009671407751739025, + 0.007247576955705881, + 0.007487135007977486, + 0.03277904912829399, + 0.016741471365094185, + 0.04561268538236618, + 0.016398046165704727, + 0.013282561674714088, + 0.02306104078888893, + 0.011188529431819916, + 0.025114867836236954, + 0.006823909468948841, + 0.0063068922609090805, + 0.004338116850703955, + 0.004685265012085438, + 0.006957876496016979, + 0.0031616927590221167, + 0.002787595149129629, + 0.0048120650462806225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 4 + ], + "selected_reward": -0.005251067224889994 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057752013206482 + ], + "opposite_logprob_tokens": [ + -3.8955774307250977 + ], + "logprob_reward": -0.02057752013206482, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057752013206482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17281512916088104, + 0.060068339109420776, + 0.04886025935411453, + 0.10045789927244186, + 0.08025716245174408, + 0.044548165053129196, + 0.2365243285894394, + 0.10311809927225113, + 0.12213847041130066, + 0.031212162226438522 + ], + "credit_map": [ + 0.019940709695219994, + 0.023281920701265335, + 0.020932337269186974, + 0.027709640562534332, + 0.016270698979496956, + 0.019573546946048737, + 0.025547515600919724, + 0.01955876499414444, + 0.0057579148560762405, + 0.0062989224679768085, + 0.005101845599710941, + 0.008763859048485756, + 0.006992094684392214, + 0.00490220682695508, + 0.0076574040576815605, + 0.014594090171158314, + 0.005795533303171396, + 0.007970347069203854, + 0.005512682255357504, + 0.006407754495739937, + 0.006071255076676607, + 0.004201281350106001, + 0.005162030458450317, + 0.007739374414086342, + 0.011388999409973621, + 0.013255641795694828, + 0.007356842048466206, + 0.015229480341076851, + 0.010538911446928978, + 0.008835559710860252, + 0.019482048228383064, + 0.014370408840477467, + 0.013982024043798447, + 0.011430424638092518, + 0.0062665510922670364, + 0.008348402567207813, + 0.011232267133891582, + 0.006801564246416092, + 0.011362941935658455, + 0.010832982137799263, + 0.007999029010534286, + 0.003936103079468012, + 0.005969741381704807, + 0.002884657820686698, + 0.004980525001883507, + 0.006679833400994539, + 0.004966944921761751, + 0.007131329737603664, + 0.06502677500247955, + 0.027308866381645203, + 0.03291420266032219, + 0.019732659682631493, + 0.014885134063661098, + 0.030746417120099068, + 0.013964519836008549, + 0.03194575384259224, + 0.012225168757140636, + 0.010841086506843567, + 0.016303865239024162, + 0.009903698228299618, + 0.012883227318525314, + 0.014836937189102173, + 0.011791778728365898, + 0.01433232706040144, + 0.014834593050181866, + 0.01073375716805458, + 0.02340731769800186, + 0.012629181146621704, + 0.014917638152837753, + 0.013714409433305264, + 0.016603656113147736, + 0.015297922305762768, + 0.005734738428145647, + 0.0029347629752010107, + 0.003999823238700628, + 0.0035149045288562775, + 0.005236928351223469, + 0.0028609586879611015, + 0.002930585527792573, + 0.003999460954219103 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 8 + ], + "selected_reward": -0.02057752013206482 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "Through a static, low-angle close-up, the camera captures the peak of athletic prowess as a balding, undersized, adult East Asian male athlete executes a flawless hurdle jump. His body, a compact engine of muscle, defined by the intense glare of the overhead stadium lights, is clad in a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts that cling to his powerful frame. Sweat gleams on his scalp and wrinkled forehead, his face a mask of pure, unwavering concentration. Trapped in mid-air, suspended for a moment, his form is a study in dynamic precision; his front leg, stretched straight and high, foot bent, just clear of the hurdle's white top bar, while his back leg is tucked tightly beside his hip. His arms swing in perfect, synchronized counter-swings, maintaining balance as he advances. The vibrant Simpson red synthetic track stretches beneath him, its grainy texture and crisp white lane lines providing a stark, high-contrast backdrop against the softly blurred stadium stands. The entire scene is imbued with an atmosphere of intense focus and explosive energy, immortalizing pure, unadulterated athletic grace and power.", + "seed": 10952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_12/index_6523_seed_10952_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.08333593774586916, + "vlm_logprob": -0.08333593774586916, + "vlm_yes_no_margin": 2.899999976158142, + "vlm_token_credit": -0.08333593774586916, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1828610748052597 + ], + "opposite_logprob_tokens": [ + -1.807861089706421 + ], + "logprob_reward": -0.1828610748052597, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1828610748052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10199803113937378, + 0.08835208415985107, + 0.13726721704006195, + 0.09790163487195969, + 0.049223221838474274, + 0.03211890906095505, + 0.1035299003124237, + 0.24044285714626312, + 0.0395820215344429, + 0.10958410799503326 + ], + "credit_map": [ + 0.020173249766230583, + 0.011168006807565689, + 0.019954262301325798, + 0.011432254686951637, + 0.007029963191598654, + 0.014288200065493584, + 0.007794295437633991, + 0.010157802142202854, + 0.015077774412930012, + 0.012176832184195518, + 0.011665686033666134, + 0.007279250770807266, + 0.006695183459669352, + 0.015314925462007523, + 0.009104724042117596, + 0.01103771012276411, + 0.017595404759049416, + 0.02420877479016781, + 0.01677577570080757, + 0.010825803503394127, + 0.01663641817867756, + 0.02412097528576851, + 0.009730403311550617, + 0.01737366057932377, + 0.009084504097700119, + 0.013540863059461117, + 0.0131691237911582, + 0.013212420046329498, + 0.01376988086849451, + 0.008850525133311749, + 0.011069650761783123, + 0.015204669907689095, + 0.005117656197398901, + 0.006406616419553757, + 0.006632319651544094, + 0.010342475958168507, + 0.004649250768125057, + 0.006655470933765173, + 0.004253090359270573, + 0.0051663401536643505, + 0.0055066850036382675, + 0.0031103186774998903, + 0.005450362339615822, + 0.0026117046363651752, + 0.0033039506524801254, + 0.00454871729016304, + 0.0037895957939326763, + 0.003797574434429407, + 0.007221716456115246, + 0.012095708400011063, + 0.013981597498059273, + 0.012033632025122643, + 0.01480092853307724, + 0.012921864166855812, + 0.011419770307838917, + 0.019054684787988663, + 0.015174700878560543, + 0.0889875739812851, + 0.09140782058238983, + 0.015061852522194386, + 0.00881256815046072, + 0.0067464569583535194, + 0.00590151222422719, + 0.008350379765033722, + 0.005348089151084423, + 0.005440105684101582, + 0.005710435565561056, + 0.007763180881738663, + 0.006078610196709633, + 0.003872971748933196, + 0.002199615351855755, + 0.0031690136529505253, + 0.013751592487096786, + 0.016084766015410423, + 0.011537996120750904, + 0.01157714705914259, + 0.020720411092042923, + 0.011896741576492786, + 0.012316408567130566, + 0.01169904787093401 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 9 + ], + "selected_reward": -0.1828610748052597 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308703541755676 + ], + "opposite_logprob_tokens": [ + -3.1680870056152344 + ], + "logprob_reward": -0.04308703541755676, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308703541755676, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08748321235179901, + 0.14073029160499573, + 0.2185036838054657, + 0.06959861516952515, + 0.06304687261581421, + 0.038565535098314285, + 0.04829086735844612, + 0.13968732953071594, + 0.04290575161576271, + 0.15118785202503204 + ], + "credit_map": [ + 0.015146925114095211, + 0.01122265588492155, + 0.012162034399807453, + 0.01085599884390831, + 0.008629385381937027, + 0.010610202327370644, + 0.008844183757901192, + 0.010011821985244751, + 0.021350370720028877, + 0.019942741841077805, + 0.020046699792146683, + 0.014292753301560879, + 0.011711034923791885, + 0.019925963133573532, + 0.013330784626305103, + 0.020129937678575516, + 0.02493140660226345, + 0.024576351046562195, + 0.02781687304377556, + 0.021022751927375793, + 0.026257384568452835, + 0.03407711163163185, + 0.028208071365952492, + 0.03161372244358063, + 0.007378598675131798, + 0.011272505857050419, + 0.008178393356502056, + 0.00971830915659666, + 0.008771023713052273, + 0.005635188892483711, + 0.007472111843526363, + 0.011172479949891567, + 0.007130674086511135, + 0.007586258929222822, + 0.009595145471394062, + 0.011685536243021488, + 0.00687188096344471, + 0.00767957279458642, + 0.004940544720739126, + 0.007557250559329987, + 0.006516528781503439, + 0.0059009818360209465, + 0.00568773690611124, + 0.00454355776309967, + 0.003976779989898205, + 0.004440412390977144, + 0.0035463955719023943, + 0.003953137435019016, + 0.004854721948504448, + 0.009780473075807095, + 0.007325079292058945, + 0.006420965306460857, + 0.004189891740679741, + 0.0037266123108565807, + 0.004945533815771341, + 0.007047589402645826, + 0.021075746044516563, + 0.029405275359749794, + 0.029132572934031487, + 0.013176729902625084, + 0.01102020125836134, + 0.009804029017686844, + 0.008577621541917324, + 0.017495131120085716, + 0.005455076694488525, + 0.006192246451973915, + 0.005890703294426203, + 0.008130445145070553, + 0.004616331774741411, + 0.004660876002162695, + 0.0036504894960671663, + 0.004309582989662886, + 0.02387414686381817, + 0.028796376660466194, + 0.0145769864320755, + 0.016586175188422203, + 0.026990685611963272, + 0.015611791983246803, + 0.011107304133474827, + 0.013644387014210224 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 1 + ], + "selected_reward": -0.04308703541755676 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14284572005271912 + ], + "opposite_logprob_tokens": [ + -2.017845630645752 + ], + "logprob_reward": -0.14284572005271912, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14284572005271912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1067984327673912, + 0.07405591011047363, + 0.09602769464254379, + 0.05612397938966751, + 0.16049732267856598, + 0.07389513403177261, + 0.1088511124253273, + 0.21095529198646545, + 0.049882881343364716, + 0.06291226297616959 + ], + "credit_map": [ + 0.016424698755145073, + 0.010344822891056538, + 0.01770312339067459, + 0.014872808940708637, + 0.008977298624813557, + 0.015086617320775986, + 0.008194636553525925, + 0.015194420702755451, + 0.010102055966854095, + 0.007534877397119999, + 0.009411188773810863, + 0.009855056181550026, + 0.008944070897996426, + 0.011357635259628296, + 0.007625251542776823, + 0.009225765243172646, + 0.010424855165183544, + 0.01873982883989811, + 0.012667358852922916, + 0.009809689596295357, + 0.012212532572448254, + 0.013766060583293438, + 0.009441970847547054, + 0.008965399116277695, + 0.005017573945224285, + 0.005159381311386824, + 0.007911324501037598, + 0.007569454610347748, + 0.006381493993103504, + 0.0053576030768454075, + 0.004720000084489584, + 0.014007145538926125, + 0.013839161023497581, + 0.017201008275151253, + 0.02631087228655815, + 0.030974969267845154, + 0.017546696588397026, + 0.019691022112965584, + 0.015539423562586308, + 0.0193941667675972, + 0.009967144578695297, + 0.01218236144632101, + 0.008856391534209251, + 0.011722996830940247, + 0.007025838829576969, + 0.01150919683277607, + 0.006598448846489191, + 0.006032750941812992, + 0.010166829451918602, + 0.02091454528272152, + 0.017102856189012527, + 0.01466840784996748, + 0.010515987873077393, + 0.010216472670435905, + 0.009456752799451351, + 0.015809258446097374, + 0.03418886661529541, + 0.044697124511003494, + 0.03661954775452614, + 0.023257242515683174, + 0.019094401970505714, + 0.018266556784510612, + 0.012782788835465908, + 0.022048745304346085, + 0.006704871542751789, + 0.007836859673261642, + 0.010939923115074635, + 0.005817373748868704, + 0.004264835733920336, + 0.006009467411786318, + 0.002916344441473484, + 0.005393204279243946, + 0.010904627852141857, + 0.013710599392652512, + 0.009128826670348644, + 0.006468179635703564, + 0.008577699773013592, + 0.005951419938355684, + 0.0041865017265081406, + 0.00398440845310688 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 6 + ], + "selected_reward": -0.14284572005271912 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009782706387341022 + ], + "opposite_logprob_tokens": [ + -4.634782791137695 + ], + "logprob_reward": -0.009782706387341022, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009782706387341022, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.057340823113918304, + 0.13512510061264038, + 0.05615019053220749, + 0.055105749517679214, + 0.07509840279817581, + 0.056226491928100586, + 0.03702738508582115, + 0.313105046749115, + 0.139183908700943, + 0.07563691586256027 + ], + "credit_map": [ + 0.011593335308134556, + 0.006369767244905233, + 0.011025414802134037, + 0.006837859749794006, + 0.004489653743803501, + 0.004808659665286541, + 0.005080231931060553, + 0.007135897409170866, + 0.01976441591978073, + 0.016661323606967926, + 0.019344596192240715, + 0.01303174439817667, + 0.013183497823774815, + 0.021632220596075058, + 0.013144438154995441, + 0.018362855538725853, + 0.007521461695432663, + 0.00961670745164156, + 0.007454624865204096, + 0.005215428303927183, + 0.005416864063590765, + 0.008386257104575634, + 0.004736312665045261, + 0.0078025320544838905, + 0.007398803252726793, + 0.008733898401260376, + 0.007370965555310249, + 0.008626051247119904, + 0.006933345925062895, + 0.0047570932656526566, + 0.0044637261889874935, + 0.006821863353252411, + 0.011662445031106472, + 0.011358041316270828, + 0.009977164678275585, + 0.011916439048945904, + 0.006322700530290604, + 0.008753836154937744, + 0.004837952554225922, + 0.010269826278090477, + 0.009780541993677616, + 0.00745818018913269, + 0.007615738548338413, + 0.00955104362219572, + 0.005009668879210949, + 0.006530536338686943, + 0.005239391699433327, + 0.005041387397795916, + 0.003961499314755201, + 0.005760158412158489, + 0.005368141457438469, + 0.006374837830662727, + 0.003243661019951105, + 0.004862640053033829, + 0.0025840681046247482, + 0.004872381687164307, + 0.03732797130942345, + 0.0937584713101387, + 0.09725474566221237, + 0.02291538193821907, + 0.01570723205804825, + 0.014615220949053764, + 0.014052819460630417, + 0.017473196610808372, + 0.022866472601890564, + 0.020427756011486053, + 0.028221329674124718, + 0.022063547745347023, + 0.011200135573744774, + 0.01379947829991579, + 0.008145792409777641, + 0.012459402903914452, + 0.012851310893893242, + 0.013481552712619305, + 0.015242258086800575, + 0.007225541863590479, + 0.01208691205829382, + 0.0062415278516709805, + 0.004592954646795988, + 0.003914862405508757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.009782706387341022 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a balding, undersized, adult East Asian male athlete?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the athlete wearing a royal blue, sleeveless, moisture-wicking sweatshirt and matching shorts?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the athlete captured in mid-air executing a hurdle jump with his front leg stretched straight and high over a white hurdle bar?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background a vibrant Simpson red synthetic track with crisp white lane lines and softly blurred stadium stands?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a static, low-angle close-up illuminated by intense overhead stadium lights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810315206646919 + ], + "opposite_logprob_tokens": [ + -3.2881031036376953 + ], + "logprob_reward": -0.03810315206646919, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810315206646919, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06485944241285324, + 0.15667149424552917, + 0.06758198887109756, + 0.03739968314766884, + 0.03892495110630989, + 0.03657101094722748, + 0.16631290316581726, + 0.22446607053279877, + 0.0872441828250885, + 0.11996829509735107 + ], + "credit_map": [ + 0.010025961324572563, + 0.006185235921293497, + 0.012141479179263115, + 0.0070920176804065704, + 0.005829528905451298, + 0.00768066244199872, + 0.006841171532869339, + 0.009063384495675564, + 0.026081083342432976, + 0.02151394449174404, + 0.021591341122984886, + 0.013847807422280312, + 0.012370124459266663, + 0.025884347036480904, + 0.015381754375994205, + 0.020001087337732315, + 0.009815976023674011, + 0.014263533987104893, + 0.007617488037794828, + 0.005892608314752579, + 0.007883446291089058, + 0.008967087604105473, + 0.006221954245120287, + 0.00691989716142416, + 0.004068097565323114, + 0.0045845177955925465, + 0.005179317202419043, + 0.00585009902715683, + 0.005471463780850172, + 0.0033834511414170265, + 0.0037808073684573174, + 0.005081928335130215, + 0.004824645351618528, + 0.006074553355574608, + 0.004957710858434439, + 0.007539169397205114, + 0.0038032757584005594, + 0.004338392056524754, + 0.0037839023862034082, + 0.0036033024080097675, + 0.00400416599586606, + 0.0036536725237965584, + 0.00534515455365181, + 0.004181940108537674, + 0.003615722293034196, + 0.005683435592800379, + 0.004506723489612341, + 0.005580196622759104, + 0.015218297019600868, + 0.0316394604742527, + 0.02539282664656639, + 0.01971147209405899, + 0.016816943883895874, + 0.016213396564126015, + 0.01441268716007471, + 0.026907822117209435, + 0.017160603776574135, + 0.07408173382282257, + 0.08178585767745972, + 0.014436684548854828, + 0.009169158525764942, + 0.009211230091750622, + 0.007080646697431803, + 0.011540141887962818, + 0.011726236902177334, + 0.014052345417439938, + 0.01398757379502058, + 0.016414344310760498, + 0.006468815729022026, + 0.010256021283566952, + 0.006187587510794401, + 0.008151261135935783, + 0.02200034074485302, + 0.027745164930820465, + 0.01609240099787712, + 0.009706316515803337, + 0.012858404777944088, + 0.014084498398005962, + 0.006571660749614239, + 0.01090950332581997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 1 + ], + "selected_reward": -0.03810315206646919 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_tensor([6556, 6556])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_tensor([6556, 6556])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..60ebdb78855d0b47f40a3adc9b70423ca9b421e2 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_tensor([6556, 6556])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "locked, static macro lens remains perfectly still, capturing a captivating succulent plant rendered in the detailed style of Pixar animation. The subject is the immaculate Echeveria, whose plump, spoon-shaped leaves are arranged in a perfect geometric rosette. These leaves have a waxy, smooth texture and display a stunning color gradient, shifting from a soft, dusty lavender at the tips to a warm, blush pink near the core. A soft, magical glow emanates from within the plant, causing the leaves to softly glow. The succulent sits in a simple, unadorned terracotta pot with a matte finish, resting on a soft-focus, warm wooden surface. In this gentle, magical time-lapse, the plant's central cluster of new growth unfolds; the tiny, almost translucent leaves at the center of the rosette steadily and gracefully separate, expanding outward in a quiet, rhythmic, and continuous motion. The entire scene is bathed in a soft, warm, diffuse light, creating subtle highlights and gentle shadows that emphasize the plant's form and focus all attention on its quiet, steady expansion, creating an atmosphere of tranquility, enchantment, and utter peace.", + "seed": 20226, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20226_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.08416719511151313, + "vlm_logprob": -0.08416719511151313, + "vlm_yes_no_margin": 3.225, + "vlm_token_credit": -0.08416719511151313, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31657618284225464 + ], + "opposite_logprob_tokens": [ + -1.3165762424468994 + ], + "logprob_reward": -0.31657618284225464, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31657618284225464, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.406960666179657, + 0.10935790836811066, + 0.01812734268605709, + 0.030433690175414085, + 0.016812719404697418, + 0.025278128683567047, + 0.15582741796970367, + 0.16271714866161346, + 0.04582912474870682, + 0.028655828908085823 + ], + "credit_map": [ + 0.02657897025346756, + 0.01576877199113369, + 0.10294715315103531, + 0.014939927496016026, + 0.011389202438294888, + 0.1769210696220398, + 0.014687188901007175, + 0.04372836649417877, + 0.008333035744726658, + 0.0025583032984286547, + 0.026738271117210388, + 0.00397811783477664, + 0.002934748772531748, + 0.017733385786414146, + 0.004254835657775402, + 0.04282721132040024, + 0.0025507162790745497, + 0.002004768466576934, + 0.004341076593846083, + 0.0016050334088504314, + 0.0012388876639306545, + 0.0019011963158845901, + 0.0014373352751135826, + 0.003048327751457691, + 0.003626701422035694, + 0.003501882776618004, + 0.0037585124373435974, + 0.0030477289110422134, + 0.004952788818627596, + 0.0026717528235167265, + 0.003479591105133295, + 0.0053947302512824535, + 0.0024550913367420435, + 0.0017923953710123897, + 0.0026230239309370518, + 0.0015951070236042142, + 0.0019799747969955206, + 0.0014451766619458795, + 0.0014999627601355314, + 0.0034219869412481785, + 0.0031652448233217, + 0.002521354937925935, + 0.005520254839211702, + 0.0018294445471838117, + 0.0024060318246483803, + 0.0018764609703794122, + 0.001955593004822731, + 0.006003743037581444, + 0.012623470276594162, + 0.0035845099482685328, + 0.008102893829345703, + 0.0028077487368136644, + 0.005492948926985264, + 0.06261199712753296, + 0.004050049465149641, + 0.05655379965901375, + 0.04424285888671875, + 0.009284461848437786, + 0.017216091975569725, + 0.007081324700266123, + 0.008225912228226662, + 0.014812681823968887, + 0.00848239753395319, + 0.05337142571806908, + 0.011286917142570019, + 0.0038535515777766705, + 0.004737528041005135, + 0.0030018792022019625, + 0.0028354772366583347, + 0.0048365783877670765, + 0.0042005605064332485, + 0.011076631024479866, + 0.007204337511211634, + 0.0028483772184699774, + 0.005782918073236942, + 0.0022392035461962223, + 0.002074403688311577, + 0.0028628376312553883, + 0.0020233590621501207, + 0.0036203907802700996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.31657618284225464 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011088063940405846 + ], + "opposite_logprob_tokens": [ + -4.511087894439697 + ], + "logprob_reward": -0.011088063940405846, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011088063940405846, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10025931894779205, + 0.12411845475435257, + 0.013014907017350197, + 0.029706276953220367, + 0.018192974850535393, + 0.01512717641890049, + 0.0903482735157013, + 0.1923806220293045, + 0.25174838304519653, + 0.1651037037372589 + ], + "credit_map": [ + 0.01322229579091072, + 0.012433051131665707, + 0.012170305475592613, + 0.008820520713925362, + 0.0180667657405138, + 0.010221928358078003, + 0.012575055472552776, + 0.012749381363391876, + 0.0061145140789449215, + 0.004857179708778858, + 0.007897202856838703, + 0.0046382322907447815, + 0.006696238182485104, + 0.050472937524318695, + 0.005392995197325945, + 0.03804915025830269, + 0.0016783007886260748, + 0.0015421062707901, + 0.0022407083306461573, + 0.001136740786023438, + 0.0020659894216805696, + 0.0014127937611192465, + 0.0013321901205927134, + 0.0016060760244727135, + 0.003422823268920183, + 0.0037004270125180483, + 0.0029497442301362753, + 0.0028262895066291094, + 0.005557315424084663, + 0.0033115430269390345, + 0.0032267228234559298, + 0.0047114077024161816, + 0.0028587377164512873, + 0.0021825178992003202, + 0.0017147413454949856, + 0.0011787546100094914, + 0.0020473275799304247, + 0.0031584626995027065, + 0.00158155825920403, + 0.0034708725288510323, + 0.0016952994046732783, + 0.0020115470979362726, + 0.0020041107200086117, + 0.0009011724614538252, + 0.0019776145927608013, + 0.001859076670370996, + 0.0011831661686301231, + 0.0034951879642903805, + 0.008068906143307686, + 0.0028460395988076925, + 0.004169986583292484, + 0.002219747519120574, + 0.00415584072470665, + 0.035157445818185806, + 0.0028962052892893553, + 0.030834097415208817, + 0.04348199814558029, + 0.012407898902893066, + 0.01265111193060875, + 0.008119573816657066, + 0.014951544813811779, + 0.022940950468182564, + 0.010068506933748722, + 0.06775903701782227, + 0.03588845208287239, + 0.035486262291669846, + 0.0242498479783535, + 0.02285800501704216, + 0.03860752657055855, + 0.022540342062711716, + 0.023053348064422607, + 0.04906456172466278, + 0.03031732141971588, + 0.02398320473730564, + 0.01713045872747898, + 0.01609266735613346, + 0.028286851942539215, + 0.016041520982980728, + 0.01617739349603653, + 0.01707426644861698 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 9 + ], + "selected_reward": -0.011088063940405846 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048652734607458115 + ], + "opposite_logprob_tokens": [ + -3.0486526489257812 + ], + "logprob_reward": -0.048652734607458115, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048652734607458115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2512494921684265, + 0.10909076035022736, + 0.06049346178770065, + 0.044252123683691025, + 0.01423887349665165, + 0.008909545838832855, + 0.06908036023378372, + 0.2147793471813202, + 0.08998642861843109, + 0.13791951537132263 + ], + "credit_map": [ + 0.01973906345665455, + 0.018196722492575645, + 0.04836219549179077, + 0.01381908729672432, + 0.02729666791856289, + 0.08247065544128418, + 0.01751844771206379, + 0.023846691474318504, + 0.006526918150484562, + 0.006465797312557697, + 0.01929202303290367, + 0.0050741033628582954, + 0.007178943138569593, + 0.020601026713848114, + 0.006249305326491594, + 0.03770264983177185, + 0.007457587402313948, + 0.007106718607246876, + 0.011342894285917282, + 0.006647408939898014, + 0.007205878384411335, + 0.004690953064709902, + 0.006402320694178343, + 0.009639706462621689, + 0.004858354572206736, + 0.005436211824417114, + 0.004447645042091608, + 0.003990611992776394, + 0.008394904434680939, + 0.00434317858889699, + 0.005188591778278351, + 0.0075926268473267555, + 0.0018386883893981576, + 0.0016994543839246035, + 0.002438439754769206, + 0.001015885267406702, + 0.002247976139187813, + 0.0011043869890272617, + 0.001502182800322771, + 0.002391860354691744, + 0.0012357999803498387, + 0.0013467601966112852, + 0.0012364104622974992, + 0.0007244230364449322, + 0.0009911928791552782, + 0.0009701556991785765, + 0.0008393882890231907, + 0.001565416227094829, + 0.005183252040296793, + 0.004140122793614864, + 0.0041952962055802345, + 0.002612239448353648, + 0.004453593865036964, + 0.022899409756064415, + 0.0032301293686032295, + 0.022366313263773918, + 0.05262261629104614, + 0.015370693989098072, + 0.015710171312093735, + 0.00902786199003458, + 0.01770288124680519, + 0.018670393154025078, + 0.014368259347975254, + 0.0713064894080162, + 0.02649296075105667, + 0.006451959256082773, + 0.010682732798159122, + 0.0052254172042012215, + 0.0076854536309838295, + 0.00562528008595109, + 0.007073895540088415, + 0.020748740062117577, + 0.03403805196285248, + 0.010976098477840424, + 0.01534313801676035, + 0.01015437487512827, + 0.013479174114763737, + 0.030790045857429504, + 0.012980347499251366, + 0.010158296674489975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.048652734607458115 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08216895163059235, + 0.2547709047794342, + 0.02903321012854576, + 0.03674064949154854, + 0.026807619258761406, + 0.013616634532809258, + 0.11485499143600464, + 0.10891921818256378, + 0.27083778381347656, + 0.0622500479221344 + ], + "credit_map": [ + 0.012243208475410938, + 0.008698406629264355, + 0.013501375913619995, + 0.006203700788319111, + 0.009073772467672825, + 0.014396514743566513, + 0.008351048454642296, + 0.009700924158096313, + 0.021570932120084763, + 0.006442635785788298, + 0.0880487710237503, + 0.012302053160965443, + 0.007445187773555517, + 0.01701061986386776, + 0.011715914122760296, + 0.09023478627204895, + 0.004243653267621994, + 0.002607276663184166, + 0.006753415334969759, + 0.0026471675373613834, + 0.0017711870605126023, + 0.003297342685982585, + 0.0024973000399768353, + 0.005215865559875965, + 0.00709460349753499, + 0.002638346515595913, + 0.006078297272324562, + 0.002469775965437293, + 0.003890693187713623, + 0.00425512483343482, + 0.0029607827309519053, + 0.007353028282523155, + 0.003568042069673538, + 0.0035228480119258165, + 0.002949038054794073, + 0.002063607331365347, + 0.0036641822662204504, + 0.004086195025593042, + 0.0026394587475806475, + 0.00431424705311656, + 0.0014576633693650365, + 0.002413045149296522, + 0.002493784064427018, + 0.0010628916788846254, + 0.001876065623946488, + 0.0009914940455928445, + 0.0013990196166560054, + 0.001922669936902821, + 0.0066933538764715195, + 0.005002006888389587, + 0.006752053275704384, + 0.002732810564339161, + 0.00437528919428587, + 0.04162455350160599, + 0.003985044080764055, + 0.043689876794815063, + 0.015230880118906498, + 0.008220953866839409, + 0.017917107790708542, + 0.008601075038313866, + 0.006044723559170961, + 0.02165904827415943, + 0.007312194909900427, + 0.023933231830596924, + 0.09720700234174728, + 0.019940968602895737, + 0.020437641069293022, + 0.00931639689952135, + 0.01667230948805809, + 0.07217088341712952, + 0.013779195956885815, + 0.0213133804500103, + 0.017940392717719078, + 0.005331035703420639, + 0.008230311796069145, + 0.004322153050452471, + 0.00425701541826129, + 0.006744345184415579, + 0.004763305652886629, + 0.010661487467586994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 6 + ], + "selected_reward": -0.01818847469985485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633051946759224 + ], + "opposite_logprob_tokens": [ + -3.6513304710388184 + ], + "logprob_reward": -0.02633051946759224, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633051946759224, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10735506564378738, + 0.5983638167381287, + 0.15083478391170502, + 0.009946297854185104, + 0.01257010642439127, + 0.015844954177737236, + 0.008326350711286068, + 0.04876355081796646, + 0.035367805510759354, + 0.012627276591956615 + ], + "credit_map": [ + 0.01042034849524498, + 0.005010379012674093, + 0.017024418339133263, + 0.00455161277204752, + 0.00546940928325057, + 0.042621392756700516, + 0.006239202804863453, + 0.01601831056177616, + 0.051048751920461655, + 0.015187756158411503, + 0.19976666569709778, + 0.027399275451898575, + 0.0165703222155571, + 0.04145301505923271, + 0.02657761052250862, + 0.2203604131937027, + 0.02633080817759037, + 0.012956936843693256, + 0.042877111583948135, + 0.014586652629077435, + 0.005817025434225798, + 0.015664180740714073, + 0.010478104464709759, + 0.02212396450340748, + 0.0009345030412077904, + 0.0008196837152354419, + 0.00147461355663836, + 0.0008454202907159925, + 0.0009379382827319205, + 0.001123518100939691, + 0.0011412739986553788, + 0.00266934628598392, + 0.002038535661995411, + 0.001037748297676444, + 0.002256918465718627, + 0.001247498206794262, + 0.0011594770476222038, + 0.0013534878380596638, + 0.0010320725850760937, + 0.0024443678557872772, + 0.0026466164272278547, + 0.0016011926345527172, + 0.0032706221099942923, + 0.0012441176222637296, + 0.0015662098303437233, + 0.0011409979779273272, + 0.0010685303714126348, + 0.003306665224954486, + 0.0011991095962002873, + 0.000736745132599026, + 0.0007763545727357268, + 0.00048825182602740824, + 0.0005926630110479891, + 0.0020819881465286016, + 0.0007380972965620458, + 0.001713141449727118, + 0.012205224484205246, + 0.0020625123288482428, + 0.0030739405192434788, + 0.0014917503576725721, + 0.0024148845113813877, + 0.006453107576817274, + 0.001948483637534082, + 0.01911364495754242, + 0.006551662925630808, + 0.005300204735249281, + 0.0035808419343084097, + 0.003220906015485525, + 0.00502573000267148, + 0.006180368363857269, + 0.002976028248667717, + 0.002532065147534013, + 0.001911116298288107, + 0.0015385533915832639, + 0.0019813715480268, + 0.0013390622334554791, + 0.0017950451001524925, + 0.0016898296307772398, + 0.0013353193644434214, + 0.001036978792399168 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.02633051946759224 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, static macro lens remains perfectly still, capturing a captivating succulent plant rendered in the detailed style of Pixar animation. The subject is the immaculate Echeveria, whose plump, spoon-shaped leaves are arranged in a perfect geometric rosette. These leaves have a waxy, smooth texture and display a stunning color gradient, shifting from a soft, dusty lavender at the tips to a warm, blush pink near the core. A soft, magical glow emanates from within the plant, causing the leaves to softly glow. The succulent sits in a simple, unadorned terracotta pot with a matte finish, resting on a soft-focus, warm wooden surface. In this gentle, magical time-lapse, the plant's central cluster of new growth unfolds; the tiny, almost translucent leaves at the center of the rosette steadily and gracefully separate, expanding outward in a quiet, rhythmic, and continuous motion. The entire scene is bathed in a soft, warm, diffuse light, creating subtle highlights and gentle shadows that emphasize the plant's form and focus all attention on its quiet, steady expansion, creating an atmosphere of tranquility, enchantment, and utter peace.", + "seed": 20227, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20227_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.11188275273889303, + "vlm_logprob": -0.11188275273889303, + "vlm_yes_no_margin": 3.325, + "vlm_token_credit": -0.11188275273889303, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4791496694087982 + ], + "opposite_logprob_tokens": [ + -0.9791496992111206 + ], + "logprob_reward": -0.4791496694087982, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4791496694087982, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.056338172405958176, + 0.03742348402738571, + 0.048448093235492706, + 0.028620310127735138, + 0.3954823315143585, + 0.03123709000647068, + 0.05621279403567314, + 0.06472305953502655, + 0.20180362462997437, + 0.07971099764108658 + ], + "credit_map": [ + 0.007239686790853739, + 0.006223815958946943, + 0.005694961175322533, + 0.007088238839060068, + 0.007420687470585108, + 0.010328601114451885, + 0.007335124537348747, + 0.005007059313356876, + 0.007643037475645542, + 0.0035134695935994387, + 0.0038908524438738823, + 0.0024531688541173935, + 0.004584199283272028, + 0.007243475876748562, + 0.003396900836378336, + 0.004698376636952162, + 0.0085914870724082, + 0.004783052951097488, + 0.0048637124709784985, + 0.0061379908584058285, + 0.004949913360178471, + 0.004145182203501463, + 0.008352907374501228, + 0.006623846013098955, + 0.005637898109853268, + 0.00458467286080122, + 0.0024436230305582285, + 0.0018247481202706695, + 0.004280178342014551, + 0.0033932889346033335, + 0.002826655749231577, + 0.0036292460281401873, + 0.05435144156217575, + 0.08798743784427643, + 0.023484352976083755, + 0.029781023040413857, + 0.0958680510520935, + 0.04642125219106674, + 0.02707531861960888, + 0.03051348589360714, + 0.005347340367734432, + 0.002718272153288126, + 0.0028198785148561, + 0.002500428818166256, + 0.004687495529651642, + 0.005678696092218161, + 0.0029474934563040733, + 0.00453748507425189, + 0.008505742065608501, + 0.0055542909540236, + 0.005123464856296778, + 0.0062062544748187065, + 0.006632697302848101, + 0.009187591262161732, + 0.008244408294558525, + 0.006758346222341061, + 0.012459544464945793, + 0.008870095945894718, + 0.004260237794369459, + 0.00364863034337759, + 0.00989639200270176, + 0.015133648179471493, + 0.004319426137953997, + 0.006135082338005304, + 0.041629765182733536, + 0.013744890689849854, + 0.012049240991473198, + 0.009740839712321758, + 0.016417553648352623, + 0.07545344531536102, + 0.01169559545814991, + 0.02107229456305504, + 0.008978870697319508, + 0.016668591648340225, + 0.005764341913163662, + 0.0036413087509572506, + 0.006780553609132767, + 0.029673628509044647, + 0.0031844526529312134, + 0.005019246134907007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.4791496694087982 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009794983081519604 + ], + "opposite_logprob_tokens": [ + -4.634795188903809 + ], + "logprob_reward": -0.009794983081519604, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009794983081519604, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30430954694747925, + 0.07185196876525879, + 0.043454501777887344, + 0.026497114449739456, + 0.1426340490579605, + 0.030784016475081444, + 0.07409626990556717, + 0.12313202768564224, + 0.08917095512151718, + 0.09406959265470505 + ], + "credit_map": [ + 0.03492104262113571, + 0.042103443294763565, + 0.028203623369336128, + 0.03127307817339897, + 0.0533277653157711, + 0.048963844776153564, + 0.03850604221224785, + 0.027010703459382057, + 0.019512420520186424, + 0.006013918202370405, + 0.005793591495603323, + 0.0038926657289266586, + 0.005930312443524599, + 0.017618566751480103, + 0.0058509353548288345, + 0.0072395591996610165, + 0.008135378360748291, + 0.005130372010171413, + 0.0037733623757958412, + 0.00474344240501523, + 0.005519096273928881, + 0.00399325555190444, + 0.007365976460278034, + 0.004793618340045214, + 0.005065061617642641, + 0.0037281475961208344, + 0.0025793728418648243, + 0.00217250338755548, + 0.0042952727526426315, + 0.0030696445610374212, + 0.0025058400351554155, + 0.0030812714248895645, + 0.01925055868923664, + 0.031914591789245605, + 0.008589987643063068, + 0.010681343264877796, + 0.0337727852165699, + 0.01780807413160801, + 0.009442448616027832, + 0.011174253188073635, + 0.0043057603761553764, + 0.00422203354537487, + 0.0030970019288361073, + 0.0031884643249213696, + 0.0046317907981574535, + 0.005183194763958454, + 0.0025488794781267643, + 0.0036068905610591173, + 0.013503682799637318, + 0.005973806604743004, + 0.005671343766152859, + 0.005276900716125965, + 0.008346982300281525, + 0.02037714794278145, + 0.007208984810858965, + 0.007737418171018362, + 0.020642541348934174, + 0.02205679751932621, + 0.007936871610581875, + 0.010674555785953999, + 0.02209370583295822, + 0.019619641825556755, + 0.01044041384011507, + 0.009667497128248215, + 0.01731674000620842, + 0.013430234044790268, + 0.005937156267464161, + 0.006713843904435635, + 0.013447077944874763, + 0.01832330785691738, + 0.006871168967336416, + 0.007131421938538551, + 0.017823899164795876, + 0.010073940269649029, + 0.006688268855214119, + 0.0037761821877211332, + 0.005577495787292719, + 0.039150916039943695, + 0.004019645974040031, + 0.006959243677556515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.009794983081519604 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818976178765297 + ], + "opposite_logprob_tokens": [ + -4.018189907073975 + ], + "logprob_reward": -0.01818976178765297, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818976178765297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09293804317712784, + 0.08469973504543304, + 0.09944523870944977, + 0.03507783263921738, + 0.3315697908401489, + 0.0245391633361578, + 0.08333402872085571, + 0.0915035754442215, + 0.10295945405960083, + 0.05393313243985176 + ], + "credit_map": [ + 0.011308382265269756, + 0.010053986683487892, + 0.009257778525352478, + 0.010423037223517895, + 0.016418835148215294, + 0.015364943072199821, + 0.010670172981917858, + 0.009440906345844269, + 0.021352702751755714, + 0.006427429616451263, + 0.006973077077418566, + 0.00610640924423933, + 0.00962737761437893, + 0.02185462787747383, + 0.005057339556515217, + 0.007300767581909895, + 0.02266043797135353, + 0.008896078914403915, + 0.007907264865934849, + 0.008666984736919403, + 0.007891394197940826, + 0.020531263202428818, + 0.01244506984949112, + 0.010446748696267605, + 0.006971674505621195, + 0.0035637784749269485, + 0.003423250513151288, + 0.0027383689302951097, + 0.0041043804958462715, + 0.004783004522323608, + 0.0043081482872366905, + 0.0051852259784936905, + 0.043474871665239334, + 0.07889051735401154, + 0.01948772370815277, + 0.026045043021440506, + 0.08218873292207718, + 0.035186413675546646, + 0.022478731349110603, + 0.023817747831344604, + 0.00383144523948431, + 0.0033801135141402483, + 0.002038008766248822, + 0.0025177313946187496, + 0.004252539481967688, + 0.0030643523205071688, + 0.003071823623031378, + 0.002383148530498147, + 0.026507243514060974, + 0.004682844504714012, + 0.004401566926389933, + 0.0027637339662760496, + 0.006184250582009554, + 0.030744336545467377, + 0.002968220505863428, + 0.0050818356685340405, + 0.019647931680083275, + 0.015129143372178078, + 0.006028502248227596, + 0.00667734257876873, + 0.015611647628247738, + 0.014003523625433445, + 0.007320274133235216, + 0.007085210178047419, + 0.025202196091413498, + 0.00563619052991271, + 0.005896016024053097, + 0.0038366315420717, + 0.008873935788869858, + 0.04156871512532234, + 0.003800738835707307, + 0.00814503338187933, + 0.00562632828950882, + 0.010063613764941692, + 0.003948050085455179, + 0.003588983556255698, + 0.00424240855500102, + 0.019866228103637695, + 0.0023997933603823185, + 0.004197729751467705 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 2 + ], + "selected_reward": -0.01818976178765297 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014197885990142822 + ], + "opposite_logprob_tokens": [ + -4.264197826385498 + ], + "logprob_reward": -0.014197885990142822, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014197885990142822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12544450163841248, + 0.047286856919527054, + 0.036625832319259644, + 0.0634288564324379, + 0.2807817757129669, + 0.053926609456539154, + 0.06337863206863403, + 0.1858505755662918, + 0.11224769055843353, + 0.03102865442633629 + ], + "credit_map": [ + 0.016717275604605675, + 0.017222709953784943, + 0.010942017659544945, + 0.0106099434196949, + 0.021507510915398598, + 0.025773009285330772, + 0.010569821111857891, + 0.012102204374969006, + 0.007971502840518951, + 0.006654567085206509, + 0.0055215200409293175, + 0.004565471783280373, + 0.006762822158634663, + 0.00506485253572464, + 0.004727557301521301, + 0.0060185631737113, + 0.0058422028087079525, + 0.005421111825853586, + 0.0038082748651504517, + 0.003088687313720584, + 0.005136882420629263, + 0.004663789179176092, + 0.003667749697342515, + 0.004997135605663061, + 0.009933074936270714, + 0.009037904441356659, + 0.0071824234910309315, + 0.00650305300951004, + 0.010427135042846203, + 0.005650473292917013, + 0.006609099917113781, + 0.008085690438747406, + 0.03829322010278702, + 0.06554745137691498, + 0.016189251095056534, + 0.02107396349310875, + 0.06796810030937195, + 0.0323014110326767, + 0.018678681924939156, + 0.0207296684384346, + 0.009188001044094563, + 0.005310973152518272, + 0.004724027588963509, + 0.004747787490487099, + 0.009816120378673077, + 0.00840366818010807, + 0.004762756172567606, + 0.006973274052143097, + 0.01242954470217228, + 0.006405186373740435, + 0.004546836484223604, + 0.004224596545100212, + 0.009173741564154625, + 0.019210210070014, + 0.003180913859978318, + 0.004207600373774767, + 0.027329403907060623, + 0.02844928577542305, + 0.011048348620533943, + 0.01554839126765728, + 0.035583361983299255, + 0.03626134991645813, + 0.018883807584643364, + 0.012746631167829037, + 0.026629112660884857, + 0.012014769949018955, + 0.0072403522208333015, + 0.007184484042227268, + 0.013143123127520084, + 0.028353897854685783, + 0.008184442296624184, + 0.009497500956058502, + 0.004781970754265785, + 0.005221158731728792, + 0.002449883846566081, + 0.002755718305706978, + 0.005424989387392998, + 0.003438834100961685, + 0.0023868780117481947, + 0.004569219425320625 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.014197885990142822 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03808146342635155 + ], + "opposite_logprob_tokens": [ + -3.288081407546997 + ], + "logprob_reward": -0.03808146342635155, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03808146342635155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09774715453386307, + 0.1298510879278183, + 0.018658867105841637, + 0.04221819341182709, + 0.486426442861557, + 0.07285413146018982, + 0.013036088086664677, + 0.030917325988411903, + 0.08150028437376022, + 0.02679038979113102 + ], + "credit_map": [ + 0.012991960160434246, + 0.01167354080826044, + 0.012714972719550133, + 0.0084204887971282, + 0.013150716200470924, + 0.011444698087871075, + 0.013447780162096024, + 0.013902994804084301, + 0.028356527909636497, + 0.00904408935457468, + 0.013877535238862038, + 0.008815312758088112, + 0.009391297586262226, + 0.030434781685471535, + 0.011546439491212368, + 0.01838509552180767, + 0.0038193247746676207, + 0.0022001939360052347, + 0.0018430054187774658, + 0.0010546286357566714, + 0.002547137439250946, + 0.003385318676009774, + 0.0015808093594387174, + 0.002228449797257781, + 0.009430226869881153, + 0.005672422703355551, + 0.0036623566411435604, + 0.0025391934905201197, + 0.006953333038836718, + 0.006349931005388498, + 0.0032193916849792004, + 0.0043913377448916435, + 0.07146661728620529, + 0.10675850510597229, + 0.03037853166460991, + 0.03658774122595787, + 0.11111032217741013, + 0.05620064586400986, + 0.03418031334877014, + 0.03974376246333122, + 0.012494221329689026, + 0.007684672251343727, + 0.007333910558372736, + 0.006592477671802044, + 0.009875049814581871, + 0.011764793656766415, + 0.00695381686091423, + 0.01015518605709076, + 0.002316115191206336, + 0.0012826164020225406, + 0.001496523735113442, + 0.0009100194438360631, + 0.0019111743895336986, + 0.0019587248098105192, + 0.0012050917139276862, + 0.0019558228086680174, + 0.007030370645225048, + 0.004287772811949253, + 0.002067445544525981, + 0.0018898396519944072, + 0.0047268676571547985, + 0.005788380745798349, + 0.0022285941522568464, + 0.002898053964599967, + 0.017005980014801025, + 0.005814220290631056, + 0.005008100066334009, + 0.004119432531297207, + 0.00752644520252943, + 0.030400188639760017, + 0.004298986867070198, + 0.007326930295675993, + 0.0036850387696176767, + 0.003217628924176097, + 0.002706083469092846, + 0.0022396931890398264, + 0.0029186184983700514, + 0.006165381986647844, + 0.0023251015227288008, + 0.0035328420344740152 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.03808146342635155 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, static macro lens remains perfectly still, capturing a captivating succulent plant rendered in the detailed style of Pixar animation. The subject is the immaculate Echeveria, whose plump, spoon-shaped leaves are arranged in a perfect geometric rosette. These leaves have a waxy, smooth texture and display a stunning color gradient, shifting from a soft, dusty lavender at the tips to a warm, blush pink near the core. A soft, magical glow emanates from within the plant, causing the leaves to softly glow. The succulent sits in a simple, unadorned terracotta pot with a matte finish, resting on a soft-focus, warm wooden surface. In this gentle, magical time-lapse, the plant's central cluster of new growth unfolds; the tiny, almost translucent leaves at the center of the rosette steadily and gracefully separate, expanding outward in a quiet, rhythmic, and continuous motion. The entire scene is bathed in a soft, warm, diffuse light, creating subtle highlights and gentle shadows that emphasize the plant's form and focus all attention on its quiet, steady expansion, creating an atmosphere of tranquility, enchantment, and utter peace.", + "seed": 20228, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20228_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.094027947075665, + "vlm_logprob": -0.094027947075665, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.094027947075665, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3522252142429352 + ], + "opposite_logprob_tokens": [ + -1.2272251844406128 + ], + "logprob_reward": -0.3522252142429352, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3522252142429352, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3329196572303772, + 0.054776210337877274, + 0.04968088120222092, + 0.07488729059696198, + 0.1623753160238266, + 0.1303313672542572, + 0.07071470469236374, + 0.04681508243083954, + 0.03692343458533287, + 0.04057602211833 + ], + "credit_map": [ + 0.037612587213516235, + 0.04956213757395744, + 0.03999686986207962, + 0.03329058736562729, + 0.06466522067785263, + 0.04599849507212639, + 0.035096537321805954, + 0.026697244495153427, + 0.007112463470548391, + 0.006445346400141716, + 0.0075147622264921665, + 0.0069050840102136135, + 0.006813463754951954, + 0.005709004122763872, + 0.00632419902831316, + 0.007951885461807251, + 0.005982455797493458, + 0.005343269091099501, + 0.006653461139649153, + 0.006083567626774311, + 0.005970621947199106, + 0.00932227447628975, + 0.006341011729091406, + 0.003984218463301659, + 0.008651151321828365, + 0.012163790874183178, + 0.009593010880053043, + 0.009835039265453815, + 0.011231249198317528, + 0.00923995953053236, + 0.007411112077534199, + 0.0067619807086884975, + 0.01994164288043976, + 0.026282131671905518, + 0.017962388694286346, + 0.014751858077943325, + 0.02113039791584015, + 0.02732296846807003, + 0.019834283739328384, + 0.015149638056755066, + 0.019439902156591415, + 0.019849684089422226, + 0.015536022372543812, + 0.012422428466379642, + 0.011563542298972607, + 0.024045521393418312, + 0.014418933540582657, + 0.01305533479899168, + 0.00986018031835556, + 0.010596457868814468, + 0.006579898763448, + 0.010508453473448753, + 0.009947008453309536, + 0.007421885617077351, + 0.009840673767030239, + 0.0059601496905088425, + 0.006945749279111624, + 0.005670903250575066, + 0.006073442753404379, + 0.0051855286583304405, + 0.0057454886846244335, + 0.0049932473339140415, + 0.006832365412265062, + 0.005368358921259642, + 0.005173738580197096, + 0.006605273578315973, + 0.003767105983570218, + 0.004669120069593191, + 0.005164818372577429, + 0.004449179396033287, + 0.00425125053152442, + 0.002842947142198682, + 0.006138859782367945, + 0.007441465277224779, + 0.005828546360135078, + 0.00443109730258584, + 0.00538282934576273, + 0.004851898644119501, + 0.0031845758203417063, + 0.003316750517114997 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.3522252142429352 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085941456258297 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011085941456258297, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085941456258297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33440613746643066, + 0.1074238121509552, + 0.04794823378324509, + 0.06464672833681107, + 0.09716124087572098, + 0.11368101835250854, + 0.11065683513879776, + 0.06709802150726318, + 0.031202085316181183, + 0.025775892660021782 + ], + "credit_map": [ + 0.04005910083651543, + 0.052281271666288376, + 0.036429453641176224, + 0.034941237419843674, + 0.06102391332387924, + 0.04503883793950081, + 0.03623934090137482, + 0.028392987325787544, + 0.01270301267504692, + 0.015511086210608482, + 0.010564046911895275, + 0.012391285970807076, + 0.017322348430752754, + 0.019195081666111946, + 0.011114787310361862, + 0.008622163906693459, + 0.005657759495079517, + 0.006158450618386269, + 0.005600868258625269, + 0.005235936492681503, + 0.0065176417119801044, + 0.008659577928483486, + 0.005659712478518486, + 0.0044582923874258995, + 0.008155923336744308, + 0.009609214961528778, + 0.005681877955794334, + 0.006562849041074514, + 0.010217317380011082, + 0.01404078770428896, + 0.006453632842749357, + 0.003925124183297157, + 0.011994695290923119, + 0.014380439184606075, + 0.011806686408817768, + 0.008776865899562836, + 0.010846240445971489, + 0.01731221377849579, + 0.012576139532029629, + 0.009467963129281998, + 0.019780203700065613, + 0.014170044101774693, + 0.01375398226082325, + 0.010046523995697498, + 0.00972241535782814, + 0.020109951496124268, + 0.012467718683183193, + 0.013630173169076443, + 0.011886779218912125, + 0.018689995631575584, + 0.009219884872436523, + 0.017788896337151527, + 0.01723126508295536, + 0.015408332459628582, + 0.014716637320816517, + 0.005715043283998966, + 0.01111416332423687, + 0.00852589588612318, + 0.008450208231806755, + 0.005621034651994705, + 0.0074362060986459255, + 0.009899941273033619, + 0.00872852560132742, + 0.007322048768401146, + 0.004136714152991772, + 0.0038753498811274767, + 0.004480574745684862, + 0.00392740685492754, + 0.0035350846592336893, + 0.006640017032623291, + 0.0027708900161087513, + 0.0018360464600846171, + 0.00425617815926671, + 0.0036608793307095766, + 0.0029864893294870853, + 0.001915339264087379, + 0.0026400256901979446, + 0.0054671927355229855, + 0.0020670369267463684, + 0.002782752737402916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 6 + ], + "selected_reward": -0.011085941456258297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187889829277992 + ], + "opposite_logprob_tokens": [ + -4.018187999725342 + ], + "logprob_reward": -0.018187889829277992, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187889829277992, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32618415355682373, + 0.0937100499868393, + 0.07286135107278824, + 0.0667508989572525, + 0.11840907484292984, + 0.07453478872776031, + 0.11359023302793503, + 0.05395589396357536, + 0.04179864004254341, + 0.03820496425032616 + ], + "credit_map": [ + 0.03979984670877457, + 0.05377764254808426, + 0.037483301013708115, + 0.03336620330810547, + 0.05419353395700455, + 0.038080278784036636, + 0.03828011825680733, + 0.031203201040625572, + 0.008642284199595451, + 0.018042797222733498, + 0.008411843329668045, + 0.010088223032653332, + 0.016999386250972748, + 0.0107721583917737, + 0.010948577895760536, + 0.009804768487811089, + 0.006766493897885084, + 0.013054407201707363, + 0.007159805856645107, + 0.008014008402824402, + 0.013236352242529392, + 0.010152144357562065, + 0.008587722666561604, + 0.005890408996492624, + 0.007028271444141865, + 0.011631850153207779, + 0.0052573708817362785, + 0.008715699426829815, + 0.010448437184095383, + 0.011596652679145336, + 0.008077120408415794, + 0.003995494917035103, + 0.012011474929749966, + 0.02343917451798916, + 0.011676061898469925, + 0.012277180328965187, + 0.019760161638259888, + 0.014946799725294113, + 0.014325396157801151, + 0.009972823783755302, + 0.01293313317000866, + 0.008778518997132778, + 0.008665761910378933, + 0.006830320227891207, + 0.007052261382341385, + 0.012676446698606014, + 0.008190488442778587, + 0.009407851845026016, + 0.015874255448579788, + 0.016579611226916313, + 0.010498079471290112, + 0.016068866476416588, + 0.015739329159259796, + 0.012967200949788094, + 0.01694861799478531, + 0.008914271369576454, + 0.006674378179013729, + 0.007795065175741911, + 0.007814854383468628, + 0.006380573380738497, + 0.005967271979898214, + 0.006462814752012491, + 0.007882344536483288, + 0.00497859064489603, + 0.00744982436299324, + 0.005442983005195856, + 0.005875715054571629, + 0.0045305113308131695, + 0.004798123612999916, + 0.005466291680932045, + 0.0037142476066946983, + 0.004520939197391272, + 0.004121927544474602, + 0.007791589479893446, + 0.0035704332403838634, + 0.004736121743917465, + 0.006086698733270168, + 0.004285357892513275, + 0.004322948399931192, + 0.0032898839563131332 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.018187889829277992 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05495745688676834 + ], + "opposite_logprob_tokens": [ + -2.929957389831543 + ], + "logprob_reward": -0.05495745688676834, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05495745688676834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3051440715789795, + 0.13983383774757385, + 0.06360303610563278, + 0.04870801791548729, + 0.15593107044696808, + 0.15587346255779266, + 0.0627521276473999, + 0.042416371405124664, + 0.01483362540602684, + 0.010904405266046524 + ], + "credit_map": [ + 0.039177246391773224, + 0.05109411105513573, + 0.03209390491247177, + 0.0288451686501503, + 0.05442719906568527, + 0.0391109474003315, + 0.02946111001074314, + 0.03093435801565647, + 0.017467275261878967, + 0.021693134680390358, + 0.014140312559902668, + 0.012694121338427067, + 0.023204628378152847, + 0.022364402189850807, + 0.013845095410943031, + 0.01442487258464098, + 0.006381453014910221, + 0.00994946714490652, + 0.0055308775044977665, + 0.0066963848657906055, + 0.011553245596587658, + 0.011708107776939869, + 0.00685966806486249, + 0.0049238246865570545, + 0.0074616107158362865, + 0.007009222637861967, + 0.0055608018301427364, + 0.005601751152426004, + 0.007450548000633717, + 0.005485325586050749, + 0.0059663341380655766, + 0.004172425251454115, + 0.02013363316655159, + 0.02302263118326664, + 0.018187925219535828, + 0.01365869864821434, + 0.017110180109739304, + 0.029023168608546257, + 0.018722079694271088, + 0.01607275754213333, + 0.02458796091377735, + 0.021491752937436104, + 0.01905573345720768, + 0.013990898616611958, + 0.013598455116152763, + 0.02927330508828163, + 0.017257988452911377, + 0.01661737449467182, + 0.0064057777635753155, + 0.012838846072554588, + 0.004331097472459078, + 0.009484155103564262, + 0.011414781212806702, + 0.008488360792398453, + 0.007125319447368383, + 0.002663790713995695, + 0.006502756383270025, + 0.005759084597229958, + 0.005644932389259338, + 0.004047983326017857, + 0.004571195226162672, + 0.005022981204092503, + 0.0063117751851677895, + 0.004555665887892246, + 0.001682668924331665, + 0.0027208670508116484, + 0.002123386599123478, + 0.001702809240669012, + 0.001960675697773695, + 0.0019518674816936255, + 0.0013962467201054096, + 0.0012951039243489504, + 0.0017837488558143377, + 0.0015531290555372834, + 0.0016603308031335473, + 0.0009212730219587684, + 0.0012278036447241902, + 0.0012170997215434909, + 0.0012927197385579348, + 0.0012482997262850404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.05495745688676834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033683232963085175 + ], + "opposite_logprob_tokens": [ + -3.4086833000183105 + ], + "logprob_reward": -0.033683232963085175, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033683232963085175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2218817174434662, + 0.13460421562194824, + 0.05936668813228607, + 0.07105060666799545, + 0.13809342682361603, + 0.15438443422317505, + 0.10131009668111801, + 0.057162001729011536, + 0.03262268751859665, + 0.029524073004722595 + ], + "credit_map": [ + 0.042419277131557465, + 0.02572747878730297, + 0.0284124705940485, + 0.022137733176350594, + 0.024995122104883194, + 0.02362281084060669, + 0.02454691007733345, + 0.030019911006093025, + 0.02078665979206562, + 0.015970829874277115, + 0.014806006103754044, + 0.011495145969092846, + 0.018058758229017258, + 0.02309899404644966, + 0.014386908151209354, + 0.01600092463195324, + 0.008497225120663643, + 0.006918622646480799, + 0.010084209032356739, + 0.0066269515082240105, + 0.0049001495353877544, + 0.007128638681024313, + 0.009356415830552578, + 0.005854479502886534, + 0.012144817039370537, + 0.005468374118208885, + 0.009788774885237217, + 0.006186883896589279, + 0.006176971830427647, + 0.015499287284910679, + 0.00937851145863533, + 0.006406988482922316, + 0.02530054561793804, + 0.012102785520255566, + 0.018358001485466957, + 0.009472041390836239, + 0.010026381351053715, + 0.034491147845983505, + 0.013693277724087238, + 0.014649257995188236, + 0.025404702872037888, + 0.018866335973143578, + 0.019966723397374153, + 0.014328578487038612, + 0.014073987491428852, + 0.030191591009497643, + 0.01606653816998005, + 0.015485965646803379, + 0.01174137182533741, + 0.014725170098245144, + 0.009795856662094593, + 0.01658560335636139, + 0.013302801176905632, + 0.011583144776523113, + 0.016339784488081932, + 0.007236356381326914, + 0.00932398997247219, + 0.007399206515401602, + 0.008062520064413548, + 0.006622657645493746, + 0.007024984806776047, + 0.0064529250375926495, + 0.0072627076879143715, + 0.005013004876673222, + 0.00543135404586792, + 0.005103561095893383, + 0.0035277530550956726, + 0.0035752966068685055, + 0.004259594716131687, + 0.003670875448733568, + 0.0034380487632006407, + 0.0036162035539746284, + 0.004780584946274757, + 0.004592251032590866, + 0.003937822300940752, + 0.0030773074831813574, + 0.0032320579048246145, + 0.00355676026083529, + 0.003195189405232668, + 0.0031521001365035772 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.033683232963085175 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "locked, static macro lens remains perfectly still, capturing a captivating succulent plant rendered in the detailed style of Pixar animation. The subject is the immaculate Echeveria, whose plump, spoon-shaped leaves are arranged in a perfect geometric rosette. These leaves have a waxy, smooth texture and display a stunning color gradient, shifting from a soft, dusty lavender at the tips to a warm, blush pink near the core. A soft, magical glow emanates from within the plant, causing the leaves to softly glow. The succulent sits in a simple, unadorned terracotta pot with a matte finish, resting on a soft-focus, warm wooden surface. In this gentle, magical time-lapse, the plant's central cluster of new growth unfolds; the tiny, almost translucent leaves at the center of the rosette steadily and gracefully separate, expanding outward in a quiet, rhythmic, and continuous motion. The entire scene is bathed in a soft, warm, diffuse light, creating subtle highlights and gentle shadows that emphasize the plant's form and focus all attention on its quiet, steady expansion, creating an atmosphere of tranquility, enchantment, and utter peace.", + "seed": 20229, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_13/index_6556_seed_20229_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.09405011404305696, + "vlm_logprob": -0.09405011404305696, + "vlm_yes_no_margin": 3.275, + "vlm_token_credit": -0.09405011404305696, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3504880964756012 + ], + "opposite_logprob_tokens": [ + -1.2254880666732788 + ], + "logprob_reward": -0.3504880964756012, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3504880964756012, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2843819856643677, + 0.09209777414798737, + 0.08265388756990433, + 0.05714520812034607, + 0.08626329153776169, + 0.04074106737971306, + 0.08799570053815842, + 0.06733167171478271, + 0.10516300797462463, + 0.09622637927532196 + ], + "credit_map": [ + 0.029367608949542046, + 0.04830028489232063, + 0.029751727357506752, + 0.03075207583606243, + 0.05644799768924713, + 0.03475110977888107, + 0.020891012623906136, + 0.03412018343806267, + 0.008902677334845066, + 0.017313892021775246, + 0.00913616456091404, + 0.014211359433829784, + 0.014064550399780273, + 0.011333545669913292, + 0.007745745591819286, + 0.009389839135110378, + 0.009708567522466183, + 0.013882748782634735, + 0.009760988876223564, + 0.00921835470944643, + 0.013738184235990047, + 0.009662707336246967, + 0.006997368298470974, + 0.009684965945780277, + 0.005428312812000513, + 0.011099063791334629, + 0.005156909115612507, + 0.006197658367455006, + 0.008695635013282299, + 0.007806994020938873, + 0.006237519904971123, + 0.0065231164917349815, + 0.009413915686309338, + 0.019078660756349564, + 0.007154957391321659, + 0.008827940560877323, + 0.012228092178702354, + 0.01289291586726904, + 0.0062716007232666016, + 0.010395207442343235, + 0.004617484752088785, + 0.006949641276150942, + 0.004035389516502619, + 0.004242404364049435, + 0.006553333718329668, + 0.004732084926217794, + 0.004045503679662943, + 0.00556523073464632, + 0.007251201197504997, + 0.014168023131787777, + 0.011764409020543098, + 0.009485078975558281, + 0.014744221232831478, + 0.011168485507369041, + 0.009275591000914574, + 0.01013869047164917, + 0.006183736491948366, + 0.011265512555837631, + 0.009832520969212055, + 0.006036210339516401, + 0.00996799860149622, + 0.009204236790537834, + 0.005406603682786226, + 0.009434849955141544, + 0.011907649226486683, + 0.012007164768874645, + 0.02009592391550541, + 0.008655235171318054, + 0.010911084711551666, + 0.01329466886818409, + 0.01174206007272005, + 0.01654922030866146, + 0.010007006116211414, + 0.014574696309864521, + 0.014496161602437496, + 0.008623432368040085, + 0.013852902688086033, + 0.012467727065086365, + 0.00838811881840229, + 0.013816334307193756 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.3504880964756012 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008648204617202282 + ], + "opposite_logprob_tokens": [ + -4.75864839553833 + ], + "logprob_reward": -0.008648204617202282, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008648204617202282, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3663806915283203, + 0.18778938055038452, + 0.09191757440567017, + 0.04474722221493721, + 0.06234726309776306, + 0.045548807829618454, + 0.048537302762269974, + 0.044312987476587296, + 0.04708736762404442, + 0.061331454664468765 + ], + "credit_map": [ + 0.03228192403912544, + 0.04358651489019394, + 0.04521709308028221, + 0.06854080408811569, + 0.046076007187366486, + 0.04902154952287674, + 0.03102993220090866, + 0.05062685161828995, + 0.01698244921863079, + 0.031150760129094124, + 0.022632144391536713, + 0.02363213151693344, + 0.026797804981470108, + 0.02383747138082981, + 0.016838757321238518, + 0.025917835533618927, + 0.008287368342280388, + 0.016048219054937363, + 0.008489362895488739, + 0.01136113703250885, + 0.016000371426343918, + 0.010470898821949959, + 0.010981358587741852, + 0.0102788545191288, + 0.004598022438585758, + 0.0072989496402442455, + 0.004144364967942238, + 0.004325143992900848, + 0.007037927396595478, + 0.006449134089052677, + 0.00377218471840024, + 0.0071214912459254265, + 0.0069090998731553555, + 0.011181537993252277, + 0.005730388220399618, + 0.006252898834645748, + 0.008844680152833462, + 0.00941993948072195, + 0.006158503238111734, + 0.007850212045013905, + 0.004807134158909321, + 0.006019670516252518, + 0.004355692304670811, + 0.004382594954222441, + 0.007502814754843712, + 0.007159714587032795, + 0.0047883642837405205, + 0.006532819010317326, + 0.004591521807014942, + 0.006250876933336258, + 0.006025027949362993, + 0.005192795768380165, + 0.009287826716899872, + 0.006597269792109728, + 0.0045583355240523815, + 0.006033645942807198, + 0.006035507190972567, + 0.005792433861643076, + 0.005641089286655188, + 0.003410079050809145, + 0.005168910138309002, + 0.007188473828136921, + 0.0034544169902801514, + 0.0076220775954425335, + 0.004655423108488321, + 0.00514378072693944, + 0.007972437888383865, + 0.003688820404931903, + 0.005461013410240412, + 0.006155975162982941, + 0.005836446303874254, + 0.008173462934792042, + 0.005310766398906708, + 0.010770861990749836, + 0.007997267879545689, + 0.006606159266084433, + 0.008923475630581379, + 0.00730767659842968, + 0.005327550694346428, + 0.009087691083550453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.008648204617202282 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069188714027405 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069188714027405, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069188714027405, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16664603352546692, + 0.09051220864057541, + 0.10858216881752014, + 0.06512752920389175, + 0.10753572732210159, + 0.0713130310177803, + 0.07436609268188477, + 0.12067709118127823, + 0.04316047951579094, + 0.15207964181900024 + ], + "credit_map": [ + 0.016826901584863663, + 0.024953732267022133, + 0.018728725612163544, + 0.018176300451159477, + 0.030093129724264145, + 0.020872950553894043, + 0.013118745759129524, + 0.02387554757297039, + 0.01118568703532219, + 0.014976209960877895, + 0.009236661717295647, + 0.012202978134155273, + 0.013842236250638962, + 0.010315562598407269, + 0.008869686163961887, + 0.009883183054625988, + 0.0090173976495862, + 0.02547004632651806, + 0.009892702102661133, + 0.015460534952580929, + 0.014688175171613693, + 0.014306401833891869, + 0.009813456796109676, + 0.009933453984558582, + 0.006531789433211088, + 0.010628337040543556, + 0.005241065751761198, + 0.00934296753257513, + 0.00948227010667324, + 0.007955539040267467, + 0.008285170421004295, + 0.007660393137484789, + 0.008549979887902737, + 0.024814581498503685, + 0.009569612331688404, + 0.013284916989505291, + 0.0159378070384264, + 0.014414743520319462, + 0.010955390520393848, + 0.010008697398006916, + 0.00797475315630436, + 0.008550917729735374, + 0.010532896034419537, + 0.008913951925933361, + 0.008548078127205372, + 0.00829388853162527, + 0.008865010924637318, + 0.009633536450564861, + 0.00816951785236597, + 0.010376989841461182, + 0.00790455099195242, + 0.009972461499273777, + 0.010083673521876335, + 0.010996549390256405, + 0.007238606922328472, + 0.009623745456337929, + 0.010008618235588074, + 0.02329520881175995, + 0.012593662366271019, + 0.014374819584190845, + 0.021785296499729156, + 0.015123036690056324, + 0.010029243305325508, + 0.013467209413647652, + 0.0049743494018912315, + 0.006325834896415472, + 0.004970865789800882, + 0.005300923716276884, + 0.006437113042920828, + 0.005003053694963455, + 0.004854210186749697, + 0.005294127855449915, + 0.014269310049712658, + 0.024531306698918343, + 0.02242363430559635, + 0.016283171251416206, + 0.023690246045589447, + 0.01789095066487789, + 0.01372147724032402, + 0.019269535318017006 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.016069188714027405 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07898680120706558 + ], + "opposite_logprob_tokens": [ + -2.578986883163452 + ], + "logprob_reward": -0.07898680120706558, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07898680120706558, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19787681102752686, + 0.1178443655371666, + 0.07540445029735565, + 0.09050924330949783, + 0.07711509615182877, + 0.04957326501607895, + 0.06855610013008118, + 0.12811732292175293, + 0.12380413711071014, + 0.0711992084980011 + ], + "credit_map": [ + 0.022578779608011246, + 0.03875147923827171, + 0.0166179072111845, + 0.023394787684082985, + 0.03582063689827919, + 0.023326072841882706, + 0.015351447276771069, + 0.022035691887140274, + 0.014369356445968151, + 0.020707406103610992, + 0.012894319370388985, + 0.013445863500237465, + 0.01617240160703659, + 0.013477739877998829, + 0.0120225315913558, + 0.014754751697182655, + 0.007616134360432625, + 0.013153161853551865, + 0.008531411178410053, + 0.006236549932509661, + 0.013591377064585686, + 0.009441975504159927, + 0.0074518523178994656, + 0.009381990879774094, + 0.008131301030516624, + 0.01409145724028349, + 0.006955740042030811, + 0.011857965029776096, + 0.017091980203986168, + 0.011662496253848076, + 0.010326849296689034, + 0.01039145328104496, + 0.007617991417646408, + 0.01393209770321846, + 0.006851365324109793, + 0.008403060957789421, + 0.015056513249874115, + 0.010909190401434898, + 0.006281166337430477, + 0.008063703775405884, + 0.0052542127668857574, + 0.00806137453764677, + 0.003846814390271902, + 0.007666525896638632, + 0.00681147538125515, + 0.006491805426776409, + 0.005500482860952616, + 0.005940571893006563, + 0.006720050238072872, + 0.010130753740668297, + 0.005695299245417118, + 0.007792110554873943, + 0.013358728028833866, + 0.009774968028068542, + 0.005930893123149872, + 0.009153299033641815, + 0.016777513548731804, + 0.014658552594482899, + 0.01599549874663353, + 0.011190347373485565, + 0.015944505110383034, + 0.019246896728873253, + 0.01009304542094469, + 0.024210968986153603, + 0.013953393325209618, + 0.017209546640515327, + 0.01830308511853218, + 0.013589313253760338, + 0.015789391472935677, + 0.015925388783216476, + 0.011629517190158367, + 0.017404496669769287, + 0.006978795863687992, + 0.009967022575438023, + 0.012537268921732903, + 0.006957572884857655, + 0.008303468115627766, + 0.00944663118571043, + 0.0056935083121061325, + 0.011314944364130497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.07898680120706558 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera completely static and locked in a macro lens view without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject an Echeveria succulent with plump, spoon-shaped leaves arranged in a geometric rosette?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the leaves display a color gradient shifting from soft dusty lavender at the tips to warm blush pink near the core?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the central cluster of new growth unfold with tiny translucent leaves separating and expanding outward in a time-lapse motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the succulent sitting in a simple unadorned terracotta pot with a matte finish on a soft-focus warm wooden surface?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01605827920138836 + ], + "opposite_logprob_tokens": [ + -4.141058444976807 + ], + "logprob_reward": -0.01605827920138836, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01605827920138836, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27131012082099915, + 0.22615645825862885, + 0.0641564428806305, + 0.04343575984239578, + 0.03386865183711052, + 0.0567932203412056, + 0.04061036556959152, + 0.03955640643835068, + 0.098789282143116, + 0.12532325088977814 + ], + "credit_map": [ + 0.03441267088055611, + 0.031445495784282684, + 0.03084886260330677, + 0.028687600046396255, + 0.03200923651456833, + 0.03179486468434334, + 0.02435207925736904, + 0.05775928869843483, + 0.03349796682596207, + 0.0315096341073513, + 0.025567419826984406, + 0.03137388825416565, + 0.02063942328095436, + 0.02419101633131504, + 0.02074591815471649, + 0.03863121196627617, + 0.00909830629825592, + 0.00803952943533659, + 0.006404539104551077, + 0.006501724012196064, + 0.00757198641076684, + 0.00866002682596445, + 0.005742025561630726, + 0.012138302437961102, + 0.0055442810989916325, + 0.007697761990129948, + 0.003327458631247282, + 0.00457840645685792, + 0.00672844797372818, + 0.005075715947896242, + 0.0028394933324307203, + 0.007644191849976778, + 0.0037322703283280134, + 0.003877628594636917, + 0.0028067519888281822, + 0.0029023163951933384, + 0.005286828149110079, + 0.004807660821825266, + 0.0028185879345983267, + 0.00763660715892911, + 0.005107637494802475, + 0.0068326592445373535, + 0.006066821049898863, + 0.004973506089299917, + 0.010375170037150383, + 0.008015555329620838, + 0.006087446119636297, + 0.009334427304565907, + 0.003669923869892955, + 0.007056997157633305, + 0.0038685051258653402, + 0.004270545672625303, + 0.006684459745883942, + 0.0059640794061124325, + 0.0035827348474413157, + 0.0055131204426288605, + 0.003506917506456375, + 0.006148310843855143, + 0.005975943524390459, + 0.0044622416608035564, + 0.0066283149644732475, + 0.00470978207886219, + 0.0035579779651015997, + 0.0045669199898839, + 0.009335989132523537, + 0.011954938061535358, + 0.017462195828557014, + 0.0082652373239398, + 0.01383628323674202, + 0.011366366408765316, + 0.010064010508358479, + 0.016504263505339622, + 0.014554903842508793, + 0.017567405477166176, + 0.020444558933377266, + 0.013175824657082558, + 0.01534330565482378, + 0.017318880185484886, + 0.010498751886188984, + 0.01641961932182312 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.01605827920138836 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_tensor([5431, 5431])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_tensor([5431, 5431])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c1eddf0e69fd728c6d7db3178bfb7a41aea7b061 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_tensor([5431, 5431])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot begins in a medium close-up, then shifts to a wide, distant vista. The frame initially focuses on a young woman standing on a grassy, breeze-swept hill, bathed in the warm, golden glow of the evening sun, casting a long, soft shadow. She has long, dark brown hair, strands of which are blown across her delicate face by the wind. She wears a simple, elegant white cotton dress. A light, translucent, and subtly shimmering scarf, held in her hand, slips delicately from her grasp as her slender, white fingers unwrap it. Just as the scarf is about to touch the ground, a stronger gust of wind suddenly picks it up. The camera smoothly tilts upward and slowly zooms out, perfectly capturing the scarf, lifted by the wind, gracefully swirling, unfurling, and dancing in the air. Like a vibrant red ribbon, it rises higher and higher against a vast, pale blue sky, dotted with wisps of wispy clouds tinged pink by the setting sun. The girl looks up, her face revealing a complex expression of surprise and a hint of melancholy, while the red scarf becomes smaller and smaller in her field of vision, eventually turning into a tiny red dot that gradually disappears. The whole picture, in the warm light and shadow, creates a poignant and peaceful atmosphere about the passing of time and gentle melancholy.", + "seed": 8277, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_5431_seed_8277_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0175449947360903, + "vlm_logprob": -0.0175449947360903, + "vlm_yes_no_margin": 4.774999952316284, + "vlm_token_credit": -0.0175449947360903, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02184741012752056 + ], + "opposite_logprob_tokens": [ + -4.021847248077393 + ], + "logprob_reward": -0.02184741012752056, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.02184741012752056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1964769810438156, + 0.04775083810091019, + 0.09467121213674545, + 0.09526363015174866, + 0.08399612456560135, + 0.062257613986730576, + 0.13729366660118103, + 0.04950406402349472, + 0.13315054774284363, + 0.09963534772396088 + ], + "credit_map": [ + 0.03991344943642616, + 0.03866775706410408, + 0.013432810083031654, + 0.021723473444581032, + 0.011790972203016281, + 0.037524957209825516, + 0.018449930474162102, + 0.014973620884120464, + 0.005797430407255888, + 0.006771245505660772, + 0.005784664303064346, + 0.005195217672735453, + 0.004086221102625132, + 0.0072702886536717415, + 0.006194178946316242, + 0.006651595234870911, + 0.016317520290613174, + 0.017466943711042404, + 0.00828784704208374, + 0.00590125285089016, + 0.008251117542386055, + 0.017848601564764977, + 0.008551152423024178, + 0.012046769261360168, + 0.017300033941864967, + 0.01269090548157692, + 0.013903350569307804, + 0.007304389029741287, + 0.006847064942121506, + 0.017709145322442055, + 0.009884782135486603, + 0.009623954072594643, + 0.007974059320986271, + 0.010869032703340054, + 0.009917470626533031, + 0.016544431447982788, + 0.011803395114839077, + 0.009685599245131016, + 0.007653836160898209, + 0.00954829528927803, + 0.00439032819122076, + 0.019654445350170135, + 0.00633660051971674, + 0.0057001314125955105, + 0.007222380489110947, + 0.008862946182489395, + 0.0059210024774074554, + 0.0041697765700519085, + 0.0052332524210214615, + 0.013842517510056496, + 0.006621381267905235, + 0.005811196751892567, + 0.053514447063207626, + 0.04059654846787453, + 0.004144793841987848, + 0.007529538124799728, + 0.0031773007940500975, + 0.006180780474096537, + 0.0028309521730989218, + 0.003820595098659396, + 0.01696249283850193, + 0.009812513366341591, + 0.003272336907684803, + 0.0034470914397388697, + 0.00629337877035141, + 0.01510658673942089, + 0.008926166221499443, + 0.010027173906564713, + 0.04967853054404259, + 0.03058810718357563, + 0.006999640259891748, + 0.00553095992654562, + 0.005044737830758095, + 0.013786235824227333, + 0.0046121240593492985, + 0.00404277304187417, + 0.04014803469181061, + 0.023131554946303368, + 0.004461974371224642, + 0.004407916218042374 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.02184741012752056 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003637841437011957 + ], + "opposite_logprob_tokens": [ + -5.628637790679932 + ], + "logprob_reward": -0.003637841437011957, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003637841437011957, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1421804428100586, + 0.0443788580596447, + 0.0431947335600853, + 0.09642891585826874, + 0.1472977101802826, + 0.05740838125348091, + 0.10639480501413345, + 0.07047323882579803, + 0.06548678874969482, + 0.22675612568855286 + ], + "credit_map": [ + 0.0263617392629385, + 0.022993179038167, + 0.015388082712888718, + 0.012320374138653278, + 0.0098918117582798, + 0.02487856335937977, + 0.012161403894424438, + 0.018185274675488472, + 0.00681722117587924, + 0.008055383339524269, + 0.007677507121115923, + 0.003346678800880909, + 0.0032437527552247047, + 0.0071733687072992325, + 0.003959060646593571, + 0.00410588551312685, + 0.005597906187176704, + 0.007278710603713989, + 0.0034038289450109005, + 0.004578934982419014, + 0.007043766789138317, + 0.004513191059231758, + 0.0051323785446584225, + 0.005646016448736191, + 0.01469874195754528, + 0.015909995883703232, + 0.00667545897886157, + 0.011512231081724167, + 0.014396460726857185, + 0.015833916142582893, + 0.01002967357635498, + 0.007372431457042694, + 0.016454437747597694, + 0.020201301202178, + 0.025959137827157974, + 0.035531457513570786, + 0.010146554559469223, + 0.011653486639261246, + 0.01352713257074356, + 0.013824207708239555, + 0.00670875608921051, + 0.010950741358101368, + 0.009232844226062298, + 0.010125471279025078, + 0.004498111549764872, + 0.005293350666761398, + 0.0051514022052288055, + 0.0054477001540362835, + 0.005186648573726416, + 0.01002907007932663, + 0.0055144126527011395, + 0.005904700141400099, + 0.03743525967001915, + 0.029755428433418274, + 0.008193207904696465, + 0.0043760803528130054, + 0.004061280284076929, + 0.008320176973938942, + 0.005416839849203825, + 0.007975082844495773, + 0.021238574758172035, + 0.012667669914662838, + 0.005177410785108805, + 0.005616200156509876, + 0.004365565255284309, + 0.007907195948064327, + 0.008308296091854572, + 0.010032953694462776, + 0.01618330180644989, + 0.009882718324661255, + 0.004381662234663963, + 0.004425095394253731, + 0.008400067687034607, + 0.0616120845079422, + 0.0050237178802490234, + 0.005103182513266802, + 0.10241685062646866, + 0.0351361408829689, + 0.005452883895486593, + 0.003611180931329727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 0 + ], + "selected_reward": -0.003637841437011957 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010456338059157133 + ], + "opposite_logprob_tokens": [ + -6.8760457038879395 + ], + "logprob_reward": -0.0010456338059157133, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010456338059157133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08187733590602875, + 0.028034696355462074, + 0.10363876074552536, + 0.061822742223739624, + 0.03749627247452736, + 0.06854874640703201, + 0.06537062674760818, + 0.02945806458592415, + 0.05947700887918472, + 0.4642758071422577 + ], + "credit_map": [ + 0.01712408848106861, + 0.016979815438389778, + 0.0066925170831382275, + 0.006306338123977184, + 0.005408048164099455, + 0.018172282725572586, + 0.006083027459681034, + 0.005111220758408308, + 0.004444925580173731, + 0.0042602382600307465, + 0.003288534004241228, + 0.0032057117205113173, + 0.0020637400448322296, + 0.003523471299558878, + 0.0038290356751531363, + 0.0034190407022833824, + 0.017882341518998146, + 0.025836903601884842, + 0.006709105335175991, + 0.007122523616999388, + 0.012283861637115479, + 0.01739686354994774, + 0.008096372708678246, + 0.008310790173709393, + 0.012141315266489983, + 0.009963207878172398, + 0.005865230690687895, + 0.005738690495491028, + 0.006778236944228411, + 0.010999084450304508, + 0.004564277827739716, + 0.005772696807980537, + 0.0036195169668644667, + 0.005344930104911327, + 0.0036720309872180223, + 0.010591602884232998, + 0.004591824021190405, + 0.0024403936695307493, + 0.004272624384611845, + 0.0029633487574756145, + 0.006009445525705814, + 0.00906175747513771, + 0.014871722087264061, + 0.015208172611892223, + 0.005357408430427313, + 0.005708423443138599, + 0.006271039135754108, + 0.006060780491679907, + 0.002106142695993185, + 0.0062488978728652, + 0.0033890006598085165, + 0.0036304432433098555, + 0.02456824667751789, + 0.020840218290686607, + 0.0022867002990096807, + 0.0023009818978607655, + 0.002274074824526906, + 0.004109050612896681, + 0.005301733035594225, + 0.006065499037504196, + 0.00387886562384665, + 0.00298706884495914, + 0.0028303605504333973, + 0.002011412288993597, + 0.0022909697145223618, + 0.007393165957182646, + 0.0036327273119241, + 0.0032334821298718452, + 0.02387399785220623, + 0.014162478037178516, + 0.002846754854544997, + 0.0020434365142136812, + 0.01944848522543907, + 0.10982533544301987, + 0.008189672604203224, + 0.008632222190499306, + 0.21563862264156342, + 0.08583920449018478, + 0.010351024568080902, + 0.006351215299218893 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.0010456338059157133 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012538060545921326 + ], + "opposite_logprob_tokens": [ + -4.387537956237793 + ], + "logprob_reward": -0.012538060545921326, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012538060545921326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048995427787303925, + 0.04947197064757347, + 0.07909209281206131, + 0.040483083575963974, + 0.05533851310610771, + 0.13069872558116913, + 0.09479151666164398, + 0.0552399642765522, + 0.24873843789100647, + 0.19715026021003723 + ], + "credit_map": [ + 0.007954703643918037, + 0.009293437004089355, + 0.0043683298863470554, + 0.0071888091042637825, + 0.00320509634912014, + 0.007435915991663933, + 0.005694258492439985, + 0.003854873590171337, + 0.006705801468342543, + 0.009957391768693924, + 0.00375547562725842, + 0.0036050190683454275, + 0.004045849200338125, + 0.008676313795149326, + 0.005286720115691423, + 0.007439397741109133, + 0.013376659713685513, + 0.021562639623880386, + 0.00699607003480196, + 0.006319379433989525, + 0.008860602974891663, + 0.011127403005957603, + 0.005868218839168549, + 0.004981124307960272, + 0.006458103656768799, + 0.005605448968708515, + 0.0032344881910830736, + 0.004416787996888161, + 0.006495298817753792, + 0.0053947060368955135, + 0.004542102571576834, + 0.00433614756911993, + 0.005864898208528757, + 0.007412413600832224, + 0.007303217425942421, + 0.013350561261177063, + 0.0034558444749563932, + 0.005452907178550959, + 0.006120762322098017, + 0.006377910263836384, + 0.011421265080571175, + 0.017346495762467384, + 0.026173854246735573, + 0.02945985645055771, + 0.010929979383945465, + 0.01061300840228796, + 0.012818055227398872, + 0.011936199851334095, + 0.007145902141928673, + 0.016384463757276535, + 0.014770024456083775, + 0.014734191820025444, + 0.014917585998773575, + 0.013114971108734608, + 0.008236744441092014, + 0.0054876296781003475, + 0.004548593889921904, + 0.011866556480526924, + 0.007543074898421764, + 0.0063502611592411995, + 0.015619850717484951, + 0.004675510339438915, + 0.002341171260923147, + 0.002294943667948246, + 0.01749081164598465, + 0.03215576335787773, + 0.03016006574034691, + 0.035292863845825195, + 0.07147373259067535, + 0.03573138639330864, + 0.012035509571433067, + 0.01439831592142582, + 0.00940127857029438, + 0.04127444326877594, + 0.0036100081633776426, + 0.004407742992043495, + 0.09064028412103653, + 0.039412740617990494, + 0.0049318610690534115, + 0.003471893258392811 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 5 + ], + "selected_reward": -0.012538060545921326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048656027764081955 + ], + "opposite_logprob_tokens": [ + -3.0486559867858887 + ], + "logprob_reward": -0.048656027764081955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048656027764081955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0775490328669548, + 0.015981948003172874, + 0.035084377974271774, + 0.027149463072419167, + 0.05242612957954407, + 0.10776836425065994, + 0.22397828102111816, + 0.07338740676641464, + 0.12223218381404877, + 0.2644428014755249 + ], + "credit_map": [ + 0.026080559939146042, + 0.005954244639724493, + 0.003313785418868065, + 0.0038851366844028234, + 0.0021881188731640577, + 0.02910320833325386, + 0.0031147829722613096, + 0.0039092013612389565, + 0.002260518493130803, + 0.0015389516483992338, + 0.0024955482222139835, + 0.002247507218271494, + 0.0012369221076369286, + 0.0020173820666968822, + 0.0016767102060839534, + 0.0025084081571549177, + 0.0047409445978701115, + 0.006538961082696915, + 0.003991304896771908, + 0.004001769237220287, + 0.004264902789145708, + 0.003201443701982498, + 0.0036586017813533545, + 0.004686454311013222, + 0.0036448556929826736, + 0.0031895735301077366, + 0.004811427555978298, + 0.0031512246932834387, + 0.0028718241956084967, + 0.003979421686381102, + 0.003086901968345046, + 0.0024142367765307426, + 0.00591634027659893, + 0.00678400881588459, + 0.00535936001688242, + 0.010517521761357784, + 0.004703981801867485, + 0.008802820928394794, + 0.004229414276778698, + 0.006112684495747089, + 0.009696303866803646, + 0.013991338200867176, + 0.022326257079839706, + 0.030780989676713943, + 0.009564102627336979, + 0.00948525033891201, + 0.0055365366861224174, + 0.006387590896338224, + 0.008190124295651913, + 0.02407621406018734, + 0.014570583589375019, + 0.011452848091721535, + 0.0836448147892952, + 0.061034128069877625, + 0.009982297196984291, + 0.011027300730347633, + 0.005583532154560089, + 0.0138784758746624, + 0.00536860479041934, + 0.0053000361658632755, + 0.02588219754397869, + 0.008496755734086037, + 0.00452053127810359, + 0.004357283469289541, + 0.007247434463351965, + 0.014038550667464733, + 0.00997900776565075, + 0.010008063167333603, + 0.03711647540330887, + 0.02349991351366043, + 0.011231130920350552, + 0.009111619554460049, + 0.011971044354140759, + 0.05847768113017082, + 0.0050094760954380035, + 0.0050690327771008015, + 0.12187305837869644, + 0.05091084912419319, + 0.006764655467122793, + 0.0043670074082911015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 8 + ], + "selected_reward": -0.048656027764081955 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in a medium close-up, then shifts to a wide, distant vista. The frame initially focuses on a young woman standing on a grassy, breeze-swept hill, bathed in the warm, golden glow of the evening sun, casting a long, soft shadow. She has long, dark brown hair, strands of which are blown across her delicate face by the wind. She wears a simple, elegant white cotton dress. A light, translucent, and subtly shimmering scarf, held in her hand, slips delicately from her grasp as her slender, white fingers unwrap it. Just as the scarf is about to touch the ground, a stronger gust of wind suddenly picks it up. The camera smoothly tilts upward and slowly zooms out, perfectly capturing the scarf, lifted by the wind, gracefully swirling, unfurling, and dancing in the air. Like a vibrant red ribbon, it rises higher and higher against a vast, pale blue sky, dotted with wisps of wispy clouds tinged pink by the setting sun. The girl looks up, her face revealing a complex expression of surprise and a hint of melancholy, while the red scarf becomes smaller and smaller in her field of vision, eventually turning into a tiny red dot that gradually disappears. The whole picture, in the warm light and shadow, creates a poignant and peaceful atmosphere about the passing of time and gentle melancholy.", + "seed": 8278, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_5431_seed_8278_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.016703382111154498, + "vlm_logprob": -0.016703382111154498, + "vlm_yes_no_margin": 5.15, + "vlm_token_credit": -0.016703382111154498, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010850241407752037 + ], + "opposite_logprob_tokens": [ + -5.010850429534912 + ], + "logprob_reward": -0.010850241407752037, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.010850241407752037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3780367076396942, + 0.0461469404399395, + 0.0111032510176301, + 0.012951996177434921, + 0.02670406550168991, + 0.015341763384640217, + 0.031335245817899704, + 0.21989968419075012, + 0.0351179875433445, + 0.22336238622665405 + ], + "credit_map": [ + 0.018357813358306885, + 0.006178822368383408, + 0.13811416923999786, + 0.013599121943116188, + 0.010011118836700916, + 0.019591767340898514, + 0.007846242748200893, + 0.1643376499414444, + 0.0037199361249804497, + 0.002182852244004607, + 0.0054898387752473354, + 0.003043567528948188, + 0.003174159675836563, + 0.013109368272125721, + 0.004371992312371731, + 0.011055223643779755, + 0.0014956629602238536, + 0.001862310920841992, + 0.001645087730139494, + 0.0009003615705296397, + 0.0011270049726590514, + 0.0013810930540785193, + 0.0013955425238236785, + 0.0012961868196725845, + 0.001603213488124311, + 0.0027920021675527096, + 0.0016114427708089352, + 0.001782838487997651, + 0.0012256780173629522, + 0.0013271261705085635, + 0.0008960895356722176, + 0.0017136054811999202, + 0.0035390418488532305, + 0.0032530217431485653, + 0.0034170858561992645, + 0.004617688246071339, + 0.003064130200073123, + 0.003413864877074957, + 0.002419872209429741, + 0.0029793609865009785, + 0.0014964004512876272, + 0.002110185567289591, + 0.001789677538909018, + 0.001711514312773943, + 0.0025038968306034803, + 0.001960414694622159, + 0.002235670341178775, + 0.0015340038808062673, + 0.0015203359071165323, + 0.007303097750991583, + 0.002561769913882017, + 0.0023455386981368065, + 0.008639008738100529, + 0.002915935590863228, + 0.0035362711641937494, + 0.002513286890462041, + 0.0048800804652273655, + 0.09423837065696716, + 0.002600584179162979, + 0.004352908115833998, + 0.09717170149087906, + 0.009532748721539974, + 0.004418832715600729, + 0.002704464364796877, + 0.0020926599390804768, + 0.010834557004272938, + 0.0024024536833167076, + 0.0032653792295604944, + 0.007833817042410374, + 0.003524911357089877, + 0.0021213721483945847, + 0.003042836906388402, + 0.009050997905433178, + 0.0797027125954628, + 0.008348604664206505, + 0.014040788635611534, + 0.06149865314364433, + 0.029362179338932037, + 0.00837618112564087, + 0.012982274405658245 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.010850241407752037 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019538376946002245 + ], + "opposite_logprob_tokens": [ + -6.251953601837158 + ], + "logprob_reward": -0.0019538376946002245, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019538376946002245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13407926261425018, + 0.06506723165512085, + 0.05478282272815704, + 0.07072725147008896, + 0.06349959969520569, + 0.03662005811929703, + 0.08108707517385483, + 0.21305714547634125, + 0.04245786368846893, + 0.23862169682979584 + ], + "credit_map": [ + 0.01058844942599535, + 0.004095553886145353, + 0.03940917178988457, + 0.005281520076096058, + 0.006028067320585251, + 0.016120081767439842, + 0.004554525017738342, + 0.04800189286470413, + 0.005567421205341816, + 0.0020611686632037163, + 0.009054985828697681, + 0.0026449947617948055, + 0.0036768862046301365, + 0.019510259851813316, + 0.00400027958676219, + 0.018551237881183624, + 0.010056956671178341, + 0.009851559065282345, + 0.007531651295721531, + 0.007471493910998106, + 0.0036195635329931974, + 0.0057077971287071705, + 0.003384864656254649, + 0.007158937398344278, + 0.008058471605181694, + 0.013766481541097164, + 0.009556903503835201, + 0.01043007429689169, + 0.00720549002289772, + 0.0068871150724589825, + 0.00562971830368042, + 0.0091929966583848, + 0.0070882029831409454, + 0.006538776680827141, + 0.009611880406737328, + 0.014054615050554276, + 0.008020857349038124, + 0.006452960427850485, + 0.004728574771434069, + 0.007003732956945896, + 0.00393639225512743, + 0.005161371547728777, + 0.0044434512965381145, + 0.0031126353424042463, + 0.005682443268597126, + 0.005270736757665873, + 0.005583000369369984, + 0.0034300254192203283, + 0.0037196206394582987, + 0.02108180709183216, + 0.006158240605145693, + 0.0070903366431593895, + 0.0242012906819582, + 0.005796935874968767, + 0.007329643238335848, + 0.0057091969065368176, + 0.005148336291313171, + 0.0825517326593399, + 0.004303840454667807, + 0.006601356435567141, + 0.0895589143037796, + 0.012878568843007088, + 0.006043390836566687, + 0.0059709977358579636, + 0.002322731539607048, + 0.014058349654078484, + 0.0027308694552630186, + 0.00280944537371397, + 0.010166224092245102, + 0.00468028848990798, + 0.002412953181192279, + 0.00327700306661427, + 0.009018367156386375, + 0.08876051008701324, + 0.008577907457947731, + 0.013966070488095284, + 0.06425057351589203, + 0.03262417018413544, + 0.00895176362246275, + 0.012472342699766159 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.0019538376946002245 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010453957365825772 + ], + "opposite_logprob_tokens": [ + -6.876045227050781 + ], + "logprob_reward": -0.0010453957365825772, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010453957365825772, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.048374127596616745, + 0.02198353409767151, + 0.028652219101786613, + 0.027305683121085167, + 0.024061188101768494, + 0.021456599235534668, + 0.030061865225434303, + 0.4760168790817261, + 0.034524738788604736, + 0.28756314516067505 + ], + "credit_map": [ + 0.004054170101881027, + 0.0015450229402631521, + 0.013259697705507278, + 0.001868873368948698, + 0.0020265120547264814, + 0.007066818419843912, + 0.0014801494544371963, + 0.017072878777980804, + 0.002751315711066127, + 0.0009272529277950525, + 0.0026827803812921047, + 0.0012778933160007, + 0.0014007992576807737, + 0.005503531079739332, + 0.002145122503861785, + 0.005294838920235634, + 0.005338343791663647, + 0.005263242404907942, + 0.00464249961078167, + 0.0027646669186651707, + 0.002070188522338867, + 0.002998888958245516, + 0.0021167148370295763, + 0.0034576731268316507, + 0.0031159708742052317, + 0.005742276553064585, + 0.004143065772950649, + 0.0033916013780981302, + 0.002808987395837903, + 0.0024383901618421078, + 0.0022371169179677963, + 0.0034282742999494076, + 0.0030820865649729967, + 0.0024834780488163233, + 0.003224837128072977, + 0.004193937871605158, + 0.002788790501654148, + 0.0031169517897069454, + 0.0026764366775751114, + 0.002494670683518052, + 0.0018347719451412559, + 0.0026613916270434856, + 0.0024819588288664818, + 0.002469742437824607, + 0.0036386079154908657, + 0.0040282560512423515, + 0.002427323954179883, + 0.001914546824991703, + 0.00172834238037467, + 0.00603108387440443, + 0.003673751140013337, + 0.003456323640421033, + 0.00676589272916317, + 0.0028056632727384567, + 0.002841884270310402, + 0.002758924849331379, + 0.01112504955381155, + 0.20811288058757782, + 0.004517251625657082, + 0.008294827304780483, + 0.21095424890518188, + 0.01845378242433071, + 0.009217417798936367, + 0.005341416224837303, + 0.001858932082541287, + 0.00864157360047102, + 0.0024292997550219297, + 0.0029512259643524885, + 0.011021001264452934, + 0.003488900139927864, + 0.002309430157765746, + 0.001824375707656145, + 0.01037687063217163, + 0.10991345345973969, + 0.008855930529534817, + 0.015455861575901508, + 0.07284869253635406, + 0.04416597634553909, + 0.010566523298621178, + 0.015379823744297028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.0010453957365825772 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11215727776288986, + 0.023878619074821472, + 0.01777467131614685, + 0.020976584404706955, + 0.10765938460826874, + 0.17693811655044556, + 0.021927906200289726, + 0.3054783046245575, + 0.11495904624462128, + 0.09825004637241364 + ], + "credit_map": [ + 0.00875268504023552, + 0.0025536101311445236, + 0.03693835064768791, + 0.004396067000925541, + 0.0037751581985503435, + 0.007867339067161083, + 0.0030837247613817453, + 0.04479033872485161, + 0.0030034424271434546, + 0.0017892904579639435, + 0.0032700125593692064, + 0.0023750532418489456, + 0.002365361899137497, + 0.0036028672475367785, + 0.0025932134594768286, + 0.004879377782344818, + 0.0024549528025090694, + 0.003765102941542864, + 0.0019819687586277723, + 0.002217757049947977, + 0.0015116207068786025, + 0.0019371564267203212, + 0.002084407489746809, + 0.0018217050237581134, + 0.002017086837440729, + 0.004884011577814817, + 0.002920303260907531, + 0.0020818032789975405, + 0.0019147343700751662, + 0.0020834626629948616, + 0.0029228730127215385, + 0.002152309287339449, + 0.01810169778764248, + 0.01321976538747549, + 0.020232969895005226, + 0.014151962473988533, + 0.014612168073654175, + 0.010975261218845844, + 0.006986623164266348, + 0.00937893707305193, + 0.01582939364016056, + 0.019721565768122673, + 0.04217987135052681, + 0.02674311213195324, + 0.030434811487793922, + 0.017031094059348106, + 0.01128171756863594, + 0.013716565445065498, + 0.0017074515344575047, + 0.002756512025371194, + 0.002976600080728531, + 0.0026538718957453966, + 0.0034508584067225456, + 0.002753034932538867, + 0.0032915666233748198, + 0.002338011981919408, + 0.007161715999245644, + 0.13315452635288239, + 0.003186316229403019, + 0.0057304673828184605, + 0.13476873934268951, + 0.01154476497322321, + 0.006223824806511402, + 0.0037079607136547565, + 0.0035632329527288675, + 0.039998676627874374, + 0.003800142090767622, + 0.0050888946279883385, + 0.043636664748191833, + 0.008663270622491837, + 0.0048036593943834305, + 0.005404502619057894, + 0.004992542788386345, + 0.030674202367663383, + 0.005018646828830242, + 0.0070322658866643906, + 0.028440535068511963, + 0.011383302509784698, + 0.004351262468844652, + 0.0063572898507118225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 8 + ], + "selected_reward": -0.007625873666256666 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204156205058098 + ], + "opposite_logprob_tokens": [ + -2.8120415210723877 + ], + "logprob_reward": -0.06204156205058098, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204156205058098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12101612985134125, + 0.019764291122555733, + 0.02387472614645958, + 0.038925908505916595, + 0.12057871371507645, + 0.046697087585926056, + 0.04222014546394348, + 0.3005538880825043, + 0.07359793782234192, + 0.21277111768722534 + ], + "credit_map": [ + 0.0067115346901118755, + 0.002458747010678053, + 0.04221635311841965, + 0.004699826240539551, + 0.0036773348692804575, + 0.007243525702506304, + 0.0033376116771250963, + 0.05067119747400284, + 0.0033148424699902534, + 0.0013389990199357271, + 0.002502240240573883, + 0.0019544607494026423, + 0.0018842426361516118, + 0.0032450275029987097, + 0.0028490042313933372, + 0.0026754741556942463, + 0.0028737783432006836, + 0.0022325213067233562, + 0.003180671716108918, + 0.0028178261127322912, + 0.0027067495975643396, + 0.0033268250990659, + 0.0028025165665894747, + 0.0039338357746601105, + 0.0031761196441948414, + 0.006426529493182898, + 0.010700293816626072, + 0.004275778774172068, + 0.003045865101739764, + 0.0030635499861091375, + 0.0023082662373781204, + 0.005929505918174982, + 0.018296442925930023, + 0.014695323072373867, + 0.023516355082392693, + 0.019047308713197708, + 0.014749440364539623, + 0.01035025529563427, + 0.008996570482850075, + 0.010927018709480762, + 0.004042462911456823, + 0.005351340398192406, + 0.00770137645304203, + 0.00624519819393754, + 0.006961079314351082, + 0.005153273232281208, + 0.006062690634280443, + 0.005179662723094225, + 0.0019433279521763325, + 0.00421205023303628, + 0.005992695223540068, + 0.004305259790271521, + 0.005333540495485067, + 0.004787728656083345, + 0.008651407435536385, + 0.006994131952524185, + 0.006955090444535017, + 0.12946660816669464, + 0.0033877091482281685, + 0.005114502273499966, + 0.1324523240327835, + 0.012357328087091446, + 0.00613071583211422, + 0.004689616616815329, + 0.002228030702099204, + 0.02705366164445877, + 0.002327226335182786, + 0.003498978214338422, + 0.025460897013545036, + 0.00612907437607646, + 0.003478387836366892, + 0.00342168309725821, + 0.007337410002946854, + 0.08147536218166351, + 0.006999078206717968, + 0.012606150470674038, + 0.05688423663377762, + 0.028561536222696304, + 0.008837603963911533, + 0.010069732554256916 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.06204156205058098 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in a medium close-up, then shifts to a wide, distant vista. The frame initially focuses on a young woman standing on a grassy, breeze-swept hill, bathed in the warm, golden glow of the evening sun, casting a long, soft shadow. She has long, dark brown hair, strands of which are blown across her delicate face by the wind. She wears a simple, elegant white cotton dress. A light, translucent, and subtly shimmering scarf, held in her hand, slips delicately from her grasp as her slender, white fingers unwrap it. Just as the scarf is about to touch the ground, a stronger gust of wind suddenly picks it up. The camera smoothly tilts upward and slowly zooms out, perfectly capturing the scarf, lifted by the wind, gracefully swirling, unfurling, and dancing in the air. Like a vibrant red ribbon, it rises higher and higher against a vast, pale blue sky, dotted with wisps of wispy clouds tinged pink by the setting sun. The girl looks up, her face revealing a complex expression of surprise and a hint of melancholy, while the red scarf becomes smaller and smaller in her field of vision, eventually turning into a tiny red dot that gradually disappears. The whole picture, in the warm light and shadow, creates a poignant and peaceful atmosphere about the passing of time and gentle melancholy.", + "seed": 8279, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_5431_seed_8279_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01474478468298912, + "vlm_logprob": -0.01474478468298912, + "vlm_yes_no_margin": 4.9, + "vlm_token_credit": -0.01474478468298912, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015408285893499851 + ], + "opposite_logprob_tokens": [ + -4.765408515930176 + ], + "logprob_reward": -0.015408285893499851, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.015408285893499851, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3254634439945221, + 0.044367559254169464, + 0.06191973760724068, + 0.03874369338154793, + 0.027297431603074074, + 0.06595995277166367, + 0.20075015723705292, + 0.07820039987564087, + 0.08594480156898499, + 0.07135281711816788 + ], + "credit_map": [ + 0.07248838245868683, + 0.04975311458110809, + 0.06677542626857758, + 0.030417270958423615, + 0.018470142036676407, + 0.03772762790322304, + 0.01744740828871727, + 0.032384082674980164, + 0.008345825597643852, + 0.007494951132684946, + 0.007934404537081718, + 0.0032888087444007397, + 0.0021181879565119743, + 0.008602121844887733, + 0.0020866813138127327, + 0.004496580921113491, + 0.01049110759049654, + 0.007902555167675018, + 0.011614901944994926, + 0.008703554049134254, + 0.0039178417064249516, + 0.007655277382582426, + 0.006108035799115896, + 0.005526472814381123, + 0.005340239033102989, + 0.004560740198940039, + 0.006421390920877457, + 0.004127212800085545, + 0.0040279473178088665, + 0.003543092170730233, + 0.006472970824688673, + 0.004250105004757643, + 0.0032533141784369946, + 0.0048452094197273254, + 0.0042024110443890095, + 0.003948284313082695, + 0.002441080752760172, + 0.0030762231908738613, + 0.002584233181551099, + 0.0029466762207448483, + 0.01393089909106493, + 0.023598020896315575, + 0.0047422125935554504, + 0.004469653591513634, + 0.005839485675096512, + 0.006451776251196861, + 0.0037348908372223377, + 0.0031930231489241123, + 0.041416045278310776, + 0.03942163661122322, + 0.01231854222714901, + 0.012144574895501137, + 0.024202648550271988, + 0.052103616297245026, + 0.012024292722344398, + 0.007118819281458855, + 0.007190737407654524, + 0.011643010191619396, + 0.007574433460831642, + 0.00672572199255228, + 0.014220013283193111, + 0.012220674194395542, + 0.013109556399285793, + 0.005516259931027889, + 0.005323503632098436, + 0.026418443769216537, + 0.004878202918916941, + 0.004098532255738974, + 0.02696511335670948, + 0.006421368103474379, + 0.007293529808521271, + 0.0045461091212928295, + 0.005405317060649395, + 0.008411295711994171, + 0.013104232959449291, + 0.006503259763121605, + 0.014541162177920341, + 0.011217976920306683, + 0.006259034387767315, + 0.005910548381507397 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.015408285893499851 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002512633567675948 + ], + "opposite_logprob_tokens": [ + -6.002512454986572 + ], + "logprob_reward": -0.002512633567675948, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002512633567675948, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11352337151765823, + 0.07885429263114929, + 0.038889702409505844, + 0.05877995118498802, + 0.10866466909646988, + 0.1743389517068863, + 0.0689820796251297, + 0.0405723936855793, + 0.10938788950443268, + 0.20800675451755524 + ], + "credit_map": [ + 0.02569827251136303, + 0.01661916822195053, + 0.02384549006819725, + 0.008928335271775723, + 0.006720389239490032, + 0.013515513390302658, + 0.0050504389218986034, + 0.013145766220986843, + 0.008448220789432526, + 0.01675967313349247, + 0.005850112531334162, + 0.006156202405691147, + 0.006088819354772568, + 0.022163541987538338, + 0.0061006066389381886, + 0.007287114392966032, + 0.0057831075973808765, + 0.005445849150419235, + 0.008173453621566296, + 0.004724364727735519, + 0.0022924647200852633, + 0.006059613078832626, + 0.003138866275548935, + 0.0032719841692596674, + 0.008578086271882057, + 0.0073394435457885265, + 0.00609064195305109, + 0.006131341680884361, + 0.007677523884922266, + 0.005910171661525965, + 0.009224752895534039, + 0.007827987894415855, + 0.016312187537550926, + 0.030093800276517868, + 0.009579871781170368, + 0.009470495395362377, + 0.00922077801078558, + 0.012463930062949657, + 0.012357356958091259, + 0.009166245348751545, + 0.02630099281668663, + 0.04177241772413254, + 0.033866655081510544, + 0.027880912646651268, + 0.012069584801793098, + 0.010562974959611893, + 0.011115404777228832, + 0.010770010761916637, + 0.01080398727208376, + 0.01000573206692934, + 0.0046364436857402325, + 0.004455958027392626, + 0.011879818513989449, + 0.016337241977453232, + 0.005364480894058943, + 0.005498416256159544, + 0.0055788494646549225, + 0.005098552908748388, + 0.005075619090348482, + 0.005377724766731262, + 0.005216493736952543, + 0.005489662755280733, + 0.0056477077305316925, + 0.003087784629315138, + 0.012907866388559341, + 0.01477806456387043, + 0.011132749728858471, + 0.008913518860936165, + 0.014914124272763729, + 0.01628444716334343, + 0.019088031724095345, + 0.011369092389941216, + 0.0114792725071311, + 0.0329836867749691, + 0.024127313867211342, + 0.013863308355212212, + 0.06429892778396606, + 0.0378657765686512, + 0.013493279926478863, + 0.009895183145999908 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.002512633567675948 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015198357868939638 + ], + "opposite_logprob_tokens": [ + -6.501519680023193 + ], + "logprob_reward": -0.0015198357868939638, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015198357868939638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21873599290847778, + 0.029087306931614876, + 0.022806329652667046, + 0.02001078426837921, + 0.03097403235733509, + 0.0973200723528862, + 0.08186864852905273, + 0.052146561443805695, + 0.05428765341639519, + 0.3927626609802246 + ], + "credit_map": [ + 0.05143142119050026, + 0.03216685354709625, + 0.04402351379394531, + 0.02016175538301468, + 0.010396972298622131, + 0.020876700058579445, + 0.011148305609822273, + 0.028530457988381386, + 0.005227883346378803, + 0.004324458073824644, + 0.003589869476854801, + 0.0027934322133660316, + 0.0024979363661259413, + 0.0038570428732782602, + 0.004089232534170151, + 0.0027074527461081743, + 0.003935506101697683, + 0.003052595304325223, + 0.004964794032275677, + 0.003416030667722225, + 0.0012922256719321012, + 0.0031364953611046076, + 0.0013651754707098007, + 0.0016435073921456933, + 0.0027836307417601347, + 0.0021525761112570763, + 0.003228380111977458, + 0.002979336539283395, + 0.0016409249510616064, + 0.001808651490136981, + 0.00297815864905715, + 0.002439125208184123, + 0.0033669627737253904, + 0.004453274887055159, + 0.00910573173314333, + 0.00647089583799243, + 0.0011906102299690247, + 0.002304323250427842, + 0.0021362355910241604, + 0.0019459996838122606, + 0.021557234227657318, + 0.03662769868969917, + 0.006038455758243799, + 0.006589947734028101, + 0.005199662409722805, + 0.010255946777760983, + 0.006316154729574919, + 0.004734972026199102, + 0.017046667635440826, + 0.01696203462779522, + 0.005275527480989695, + 0.004760319367051125, + 0.00893829669803381, + 0.022646142169833183, + 0.003808996407315135, + 0.0024306660052388906, + 0.005320850759744644, + 0.006201983895152807, + 0.005293822847306728, + 0.0054423012770712376, + 0.010035347193479538, + 0.0082009918987751, + 0.00812460109591484, + 0.0035266648046672344, + 0.005827921908348799, + 0.008775335736572742, + 0.00864323042333126, + 0.006224213633686304, + 0.0077808331698179245, + 0.0068244352005422115, + 0.005061749368906021, + 0.005149933509528637, + 0.02959304116666317, + 0.06895403563976288, + 0.06977523863315582, + 0.024809056892991066, + 0.1104850172996521, + 0.06534411758184433, + 0.010701862163841724, + 0.013100308366119862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 5 + ], + "selected_reward": -0.0015198357868939638 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020581956952810287 + ], + "opposite_logprob_tokens": [ + -3.8955819606781006 + ], + "logprob_reward": -0.020581956952810287, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020581956952810287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11600357294082642, + 0.03183209151029587, + 0.05727769061923027, + 0.0277966670691967, + 0.02127433754503727, + 0.026231935247778893, + 0.2075119912624359, + 0.04793286323547363, + 0.09550449997186661, + 0.36863434314727783 + ], + "credit_map": [ + 0.028300728648900986, + 0.017680512741208076, + 0.02705729566514492, + 0.007965421304106712, + 0.004855566658079624, + 0.009430907666683197, + 0.0042726569809019566, + 0.016440492123365402, + 0.007577840238809586, + 0.005680440925061703, + 0.0038036848418414593, + 0.002095618983730674, + 0.0020213627722114325, + 0.005312731023877859, + 0.002434155670925975, + 0.002906257286667824, + 0.01306698564440012, + 0.009334569796919823, + 0.010417900048196316, + 0.007402114570140839, + 0.0021726228296756744, + 0.009073926135897636, + 0.002056425204500556, + 0.003753146855160594, + 0.00635541882365942, + 0.004376432858407497, + 0.0029528741724789143, + 0.003563646925613284, + 0.0017989969346672297, + 0.004831156227737665, + 0.0021224969532340765, + 0.0017956460360437632, + 0.002791706705465913, + 0.0049522314220666885, + 0.0022734098602086306, + 0.0024384569842368364, + 0.001620827242732048, + 0.0038929819129407406, + 0.0018985926872119308, + 0.0014061329420655966, + 0.004782491363584995, + 0.007489355746656656, + 0.0037267981097102165, + 0.0028625468257814646, + 0.001689839526079595, + 0.002224854426458478, + 0.0017142788274213672, + 0.001741773565299809, + 0.04608834907412529, + 0.03322717547416687, + 0.008784398436546326, + 0.010490277782082558, + 0.024484604597091675, + 0.06678495556116104, + 0.010384461842477322, + 0.007267780136317015, + 0.005950655788183212, + 0.00575824361294508, + 0.005077692214399576, + 0.004644571337848902, + 0.00864754244685173, + 0.008325880393385887, + 0.005556474905461073, + 0.003971800673753023, + 0.008638536557555199, + 0.020899653434753418, + 0.0073936753906309605, + 0.007239721715450287, + 0.01878035068511963, + 0.011809630319476128, + 0.011536503210663795, + 0.009206431917846203, + 0.018479395657777786, + 0.049296557903289795, + 0.04076454043388367, + 0.01937776803970337, + 0.1321079432964325, + 0.08813498914241791, + 0.01186202559620142, + 0.008611160330474377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.020581956952810287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370121121406555 + ], + "opposite_logprob_tokens": [ + -3.408701181411743 + ], + "logprob_reward": -0.03370121121406555, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370121121406555, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07253195345401764, + 0.03383393585681915, + 0.04629082977771759, + 0.03308872878551483, + 0.04306325316429138, + 0.051984190940856934, + 0.08270320296287537, + 0.13308027386665344, + 0.34083622694015503, + 0.16258744895458221 + ], + "credit_map": [ + 0.013692289590835571, + 0.009600612334907055, + 0.012786180712282658, + 0.007712636608630419, + 0.005590371321886778, + 0.010007813572883606, + 0.004798706620931625, + 0.008343338035047054, + 0.004533641505986452, + 0.00609097583219409, + 0.0061059375293552876, + 0.0032747092191129923, + 0.002414217684417963, + 0.005086603108793497, + 0.002593075390905142, + 0.003734773490577936, + 0.008941548876464367, + 0.007905608974397182, + 0.008322477340698242, + 0.004564136732369661, + 0.0020233544055372477, + 0.009209394454956055, + 0.0020185767207294703, + 0.0033057357650250196, + 0.00787598267197609, + 0.0053374203853309155, + 0.0036206848453730345, + 0.003360769245773554, + 0.001952437567524612, + 0.0055218655616045, + 0.0027382955886423588, + 0.002681269310414791, + 0.0060537937097251415, + 0.011384917423129082, + 0.005524543579667807, + 0.005974080413579941, + 0.00328604644164443, + 0.0034346827305853367, + 0.00413762591779232, + 0.0032675592228770256, + 0.010606569238007069, + 0.013769871555268764, + 0.006901986431330442, + 0.005423568654805422, + 0.0033524830359965563, + 0.005611148662865162, + 0.003524508560076356, + 0.0027940550353378057, + 0.013740826398134232, + 0.013558641076087952, + 0.0073289466090500355, + 0.00662007974460721, + 0.011120980605483055, + 0.013042481616139412, + 0.009417911060154438, + 0.007873335853219032, + 0.013828071765601635, + 0.018433229997754097, + 0.011454859748482704, + 0.011894844472408295, + 0.02685791626572609, + 0.021477822214365005, + 0.017398525029420853, + 0.011735016480088234, + 0.018375389277935028, + 0.09356658905744553, + 0.022378703579306602, + 0.021136051043868065, + 0.08951430767774582, + 0.02969934418797493, + 0.04451114311814308, + 0.02165469154715538, + 0.009236328303813934, + 0.0228443406522274, + 0.019639836624264717, + 0.01616421900689602, + 0.046442147344350815, + 0.026025226339697838, + 0.013459756039083004, + 0.008775594644248486 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.03370121121406555 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot begins in a medium close-up, then shifts to a wide, distant vista. The frame initially focuses on a young woman standing on a grassy, breeze-swept hill, bathed in the warm, golden glow of the evening sun, casting a long, soft shadow. She has long, dark brown hair, strands of which are blown across her delicate face by the wind. She wears a simple, elegant white cotton dress. A light, translucent, and subtly shimmering scarf, held in her hand, slips delicately from her grasp as her slender, white fingers unwrap it. Just as the scarf is about to touch the ground, a stronger gust of wind suddenly picks it up. The camera smoothly tilts upward and slowly zooms out, perfectly capturing the scarf, lifted by the wind, gracefully swirling, unfurling, and dancing in the air. Like a vibrant red ribbon, it rises higher and higher against a vast, pale blue sky, dotted with wisps of wispy clouds tinged pink by the setting sun. The girl looks up, her face revealing a complex expression of surprise and a hint of melancholy, while the red scarf becomes smaller and smaller in her field of vision, eventually turning into a tiny red dot that gradually disappears. The whole picture, in the warm light and shadow, creates a poignant and peaceful atmosphere about the passing of time and gentle melancholy.", + "seed": 8280, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_14/index_5431_seed_8280_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.014826919860206544, + "vlm_logprob": -0.014826919860206544, + "vlm_yes_no_margin": 4.85, + "vlm_token_credit": -0.014826919860206544, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020189495757222176 + ], + "opposite_logprob_tokens": [ + -4.27018928527832 + ], + "logprob_reward": -0.020189495757222176, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.020189495757222176, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6095892786979675, + 0.03597402945160866, + 0.027840785682201385, + 0.027514958754181862, + 0.03085784614086151, + 0.05978888273239136, + 0.03758404031395912, + 0.03310389071702957, + 0.044759854674339294, + 0.09298647195100784 + ], + "credit_map": [ + 0.07407236844301224, + 0.041461072862148285, + 0.19880411028862, + 0.04993073642253876, + 0.03024587593972683, + 0.04849571734666824, + 0.019118856638669968, + 0.14746055006980896, + 0.0043943594209849834, + 0.003850255161523819, + 0.004661161918193102, + 0.0029372817371040583, + 0.00498119555413723, + 0.004563967231661081, + 0.005623719654977322, + 0.004962089471518993, + 0.0038180185947567225, + 0.003347956109791994, + 0.003745428519323468, + 0.002545034745708108, + 0.004692758433520794, + 0.004602445289492607, + 0.002275743056088686, + 0.0028133993037045, + 0.0033786287531256676, + 0.004826107062399387, + 0.002661700826138258, + 0.002703270176425576, + 0.004871389828622341, + 0.0030758881475776434, + 0.003464668057858944, + 0.0025333063676953316, + 0.0032672861125320196, + 0.0038855411112308502, + 0.0025237342342734337, + 0.00268651288934052, + 0.006083215121179819, + 0.004381520673632622, + 0.004435774404555559, + 0.0035942622926086187, + 0.007162746042013168, + 0.011312411166727543, + 0.004577354993671179, + 0.004694558214396238, + 0.013816711492836475, + 0.009362922050058842, + 0.005293951835483313, + 0.003568227868527174, + 0.0033892174251377583, + 0.006544351577758789, + 0.0037553696893155575, + 0.0038307546637952328, + 0.003718380583450198, + 0.006960786879062653, + 0.005353636108338833, + 0.004031543154269457, + 0.0030706007964909077, + 0.005127570126205683, + 0.0024773043114691973, + 0.0024202235508710146, + 0.0064283874817192554, + 0.0059715760871768, + 0.002817785367369652, + 0.0047904434613883495, + 0.0039261989295482635, + 0.009454893879592419, + 0.004065781831741333, + 0.003238147124648094, + 0.007120097987353802, + 0.00929117202758789, + 0.003937332890927792, + 0.0037262265104800463, + 0.006934246048331261, + 0.011062792502343655, + 0.0075750709511339664, + 0.004053233191370964, + 0.018369384109973907, + 0.017134051769971848, + 0.006651592440903187, + 0.02120610699057579 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.020189495757222176 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028388698119670153 + ], + "opposite_logprob_tokens": [ + -5.877839088439941 + ], + "logprob_reward": -0.0028388698119670153, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028388698119670153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38781511783599854, + 0.06490985304117203, + 0.04165077582001686, + 0.11176388710737228, + 0.06548872590065002, + 0.04645940288901329, + 0.026514723896980286, + 0.039110928773880005, + 0.0318901427090168, + 0.184396430850029 + ], + "credit_map": [ + 0.04546047002077103, + 0.029134009033441544, + 0.1195753738284111, + 0.032716669142246246, + 0.018914207816123962, + 0.03077981248497963, + 0.014467300847172737, + 0.09676728397607803, + 0.013209165073931217, + 0.006995867937803268, + 0.006137696094810963, + 0.0043001421727240086, + 0.0055014402605593204, + 0.016113312914967537, + 0.0058288974687457085, + 0.006823329720646143, + 0.005733456928282976, + 0.004327303729951382, + 0.009220864623785019, + 0.0039962707087397575, + 0.004255227744579315, + 0.006483225151896477, + 0.0033267547842115164, + 0.004307672381401062, + 0.010793635621666908, + 0.01598307304084301, + 0.012751751579344273, + 0.011131558567285538, + 0.01796516589820385, + 0.012323737144470215, + 0.01654515415430069, + 0.014269816689193249, + 0.006088018883019686, + 0.008273819461464882, + 0.004643385298550129, + 0.006229460705071688, + 0.014605763368308544, + 0.008890300989151001, + 0.010217114351689816, + 0.006540860515087843, + 0.006169546861201525, + 0.008667202666401863, + 0.003428948111832142, + 0.004193565342575312, + 0.008726689033210278, + 0.006064444314688444, + 0.0041230893693864346, + 0.0050859153270721436, + 0.003356111003085971, + 0.003038882277905941, + 0.0028972686268389225, + 0.002648179419338703, + 0.003754469333216548, + 0.0031684287823736668, + 0.004319815430790186, + 0.0033315697219222784, + 0.0037867887876927853, + 0.008512438274919987, + 0.0026783901266753674, + 0.0026384054217487574, + 0.006027385592460632, + 0.0075284866616129875, + 0.0041815596632659435, + 0.0037574716843664646, + 0.0025200520176440477, + 0.003811221569776535, + 0.0029157069511711597, + 0.0025978884659707546, + 0.006094552110880613, + 0.005876233335584402, + 0.004491422791033983, + 0.003583066165447235, + 0.014689155854284763, + 0.029836395755410194, + 0.00850012619048357, + 0.008292267099022865, + 0.05592488870024681, + 0.03966989740729332, + 0.007639091927558184, + 0.019844604656100273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.0028388698119670153 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013406822690740228 + ], + "opposite_logprob_tokens": [ + -6.626340866088867 + ], + "logprob_reward": -0.0013406822690740228, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013406822690740228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2717415988445282, + 0.09223830699920654, + 0.04019876942038536, + 0.2003318816423416, + 0.09049797058105469, + 0.06001587212085724, + 0.0579078271985054, + 0.022159365937113762, + 0.025682849809527397, + 0.13922558724880219 + ], + "credit_map": [ + 0.03325777128338814, + 0.019061092287302017, + 0.09089552611112595, + 0.024137822911143303, + 0.013986412435770035, + 0.022258048877120018, + 0.011021692305803299, + 0.057123225182294846, + 0.02459030970931053, + 0.005439134314656258, + 0.0072442214004695415, + 0.004357622470706701, + 0.005351169966161251, + 0.03295626863837242, + 0.004887056071311235, + 0.007412528619170189, + 0.004356854595243931, + 0.004977427423000336, + 0.007718039210885763, + 0.003911507781594992, + 0.005721196066588163, + 0.0045789978466928005, + 0.0042199562303721905, + 0.004714791662991047, + 0.023455705493688583, + 0.03793298825621605, + 0.012193315662443638, + 0.015767278149724007, + 0.048782024532556534, + 0.027990171685814857, + 0.022387227043509483, + 0.01182317454367876, + 0.00705620925873518, + 0.011281014420092106, + 0.0068570892326533794, + 0.01186169683933258, + 0.01749177649617195, + 0.010202132165431976, + 0.017045192420482635, + 0.008702862076461315, + 0.007980704307556152, + 0.009187412448227406, + 0.004265098832547665, + 0.0038598631508648396, + 0.010616844519972801, + 0.010650509968400002, + 0.006105923093855381, + 0.007349517662078142, + 0.004705342464148998, + 0.008209007792174816, + 0.006943954154849052, + 0.004932883195579052, + 0.010136241093277931, + 0.009650876745581627, + 0.006482572760432959, + 0.006846947595477104, + 0.0017315354198217392, + 0.002475144574418664, + 0.00205004820600152, + 0.0023735181894153357, + 0.004316715989261866, + 0.003883663099259138, + 0.0028410854283720255, + 0.0024876557290554047, + 0.0022734159138053656, + 0.0032234624959528446, + 0.002845399547368288, + 0.0023307313676923513, + 0.004558459855616093, + 0.003833318827673793, + 0.0033956612460315228, + 0.003222399391233921, + 0.012700281105935574, + 0.026504414156079292, + 0.011106986552476883, + 0.007669338956475258, + 0.026231013238430023, + 0.0239341352134943, + 0.009308397769927979, + 0.02177102491259575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.0013406822690740228 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016068601980805397 + ], + "opposite_logprob_tokens": [ + -4.141068458557129 + ], + "logprob_reward": -0.016068601980805397, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016068601980805397, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07528793066740036, + 0.044092681258916855, + 0.06549660116434097, + 0.08907126635313034, + 0.040383439511060715, + 0.12267441302537918, + 0.06935112178325653, + 0.15138106048107147, + 0.07533575594425201, + 0.26692575216293335 + ], + "credit_map": [ + 0.006632720120251179, + 0.005844377912580967, + 0.012557129375636578, + 0.012384840287268162, + 0.009438464418053627, + 0.00899121630936861, + 0.007877242751419544, + 0.011561939492821693, + 0.006904136389493942, + 0.004496619571000338, + 0.005233597941696644, + 0.003544025355949998, + 0.004432873800396919, + 0.00997067429125309, + 0.004541297908872366, + 0.00496945483610034, + 0.005763479042798281, + 0.0070748780854046345, + 0.01338122133165598, + 0.008000548928976059, + 0.009787425398826599, + 0.006214786786586046, + 0.007840531878173351, + 0.007433726917952299, + 0.011239161714911461, + 0.017968768253922462, + 0.005473004654049873, + 0.006601699162274599, + 0.01905096136033535, + 0.011730161495506763, + 0.010404553264379501, + 0.006602956913411617, + 0.004336158744990826, + 0.004954448901116848, + 0.003133527934551239, + 0.004188384395092726, + 0.006340731400996447, + 0.005969157908111811, + 0.006955826189368963, + 0.004505201242864132, + 0.016579417511820793, + 0.014426946640014648, + 0.011737827211618423, + 0.009345694445073605, + 0.018396882340312004, + 0.027849800884723663, + 0.01155044510960579, + 0.012787400744855404, + 0.008847414515912533, + 0.011478850618004799, + 0.00520664406940341, + 0.005483537446707487, + 0.013272333890199661, + 0.012629334814846516, + 0.006317371968179941, + 0.0061156307347118855, + 0.016737738624215126, + 0.03596769645810127, + 0.008976706303656101, + 0.008924274705350399, + 0.020795511081814766, + 0.02998882159590721, + 0.01594228297472, + 0.014048021286725998, + 0.006468608975410461, + 0.017056409269571304, + 0.00687710614874959, + 0.0053862836211919785, + 0.010002636350691319, + 0.014679880812764168, + 0.006207651924341917, + 0.008657184429466724, + 0.024635672569274902, + 0.05094398185610771, + 0.022846447303891182, + 0.016463736072182655, + 0.04496053606271744, + 0.05211976543068886, + 0.018779147416353226, + 0.036176443099975586 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 5 + ], + "selected_reward": -0.016068601980805397 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a young woman with long dark brown hair wearing a white cotton dress?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman holding a light translucent scarf that slips from her grasp and is lifted by the wind?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scarf appear vibrant red and swirl in the air against a pale blue sky with pink-tinged clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera smoothly tilt upward and slowly zoom out to capture the scarf rising?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the woman look up with an expression of surprise and a hint of melancholy as the scarf disappears?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369694948196411 + ], + "opposite_logprob_tokens": [ + -3.4086968898773193 + ], + "logprob_reward": -0.03369694948196411, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369694948196411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26996713876724243, + 0.10940831154584885, + 0.021248487755656242, + 0.06702909618616104, + 0.0612836591899395, + 0.04324556887149811, + 0.04391765594482422, + 0.10870958119630814, + 0.1418953388929367, + 0.1332951784133911 + ], + "credit_map": [ + 0.030406875535845757, + 0.01985578052699566, + 0.08141953498125076, + 0.02176074869930744, + 0.014313395135104656, + 0.02506619319319725, + 0.010150921531021595, + 0.06699369102716446, + 0.03176569938659668, + 0.006325280759483576, + 0.008082076907157898, + 0.003809125628322363, + 0.005950292572379112, + 0.039948172867298126, + 0.004676453769207001, + 0.008851207792758942, + 0.0024215674493461847, + 0.0023154625669121742, + 0.0036410244647413492, + 0.0021722521632909775, + 0.002540928078815341, + 0.002828132826834917, + 0.0026164203882217407, + 0.002712699817493558, + 0.009982054121792316, + 0.009582915343344212, + 0.006582288071513176, + 0.005179139785468578, + 0.011404643766582012, + 0.009269686415791512, + 0.009689323604106903, + 0.00533903855830431, + 0.005389263387769461, + 0.0073473528027534485, + 0.005288636777549982, + 0.008467449806630611, + 0.010498682036995888, + 0.007597492076456547, + 0.011127721518278122, + 0.005567059852182865, + 0.006529578473418951, + 0.005606657359749079, + 0.004121995531022549, + 0.004138158168643713, + 0.006570183672010899, + 0.006598589010536671, + 0.00463400362059474, + 0.005046403966844082, + 0.0037882570177316666, + 0.004517558962106705, + 0.0038665621541440487, + 0.0045068711042404175, + 0.006038385909050703, + 0.005582012236118317, + 0.00989268347620964, + 0.00572532182559371, + 0.009202797897160053, + 0.023453189060091972, + 0.009389009326696396, + 0.007515823468565941, + 0.015681492164731026, + 0.022992843762040138, + 0.009588128887116909, + 0.010886291973292828, + 0.010516435839235783, + 0.022497663274407387, + 0.017574353143572807, + 0.011080753989517689, + 0.020688079297542572, + 0.025587737560272217, + 0.015104107558727264, + 0.01884619891643524, + 0.010845956392586231, + 0.019288482144474983, + 0.008362740278244019, + 0.008543530479073524, + 0.03214189410209656, + 0.02874203585088253, + 0.009114811196923256, + 0.016255725175142288 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 9 + ], + "selected_reward": -0.03369694948196411 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_tensor([9714, 9714])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_tensor([9714, 9714])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..48329cd1dd6e23fbfd18722f4c52823757e0e9b6 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_tensor([9714, 9714])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "supremely luxurious close-up, captured in an uninterrupted, extremely slow horizontal tracking shot, smoothly sweeps from left to right across a custom-made M4A1 rifle. The rifle rests silently on a deep, light-absorbing black velvet display stand, its background completely blurred, creating a sense of luxury suspended in space. The camera begins at the end of the stock, where soft, diffused studio light from a single direction highlights the sturdy contours of the tactical stock and the delicate texture of its composite material. As the camera steadily and deliberately pans to the right, light flows across the cool, matte gunmetal-gray metal receiver, revealing the intricately carved rose relief pattern. These reliefs are not simple etchings, but three-dimensional flowers and vines, seemingly growing from the metal. Each delicate petal and thorny stem is meticulously polished, contrasting sharply with the rough, matte texture of the main body of the rifle, reflecting the light with a fluid, mercury-like luster. The camera continued to slide, capturing the precise mechanical structure of the magazine well and trigger guard, and finally slowly stopped on the handguard and muzzle brake covered with a Picatinny rail. The entire process was without a trace of shaking, perfectly blending the weapon's lethal functionality with its unparalleled artistic beauty, presenting a breathtaking atmosphere of tranquility and luxury unique to top luxury goods advertisements.", + "seed": 13251, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_9714_seed_13251_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.19994422178715468, + "vlm_logprob": -0.19994422178715468, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.19994422178715468, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04216881841421127 + ], + "opposite_logprob_tokens": [ + -3.2921688556671143 + ], + "logprob_reward": -0.04216881841421127, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.04216881841421127, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16108763217926025, + 0.05924613028764725, + 0.11346224695444107, + 0.07900366187095642, + 0.17216435074806213, + 0.05663298815488815, + 0.09532836079597473, + 0.13355126976966858, + 0.0685066431760788, + 0.06101668253540993 + ], + "credit_map": [ + 0.018071841448545456, + 0.012643243186175823, + 0.031002236530184746, + 0.03797118738293648, + 0.010346967726945877, + 0.014065993949770927, + 0.011282560415565968, + 0.025703610852360725, + 0.00542199844494462, + 0.009148376062512398, + 0.0059156278148293495, + 0.006211023312062025, + 0.01002113614231348, + 0.008045791648328304, + 0.006145156919956207, + 0.008337019011378288, + 0.009309264831244946, + 0.00795687735080719, + 0.011601806618273258, + 0.018089327961206436, + 0.011022558435797691, + 0.010927952826023102, + 0.027916500344872475, + 0.016637951135635376, + 0.005382932722568512, + 0.003369176760315895, + 0.02651127614080906, + 0.01449801679700613, + 0.0026834530290216208, + 0.005539577454328537, + 0.011178449727594852, + 0.009840771555900574, + 0.010097641497850418, + 0.02446344494819641, + 0.014895513653755188, + 0.033010225743055344, + 0.015564355999231339, + 0.017665386199951172, + 0.04290368780493736, + 0.0135640949010849, + 0.0036727648694068193, + 0.005946055520325899, + 0.008636390790343285, + 0.006853998638689518, + 0.007972832769155502, + 0.0060718185268342495, + 0.008784382604062557, + 0.008694740943610668, + 0.010695883072912693, + 0.017739295959472656, + 0.013340577483177185, + 0.01255251094698906, + 0.010156404227018356, + 0.012607424519956112, + 0.00998778734356165, + 0.008248475380241871, + 0.018610578030347824, + 0.01812591589987278, + 0.017138557508587837, + 0.012212476693093777, + 0.015912910923361778, + 0.018457960337400436, + 0.0183247197419405, + 0.014768141321837902, + 0.009216097183525562, + 0.005142004694789648, + 0.010128219611942768, + 0.009268377907574177, + 0.007281039375811815, + 0.005769561976194382, + 0.011592045426368713, + 0.010109296068549156, + 0.00820433720946312, + 0.007493665441870689, + 0.007896349765360355, + 0.008891412988305092, + 0.009227468632161617, + 0.009012185968458652, + 0.004332636948674917, + 0.005958631169050932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 7 + ], + "selected_reward": -0.04216881841421127 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904284238815308 + ], + "opposite_logprob_tokens": [ + -2.464042901992798 + ], + "logprob_reward": -0.08904284238815308, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904284238815308, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03829888999462128, + 0.039703819900751114, + 0.04070892930030823, + 0.2982611060142517, + 0.3206079602241516, + 0.05052250623703003, + 0.035949185490608215, + 0.09051169455051422, + 0.03802894428372383, + 0.04740700125694275 + ], + "credit_map": [ + 0.005598419811576605, + 0.0036825856659561396, + 0.006990001071244478, + 0.006446506828069687, + 0.0030354387126863003, + 0.0030925257597118616, + 0.0031423154287040234, + 0.0063110943883657455, + 0.003464900655671954, + 0.005699525121599436, + 0.0045556905679404736, + 0.0064775883220136166, + 0.004825974814593792, + 0.004026512149721384, + 0.005426385905593634, + 0.005227241199463606, + 0.0030017339158803225, + 0.0036809525918215513, + 0.0042622992768883705, + 0.008557788096368313, + 0.003124697133898735, + 0.0036263691727072, + 0.0088895158842206, + 0.005565566476434469, + 0.019528904929757118, + 0.010375739075243473, + 0.10600722581148148, + 0.05404020845890045, + 0.010318627581000328, + 0.017466099932789803, + 0.0434546023607254, + 0.03706967085599899, + 0.01640697754919529, + 0.021869752556085587, + 0.059272173792123795, + 0.048370346426963806, + 0.020235786214470863, + 0.026555847376585007, + 0.0998527929186821, + 0.02804424986243248, + 0.0034242321271449327, + 0.0062441397458314896, + 0.008186948485672474, + 0.010377222672104836, + 0.003630400635302067, + 0.003370437305420637, + 0.00589392613619566, + 0.009395196102559566, + 0.003815066535025835, + 0.007936730980873108, + 0.0045921532437205315, + 0.004252437502145767, + 0.003918624483048916, + 0.0035716884303838015, + 0.004084807820618153, + 0.0037776753306388855, + 0.015150798484683037, + 0.015289772301912308, + 0.008974052034318447, + 0.010574761778116226, + 0.011311091482639313, + 0.0101345619186759, + 0.009948031976819038, + 0.00912861805409193, + 0.005683558993041515, + 0.003621838754042983, + 0.005805734544992447, + 0.004970748443156481, + 0.003897374961525202, + 0.0036506326869130135, + 0.006585706956684589, + 0.003813347080722451, + 0.005142105743288994, + 0.005150701850652695, + 0.010210013948380947, + 0.007836331613361835, + 0.003893411485478282, + 0.004189069848507643, + 0.005025459453463554, + 0.005959902890026569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 7 + ], + "selected_reward": -0.08904284238815308 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033716194331645966 + ], + "opposite_logprob_tokens": [ + -3.4087162017822266 + ], + "logprob_reward": -0.033716194331645966, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033716194331645966, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13093696534633636, + 0.06630850583314896, + 0.09049029648303986, + 0.087678924202919, + 0.17825382947921753, + 0.08646087348461151, + 0.0769771859049797, + 0.18074285984039307, + 0.059633541852235794, + 0.04251701012253761 + ], + "credit_map": [ + 0.018806014209985733, + 0.012036431580781937, + 0.015292763710021973, + 0.016476992517709732, + 0.012383694760501385, + 0.014133535325527191, + 0.01698824018239975, + 0.02481929399073124, + 0.005207703448832035, + 0.010628415271639824, + 0.007025988306850195, + 0.009659135714173317, + 0.00948304496705532, + 0.006097864825278521, + 0.009505894035100937, + 0.008700459264218807, + 0.006929339841008186, + 0.004959031939506531, + 0.006750328931957483, + 0.01755107007920742, + 0.00530634680762887, + 0.005842479877173901, + 0.027925875037908554, + 0.015225815586745739, + 0.005952257663011551, + 0.00395312812179327, + 0.02390185184776783, + 0.01353486068546772, + 0.005484034772962332, + 0.006394767668098211, + 0.015250985510647297, + 0.013207041658461094, + 0.011309300549328327, + 0.035753387957811356, + 0.03854271024465561, + 0.023804064840078354, + 0.013699056580662727, + 0.017847837880253792, + 0.02463638409972191, + 0.012661091051995754, + 0.006390547379851341, + 0.00772616034373641, + 0.014241229742765427, + 0.013125184923410416, + 0.011488072574138641, + 0.006342846434563398, + 0.013342875055968761, + 0.01380396168678999, + 0.008680245839059353, + 0.013326548971235752, + 0.008303174749016762, + 0.011832436546683311, + 0.007310062646865845, + 0.011723832227289677, + 0.008753640577197075, + 0.007047242019325495, + 0.026470275595784187, + 0.02955552004277706, + 0.019646285101771355, + 0.01673237234354019, + 0.02348499745130539, + 0.023890461772680283, + 0.0251007117331028, + 0.015862228348851204, + 0.009419831447303295, + 0.006490724161267281, + 0.008321941830217838, + 0.007303995080292225, + 0.00632029352709651, + 0.0058312746696174145, + 0.009836211800575256, + 0.006109269335865974, + 0.004229709971696138, + 0.00575789762660861, + 0.008276894688606262, + 0.005895450711250305, + 0.004579776898026466, + 0.005370877683162689, + 0.004406818188726902, + 0.00399958249181509 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 4, + 0 + ], + "selected_reward": -0.033716194331645966 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008634849451482296 + ], + "opposite_logprob_tokens": [ + -4.7586350440979 + ], + "logprob_reward": -0.008634849451482296, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008634849451482296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13106879591941833, + 0.054259274154901505, + 0.047828853130340576, + 0.09632202237844467, + 0.19960926473140717, + 0.12413240224123001, + 0.06982961297035217, + 0.14514225721359253, + 0.07023542374372482, + 0.0615721121430397 + ], + "credit_map": [ + 0.01489837933331728, + 0.007741705980151892, + 0.031138328835368156, + 0.029816292226314545, + 0.00943696592003107, + 0.011446096934378147, + 0.009578729048371315, + 0.01701228693127632, + 0.0036520713474601507, + 0.007103916723281145, + 0.006391747388988733, + 0.010408363305032253, + 0.004612227901816368, + 0.005937471985816956, + 0.00922886561602354, + 0.0069246115162968636, + 0.005791023373603821, + 0.0030208276584744453, + 0.00489009078592062, + 0.008322327397763729, + 0.0033589773811399937, + 0.0038443622179329395, + 0.010783255100250244, + 0.007817992940545082, + 0.00673658074811101, + 0.004031830467283726, + 0.026117948815226555, + 0.016033535823225975, + 0.004461696371436119, + 0.006331318523734808, + 0.017626842483878136, + 0.014982270076870918, + 0.010231728665530682, + 0.006597853731364012, + 0.046702317893505096, + 0.027413420379161835, + 0.013810304924845695, + 0.009414305910468102, + 0.06664738804101944, + 0.0187919270247221, + 0.00783237162977457, + 0.012520250864326954, + 0.020215759053826332, + 0.022583061829209328, + 0.010090327821671963, + 0.010720377787947655, + 0.01633404940366745, + 0.023836197331547737, + 0.006826311815530062, + 0.008219175972044468, + 0.012916859239339828, + 0.009110487066209316, + 0.006549217738211155, + 0.014748000539839268, + 0.005323965102434158, + 0.006135599222034216, + 0.020427554845809937, + 0.02090919204056263, + 0.018982471898198128, + 0.013320405967533588, + 0.01784403994679451, + 0.02124597132205963, + 0.019029734656214714, + 0.01338288001716137, + 0.011203687638044357, + 0.008377976715564728, + 0.011589250527322292, + 0.009243502281606197, + 0.005828352645039558, + 0.007185162045061588, + 0.010123142972588539, + 0.006684354972094297, + 0.004789290949702263, + 0.0066450429148972034, + 0.016423635184764862, + 0.010981655679643154, + 0.004858214873820543, + 0.005044801626354456, + 0.005949107930064201, + 0.0068803587928414345 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.008634849451482296 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261584043502808 + ], + "opposite_logprob_tokens": [ + -0.5761584043502808 + ], + "logprob_reward": -0.8261584043502808, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261584043502808, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.06364821642637253, + 0.0638434886932373, + 0.0657440721988678, + 0.0499391183257103, + 0.37762585282325745, + 0.09841031581163406, + 0.028290268033742905, + 0.10265395045280457, + 0.03394195809960365, + 0.11590272933244705 + ], + "credit_map": [ + 0.011162271723151207, + 0.006530471611768007, + 0.009837848134338856, + 0.008442075923085213, + 0.005786091089248657, + 0.005372297018766403, + 0.006155063863843679, + 0.01036209985613823, + 0.005169392097741365, + 0.010309536010026932, + 0.007488822564482689, + 0.006683731451630592, + 0.01069220807403326, + 0.007572895381599665, + 0.006064646877348423, + 0.009862257167696953, + 0.005965807009488344, + 0.004439117386937141, + 0.005468093790113926, + 0.00852438434958458, + 0.005452933255583048, + 0.005529389251023531, + 0.019975973293185234, + 0.010388372465968132, + 0.003651364240795374, + 0.0037682047113776207, + 0.013364717364311218, + 0.009841104038059711, + 0.003433263162150979, + 0.005568667780607939, + 0.006322619970887899, + 0.0039891754277050495, + 0.025798125192523003, + 0.04327438771724701, + 0.1328049600124359, + 0.0748668983578682, + 0.01894272491335869, + 0.014182954095304012, + 0.04504816234111786, + 0.022707652300596237, + 0.007561491336673498, + 0.011532722041010857, + 0.020313335582613945, + 0.017138991504907608, + 0.009357087314128876, + 0.006846422795206308, + 0.01181199960410595, + 0.013848261907696724, + 0.004225484095513821, + 0.0040288264863193035, + 0.004140045959502459, + 0.004598446190357208, + 0.002293726895004511, + 0.0019453611457720399, + 0.0034991041757166386, + 0.003559274133294821, + 0.015195549465715885, + 0.01630769670009613, + 0.012756161391735077, + 0.009801658801734447, + 0.01163201592862606, + 0.013809511438012123, + 0.01378381997346878, + 0.009367537684738636, + 0.0052940016612410545, + 0.004769936669617891, + 0.00467115780338645, + 0.004402116406708956, + 0.0030966447666287422, + 0.0036841160617768764, + 0.0046163215301930904, + 0.003407664131373167, + 0.013096263632178307, + 0.011744444258511066, + 0.01702871546149254, + 0.016989506781101227, + 0.012729533948004246, + 0.013410478830337524, + 0.010692119598388672, + 0.02021166868507862 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 7 + ], + "selected_reward": -0.8261584043502808 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "supremely luxurious close-up, captured in an uninterrupted, extremely slow horizontal tracking shot, smoothly sweeps from left to right across a custom-made M4A1 rifle. The rifle rests silently on a deep, light-absorbing black velvet display stand, its background completely blurred, creating a sense of luxury suspended in space. The camera begins at the end of the stock, where soft, diffused studio light from a single direction highlights the sturdy contours of the tactical stock and the delicate texture of its composite material. As the camera steadily and deliberately pans to the right, light flows across the cool, matte gunmetal-gray metal receiver, revealing the intricately carved rose relief pattern. These reliefs are not simple etchings, but three-dimensional flowers and vines, seemingly growing from the metal. Each delicate petal and thorny stem is meticulously polished, contrasting sharply with the rough, matte texture of the main body of the rifle, reflecting the light with a fluid, mercury-like luster. The camera continued to slide, capturing the precise mechanical structure of the magazine well and trigger guard, and finally slowly stopped on the handguard and muzzle brake covered with a Picatinny rail. The entire process was without a trace of shaking, perfectly blending the weapon's lethal functionality with its unparalleled artistic beauty, presenting a breathtaking atmosphere of tranquility and luxury unique to top luxury goods advertisements.", + "seed": 13252, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_9714_seed_13252_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.22214613296091557, + "vlm_logprob": -0.22214613296091557, + "vlm_yes_no_margin": 2.35, + "vlm_token_credit": -0.22214613296091557, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.053880661725997925 + ], + "opposite_logprob_tokens": [ + -3.0538806915283203 + ], + "logprob_reward": -0.053880661725997925, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.053880661725997925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11862386018037796, + 0.6245698928833008, + 0.06352662295103073, + 0.032276153564453125, + 0.015611435286700726, + 0.04325287789106369, + 0.0180796068161726, + 0.020125502720475197, + 0.043584488332271576, + 0.020349552854895592 + ], + "credit_map": [ + 0.013813890516757965, + 0.019177498295903206, + 0.017162827774882317, + 0.010594199411571026, + 0.015058533288538456, + 0.01673467829823494, + 0.010340469889342785, + 0.015741746872663498, + 0.14018698036670685, + 0.2533475458621979, + 0.13535794615745544, + 0.01888081431388855, + 0.02819765731692314, + 0.015244879759848118, + 0.012816398404538631, + 0.020537633448839188, + 0.009839900769293308, + 0.010989426635205746, + 0.005535879638046026, + 0.004585891030728817, + 0.006192822940647602, + 0.014437435194849968, + 0.004634125158190727, + 0.007311136927455664, + 0.005899395328015089, + 0.0027877995744347572, + 0.0024368390440940857, + 0.0017960075056180358, + 0.005304369609802961, + 0.008233345113694668, + 0.0023553527425974607, + 0.0034630426671355963, + 0.001856258139014244, + 0.001955366460606456, + 0.0009400594281032681, + 0.001156761427409947, + 0.003777876729145646, + 0.0026100839022547007, + 0.0017364860977977514, + 0.0015785415889695287, + 0.006983091589063406, + 0.008485907688736916, + 0.003337249858304858, + 0.003975500352680683, + 0.004383525811135769, + 0.007182777859270573, + 0.004313760902732611, + 0.004591057542711496, + 0.0019648501183837652, + 0.001763795386068523, + 0.0018407735042273998, + 0.003686025273054838, + 0.002597292186692357, + 0.0021736319176852703, + 0.001848386018536985, + 0.0022048514802008867, + 0.002627279609441757, + 0.0025859938468784094, + 0.002755966270342469, + 0.0019040630431845784, + 0.0021968306973576546, + 0.0038814980071038008, + 0.0015756618231534958, + 0.0025982067454606295, + 0.008145560510456562, + 0.0029156003147363663, + 0.008117246441543102, + 0.003331097774207592, + 0.004502449184656143, + 0.006747981533408165, + 0.00460014957934618, + 0.005224398337304592, + 0.0022839203011244535, + 0.001540247816592455, + 0.0019452451961115003, + 0.0014622894814237952, + 0.003140414832159877, + 0.003940836526453495, + 0.0024737385101616383, + 0.003562858561053872 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.053880661725997925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905625343322754 + ], + "opposite_logprob_tokens": [ + -2.4640562534332275 + ], + "logprob_reward": -0.08905625343322754, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905625343322754, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12415628880262375, + 0.5527080297470093, + 0.04188361391425133, + 0.04061571881175041, + 0.024066541343927383, + 0.05199335142970085, + 0.05502799153327942, + 0.02207518368959427, + 0.027764366939663887, + 0.05970892310142517 + ], + "credit_map": [ + 0.014231480658054352, + 0.019608190283179283, + 0.014197512529790401, + 0.008448068983852863, + 0.023051682859659195, + 0.012354396283626556, + 0.017516138032078743, + 0.014748810790479183, + 0.12253539264202118, + 0.22218091785907745, + 0.12210093438625336, + 0.016595469787716866, + 0.02580106072127819, + 0.01219797320663929, + 0.012651562690734863, + 0.0186446700245142, + 0.006978228222578764, + 0.007619664538651705, + 0.0036595771089196205, + 0.003317582653835416, + 0.004199695307761431, + 0.00905105285346508, + 0.002712232293561101, + 0.004345576278865337, + 0.008477076888084412, + 0.0030256153549999, + 0.002596034435555339, + 0.002176553476601839, + 0.006878715939819813, + 0.012121954932808876, + 0.0026345127262175083, + 0.0027052517980337143, + 0.004259433131664991, + 0.002892154734581709, + 0.0026949369348585606, + 0.002387854503467679, + 0.0031922657508403063, + 0.0034351195208728313, + 0.0024389552418142557, + 0.0027658191975206137, + 0.007636058144271374, + 0.008040905930101871, + 0.004928611684590578, + 0.0052926205098629, + 0.005733304191380739, + 0.007490817457437515, + 0.007193736266344786, + 0.0056772916577756405, + 0.006615812424570322, + 0.005647596437484026, + 0.006661938037723303, + 0.009011086076498032, + 0.007117524277418852, + 0.0065557630732655525, + 0.008352049626410007, + 0.00506622064858675, + 0.0029373494908213615, + 0.002805369673296809, + 0.003838075790554285, + 0.001641237991861999, + 0.0019844910129904747, + 0.003137317020446062, + 0.0019386891508474946, + 0.0037926535587757826, + 0.0037423644680529833, + 0.0022389988880604506, + 0.0035000855568796396, + 0.0022285678423941135, + 0.0031536107417196035, + 0.005439152475446463, + 0.003566646482795477, + 0.0038949388545006514, + 0.007033936679363251, + 0.003020885633304715, + 0.0051185255870223045, + 0.0033479847479611635, + 0.007305737119168043, + 0.008888929150998592, + 0.005618259310722351, + 0.019374659284949303 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.08905625343322754 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309125989675522 + ], + "opposite_logprob_tokens": [ + -3.168091297149658 + ], + "logprob_reward": -0.04309125989675522, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309125989675522, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1711585372686386, + 0.36306601762771606, + 0.1650734692811966, + 0.028438150882720947, + 0.024156570434570312, + 0.026695754379034042, + 0.0480155386030674, + 0.025349538773298264, + 0.11588982492685318, + 0.0321565642952919 + ], + "credit_map": [ + 0.021094897761940956, + 0.03548988327383995, + 0.016360443085432053, + 0.019025156274437904, + 0.02847272716462612, + 0.02285928837954998, + 0.013774830847978592, + 0.014081338420510292, + 0.07913174480199814, + 0.14178399741649628, + 0.0778045728802681, + 0.011891162022948265, + 0.018339117988944054, + 0.011024583131074905, + 0.009980355389416218, + 0.013110514730215073, + 0.02689618617296219, + 0.027737069875001907, + 0.014179497957229614, + 0.010053166188299656, + 0.016368528828024864, + 0.04065496101975441, + 0.01091409008949995, + 0.018269985914230347, + 0.0045264773070812225, + 0.002171649830415845, + 0.0025639808736741543, + 0.0023506488651037216, + 0.005064339842647314, + 0.005429910961538553, + 0.001972104888409376, + 0.004359045065939426, + 0.0025110673159360886, + 0.00276958174072206, + 0.003074208740144968, + 0.002843733411282301, + 0.004260658752173185, + 0.002713379915803671, + 0.003614060813561082, + 0.0023698832374066114, + 0.0035138402599841356, + 0.004832476377487183, + 0.002451363718137145, + 0.003107914701104164, + 0.0022980626672506332, + 0.0041631776839494705, + 0.0035562997218221426, + 0.002772624371573329, + 0.005534391384571791, + 0.006547908298671246, + 0.005333212204277515, + 0.00832702498883009, + 0.006318250671029091, + 0.00656288955360651, + 0.004409511107951403, + 0.004982354585081339, + 0.0022112594451755285, + 0.002353479852899909, + 0.0032372146379202604, + 0.001795818330720067, + 0.0022755430545657873, + 0.004142759833484888, + 0.0037524900399148464, + 0.005580975208431482, + 0.018697092309594154, + 0.009404252283275127, + 0.01588493399322033, + 0.0072603169828653336, + 0.014266811311244965, + 0.021013563498854637, + 0.013340217992663383, + 0.01602264866232872, + 0.0026318884920328856, + 0.0021640919148921967, + 0.0036747532431036234, + 0.003897673450410366, + 0.003702321322634816, + 0.0057395100593566895, + 0.0028988022822886705, + 0.007447525858879089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04309125989675522 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026342594996094704 + ], + "opposite_logprob_tokens": [ + -3.6513426303863525 + ], + "logprob_reward": -0.026342594996094704, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026342594996094704, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30449312925338745, + 0.26420992612838745, + 0.04586312919855118, + 0.021710282191634178, + 0.04067963734269142, + 0.03574398159980774, + 0.06473418325185776, + 0.05140204355120659, + 0.07020121067762375, + 0.10096248239278793 + ], + "credit_map": [ + 0.03699733316898346, + 0.0725877657532692, + 0.036056920886039734, + 0.023756779730319977, + 0.05332276597619057, + 0.030393272638320923, + 0.022557532414793968, + 0.02882075309753418, + 0.05392984673380852, + 0.09513065963983536, + 0.059235990047454834, + 0.010831675492227077, + 0.01322958618402481, + 0.00863519124686718, + 0.0114125432446599, + 0.01180441677570343, + 0.006534479558467865, + 0.007683948613703251, + 0.004651524592190981, + 0.0036145998165011406, + 0.005443492438644171, + 0.007574259303510189, + 0.0037973811849951744, + 0.006563444621860981, + 0.003150671487674117, + 0.0020025442354381084, + 0.0025853291153907776, + 0.002354476135224104, + 0.002922399900853634, + 0.003915958106517792, + 0.002118470147252083, + 0.002660432830452919, + 0.006695819552987814, + 0.004576605744659901, + 0.0034076462034136057, + 0.004795097280293703, + 0.006351710297167301, + 0.005382244940847158, + 0.005117200314998627, + 0.0043533118441700935, + 0.004856229294091463, + 0.006540318485349417, + 0.003030933439731598, + 0.0035575157962739468, + 0.003419990185648203, + 0.005526542663574219, + 0.004181133583188057, + 0.004631320014595985, + 0.007436719257384539, + 0.00569965410977602, + 0.005993364844471216, + 0.014511526562273502, + 0.008891932666301727, + 0.008138895966112614, + 0.007206156849861145, + 0.006855933461338282, + 0.005236040335148573, + 0.004877765662968159, + 0.007195161655545235, + 0.003423321759328246, + 0.005654297303408384, + 0.009663291275501251, + 0.004413902293890715, + 0.010938264429569244, + 0.0067834812216460705, + 0.003188551403582096, + 0.01900440640747547, + 0.003989115357398987, + 0.006325196009129286, + 0.010810598731040955, + 0.013688330538570881, + 0.00641153147444129, + 0.009626451879739761, + 0.005049312021583319, + 0.010023902170360088, + 0.005787834525108337, + 0.014533023349940777, + 0.014935673214495182, + 0.008897220715880394, + 0.03210906684398651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.026342594996094704 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983598947525024 + ], + "opposite_logprob_tokens": [ + -0.5233598947525024 + ], + "logprob_reward": -0.8983598947525024, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983598947525024, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.10746829956769943, + 0.5791727900505066, + 0.0795748308300972, + 0.018951771780848503, + 0.03729274123907089, + 0.04373989999294281, + 0.03290359675884247, + 0.053236450999975204, + 0.01883639767765999, + 0.02882319502532482 + ], + "credit_map": [ + 0.013370099477469921, + 0.01686200499534607, + 0.013487241230905056, + 0.02330402098596096, + 0.012087984941899776, + 0.009517213329672813, + 0.008195576258003712, + 0.010644158348441124, + 0.13856105506420135, + 0.2329229712486267, + 0.11027845740318298, + 0.018210452049970627, + 0.028839383274316788, + 0.0177856907248497, + 0.0143482256680727, + 0.018226584419608116, + 0.012895685620605946, + 0.012046635150909424, + 0.009089604951441288, + 0.005018786061555147, + 0.007116758264601231, + 0.01869761385023594, + 0.00568203441798687, + 0.009027710184454918, + 0.002743764081969857, + 0.002531795296818018, + 0.00224379263818264, + 0.0016018288442865014, + 0.0023143773432821035, + 0.003390387399122119, + 0.0017788640689104795, + 0.002346963854506612, + 0.0049164025112986565, + 0.004393671173602343, + 0.0032127033919095993, + 0.003981124144047499, + 0.0062253959476947784, + 0.005231176503002644, + 0.004977324046194553, + 0.004354944918304682, + 0.006396449636667967, + 0.009713863953948021, + 0.0034248684532940388, + 0.0035595616791397333, + 0.00559202441945672, + 0.0072822049260139465, + 0.0037217456847429276, + 0.004049176815897226, + 0.0048143998719751835, + 0.003281728597357869, + 0.004473251756280661, + 0.004403506405651569, + 0.0043853940442204475, + 0.0050421953201293945, + 0.0025802706368267536, + 0.00392284756526351, + 0.004175309557467699, + 0.004447360523045063, + 0.007571007125079632, + 0.0034656249918043613, + 0.006221248302608728, + 0.009940313175320625, + 0.005154260899871588, + 0.012261328287422657, + 0.0030010140035301447, + 0.001478091231547296, + 0.004451364278793335, + 0.0012801078846678138, + 0.002164343139156699, + 0.0018796217627823353, + 0.0026257128920406103, + 0.001956143183633685, + 0.002875340636819601, + 0.0018264612881466746, + 0.0024217376485466957, + 0.0024393070489168167, + 0.0034635071642696857, + 0.004159413743764162, + 0.0029638076666742563, + 0.008673619478940964 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.8983598947525024 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "No" + ] + }, + { + "prompt": "supremely luxurious close-up, captured in an uninterrupted, extremely slow horizontal tracking shot, smoothly sweeps from left to right across a custom-made M4A1 rifle. The rifle rests silently on a deep, light-absorbing black velvet display stand, its background completely blurred, creating a sense of luxury suspended in space. The camera begins at the end of the stock, where soft, diffused studio light from a single direction highlights the sturdy contours of the tactical stock and the delicate texture of its composite material. As the camera steadily and deliberately pans to the right, light flows across the cool, matte gunmetal-gray metal receiver, revealing the intricately carved rose relief pattern. These reliefs are not simple etchings, but three-dimensional flowers and vines, seemingly growing from the metal. Each delicate petal and thorny stem is meticulously polished, contrasting sharply with the rough, matte texture of the main body of the rifle, reflecting the light with a fluid, mercury-like luster. The camera continued to slide, capturing the precise mechanical structure of the magazine well and trigger guard, and finally slowly stopped on the handguard and muzzle brake covered with a Picatinny rail. The entire process was without a trace of shaking, perfectly blending the weapon's lethal functionality with its unparalleled artistic beauty, presenting a breathtaking atmosphere of tranquility and luxury unique to top luxury goods advertisements.", + "seed": 13253, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_9714_seed_13253_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15043936297297478, + "vlm_logprob": -0.15043936297297478, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.15043936297297478, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03897259756922722 + ], + "opposite_logprob_tokens": [ + -3.4139726161956787 + ], + "logprob_reward": -0.03897259756922722, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03897259756922722, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12631790339946747, + 0.07342240959405899, + 0.11261969059705734, + 0.03937903046607971, + 0.05413287505507469, + 0.09696274995803833, + 0.04828114062547684, + 0.2413298785686493, + 0.12376131117343903, + 0.08379299938678741 + ], + "credit_map": [ + 0.015987323597073555, + 0.012944350019097328, + 0.0169698316603899, + 0.025288529694080353, + 0.01236923411488533, + 0.011137346737086773, + 0.01909993588924408, + 0.012521353550255299, + 0.008854558691382408, + 0.005988364107906818, + 0.012909776531159878, + 0.009365496225655079, + 0.011514846235513687, + 0.006852064281702042, + 0.008075647056102753, + 0.0098616573959589, + 0.009995121508836746, + 0.009837893769145012, + 0.020387083292007446, + 0.011556989513337612, + 0.01222303882241249, + 0.011789456941187382, + 0.023670731112360954, + 0.013159383088350296, + 0.004992860835045576, + 0.0052980282343924046, + 0.006838028784841299, + 0.0054849679581820965, + 0.003958115354180336, + 0.002162882126867771, + 0.0056140003725886345, + 0.005030145402997732, + 0.006682245060801506, + 0.00585642084479332, + 0.007613875437527895, + 0.010204731486737728, + 0.004581028129905462, + 0.004284490365535021, + 0.007066396065056324, + 0.007843685336411, + 0.010422618128359318, + 0.013262459076941013, + 0.013660854659974575, + 0.011210598051548004, + 0.01133966725319624, + 0.008953476324677467, + 0.013662723824381828, + 0.014450355432927608, + 0.004978918470442295, + 0.003450080519542098, + 0.007191000040620565, + 0.008175776340067387, + 0.0035575891379266977, + 0.002854566555470228, + 0.00970448087900877, + 0.00836873147636652, + 0.025454072281718254, + 0.025692706927657127, + 0.020695973187685013, + 0.03719963878393173, + 0.021582413464784622, + 0.028030836954712868, + 0.040410116314888, + 0.04226412996649742, + 0.013547683134675026, + 0.007971901446580887, + 0.025495631620287895, + 0.017285283654928207, + 0.009998518042266369, + 0.009115930646657944, + 0.019894346594810486, + 0.020452013239264488, + 0.0035847965627908707, + 0.0038156069349497557, + 0.015243529342114925, + 0.009452351368963718, + 0.003784853033721447, + 0.016383258625864983, + 0.010817433707416058, + 0.020711174234747887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.03897259756922722 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08905723690986633 + ], + "opposite_logprob_tokens": [ + -2.464057207107544 + ], + "logprob_reward": -0.08905723690986633, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08905723690986633, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11923318356275558, + 0.15427793562412262, + 0.0988336130976677, + 0.02204461768269539, + 0.038217607885599136, + 0.12297184765338898, + 0.11885255575180054, + 0.0838036835193634, + 0.09517563879489899, + 0.14658933877944946 + ], + "credit_map": [ + 0.013282153755426407, + 0.013364855200052261, + 0.025487937033176422, + 0.020142575725913048, + 0.009630738757550716, + 0.009486289694905281, + 0.013575708493590355, + 0.014262931421399117, + 0.01689523458480835, + 0.014177661389112473, + 0.02992602251470089, + 0.01982288435101509, + 0.021374071016907692, + 0.017198028042912483, + 0.015013474971055984, + 0.019870560616254807, + 0.011542296968400478, + 0.00990552082657814, + 0.015774328261613846, + 0.016552485525608063, + 0.008942247368395329, + 0.00917086936533451, + 0.014934207312762737, + 0.012011656537652016, + 0.003572510788217187, + 0.003103908384218812, + 0.0031797396950423717, + 0.0032067291904240847, + 0.002056858269497752, + 0.001406756928190589, + 0.0029593317303806543, + 0.0025587829295545816, + 0.005009736865758896, + 0.004585526417940855, + 0.004872417543083429, + 0.0062790485098958015, + 0.003220107639208436, + 0.004208482336252928, + 0.004884752910584211, + 0.00515753822401166, + 0.014522658661007881, + 0.020403068512678146, + 0.02296091988682747, + 0.017769550904631615, + 0.0108591727912426, + 0.008096740581095219, + 0.012844115495681763, + 0.015515623614192009, + 0.014328688383102417, + 0.009437731467187405, + 0.009175905957818031, + 0.010387456975877285, + 0.005697574932128191, + 0.007333829067647457, + 0.03101266361773014, + 0.031478699296712875, + 0.0095130056142807, + 0.008524708449840546, + 0.006455788854509592, + 0.012823249213397503, + 0.006973639130592346, + 0.0064836828969419, + 0.01592784933745861, + 0.01710176281630993, + 0.007770076859742403, + 0.005549732130020857, + 0.011629277840256691, + 0.008797657676041126, + 0.006520990282297134, + 0.011652735993266106, + 0.020839685574173927, + 0.02241547591984272, + 0.00893342774361372, + 0.0066481041721999645, + 0.016501111909747124, + 0.011356260627508163, + 0.007758003659546375, + 0.043198224157094955, + 0.02822750248014927, + 0.023966701701283455 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 5 + ], + "selected_reward": -0.08905723690986633 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018201468512415886 + ], + "opposite_logprob_tokens": [ + -4.0182013511657715 + ], + "logprob_reward": -0.018201468512415886, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018201468512415886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1295337677001953, + 0.2933848798274994, + 0.07790999859571457, + 0.07307927310466766, + 0.06457477807998657, + 0.06757558137178421, + 0.09606295078992844, + 0.0661904588341713, + 0.06906147301197052, + 0.06262682378292084 + ], + "credit_map": [ + 0.015357804484665394, + 0.015242353081703186, + 0.022297972813248634, + 0.026909450069069862, + 0.008511739782989025, + 0.009803186170756817, + 0.016592826694250107, + 0.01481844112277031, + 0.044249776750802994, + 0.022131284698843956, + 0.036139409989118576, + 0.036680642515420914, + 0.06355521827936172, + 0.026677485555410385, + 0.030451981350779533, + 0.03349908068776131, + 0.005815759301185608, + 0.005575011949986219, + 0.011972145177423954, + 0.01242732536047697, + 0.009384560398757458, + 0.012666376307606697, + 0.010667103342711926, + 0.009401720017194748, + 0.00787074863910675, + 0.008252697996795177, + 0.011310873553156853, + 0.012197074480354786, + 0.006761160213500261, + 0.005676457658410072, + 0.012352489866316319, + 0.008657772094011307, + 0.005912823602557182, + 0.008588096126914024, + 0.007555379532277584, + 0.009970053099095821, + 0.010800057090818882, + 0.0071144248358905315, + 0.006642947439104319, + 0.007990997284650803, + 0.007313994225114584, + 0.0090135233476758, + 0.012993823736906052, + 0.008829987607896328, + 0.007139598950743675, + 0.00418478436768055, + 0.008485649712383747, + 0.009614218957722187, + 0.00904480367898941, + 0.008119478821754456, + 0.010840415954589844, + 0.014224999584257603, + 0.0067861261777579784, + 0.008514496497809887, + 0.019615983590483665, + 0.018916651606559753, + 0.006404704414308071, + 0.005070798564702272, + 0.005236332770437002, + 0.0073053413070738316, + 0.006312425248324871, + 0.007508296985179186, + 0.013037541881203651, + 0.01531501580029726, + 0.007656009402126074, + 0.006048442330211401, + 0.01140078529715538, + 0.009231898933649063, + 0.008187426254153252, + 0.0072835455648601055, + 0.009444245137274265, + 0.009809116832911968, + 0.0029747693333774805, + 0.003947930410504341, + 0.012449551373720169, + 0.007776449900120497, + 0.0035663009621202946, + 0.011782565154135227, + 0.009958085604012012, + 0.010171172209084034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.018201468512415886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029799478128552437 + ], + "opposite_logprob_tokens": [ + -3.529799461364746 + ], + "logprob_reward": -0.029799478128552437, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029799478128552437, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10866305232048035, + 0.10238843411207199, + 0.14192570745944977, + 0.04324943572282791, + 0.0624755322933197, + 0.05484318733215332, + 0.12090300023555756, + 0.2535543441772461, + 0.07167418301105499, + 0.04032308608293533 + ], + "credit_map": [ + 0.01257249154150486, + 0.009455055929720402, + 0.022734377533197403, + 0.020952047780156136, + 0.007859705947339535, + 0.00960745383054018, + 0.012817075476050377, + 0.012664848007261753, + 0.012942655012011528, + 0.00890988577157259, + 0.02745693176984787, + 0.014685841277241707, + 0.010355869308114052, + 0.010722892358899117, + 0.00865508709102869, + 0.00865927804261446, + 0.012442423962056637, + 0.013083857484161854, + 0.01965840347111225, + 0.02137206308543682, + 0.014385509304702282, + 0.0173749178647995, + 0.026363268494606018, + 0.017245257273316383, + 0.004081814084202051, + 0.004604065325111151, + 0.006366712041199207, + 0.00773974135518074, + 0.004041503183543682, + 0.003515019081532955, + 0.00716733280569315, + 0.0057332501746714115, + 0.008044557645916939, + 0.0069888257421553135, + 0.00625061197206378, + 0.009395009838044643, + 0.005353410262614489, + 0.0055181486532092094, + 0.008923604153096676, + 0.01200136449187994, + 0.005269596353173256, + 0.0073784152045845985, + 0.011313912458717823, + 0.006493372842669487, + 0.006346875801682472, + 0.003405696712434292, + 0.0070564281195402145, + 0.007578890770673752, + 0.01395385805517435, + 0.009619133546948433, + 0.01240920927375555, + 0.010141272097826004, + 0.0060757738538086414, + 0.007674087304621935, + 0.030011337250471115, + 0.0310183297842741, + 0.022858968004584312, + 0.022314995527267456, + 0.023524697870016098, + 0.045048318803310394, + 0.019136440008878708, + 0.01653563790023327, + 0.05036362633109093, + 0.05377162992954254, + 0.00546674570068717, + 0.006369089707732201, + 0.005300777032971382, + 0.00656937574967742, + 0.007291873916983604, + 0.009475598111748695, + 0.014909396879374981, + 0.01629132218658924, + 0.0035839981865137815, + 0.002188841812312603, + 0.006517062429338694, + 0.0048677558079361916, + 0.00276137120090425, + 0.00784919410943985, + 0.0057351780124008656, + 0.006819685455411673 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 6 + ], + "selected_reward": -0.029799478128552437 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.576166033744812 + ], + "opposite_logprob_tokens": [ + -0.826166033744812 + ], + "logprob_reward": -0.576166033744812, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.576166033744812, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18816739320755005, + 0.0784575343132019, + 0.10346514731645584, + 0.04966450110077858, + 0.041479550302028656, + 0.04312251880764961, + 0.12303868681192398, + 0.04623447731137276, + 0.06537965685129166, + 0.2609905004501343 + ], + "credit_map": [ + 0.02252832241356373, + 0.028751982375979424, + 0.029838699847459793, + 0.03213828429579735, + 0.015720609575510025, + 0.015982026234269142, + 0.02161506563425064, + 0.02159242145717144, + 0.011239450424909592, + 0.006667362526059151, + 0.018420321866869926, + 0.010965459048748016, + 0.006184931844472885, + 0.009591641835868359, + 0.00720160361379385, + 0.008186767809092999, + 0.007122761569917202, + 0.009918763302266598, + 0.020451616495847702, + 0.010794360190629959, + 0.010222177021205425, + 0.012480043806135654, + 0.019515037536621094, + 0.012960400432348251, + 0.00652112765237689, + 0.006741911172866821, + 0.007049547042697668, + 0.006184653844684362, + 0.004819037392735481, + 0.0029058349318802357, + 0.009471611119806767, + 0.0059707812033593655, + 0.004584553185850382, + 0.004027317278087139, + 0.00502614863216877, + 0.007092119660228491, + 0.0032690083608031273, + 0.004092330113053322, + 0.006511799059808254, + 0.006876278202980757, + 0.00567705649882555, + 0.005534162744879723, + 0.008764018304646015, + 0.006184211000800133, + 0.003321421332657337, + 0.002339480211958289, + 0.005747542716562748, + 0.0055546327494084835, + 0.012170037254691124, + 0.0077846539206802845, + 0.015816347673535347, + 0.016468483954668045, + 0.007342012133449316, + 0.009025286883115768, + 0.0267855953425169, + 0.027646278962492943, + 0.005943618714809418, + 0.004158779978752136, + 0.006189864594489336, + 0.008348935283720493, + 0.0037587014958262444, + 0.0032889803405851126, + 0.007333077490329742, + 0.007212525699287653, + 0.008479422889649868, + 0.005397095810621977, + 0.009867830201983452, + 0.00807814672589302, + 0.008583227172493935, + 0.006796924863010645, + 0.008583825081586838, + 0.009593191556632519, + 0.016062982380390167, + 0.012611192651093006, + 0.051850974559783936, + 0.02170279435813427, + 0.014244678430259228, + 0.046792320907115936, + 0.03378380462527275, + 0.06394176930189133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.576166033744812 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "supremely luxurious close-up, captured in an uninterrupted, extremely slow horizontal tracking shot, smoothly sweeps from left to right across a custom-made M4A1 rifle. The rifle rests silently on a deep, light-absorbing black velvet display stand, its background completely blurred, creating a sense of luxury suspended in space. The camera begins at the end of the stock, where soft, diffused studio light from a single direction highlights the sturdy contours of the tactical stock and the delicate texture of its composite material. As the camera steadily and deliberately pans to the right, light flows across the cool, matte gunmetal-gray metal receiver, revealing the intricately carved rose relief pattern. These reliefs are not simple etchings, but three-dimensional flowers and vines, seemingly growing from the metal. Each delicate petal and thorny stem is meticulously polished, contrasting sharply with the rough, matte texture of the main body of the rifle, reflecting the light with a fluid, mercury-like luster. The camera continued to slide, capturing the precise mechanical structure of the magazine well and trigger guard, and finally slowly stopped on the handguard and muzzle brake covered with a Picatinny rail. The entire process was without a trace of shaking, perfectly blending the weapon's lethal functionality with its unparalleled artistic beauty, presenting a breathtaking atmosphere of tranquility and luxury unique to top luxury goods advertisements.", + "seed": 13254, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_15/index_9714_seed_13254_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.1725240670144558, + "vlm_logprob": -0.1725240670144558, + "vlm_yes_no_margin": 2.375, + "vlm_token_credit": -0.1725240670144558, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05271732434630394 + ], + "opposite_logprob_tokens": [ + -3.0527172088623047 + ], + "logprob_reward": -0.05271732434630394, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05271732434630394, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5598180294036865, + 0.04546678066253662, + 0.021728310734033585, + 0.0243645329028368, + 0.1553804725408554, + 0.02361631765961647, + 0.022518832236528397, + 0.03360530361533165, + 0.07985921949148178, + 0.0336422324180603 + ], + "credit_map": [ + 0.21924415230751038, + 0.029238009825348854, + 0.10666479170322418, + 0.025128018110990524, + 0.033678341656923294, + 0.07136158645153046, + 0.019791901111602783, + 0.05471120774745941, + 0.006379749625921249, + 0.005437023472040892, + 0.008997885510325432, + 0.00431310897693038, + 0.006955988705158234, + 0.0031913809943944216, + 0.003977168817073107, + 0.006214473396539688, + 0.0031703391578048468, + 0.0026926591526716948, + 0.003403019392862916, + 0.002316681668162346, + 0.0034593832679092884, + 0.001814880408346653, + 0.0020667947828769684, + 0.0028045540675520897, + 0.004135563969612122, + 0.0017892868490889668, + 0.00238969037309289, + 0.002284843707457185, + 0.006348941940814257, + 0.00242540892213583, + 0.0019197185756638646, + 0.003071077633649111, + 0.012335462495684624, + 0.014257600530982018, + 0.012159502133727074, + 0.016668597236275673, + 0.01746436581015587, + 0.034045733511447906, + 0.018262522295117378, + 0.030186686664819717, + 0.0026336954906582832, + 0.0042572119273245335, + 0.0025814571417868137, + 0.002809640485793352, + 0.0026296682190150023, + 0.0018387610325589776, + 0.0027973814867436886, + 0.004068499896675348, + 0.002250717021524906, + 0.003044778946787119, + 0.002245935145765543, + 0.0023294442798942327, + 0.0018564066849648952, + 0.0020926035940647125, + 0.004015577957034111, + 0.004683369770646095, + 0.00281997537240386, + 0.003438964020460844, + 0.005221852567046881, + 0.005010177381336689, + 0.003530779154971242, + 0.0032070695888251066, + 0.004904876463115215, + 0.005471611395478249, + 0.008252525702118874, + 0.006199427880346775, + 0.005968857556581497, + 0.01700131595134735, + 0.006424839608371258, + 0.008133457973599434, + 0.015041250735521317, + 0.012837542220950127, + 0.0031476265285164118, + 0.002367312554270029, + 0.003984515555202961, + 0.006876521278172731, + 0.0017146972240880132, + 0.00199125031940639, + 0.007013233378529549, + 0.006547078490257263 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.05271732434630394 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.281321257352829 + ], + "opposite_logprob_tokens": [ + -1.4063212871551514 + ], + "logprob_reward": -0.281321257352829, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.281321257352829, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24250806868076324, + 0.0662464126944542, + 0.050422899425029755, + 0.18746350705623627, + 0.07737449556589127, + 0.048843659460544586, + 0.044642772525548935, + 0.09210735559463501, + 0.10639619082212448, + 0.08399471640586853 + ], + "credit_map": [ + 0.09451751410961151, + 0.013061380945146084, + 0.03969893231987953, + 0.011067286133766174, + 0.014726946130394936, + 0.032103292644023895, + 0.008852154016494751, + 0.02848055772483349, + 0.008668888360261917, + 0.0063677458092570305, + 0.015260005369782448, + 0.0061513520777225494, + 0.008715716190636158, + 0.005600859876722097, + 0.0049377125687897205, + 0.010544124990701675, + 0.008409053087234497, + 0.005632850341498852, + 0.007867835462093353, + 0.004093805328011513, + 0.00839999970048666, + 0.005045134574174881, + 0.004070301074534655, + 0.006903919391334057, + 0.03663799539208412, + 0.01724328100681305, + 0.02570902369916439, + 0.02103455364704132, + 0.026029307395219803, + 0.013746164739131927, + 0.015113339759409428, + 0.031949836760759354, + 0.011157425120472908, + 0.008286556228995323, + 0.008046029135584831, + 0.008284458890557289, + 0.010407570749521255, + 0.010663174092769623, + 0.009674504399299622, + 0.010854780673980713, + 0.005863289814442396, + 0.005921714473515749, + 0.006038039922714233, + 0.005100300069898367, + 0.005538371857255697, + 0.005062867887318134, + 0.005963955074548721, + 0.00935511663556099, + 0.0035436986945569515, + 0.0048834881745278835, + 0.005067727062851191, + 0.004104921128600836, + 0.0037177386693656445, + 0.006330867763608694, + 0.006632018368691206, + 0.010362309403717518, + 0.009023712947964668, + 0.013505822978913784, + 0.013490409590303898, + 0.010080992244184017, + 0.009318526834249496, + 0.011401746422052383, + 0.01135120540857315, + 0.013934937305748463, + 0.009940316900610924, + 0.009736734442412853, + 0.0071473438292741776, + 0.029203617945313454, + 0.007458665408194065, + 0.007171082776039839, + 0.02512119524180889, + 0.010617231018841267, + 0.008912043645977974, + 0.005892874673008919, + 0.006996281910687685, + 0.018484028056263924, + 0.0050025577656924725, + 0.005721508525311947, + 0.018060041591525078, + 0.014925382100045681 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 8 + ], + "selected_reward": -0.281321257352829 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020598072558641434 + ], + "opposite_logprob_tokens": [ + -3.8955981731414795 + ], + "logprob_reward": -0.020598072558641434, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020598072558641434, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.309352308511734, + 0.04025173932313919, + 0.07078234851360321, + 0.043571751564741135, + 0.0837940126657486, + 0.02532201260328293, + 0.11384076625108719, + 0.10655258595943451, + 0.135726660490036, + 0.07080581784248352 + ], + "credit_map": [ + 0.08004100620746613, + 0.022619666531682014, + 0.039797961711883545, + 0.01807522028684616, + 0.02884751558303833, + 0.045547112822532654, + 0.023389112204313278, + 0.05103471875190735, + 0.005660850089043379, + 0.004232081118971109, + 0.0069089788012206554, + 0.0052894712425768375, + 0.00517431553453207, + 0.003983577713370323, + 0.003447113325819373, + 0.005555351264774799, + 0.010909423232078552, + 0.005880335345864296, + 0.013098840601742268, + 0.0058889533393085, + 0.012232879176735878, + 0.00809171237051487, + 0.005174958612769842, + 0.009505247697234154, + 0.008284750394523144, + 0.004181277006864548, + 0.005149381700903177, + 0.005372517742216587, + 0.0064043267630040646, + 0.003609627718105912, + 0.004240448586642742, + 0.006329420488327742, + 0.008598947897553444, + 0.009381143376231194, + 0.008919109590351582, + 0.008418307639658451, + 0.011142074130475521, + 0.017579546198248863, + 0.010031521320343018, + 0.00972336158156395, + 0.0029748380184173584, + 0.003451666794717312, + 0.003235150594264269, + 0.002869930351153016, + 0.0033079637214541435, + 0.002502887276932597, + 0.002895187120884657, + 0.004084388259798288, + 0.012270915322005749, + 0.015586388297379017, + 0.014237483963370323, + 0.011498283594846725, + 0.010941535234451294, + 0.011980244889855385, + 0.01648770086467266, + 0.020838212221860886, + 0.009966472163796425, + 0.012852098792791367, + 0.010746540501713753, + 0.015413489192724228, + 0.011281309649348259, + 0.012928178533911705, + 0.01599314622581005, + 0.017371349036693573, + 0.01225781999528408, + 0.01306390855461359, + 0.010869291611015797, + 0.02995344251394272, + 0.009984288364648819, + 0.012171817943453789, + 0.02891710214316845, + 0.01850898005068302, + 0.006251272279769182, + 0.005931471008807421, + 0.0044445800594985485, + 0.015041392296552658, + 0.004361327737569809, + 0.004975861869752407, + 0.014558549039065838, + 0.015241355635225773 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.020598072558641434 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369521722197533 + ], + "opposite_logprob_tokens": [ + -3.4086952209472656 + ], + "logprob_reward": -0.03369521722197533, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369521722197533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.43962734937667847, + 0.10660620778799057, + 0.045318469405174255, + 0.05457207188010216, + 0.05307355150580406, + 0.04228869080543518, + 0.031242921948432922, + 0.05984705314040184, + 0.08093562722206116, + 0.08648804575204849 + ], + "credit_map": [ + 0.13877762854099274, + 0.01877135969698429, + 0.14069153368473053, + 0.0237513966858387, + 0.02151700109243393, + 0.03404361754655838, + 0.019536305218935013, + 0.04253853112459183, + 0.01175482477992773, + 0.013779697939753532, + 0.0153569420799613, + 0.023345595225691795, + 0.01485536340624094, + 0.008475478738546371, + 0.006339577957987785, + 0.01269872859120369, + 0.007485205307602882, + 0.005693669896572828, + 0.00564572773873806, + 0.005033658817410469, + 0.00821816734969616, + 0.004414116032421589, + 0.003393744118511677, + 0.005434179212898016, + 0.009494385682046413, + 0.005661382805556059, + 0.007103183306753635, + 0.006040540989488363, + 0.008294794708490372, + 0.0050131781026721, + 0.005099206231534481, + 0.007865406572818756, + 0.004847685340791941, + 0.004757686983793974, + 0.004060575272887945, + 0.003719198750331998, + 0.004798271227627993, + 0.006863028276711702, + 0.00995934009552002, + 0.014067769050598145, + 0.003889985615387559, + 0.0051765646785497665, + 0.0037971658166497946, + 0.0056698257103562355, + 0.004088637884706259, + 0.0040472825057804585, + 0.005524113774299622, + 0.010095113888382912, + 0.002518370980396867, + 0.00450525339692831, + 0.004292224068194628, + 0.003145812312141061, + 0.0031202661339193583, + 0.0031421741005033255, + 0.004663950763642788, + 0.005854873452335596, + 0.004157667048275471, + 0.007597353309392929, + 0.007450238335877657, + 0.0074898432940244675, + 0.006007855292409658, + 0.006677715107798576, + 0.008916937746107578, + 0.011549445800483227, + 0.006325136870145798, + 0.0060493010096251965, + 0.007591580040752888, + 0.015590565279126167, + 0.006589891854673624, + 0.009082251228392124, + 0.017736561596393585, + 0.011970339342951775, + 0.009342452511191368, + 0.005334137938916683, + 0.004965486004948616, + 0.024366559460759163, + 0.004677984863519669, + 0.004968069493770599, + 0.016462720930576324, + 0.01637064479291439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.03369521722197533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a custom-made M4A1 rifle resting on a black velvet display stand?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the camera execute a smooth, uninterrupted horizontal tracking shot moving from left to right?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the rifle's metal receiver gunmetal-gray with three-dimensional rose relief patterns featuring polished petals and thorny stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the video show the camera starting at the tactical stock and ending at the handguard and muzzle brake with a Picatinny rail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background completely blurred with soft, diffused studio lighting coming from a single direction?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4742884635925293 + ], + "opposite_logprob_tokens": [ + -0.9742884635925293 + ], + "logprob_reward": -0.4742884635925293, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4742884635925293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4309979975223541, + 0.056048229336738586, + 0.08949093520641327, + 0.045547615736722946, + 0.053585343062877655, + 0.05753222480416298, + 0.049668945372104645, + 0.03358913213014603, + 0.03922215476632118, + 0.14431744813919067 + ], + "credit_map": [ + 0.11218972504138947, + 0.02137499675154686, + 0.16446377336978912, + 0.027013912796974182, + 0.017389409244060516, + 0.017159614711999893, + 0.012997172772884369, + 0.058409374207258224, + 0.006651079747825861, + 0.007645624224096537, + 0.00944182462990284, + 0.00652801850810647, + 0.007434200029820204, + 0.0063148424960672855, + 0.004956007935106754, + 0.007076628971844912, + 0.013088757172226906, + 0.0072713796980679035, + 0.02101113088428974, + 0.00930446945130825, + 0.010665284469723701, + 0.006231493782252073, + 0.00734564708545804, + 0.014572774060070515, + 0.008923953399062157, + 0.0038745154161006212, + 0.004938412923365831, + 0.005643371492624283, + 0.00982979778200388, + 0.003002044977620244, + 0.003741543274372816, + 0.005593976937234402, + 0.005262748803943396, + 0.006363717373460531, + 0.006068668328225613, + 0.005118845496326685, + 0.009315187111496925, + 0.006563096307218075, + 0.006787363439798355, + 0.008105716668069363, + 0.006318477913737297, + 0.006831438280642033, + 0.006823656614869833, + 0.008141911588609219, + 0.006517424713820219, + 0.005093215964734554, + 0.007865742780268192, + 0.009940359741449356, + 0.00475001335144043, + 0.006343063898384571, + 0.00574439438059926, + 0.004945672582834959, + 0.005917605012655258, + 0.006280392874032259, + 0.007507472764700651, + 0.008180328644812107, + 0.002523665549233556, + 0.004882222507148981, + 0.003652246668934822, + 0.004234324675053358, + 0.004385028965771198, + 0.0036397308576852083, + 0.004530858714133501, + 0.005741055589169264, + 0.0044481088407337666, + 0.004694389645010233, + 0.004778287839144468, + 0.00538012245669961, + 0.0048974668607115746, + 0.004322963301092386, + 0.00591018283739686, + 0.004790635779500008, + 0.009901399724185467, + 0.006973506882786751, + 0.008285240270197392, + 0.046962495893239975, + 0.006523735821247101, + 0.00639656325802207, + 0.03528270870447159, + 0.023991793394088745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 2 + ], + "selected_reward": -0.4742884635925293 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_tensor([7011, 7011])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_tensor([7011, 7011])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..6c20fac5a9e44c8f04d271a62051649cccab7192 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_tensor([7011, 7011])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, long shot captured from an extremely low angle, the composition exhibits a near-perfect, unsettling symmetry. On a pitch-black asphalt street, thoroughly drenched by fresh rain, two classic black sedans stand head-to-head, motionless. Each with its high-intensity high beams on, emitting a cold, almost white glow, these two sharp beams pierce the humid night sky, illuminating the still-falling fine raindrops like dancing dust motes. The slippery road, like a giant black mirror, perfectly reflects the two piercing beams and the silhouettes of the cars, creating a double glare effect that intensifies the image's visual impact. Silent buildings, blurred in the darkness, line the street. With no other light sources, these two opposing beams become the sole focal point of the world. The stark contrast between light and darkness, combined with the absolute stillness, creates a tense and suspenseful atmosphere of confrontation.", + "seed": 3388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3388_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.05425583617761731, + "vlm_logprob": -0.05425583617761731, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.05425583617761731, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11946877092123032 + ], + "opposite_logprob_tokens": [ + -2.2444686889648438 + ], + "logprob_reward": -0.11946877092123032, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11946877092123032, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09571950137615204, + 0.02488655596971512, + 0.07958175987005234, + 0.22559675574302673, + 0.3075897991657257, + 0.07424955070018768, + 0.016607511788606644, + 0.04369319602847099, + 0.022482113912701607, + 0.10959325730800629 + ], + "credit_map": [ + 0.015049266628921032, + 0.010194838047027588, + 0.014288215897977352, + 0.01440394390374422, + 0.007833120413124561, + 0.011852050200104713, + 0.011328715831041336, + 0.01076934952288866, + 0.00355014530941844, + 0.0024732970632612705, + 0.003903122851625085, + 0.0038851893041282892, + 0.002304050140082836, + 0.003070906037464738, + 0.002333232434466481, + 0.0033666144590824842, + 0.01605086773633957, + 0.007296539377421141, + 0.006557433400303125, + 0.005842399783432484, + 0.008204150944948196, + 0.012517114169895649, + 0.011420663446187973, + 0.011692600324749947, + 0.08041080832481384, + 0.09725877642631531, + 0.005641258787363768, + 0.007065945770591497, + 0.012012827210128307, + 0.0128282206133008, + 0.006035314407199621, + 0.004343611188232899, + 0.06865353882312775, + 0.12173691391944885, + 0.029849432408809662, + 0.013608788140118122, + 0.023841233924031258, + 0.02191481925547123, + 0.010833578184247017, + 0.017151515930891037, + 0.00994518119841814, + 0.02715568244457245, + 0.009084325283765793, + 0.004694418981671333, + 0.005316124763339758, + 0.005761985667049885, + 0.004599232226610184, + 0.007692602463066578, + 0.001922363298945129, + 0.002543185604736209, + 0.002506262855604291, + 0.0025300492998212576, + 0.0012366176815703511, + 0.0035687158815562725, + 0.0011297562159597874, + 0.0011705610668286681, + 0.008543762378394604, + 0.010404258966445923, + 0.0061003961600363255, + 0.0043042185716331005, + 0.003212140640243888, + 0.005533847026526928, + 0.0029212862718850374, + 0.0026732897385954857, + 0.0026141402777284384, + 0.002304835245013237, + 0.003924556076526642, + 0.003028104081749916, + 0.0033981746528297663, + 0.0031835606787353754, + 0.0021995489951223135, + 0.0018291950691491365, + 0.01870627887547016, + 0.03507910296320915, + 0.012921844609081745, + 0.0160367451608181, + 0.007682423107326031, + 0.01049676164984703, + 0.0038859283085912466, + 0.0047841728664934635 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 9 + ], + "selected_reward": -0.11946877092123032 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208469718694687 + ], + "opposite_logprob_tokens": [ + -2.812084674835205 + ], + "logprob_reward": -0.06208469718694687, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208469718694687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26695582270622253, + 0.09481920301914215, + 0.027929946780204773, + 0.15427279472351074, + 0.21430577337741852, + 0.11297949403524399, + 0.03603140637278557, + 0.020730314776301384, + 0.00997790414839983, + 0.061997342854738235 + ], + "credit_map": [ + 0.03904908895492554, + 0.020531469956040382, + 0.04361063987016678, + 0.05250146985054016, + 0.01942831091582775, + 0.025965183973312378, + 0.029500162228941917, + 0.03636949509382248, + 0.01382516510784626, + 0.010563433170318604, + 0.00976891815662384, + 0.013429508544504642, + 0.008816765621304512, + 0.010572981089353561, + 0.01043330691754818, + 0.017409123480319977, + 0.004030085634440184, + 0.004293265286833048, + 0.0022994293831288815, + 0.0028748870827257633, + 0.00356785929761827, + 0.0036141970194876194, + 0.002131278160959482, + 0.005118942353874445, + 0.05337712541222572, + 0.06657196581363678, + 0.0037688962183892727, + 0.004637818783521652, + 0.008107409812510014, + 0.009046266786754131, + 0.004099694080650806, + 0.004663616884499788, + 0.050340499728918076, + 0.07716811448335648, + 0.014999103732407093, + 0.009043892845511436, + 0.014736638404428959, + 0.020778730511665344, + 0.010416632518172264, + 0.016822168603539467, + 0.014519746415317059, + 0.04108899086713791, + 0.015960093587636948, + 0.0055658891797065735, + 0.008310501463711262, + 0.0092760706320405, + 0.0060174488462507725, + 0.012240747921168804, + 0.0042855399660766125, + 0.005367649719119072, + 0.0035850831773132086, + 0.0028064653743058443, + 0.00672357389703393, + 0.007820594124495983, + 0.0022973557934165, + 0.0031451459508389235, + 0.0022201859392225742, + 0.0032994416542351246, + 0.002972156275063753, + 0.0024884624872356653, + 0.0027912051882594824, + 0.0024872184731066227, + 0.0023687046486884356, + 0.002102940110489726, + 0.001376185566186905, + 0.001220707199536264, + 0.0012086055940017104, + 0.0008437085198238492, + 0.001970227574929595, + 0.0015736951027065516, + 0.0008691315306350589, + 0.0009156432352028787, + 0.00933036208152771, + 0.014334979467093945, + 0.01140413898974657, + 0.006942399777472019, + 0.005402686074376106, + 0.008107501082122326, + 0.0030118245631456375, + 0.0034634501207619905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.06208469718694687 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016088780015707016 + ], + "opposite_logprob_tokens": [ + -4.141088962554932 + ], + "logprob_reward": -0.016088780015707016, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016088780015707016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12049157172441483, + 0.08921422064304352, + 0.13228300213813782, + 0.14493800699710846, + 0.31384503841400146, + 0.0326269157230854, + 0.03943612426519394, + 0.015676436945796013, + 0.015204797498881817, + 0.09628384560346603 + ], + "credit_map": [ + 0.015494145452976227, + 0.01384773850440979, + 0.02473209612071514, + 0.02071201615035534, + 0.008120612241327763, + 0.011744924820959568, + 0.012899077497422695, + 0.01294095627963543, + 0.021486250683665276, + 0.009650456719100475, + 0.01619953289628029, + 0.008532705716788769, + 0.00509247463196516, + 0.007973908446729183, + 0.009208408184349537, + 0.011070477776229382, + 0.023244298994541168, + 0.012019112706184387, + 0.011403375305235386, + 0.011400331743061543, + 0.012930260971188545, + 0.01975526288151741, + 0.020839137956500053, + 0.02069120854139328, + 0.05197412893176079, + 0.05485914275050163, + 0.005154464393854141, + 0.005048852879554033, + 0.010228819213807583, + 0.008135891519486904, + 0.004503996577113867, + 0.005032708868384361, + 0.06341080367565155, + 0.12769724428653717, + 0.03391113132238388, + 0.012275676243007183, + 0.02221643179655075, + 0.02192564681172371, + 0.007497559301555157, + 0.024910561740398407, + 0.003723714267835021, + 0.006749045103788376, + 0.004607880488038063, + 0.00241941399872303, + 0.0024432758800685406, + 0.0057168034836649895, + 0.002070740098133683, + 0.004896041937172413, + 0.004846360068768263, + 0.011153388768434525, + 0.003910336177796125, + 0.0029889517463743687, + 0.005225155968219042, + 0.005580287892371416, + 0.0023406229447573423, + 0.003391020465642214, + 0.0017154221422970295, + 0.0034199648071080446, + 0.002226258860900998, + 0.0016686501912772655, + 0.0018966193310916424, + 0.0017275845166295767, + 0.0012194039300084114, + 0.0018025352619588375, + 0.0018639828776940703, + 0.0017764990916475654, + 0.0017724441131576896, + 0.0021522340830415487, + 0.0017573873046785593, + 0.002760743023827672, + 0.0018089083023369312, + 0.0013125992845743895, + 0.017737675458192825, + 0.03006024658679962, + 0.012795036658644676, + 0.009551431983709335, + 0.007829009555280209, + 0.009777361527085304, + 0.0035542629193514585, + 0.004978821612894535 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 2 + ], + "selected_reward": -0.016088780015707016 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036214501596987247 + ], + "opposite_logprob_tokens": [ + -5.628621578216553 + ], + "logprob_reward": -0.0036214501596987247, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036214501596987247, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15086297690868378, + 0.06910672783851624, + 0.062281835824251175, + 0.24011503159999847, + 0.26698049902915955, + 0.0728282555937767, + 0.022166015580296516, + 0.03270065784454346, + 0.016506560146808624, + 0.06645145267248154 + ], + "credit_map": [ + 0.022874483838677406, + 0.012222892604768276, + 0.022725123912096024, + 0.025325825437903404, + 0.014505152590572834, + 0.014298571273684502, + 0.01791023276746273, + 0.0210006944835186, + 0.0143543416634202, + 0.008067739196121693, + 0.014898250810801983, + 0.006871752440929413, + 0.005517173558473587, + 0.00531997624784708, + 0.006662881933152676, + 0.007414616644382477, + 0.011159704066812992, + 0.006034719292074442, + 0.005410898942500353, + 0.005133562721312046, + 0.007146773394197226, + 0.009531906805932522, + 0.008028276264667511, + 0.009835994802415371, + 0.0838121846318245, + 0.1042148768901825, + 0.006262718699872494, + 0.007159407250583172, + 0.013301217928528786, + 0.01361286174505949, + 0.006346572656184435, + 0.0054052118211984634, + 0.08593703806400299, + 0.06322605162858963, + 0.029287902638316154, + 0.013325135223567486, + 0.018210401758551598, + 0.018695570528507233, + 0.010459020733833313, + 0.027839364483952522, + 0.014270040206611156, + 0.011814182624220848, + 0.013078072108328342, + 0.007898088544607162, + 0.005763516295701265, + 0.00888238288462162, + 0.0047608609311282635, + 0.006361110135912895, + 0.002262563444674015, + 0.0033433567732572556, + 0.0022547964472323656, + 0.001972444821149111, + 0.004501650109887123, + 0.004154801368713379, + 0.0016246380982920527, + 0.0020517664961516857, + 0.004517479799687862, + 0.006822370924055576, + 0.003913666121661663, + 0.002835734747350216, + 0.0037782141007483006, + 0.005230468697845936, + 0.003412532852962613, + 0.002190189901739359, + 0.0023749906104058027, + 0.0017002879176288843, + 0.0019588449504226446, + 0.002198130125179887, + 0.003104486968368292, + 0.0016696196980774403, + 0.0014781770296394825, + 0.002022023079916835, + 0.007736359257251024, + 0.010890805162489414, + 0.011914619244635105, + 0.010060911066830158, + 0.008084267377853394, + 0.007273909170180559, + 0.004953823518007994, + 0.005536759737879038 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.0036214501596987247 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001548260450363 + ], + "opposite_logprob_tokens": [ + -2.6950154304504395 + ], + "logprob_reward": -0.07001548260450363, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001548260450363, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0666215568780899, + 0.02688252367079258, + 0.04340570420026779, + 0.3486506938934326, + 0.2801898419857025, + 0.06856238842010498, + 0.023717055097222328, + 0.049898698925971985, + 0.016538439318537712, + 0.07553309202194214 + ], + "credit_map": [ + 0.009493003599345684, + 0.005505653098225594, + 0.011156001128256321, + 0.01648668944835663, + 0.005724141839891672, + 0.006404388230293989, + 0.005597010254859924, + 0.006254672538489103, + 0.003464499721303582, + 0.0034325397573411465, + 0.0033860730472952127, + 0.003150802804157138, + 0.002412272384390235, + 0.003515125485137105, + 0.0027529310900717974, + 0.004768280778080225, + 0.006908842828124762, + 0.004366772715002298, + 0.003224582877010107, + 0.0036862995475530624, + 0.00493581360206008, + 0.006757729686796665, + 0.006398709025233984, + 0.007126954849809408, + 0.1211128979921341, + 0.1494719684123993, + 0.009733861312270164, + 0.010887154377996922, + 0.020780472084879875, + 0.01942904107272625, + 0.009142477996647358, + 0.00809281226247549, + 0.07147655636072159, + 0.11811468005180359, + 0.022221272811293602, + 0.00891724694520235, + 0.018235523253679276, + 0.01841777190566063, + 0.006672190967947245, + 0.016134584322571754, + 0.015227599069476128, + 0.009523053653538227, + 0.014735044911503792, + 0.00736027117818594, + 0.004988024942576885, + 0.006406374741345644, + 0.004676886834204197, + 0.005645129829645157, + 0.003323524259030819, + 0.004848772659897804, + 0.0033392230980098248, + 0.0023594943340867758, + 0.003229772672057152, + 0.0031421512831002474, + 0.0017969431355595589, + 0.0016771735390648246, + 0.011008672416210175, + 0.014932743273675442, + 0.005640408024191856, + 0.0037107605021446943, + 0.0035063305404037237, + 0.005885484628379345, + 0.00288380472920835, + 0.00233049551025033, + 0.0021484808530658484, + 0.0021308870054781437, + 0.0033505382016301155, + 0.001699301996268332, + 0.0023820013739168644, + 0.0026059073861688375, + 0.000941069854889065, + 0.0012802521232515574, + 0.009752696380019188, + 0.006376713514328003, + 0.008951732888817787, + 0.023629192262887955, + 0.007202881388366222, + 0.00532431760802865, + 0.004967723973095417, + 0.009327835403382778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 4, + 9 + ], + "selected_reward": -0.07001548260450363 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captured from an extremely low angle, the composition exhibits a near-perfect, unsettling symmetry. On a pitch-black asphalt street, thoroughly drenched by fresh rain, two classic black sedans stand head-to-head, motionless. Each with its high-intensity high beams on, emitting a cold, almost white glow, these two sharp beams pierce the humid night sky, illuminating the still-falling fine raindrops like dancing dust motes. The slippery road, like a giant black mirror, perfectly reflects the two piercing beams and the silhouettes of the cars, creating a double glare effect that intensifies the image's visual impact. Silent buildings, blurred in the darkness, line the street. With no other light sources, these two opposing beams become the sole focal point of the world. The stark contrast between light and darkness, combined with the absolute stillness, creates a tense and suspenseful atmosphere of confrontation.", + "seed": 3389, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3389_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.05621015606448054, + "vlm_logprob": -0.05621015606448054, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.05621015606448054, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.060212887823581696 + ], + "opposite_logprob_tokens": [ + -2.9352128505706787 + ], + "logprob_reward": -0.060212887823581696, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.060212887823581696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0888981744647026, + 0.16390058398246765, + 0.10279424488544464, + 0.17213591933250427, + 0.14701662957668304, + 0.218166321516037, + 0.02535073831677437, + 0.023402608931064606, + 0.025471659377217293, + 0.0328630767762661 + ], + "credit_map": [ + 0.022179031744599342, + 0.007094919681549072, + 0.008514168672263622, + 0.0058323657140135765, + 0.006863851100206375, + 0.007104876916855574, + 0.0089827636256814, + 0.02232619933784008, + 0.011488803662359715, + 0.01213616319000721, + 0.020903293043375015, + 0.007115525659173727, + 0.023444168269634247, + 0.012208028696477413, + 0.03456810116767883, + 0.04203649237751961, + 0.011566522531211376, + 0.009633435867726803, + 0.0054985033348202705, + 0.008189144544303417, + 0.02385437861084938, + 0.008601698093116283, + 0.027629654854536057, + 0.00782090611755848, + 0.03629695624113083, + 0.012854778207838535, + 0.01740366220474243, + 0.04890508949756622, + 0.010785180144011974, + 0.010885126888751984, + 0.025497805327177048, + 0.009507320821285248, + 0.02051456645131111, + 0.012815211899578571, + 0.04470504820346832, + 0.01281898282468319, + 0.023069439455866814, + 0.010735815390944481, + 0.009339665062725544, + 0.01301790215075016, + 0.017540933564305305, + 0.09288522601127625, + 0.010339273139834404, + 0.00934569537639618, + 0.0491236113011837, + 0.018529396504163742, + 0.004887125454843044, + 0.015515059232711792, + 0.0027893641963601112, + 0.0040086861699819565, + 0.003946148790419102, + 0.00414471747353673, + 0.0025972328148782253, + 0.0027435850352048874, + 0.003244553692638874, + 0.0018764508422464132, + 0.003679040353745222, + 0.004287724383175373, + 0.0029906744603067636, + 0.002340573351830244, + 0.0021392505150288343, + 0.002573649864643812, + 0.002609053859487176, + 0.0027826435398310423, + 0.0030404964927583933, + 0.004229375161230564, + 0.003198362188413739, + 0.004445853177458048, + 0.00268369959667325, + 0.0013956258771941066, + 0.004907925613224506, + 0.001570322667248547, + 0.0049031744711101055, + 0.002639011712744832, + 0.005180432926863432, + 0.005240851081907749, + 0.003146333387121558, + 0.0041696117259562016, + 0.004739139229059219, + 0.0028445213101804256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 1 + ], + "selected_reward": -0.060212887823581696 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10033771395683289 + ], + "opposite_logprob_tokens": [ + -2.3503377437591553 + ], + "logprob_reward": -0.10033771395683289, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10033771395683289, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19155119359493256, + 0.06346907466650009, + 0.0720919668674469, + 0.161031574010849, + 0.07329723238945007, + 0.32612860202789307, + 0.031414005905389786, + 0.02705468237400055, + 0.029935529455542564, + 0.024026120081543922 + ], + "credit_map": [ + 0.014036034233868122, + 0.011822291649878025, + 0.012918912805616856, + 0.009586664848029613, + 0.0190484169870615, + 0.015086333267390728, + 0.050253015011548996, + 0.058799512684345245, + 0.007556776516139507, + 0.005150275770574808, + 0.004721301142126322, + 0.004305117763578892, + 0.0037560584023594856, + 0.00543634220957756, + 0.010750007815659046, + 0.021793195977807045, + 0.006717908661812544, + 0.0061286319978535175, + 0.005098155699670315, + 0.008168764412403107, + 0.017670564353466034, + 0.005748705938458443, + 0.016012810170650482, + 0.006546424701809883, + 0.04380372166633606, + 0.012449909932911396, + 0.01326720230281353, + 0.04210996255278587, + 0.010058904998004436, + 0.01124549563974142, + 0.01912103407084942, + 0.008975334465503693, + 0.006702123675495386, + 0.007898353040218353, + 0.013125479221343994, + 0.003125315997749567, + 0.026941431686282158, + 0.007421568967401981, + 0.003877543145790696, + 0.004205417353659868, + 0.027509329840540886, + 0.13134361803531647, + 0.014651103876531124, + 0.013835704885423183, + 0.06908328086137772, + 0.02652893401682377, + 0.01284210104495287, + 0.03033452108502388, + 0.0039038110990077257, + 0.004776038695126772, + 0.004676118493080139, + 0.003369601210579276, + 0.0039891586638987064, + 0.002931583672761917, + 0.004701675847172737, + 0.0030660212505608797, + 0.00330017670057714, + 0.004103548359125853, + 0.004326615948230028, + 0.0029371746350079775, + 0.00396369444206357, + 0.0028189225122332573, + 0.002772317733615637, + 0.0028322329744696617, + 0.00293828290887177, + 0.004215017426759005, + 0.003884214209392667, + 0.004091145470738411, + 0.005042455159127712, + 0.002363823354244232, + 0.004848775919526815, + 0.0025518147740513086, + 0.004383313935250044, + 0.002125392435118556, + 0.003612156957387924, + 0.0027604082133620977, + 0.0022166394628584385, + 0.0033272667787969112, + 0.0026489454321563244, + 0.002951997332274914 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.10033771395683289 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160913597792387 + ], + "opposite_logprob_tokens": [ + -4.141091346740723 + ], + "logprob_reward": -0.0160913597792387, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160913597792387, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04763997346162796, + 0.21252429485321045, + 0.1163075789809227, + 0.01478969119489193, + 0.021348992362618446, + 0.3327251672744751, + 0.08134135603904724, + 0.04833540320396423, + 0.08301354944705963, + 0.041973937302827835 + ], + "credit_map": [ + 0.004793982952833176, + 0.004325344227254391, + 0.004828826058655977, + 0.006937498692423105, + 0.005884175654500723, + 0.004702264908701181, + 0.007765795569866896, + 0.008402087725698948, + 0.014350580051541328, + 0.013364112935960293, + 0.01661541685461998, + 0.013178604654967785, + 0.013918334618210793, + 0.01755855791270733, + 0.0500292107462883, + 0.07350950688123703, + 0.009636376984417439, + 0.010037812404334545, + 0.007007579784840345, + 0.013540652580559254, + 0.03168105334043503, + 0.009730441495776176, + 0.016010204330086708, + 0.018663467839360237, + 0.0013440399197861552, + 0.0012599935289472342, + 0.0016850357642397285, + 0.002448124112561345, + 0.0039372225292027, + 0.0012502616737037897, + 0.0017224481562152505, + 0.0011425659758970141, + 0.003219437552616, + 0.0021788040176033974, + 0.0029501819517463446, + 0.0022079513873904943, + 0.006035435479134321, + 0.0021709802094846964, + 0.0011674299603328109, + 0.0014187738997861743, + 0.024539116770029068, + 0.12563152611255646, + 0.017643125727772713, + 0.014851081185042858, + 0.050606418401002884, + 0.03016245923936367, + 0.046558745205402374, + 0.02273271232843399, + 0.009116844274103642, + 0.03070479817688465, + 0.007101832889020443, + 0.007732309401035309, + 0.01065076794475317, + 0.006427260115742683, + 0.005630893167108297, + 0.0039766631089150906, + 0.004280075430870056, + 0.009557180106639862, + 0.015580574981868267, + 0.004066071007400751, + 0.003712946083396673, + 0.0029044775292277336, + 0.00364436162635684, + 0.004589723888784647, + 0.009739653207361698, + 0.011426319368183613, + 0.011417089030146599, + 0.015587517991662025, + 0.010046908631920815, + 0.004344153683632612, + 0.01503173727542162, + 0.00542017538100481, + 0.00610888097435236, + 0.003829974913969636, + 0.0066583361476659775, + 0.0058354889042675495, + 0.004317588172852993, + 0.0051519316621124744, + 0.00525282695889473, + 0.004818913526833057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.0160913597792387 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004103455226868391 + ], + "opposite_logprob_tokens": [ + -5.504103660583496 + ], + "logprob_reward": -0.004103455226868391, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004103455226868391, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18612000346183777, + 0.11909593641757965, + 0.03567471355199814, + 0.07041376084089279, + 0.14310084283351898, + 0.2920341491699219, + 0.07373400032520294, + 0.031986769288778305, + 0.023817293345928192, + 0.024022504687309265 + ], + "credit_map": [ + 0.020284732803702354, + 0.016060784459114075, + 0.02295750379562378, + 0.014997756108641624, + 0.013483403250575066, + 0.014670797623693943, + 0.03357981517910957, + 0.05008520558476448, + 0.01839854195713997, + 0.011551761999726295, + 0.007851365022361279, + 0.00674189580604434, + 0.01577095128595829, + 0.008972598239779472, + 0.029512856155633926, + 0.020295970141887665, + 0.002921219915151596, + 0.0030136064160615206, + 0.004833828657865524, + 0.0022256665397435427, + 0.008631373755633831, + 0.0033095935359597206, + 0.006231039762496948, + 0.004508384969085455, + 0.0205454733222723, + 0.005930663552135229, + 0.008542578667402267, + 0.016517166048288345, + 0.004742408636957407, + 0.005068490281701088, + 0.00530961761251092, + 0.0037573662120848894, + 0.011838137172162533, + 0.01779758930206299, + 0.004778682254254818, + 0.004407237283885479, + 0.06830298155546188, + 0.01814960315823555, + 0.007758485618978739, + 0.010068134404718876, + 0.0229981429874897, + 0.0638396218419075, + 0.015889253467321396, + 0.010673842392861843, + 0.04298316687345505, + 0.02051626332104206, + 0.042130667716264725, + 0.07300321012735367, + 0.008949658833444118, + 0.01324431411921978, + 0.010038480162620544, + 0.009351721033453941, + 0.009759115986526012, + 0.006456124130636454, + 0.00826761033385992, + 0.007666982244700193, + 0.0047996994107961655, + 0.004388216882944107, + 0.00479085510596633, + 0.0032571451738476753, + 0.003157199826091528, + 0.0035014210734516382, + 0.0044304439797997475, + 0.003661789931356907, + 0.0028268208261579275, + 0.003199605969712138, + 0.0027595546562224627, + 0.00407140189781785, + 0.0031719976104795933, + 0.0016078685875982046, + 0.003938128240406513, + 0.0022419155575335026, + 0.0030770471785217524, + 0.002597617916762829, + 0.0031631302554160357, + 0.003865876467898488, + 0.002167653990909457, + 0.0028531893622130156, + 0.0039412048645317554, + 0.0023567848838865757 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004103455226868391 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030536353588104 + ], + "opposite_logprob_tokens": [ + -2.3503053188323975 + ], + "logprob_reward": -0.10030536353588104, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030536353588104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12356499582529068, + 0.06197257339954376, + 0.0742124542593956, + 0.06341727077960968, + 0.06915921717882156, + 0.4101291298866272, + 0.09616082161664963, + 0.034781381487846375, + 0.046172525733709335, + 0.020429588854312897 + ], + "credit_map": [ + 0.009552371688187122, + 0.009723670780658722, + 0.0073820119723677635, + 0.006895026192069054, + 0.02072293497622013, + 0.00936899147927761, + 0.03252624347805977, + 0.027393752709031105, + 0.0087891248986125, + 0.005541563965380192, + 0.00845099613070488, + 0.003407727461308241, + 0.013615689240396023, + 0.0046021356247365475, + 0.011659406125545502, + 0.005905928555876017, + 0.0047129797749221325, + 0.007958030328154564, + 0.00594578729942441, + 0.005598495248705149, + 0.009960601106286049, + 0.006119561847299337, + 0.016099654138088226, + 0.01781734824180603, + 0.0141885532066226, + 0.00535713043063879, + 0.007453272119164467, + 0.015956878662109375, + 0.006830241996794939, + 0.00427375128492713, + 0.005162240471690893, + 0.0041952054016292095, + 0.0067498041316866875, + 0.007710665930062532, + 0.006174482870846987, + 0.0022010458633303642, + 0.030269479379057884, + 0.007876362651586533, + 0.0038277804851531982, + 0.004349597729742527, + 0.02689731866121292, + 0.1593071073293686, + 0.018861720338463783, + 0.016679730266332626, + 0.08529471606016159, + 0.03279287740588188, + 0.019260846078395844, + 0.0510348379611969, + 0.010715408250689507, + 0.017211850732564926, + 0.018466949462890625, + 0.008938955143094063, + 0.012510433793067932, + 0.008187520317733288, + 0.012045495212078094, + 0.008084218017756939, + 0.004414813127368689, + 0.005002948921173811, + 0.004451012704521418, + 0.004823875613510609, + 0.004006393253803253, + 0.0036625289358198643, + 0.004443100653588772, + 0.003976708278059959, + 0.0056151351891458035, + 0.0066558788530528545, + 0.00868193618953228, + 0.007720365654677153, + 0.005195053294301033, + 0.0031941207125782967, + 0.0064440276473760605, + 0.002666008658707142, + 0.0025281263515353203, + 0.0018647875403985381, + 0.0029028141871094704, + 0.0034412939567118883, + 0.002319018356502056, + 0.0024754987098276615, + 0.002415412338450551, + 0.0024826377630233765 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.10030536353588104 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captured from an extremely low angle, the composition exhibits a near-perfect, unsettling symmetry. On a pitch-black asphalt street, thoroughly drenched by fresh rain, two classic black sedans stand head-to-head, motionless. Each with its high-intensity high beams on, emitting a cold, almost white glow, these two sharp beams pierce the humid night sky, illuminating the still-falling fine raindrops like dancing dust motes. The slippery road, like a giant black mirror, perfectly reflects the two piercing beams and the silhouettes of the cars, creating a double glare effect that intensifies the image's visual impact. Silent buildings, blurred in the darkness, line the street. With no other light sources, these two opposing beams become the sole focal point of the world. The stark contrast between light and darkness, combined with the absolute stillness, creates a tense and suspenseful atmosphere of confrontation.", + "seed": 3390, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3390_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.053279528860002756, + "vlm_logprob": -0.053279528860002756, + "vlm_yes_no_margin": 3.475, + "vlm_token_credit": -0.053279528860002756, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12242384999990463 + ], + "opposite_logprob_tokens": [ + -2.2474238872528076 + ], + "logprob_reward": -0.12242384999990463, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.12242384999990463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21400295197963715, + 0.10107424855232239, + 0.06046346575021744, + 0.04229700192809105, + 0.18457964062690735, + 0.04237206652760506, + 0.05352068692445755, + 0.04469311982393265, + 0.0584966205060482, + 0.19850018620491028 + ], + "credit_map": [ + 0.03633187711238861, + 0.03763013705611229, + 0.020430583506822586, + 0.020403878763318062, + 0.0372861847281456, + 0.021154139190912247, + 0.02463642507791519, + 0.01612973026931286, + 0.019685929641127586, + 0.01615220680832863, + 0.008455922827124596, + 0.008922701701521873, + 0.02153598703444004, + 0.010109574533998966, + 0.00979362241923809, + 0.006418307777494192, + 0.01200096309185028, + 0.010176410898566246, + 0.0036298062186688185, + 0.004131906200200319, + 0.013673379085958004, + 0.008013845421373844, + 0.004769851919263601, + 0.004067301284521818, + 0.007612667977809906, + 0.005806960165500641, + 0.002837770851328969, + 0.003636762499809265, + 0.0073492019437253475, + 0.0067923651076853275, + 0.004140838980674744, + 0.004120435565710068, + 0.03484232351183891, + 0.020943675190210342, + 0.01870255544781685, + 0.021537574008107185, + 0.027237411588430405, + 0.02551807090640068, + 0.01880938559770584, + 0.016988633200526237, + 0.007916130125522614, + 0.003731350414454937, + 0.0050404248759150505, + 0.003921229392290115, + 0.008138054050505161, + 0.0035047486890107393, + 0.006080596707761288, + 0.004039532504975796, + 0.009003148414194584, + 0.003883878467604518, + 0.005530474707484245, + 0.005185648333281279, + 0.007856405340135098, + 0.00796375423669815, + 0.008026296272873878, + 0.0060710785910487175, + 0.011244340799748898, + 0.0032548080198466778, + 0.004755148198455572, + 0.003322367090731859, + 0.005157142877578735, + 0.0035335521679371595, + 0.007016268093138933, + 0.0064094895496964455, + 0.007519835140556097, + 0.005730032920837402, + 0.005896939896047115, + 0.005447711795568466, + 0.004657452926039696, + 0.009146914817392826, + 0.009850681759417057, + 0.01024705171585083, + 0.02190844528377056, + 0.01850653998553753, + 0.013875048607587814, + 0.031363267451524734, + 0.01974412240087986, + 0.02197207883000374, + 0.049307651817798615, + 0.021823029965162277 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 4 + ], + "selected_reward": -0.12242384999990463 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054992545396089554 + ], + "opposite_logprob_tokens": [ + -2.929992437362671 + ], + "logprob_reward": -0.054992545396089554, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054992545396089554, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13993212580680847, + 0.061750177294015884, + 0.15002799034118652, + 0.1559954285621643, + 0.031189389526844025, + 0.1011156439781189, + 0.04614639654755592, + 0.18245643377304077, + 0.082344189286232, + 0.049042247235774994 + ], + "credit_map": [ + 0.027321714907884598, + 0.019788052886724472, + 0.018701709806919098, + 0.014000185765326023, + 0.01522279717028141, + 0.01960194669663906, + 0.011821535415947437, + 0.01347417663782835, + 0.011273742653429508, + 0.009120640344917774, + 0.007128851488232613, + 0.005361615214496851, + 0.00826682522892952, + 0.007721449714154005, + 0.0064971731044352055, + 0.0063798800110816956, + 0.027180511504411697, + 0.019862918183207512, + 0.010192754678428173, + 0.010855748318135738, + 0.033247675746679306, + 0.024725643917918205, + 0.012859761714935303, + 0.01110298652201891, + 0.028359917923808098, + 0.02199425920844078, + 0.013473736122250557, + 0.01969013921916485, + 0.022965336218476295, + 0.01827320270240307, + 0.01652439869940281, + 0.014714442193508148, + 0.006757003720849752, + 0.0037520253099501133, + 0.0018663189839571714, + 0.0026103658601641655, + 0.0041256980039179325, + 0.005049617029726505, + 0.0026303494814783335, + 0.004398011602461338, + 0.017342468723654747, + 0.008985663764178753, + 0.010226134210824966, + 0.008795992471277714, + 0.024834338575601578, + 0.008288484998047352, + 0.015193819999694824, + 0.007448738906532526, + 0.008576183579862118, + 0.004408384207636118, + 0.005961553659290075, + 0.004433711525052786, + 0.006072873715311289, + 0.004783736076205969, + 0.006834292318671942, + 0.005075663793832064, + 0.028262564912438393, + 0.017865853384137154, + 0.021745692938566208, + 0.02051869034767151, + 0.02284950204193592, + 0.011812523938715458, + 0.04272257164120674, + 0.016679035499691963, + 0.010478970594704151, + 0.005493713077157736, + 0.008683232590556145, + 0.008380773477256298, + 0.008526064455509186, + 0.007479811552911997, + 0.02047727443277836, + 0.012824342586100101, + 0.004802567418664694, + 0.0038742274045944214, + 0.0066367401741445065, + 0.008766831830143929, + 0.006005911622196436, + 0.005884960759431124, + 0.006291148252785206, + 0.006779859308153391 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 3, + 2 + ], + "selected_reward": -0.054992545396089554 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04867646470665932 + ], + "opposite_logprob_tokens": [ + -3.0486764907836914 + ], + "logprob_reward": -0.04867646470665932, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04867646470665932, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.244186133146286, + 0.21335221827030182, + 0.11602739244699478, + 0.030875377357006073, + 0.1631634682416916, + 0.03274954855442047, + 0.023533103987574577, + 0.04444331303238869, + 0.05517231673002243, + 0.07649709284305573 + ], + "credit_map": [ + 0.03958074375987053, + 0.03662596642971039, + 0.023677198216319084, + 0.018742216750979424, + 0.04976888746023178, + 0.029921436682343483, + 0.02722558006644249, + 0.018644092604517937, + 0.031840577721595764, + 0.02431999333202839, + 0.02498946525156498, + 0.02761244587600231, + 0.030455105006694794, + 0.02274315431714058, + 0.028971314430236816, + 0.02242017164826393, + 0.02126544713973999, + 0.017564279958605766, + 0.006329734809696674, + 0.007519758306443691, + 0.027604440227150917, + 0.01550010684877634, + 0.011553093791007996, + 0.008690525777637959, + 0.005869103595614433, + 0.0034692189656198025, + 0.0025977345649152994, + 0.0033258977346122265, + 0.004740589763969183, + 0.00488261179998517, + 0.0028213621117174625, + 0.0031688592862337828, + 0.03181508556008339, + 0.01894683577120304, + 0.015757044777274132, + 0.01795138418674469, + 0.023405836895108223, + 0.02311098389327526, + 0.016255974769592285, + 0.01592031866312027, + 0.006138324737548828, + 0.0030348908621817827, + 0.003160487860441208, + 0.0028099024202674627, + 0.007105233613401651, + 0.0030194397550076246, + 0.004623453598469496, + 0.00285781710408628, + 0.004507740493863821, + 0.0018248589476570487, + 0.0030440317932516336, + 0.0029444729443639517, + 0.0029480070807039738, + 0.002181688556447625, + 0.004008289892226458, + 0.0020740136969834566, + 0.00884241983294487, + 0.0036849635653197765, + 0.0054156482219696045, + 0.00561649352312088, + 0.004247310105711222, + 0.003922132775187492, + 0.0073533132672309875, + 0.00536103080958128, + 0.007711643818765879, + 0.004043275956064463, + 0.004630255047231913, + 0.0055099292658269405, + 0.004487829282879829, + 0.007266731467097998, + 0.010801119729876518, + 0.010721536353230476, + 0.00785358902066946, + 0.006167827174067497, + 0.0051598260179162025, + 0.013608062639832497, + 0.007140859961509705, + 0.00901978462934494, + 0.01894471049308777, + 0.008602435700595379 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.04867646470665932 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0022047278471291065 + ], + "opposite_logprob_tokens": [ + -6.127204895019531 + ], + "logprob_reward": -0.0022047278471291065, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0022047278471291065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21654726564884186, + 0.07859747856855392, + 0.05640248581767082, + 0.10749494284391403, + 0.29466620087623596, + 0.03585745394229889, + 0.04074322432279587, + 0.044528547674417496, + 0.046697426587343216, + 0.07846494019031525 + ], + "credit_map": [ + 0.031714778393507004, + 0.03328818827867508, + 0.025592757388949394, + 0.018850117921829224, + 0.02904740907251835, + 0.03209691867232323, + 0.0215840395539999, + 0.024373086169362068, + 0.012201797217130661, + 0.010397477075457573, + 0.005868277978152037, + 0.008028805255889893, + 0.012843680568039417, + 0.010497194714844227, + 0.010561356320977211, + 0.008198898285627365, + 0.009092176333069801, + 0.005764506291598082, + 0.0037736331578344107, + 0.00458140904083848, + 0.013694524765014648, + 0.009711794555187225, + 0.004906264133751392, + 0.004878181964159012, + 0.017895646393299103, + 0.01712265983223915, + 0.006943325512111187, + 0.008894084021449089, + 0.026235457509756088, + 0.012991649098694324, + 0.008902641013264656, + 0.008509484119713306, + 0.05532708391547203, + 0.03552941977977753, + 0.02762630581855774, + 0.03183797746896744, + 0.04233674705028534, + 0.04377776011824608, + 0.028990739956498146, + 0.029240187257528305, + 0.006474565248936415, + 0.0034889476373791695, + 0.0029268928337842226, + 0.0028994055464863777, + 0.008458983153104782, + 0.0033207868691533804, + 0.005633786786347628, + 0.0026540893595665693, + 0.007065662648528814, + 0.002959093078970909, + 0.004847647622227669, + 0.004126404412090778, + 0.0061644259840250015, + 0.004905896261334419, + 0.0067382073029875755, + 0.003935891203582287, + 0.00701903598383069, + 0.0040529347024858, + 0.004521752241998911, + 0.0033816613722592592, + 0.00718352198600769, + 0.004296377766877413, + 0.00875730998814106, + 0.005315961316227913, + 0.007221872452646494, + 0.0033936561085283756, + 0.0034411733504384756, + 0.0044374121353030205, + 0.004984202329069376, + 0.006177456118166447, + 0.007135268300771713, + 0.009906387887895107, + 0.009414068423211575, + 0.004913410637527704, + 0.0096448278054595, + 0.014206425286829472, + 0.007142918650060892, + 0.00927997101098299, + 0.012739426456391811, + 0.011123902164399624 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 3 + ], + "selected_reward": -0.0022047278471291065 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810005635023117 + ], + "opposite_logprob_tokens": [ + -3.288100004196167 + ], + "logprob_reward": -0.03810005635023117, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810005635023117, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10266152024269104, + 0.026682306081056595, + 0.05901101231575012, + 0.1656413972377777, + 0.39438530802726746, + 0.05653993412852287, + 0.025074845179915428, + 0.04414815828204155, + 0.04733824357390404, + 0.07851723581552505 + ], + "credit_map": [ + 0.01557866670191288, + 0.014041678048670292, + 0.012384289875626564, + 0.00843687541782856, + 0.01713567040860653, + 0.010960929095745087, + 0.011018569581210613, + 0.01310485415160656, + 0.004549400415271521, + 0.0037293978966772556, + 0.0019920910708606243, + 0.002450943924486637, + 0.004522634670138359, + 0.003156165359541774, + 0.0031318068504333496, + 0.0031498668249696493, + 0.010103886015713215, + 0.0077017913572490215, + 0.004207472782582045, + 0.0039711808785796165, + 0.012966128997504711, + 0.009645809419453144, + 0.005849184934049845, + 0.004565561190247536, + 0.028867745772004128, + 0.02005586586892605, + 0.00896404217928648, + 0.012547089718282223, + 0.03814054653048515, + 0.024769550189375877, + 0.015057822689414024, + 0.01723875291645527, + 0.07105246931314468, + 0.04406784102320671, + 0.04194808751344681, + 0.04448065534234047, + 0.057433437556028366, + 0.05940160155296326, + 0.04204544052481651, + 0.033955808728933334, + 0.010203294456005096, + 0.005129798781126738, + 0.006115127820521593, + 0.005202499683946371, + 0.012192784808576107, + 0.004686507862061262, + 0.008625361137092113, + 0.004384559579193592, + 0.004277416039258242, + 0.0020849276334047318, + 0.002962918486446142, + 0.002260846784338355, + 0.0035960718523710966, + 0.0027357537765055895, + 0.004144303034991026, + 0.0030126096680760384, + 0.006458107382059097, + 0.0036492436192929745, + 0.004739874042570591, + 0.0048478031530976295, + 0.004629725124686956, + 0.004560744855552912, + 0.009182841517031193, + 0.006079820450395346, + 0.006294673308730125, + 0.003282176097854972, + 0.004998891614377499, + 0.004527595825493336, + 0.00439146114513278, + 0.006720882374793291, + 0.009163626469671726, + 0.007958939298987389, + 0.00912422314286232, + 0.005425649229437113, + 0.008606872521340847, + 0.012612386606633663, + 0.00778356846421957, + 0.010322587564587593, + 0.01399269886314869, + 0.01064925268292427 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 0 + ], + "selected_reward": -0.03810005635023117 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, long shot captured from an extremely low angle, the composition exhibits a near-perfect, unsettling symmetry. On a pitch-black asphalt street, thoroughly drenched by fresh rain, two classic black sedans stand head-to-head, motionless. Each with its high-intensity high beams on, emitting a cold, almost white glow, these two sharp beams pierce the humid night sky, illuminating the still-falling fine raindrops like dancing dust motes. The slippery road, like a giant black mirror, perfectly reflects the two piercing beams and the silhouettes of the cars, creating a double glare effect that intensifies the image's visual impact. Silent buildings, blurred in the darkness, line the street. With no other light sources, these two opposing beams become the sole focal point of the world. The stark contrast between light and darkness, combined with the absolute stillness, creates a tense and suspenseful atmosphere of confrontation.", + "seed": 3391, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_2/index_7011_seed_3391_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03834288700018078, + "vlm_logprob": -0.03834288700018078, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03834288700018078, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09763187170028687 + ], + "opposite_logprob_tokens": [ + -2.4726319313049316 + ], + "logprob_reward": -0.09763187170028687, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.09763187170028687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21582810580730438, + 0.06968074291944504, + 0.03442823886871338, + 0.09620239585638046, + 0.07383382320404053, + 0.347271591424942, + 0.044276174157857895, + 0.04250454157590866, + 0.05734079331159592, + 0.018633635714650154 + ], + "credit_map": [ + 0.02649375982582569, + 0.028146885335445404, + 0.019856853410601616, + 0.019146081060171127, + 0.029632827267050743, + 0.034581128507852554, + 0.03052082471549511, + 0.027449743822216988, + 0.008928801864385605, + 0.009793803095817566, + 0.006665297783911228, + 0.005605420097708702, + 0.011755217798054218, + 0.008783682249486446, + 0.00900584738701582, + 0.009142670780420303, + 0.0037240416277199984, + 0.004124781116843224, + 0.003452393226325512, + 0.003041773336008191, + 0.005394369829446077, + 0.0044042691588401794, + 0.005820579826831818, + 0.004466029349714518, + 0.009475184604525566, + 0.01868354342877865, + 0.008506509475409985, + 0.014002302661538124, + 0.011079239659011364, + 0.011379767209291458, + 0.01209932379424572, + 0.010976528748869896, + 0.008321892470121384, + 0.013882211409509182, + 0.006773197092115879, + 0.009204821661114693, + 0.00984216295182705, + 0.006240387912839651, + 0.00966927781701088, + 0.009899875149130821, + 0.03238889202475548, + 0.06240924447774887, + 0.029190510511398315, + 0.05094132572412491, + 0.042436886578798294, + 0.03766060993075371, + 0.04741242155432701, + 0.044831663370132446, + 0.005247851833701134, + 0.0056040240451693535, + 0.004810204263776541, + 0.005016650538891554, + 0.004835523199290037, + 0.005205648951232433, + 0.00613147858530283, + 0.0074247932061553, + 0.004753969144076109, + 0.004164331592619419, + 0.005097436252981424, + 0.003926933743059635, + 0.005750334355980158, + 0.00447107944637537, + 0.008125832304358482, + 0.006214626133441925, + 0.0059851291589438915, + 0.005642016883939505, + 0.005577771924436092, + 0.004915900062769651, + 0.008497199043631554, + 0.008947525173425674, + 0.010172323323786259, + 0.007602929603308439, + 0.0023978983517736197, + 0.0024799080565571785, + 0.0013810025993734598, + 0.001122345682233572, + 0.003915963228791952, + 0.0024965722113847733, + 0.0021420319098979235, + 0.0026979136746376753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 3 + ], + "selected_reward": -0.09763187170028687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02983962930738926 + ], + "opposite_logprob_tokens": [ + -3.529839515686035 + ], + "logprob_reward": -0.02983962930738926, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02983962930738926, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1559973508119583, + 0.04011336714029312, + 0.10757219046354294, + 0.05851757526397705, + 0.21252590417861938, + 0.12884359061717987, + 0.045847341418266296, + 0.06837683171033859, + 0.07010278105735779, + 0.11210308223962784 + ], + "credit_map": [ + 0.018312864005565643, + 0.022355208173394203, + 0.018247531726956367, + 0.013406873680651188, + 0.019619997590780258, + 0.02037218026816845, + 0.021725624799728394, + 0.021957075223326683, + 0.004111829213798046, + 0.003765955101698637, + 0.003711589379236102, + 0.005873186979442835, + 0.004921975079923868, + 0.0037814201787114143, + 0.008355671539902687, + 0.005591738037765026, + 0.010825689882040024, + 0.011935079470276833, + 0.010395747609436512, + 0.008085279725492, + 0.016704946756362915, + 0.014215433038771152, + 0.021705999970436096, + 0.013704011216759682, + 0.006610599346458912, + 0.0075607141479849815, + 0.00421808660030365, + 0.005936367437243462, + 0.01021923590451479, + 0.009797928854823112, + 0.006827037315815687, + 0.007347602862864733, + 0.02206486277282238, + 0.04005357623100281, + 0.015546162612736225, + 0.028644243255257607, + 0.023776009678840637, + 0.023849213495850563, + 0.033918533474206924, + 0.02467329613864422, + 0.012991968542337418, + 0.022929152473807335, + 0.010958461090922356, + 0.019492564722895622, + 0.014966909773647785, + 0.013098512776196003, + 0.01845659129321575, + 0.015949437394738197, + 0.005467158276587725, + 0.0057901665568351746, + 0.00515950471162796, + 0.004544515628367662, + 0.004898201208561659, + 0.005649909842759371, + 0.006380300968885422, + 0.007957584224641323, + 0.007279223296791315, + 0.006329821888357401, + 0.00832364708185196, + 0.0055784606374800205, + 0.008349379524588585, + 0.008791530504822731, + 0.013487781397998333, + 0.01023698877543211, + 0.006946835201233625, + 0.008470876142382622, + 0.006854929029941559, + 0.007939241826534271, + 0.012049009092152119, + 0.008055470883846283, + 0.010421228595077991, + 0.009365190751850605, + 0.01155495923012495, + 0.019749179482460022, + 0.00805431418120861, + 0.012979183346033096, + 0.017199605703353882, + 0.01542939804494381, + 0.012709841132164001, + 0.01442659366875887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.02983962930738926 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371884673833847 + ], + "opposite_logprob_tokens": [ + -3.4087188243865967 + ], + "logprob_reward": -0.03371884673833847, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371884673833847, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3848608732223511, + 0.10388258844614029, + 0.08968096971511841, + 0.05196133628487587, + 0.04355429485440254, + 0.1571015566587448, + 0.03142385929822922, + 0.05125967413187027, + 0.06373050063848495, + 0.022544322535395622 + ], + "credit_map": [ + 0.046517856419086456, + 0.074132539331913, + 0.036140501499176025, + 0.031006954610347748, + 0.052541207522153854, + 0.05390368774533272, + 0.03186335414648056, + 0.05875478684902191, + 0.011504912748932838, + 0.018714018166065216, + 0.009557741694152355, + 0.008738771080970764, + 0.01516193337738514, + 0.012469610199332237, + 0.011393788270652294, + 0.01634182035923004, + 0.008227289654314518, + 0.015408138744533062, + 0.008232973515987396, + 0.011483876965939999, + 0.009908419102430344, + 0.009267237037420273, + 0.015311993658542633, + 0.01184104848653078, + 0.004990553483366966, + 0.006373680662363768, + 0.004161733668297529, + 0.0073044179007411, + 0.005833090748637915, + 0.006242862902581692, + 0.00954663660377264, + 0.007508364040404558, + 0.004046189598739147, + 0.006503259297460318, + 0.005131034180521965, + 0.0051642972975969315, + 0.005648579448461533, + 0.002945901593193412, + 0.007183466572314501, + 0.006931567098945379, + 0.014372526668012142, + 0.02743537910282612, + 0.01321254763752222, + 0.02233572117984295, + 0.01938319206237793, + 0.01835370622575283, + 0.020715530961751938, + 0.021292971447110176, + 0.003881345735862851, + 0.0035727585200220346, + 0.003218765137717128, + 0.0029009224381297827, + 0.004154818132519722, + 0.004103725776076317, + 0.004330786876380444, + 0.005260743200778961, + 0.005524736829102039, + 0.0050348155200481415, + 0.006087278015911579, + 0.004330302122980356, + 0.006996653042733669, + 0.0062868632376194, + 0.00956906471401453, + 0.007429968100041151, + 0.006631885655224323, + 0.006528670433908701, + 0.0059723095037043095, + 0.0054358369670808315, + 0.009693422354757786, + 0.009564343839883804, + 0.011402906849980354, + 0.008501133881509304, + 0.0027232482098042965, + 0.0027230919804424047, + 0.0019950871355831623, + 0.0018342830007895827, + 0.003607757156714797, + 0.0037430364172905684, + 0.002826348412781954, + 0.0030914717353880405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.03371884673833847 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007232195930555463 + ], + "opposite_logprob_tokens": [ + -7.250723361968994 + ], + "logprob_reward": -0.0007232195930555463, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007232195930555463, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17056426405906677, + 0.06018562614917755, + 0.0630243793129921, + 0.05130995064973831, + 0.22522717714309692, + 0.1635000854730606, + 0.051424648612737656, + 0.1431291252374649, + 0.04463732987642288, + 0.02699742279946804 + ], + "credit_map": [ + 0.02008933760225773, + 0.020247606560587883, + 0.01439631637185812, + 0.013500098139047623, + 0.028742540627717972, + 0.02497353032231331, + 0.024032246321439743, + 0.024582594633102417, + 0.005491008050739765, + 0.006904708221554756, + 0.005738785024732351, + 0.008376587182283401, + 0.008595168590545654, + 0.005611710250377655, + 0.011565832421183586, + 0.00790182501077652, + 0.006649516057223082, + 0.007752056233584881, + 0.004788223188370466, + 0.007433581165969372, + 0.011431563645601273, + 0.008585610426962376, + 0.007886877283453941, + 0.008496950380504131, + 0.005208903457969427, + 0.009239839389920235, + 0.003457484534010291, + 0.0057216063141822815, + 0.009161066263914108, + 0.004853384103626013, + 0.005941953510046005, + 0.007725711911916733, + 0.023792972788214684, + 0.04294794425368309, + 0.017904354259371758, + 0.029447482898831367, + 0.026539087295532227, + 0.025775132700800896, + 0.03126533329486847, + 0.027554864063858986, + 0.016300607472658157, + 0.027357349172234535, + 0.011207206174731255, + 0.018866071477532387, + 0.02269946038722992, + 0.02264476753771305, + 0.02321692369878292, + 0.021207714453339577, + 0.00439962325617671, + 0.0066437856294214725, + 0.00477241724729538, + 0.007352047134190798, + 0.00690503092482686, + 0.004995241761207581, + 0.009814304299652576, + 0.0065421974286437035, + 0.013480444438755512, + 0.015306474640965462, + 0.010263185948133469, + 0.010766519233584404, + 0.02751842327415943, + 0.026081133633852005, + 0.02176850661635399, + 0.01794443465769291, + 0.004373371601104736, + 0.004504023585468531, + 0.004260391928255558, + 0.0042336806654930115, + 0.006780513562262058, + 0.006791671738028526, + 0.008043838664889336, + 0.005649838130921125, + 0.00295218825340271, + 0.005745620932430029, + 0.002292656572535634, + 0.002858357969671488, + 0.003606769721955061, + 0.0035197485703974962, + 0.0024376153014600277, + 0.003584466176107526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.0007232195930555463 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Are two classic black sedans visible standing head-to-head on the street?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the street surface pitch-black asphalt thoroughly drenched by fresh rain?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do both cars have their high-intensity high beams on, emitting a cold white glow?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the road surface reflect the car beams and car silhouettes to create a mirror-like double glare effect?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot captured from an extremely low angle with a static position?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029800867661833763 + ], + "opposite_logprob_tokens": [ + -3.5298008918762207 + ], + "logprob_reward": -0.029800867661833763, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029800867661833763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15615350008010864, + 0.03516345098614693, + 0.047175049781799316, + 0.0782330185174942, + 0.22358110547065735, + 0.33475613594055176, + 0.02707701548933983, + 0.019355805590748787, + 0.01461752224713564, + 0.06388738006353378 + ], + "credit_map": [ + 0.018572697415947914, + 0.01935359090566635, + 0.020029854029417038, + 0.012944286689162254, + 0.021882761269807816, + 0.018695464357733727, + 0.01891152560710907, + 0.025763317942619324, + 0.0029632183723151684, + 0.004761734511703253, + 0.003387691453099251, + 0.0058383881114423275, + 0.004212749190628529, + 0.0027197888121008873, + 0.0060889143496751785, + 0.0051909638568758965, + 0.00528969569131732, + 0.0057144747115671635, + 0.003752158721908927, + 0.004568392876535654, + 0.009098796173930168, + 0.006835442967712879, + 0.005601579789072275, + 0.006314509082585573, + 0.00813381839543581, + 0.011875038966536522, + 0.005303783342242241, + 0.008166447281837463, + 0.014361795969307423, + 0.01002304907888174, + 0.009150642901659012, + 0.011218440718948841, + 0.02410029247403145, + 0.04336891695857048, + 0.018817953765392303, + 0.028561530634760857, + 0.027309665456414223, + 0.023268025368452072, + 0.030521992594003677, + 0.02763272449374199, + 0.030559007078409195, + 0.05705884471535683, + 0.024776680395007133, + 0.04731450974941254, + 0.03918437287211418, + 0.043821629136800766, + 0.05018359422683716, + 0.041857507079839706, + 0.002267728094011545, + 0.0032377808820456266, + 0.0029506213031709194, + 0.0027616533916443586, + 0.00363580952398479, + 0.003528211498633027, + 0.004997518844902515, + 0.003697692882269621, + 0.0019598393701016903, + 0.0021386421285569668, + 0.0016949968412518501, + 0.0017017959617078304, + 0.00404403917491436, + 0.0024931826628744602, + 0.0023367840331047773, + 0.002986524486914277, + 0.0013418779708445072, + 0.00240528816357255, + 0.001661202171817422, + 0.00314746773801744, + 0.0014087039744481444, + 0.0009865324245765805, + 0.001856106799095869, + 0.0018103433540090919, + 0.006568515673279762, + 0.01097386609762907, + 0.004809841047972441, + 0.0072634476236999035, + 0.00964339729398489, + 0.009225830435752869, + 0.007266870234161615, + 0.00813561026006937 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.029800867661833763 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_tensor([7068, 7068])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_tensor([7068, 7068])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3a7cb8189a508e679a3c69a2a4d54a9bd8b21c60 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_tensor([7068, 7068])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "realistic, eye-level medium shot, held perfectly still, captures a blonde woman partially submerged in a swimming pool, emanating a profound sense of peace. Her long, honey-blond hair, wet and clinging to her bare shoulders, floats like silk in the crystal-clear water. Her eyes are gently closed, a subtle, serene smile playing on her lips, a picture of blissful relaxation. Submerged to her shoulders, her arms resting lightly on the surface, creating the faintest ripples. The brilliant midday sun bathes the landscape, causing countless tiny water droplets on her sun-soaked skin to sparkle like diamonds. The water is so transparent that beneath her, the small, iridescent blue and turquoise mosaic tiles lining the pool can be clearly seen, their colors shimmering and distorted by dancing, corroded patterns of light. The soft-focus composition of the clean, pale stone poolside deck in the background situates the scene in a luxurious, private setting. The cool blue of the water, the warm tones of her skin, and the sunlight dominate the composition, emanating a tranquil, almost meditative atmosphere.", + "seed": 29949, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29949_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.260662748478353, + "vlm_logprob": -0.260662748478353, + "vlm_yes_no_margin": 2.775, + "vlm_token_credit": -0.260662748478353, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05218476802110672 + ], + "opposite_logprob_tokens": [ + -4.302184581756592 + ], + "logprob_reward": -0.05218476802110672, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.05218476802110672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1431732326745987, + 0.07809147983789444, + 0.058269258588552475, + 0.1710842102766037, + 0.12917868793010712, + 0.06544768065214157, + 0.05849698558449745, + 0.08329440653324127, + 0.1643269807100296, + 0.04863707721233368 + ], + "credit_map": [ + 0.023246532306075096, + 0.013552767224609852, + 0.015324639156460762, + 0.01302819699048996, + 0.018084706738591194, + 0.018565712496638298, + 0.01951432041823864, + 0.021856343373656273, + 0.007430103607475758, + 0.011309032328426838, + 0.007860216312110424, + 0.005560994613915682, + 0.016201060265302658, + 0.018340645357966423, + 0.006032305769622326, + 0.005357121117413044, + 0.009760832414031029, + 0.00915342103689909, + 0.005225833039730787, + 0.005596523638814688, + 0.008998367935419083, + 0.00632482161745429, + 0.006691860966384411, + 0.006517595145851374, + 0.018360480666160583, + 0.01653476245701313, + 0.01356988213956356, + 0.00979453045874834, + 0.03918318450450897, + 0.03764023259282112, + 0.013184644281864166, + 0.022816484794020653, + 0.014613532461225986, + 0.009607630781829357, + 0.0109002236276865, + 0.006774535868316889, + 0.02370435930788517, + 0.032536983489990234, + 0.011895477771759033, + 0.01914595253765583, + 0.014070210047066212, + 0.004810145124793053, + 0.006443742662668228, + 0.00558985723182559, + 0.010267228819429874, + 0.01109407376497984, + 0.005641512107104063, + 0.007530911825597286, + 0.00771297188475728, + 0.006093502510339022, + 0.008293170481920242, + 0.004840395413339138, + 0.007400089874863625, + 0.007958739064633846, + 0.005479777697473764, + 0.0107183326035738, + 0.00981365330517292, + 0.012320023961365223, + 0.005670515354722738, + 0.007314055226743221, + 0.01758752390742302, + 0.013832221738994122, + 0.009906002320349216, + 0.006850406527519226, + 0.01644660159945488, + 0.010800345800817013, + 0.009089173749089241, + 0.006806093733757734, + 0.05151527002453804, + 0.0502367727458477, + 0.01048676110804081, + 0.008945971727371216, + 0.009189777076244354, + 0.004970700014382601, + 0.004061718005686998, + 0.0033563936594873667, + 0.008558922447264194, + 0.008668110705912113, + 0.004639778286218643, + 0.005191677715629339 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 0 + ], + "selected_reward": -0.05218476802110672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014219510369002819 + ], + "opposite_logprob_tokens": [ + -4.264219284057617 + ], + "logprob_reward": -0.014219510369002819, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014219510369002819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04826847463846207, + 0.1737445741891861, + 0.19437220692634583, + 0.19134403765201569, + 0.07698701322078705, + 0.03910704329609871, + 0.044822510331869125, + 0.10203436762094498, + 0.054713018238544464, + 0.07460677623748779 + ], + "credit_map": [ + 0.008200468495488167, + 0.005224897526204586, + 0.006597575731575489, + 0.004633508622646332, + 0.007924548350274563, + 0.004863561596721411, + 0.005252856761217117, + 0.005571055691689253, + 0.0170296523720026, + 0.014057193882763386, + 0.0175149105489254, + 0.012282310053706169, + 0.03630635887384415, + 0.040325358510017395, + 0.016992930322885513, + 0.019235869869589806, + 0.020432937890291214, + 0.01399502344429493, + 0.02024497464299202, + 0.012944785878062248, + 0.03365425392985344, + 0.03703392297029495, + 0.024448825046420097, + 0.03161747753620148, + 0.017103854566812515, + 0.016503261402249336, + 0.014120805077254772, + 0.010014591738581657, + 0.049653515219688416, + 0.046004101634025574, + 0.01726650632917881, + 0.020677410066127777, + 0.008713170886039734, + 0.009325358085334301, + 0.007738311775028706, + 0.010166588239371777, + 0.010850036516785622, + 0.012434679083526134, + 0.008822577074170113, + 0.00893628504127264, + 0.009555705823004246, + 0.0035531872417777777, + 0.004202740266919136, + 0.002611748641356826, + 0.005474989302456379, + 0.006074060685932636, + 0.0032409215345978737, + 0.004393688403069973, + 0.005176486913114786, + 0.0063480194658041, + 0.0044346945360302925, + 0.004882659297436476, + 0.007120135705918074, + 0.007104980759322643, + 0.004551031161099672, + 0.0052044992335140705, + 0.01065711211413145, + 0.008530297316610813, + 0.008869302459061146, + 0.006348194554448128, + 0.021575620397925377, + 0.024631470441818237, + 0.009194386191666126, + 0.012227986939251423, + 0.0058359187096357346, + 0.005043834913522005, + 0.004084015730768442, + 0.003666345262899995, + 0.013071177527308464, + 0.011557314544916153, + 0.005605197511613369, + 0.005849217530339956, + 0.00863427110016346, + 0.0038911066949367523, + 0.005810770206153393, + 0.0033844239078462124, + 0.02181055210530758, + 0.021955866366624832, + 0.0031055754516273737, + 0.0060142031870782375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.014219510369002819 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02979855239391327 + ], + "opposite_logprob_tokens": [ + -3.5297985076904297 + ], + "logprob_reward": -0.02979855239391327, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02979855239391327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16912873089313507, + 0.13297796249389648, + 0.03650755435228348, + 0.09575597196817398, + 0.053022630512714386, + 0.19186589121818542, + 0.1905355155467987, + 0.045236267149448395, + 0.037934672087430954, + 0.04703481122851372 + ], + "credit_map": [ + 0.02189207822084427, + 0.009935599751770496, + 0.01427372358739376, + 0.009057893417775631, + 0.04004115238785744, + 0.04156329855322838, + 0.013245666399598122, + 0.019119305536150932, + 0.015946606174111366, + 0.012118383310735226, + 0.015889769420027733, + 0.009213387966156006, + 0.021706588566303253, + 0.0247805155813694, + 0.015846341848373413, + 0.01747637428343296, + 0.008196108043193817, + 0.003004017286002636, + 0.0038674643728882074, + 0.0028579041827470064, + 0.005920969881117344, + 0.006447685416787863, + 0.002783513627946377, + 0.0034298941027373075, + 0.013401555828750134, + 0.00877672154456377, + 0.00966207217425108, + 0.0058918423019349575, + 0.018558157607913017, + 0.02130197361111641, + 0.007511645555496216, + 0.010652007535099983, + 0.008059365674853325, + 0.005501954350620508, + 0.006156222429126501, + 0.005009673535823822, + 0.009301598183810711, + 0.010052529163658619, + 0.004228153731673956, + 0.004713134840130806, + 0.017650194466114044, + 0.02509859949350357, + 0.01976824551820755, + 0.024386674165725708, + 0.031148161739110947, + 0.025467760860919952, + 0.0207839235663414, + 0.02756233513355255, + 0.019531240686774254, + 0.027826417237520218, + 0.014912335202097893, + 0.020160071551799774, + 0.0358441062271595, + 0.03266817331314087, + 0.0195763036608696, + 0.020016873255372047, + 0.0038753331173211336, + 0.005297081079334021, + 0.005211237818002701, + 0.0033754240721464157, + 0.008963231928646564, + 0.007388236001133919, + 0.00445328326895833, + 0.00667243916541338, + 0.004994294140487909, + 0.002914843615144491, + 0.004927666857838631, + 0.002393609844148159, + 0.005110111087560654, + 0.006285555195063353, + 0.0036448847968131304, + 0.0076637063175439835, + 0.005998501554131508, + 0.002853068755939603, + 0.0031435799319297075, + 0.0021443720906972885, + 0.012738202698528767, + 0.013073954731225967, + 0.0030047937761992216, + 0.004078337922692299 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 0 + ], + "selected_reward": -0.02979855239391327 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.1370902061462402 + ], + "opposite_logprob_tokens": [ + -0.3870902359485626 + ], + "logprob_reward": -1.1370902061462402, + "yes_no_margin_reward": -0.75, + "token_credit_reward": -1.1370902061462402, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.15069353580474854, + 0.0631832405924797, + 0.07688525319099426, + 0.04959362372756004, + 0.11807212978601456, + 0.08079054951667786, + 0.02884599007666111, + 0.13298699259757996, + 0.16801974177360535, + 0.13092893362045288 + ], + "credit_map": [ + 0.02734239026904106, + 0.016152214258909225, + 0.023415440693497658, + 0.014963784255087376, + 0.018639463931322098, + 0.0160993579775095, + 0.013916932977735996, + 0.02016395516693592, + 0.009212394244968891, + 0.007754892576485872, + 0.0055292691104114056, + 0.005355435889214277, + 0.012636137194931507, + 0.009188040159642696, + 0.00709571735933423, + 0.006411351729184389, + 0.013557979837059975, + 0.007681545801460743, + 0.006045874208211899, + 0.004043246153742075, + 0.016129767522215843, + 0.014389485120773315, + 0.0065719387494027615, + 0.0084654176607728, + 0.0058707077987492085, + 0.005592618137598038, + 0.007098629139363766, + 0.006700955796986818, + 0.0038075551856309175, + 0.008178425952792168, + 0.005024873185902834, + 0.007319858763366938, + 0.013702909462153912, + 0.008173272013664246, + 0.009174241684377193, + 0.006902073044329882, + 0.02387939766049385, + 0.03141668811440468, + 0.009763996116816998, + 0.015059544704854488, + 0.012399828992784023, + 0.009966990910470486, + 0.008221643045544624, + 0.010899079963564873, + 0.012567157857120037, + 0.011988354846835136, + 0.008214679546654224, + 0.006532812025398016, + 0.0077261971309781075, + 0.0034983542282134295, + 0.002879568375647068, + 0.0025122566148638725, + 0.0024775578640401363, + 0.0032810410484671593, + 0.002413493348285556, + 0.004057521466165781, + 0.01940603367984295, + 0.009827838279306889, + 0.013110755011439323, + 0.008371766656637192, + 0.02162863127887249, + 0.030551712960004807, + 0.00985539611428976, + 0.0202348530292511, + 0.015052590519189835, + 0.013280444778501987, + 0.009143353439867496, + 0.009175695478916168, + 0.04814585670828819, + 0.045526571571826935, + 0.014541959390044212, + 0.01315326802432537, + 0.012796496972441673, + 0.013516223058104515, + 0.008450211025774479, + 0.00998678058385849, + 0.03445182368159294, + 0.03254812955856323, + 0.010212008841335773, + 0.0089672626927495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -1.1370902061462402 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002070546150208 + ], + "opposite_logprob_tokens": [ + -2.6950206756591797 + ], + "logprob_reward": -0.07002070546150208, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002070546150208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10174136608839035, + 0.07478149980306625, + 0.04268195852637291, + 0.10369464755058289, + 0.135331392288208, + 0.15920096635818481, + 0.13595466315746307, + 0.140409916639328, + 0.0333167240023613, + 0.07288684695959091 + ], + "credit_map": [ + 0.023985987529158592, + 0.008467433042824268, + 0.00857415609061718, + 0.008129212073981762, + 0.019241951406002045, + 0.01616002805531025, + 0.009406225755810738, + 0.007776367478072643, + 0.006756597198545933, + 0.007993391714990139, + 0.007462885696440935, + 0.005798365920782089, + 0.015525264665484428, + 0.013634049333631992, + 0.008091602474451065, + 0.009519343264400959, + 0.006665832828730345, + 0.0036662588827311993, + 0.004508570767939091, + 0.0034380503930151463, + 0.006444920785725117, + 0.006032193545252085, + 0.00494451355189085, + 0.006981615908443928, + 0.011547453701496124, + 0.008712123148143291, + 0.00719413859769702, + 0.00465399632230401, + 0.026637613773345947, + 0.02570045366883278, + 0.007705019321292639, + 0.011543849483132362, + 0.02925282157957554, + 0.010436571203172207, + 0.012970800511538982, + 0.007381439208984375, + 0.01838568039238453, + 0.030014976859092712, + 0.009098116308450699, + 0.017790980637073517, + 0.012224715203046799, + 0.018498599529266357, + 0.011370370164513588, + 0.017625268548727036, + 0.03459766134619713, + 0.026642857119441032, + 0.016105690971016884, + 0.022135786712169647, + 0.021238751709461212, + 0.014915045350790024, + 0.010409963317215443, + 0.013339363038539886, + 0.024680420756340027, + 0.022220799699425697, + 0.014057787135243416, + 0.01509254053235054, + 0.014065315015614033, + 0.013876436278223991, + 0.012395737692713737, + 0.010342809371650219, + 0.02889779768884182, + 0.02777109295129776, + 0.014596023596823215, + 0.018464699387550354, + 0.004022435750812292, + 0.002458882285282016, + 0.00203633401542902, + 0.001536609954200685, + 0.00892663188278675, + 0.009413053281605244, + 0.002638649195432663, + 0.0022841268219053745, + 0.006491756532341242, + 0.0040298872627317905, + 0.004630405921489, + 0.0029555417131632566, + 0.022888358682394028, + 0.022753169760107994, + 0.004223215859383345, + 0.004914513789117336 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 7, + 6 + ], + "selected_reward": -0.07002070546150208 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "realistic, eye-level medium shot, held perfectly still, captures a blonde woman partially submerged in a swimming pool, emanating a profound sense of peace. Her long, honey-blond hair, wet and clinging to her bare shoulders, floats like silk in the crystal-clear water. Her eyes are gently closed, a subtle, serene smile playing on her lips, a picture of blissful relaxation. Submerged to her shoulders, her arms resting lightly on the surface, creating the faintest ripples. The brilliant midday sun bathes the landscape, causing countless tiny water droplets on her sun-soaked skin to sparkle like diamonds. The water is so transparent that beneath her, the small, iridescent blue and turquoise mosaic tiles lining the pool can be clearly seen, their colors shimmering and distorted by dancing, corroded patterns of light. The soft-focus composition of the clean, pale stone poolside deck in the background situates the scene in a luxurious, private setting. The cool blue of the water, the warm tones of her skin, and the sunlight dominate the composition, emanating a tranquil, almost meditative atmosphere.", + "seed": 29950, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29950_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2070893682539463, + "vlm_logprob": -0.2070893682539463, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.2070893682539463, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04478086158633232 + ], + "opposite_logprob_tokens": [ + -4.544780731201172 + ], + "logprob_reward": -0.04478086158633232, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.04478086158633232, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28642964363098145, + 0.17378342151641846, + 0.06896553188562393, + 0.08489639312028885, + 0.08699911832809448, + 0.05699839070439339, + 0.052785903215408325, + 0.09699927270412445, + 0.05295267328619957, + 0.03918964043259621 + ], + "credit_map": [ + 0.03797505795955658, + 0.016937952488660812, + 0.022160883992910385, + 0.016987662762403488, + 0.02673867717385292, + 0.08720234781503677, + 0.022206051275134087, + 0.05622101575136185, + 0.019527709111571312, + 0.012602831237018108, + 0.02604762837290764, + 0.013637937605381012, + 0.013575606979429722, + 0.04525301605463028, + 0.01356170978397131, + 0.029576990753412247, + 0.006175922695547342, + 0.004467426333576441, + 0.007417961489409208, + 0.0053384373895823956, + 0.008319689892232418, + 0.01967713236808777, + 0.007747884839773178, + 0.009821075946092606, + 0.006793666165322065, + 0.004772086162120104, + 0.017570549622178078, + 0.009650648571550846, + 0.008799776434898376, + 0.01509016565978527, + 0.006451239809393883, + 0.0157682616263628, + 0.010977005586028099, + 0.013070785440504551, + 0.013909174129366875, + 0.011898430064320564, + 0.01147288829088211, + 0.009174488484859467, + 0.007073101121932268, + 0.009423237293958664, + 0.004680377896875143, + 0.0042535909451544285, + 0.005278856959193945, + 0.008030399680137634, + 0.012830590829253197, + 0.007850057445466518, + 0.007380948401987553, + 0.006693567149341106, + 0.008351325988769531, + 0.007102197967469692, + 0.00569726200774312, + 0.006466797087341547, + 0.009392033331096172, + 0.00536929490044713, + 0.006140017416328192, + 0.004266973119229078, + 0.013297095894813538, + 0.012801874428987503, + 0.0171769168227911, + 0.01126455795019865, + 0.012230630964040756, + 0.010446231812238693, + 0.008809560909867287, + 0.01097240298986435, + 0.010523511096835136, + 0.011708725243806839, + 0.003989086020737886, + 0.006159448996186256, + 0.006965578068047762, + 0.006087305955588818, + 0.004002866800874472, + 0.0035161548294126987, + 0.007474575657397509, + 0.008313184604048729, + 0.003806260647252202, + 0.0038038676138967276, + 0.0048012458719313145, + 0.004295912571251392, + 0.00392325920984149, + 0.0027713344898074865 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.04478086158633232 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016087606549263 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.016087606549263, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016087606549263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2519989311695099, + 0.18023520708084106, + 0.1507243812084198, + 0.070912204682827, + 0.06001100689172745, + 0.07355427742004395, + 0.04989366978406906, + 0.05325504392385483, + 0.042447466403245926, + 0.06696774065494537 + ], + "credit_map": [ + 0.023656152188777924, + 0.014418894425034523, + 0.013398193754255772, + 0.015664951875805855, + 0.024518534541130066, + 0.09040651470422745, + 0.01892685890197754, + 0.05100882053375244, + 0.02052665688097477, + 0.014790434390306473, + 0.020415976643562317, + 0.013254298828542233, + 0.014333024621009827, + 0.05441974848508835, + 0.013128925114870071, + 0.02936614863574505, + 0.010383968241512775, + 0.008723504841327667, + 0.010357088409364223, + 0.008162308484315872, + 0.01409728080034256, + 0.0552092008292675, + 0.01254891138523817, + 0.0312421265989542, + 0.005738350097090006, + 0.005063669290393591, + 0.009537680074572563, + 0.007672320585697889, + 0.010696678422391415, + 0.01000118162482977, + 0.01097356528043747, + 0.011228756979107857, + 0.007119604852050543, + 0.006126023828983307, + 0.008505935780704021, + 0.007986163720488548, + 0.01103515550494194, + 0.006115308962762356, + 0.007548747584223747, + 0.005574065260589123, + 0.010461685247719288, + 0.006240061484277248, + 0.009015164338052273, + 0.007385233882814646, + 0.013511540368199348, + 0.007719679735600948, + 0.010749160312116146, + 0.008471755310893059, + 0.005795420613139868, + 0.007036269176751375, + 0.004729981534183025, + 0.005797470919787884, + 0.009471438825130463, + 0.006034334655851126, + 0.0057735866867005825, + 0.0052551687695086, + 0.007474404294043779, + 0.004901478067040443, + 0.007822403684258461, + 0.005767017602920532, + 0.007755514234304428, + 0.006636830512434244, + 0.005601617973297834, + 0.007295778952538967, + 0.008306900039315224, + 0.006336160469800234, + 0.0038647593464702368, + 0.004044699482619762, + 0.006668105721473694, + 0.0043835388496518135, + 0.005529248155653477, + 0.003314054338261485, + 0.008310830220580101, + 0.010159758850932121, + 0.005717151798307896, + 0.007397543638944626, + 0.012680824846029282, + 0.00652211531996727, + 0.009195149876177311, + 0.006984371226280928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.016087606549263 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014194242656230927 + ], + "opposite_logprob_tokens": [ + -4.264194011688232 + ], + "logprob_reward": -0.014194242656230927, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014194242656230927, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3319804072380066, + 0.1336967647075653, + 0.04881493002176285, + 0.06412289291620255, + 0.026702385395765305, + 0.07185734063386917, + 0.04624917358160019, + 0.09392181783914566, + 0.05751228332519531, + 0.12514200806617737 + ], + "credit_map": [ + 0.028290532529354095, + 0.018070202320814133, + 0.037580329924821854, + 0.02309330739080906, + 0.028371622785925865, + 0.11361029744148254, + 0.024885164573788643, + 0.058078959584236145, + 0.013643736951053143, + 0.008853662759065628, + 0.029666800051927567, + 0.014552823267877102, + 0.01208207942545414, + 0.02491033263504505, + 0.009776851162314415, + 0.020210472866892815, + 0.004086730070412159, + 0.003971160855144262, + 0.00696626678109169, + 0.005467712879180908, + 0.00509186927229166, + 0.011002295650541782, + 0.005201851483434439, + 0.007027043495327234, + 0.005595289636403322, + 0.0039368318393826485, + 0.009674192406237125, + 0.006775721907615662, + 0.005229146685451269, + 0.014952084049582481, + 0.005208146292716265, + 0.012751477770507336, + 0.004105899482965469, + 0.0027254882734268904, + 0.004789840430021286, + 0.0030972675886005163, + 0.00368001707829535, + 0.0025737411342561245, + 0.0023624892346560955, + 0.003367641009390354, + 0.00831276923418045, + 0.006245218217372894, + 0.011038863100111485, + 0.0077031273394823074, + 0.01085702981799841, + 0.013136244378983974, + 0.0061098854057490826, + 0.008454198017716408, + 0.008147437125444412, + 0.010453042574226856, + 0.003941203001886606, + 0.005762598942965269, + 0.00715595344081521, + 0.004400456324219704, + 0.0031595376785844564, + 0.0032289426308125257, + 0.016541801393032074, + 0.013096528127789497, + 0.012863061390817165, + 0.008933267556130886, + 0.010571933351457119, + 0.011450053192675114, + 0.007997505366802216, + 0.012467659078538418, + 0.010619480162858963, + 0.009690560400485992, + 0.0069192261435091496, + 0.006364498287439346, + 0.008482064120471478, + 0.006429519038647413, + 0.004409315530210733, + 0.004597620572894812, + 0.02480555884540081, + 0.027758272364735603, + 0.013685240410268307, + 0.01369806844741106, + 0.01184302382171154, + 0.014927355572581291, + 0.008756529539823532, + 0.009667966514825821 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.014194242656230927 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8983446955680847 + ], + "opposite_logprob_tokens": [ + -0.5233446955680847 + ], + "logprob_reward": -0.8983446955680847, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.8983446955680847, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.17201872169971466, + 0.17173625528812408, + 0.046580709517002106, + 0.24701514840126038, + 0.08727520704269409, + 0.06787943840026855, + 0.05472905933856964, + 0.044348232448101044, + 0.07355111092329025, + 0.034866102039813995 + ], + "credit_map": [ + 0.014674447476863861, + 0.012141621671617031, + 0.013758021406829357, + 0.019583387300372124, + 0.021745989099144936, + 0.04345044121146202, + 0.01814972795546055, + 0.028515083715319633, + 0.014385300688445568, + 0.011145933531224728, + 0.013436899520456791, + 0.008090361952781677, + 0.011582520790398121, + 0.06659458577632904, + 0.008692645467817783, + 0.03780800476670265, + 0.006000131834298372, + 0.003773226635530591, + 0.008867008611559868, + 0.005270258989185095, + 0.0065561020746827126, + 0.0038684906903654337, + 0.006740369368344545, + 0.005505124572664499, + 0.0205305777490139, + 0.016398854553699493, + 0.04662514850497246, + 0.022948868572711945, + 0.01942610926926136, + 0.060131318867206573, + 0.01332435104995966, + 0.047629911452531815, + 0.011582572013139725, + 0.01314123161137104, + 0.012475695461034775, + 0.010475675575435162, + 0.009493938647210598, + 0.012369432486593723, + 0.0071791778318583965, + 0.010557486675679684, + 0.008164073340594769, + 0.0062075452879071236, + 0.008113471791148186, + 0.007926208898425102, + 0.012873291037976742, + 0.009531373158097267, + 0.007381038274616003, + 0.007682431489229202, + 0.0050754486583173275, + 0.004813201725482941, + 0.0058927880600094795, + 0.006358899176120758, + 0.009779590182006359, + 0.005207774229347706, + 0.008006107993423939, + 0.009595249779522419, + 0.006317171733826399, + 0.0056363665498793125, + 0.007177900988608599, + 0.004740109201520681, + 0.007461307104676962, + 0.004823349881917238, + 0.003617093898355961, + 0.004574933089315891, + 0.015236045233905315, + 0.014498118311166763, + 0.009643116034567356, + 0.008096493780612946, + 0.009028222411870956, + 0.007725321222096682, + 0.004769507329910994, + 0.004554285202175379, + 0.005225612316280603, + 0.004591937642544508, + 0.005166910123080015, + 0.005070853047072887, + 0.004014104604721069, + 0.0038724520709365606, + 0.0031626953277736902, + 0.0037615341134369373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.8983446955680847 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203943490982056 + ], + "opposite_logprob_tokens": [ + -2.812039375305176 + ], + "logprob_reward": -0.06203943490982056, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203943490982056, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10923561453819275, + 0.09335270524024963, + 0.1812368631362915, + 0.074368417263031, + 0.061528850346803665, + 0.1916578859090805, + 0.0491114966571331, + 0.09738190472126007, + 0.014194337651133537, + 0.12793195247650146 + ], + "credit_map": [ + 0.017139578238129616, + 0.007473601959645748, + 0.01292124018073082, + 0.012124059721827507, + 0.015563232824206352, + 0.0152249401435256, + 0.01578798145055771, + 0.013000971637666225, + 0.011204141192138195, + 0.006588224787265062, + 0.015705851837992668, + 0.00846041925251484, + 0.009213397279381752, + 0.0206723902374506, + 0.008428509347140789, + 0.013079769909381866, + 0.02948354184627533, + 0.01966400071978569, + 0.04533560946583748, + 0.018035506829619408, + 0.015308380126953125, + 0.01496982853859663, + 0.012901460751891136, + 0.025538524612784386, + 0.00789379421621561, + 0.0066018858924508095, + 0.013351785950362682, + 0.006427471991628408, + 0.004484404344111681, + 0.018676625564694405, + 0.003765273606404662, + 0.01316717080771923, + 0.0093808239325881, + 0.00722649646922946, + 0.010961180552840233, + 0.008529008366167545, + 0.007150175049901009, + 0.006970548536628485, + 0.005189894698560238, + 0.006120720878243446, + 0.02398429811000824, + 0.019911931827664375, + 0.027241433039307594, + 0.019981997087597847, + 0.026880748569965363, + 0.030661890283226967, + 0.019019953906536102, + 0.023975644260644913, + 0.007901391014456749, + 0.0045531149953603745, + 0.009637902490794659, + 0.004048430360853672, + 0.005012288223952055, + 0.006427354644984007, + 0.004623691085726023, + 0.006907323375344276, + 0.019081655889749527, + 0.007326880935579538, + 0.023096084594726562, + 0.00918380357325077, + 0.009800863452255726, + 0.012303228490054607, + 0.007650682702660561, + 0.008938703685998917, + 0.002525320742279291, + 0.0012704954715445638, + 0.001825646380893886, + 0.0013869006652384996, + 0.0023756735026836395, + 0.0016361374873667955, + 0.0017046591965481639, + 0.0014695044374093413, + 0.024038178846240044, + 0.0247160904109478, + 0.015430046245455742, + 0.014143806882202625, + 0.013552635908126831, + 0.012688345275819302, + 0.010512471199035645, + 0.012850381433963776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.06203943490982056 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "realistic, eye-level medium shot, held perfectly still, captures a blonde woman partially submerged in a swimming pool, emanating a profound sense of peace. Her long, honey-blond hair, wet and clinging to her bare shoulders, floats like silk in the crystal-clear water. Her eyes are gently closed, a subtle, serene smile playing on her lips, a picture of blissful relaxation. Submerged to her shoulders, her arms resting lightly on the surface, creating the faintest ripples. The brilliant midday sun bathes the landscape, causing countless tiny water droplets on her sun-soaked skin to sparkle like diamonds. The water is so transparent that beneath her, the small, iridescent blue and turquoise mosaic tiles lining the pool can be clearly seen, their colors shimmering and distorted by dancing, corroded patterns of light. The soft-focus composition of the clean, pale stone poolside deck in the background situates the scene in a luxurious, private setting. The cool blue of the water, the warm tones of her skin, and the sunlight dominate the composition, emanating a tranquil, almost meditative atmosphere.", + "seed": 29951, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29951_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.19814466889947652, + "vlm_logprob": -0.19814466889947652, + "vlm_yes_no_margin": 3.05, + "vlm_token_credit": -0.19814466889947652, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03397664427757263 + ], + "opposite_logprob_tokens": [ + -4.9089765548706055 + ], + "logprob_reward": -0.03397664427757263, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.03397664427757263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18515826761722565, + 0.06165220960974693, + 0.19297540187835693, + 0.057026010006666183, + 0.05087495222687721, + 0.06160551682114601, + 0.14368614554405212, + 0.14361509680747986, + 0.04769415035843849, + 0.05571224167943001 + ], + "credit_map": [ + 0.02613808773458004, + 0.0220487117767334, + 0.017461657524108887, + 0.02098829485476017, + 0.025208979845046997, + 0.025477008894085884, + 0.023594936355948448, + 0.024240581318736076, + 0.009207597002387047, + 0.008322036825120449, + 0.005707860458642244, + 0.007533136755228043, + 0.008370722644031048, + 0.006174001842737198, + 0.006549774669110775, + 0.009787073358893394, + 0.022640006616711617, + 0.023707272484898567, + 0.028614619746804237, + 0.019370464608073235, + 0.03026730939745903, + 0.019237596541643143, + 0.020849648863077164, + 0.02828849107027054, + 0.005809512455016375, + 0.005665904376655817, + 0.0049554831348359585, + 0.010181800462305546, + 0.006260508671402931, + 0.00919373333454132, + 0.005538636818528175, + 0.009420430287718773, + 0.006606375332921743, + 0.006325273308902979, + 0.006253104656934738, + 0.006961483042687178, + 0.005011737812310457, + 0.006467466242611408, + 0.005389907862991095, + 0.007859604433178902, + 0.007090525235980749, + 0.004681545775383711, + 0.0074640968814492226, + 0.006739695556461811, + 0.008450866676867008, + 0.009243967942893505, + 0.008794466964900494, + 0.009140350855886936, + 0.017367295920848846, + 0.01595054753124714, + 0.014534762129187584, + 0.01602322794497013, + 0.022280817851424217, + 0.014099064283072948, + 0.021925346925854683, + 0.02150508388876915, + 0.015429459512233734, + 0.010699326172471046, + 0.01602908968925476, + 0.01649000309407711, + 0.01777719520032406, + 0.020482812076807022, + 0.02163151651620865, + 0.025075698271393776, + 0.005340988282114267, + 0.007225106004625559, + 0.004021122585982084, + 0.005538550205528736, + 0.0068314094096422195, + 0.005648937076330185, + 0.004275868181139231, + 0.00881216675043106, + 0.007056124974042177, + 0.006177451461553574, + 0.005376438610255718, + 0.00593149196356535, + 0.00764456856995821, + 0.006762867793440819, + 0.007395610678941011, + 0.009367685765028 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.03397664427757263 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016081858426332474 + ], + "opposite_logprob_tokens": [ + -4.141081809997559 + ], + "logprob_reward": -0.016081858426332474, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016081858426332474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10535173118114471, + 0.09411787241697311, + 0.5256916880607605, + 0.042436208575963974, + 0.02148667722940445, + 0.03924327716231346, + 0.03365076705813408, + 0.07542261481285095, + 0.02639683708548546, + 0.036202263087034225 + ], + "credit_map": [ + 0.008802304975688457, + 0.01155247911810875, + 0.008832870982587337, + 0.026870423927903175, + 0.012652126140892506, + 0.010581606067717075, + 0.01162662822753191, + 0.014433289878070354, + 0.008753233589231968, + 0.010841568000614643, + 0.006794819142669439, + 0.021356170997023582, + 0.014844754710793495, + 0.013307432644069195, + 0.008417989127337933, + 0.009801903739571571, + 0.0550626702606678, + 0.07054755836725235, + 0.08056004345417023, + 0.058708567172288895, + 0.07910715788602829, + 0.052608147263526917, + 0.05566731467843056, + 0.07343022525310516, + 0.004210912622511387, + 0.0037340386770665646, + 0.0031326247844845057, + 0.006940059829503298, + 0.0044088964350521564, + 0.00908580794930458, + 0.0036991890519857407, + 0.007224679458886385, + 0.0023269362282007933, + 0.00317115499638021, + 0.0019809016957879066, + 0.0027740811929106712, + 0.004005156457424164, + 0.0019986219704151154, + 0.0026479626540094614, + 0.002581861801445484, + 0.004581396467983723, + 0.002900264458730817, + 0.005783348809927702, + 0.0037149451673030853, + 0.005312053021043539, + 0.005415750201791525, + 0.004064634442329407, + 0.0074708834290504456, + 0.004224166739732027, + 0.003757679834961891, + 0.003866492537781596, + 0.0029398046899586916, + 0.005869692657142878, + 0.004658870864659548, + 0.0038709533400833607, + 0.004463110584765673, + 0.007496651262044907, + 0.006815284490585327, + 0.006917616352438927, + 0.009692358784377575, + 0.009120478294789791, + 0.009196793660521507, + 0.012852816842496395, + 0.013330613262951374, + 0.002901124069467187, + 0.002143322490155697, + 0.0031738318502902985, + 0.0027125123888254166, + 0.004030116833746433, + 0.0030784495174884796, + 0.0027336108032613993, + 0.005623870063573122, + 0.004869121126830578, + 0.0035448751877993345, + 0.003608870320022106, + 0.0036228587850928307, + 0.0056586056016385555, + 0.003919529262930155, + 0.00426420709118247, + 0.006714195944368839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.016081858426332474 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193772338330746 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193772338330746, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193772338330746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31888845562934875, + 0.09004022926092148, + 0.07586585730314255, + 0.047707073390483856, + 0.15538334846496582, + 0.0397980660200119, + 0.12456093728542328, + 0.0644468143582344, + 0.061517126858234406, + 0.021792085841298103 + ], + "credit_map": [ + 0.022513829171657562, + 0.025198720395565033, + 0.01776440627872944, + 0.0943157747387886, + 0.020275862887501717, + 0.08469117432832718, + 0.021280255168676376, + 0.03284844011068344, + 0.007978628389537334, + 0.009037409909069538, + 0.0073580509051680565, + 0.014092648401856422, + 0.010501851327717304, + 0.01935306377708912, + 0.009751901030540466, + 0.011966684833168983, + 0.006588787771761417, + 0.010000011883676052, + 0.00934206135571003, + 0.010367672890424728, + 0.013663480058312416, + 0.007739496883004904, + 0.007728686556220055, + 0.010435654781758785, + 0.004488704726099968, + 0.004906903486698866, + 0.0038933507166802883, + 0.007104318588972092, + 0.006189377512782812, + 0.010002166964113712, + 0.005022492725402117, + 0.00609975541010499, + 0.014913566410541534, + 0.0203835628926754, + 0.01546220388263464, + 0.02158094011247158, + 0.021278932690620422, + 0.016570212319493294, + 0.021632615476846695, + 0.02356131561100483, + 0.0045537990517914295, + 0.004793174564838409, + 0.0033193419221788645, + 0.003921144641935825, + 0.00665351003408432, + 0.004910809453576803, + 0.004868217743933201, + 0.006778069771826267, + 0.015559450723230839, + 0.011182419024407864, + 0.013868036679923534, + 0.01347511075437069, + 0.015332912094891071, + 0.011922699399292469, + 0.019729653373360634, + 0.023490656167268753, + 0.007954210042953491, + 0.005182597786188126, + 0.006566779688000679, + 0.006986833643168211, + 0.008245022036135197, + 0.009707664139568806, + 0.008152787573635578, + 0.011650917120277882, + 0.006662061903625727, + 0.007543290965259075, + 0.00682100560516119, + 0.0062352754175662994, + 0.008820721879601479, + 0.006626366171985865, + 0.006342642940580845, + 0.012465760111808777, + 0.0020005737897008657, + 0.002629320602864027, + 0.002689013723284006, + 0.0029053559992462397, + 0.0032581808045506477, + 0.0022380934096872807, + 0.0032329654786735773, + 0.0028385822661221027 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.014193772338330746 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261648416519165 + ], + "opposite_logprob_tokens": [ + -0.5761648416519165 + ], + "logprob_reward": -0.8261648416519165, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261648416519165, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.28884607553482056, + 0.04822295904159546, + 0.27201780676841736, + 0.07962269335985184, + 0.09113842993974686, + 0.034574732184410095, + 0.05522541701793671, + 0.03950360417366028, + 0.03974244371056557, + 0.05110585317015648 + ], + "credit_map": [ + 0.028503669425845146, + 0.03192412853240967, + 0.027103420346975327, + 0.05834135785698891, + 0.03774326294660568, + 0.044502634555101395, + 0.03306378796696663, + 0.027663804590702057, + 0.004871425684541464, + 0.006675748620182276, + 0.004095588810741901, + 0.006898857187479734, + 0.006818871945142746, + 0.0067553832195699215, + 0.005450497847050428, + 0.006656585726886988, + 0.029233403503894806, + 0.03742086514830589, + 0.04054028540849686, + 0.03165152668952942, + 0.04241371899843216, + 0.02634471096098423, + 0.027367003262043, + 0.03704630210995674, + 0.007811906281858683, + 0.009696150198578835, + 0.006760192569345236, + 0.011665054596960545, + 0.009879710152745247, + 0.018821224570274353, + 0.005790450144559145, + 0.009198003448545933, + 0.010034153237938881, + 0.009035937488079071, + 0.012716122902929783, + 0.009271431714296341, + 0.01468812208622694, + 0.011558666825294495, + 0.00766505254432559, + 0.016168944537639618, + 0.004224022384732962, + 0.004111750517040491, + 0.002434826223179698, + 0.0035567390732467175, + 0.006184412632137537, + 0.004706012085080147, + 0.003396055893972516, + 0.005960914306342602, + 0.006771369371563196, + 0.0054356106556952, + 0.005015809088945389, + 0.005171104334294796, + 0.008138123899698257, + 0.00577533058822155, + 0.008637672290205956, + 0.010280396789312363, + 0.004132089205086231, + 0.004847781267017126, + 0.004139697179198265, + 0.004771699197590351, + 0.004666281398385763, + 0.005098054651170969, + 0.004610904026776552, + 0.0072370958514511585, + 0.004095142241567373, + 0.004159810021519661, + 0.0038107624277472496, + 0.004840902518481016, + 0.004964624065905809, + 0.004241040907800198, + 0.005105269141495228, + 0.00852489285171032, + 0.006147601641714573, + 0.005241380538791418, + 0.00629868358373642, + 0.0047242725268006325, + 0.009510945528745651, + 0.005696022883057594, + 0.005458647385239601, + 0.008028296753764153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.8261648416519165 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030622780323029 + ], + "opposite_logprob_tokens": [ + -2.350306272506714 + ], + "logprob_reward": -0.10030622780323029, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030622780323029, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15916235744953156, + 0.06612366437911987, + 0.19905662536621094, + 0.07839849591255188, + 0.09815935045480728, + 0.10121278464794159, + 0.13170349597930908, + 0.09487417340278625, + 0.023172805085778236, + 0.04813623055815697 + ], + "credit_map": [ + 0.009510756470263004, + 0.015532064251601696, + 0.01178563479334116, + 0.04121551290154457, + 0.016711445525288582, + 0.02703181654214859, + 0.020283633843064308, + 0.017091499641537666, + 0.0066986242309212685, + 0.006665188353508711, + 0.005661496426910162, + 0.00821000523865223, + 0.007005137391388416, + 0.013906327076256275, + 0.0059168157167732716, + 0.012060058303177357, + 0.02041560970246792, + 0.027337906882166862, + 0.029766544699668884, + 0.02307177148759365, + 0.031104452908039093, + 0.018829060718417168, + 0.02044421248137951, + 0.028087051585316658, + 0.006864564027637243, + 0.005706203170120716, + 0.004362765233963728, + 0.01578107289969921, + 0.006146172061562538, + 0.023252388462424278, + 0.005221609026193619, + 0.011063708923757076, + 0.011389090679585934, + 0.009874415583908558, + 0.011998836882412434, + 0.01144645269960165, + 0.012271592393517494, + 0.00997865293174982, + 0.014866096898913383, + 0.01633419655263424, + 0.007908475585281849, + 0.013912014663219452, + 0.01127920113503933, + 0.012441026978194714, + 0.012377501465380192, + 0.016341466456651688, + 0.009520445950329304, + 0.017432644963264465, + 0.01586318202316761, + 0.011485131457448006, + 0.014682409353554249, + 0.013311617076396942, + 0.015143354423344135, + 0.014803227037191391, + 0.01958189718425274, + 0.026832660660147667, + 0.009450411424040794, + 0.008998177945613861, + 0.009871721267700195, + 0.011013427749276161, + 0.013308203779160976, + 0.012209407053887844, + 0.012442277744412422, + 0.017580529674887657, + 0.0024801092222332954, + 0.0029949171002954245, + 0.002426538383588195, + 0.002285053487867117, + 0.003552932059392333, + 0.0027489603962749243, + 0.0022094750311225653, + 0.004474816843867302, + 0.006184421479701996, + 0.00516815809533, + 0.006115434225648642, + 0.0053938874043524265, + 0.008035434409976006, + 0.004895200952887535, + 0.004395409021526575, + 0.007948284037411213 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 6 + ], + "selected_reward": -0.10030622780323029 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "realistic, eye-level medium shot, held perfectly still, captures a blonde woman partially submerged in a swimming pool, emanating a profound sense of peace. Her long, honey-blond hair, wet and clinging to her bare shoulders, floats like silk in the crystal-clear water. Her eyes are gently closed, a subtle, serene smile playing on her lips, a picture of blissful relaxation. Submerged to her shoulders, her arms resting lightly on the surface, creating the faintest ripples. The brilliant midday sun bathes the landscape, causing countless tiny water droplets on her sun-soaked skin to sparkle like diamonds. The water is so transparent that beneath her, the small, iridescent blue and turquoise mosaic tiles lining the pool can be clearly seen, their colors shimmering and distorted by dancing, corroded patterns of light. The soft-focus composition of the clean, pale stone poolside deck in the background situates the scene in a luxurious, private setting. The cool blue of the water, the warm tones of her skin, and the sunlight dominate the composition, emanating a tranquil, almost meditative atmosphere.", + "seed": 29952, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_3/index_7068_seed_29952_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3055019184947014, + "vlm_logprob": -0.3055019184947014, + "vlm_yes_no_margin": 2.7, + "vlm_token_credit": -0.3055019184947014, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04042854532599449 + ], + "opposite_logprob_tokens": [ + -5.040428638458252 + ], + "logprob_reward": -0.04042854532599449, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.04042854532599449, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08719915896654129, + 0.17134283483028412, + 0.0309513658285141, + 0.24711579084396362, + 0.09778345376253128, + 0.1329990029335022, + 0.08569888025522232, + 0.04481378570199013, + 0.026591015979647636, + 0.07550469785928726 + ], + "credit_map": [ + 0.00998741015791893, + 0.01579079031944275, + 0.007405442651361227, + 0.008162382058799267, + 0.011789812706410885, + 0.013059969060122967, + 0.011143682524561882, + 0.00985967181622982, + 0.011781019158661366, + 0.01656876504421234, + 0.016634654253721237, + 0.006706166546791792, + 0.052470557391643524, + 0.012629867531359196, + 0.006190067157149315, + 0.048361729830503464, + 0.002620500745251775, + 0.0037051327526569366, + 0.00445761252194643, + 0.003988536540418863, + 0.00638982281088829, + 0.0033811964094638824, + 0.004166221246123314, + 0.0022423435002565384, + 0.018127132207155228, + 0.07041602581739426, + 0.009080450050532818, + 0.022485841065645218, + 0.07671034336090088, + 0.021004794165492058, + 0.009203851222991943, + 0.020087363198399544, + 0.009162274189293385, + 0.019780686125159264, + 0.006845348514616489, + 0.010706557892262936, + 0.025356575846672058, + 0.008982360363006592, + 0.009244863875210285, + 0.007704789750277996, + 0.012825554236769676, + 0.017916979268193245, + 0.019319040700793266, + 0.023333560675382614, + 0.015288186259567738, + 0.009886366315186024, + 0.01823483221232891, + 0.016194485127925873, + 0.007473020348697901, + 0.010505120269954205, + 0.015812493860721588, + 0.01340759452432394, + 0.00613384647294879, + 0.011655338108539581, + 0.008579199202358723, + 0.012132271192967892, + 0.003379239933565259, + 0.005686815362423658, + 0.005566050298511982, + 0.006536043249070644, + 0.0030882623977959156, + 0.004871996119618416, + 0.006262151058763266, + 0.009423229843378067, + 0.0033634880091995, + 0.0029561659321188927, + 0.0022955969907343388, + 0.004359211772680283, + 0.002648185007274151, + 0.003877875627949834, + 0.003504987107589841, + 0.0035855057649314404, + 0.006165745202451944, + 0.01911655254662037, + 0.007814409211277962, + 0.006142208352684975, + 0.021614979952573776, + 0.006345449015498161, + 0.003697814419865608, + 0.004607546143233776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.04042854532599449 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023300303146243095 + ], + "opposite_logprob_tokens": [ + -3.7733004093170166 + ], + "logprob_reward": -0.023300303146243095, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023300303146243095, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1265455037355423, + 0.27702122926712036, + 0.09892742335796356, + 0.12162799388170242, + 0.05871313810348511, + 0.05935383588075638, + 0.008261453360319138, + 0.02149832248687744, + 0.009395034983754158, + 0.21865609288215637 + ], + "credit_map": [ + 0.012405181303620338, + 0.01830212026834488, + 0.010986979119479656, + 0.021556584164500237, + 0.026235336437821388, + 0.01257561519742012, + 0.012017889879643917, + 0.01246579922735691, + 0.01777040585875511, + 0.02374931238591671, + 0.02084498293697834, + 0.0077741737477481365, + 0.09602423757314682, + 0.021037578582763672, + 0.0073044924065470695, + 0.08251603692770004, + 0.006663977634161711, + 0.008117914199829102, + 0.013105364516377449, + 0.007682980503886938, + 0.027315516024827957, + 0.00833562295883894, + 0.008171692490577698, + 0.01953435130417347, + 0.00842356588691473, + 0.031086036935448647, + 0.004239147994667292, + 0.008844955824315548, + 0.04054839909076691, + 0.00949133187532425, + 0.004459824413061142, + 0.014534725807607174, + 0.004998283460736275, + 0.010876249521970749, + 0.007935273461043835, + 0.008428175002336502, + 0.0107883894816041, + 0.00431049894541502, + 0.006409266032278538, + 0.004967004060745239, + 0.004811514634639025, + 0.008855625987052917, + 0.010377796366810799, + 0.008096959441900253, + 0.007686507422477007, + 0.007311573717743158, + 0.006307410076260567, + 0.0059064459055662155, + 0.0008230258245021105, + 0.001047420664690435, + 0.0012667712289839983, + 0.000984824844636023, + 0.0006400048150680959, + 0.001307265367358923, + 0.0011025624116882682, + 0.001089578028768301, + 0.00168316881172359, + 0.002607404487207532, + 0.003838585689663887, + 0.0026003296952694654, + 0.0017908884910866618, + 0.003313795430585742, + 0.0021884196903556585, + 0.0034757324028760195, + 0.0012631543213501573, + 0.0011015882482752204, + 0.0013309279456734657, + 0.0013496429892256856, + 0.0009244755492545664, + 0.0010797204449772835, + 0.0011781963985413313, + 0.0011673294939100742, + 0.01654118113219738, + 0.06008683517575264, + 0.025639733299613, + 0.015287134796380997, + 0.06685850769281387, + 0.013640969060361385, + 0.009033462963998318, + 0.011568273417651653 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.023300303146243095 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023282600566744804 + ], + "opposite_logprob_tokens": [ + -3.773282527923584 + ], + "logprob_reward": -0.023282600566744804, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023282600566744804, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09515149891376495, + 0.045540791004896164, + 0.11016486585140228, + 0.3985423147678375, + 0.06211508437991142, + 0.08061759918928146, + 0.0641336664557457, + 0.006721568293869495, + 0.014400807209312916, + 0.12261182814836502 + ], + "credit_map": [ + 0.008620516397058964, + 0.014726344496011734, + 0.010635667480528355, + 0.013138599693775177, + 0.012231949716806412, + 0.015500999055802822, + 0.010316204279661179, + 0.00998120941221714, + 0.002878266852349043, + 0.00416547479107976, + 0.00464640511199832, + 0.002520299283787608, + 0.013289323076605797, + 0.005173997022211552, + 0.003046603873372078, + 0.009820425882935524, + 0.006760014686733484, + 0.007838858291506767, + 0.011622581630945206, + 0.005897513125091791, + 0.03507440164685249, + 0.008363776840269566, + 0.0055095236748456955, + 0.02909819409251213, + 0.03244069963693619, + 0.1255042403936386, + 0.01571083627641201, + 0.05322611704468727, + 0.09936252981424332, + 0.033211540430784225, + 0.018114369362592697, + 0.020972002297639847, + 0.005280130077153444, + 0.011415746062994003, + 0.009147386997938156, + 0.008187220431864262, + 0.01254283171147108, + 0.0051859053783118725, + 0.005988210439682007, + 0.004367648158222437, + 0.006959103047847748, + 0.012441834434866905, + 0.014366802759468555, + 0.011423546820878983, + 0.00951098557561636, + 0.010545925237238407, + 0.00813205074518919, + 0.007237350568175316, + 0.006348907481878996, + 0.00695749232545495, + 0.010493913665413857, + 0.012774191796779633, + 0.005509759299457073, + 0.007124215830117464, + 0.007831393741071224, + 0.007093797903507948, + 0.0008161490550264716, + 0.0008093032520264387, + 0.0006486335187219083, + 0.0009712015744298697, + 0.0005185029585845768, + 0.0007887068204581738, + 0.0008944534347392619, + 0.0012746178545057774, + 0.0012526315404102206, + 0.0017259237356483936, + 0.001981972483918071, + 0.002452546963468194, + 0.0011761753121390939, + 0.0019003782654181123, + 0.0018340531969442964, + 0.002077126642689109, + 0.009164810180664062, + 0.03377328813076019, + 0.014309955760836601, + 0.008193032816052437, + 0.03777008131146431, + 0.0077384235337376595, + 0.00521040428429842, + 0.006451823748648167 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 9, + 2 + ], + "selected_reward": -0.023282600566744804 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3134359121322632 + ], + "opposite_logprob_tokens": [ + -0.31343594193458557 + ], + "logprob_reward": -1.3134359121322632, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3134359121322632, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13638821244239807, + 0.18814478814601898, + 0.02640703320503235, + 0.16811054944992065, + 0.11537668108940125, + 0.0972873717546463, + 0.0907394289970398, + 0.019254840910434723, + 0.016280891373753548, + 0.1420101672410965 + ], + "credit_map": [ + 0.008077891543507576, + 0.019047530367970467, + 0.01257239282131195, + 0.030764641240239143, + 0.02355479821562767, + 0.010014460422098637, + 0.016640927642583847, + 0.015715574845671654, + 0.011731157079339027, + 0.014964835718274117, + 0.014844425022602081, + 0.005376098211854696, + 0.06504808366298676, + 0.0146150141954422, + 0.006556249689310789, + 0.05500892922282219, + 0.0033319422509521246, + 0.003798065008595586, + 0.003074903041124344, + 0.002499246271327138, + 0.004677566699683666, + 0.0031646783463656902, + 0.002756476402282715, + 0.003104154486209154, + 0.011711410246789455, + 0.0340152345597744, + 0.012809854000806808, + 0.029588919132947922, + 0.02465512417256832, + 0.009591496549546719, + 0.010221381671726704, + 0.035517144948244095, + 0.010107009671628475, + 0.022554069757461548, + 0.013824095949530602, + 0.012701228260993958, + 0.025949671864509583, + 0.012798155657947063, + 0.010112574324011803, + 0.007329879328608513, + 0.009928020648658276, + 0.013064184226095676, + 0.017141666263341904, + 0.0179472416639328, + 0.011116624809801579, + 0.009025690145790577, + 0.011385573074221611, + 0.007678365334868431, + 0.007342735771089792, + 0.010820995084941387, + 0.016064250841736794, + 0.01792963594198227, + 0.0065583232790231705, + 0.010279180482029915, + 0.01106533594429493, + 0.010678972117602825, + 0.002104541752487421, + 0.0023289036471396685, + 0.00221363827586174, + 0.0017169704660773277, + 0.0018834826769307256, + 0.0018844726728275418, + 0.0021245654206722975, + 0.004998264834284782, + 0.0020607050973922014, + 0.0019191743340343237, + 0.0029153272043913603, + 0.0024729163851588964, + 0.0010979105718433857, + 0.0016809194348752499, + 0.0022569491993635893, + 0.001876990427263081, + 0.011120306327939034, + 0.037687864154577255, + 0.01753624528646469, + 0.010191228240728378, + 0.037544894963502884, + 0.00947167631238699, + 0.0077274455688893795, + 0.010730503126978874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 9 + ], + "selected_reward": -1.3134359121322632 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot an eye-level medium shot of a blonde woman?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the woman partially submerged in a swimming pool up to her shoulders?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are iridescent blue and turquoise mosaic tiles visible beneath the water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the camera held perfectly still throughout the shot?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a clean, pale stone poolside deck?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12706223130226135 + ], + "opposite_logprob_tokens": [ + -2.1270623207092285 + ], + "logprob_reward": -0.12706223130226135, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12706223130226135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16440477967262268, + 0.042195674031972885, + 0.013065226376056671, + 0.35380932688713074, + 0.03295573219656944, + 0.1831599920988083, + 0.020745983347296715, + 0.011880403384566307, + 0.016302812844514847, + 0.1614801585674286 + ], + "credit_map": [ + 0.01576763577759266, + 0.017199868336319923, + 0.013306156732141972, + 0.03596585988998413, + 0.03936632350087166, + 0.018322711810469627, + 0.01190211996436119, + 0.01257409993559122, + 0.002764030359685421, + 0.00346539169549942, + 0.0038245036266744137, + 0.0018607631791383028, + 0.013262777589261532, + 0.0037989600095897913, + 0.0020393752492964268, + 0.01117987185716629, + 0.0008005236159078777, + 0.0015814928337931633, + 0.001649656449444592, + 0.0022325122263282537, + 0.002756600733846426, + 0.0015689427964389324, + 0.0009013917879201472, + 0.0015741040697321296, + 0.029082223773002625, + 0.11291316896677017, + 0.014759756624698639, + 0.04911522939801216, + 0.08317898958921432, + 0.02957547828555107, + 0.016711203381419182, + 0.01847320981323719, + 0.0028726609889417887, + 0.005525845568627119, + 0.004589538089931011, + 0.004334489814937115, + 0.005851526744663715, + 0.003698439337313175, + 0.002910471986979246, + 0.0031727540772408247, + 0.015314933843910694, + 0.02618137188255787, + 0.03224433213472366, + 0.02962522581219673, + 0.022813353687524796, + 0.017901763319969177, + 0.020676130428910255, + 0.018402857705950737, + 0.002479788614436984, + 0.0027253427542746067, + 0.0024872813373804092, + 0.0030336943455040455, + 0.001817591255530715, + 0.0030114762485027313, + 0.0026478718500584364, + 0.0025429355446249247, + 0.0009756775689311326, + 0.0013573078904300928, + 0.001176672289147973, + 0.0013891649432480335, + 0.0011914778733626008, + 0.0011391607113182545, + 0.0014293373096734285, + 0.003221604274585843, + 0.0014095985097810626, + 0.0021089110523462296, + 0.002553042722865939, + 0.00242748879827559, + 0.0014147530309855938, + 0.0024316003546118736, + 0.0018561907345429063, + 0.0021012267097830772, + 0.012108090333640575, + 0.04429803788661957, + 0.019038420170545578, + 0.01086477655917406, + 0.04990331083536148, + 0.010289907455444336, + 0.0064787124283611774, + 0.008498878218233585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 5, + 0 + ], + "selected_reward": -0.12706223130226135 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_tensor([8875, 8875])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_tensor([8875, 8875])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..41762205f690ae303b85876bf53140d7109a35cb --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_tensor([8875, 8875])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level close-up captures the tranquil world of a solitary Ranchu goldfish, suspended in a glass aquarium with algae growth. This goldfish, with its distinctive egg-shaped body and lack of a dorsal fin, displays a vibrant pattern of deep orange and pearlescent white patches on its shimmering scales. Its prominent, bumpy head growth, or wen, partially obscures its dark, calm eyes, giving it a contemplative expression. The fish hovers almost motionless, its delicate pectoral fins making tiny, steady adjustments while its twin tail fins occasionally swish slowly, listlessly, propelling it aimlessly through the murky water. The tank's glass walls are coated in a soft green film, and long, ethereal strands of hair algae drift from the dark driftwood in the background. A faint light source from above penetrates the turquoise water, casting a hazy crepuscular glow that illuminates specks of floating debris and captures the metallic sheen of the goldfish's scales. The main color palette is a soft green and brown spectrum, which makes for a striking contrast to the bright colors of the goldfish, creating a strong atmosphere of tranquility and a profound sense of solitude.", + "seed": 22232, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_8875_seed_22232_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03145181271247566, + "vlm_logprob": -0.03145181271247566, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.03145181271247566, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007925373502075672 + ], + "opposite_logprob_tokens": [ + -5.007925510406494 + ], + "logprob_reward": -0.007925373502075672, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.007925373502075672, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23207172751426697, + 0.3487718105316162, + 0.10842829942703247, + 0.10043902695178986, + 0.0369953028857708, + 0.03088468313217163, + 0.04290644824504852, + 0.01845831423997879, + 0.05361904948949814, + 0.027425292879343033 + ], + "credit_map": [ + 0.05887053906917572, + 0.021256791427731514, + 0.04372149705886841, + 0.01786760427057743, + 0.0174744613468647, + 0.022106047719717026, + 0.02003367803990841, + 0.030741136521100998, + 0.041854724287986755, + 0.04327917844057083, + 0.04274847358465195, + 0.04065285995602608, + 0.06031597778201103, + 0.03267035633325577, + 0.05110060051083565, + 0.03614965081214905, + 0.01261925045400858, + 0.014639206230640411, + 0.01618274301290512, + 0.013442076742649078, + 0.011949408799409866, + 0.012951373122632504, + 0.015030081383883953, + 0.011614168994128704, + 0.020194724202156067, + 0.007786684203892946, + 0.0300107654184103, + 0.005355804692953825, + 0.005374818108975887, + 0.008486301638185978, + 0.005273285787552595, + 0.017956648021936417, + 0.007665693294256926, + 0.0033161973115056753, + 0.006723539903759956, + 0.003916904330253601, + 0.0037021669559180737, + 0.003670684527605772, + 0.004137017764151096, + 0.0038631008937954903, + 0.004472136497497559, + 0.0026568484026938677, + 0.00665463600307703, + 0.0033094610553234816, + 0.002625345950946212, + 0.0026048794388771057, + 0.0031383689492940903, + 0.005423008464276791, + 0.009994938038289547, + 0.004299531225115061, + 0.0073172468692064285, + 0.0029692354146391153, + 0.0043797907419502735, + 0.003884614212438464, + 0.005649878643453121, + 0.004411218222230673, + 0.0024697778280824423, + 0.0019309563795104623, + 0.002806382719427347, + 0.0017715132562443614, + 0.001932572340592742, + 0.002212724182754755, + 0.0030486881732940674, + 0.0022856995929032564, + 0.008348409086465836, + 0.0032123872078955173, + 0.015117567032575607, + 0.005607569590210915, + 0.0034305728040635586, + 0.004067096393555403, + 0.004669228568673134, + 0.00916622020304203, + 0.006615371908992529, + 0.002789764665067196, + 0.0036560194566845894, + 0.002917112782597542, + 0.0024856673553586006, + 0.002755096647888422, + 0.0033042237628251314, + 0.0029020390938967466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.007925373502075672 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003627388970926404 + ], + "opposite_logprob_tokens": [ + -5.628627300262451 + ], + "logprob_reward": -0.003627388970926404, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.003627388970926404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38566505908966064, + 0.05354383587837219, + 0.052805423736572266, + 0.04907587543129921, + 0.0846591591835022, + 0.09904934465885162, + 0.03421863168478012, + 0.13668160140514374, + 0.05943584442138672, + 0.0448651984333992 + ], + "credit_map": [ + 0.08976852148771286, + 0.039676811546087265, + 0.0592472217977047, + 0.029418297111988068, + 0.039085421711206436, + 0.0423445999622345, + 0.04261988401412964, + 0.04350431263446808, + 0.01607624441385269, + 0.006341748405247927, + 0.007153815124183893, + 0.003084247699007392, + 0.004147246479988098, + 0.006291450932621956, + 0.0044278777204453945, + 0.0060212076641619205, + 0.00644027441740036, + 0.008074095472693443, + 0.008365340530872345, + 0.005517315585166216, + 0.006293757818639278, + 0.0071545992977917194, + 0.005839258432388306, + 0.005120778456330299, + 0.008940869942307472, + 0.0055129649117589, + 0.006867812480777502, + 0.006368751171976328, + 0.005817054305225611, + 0.004411930218338966, + 0.006099191959947348, + 0.005057299509644508, + 0.019600842148065567, + 0.006643469445407391, + 0.01671655848622322, + 0.008768697269260883, + 0.0075110504403710365, + 0.006049012299627066, + 0.009491080418229103, + 0.009878449141979218, + 0.009587586857378483, + 0.010669896379113197, + 0.021437399089336395, + 0.008845177479088306, + 0.01292052399367094, + 0.007669993210583925, + 0.015526236966252327, + 0.012392533011734486, + 0.005565413273870945, + 0.003802977269515395, + 0.009542828425765038, + 0.0026826548855751753, + 0.00216744770295918, + 0.0022032957058399916, + 0.002965538762509823, + 0.005288478918373585, + 0.014521809294819832, + 0.01291788648813963, + 0.016764111816883087, + 0.01685468666255474, + 0.02008395828306675, + 0.0184337068349123, + 0.023668475449085236, + 0.013436974957585335, + 0.008160922676324844, + 0.004608434159308672, + 0.012428594753146172, + 0.006833283696323633, + 0.006045555230230093, + 0.0048282938078045845, + 0.007921232841908932, + 0.008609525859355927, + 0.0074283950962126255, + 0.004450191278010607, + 0.00811704434454441, + 0.004896081052720547, + 0.0036539447028189898, + 0.006570453755557537, + 0.0043489267118275166, + 0.005400161724537611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.003627388970926404 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011072029359638691 + ], + "opposite_logprob_tokens": [ + -4.511072158813477 + ], + "logprob_reward": -0.011072029359638691, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011072029359638691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14244729280471802, + 0.3363926112651825, + 0.16288071870803833, + 0.045377038419246674, + 0.054285064339637756, + 0.05667657405138016, + 0.05387641862034798, + 0.028447961434721947, + 0.050682760775089264, + 0.06893359869718552 + ], + "credit_map": [ + 0.030488956719636917, + 0.015079101547598839, + 0.017172060906887054, + 0.015028533525764942, + 0.021042751148343086, + 0.011347739957273006, + 0.018279828131198883, + 0.014008313417434692, + 0.04284057021141052, + 0.042598944157361984, + 0.039263766258955, + 0.03912229835987091, + 0.05627527832984924, + 0.032664164900779724, + 0.04852435737848282, + 0.03510322421789169, + 0.024225596338510513, + 0.024385109543800354, + 0.02140013873577118, + 0.020535267889499664, + 0.016685059294104576, + 0.020967120304703712, + 0.019894786179065704, + 0.014787639491260052, + 0.008704273030161858, + 0.005942640360444784, + 0.00455327145755291, + 0.005357820075005293, + 0.004257796332240105, + 0.003967789467424154, + 0.007703052368015051, + 0.004890393931418657, + 0.007254232186824083, + 0.004763251170516014, + 0.011673405766487122, + 0.006972647272050381, + 0.0039023105055093765, + 0.004888498689979315, + 0.006711205467581749, + 0.008119514212012291, + 0.010539965704083443, + 0.005046197213232517, + 0.008169679902493954, + 0.007160704582929611, + 0.007184719666838646, + 0.005011001136153936, + 0.00808416586369276, + 0.0054801409132778645, + 0.012411270290613174, + 0.005969705060124397, + 0.007656315341591835, + 0.005776809994131327, + 0.004903445951640606, + 0.005779996979981661, + 0.006038906518369913, + 0.005339964758604765, + 0.003704471979290247, + 0.002614357275888324, + 0.004698931705206633, + 0.004745360463857651, + 0.0021141208708286285, + 0.0033915513195097446, + 0.0038385188672691584, + 0.0033406484872102737, + 0.007447867188602686, + 0.00422285171225667, + 0.01253854762762785, + 0.005624673329293728, + 0.004479223862290382, + 0.003918677568435669, + 0.004928054288029671, + 0.007522860541939735, + 0.011060573160648346, + 0.007391105405986309, + 0.009920314885675907, + 0.006036682054400444, + 0.007856839336454868, + 0.01094038411974907, + 0.00790488626807928, + 0.007822815328836441 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.011072029359638691 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007623270619660616 + ], + "opposite_logprob_tokens": [ + -4.882623195648193 + ], + "logprob_reward": -0.007623270619660616, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007623270619660616, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4743345379829407, + 0.2568199038505554, + 0.0462581105530262, + 0.050481460988521576, + 0.036061350256204605, + 0.03915063291788101, + 0.03937245160341263, + 0.027438128367066383, + 0.014345232397317886, + 0.015738148242235184 + ], + "credit_map": [ + 0.08376814424991608, + 0.047812819480895996, + 0.08172757178544998, + 0.040991563349962234, + 0.058062829077243805, + 0.05083867162466049, + 0.05396212637424469, + 0.0571708157658577, + 0.03418669477105141, + 0.030215220525860786, + 0.032554831355810165, + 0.029729872941970825, + 0.04170816391706467, + 0.022414252161979675, + 0.03789249807596207, + 0.028118396177887917, + 0.007013990543782711, + 0.006440578028559685, + 0.0070847137831151485, + 0.0052394745871424675, + 0.005597066134214401, + 0.005043730139732361, + 0.005328996106982231, + 0.004509564954787493, + 0.011865679174661636, + 0.005504321772605181, + 0.00591279799118638, + 0.0054266140796244144, + 0.005159758497029543, + 0.005004118196666241, + 0.005864851176738739, + 0.005743322428315878, + 0.008206406608223915, + 0.003356101457029581, + 0.006195874884724617, + 0.0037757675163447857, + 0.0031392276287078857, + 0.0022895867004990578, + 0.004903009627014399, + 0.004195380490273237, + 0.005657453555613756, + 0.0031718863174319267, + 0.005918078124523163, + 0.010511266998946667, + 0.002561457920819521, + 0.002892702352255583, + 0.0046230885200202465, + 0.003814703319221735, + 0.0050389752723276615, + 0.0037392321974039078, + 0.009022688493132591, + 0.007526782341301441, + 0.0022691055200994015, + 0.003989271819591522, + 0.0033781914971768856, + 0.004408208187669516, + 0.005438712891191244, + 0.0036758717615157366, + 0.0037181132938712835, + 0.003259955672547221, + 0.002981590572744608, + 0.0023070713505148888, + 0.00346618564799428, + 0.0025906283408403397, + 0.001709671108983457, + 0.001658934517763555, + 0.0015258645871654153, + 0.00234465510584414, + 0.0017341317143291235, + 0.0015185714000836015, + 0.0023252354003489017, + 0.0015281692612916231, + 0.002976552117615938, + 0.0014920334797352552, + 0.00247076153755188, + 0.0014627709751948714, + 0.0012105132918804884, + 0.001978848595172167, + 0.002086624735966325, + 0.0020600457210093737 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.007623270619660616 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270110011100769 + ], + "opposite_logprob_tokens": [ + -2.1270110607147217 + ], + "logprob_reward": -0.1270110011100769, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270110011100769, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07379288971424103, + 0.6825212240219116, + 0.03994143009185791, + 0.019239699468016624, + 0.04173070192337036, + 0.027158569544553757, + 0.024529756978154182, + 0.04296113923192024, + 0.02410745993256569, + 0.02401711419224739 + ], + "credit_map": [ + 0.012689600698649883, + 0.008904947899281979, + 0.009731256403028965, + 0.008161169476807117, + 0.006685520522296429, + 0.00999727938324213, + 0.008025219663977623, + 0.009597890079021454, + 0.09263575822114944, + 0.08055666834115982, + 0.08908609300851822, + 0.07901494950056076, + 0.11003866046667099, + 0.059015702456235886, + 0.09904687851667404, + 0.07312652468681335, + 0.006278804503381252, + 0.007318579126149416, + 0.006041802931576967, + 0.00469225225970149, + 0.003007694147527218, + 0.0041921972297132015, + 0.003990472760051489, + 0.004419626668095589, + 0.004910887219011784, + 0.0025921319611370564, + 0.0025764263700693846, + 0.0017625144682824612, + 0.0017055212520062923, + 0.0020110609475523233, + 0.0017875292105600238, + 0.0018936265259981155, + 0.007869703695178032, + 0.0040829963982105255, + 0.0045936135575175285, + 0.00566922128200531, + 0.00507658114656806, + 0.0031234263442456722, + 0.007175816688686609, + 0.004139341879636049, + 0.0060844519175589085, + 0.002854288090020418, + 0.0048246984370052814, + 0.0038088641595095396, + 0.002353621879592538, + 0.0017348190303891897, + 0.0030349348671734333, + 0.002462890464812517, + 0.005627560429275036, + 0.0025490610860288143, + 0.006428747903555632, + 0.0018616163870319724, + 0.0016512201400473714, + 0.0016219118842855096, + 0.001824444392696023, + 0.002965191612020135, + 0.008386851288378239, + 0.004587060771882534, + 0.004585833288729191, + 0.00450562871992588, + 0.006459364667534828, + 0.0043815611861646175, + 0.006346213165670633, + 0.003708624979481101, + 0.0048431651666760445, + 0.002439557109028101, + 0.004696860909461975, + 0.0024702742230147123, + 0.002263653790578246, + 0.0017827716656029224, + 0.002938364865258336, + 0.002672809874638915, + 0.002654370851814747, + 0.002404205035418272, + 0.004018663894385099, + 0.002225213684141636, + 0.0027795142959803343, + 0.003954751882702112, + 0.003137331223115325, + 0.002843061927706003 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 7 + ], + "selected_reward": -0.1270110011100769 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the tranquil world of a solitary Ranchu goldfish, suspended in a glass aquarium with algae growth. This goldfish, with its distinctive egg-shaped body and lack of a dorsal fin, displays a vibrant pattern of deep orange and pearlescent white patches on its shimmering scales. Its prominent, bumpy head growth, or wen, partially obscures its dark, calm eyes, giving it a contemplative expression. The fish hovers almost motionless, its delicate pectoral fins making tiny, steady adjustments while its twin tail fins occasionally swish slowly, listlessly, propelling it aimlessly through the murky water. The tank's glass walls are coated in a soft green film, and long, ethereal strands of hair algae drift from the dark driftwood in the background. A faint light source from above penetrates the turquoise water, casting a hazy crepuscular glow that illuminates specks of floating debris and captures the metallic sheen of the goldfish's scales. The main color palette is a soft green and brown spectrum, which makes for a striking contrast to the bright colors of the goldfish, creating a strong atmosphere of tranquility and a profound sense of solitude.", + "seed": 22233, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_8875_seed_22233_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.02573545784689486, + "vlm_logprob": -0.02573545784689486, + "vlm_yes_no_margin": 4.525, + "vlm_token_credit": -0.02573545784689486, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010160863399505615 + ], + "opposite_logprob_tokens": [ + -4.76016092300415 + ], + "logprob_reward": -0.010160863399505615, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.010160863399505615, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.42974984645843506, + 0.09377705305814743, + 0.06907878071069717, + 0.04836585745215416, + 0.06675364077091217, + 0.051178865134716034, + 0.03842164948582649, + 0.05454540252685547, + 0.09205292165279388, + 0.05607593059539795 + ], + "credit_map": [ + 0.04509591683745384, + 0.0730413869023323, + 0.06303995847702026, + 0.10130265355110168, + 0.03793124854564667, + 0.024823786690831184, + 0.04916749894618988, + 0.03534739464521408, + 0.010624421760439873, + 0.009274983778595924, + 0.015081043355166912, + 0.01546203438192606, + 0.010143079794943333, + 0.009649301879107952, + 0.014058142900466919, + 0.00948403961956501, + 0.006220719311386347, + 0.007030677981674671, + 0.009509801864624023, + 0.008270141668617725, + 0.0078090522438287735, + 0.0069419387727975845, + 0.017136799171566963, + 0.006159645970910788, + 0.005527925677597523, + 0.005475070793181658, + 0.005998408421874046, + 0.006725475657731295, + 0.0052344235591590405, + 0.004622833803296089, + 0.009732922539114952, + 0.005048800725489855, + 0.006794019136577845, + 0.008037510327994823, + 0.007670928258448839, + 0.008704782463610172, + 0.009824167937040329, + 0.005092085339128971, + 0.014540550298988819, + 0.006089598406106234, + 0.00563270878046751, + 0.0064402418211102486, + 0.00625629024580121, + 0.0059414212591946125, + 0.006220873445272446, + 0.005837019998580217, + 0.009193380363285542, + 0.0056569259613752365, + 0.004101120866835117, + 0.006540939677506685, + 0.004698748234659433, + 0.0036386363208293915, + 0.004883193876594305, + 0.0038229995407164097, + 0.006649225018918514, + 0.004086787346750498, + 0.006910474039614201, + 0.008499273099005222, + 0.0064373016357421875, + 0.007084015291184187, + 0.0064523788169026375, + 0.005440542474389076, + 0.00827905535697937, + 0.005442362278699875, + 0.007328423671424389, + 0.011243819259107113, + 0.011773594655096531, + 0.022439025342464447, + 0.009840111248195171, + 0.0072655947878956795, + 0.012218865565955639, + 0.009943484328687191, + 0.005634617060422897, + 0.0073984600603580475, + 0.006636875215917826, + 0.007874011993408203, + 0.008106962777674198, + 0.007553377188742161, + 0.007591420318931341, + 0.00528020691126585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.010160863399505615 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032063524704426527 + ], + "opposite_logprob_tokens": [ + -5.753206253051758 + ], + "logprob_reward": -0.0032063524704426527, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0032063524704426527, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4139013886451721, + 0.08082292973995209, + 0.11160393059253693, + 0.0662974938750267, + 0.04192136973142624, + 0.0598030686378479, + 0.03801828995347023, + 0.05422940105199814, + 0.08495970070362091, + 0.04844241589307785 + ], + "credit_map": [ + 0.04485192894935608, + 0.08319395780563354, + 0.05640634149312973, + 0.11366229504346848, + 0.024707913398742676, + 0.018473362550139427, + 0.038413189351558685, + 0.034192413091659546, + 0.009147507138550282, + 0.010480104014277458, + 0.012585229240357876, + 0.013668187893927097, + 0.006894371472299099, + 0.009105799719691277, + 0.010882286354899406, + 0.008059442974627018, + 0.010560044087469578, + 0.010509846732020378, + 0.016229132190346718, + 0.020172011107206345, + 0.011523155495524406, + 0.013893303461372852, + 0.014007454738020897, + 0.014708982780575752, + 0.006508432794362307, + 0.010805686935782433, + 0.007689660415053368, + 0.008397105149924755, + 0.007297217380255461, + 0.0058778878301382065, + 0.010736586526036263, + 0.008984919637441635, + 0.003132350044324994, + 0.006276092957705259, + 0.003879087045788765, + 0.00608732458204031, + 0.004108056426048279, + 0.0034923809580504894, + 0.01002961304038763, + 0.00491646071895957, + 0.006497641559690237, + 0.008213928900659084, + 0.007995505817234516, + 0.006813094485551119, + 0.006832432467490435, + 0.0063412911258637905, + 0.01141411904245615, + 0.00569505337625742, + 0.003959738649427891, + 0.00438624806702137, + 0.004674641881138086, + 0.00424141064286232, + 0.005098898895084858, + 0.003962425515055656, + 0.006171845365315676, + 0.0055230832658708096, + 0.005314299836754799, + 0.008976527489721775, + 0.006621236447244883, + 0.007271254900842905, + 0.007062401156872511, + 0.004102188162505627, + 0.009493397548794746, + 0.005388095509260893, + 0.00769847584888339, + 0.009409460239112377, + 0.01130327396094799, + 0.022141871973872185, + 0.008827951736748219, + 0.005433721002191305, + 0.011936796829104424, + 0.008208150044083595, + 0.004528399556875229, + 0.006335964426398277, + 0.006047017872333527, + 0.006188154220581055, + 0.0074785067699849606, + 0.004566298332065344, + 0.007178526371717453, + 0.006119545549154282 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 8 + ], + "selected_reward": -0.0032063524704426527 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978105328977108 + ], + "opposite_logprob_tokens": [ + -4.6347808837890625 + ], + "logprob_reward": -0.00978105328977108, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978105328977108, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27580326795578003, + 0.13737507164478302, + 0.14446552097797394, + 0.08618512004613876, + 0.04516209661960602, + 0.07898682355880737, + 0.06328707188367844, + 0.028708940371870995, + 0.10042095184326172, + 0.039605144411325455 + ], + "credit_map": [ + 0.030933499336242676, + 0.041956882923841476, + 0.04429697245359421, + 0.055105384439229965, + 0.026713630184531212, + 0.018705228343605995, + 0.033670686185359955, + 0.024420982226729393, + 0.01461555901914835, + 0.02111200988292694, + 0.015622175298631191, + 0.024901781231164932, + 0.012631561607122421, + 0.014260783791542053, + 0.019194234162569046, + 0.015036968514323235, + 0.01410187128931284, + 0.014399874024093151, + 0.023780759423971176, + 0.018113693222403526, + 0.017168408259749413, + 0.014859002083539963, + 0.029818642884492874, + 0.01222327258437872, + 0.0075943381525576115, + 0.010319462977349758, + 0.011210205964744091, + 0.01056789979338646, + 0.010543926618993282, + 0.010650546289980412, + 0.016979258507490158, + 0.008319483138620853, + 0.0034716855734586716, + 0.008939637802541256, + 0.005013431888073683, + 0.005485412199050188, + 0.00677480036392808, + 0.004526000469923019, + 0.006303766276687384, + 0.00464736158028245, + 0.007360956631600857, + 0.010869587771594524, + 0.010403012856841087, + 0.00902743823826313, + 0.010377597063779831, + 0.008665651082992554, + 0.014866858720779419, + 0.0074157193303108215, + 0.00605705427005887, + 0.007392848376184702, + 0.007568231783807278, + 0.006683495361357927, + 0.008925327099859715, + 0.006760396994650364, + 0.011581073515117168, + 0.008318645879626274, + 0.0033973052632063627, + 0.003267718944698572, + 0.0029223652090877295, + 0.004859765060245991, + 0.0033613049890846014, + 0.0037320605479180813, + 0.003927602432668209, + 0.0032408167608082294, + 0.008141908794641495, + 0.011033597402274609, + 0.013020484708249569, + 0.025315241888165474, + 0.01066538318991661, + 0.008702320046722889, + 0.013440380804240704, + 0.01010163314640522, + 0.004001527559012175, + 0.005474689416587353, + 0.004861738998442888, + 0.005187248345464468, + 0.00581557909026742, + 0.004558064043521881, + 0.004796938970685005, + 0.004909356124699116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.00978105328977108 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005249644163995981 + ], + "opposite_logprob_tokens": [ + -5.255249500274658 + ], + "logprob_reward": -0.005249644163995981, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005249644163995981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37180522084236145, + 0.10181526094675064, + 0.08349204808473587, + 0.05460356920957565, + 0.07431545853614807, + 0.06840722262859344, + 0.054722804576158524, + 0.08346100151538849, + 0.044033490121364594, + 0.06334397196769714 + ], + "credit_map": [ + 0.0383855476975441, + 0.07472018152475357, + 0.04865172132849693, + 0.10046640783548355, + 0.028820928186178207, + 0.016711050644516945, + 0.0334353931248188, + 0.03061399981379509, + 0.011736402288079262, + 0.014457013458013535, + 0.017713764682412148, + 0.019786348566412926, + 0.008595110848546028, + 0.008297758176922798, + 0.012800666503608227, + 0.008428191766142845, + 0.009719625115394592, + 0.012402573600411415, + 0.010912257246673107, + 0.01404842734336853, + 0.00850594975054264, + 0.006602741312235594, + 0.01352546364068985, + 0.007775008212774992, + 0.007366284262388945, + 0.014465770684182644, + 0.006336257327347994, + 0.008262516930699348, + 0.003703740891069174, + 0.004667013883590698, + 0.004224512260407209, + 0.005577476229518652, + 0.009757870808243752, + 0.013451224192976952, + 0.01054629497230053, + 0.010207593441009521, + 0.0063818651251494884, + 0.00838461797684431, + 0.0075698005966842175, + 0.008016186766326427, + 0.00912049412727356, + 0.00970709789544344, + 0.008602729998528957, + 0.009219945408403873, + 0.0067392876371741295, + 0.009957772679626942, + 0.006995311006903648, + 0.008064580149948597, + 0.005887352861464024, + 0.007147214841097593, + 0.007059871684759855, + 0.005756665952503681, + 0.006795067340135574, + 0.006301613990217447, + 0.0079950587823987, + 0.007779955398291349, + 0.010876540094614029, + 0.011783676221966743, + 0.010919672437012196, + 0.00905450526624918, + 0.008494524285197258, + 0.009590628556907177, + 0.011108162812888622, + 0.011633285321295261, + 0.004805786535143852, + 0.004633110482245684, + 0.005414480809122324, + 0.005299352575093508, + 0.004725108854472637, + 0.009043073281645775, + 0.004839323926717043, + 0.005273252725601196, + 0.007034946698695421, + 0.007571082562208176, + 0.0078084408305585384, + 0.007958775386214256, + 0.007412196137011051, + 0.009604771621525288, + 0.007447579875588417, + 0.008506180718541145 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.005249644163995981 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10027937591075897 + ], + "opposite_logprob_tokens": [ + -2.3502793312072754 + ], + "logprob_reward": -0.10027937591075897, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10027937591075897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3503592908382416, + 0.07309839129447937, + 0.07950232177972794, + 0.14095079898834229, + 0.0760841965675354, + 0.04566917568445206, + 0.031183116137981415, + 0.0877755880355835, + 0.04265287145972252, + 0.07272417843341827 + ], + "credit_map": [ + 0.03742295131087303, + 0.054943013936281204, + 0.05930028110742569, + 0.06565959751605988, + 0.033337660133838654, + 0.02339927852153778, + 0.040489498525857925, + 0.035807035863399506, + 0.008468200452625751, + 0.0056360624730587006, + 0.013934626244008541, + 0.010875781998038292, + 0.007801508530974388, + 0.006790996994823217, + 0.012458715587854385, + 0.0071324980817735195, + 0.00839091744273901, + 0.011994821019470692, + 0.013407457619905472, + 0.010791429318487644, + 0.00817672349512577, + 0.00722112413495779, + 0.012121233157813549, + 0.0073986114002764225, + 0.013736157678067684, + 0.020616095513105392, + 0.015676770359277725, + 0.02065308205783367, + 0.016263773664832115, + 0.011890769936144352, + 0.026912683621048927, + 0.01520145870745182, + 0.011157148517668247, + 0.00818303506821394, + 0.008911598473787308, + 0.009648947976529598, + 0.008837345987558365, + 0.006092555355280638, + 0.015624276362359524, + 0.007629291620105505, + 0.004025513771921396, + 0.006545442156493664, + 0.005902018863707781, + 0.0069030774757266045, + 0.004915506113320589, + 0.0046230158768594265, + 0.005978321190923452, + 0.0067762816324830055, + 0.0033730643335729837, + 0.003339342074468732, + 0.004039864521473646, + 0.004618060775101185, + 0.003781093517318368, + 0.003970501013100147, + 0.0036391548346728086, + 0.00442203413695097, + 0.011352306231856346, + 0.010844238102436066, + 0.012179292738437653, + 0.007828442379832268, + 0.01295449398458004, + 0.007407450582832098, + 0.015766089782118797, + 0.009443272836506367, + 0.005145153030753136, + 0.005021682009100914, + 0.004786822944879532, + 0.004964000545442104, + 0.005209862720221281, + 0.006205656100064516, + 0.005804780405014753, + 0.0055149151012301445, + 0.008440052159130573, + 0.010911515913903713, + 0.009844438172876835, + 0.009754310362040997, + 0.00811768975108862, + 0.008299337700009346, + 0.01017051748931408, + 0.007186317816376686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.10027937591075897 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the tranquil world of a solitary Ranchu goldfish, suspended in a glass aquarium with algae growth. This goldfish, with its distinctive egg-shaped body and lack of a dorsal fin, displays a vibrant pattern of deep orange and pearlescent white patches on its shimmering scales. Its prominent, bumpy head growth, or wen, partially obscures its dark, calm eyes, giving it a contemplative expression. The fish hovers almost motionless, its delicate pectoral fins making tiny, steady adjustments while its twin tail fins occasionally swish slowly, listlessly, propelling it aimlessly through the murky water. The tank's glass walls are coated in a soft green film, and long, ethereal strands of hair algae drift from the dark driftwood in the background. A faint light source from above penetrates the turquoise water, casting a hazy crepuscular glow that illuminates specks of floating debris and captures the metallic sheen of the goldfish's scales. The main color palette is a soft green and brown spectrum, which makes for a striking contrast to the bright colors of the goldfish, creating a strong atmosphere of tranquility and a profound sense of solitude.", + "seed": 22234, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_8875_seed_22234_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0437954006716609, + "vlm_logprob": -0.0437954006716609, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.0437954006716609, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.027398407459259033 + ], + "opposite_logprob_tokens": [ + -3.6523983478546143 + ], + "logprob_reward": -0.027398407459259033, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.027398407459259033, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2159273773431778, + 0.061156705021858215, + 0.07274305820465088, + 0.06216156482696533, + 0.0770384892821312, + 0.10500988364219666, + 0.13085079193115234, + 0.14196020364761353, + 0.0453076995909214, + 0.08784422278404236 + ], + "credit_map": [ + 0.03146772459149361, + 0.028376290574669838, + 0.03574138879776001, + 0.0286360215395689, + 0.017228761687874794, + 0.02528741955757141, + 0.027730319648981094, + 0.02145947329699993, + 0.007978983223438263, + 0.005944996140897274, + 0.00895786751061678, + 0.008157887496054173, + 0.005714583210647106, + 0.0091681694611907, + 0.008424559608101845, + 0.006809661164879799, + 0.011015219613909721, + 0.012558598071336746, + 0.010684655047953129, + 0.008442215621471405, + 0.006088638212531805, + 0.008510627783834934, + 0.007250174414366484, + 0.008192933164536953, + 0.007532505318522453, + 0.004137629177421331, + 0.007077046670019627, + 0.005375307984650135, + 0.005576668307185173, + 0.01696613058447838, + 0.006123927887529135, + 0.009372349828481674, + 0.009786558337509632, + 0.006886136718094349, + 0.009224442765116692, + 0.01251116581261158, + 0.008090143091976643, + 0.013080994598567486, + 0.009442126378417015, + 0.0080169178545475, + 0.016072245314717293, + 0.009125987999141216, + 0.017395690083503723, + 0.015330079942941666, + 0.008380889892578125, + 0.014395399950444698, + 0.011527814902365208, + 0.0127817802131176, + 0.020134393125772476, + 0.010396965779364109, + 0.01647168956696987, + 0.014916409738361835, + 0.016027923673391342, + 0.026914633810520172, + 0.01059939805418253, + 0.015389371663331985, + 0.01587512344121933, + 0.016170045360922813, + 0.0144615164026618, + 0.015585191547870636, + 0.018369875848293304, + 0.032613981515169144, + 0.01693781279027462, + 0.011946657672524452, + 0.005783108063042164, + 0.0036926635075360537, + 0.006470425985753536, + 0.004866358824074268, + 0.006915947422385216, + 0.005990108009427786, + 0.005686883348971605, + 0.005902205593883991, + 0.009585866704583168, + 0.012203254736959934, + 0.007913574576377869, + 0.01124065462499857, + 0.012673868797719479, + 0.012314884923398495, + 0.011160293594002724, + 0.010751820169389248 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.027398407459259033 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01608455553650856 + ], + "opposite_logprob_tokens": [ + -4.141084671020508 + ], + "logprob_reward": -0.01608455553650856, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01608455553650856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20985616743564606, + 0.10362765192985535, + 0.0731113851070404, + 0.06679119914770126, + 0.05296589061617851, + 0.07534482330083847, + 0.0378083735704422, + 0.15967102348804474, + 0.06061170995235443, + 0.16021180152893066 + ], + "credit_map": [ + 0.028196999803185463, + 0.04078152775764465, + 0.03123318776488304, + 0.023679884150624275, + 0.022190622985363007, + 0.02021240070462227, + 0.022924000397324562, + 0.020637528970837593, + 0.01518949493765831, + 0.010182982310652733, + 0.01784149743616581, + 0.010621833615005016, + 0.011996056884527206, + 0.012843926437199116, + 0.013413077220320702, + 0.011538790538907051, + 0.011161770671606064, + 0.008220293559134007, + 0.010401167906820774, + 0.007369608152657747, + 0.0075338720344007015, + 0.008492019027471542, + 0.008350875228643417, + 0.011581775732338428, + 0.014997740276157856, + 0.004514005966484547, + 0.010334057733416557, + 0.00766910333186388, + 0.004707425832748413, + 0.011343758553266525, + 0.006264065392315388, + 0.0069610378704965115, + 0.007586462423205376, + 0.004356769844889641, + 0.0050592259503901005, + 0.006716863717883825, + 0.003963269293308258, + 0.01318968553096056, + 0.005725313909351826, + 0.006368303671479225, + 0.011300311423838139, + 0.005879098549485207, + 0.011987976729869843, + 0.009217189624905586, + 0.0072258575819432735, + 0.011765575036406517, + 0.009455804713070393, + 0.008513012900948524, + 0.004951006732881069, + 0.0032251509837806225, + 0.003619343973696232, + 0.004305094946175814, + 0.00462898425757885, + 0.008382751606404781, + 0.004706101026386023, + 0.003989940043538809, + 0.021051352843642235, + 0.013887007720768452, + 0.013727176934480667, + 0.018838604912161827, + 0.018018050119280815, + 0.03735846281051636, + 0.015422487631440163, + 0.021367862820625305, + 0.009095978923141956, + 0.0052911206148564816, + 0.008033204823732376, + 0.008383916690945625, + 0.006525177508592606, + 0.008818975649774075, + 0.008311119861900806, + 0.006152215413749218, + 0.02393096685409546, + 0.01901920884847641, + 0.013014620169997215, + 0.017537148669362068, + 0.021048689261078835, + 0.021869895979762077, + 0.018019868060946465, + 0.02577139250934124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.01608455553650856 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011073680594563484 + ], + "opposite_logprob_tokens": [ + -4.511073589324951 + ], + "logprob_reward": -0.011073680594563484, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011073680594563484, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17939913272857666, + 0.20472487807273865, + 0.08466484397649765, + 0.07351244240999222, + 0.06686478108167648, + 0.08830247074365616, + 0.1519887000322342, + 0.03727231174707413, + 0.044623229652643204, + 0.06864724308252335 + ], + "credit_map": [ + 0.019536977633833885, + 0.02986839972436428, + 0.027358314022421837, + 0.02575446479022503, + 0.016399672254920006, + 0.022080250084400177, + 0.023177769035100937, + 0.015223290771245956, + 0.01977265626192093, + 0.013247030787169933, + 0.028233839198946953, + 0.03081382066011429, + 0.02200286090373993, + 0.030217397958040237, + 0.034016888588666916, + 0.02642037905752659, + 0.010431372560560703, + 0.007051839493215084, + 0.012392672710120678, + 0.010889108292758465, + 0.008862827904522419, + 0.011382097378373146, + 0.011804365552961826, + 0.011850561946630478, + 0.008791087195277214, + 0.006690678186714649, + 0.00862615741789341, + 0.01066899299621582, + 0.006271821912378073, + 0.014813252724707127, + 0.008862261660397053, + 0.00878819264471531, + 0.008411962538957596, + 0.006909871939569712, + 0.007101776544004679, + 0.009469454176723957, + 0.006493718363344669, + 0.013130277395248413, + 0.00761064700782299, + 0.00773707265034318, + 0.012086652219295502, + 0.008330713026225567, + 0.013051053509116173, + 0.013389693573117256, + 0.007412488106638193, + 0.014278165064752102, + 0.010288074612617493, + 0.009465627372264862, + 0.01871798001229763, + 0.01592542789876461, + 0.018447617068886757, + 0.0180220827460289, + 0.0260607972741127, + 0.020112473517656326, + 0.01746722310781479, + 0.01723509281873703, + 0.005673275329172611, + 0.0041432008147239685, + 0.0047758170403540134, + 0.005209073424339294, + 0.0034647199790924788, + 0.004426126834005117, + 0.006098870187997818, + 0.0034812290687114, + 0.006025044713169336, + 0.0031885954085737467, + 0.007204809226095676, + 0.005219252780079842, + 0.005267108790576458, + 0.00600167503580451, + 0.006361468695104122, + 0.005355273373425007, + 0.008421473205089569, + 0.005840288009494543, + 0.008131736889481544, + 0.008298803120851517, + 0.008554993197321892, + 0.011380486190319061, + 0.010038773529231548, + 0.00798068568110466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.011073680594563484 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091345705091953 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091345705091953, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091345705091953, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3165924549102783, + 0.05178508907556534, + 0.09344455599784851, + 0.03901539370417595, + 0.11811350286006927, + 0.09808500856161118, + 0.061097756028175354, + 0.07341022044420242, + 0.08463263511657715, + 0.06382341682910919 + ], + "credit_map": [ + 0.042041532695293427, + 0.045255426317453384, + 0.05647169426083565, + 0.04844983294606209, + 0.024617692455649376, + 0.029246598482131958, + 0.04344400763511658, + 0.02706562541425228, + 0.00558923464268446, + 0.005174640100449324, + 0.007030317559838295, + 0.007003724109381437, + 0.005268174689263105, + 0.006764330435544252, + 0.008326193317770958, + 0.0066284737549722195, + 0.014563178643584251, + 0.016965987160801888, + 0.01207887101918459, + 0.008835994638502598, + 0.007980349473655224, + 0.009100514464080334, + 0.010500562377274036, + 0.013419091701507568, + 0.0066404263488948345, + 0.003526014043018222, + 0.005100612994283438, + 0.003893292974680662, + 0.0031604082323610783, + 0.008244654163718224, + 0.004495392553508282, + 0.003954590298235416, + 0.01430668868124485, + 0.01223466545343399, + 0.012227572500705719, + 0.01747293956577778, + 0.01293659396469593, + 0.02287956140935421, + 0.014347716234624386, + 0.011707760393619537, + 0.013620706275105476, + 0.008500857278704643, + 0.013767571188509464, + 0.01327259186655283, + 0.009361013770103455, + 0.013837037608027458, + 0.013764992356300354, + 0.011960229836404324, + 0.01061658188700676, + 0.007241859566420317, + 0.006989484652876854, + 0.006688421592116356, + 0.006384748034179211, + 0.009884554892778397, + 0.006227416917681694, + 0.00706468103453517, + 0.009409364312887192, + 0.00601228978484869, + 0.008009830489754677, + 0.008565998636186123, + 0.011095508933067322, + 0.013721106573939323, + 0.008217930793762207, + 0.008378181606531143, + 0.011884115636348724, + 0.004766767844557762, + 0.013443912379443645, + 0.009133358485996723, + 0.011990011669695377, + 0.009396770037710667, + 0.012137942016124725, + 0.011879750527441502, + 0.008563521318137646, + 0.0058819870464503765, + 0.009056327864527702, + 0.010538679547607899, + 0.007605697028338909, + 0.007137275766581297, + 0.008073694072663784, + 0.006966226268559694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.004091345705091953 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16032901406288147 + ], + "opposite_logprob_tokens": [ + -1.910328984260559 + ], + "logprob_reward": -0.16032901406288147, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16032901406288147, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1761496514081955, + 0.04146251827478409, + 0.041432205587625504, + 0.04606735706329346, + 0.12561637163162231, + 0.11622148752212524, + 0.16073547303676605, + 0.17751853168010712, + 0.06475215405225754, + 0.0500442311167717 + ], + "credit_map": [ + 0.02142351120710373, + 0.027763275429606438, + 0.03333864361047745, + 0.024235421791672707, + 0.01646251603960991, + 0.013058427721261978, + 0.020809724926948547, + 0.019058121368288994, + 0.0049078501760959625, + 0.006788740400224924, + 0.005963024217635393, + 0.004421830177307129, + 0.004309140145778656, + 0.0033407090231776237, + 0.005822698585689068, + 0.005908522754907608, + 0.007051725406199694, + 0.007602256257086992, + 0.005439189728349447, + 0.005103239789605141, + 0.0037550132256001234, + 0.003569288644939661, + 0.004268523305654526, + 0.004642967134714127, + 0.00643585529178381, + 0.00491403229534626, + 0.005066318437457085, + 0.006375741679221392, + 0.004692340735346079, + 0.0075408159755170345, + 0.006400934886187315, + 0.004641313571482897, + 0.020210498943924904, + 0.01207993645220995, + 0.01612427458167076, + 0.017695052549242973, + 0.011074936017394066, + 0.019922327250242233, + 0.014240001328289509, + 0.014269331470131874, + 0.01945776492357254, + 0.008993654511868954, + 0.016794195398688316, + 0.01627480424940586, + 0.007772637996822596, + 0.0237919669598341, + 0.011251023039221764, + 0.011885436251759529, + 0.021989375352859497, + 0.016765668988227844, + 0.01769542321562767, + 0.01890634000301361, + 0.029454560950398445, + 0.021228067576885223, + 0.017725005745887756, + 0.016971029341220856, + 0.024996604770421982, + 0.013823829591274261, + 0.024973668158054352, + 0.021132130175828934, + 0.029238086193799973, + 0.02260860614478588, + 0.020618295297026634, + 0.020127300173044205, + 0.008281332440674305, + 0.004846533760428429, + 0.009728516452014446, + 0.0072508095763623714, + 0.009087606333196163, + 0.007593086455017328, + 0.009495193138718605, + 0.008469070307910442, + 0.0061819301918148994, + 0.0037416464183479548, + 0.010082314722239971, + 0.006935490760952234, + 0.0056642284616827965, + 0.006571653764694929, + 0.005089557729661465, + 0.005777404643595219 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 6 + ], + "selected_reward": -0.16032901406288147 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level close-up captures the tranquil world of a solitary Ranchu goldfish, suspended in a glass aquarium with algae growth. This goldfish, with its distinctive egg-shaped body and lack of a dorsal fin, displays a vibrant pattern of deep orange and pearlescent white patches on its shimmering scales. Its prominent, bumpy head growth, or wen, partially obscures its dark, calm eyes, giving it a contemplative expression. The fish hovers almost motionless, its delicate pectoral fins making tiny, steady adjustments while its twin tail fins occasionally swish slowly, listlessly, propelling it aimlessly through the murky water. The tank's glass walls are coated in a soft green film, and long, ethereal strands of hair algae drift from the dark driftwood in the background. A faint light source from above penetrates the turquoise water, casting a hazy crepuscular glow that illuminates specks of floating debris and captures the metallic sheen of the goldfish's scales. The main color palette is a soft green and brown spectrum, which makes for a striking contrast to the bright colors of the goldfish, creating a strong atmosphere of tranquility and a profound sense of solitude.", + "seed": 22235, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_4/index_8875_seed_22235_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02425012793391943, + "vlm_logprob": -0.02425012793391943, + "vlm_yes_no_margin": 4.375, + "vlm_token_credit": -0.02425012793391943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009460381232202053 + ], + "opposite_logprob_tokens": [ + -4.75946044921875 + ], + "logprob_reward": -0.009460381232202053, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.009460381232202053, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2805096507072449, + 0.07020003348588943, + 0.046229735016822815, + 0.0836416706442833, + 0.047780901193618774, + 0.05956946685910225, + 0.15589448809623718, + 0.0586189329624176, + 0.15321844816207886, + 0.04433669522404671 + ], + "credit_map": [ + 0.046499427407979965, + 0.02722601220011711, + 0.04219486191868782, + 0.03190788999199867, + 0.02339118905365467, + 0.0488847978413105, + 0.026785634458065033, + 0.03361982852220535, + 0.013495994731783867, + 0.006886846851557493, + 0.010042085312306881, + 0.008901233784854412, + 0.00517655722796917, + 0.012490563094615936, + 0.006934920325875282, + 0.006271834950894117, + 0.009960973635315895, + 0.003706979565322399, + 0.006515749264508486, + 0.005984900984913111, + 0.004498897586017847, + 0.004944398999214172, + 0.0042414916679263115, + 0.006376341916620731, + 0.012256139889359474, + 0.005246098153293133, + 0.011465795338153839, + 0.013681963086128235, + 0.00523131899535656, + 0.0162022914737463, + 0.010704023763537407, + 0.008854040876030922, + 0.009009983390569687, + 0.005466712173074484, + 0.005709843244403601, + 0.004775578621774912, + 0.004904297646135092, + 0.006174860987812281, + 0.005253486335277557, + 0.00648614251986146, + 0.010090474039316177, + 0.006323904264718294, + 0.008302628993988037, + 0.008542025461792946, + 0.00333777186460793, + 0.009628198109567165, + 0.006663456559181213, + 0.0066810050047934055, + 0.031381603330373764, + 0.015145980753004551, + 0.021852809935808182, + 0.025006648153066635, + 0.012976658530533314, + 0.020349761471152306, + 0.01430192869156599, + 0.014879083260893822, + 0.007347840815782547, + 0.006343265064060688, + 0.009114330634474754, + 0.0066856397315859795, + 0.005840395111590624, + 0.007489397190511227, + 0.009803617373108864, + 0.00599445216357708, + 0.026038827374577522, + 0.015488514676690102, + 0.01920331083238125, + 0.017723601311445236, + 0.012636767700314522, + 0.02254333347082138, + 0.01719696633517742, + 0.022387132048606873, + 0.00722489831969142, + 0.004636881407350302, + 0.005467159673571587, + 0.004608913790434599, + 0.0052080596797168255, + 0.008181029930710793, + 0.005177578888833523, + 0.0038321742322295904 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.009460381232202053 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028305486775934696 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.0028305486775934696, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028305486775934696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10568773001432419, + 0.07143425196409225, + 0.07729451358318329, + 0.08162336051464081, + 0.06925077736377716, + 0.03880388289690018, + 0.09729285538196564, + 0.12095814198255539, + 0.23668015003204346, + 0.10097435116767883 + ], + "credit_map": [ + 0.015224199742078781, + 0.008609892800450325, + 0.015629926696419716, + 0.015538105741143227, + 0.008623704314231873, + 0.01317562535405159, + 0.013946449384093285, + 0.014939826913177967, + 0.014568683691322803, + 0.004789587110280991, + 0.01052874606102705, + 0.007404209114611149, + 0.00488649308681488, + 0.009988157078623772, + 0.008063450455665588, + 0.01120492722839117, + 0.010174062103033066, + 0.007657049689441919, + 0.011728515848517418, + 0.009573139250278473, + 0.007434370927512646, + 0.012370012700557709, + 0.010485232807695866, + 0.007872129790484905, + 0.018698422238230705, + 0.005724486894905567, + 0.009425998665392399, + 0.009110638871788979, + 0.00641118548810482, + 0.010394943878054619, + 0.007043692748993635, + 0.014813987538218498, + 0.008460523560643196, + 0.006355345714837313, + 0.01074990350753069, + 0.01209980621933937, + 0.005098193418234587, + 0.009341520257294178, + 0.009636687114834785, + 0.007508798502385616, + 0.006735885515809059, + 0.005371284205466509, + 0.005280456971377134, + 0.005033858586102724, + 0.003777403151616454, + 0.003888447303324938, + 0.0046307360753417015, + 0.0040858094580471516, + 0.01282871887087822, + 0.010956636629998684, + 0.01566113531589508, + 0.014786484651267529, + 0.007615599315613508, + 0.011436075903475285, + 0.012038998305797577, + 0.011969207786023617, + 0.015176682732999325, + 0.015448360703885555, + 0.016614841297268867, + 0.013612787239253521, + 0.01514276210218668, + 0.01355927437543869, + 0.013929974287748337, + 0.017473449930548668, + 0.04333419352769852, + 0.024070631712675095, + 0.0334644578397274, + 0.031234726309776306, + 0.018106890842318535, + 0.025345424190163612, + 0.02557956427335739, + 0.03554427623748779, + 0.013381130993366241, + 0.013056200928986073, + 0.012708163820207119, + 0.012411728501319885, + 0.011797855608165264, + 0.018123788759112358, + 0.008387994952499866, + 0.011107485741376877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 0 + ], + "selected_reward": -0.0028305486775934696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023280736058950424 + ], + "opposite_logprob_tokens": [ + -3.773280620574951 + ], + "logprob_reward": -0.023280736058950424, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023280736058950424, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13543294370174408, + 0.04231053590774536, + 0.07473691552877426, + 0.08725859224796295, + 0.055858857929706573, + 0.1326046735048294, + 0.1159278154373169, + 0.07136759907007217, + 0.19114524126052856, + 0.09335680305957794 + ], + "credit_map": [ + 0.029793623834848404, + 0.00831016805022955, + 0.020486777648329735, + 0.017621386796236038, + 0.0068609099835157394, + 0.019652720540761948, + 0.015779666602611542, + 0.01692771539092064, + 0.006662069819867611, + 0.0043044849298894405, + 0.006183044984936714, + 0.006984814070165157, + 0.003093498758971691, + 0.004940280225127935, + 0.004518633242696524, + 0.005623712204396725, + 0.014183550141751766, + 0.006554283667355776, + 0.010637992061674595, + 0.00855758972465992, + 0.004239355213940144, + 0.010510814376175404, + 0.008654371835291386, + 0.011398964561522007, + 0.017750060185790062, + 0.005786119494587183, + 0.013093989342451096, + 0.011537101119756699, + 0.005864896811544895, + 0.012835251167416573, + 0.007984918542206287, + 0.012406258843839169, + 0.00816695112735033, + 0.004579836502671242, + 0.007317736279219389, + 0.009303281083703041, + 0.003546197433024645, + 0.00484785670414567, + 0.009619641117751598, + 0.00847736094146967, + 0.029068121686577797, + 0.010605709627270699, + 0.01969948224723339, + 0.017334068194031715, + 0.010455549694597721, + 0.019146079197525978, + 0.013620554469525814, + 0.012675128877162933, + 0.01346869207918644, + 0.010769854299724102, + 0.01835799776017666, + 0.019934862852096558, + 0.009063620120286942, + 0.01916341856122017, + 0.014494223520159721, + 0.010675160214304924, + 0.006536268629133701, + 0.007552689407020807, + 0.013618048280477524, + 0.012414167635142803, + 0.006950348149985075, + 0.009183045476675034, + 0.008477622643113136, + 0.006635418627411127, + 0.036486055701971054, + 0.01628856547176838, + 0.025353917852044106, + 0.025094415992498398, + 0.014027994126081467, + 0.0221091378480196, + 0.021044479683041573, + 0.03074069879949093, + 0.014741632156074047, + 0.00711388373747468, + 0.013424382545053959, + 0.0104842409491539, + 0.010117491707205772, + 0.017486289143562317, + 0.008796933107078075, + 0.011191959492862225 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 5 + ], + "selected_reward": -0.023280736058950424 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006730505730956793 + ], + "opposite_logprob_tokens": [ + -5.006730556488037 + ], + "logprob_reward": -0.006730505730956793, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006730505730956793, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16347941756248474, + 0.08363225311040878, + 0.062303576618433, + 0.08453895896673203, + 0.05151428282260895, + 0.20922333002090454, + 0.0779922604560852, + 0.10977303236722946, + 0.08417265862226486, + 0.07337027788162231 + ], + "credit_map": [ + 0.032989293336868286, + 0.012939690612256527, + 0.0240543894469738, + 0.022235363721847534, + 0.01124810241162777, + 0.026453236117959023, + 0.017716966569423676, + 0.01584235206246376, + 0.022707130759954453, + 0.006097517441958189, + 0.014849018305540085, + 0.009351450949907303, + 0.005270321853458881, + 0.008577796630561352, + 0.009150932542979717, + 0.007628076244145632, + 0.01908852718770504, + 0.004597761202603579, + 0.00953851081430912, + 0.008029811084270477, + 0.0037863655015826225, + 0.006076734513044357, + 0.003738769795745611, + 0.007447090465575457, + 0.01788870431482792, + 0.004844106733798981, + 0.0156227583065629, + 0.010392524302005768, + 0.004978367127478123, + 0.012339566834270954, + 0.007431581616401672, + 0.011041342280805111, + 0.007761286571621895, + 0.0035954725462943316, + 0.0063107614405453205, + 0.00818390492349863, + 0.0031054106075316668, + 0.007915647700428963, + 0.0047914511524140835, + 0.00985034555196762, + 0.04581804946064949, + 0.016407927498221397, + 0.032276544719934464, + 0.026864871382713318, + 0.018110085278749466, + 0.0296695064753294, + 0.020092206075787544, + 0.019984109327197075, + 0.01397060789167881, + 0.00674875034019351, + 0.010923946276307106, + 0.009626070968806744, + 0.004906403366476297, + 0.013430312275886536, + 0.009367765858769417, + 0.00901839230209589, + 0.011131797917187214, + 0.007473616860806942, + 0.019276263192296028, + 0.022624842822551727, + 0.007363153621554375, + 0.018212534487247467, + 0.014233563095331192, + 0.009457250125706196, + 0.011824405752122402, + 0.006081311497837305, + 0.013703121803700924, + 0.014941918663680553, + 0.005499499384313822, + 0.013333615846931934, + 0.00945821963250637, + 0.009330558590590954, + 0.007527508307248354, + 0.004794256296008825, + 0.011031504720449448, + 0.008762550540268421, + 0.006884437054395676, + 0.01596330851316452, + 0.01083572581410408, + 0.00757098151370883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.006730505730956793 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a solitary Ranchu goldfish with an egg-shaped body and no dorsal fin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the goldfish display a vibrant pattern of deep orange and pearlescent white patches on its scales?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the goldfish's head covered by a prominent, bumpy growth (wen) that partially obscures its eyes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the setting a glass aquarium containing dark driftwood, long strands of hair algae, and murky turquoise water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera positioned at eye-level in a static close-up shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07894846796989441 + ], + "opposite_logprob_tokens": [ + -2.578948497772217 + ], + "logprob_reward": -0.07894846796989441, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07894846796989441, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20303045213222504, + 0.04458250105381012, + 0.05371475964784622, + 0.04907839745283127, + 0.1194755882024765, + 0.13907858729362488, + 0.11753083020448685, + 0.04429693892598152, + 0.1384837031364441, + 0.09072820842266083 + ], + "credit_map": [ + 0.038402024656534195, + 0.02021263726055622, + 0.02978650853037834, + 0.02753601036965847, + 0.016219796612858772, + 0.029040327295660973, + 0.01925811544060707, + 0.022575030103325844, + 0.010231105610728264, + 0.0028799711726605892, + 0.005520970560610294, + 0.005667548161000013, + 0.002700167242437601, + 0.006610647309571505, + 0.004680330399423838, + 0.006291764322668314, + 0.009774361737072468, + 0.004553359467536211, + 0.009528061375021935, + 0.006801025476306677, + 0.003602187614887953, + 0.006940321531146765, + 0.005903261713683605, + 0.00661217886954546, + 0.011320295743644238, + 0.0033606351353228092, + 0.007577992510050535, + 0.006377786863595247, + 0.0027519857976585627, + 0.005930040962994099, + 0.005376057233661413, + 0.006383602973073721, + 0.018669327720999718, + 0.012888484634459019, + 0.015918126329779625, + 0.018397707492113113, + 0.010955704376101494, + 0.018383868038654327, + 0.013248482719063759, + 0.011013887822628021, + 0.03108220174908638, + 0.01017650868743658, + 0.02302572876214981, + 0.01849261485040188, + 0.01063706073909998, + 0.01941242255270481, + 0.012637711130082607, + 0.01361433882266283, + 0.02465886063873768, + 0.011052506975829601, + 0.018116828054189682, + 0.01897694170475006, + 0.008764469064772129, + 0.01543206162750721, + 0.010450408793985844, + 0.01007874682545662, + 0.006270055193454027, + 0.003222088562324643, + 0.006363590247929096, + 0.004722961224615574, + 0.0031563350930809975, + 0.004586797207593918, + 0.009105309844017029, + 0.0068697999231517315, + 0.024210508912801743, + 0.012664774432778358, + 0.017674051225185394, + 0.01818206161260605, + 0.010417145676910877, + 0.01894972287118435, + 0.015887830406427383, + 0.02049761451780796, + 0.018762607127428055, + 0.004189983941614628, + 0.014325357973575592, + 0.010775668546557426, + 0.006471049971878529, + 0.017305532470345497, + 0.007077950052917004, + 0.011820056475698948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.07894846796989441 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_tensor([9503, 9503])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_tensor([9503, 9503])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..ddd6e5511fe55eb145ec2a5ab02708e85909a87c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_tensor([9503, 9503])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot, shot from a slightly upward angle, captures a brilliant golden ginkgo leaf, slowly swirling and falling from the sky. Warm afternoon sunlight penetrates the sparse canopy, softly backlighting the leaf, making its thin, cicada-wing texture appear crystal clear, its signature scalloped edges shimmering with a dazzling halo. The camera pans forward with an incredibly smooth and steady pace, following the leaf's leisurely descent as it gracefully and deliberately spins through the air, as if performing a silent solo. As the camera zooms in, the image gradually shifts from a mid-ground depicting the falling leaf's movement to an extreme close-up, revealing the leaf's forked veins, like an intricate map. Each fine vein shimmers subtly in the light, its edges tinged with a caramel-brown hue. The scene is enveloped in a hazy, shallow depth of field, while the blurred autumn forest colors in the background contrast sharply with the subject, creating a tranquil, beautiful, and poetic atmosphere.", + "seed": 26259, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_9503_seed_26259_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.049687453312799335, + "vlm_logprob": -0.049687453312799335, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.049687453312799335, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014706186018884182 + ], + "opposite_logprob_tokens": [ + -4.514706134796143 + ], + "logprob_reward": -0.014706186018884182, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.014706186018884182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17031441628932953, + 0.04683220759034157, + 0.09135578572750092, + 0.06874976307153702, + 0.3570594787597656, + 0.1744793802499771, + 0.02577466145157814, + 0.01330525428056717, + 0.02787105180323124, + 0.024258015677332878 + ], + "credit_map": [ + 0.010949837043881416, + 0.01553569920361042, + 0.024031687527894974, + 0.034620024263858795, + 0.02084388956427574, + 0.01722472533583641, + 0.021141281351447105, + 0.025967257097363472, + 0.004410178400576115, + 0.004301338456571102, + 0.009453555569052696, + 0.00860509742051363, + 0.004147171508520842, + 0.0053320531733334064, + 0.005817710421979427, + 0.004765097983181477, + 0.005081137642264366, + 0.00605690898373723, + 0.031258415430784225, + 0.02509542740881443, + 0.006021774373948574, + 0.006984212435781956, + 0.004959261044859886, + 0.005898638628423214, + 0.004897784441709518, + 0.00822617020457983, + 0.012637504376471043, + 0.010334780439734459, + 0.007568757515400648, + 0.011320904828608036, + 0.0078597916290164, + 0.005904067307710648, + 0.02333115041255951, + 0.013679523020982742, + 0.1517379879951477, + 0.08907709270715714, + 0.006109203211963177, + 0.041021209210157394, + 0.010054605081677437, + 0.022048674523830414, + 0.006871869787573814, + 0.012224063277244568, + 0.0381571501493454, + 0.07416396588087082, + 0.005227771587669849, + 0.015198628418147564, + 0.005707699805498123, + 0.01692822389304638, + 0.0022298553958535194, + 0.004480625968426466, + 0.005152410827577114, + 0.0033275990281254053, + 0.0024254319723695517, + 0.0038317411672323942, + 0.0014866493875160813, + 0.0028403468895703554, + 0.001187172601930797, + 0.0015274271136149764, + 0.002192517276853323, + 0.001862795208580792, + 0.0012196886818856, + 0.0022064191289246082, + 0.0014038097579032183, + 0.0017054227646440268, + 0.002637289697304368, + 0.002926318207755685, + 0.003755129175260663, + 0.004557361360639334, + 0.0036771814338862896, + 0.0036129208747297525, + 0.0021634548902511597, + 0.004541392903774977, + 0.002630353206768632, + 0.002111604670062661, + 0.002774677472189069, + 0.002519635483622551, + 0.002642917213961482, + 0.00457883533090353, + 0.0031369163189083338, + 0.0038630706258118153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.014706186018884182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270446926355362 + ], + "opposite_logprob_tokens": [ + -2.127044677734375 + ], + "logprob_reward": -0.1270446926355362, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270446926355362, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17180706560611725, + 0.03582713007926941, + 0.3252076804637909, + 0.12386522442102432, + 0.1635771095752716, + 0.048798128962516785, + 0.03204163908958435, + 0.015282568521797657, + 0.03315030783414841, + 0.050443198531866074 + ], + "credit_map": [ + 0.007419095374643803, + 0.014112737029790878, + 0.018458496779203415, + 0.05461316555738449, + 0.018842656165361404, + 0.01243811659514904, + 0.02152455598115921, + 0.02439824864268303, + 0.0034451524261385202, + 0.0033369313459843397, + 0.00689280079677701, + 0.007658947724848986, + 0.0028496473096311092, + 0.0040964786894619465, + 0.003880611853674054, + 0.0036665580701082945, + 0.014366762712597847, + 0.017263107001781464, + 0.0911012813448906, + 0.12446559220552444, + 0.011687894351780415, + 0.020922347903251648, + 0.026081185787916183, + 0.019319510087370872, + 0.009832660667598248, + 0.009631113149225712, + 0.027870692312717438, + 0.01650194451212883, + 0.017022430896759033, + 0.01740991324186325, + 0.014902302995324135, + 0.010694156400859356, + 0.008841563947498798, + 0.014591752551496029, + 0.06119294837117195, + 0.04174422100186348, + 0.00606512650847435, + 0.015212923288345337, + 0.007011058274656534, + 0.008917511440813541, + 0.002358193974941969, + 0.004621095955371857, + 0.008952329866588116, + 0.017895422875881195, + 0.003203908447176218, + 0.004608702380210161, + 0.003265785053372383, + 0.00389269320294261, + 0.003799966536462307, + 0.004986538551747799, + 0.0044981990940868855, + 0.004266634117811918, + 0.0026823151856660843, + 0.006120692938566208, + 0.002895490964874625, + 0.0027918019331991673, + 0.0011456772917881608, + 0.0015306241111829877, + 0.0020541062112897635, + 0.0028040397446602583, + 0.0017616139957681298, + 0.00233128247782588, + 0.0016773404786363244, + 0.001977883279323578, + 0.0037374894600361586, + 0.003475944045931101, + 0.004959124606102705, + 0.0037886209320276976, + 0.004255627281963825, + 0.004506929311901331, + 0.002322434214875102, + 0.006104137282818556, + 0.005368932615965605, + 0.005686040967702866, + 0.0058332784101367, + 0.003924134653061628, + 0.00830369908362627, + 0.009341442957520485, + 0.003998910542577505, + 0.007986757904291153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 4 + ], + "selected_reward": -0.1270446926355362 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031949449330568314 + ], + "opposite_logprob_tokens": [ + -5.753194808959961 + ], + "logprob_reward": -0.0031949449330568314, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031949449330568314, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11061431467533112, + 0.3225095272064209, + 0.12964215874671936, + 0.051764458417892456, + 0.16886499524116516, + 0.11154866963624954, + 0.035359937697649, + 0.018032142892479897, + 0.02882910519838333, + 0.022834744304418564 + ], + "credit_map": [ + 0.008904711343348026, + 0.012083347886800766, + 0.01899336650967598, + 0.020719420164823532, + 0.008671061135828495, + 0.011730002239346504, + 0.012919319793581963, + 0.016593087464571, + 0.027653636410832405, + 0.03198889270424843, + 0.05216488614678383, + 0.07509052008390427, + 0.030873361974954605, + 0.034916359931230545, + 0.039756890386343, + 0.03006497211754322, + 0.00794136244803667, + 0.017214588820934296, + 0.03624703735113144, + 0.03239452838897705, + 0.004626404494047165, + 0.011533142998814583, + 0.009961295872926712, + 0.009723802097141743, + 0.00407882547006011, + 0.0059535619802773, + 0.007930425927042961, + 0.01160059031099081, + 0.00401836633682251, + 0.008827069774270058, + 0.00512891449034214, + 0.004226704593747854, + 0.010738853365182877, + 0.00819015596061945, + 0.06966517865657806, + 0.042112309485673904, + 0.0038751494139432907, + 0.021309033036231995, + 0.0042067826725542545, + 0.008767538703978062, + 0.004135987721383572, + 0.009880078956484795, + 0.023449549451470375, + 0.04794919490814209, + 0.0029012064915150404, + 0.009642091579735279, + 0.003393343184143305, + 0.010197216644883156, + 0.004187093582004309, + 0.0050654662773013115, + 0.005275676026940346, + 0.0043819681741297245, + 0.002216695575043559, + 0.008061378262937069, + 0.0019401288591325283, + 0.004231531172990799, + 0.0010312814265489578, + 0.0016344035975635052, + 0.0029681273736059666, + 0.0021178051829338074, + 0.002788137411698699, + 0.0023493533954024315, + 0.0013866627123206854, + 0.0037563710939139128, + 0.0019515103194862604, + 0.00258467486128211, + 0.003734615398570895, + 0.005755595862865448, + 0.0032146526500582695, + 0.0029270413797348738, + 0.0032910217996686697, + 0.005369993392378092, + 0.0033048682380467653, + 0.003318505361676216, + 0.003156693186610937, + 0.0019315672107040882, + 0.00232190964743495, + 0.0028414905536919832, + 0.002418144838884473, + 0.003541564801707864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 2 + ], + "selected_reward": -0.0031949449330568314 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10030137002468109 + ], + "opposite_logprob_tokens": [ + -2.3503012657165527 + ], + "logprob_reward": -0.10030137002468109, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10030137002468109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20691896975040436, + 0.04410291090607643, + 0.1181386262178421, + 0.12375672161579132, + 0.15451008081436157, + 0.10930227488279343, + 0.04803670197725296, + 0.025134677067399025, + 0.04660657420754433, + 0.12349246442317963 + ], + "credit_map": [ + 0.01547620352357626, + 0.020893236622214317, + 0.030597420409321785, + 0.03279022127389908, + 0.022875692695379257, + 0.02532309666275978, + 0.027415676042437553, + 0.03154740482568741, + 0.0031280014663934708, + 0.0030613690614700317, + 0.006841633468866348, + 0.012830891646444798, + 0.002140116412192583, + 0.003668086836114526, + 0.00424146605655551, + 0.008191348984837532, + 0.005015066359192133, + 0.009532312862575054, + 0.03182414174079895, + 0.041205696761608124, + 0.004495847038924694, + 0.008047404699027538, + 0.009249466471374035, + 0.00876869447529316, + 0.00893543940037489, + 0.012988559901714325, + 0.022626405581831932, + 0.02481159381568432, + 0.007246897555887699, + 0.022780615836381912, + 0.013653638772666454, + 0.01071357075124979, + 0.009609073400497437, + 0.013016043230891228, + 0.05286234989762306, + 0.03619566187262535, + 0.0061345649883151054, + 0.014675906859338284, + 0.008245456032454967, + 0.013771031983196735, + 0.004849215038120747, + 0.008416764438152313, + 0.026447290554642677, + 0.0436006523668766, + 0.004295088350772858, + 0.01092818845063448, + 0.0036650055553764105, + 0.007100076414644718, + 0.004715863615274429, + 0.007196785882115364, + 0.006285545881837606, + 0.009454220533370972, + 0.005132944788783789, + 0.007037284318357706, + 0.0037662575487047434, + 0.00444779871031642, + 0.002099005738273263, + 0.003129374235868454, + 0.004130927845835686, + 0.0023090424947440624, + 0.002650019945576787, + 0.004674081224948168, + 0.002168292412534356, + 0.003973931074142456, + 0.0036918686237186193, + 0.004766739439219236, + 0.006407169159501791, + 0.00765214953571558, + 0.005681606475263834, + 0.005352270323783159, + 0.00434931879863143, + 0.008705454878509045, + 0.013458251021802425, + 0.013331812806427479, + 0.019341478124260902, + 0.01470476109534502, + 0.015395702794194221, + 0.011736193671822548, + 0.018527662381529808, + 0.01699659787118435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.10030137002468109 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190072951838374 + ], + "opposite_logprob_tokens": [ + -5.753190040588379 + ], + "logprob_reward": -0.003190072951838374, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190072951838374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21140339970588684, + 0.09437797963619232, + 0.165696382522583, + 0.11775846779346466, + 0.13934233784675598, + 0.13144534826278687, + 0.05961234122514725, + 0.015506499446928501, + 0.03699673339724541, + 0.027860531583428383 + ], + "credit_map": [ + 0.015799298882484436, + 0.019041208550333977, + 0.03413126990199089, + 0.03570473939180374, + 0.023234454914927483, + 0.02321912907063961, + 0.028176119551062584, + 0.03209718316793442, + 0.008345321752130985, + 0.006724715698510408, + 0.020865026861429214, + 0.01780983805656433, + 0.006304658018052578, + 0.011898687109351158, + 0.008705979213118553, + 0.013723758049309254, + 0.009323377162218094, + 0.01801370270550251, + 0.034566447138786316, + 0.05817302316427231, + 0.005849392153322697, + 0.014365731738507748, + 0.012959768064320087, + 0.01244492456316948, + 0.008497965522110462, + 0.013131319545209408, + 0.02753271907567978, + 0.0214993879199028, + 0.006853906437754631, + 0.023092500865459442, + 0.007817854173481464, + 0.009332816116511822, + 0.008777840062975883, + 0.01038102712482214, + 0.05302678421139717, + 0.0315619632601738, + 0.003968814387917519, + 0.01894216053187847, + 0.005221419036388397, + 0.007462335284799337, + 0.005344516132026911, + 0.010532943531870842, + 0.03039199486374855, + 0.05522651970386505, + 0.004112112335860729, + 0.01129977311939001, + 0.0038336857687681913, + 0.010703805834054947, + 0.007327548693865538, + 0.01038351096212864, + 0.008266412653028965, + 0.006922297179698944, + 0.004478374496102333, + 0.01296924613416195, + 0.003348805708810687, + 0.005916145164519548, + 0.001147016417235136, + 0.0013746479526162148, + 0.002520543523132801, + 0.0019884088542312384, + 0.0019700895063579082, + 0.0019111260771751404, + 0.0014357969630509615, + 0.00315886945463717, + 0.0032559523824602365, + 0.004056247416883707, + 0.005757700651884079, + 0.005061710719019175, + 0.0041968910954892635, + 0.004094278905540705, + 0.002949352143332362, + 0.007624599616974592, + 0.0026552234776318073, + 0.002838744781911373, + 0.004207747057080269, + 0.0029780035838484764, + 0.003995803184807301, + 0.0033354200422763824, + 0.0037961285561323166, + 0.0040534608997404575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 4 + ], + "selected_reward": -0.003190072951838374 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures a brilliant golden ginkgo leaf, slowly swirling and falling from the sky. Warm afternoon sunlight penetrates the sparse canopy, softly backlighting the leaf, making its thin, cicada-wing texture appear crystal clear, its signature scalloped edges shimmering with a dazzling halo. The camera pans forward with an incredibly smooth and steady pace, following the leaf's leisurely descent as it gracefully and deliberately spins through the air, as if performing a silent solo. As the camera zooms in, the image gradually shifts from a mid-ground depicting the falling leaf's movement to an extreme close-up, revealing the leaf's forked veins, like an intricate map. Each fine vein shimmers subtly in the light, its edges tinged with a caramel-brown hue. The scene is enveloped in a hazy, shallow depth of field, while the blurred autumn forest colors in the background contrast sharply with the subject, creating a tranquil, beautiful, and poetic atmosphere.", + "seed": 26260, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_9503_seed_26260_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.049369503033813086, + "vlm_logprob": -0.049369503033813086, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.049369503033813086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01226525753736496 + ], + "opposite_logprob_tokens": [ + -4.762265205383301 + ], + "logprob_reward": -0.01226525753736496, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.01226525753736496, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1137692779302597, + 0.057939231395721436, + 0.07607798278331757, + 0.19988574087619781, + 0.054292596876621246, + 0.09924393147230148, + 0.07816127687692642, + 0.15566980838775635, + 0.08538436889648438, + 0.07957582175731659 + ], + "credit_map": [ + 0.010692696087062359, + 0.011407165788114071, + 0.015939926728606224, + 0.022853240370750427, + 0.009683608077466488, + 0.013172823935747147, + 0.011790989898145199, + 0.018228810280561447, + 0.005186178255826235, + 0.0033287364058196545, + 0.008920863270759583, + 0.013731936924159527, + 0.005772935226559639, + 0.004983460996299982, + 0.0070871757343411446, + 0.008927940391004086, + 0.008086204528808594, + 0.0063829002901911736, + 0.013966456055641174, + 0.011255214922130108, + 0.008871274068951607, + 0.005083649419248104, + 0.012593772262334824, + 0.009838506579399109, + 0.013615970499813557, + 0.020918719470500946, + 0.04991978034377098, + 0.05855245888233185, + 0.015028136782348156, + 0.012510518543422222, + 0.016876978799700737, + 0.01246314775198698, + 0.006905076559633017, + 0.00569944828748703, + 0.012034302577376366, + 0.006354744546115398, + 0.005729465279728174, + 0.005602785386145115, + 0.007019592449069023, + 0.004947175271809101, + 0.0075308470986783504, + 0.005175061523914337, + 0.011152872815728188, + 0.016042351722717285, + 0.027254799380898476, + 0.007953188382089138, + 0.011104783974587917, + 0.013030015863478184, + 0.012326210737228394, + 0.01034396793693304, + 0.011840316466987133, + 0.012364858761429787, + 0.009924226440489292, + 0.006763012148439884, + 0.008085602894425392, + 0.006513070780783892, + 0.015203080140054226, + 0.01215757243335247, + 0.024440951645374298, + 0.021313009783625603, + 0.01914261095225811, + 0.008534385822713375, + 0.0383104532957077, + 0.016567731276154518, + 0.011208327487111092, + 0.010534427128732204, + 0.009916402399539948, + 0.007504225708544254, + 0.012556053698062897, + 0.013525898568332195, + 0.012966055423021317, + 0.007172976154834032, + 0.008222803473472595, + 0.009929375723004341, + 0.010143089108169079, + 0.011717844754457474, + 0.010425769723951817, + 0.007581943646073341, + 0.010164744220674038, + 0.011390239000320435 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 0 + ], + "selected_reward": -0.01226525753736496 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156264305114746 + ], + "opposite_logprob_tokens": [ + -1.7015626430511475 + ], + "logprob_reward": -0.20156264305114746, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156264305114746, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06606322526931763, + 0.047004781663417816, + 0.05886128544807434, + 0.11896522343158722, + 0.06814464181661606, + 0.14884111285209656, + 0.2640232443809509, + 0.1593834012746811, + 0.03962010517716408, + 0.02909299172461033 + ], + "credit_map": [ + 0.0073561291210353374, + 0.005149635951966047, + 0.009506586007773876, + 0.010158390738070011, + 0.0074434345588088036, + 0.00595476571470499, + 0.010234522633254528, + 0.010259762406349182, + 0.003706854535266757, + 0.005812698509544134, + 0.007394244894385338, + 0.00664266012609005, + 0.004856226034462452, + 0.006998138967901468, + 0.005221391562372446, + 0.006372569128870964, + 0.005321758799254894, + 0.004479675553739071, + 0.011838207021355629, + 0.008654292672872543, + 0.00753565551713109, + 0.0036429271567612886, + 0.009388003498315811, + 0.008000764064490795, + 0.009326090104877949, + 0.011797741986811161, + 0.02413984388113022, + 0.025684108957648277, + 0.011392714455723763, + 0.010254175402224064, + 0.012335464358329773, + 0.014035090804100037, + 0.007004179526120424, + 0.008262348361313343, + 0.008965641260147095, + 0.012643460184335709, + 0.011945490725338459, + 0.006310123018920422, + 0.008039346896111965, + 0.004974049050360918, + 0.011911721900105476, + 0.009666125290095806, + 0.021338747814297676, + 0.02192816697061062, + 0.040117744356393814, + 0.010365021415054798, + 0.017839346081018448, + 0.01567424088716507, + 0.01818607933819294, + 0.0352577343583107, + 0.05106094107031822, + 0.0485016405582428, + 0.025926750153303146, + 0.027723193168640137, + 0.035718709230422974, + 0.021648211404681206, + 0.023931005969643593, + 0.01899619773030281, + 0.022794287651777267, + 0.02094789408147335, + 0.018322976306080818, + 0.013082197867333889, + 0.024687713012099266, + 0.01662113144993782, + 0.007431602105498314, + 0.004525323398411274, + 0.005664037074893713, + 0.0035342127084732056, + 0.005598363932222128, + 0.003763829357922077, + 0.0033707600086927414, + 0.005731973797082901, + 0.003689471399411559, + 0.003225951688364148, + 0.004440325312316418, + 0.004013438243418932, + 0.0035217483527958393, + 0.003550598630681634, + 0.003476985963061452, + 0.0031744723673909903 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 5 + ], + "selected_reward": -0.20156264305114746 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954977124929428 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954977124929428, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954977124929428, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06540410965681076, + 0.055860769003629684, + 0.14491167664527893, + 0.13270613551139832, + 0.07982785999774933, + 0.12730376422405243, + 0.05992567166686058, + 0.16618622839450836, + 0.04310818761587143, + 0.12476553022861481 + ], + "credit_map": [ + 0.004826133605092764, + 0.003971177153289318, + 0.007749761454761028, + 0.012322692200541496, + 0.007194434758275747, + 0.006847194395959377, + 0.007738935761153698, + 0.014753789640963078, + 0.00576799688860774, + 0.007165494840592146, + 0.00911855511367321, + 0.007736217230558395, + 0.0044130077585577965, + 0.009331167675554752, + 0.006054833997040987, + 0.006273502018302679, + 0.01721508987247944, + 0.014807573519647121, + 0.01796695776283741, + 0.023039013147354126, + 0.018821194767951965, + 0.013405120931565762, + 0.022606397047638893, + 0.01705033704638481, + 0.014325450174510479, + 0.012762880884110928, + 0.030557094141840935, + 0.024990247562527657, + 0.013023674488067627, + 0.008924521505832672, + 0.013568686321377754, + 0.01455360371619463, + 0.006885332986712456, + 0.01360583957284689, + 0.015461456961929798, + 0.010766577906906605, + 0.00852194707840681, + 0.00517449202015996, + 0.008908664807677269, + 0.010503549128770828, + 0.01057138666510582, + 0.009248635731637478, + 0.01775607280433178, + 0.023944314569234848, + 0.030451036989688873, + 0.007716995198279619, + 0.010726484470069408, + 0.016888858750462532, + 0.004191969055682421, + 0.005143336486071348, + 0.011343256570398808, + 0.011121624149382114, + 0.0079047791659832, + 0.007911729626357555, + 0.008127838373184204, + 0.004181142896413803, + 0.02616156078875065, + 0.022628700360655785, + 0.024137914180755615, + 0.029332373291254044, + 0.01609019562602043, + 0.008766950108110905, + 0.023025667294859886, + 0.016042878851294518, + 0.009621314704418182, + 0.0037383721210062504, + 0.007003629580140114, + 0.005012156907469034, + 0.005538879428058863, + 0.002821575617417693, + 0.004845287650823593, + 0.00452697416767478, + 0.014175012707710266, + 0.01846049539744854, + 0.015473609790205956, + 0.020651213824748993, + 0.018688956275582314, + 0.009101174771785736, + 0.016405902802944183, + 0.011809173040091991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 3 + ], + "selected_reward": -0.005954977124929428 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026346659287810326 + ], + "opposite_logprob_tokens": [ + -3.6513466835021973 + ], + "logprob_reward": -0.026346659287810326, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026346659287810326, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06727208197116852, + 0.020369339734315872, + 0.17886982858181, + 0.07050677388906479, + 0.1558619737625122, + 0.05781126767396927, + 0.0716448649764061, + 0.11997054517269135, + 0.14282777905464172, + 0.11486557126045227 + ], + "credit_map": [ + 0.006520040798932314, + 0.0061005763709545135, + 0.008955912664532661, + 0.013670668937265873, + 0.00991194136440754, + 0.005021766759455204, + 0.008379488252103329, + 0.00871168915182352, + 0.0021293479949235916, + 0.0017857215134426951, + 0.0026270248927176, + 0.002997415838763118, + 0.0023331583943217993, + 0.0021491404622793198, + 0.003026740625500679, + 0.0033207887317985296, + 0.020703045651316643, + 0.023404601961374283, + 0.029016345739364624, + 0.02464783377945423, + 0.02080179564654827, + 0.01662026159465313, + 0.02291703410446644, + 0.020758915692567825, + 0.006065232213586569, + 0.008846418932080269, + 0.009847081266343594, + 0.012220650911331177, + 0.008717565797269344, + 0.0068254792131483555, + 0.007786978036165237, + 0.010197367519140244, + 0.02110172249376774, + 0.018467698246240616, + 0.039195992052555084, + 0.019949454814195633, + 0.0172693133354187, + 0.006905566900968552, + 0.012791602872312069, + 0.02018062025308609, + 0.00569014111533761, + 0.0036809472367167473, + 0.00982190016657114, + 0.008886528201401234, + 0.007466125767678022, + 0.006225425750017166, + 0.0077376822009682655, + 0.008302518166601658, + 0.005451907869428396, + 0.009302208200097084, + 0.01100915390998125, + 0.010232573375105858, + 0.012591288425028324, + 0.007141188718378544, + 0.009014155715703964, + 0.006902385037392378, + 0.01687454804778099, + 0.01885310560464859, + 0.019612183794379234, + 0.01938243955373764, + 0.008796712383627892, + 0.013885400258004665, + 0.013069245964288712, + 0.009496904909610748, + 0.03994980454444885, + 0.010810252279043198, + 0.026959259063005447, + 0.0148762296885252, + 0.014353938400745392, + 0.011117372661828995, + 0.012158568948507309, + 0.012602346017956734, + 0.01128586195409298, + 0.015938924625515938, + 0.016941845417022705, + 0.016542324796319008, + 0.01653161831200123, + 0.008183753117918968, + 0.01462214719504118, + 0.01481909304857254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 8 + ], + "selected_reward": -0.026346659287810326 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0007179781678132713 + ], + "opposite_logprob_tokens": [ + -7.250718116760254 + ], + "logprob_reward": -0.0007179781678132713, + "yes_no_margin_reward": 7.25, + "token_credit_reward": -0.0007179781678132713, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20661844313144684, + 0.06836723536252975, + 0.10481137037277222, + 0.08019985258579254, + 0.12685932219028473, + 0.1146320253610611, + 0.06492046266794205, + 0.08315231651067734, + 0.06726518273353577, + 0.08317380398511887 + ], + "credit_map": [ + 0.019387440755963326, + 0.01764192432165146, + 0.029204316437244415, + 0.043727271258831024, + 0.02140938490629196, + 0.0248413048684597, + 0.020933255553245544, + 0.02947354130446911, + 0.007976415567100048, + 0.006036523263901472, + 0.010761675424873829, + 0.010347731411457062, + 0.006729949731379747, + 0.008571946062147617, + 0.00784833263605833, + 0.010094658471643925, + 0.012522812932729721, + 0.015024648979306221, + 0.017154986038804054, + 0.014297043904662132, + 0.01330226007848978, + 0.007802452892065048, + 0.011261556297540665, + 0.013445605523884296, + 0.008965160697698593, + 0.008436795324087143, + 0.012736049480736256, + 0.011565283872187138, + 0.009512013755738735, + 0.0067294598557055, + 0.009582686237990856, + 0.01267240010201931, + 0.014516061171889305, + 0.017448995262384415, + 0.0283012967556715, + 0.017391420900821686, + 0.013088307343423367, + 0.0069109369069337845, + 0.010751389898359776, + 0.018450919538736343, + 0.009365708567202091, + 0.009080921299755573, + 0.015518353320658207, + 0.025498226284980774, + 0.02830701693892479, + 0.006895480211824179, + 0.011553539894521236, + 0.008412783034145832, + 0.005576040130108595, + 0.011465713381767273, + 0.008806075900793076, + 0.008992287330329418, + 0.01002863235771656, + 0.006570011377334595, + 0.008173265494406223, + 0.0053084311075508595, + 0.017445890232920647, + 0.008886349387466908, + 0.011215421371161938, + 0.013743777759373188, + 0.0071825021877884865, + 0.005065722856670618, + 0.011858025565743446, + 0.0077546266838908195, + 0.01575305312871933, + 0.006062358617782593, + 0.010243937373161316, + 0.006394357420504093, + 0.008815797045826912, + 0.006299471948295832, + 0.007853345014154911, + 0.005842866376042366, + 0.00951344519853592, + 0.009955069981515408, + 0.010679295286536217, + 0.013214658945798874, + 0.010590439662337303, + 0.006360842380672693, + 0.010933379642665386, + 0.011926672421395779 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0007179781678132713 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures a brilliant golden ginkgo leaf, slowly swirling and falling from the sky. Warm afternoon sunlight penetrates the sparse canopy, softly backlighting the leaf, making its thin, cicada-wing texture appear crystal clear, its signature scalloped edges shimmering with a dazzling halo. The camera pans forward with an incredibly smooth and steady pace, following the leaf's leisurely descent as it gracefully and deliberately spins through the air, as if performing a silent solo. As the camera zooms in, the image gradually shifts from a mid-ground depicting the falling leaf's movement to an extreme close-up, revealing the leaf's forked veins, like an intricate map. Each fine vein shimmers subtly in the light, its edges tinged with a caramel-brown hue. The scene is enveloped in a hazy, shallow depth of field, while the blurred autumn forest colors in the background contrast sharply with the subject, creating a tranquil, beautiful, and poetic atmosphere.", + "seed": 26261, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_9503_seed_26261_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.15341627663001417, + "vlm_logprob": -0.15341627663001417, + "vlm_yes_no_margin": 3.125, + "vlm_token_credit": -0.15341627663001417, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011147127486765385 + ], + "opposite_logprob_tokens": [ + -4.7611470222473145 + ], + "logprob_reward": -0.011147127486765385, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011147127486765385, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2572607100009918, + 0.08028733730316162, + 0.061570875346660614, + 0.04970927536487579, + 0.2858761250972748, + 0.10180581361055374, + 0.04220985993742943, + 0.030572643503546715, + 0.04965640977025032, + 0.04105101898312569 + ], + "credit_map": [ + 0.012901494279503822, + 0.0070040407590568066, + 0.08267408609390259, + 0.008303450420498848, + 0.004299728199839592, + 0.01737440750002861, + 0.01834293082356453, + 0.10636056214570999, + 0.008650214411318302, + 0.00956282950937748, + 0.01628985069692135, + 0.00687419855967164, + 0.0054169935174286366, + 0.008690250106155872, + 0.007487880997359753, + 0.01731511764228344, + 0.004926169291138649, + 0.00537386629730463, + 0.008625386282801628, + 0.005352837964892387, + 0.005416168365627527, + 0.011425076983869076, + 0.010293947532773018, + 0.010157426819205284, + 0.005459392908960581, + 0.004007343202829361, + 0.0055311257019639015, + 0.005076349247246981, + 0.004777280148118734, + 0.004900520667433739, + 0.014679351821541786, + 0.005277911201119423, + 0.014575720764696598, + 0.005690774414688349, + 0.1126566231250763, + 0.009633678011596203, + 0.013402908109128475, + 0.11609608680009842, + 0.006984146311879158, + 0.006836161948740482, + 0.01831742748618126, + 0.007946551777422428, + 0.019604187458753586, + 0.00714427325874567, + 0.007922589778900146, + 0.012662572786211967, + 0.010725862346589565, + 0.017482344061136246, + 0.008454889990389347, + 0.003742047818377614, + 0.008092635311186314, + 0.003350566141307354, + 0.004164909012615681, + 0.004942890722304583, + 0.0034915467258542776, + 0.005970376078039408, + 0.0072192116640508175, + 0.004237036686390638, + 0.0041827852837741375, + 0.002221304690465331, + 0.0034205566626042128, + 0.003149439347907901, + 0.0030805403366684914, + 0.0030617681331932545, + 0.007538589183241129, + 0.00434187613427639, + 0.005274370778352022, + 0.006375247146934271, + 0.007389239501208067, + 0.0059621939435601234, + 0.007097376510500908, + 0.005677519831806421, + 0.006185798905789852, + 0.004571215715259314, + 0.004875113256275654, + 0.006050839088857174, + 0.004721222445368767, + 0.0045315357856452465, + 0.006156712304800749, + 0.003958579618483782 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.011147127486765385 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486643135547638 + ], + "opposite_logprob_tokens": [ + -1.2236642837524414 + ], + "logprob_reward": -0.3486643135547638, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486643135547638, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2818445861339569, + 0.08577068895101547, + 0.057688139379024506, + 0.030033618211746216, + 0.1818261295557022, + 0.04694894701242447, + 0.03645806759595871, + 0.0631243884563446, + 0.07636420428752899, + 0.13994121551513672 + ], + "credit_map": [ + 0.011455773375928402, + 0.006061102729290724, + 0.0908873975276947, + 0.008075623773038387, + 0.004436361603438854, + 0.019796915352344513, + 0.02261991798877716, + 0.11851148307323456, + 0.007619101088494062, + 0.006857879925519228, + 0.017779449000954628, + 0.007033867761492729, + 0.005468061659485102, + 0.008702941238880157, + 0.010864412412047386, + 0.021444974467158318, + 0.00555222574621439, + 0.005269416142255068, + 0.0075676157139241695, + 0.008396320044994354, + 0.0040853749960660934, + 0.009245786815881729, + 0.009653816930949688, + 0.007917585782706738, + 0.003241355763748288, + 0.0031024236232042313, + 0.005436571780592203, + 0.003188659204170108, + 0.002850292017683387, + 0.002495300956070423, + 0.004732303321361542, + 0.004986714106053114, + 0.00877347681671381, + 0.007114973850548267, + 0.06554564833641052, + 0.005815260577946901, + 0.0113889891654253, + 0.06692123413085938, + 0.007574170362204313, + 0.008692391216754913, + 0.009909485466778278, + 0.002369456458836794, + 0.015451304614543915, + 0.0020829143468290567, + 0.002669593086466193, + 0.009308276697993279, + 0.0026114715728908777, + 0.0025464429054409266, + 0.009333595633506775, + 0.003034517401829362, + 0.006638383958488703, + 0.0034045004285871983, + 0.002670824993401766, + 0.0030286412220448256, + 0.002414172748103738, + 0.005933430977165699, + 0.014266742393374443, + 0.0077063050121068954, + 0.009679402224719524, + 0.006314270198345184, + 0.007332269102334976, + 0.005029463674873114, + 0.00551771093159914, + 0.00727822445333004, + 0.01170967798680067, + 0.009501096792519093, + 0.00952237006276846, + 0.00909404642879963, + 0.010060162283480167, + 0.009777559898793697, + 0.008228152059018612, + 0.008471136912703514, + 0.018014274537563324, + 0.014948255382478237, + 0.014948192052543163, + 0.0222823116928339, + 0.015691114589571953, + 0.014345447532832623, + 0.025868626311421394, + 0.013842985965311527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.3486643135547638 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012551010586321354 + ], + "opposite_logprob_tokens": [ + -4.3875508308410645 + ], + "logprob_reward": -0.012551010586321354, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012551010586321354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08951101452112198, + 0.07544071972370148, + 0.023219622671604156, + 0.024477455765008926, + 0.5211525559425354, + 0.09931791573762894, + 0.04402702674269676, + 0.01482783816754818, + 0.059753648936748505, + 0.04827230051159859 + ], + "credit_map": [ + 0.009778826497495174, + 0.004889166913926601, + 0.026048416271805763, + 0.004834133666008711, + 0.002649306785315275, + 0.008872401900589466, + 0.008585372008383274, + 0.023853396996855736, + 0.005201187450438738, + 0.006880265194922686, + 0.018737908452749252, + 0.00641394779086113, + 0.00447892677038908, + 0.006713134236633778, + 0.007558467332273722, + 0.019456882029771805, + 0.0021034174133092165, + 0.0020746991503983736, + 0.004217746201902628, + 0.0018802169943228364, + 0.0016721379943192005, + 0.004081593360751867, + 0.0035658820997923613, + 0.003623928874731064, + 0.0026483172550797462, + 0.0018961112946271896, + 0.0037892363034188747, + 0.0023717426229268312, + 0.001608360093086958, + 0.0022111020516604185, + 0.003858815645799041, + 0.006093769799917936, + 0.028639189898967743, + 0.009212610311806202, + 0.20543476939201355, + 0.016024157404899597, + 0.022911660373210907, + 0.20747852325439453, + 0.015846116468310356, + 0.015605482272803783, + 0.01279979757964611, + 0.008677354082465172, + 0.01616104692220688, + 0.010019894689321518, + 0.007963031530380249, + 0.013946683146059513, + 0.007817506790161133, + 0.02193259634077549, + 0.007506996393203735, + 0.004318143241107464, + 0.005464030429720879, + 0.005542273633182049, + 0.005582056939601898, + 0.003942782524973154, + 0.0051910001784563065, + 0.00647974107414484, + 0.002510549034923315, + 0.0014114858349785209, + 0.0015870279166847467, + 0.0014777283649891615, + 0.0014550078194588423, + 0.002461701398715377, + 0.002538007218390703, + 0.0013863317435607314, + 0.009273610077798367, + 0.005457540042698383, + 0.005852223839610815, + 0.009478415362536907, + 0.007451350800693035, + 0.006956306751817465, + 0.008961341343820095, + 0.006322859786450863, + 0.006551289930939674, + 0.005344375967979431, + 0.005301072262227535, + 0.006825241260230541, + 0.00522844772785902, + 0.005369765218347311, + 0.008782743476331234, + 0.004869365599006414 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.012551010586321354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38708797097206116 + ], + "opposite_logprob_tokens": [ + -1.1370879411697388 + ], + "logprob_reward": -0.38708797097206116, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38708797097206116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08622583001852036, + 0.029064727947115898, + 0.05491403490304947, + 0.04692062735557556, + 0.3994697630405426, + 0.04168718680739403, + 0.049222953617572784, + 0.040612492710351944, + 0.18868368864059448, + 0.06319877505302429 + ], + "credit_map": [ + 0.004495508968830109, + 0.004684416577219963, + 0.022677678614854813, + 0.004232895560562611, + 0.0027655106969177723, + 0.006902101915329695, + 0.007707931566983461, + 0.03275977447628975, + 0.003608011407777667, + 0.00255506276153028, + 0.005740321706980467, + 0.0028224042616784573, + 0.00212722085416317, + 0.003226337255910039, + 0.0034076226875185966, + 0.005577743984758854, + 0.0032836529426276684, + 0.005847297143191099, + 0.0060189287178218365, + 0.006972631439566612, + 0.005892914719879627, + 0.00914332177489996, + 0.01019222754985094, + 0.007563059683889151, + 0.0049937027506530285, + 0.0038076911587268114, + 0.006602220702916384, + 0.006187573075294495, + 0.00435596751049161, + 0.004037626087665558, + 0.009454868733882904, + 0.007480977103114128, + 0.02488642744719982, + 0.013110524974763393, + 0.1388506293296814, + 0.01698242872953415, + 0.021207015961408615, + 0.14130799472332, + 0.02601916901767254, + 0.01710556074976921, + 0.0063820029608905315, + 0.003705172101035714, + 0.006718343589454889, + 0.004060093313455582, + 0.0032689145300537348, + 0.004705015569925308, + 0.003868469735607505, + 0.00897916965186596, + 0.0063087549060583115, + 0.004190011415630579, + 0.005945667624473572, + 0.006869863253086805, + 0.006329825613647699, + 0.005356585141271353, + 0.007178339175879955, + 0.007043899968266487, + 0.0076751443557441235, + 0.005570296663790941, + 0.0053643519058823586, + 0.0032413480803370476, + 0.005113639868795872, + 0.004892344120889902, + 0.002923767315223813, + 0.005831596907228231, + 0.03645269572734833, + 0.017021872103214264, + 0.027294056490063667, + 0.02806117758154869, + 0.021151114255189896, + 0.016498301178216934, + 0.027173103764653206, + 0.015031364746391773, + 0.010137062519788742, + 0.006996666081249714, + 0.0073725082911551, + 0.008869068697094917, + 0.00673942593857646, + 0.005701485089957714, + 0.010016529820859432, + 0.0073660206981003284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 0 + ], + "selected_reward": -0.38708797097206116 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630960550159216 + ], + "opposite_logprob_tokens": [ + -4.882630825042725 + ], + "logprob_reward": -0.007630960550159216, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630960550159216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1595442146062851, + 0.03963537886738777, + 0.03027157485485077, + 0.03236619755625725, + 0.5435037612915039, + 0.043966904282569885, + 0.02537113055586815, + 0.03294365108013153, + 0.05104829743504524, + 0.04134887084364891 + ], + "credit_map": [ + 0.010938200168311596, + 0.0065228804014623165, + 0.035189006477594376, + 0.007065760903060436, + 0.004058585502207279, + 0.015745164826512337, + 0.016322964802384377, + 0.06370165199041367, + 0.0036479062400758266, + 0.004807441495358944, + 0.006697909440845251, + 0.0035606734454631805, + 0.0028339147102087736, + 0.004260568879544735, + 0.004185487981885672, + 0.009641478769481182, + 0.003710199147462845, + 0.002495210152119398, + 0.005749695003032684, + 0.0027848484460264444, + 0.0021348432637751102, + 0.004113796632736921, + 0.0037587378174066544, + 0.005524245090782642, + 0.004818562883883715, + 0.002694678958505392, + 0.005643917713314295, + 0.0036928232293576, + 0.003313436172902584, + 0.004639062099158764, + 0.0031075342558324337, + 0.004456185735762119, + 0.031783681362867355, + 0.009428253397345543, + 0.2162984311580658, + 0.016899866983294487, + 0.023225972428917885, + 0.21648351848125458, + 0.015471937134861946, + 0.01391211524605751, + 0.006612607277929783, + 0.004184501711279154, + 0.010527023114264011, + 0.0030672610737383366, + 0.003854339476674795, + 0.0061776298098266125, + 0.003919017035514116, + 0.005624526180326939, + 0.006324368994683027, + 0.0028107655234634876, + 0.00382040161639452, + 0.0020681980531662703, + 0.0019469066755846143, + 0.001783182262443006, + 0.0021567409858107567, + 0.004460566211491823, + 0.008320690132677555, + 0.004755752626806498, + 0.004945128690451384, + 0.0030309956055134535, + 0.004918758291751146, + 0.002232577186077833, + 0.0024315991904586554, + 0.00230815215036273, + 0.008723127655684948, + 0.005169176030904055, + 0.005608638282865286, + 0.008673559874296188, + 0.005692953243851662, + 0.0058241491205990314, + 0.006028312724083662, + 0.0053283823654055595, + 0.00683972891420126, + 0.004133040551096201, + 0.004967030603438616, + 0.00684732012450695, + 0.004172054119408131, + 0.0038159689866006374, + 0.005600017495453358, + 0.0049737077206373215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.007630960550159216 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot, shot from a slightly upward angle, captures a brilliant golden ginkgo leaf, slowly swirling and falling from the sky. Warm afternoon sunlight penetrates the sparse canopy, softly backlighting the leaf, making its thin, cicada-wing texture appear crystal clear, its signature scalloped edges shimmering with a dazzling halo. The camera pans forward with an incredibly smooth and steady pace, following the leaf's leisurely descent as it gracefully and deliberately spins through the air, as if performing a silent solo. As the camera zooms in, the image gradually shifts from a mid-ground depicting the falling leaf's movement to an extreme close-up, revealing the leaf's forked veins, like an intricate map. Each fine vein shimmers subtly in the light, its edges tinged with a caramel-brown hue. The scene is enveloped in a hazy, shallow depth of field, while the blurred autumn forest colors in the background contrast sharply with the subject, creating a tranquil, beautiful, and poetic atmosphere.", + "seed": 26262, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_5/index_9503_seed_26262_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.036882797232829036, + "vlm_logprob": -0.036882797232829036, + "vlm_yes_no_margin": 4.35, + "vlm_token_credit": -0.036882797232829036, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017024332657456398 + ], + "opposite_logprob_tokens": [ + -4.517024517059326 + ], + "logprob_reward": -0.017024332657456398, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.017024332657456398, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.100918710231781, + 0.05142763629555702, + 0.11432068794965744, + 0.10030221194028854, + 0.06491399556398392, + 0.06141306459903717, + 0.0746721550822258, + 0.09050925076007843, + 0.042769674211740494, + 0.29875263571739197 + ], + "credit_map": [ + 0.013310849666595459, + 0.0123965535312891, + 0.012848875485360622, + 0.010639752261340618, + 0.007626568898558617, + 0.011802181601524353, + 0.014806256629526615, + 0.017487674951553345, + 0.004048444796353579, + 0.007201141677796841, + 0.007474043872207403, + 0.007831910625100136, + 0.005220180843025446, + 0.00514353858307004, + 0.007215389981865883, + 0.007292985916137695, + 0.012188825756311417, + 0.012750530615448952, + 0.017390571534633636, + 0.01462184265255928, + 0.013027223758399487, + 0.013040975667536259, + 0.011043349280953407, + 0.020257364958524704, + 0.014098124578595161, + 0.010594990104436874, + 0.016460591927170753, + 0.012367496266961098, + 0.01099255122244358, + 0.012955996207892895, + 0.008400555700063705, + 0.014431907795369625, + 0.0043700262904167175, + 0.00647232448682189, + 0.008992784656584263, + 0.010606158524751663, + 0.007931651547551155, + 0.007453696336597204, + 0.006668971385806799, + 0.012418377213180065, + 0.007412847131490707, + 0.007867182604968548, + 0.008086098358035088, + 0.012449843809008598, + 0.00630952650681138, + 0.006941965315490961, + 0.00587426358833909, + 0.0064713358879089355, + 0.008443384431302547, + 0.009003191255033016, + 0.015326954424381256, + 0.007226210553199053, + 0.006746269296854734, + 0.010233104228973389, + 0.007100742310285568, + 0.01059229951351881, + 0.008555212058126926, + 0.010156134143471718, + 0.01531857531517744, + 0.010092148557305336, + 0.010957470163702965, + 0.00981027353554964, + 0.01338678877800703, + 0.012232649140059948, + 0.0032577516976743937, + 0.00448604766279459, + 0.005328755360096693, + 0.008666595444083214, + 0.005318580660969019, + 0.00660680653527379, + 0.0047814492136240005, + 0.00432368740439415, + 0.015091555193066597, + 0.017639514058828354, + 0.09691210836172104, + 0.09134604781866074, + 0.011462938040494919, + 0.02269226498901844, + 0.01687222719192505, + 0.026735978201031685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.017024332657456398 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048661138862371445 + ], + "opposite_logprob_tokens": [ + -3.048661231994629 + ], + "logprob_reward": -0.048661138862371445, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048661138862371445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12476035952568054, + 0.15533915162086487, + 0.21609343588352203, + 0.15664127469062805, + 0.06718546897172928, + 0.07769042998552322, + 0.03409024700522423, + 0.024562189355492592, + 0.02235422097146511, + 0.1212831661105156 + ], + "credit_map": [ + 0.008381436578929424, + 0.013823961839079857, + 0.024471305310726166, + 0.02408725768327713, + 0.010659066960215569, + 0.012364563532173634, + 0.016501784324645996, + 0.014470982365310192, + 0.00936862826347351, + 0.02489440329372883, + 0.0373762845993042, + 0.03182516247034073, + 0.010513982735574245, + 0.012124603614211082, + 0.01378236711025238, + 0.015453727915883064, + 0.02104482613503933, + 0.022961771115660667, + 0.02772655338048935, + 0.023950079455971718, + 0.023780182003974915, + 0.022263772785663605, + 0.0206363033503294, + 0.053729936480522156, + 0.016135133802890778, + 0.014947700314223766, + 0.020417455583810806, + 0.018476536497473717, + 0.02034212462604046, + 0.02644410915672779, + 0.017182623967528343, + 0.022695574909448624, + 0.0047057741321623325, + 0.006598677486181259, + 0.008252508006989956, + 0.00920333806425333, + 0.00807635672390461, + 0.010138912126421928, + 0.009739619679749012, + 0.010470282286405563, + 0.007068577688187361, + 0.010899761691689491, + 0.010061748325824738, + 0.016909345984458923, + 0.007985126227140427, + 0.010392279364168644, + 0.006796257104724646, + 0.007577337324619293, + 0.0029676237609237432, + 0.003541527781635523, + 0.0069237155839800835, + 0.0053262715227901936, + 0.0030869776383042336, + 0.0033945778850466013, + 0.0034346296451985836, + 0.005414921324700117, + 0.0022586786653846502, + 0.0021071757655590773, + 0.0036581496242433786, + 0.003514669369906187, + 0.002467105397954583, + 0.003094510640949011, + 0.003873415756970644, + 0.0035884834360331297, + 0.0019579667132347822, + 0.0020297111477702856, + 0.0026199878193438053, + 0.003587063169106841, + 0.0030782774556428194, + 0.0038437896873801947, + 0.0025207302533090115, + 0.0027166935615241528, + 0.006577715743333101, + 0.00812377966940403, + 0.03707807511091232, + 0.03453638777136803, + 0.007169157266616821, + 0.00966651737689972, + 0.006650249008089304, + 0.011481281369924545 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 1 + ], + "selected_reward": -0.048661138862371445 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00464016804471612 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00464016804471612, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00464016804471612, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08248785883188248, + 0.12623348832130432, + 0.04775185510516167, + 0.1325502246618271, + 0.19569182395935059, + 0.11068949103355408, + 0.04683110490441322, + 0.026586703956127167, + 0.08092235028743744, + 0.15025512874126434 + ], + "credit_map": [ + 0.00828795600682497, + 0.01391095481812954, + 0.012635668739676476, + 0.011556711047887802, + 0.005175731610506773, + 0.012106990441679955, + 0.010204343125224113, + 0.008609505370259285, + 0.008574469946324825, + 0.020844686776399612, + 0.024442298337817192, + 0.02549014985561371, + 0.010090910829603672, + 0.010931242257356644, + 0.010268469341099262, + 0.015591259114444256, + 0.00596470246091485, + 0.0055060177110135555, + 0.006057083141058683, + 0.005690469406545162, + 0.0060678753070533276, + 0.0056807552464306355, + 0.005706507712602615, + 0.0070784431882202625, + 0.010705284774303436, + 0.026913588866591454, + 0.019051605835556984, + 0.02283194661140442, + 0.009767136536538601, + 0.012325620278716087, + 0.009912458248436451, + 0.021042587235569954, + 0.017241423949599266, + 0.021052684634923935, + 0.02684026025235653, + 0.05161525309085846, + 0.01500407513231039, + 0.015293891541659832, + 0.018195895478129387, + 0.03044833056628704, + 0.009746957570314407, + 0.01087383646517992, + 0.018256694078445435, + 0.02759944088757038, + 0.007676827721297741, + 0.01023313868790865, + 0.007535549346357584, + 0.018767043948173523, + 0.006203098688274622, + 0.003896894631907344, + 0.00965891219675541, + 0.003713356563821435, + 0.004630699288100004, + 0.007519514765590429, + 0.004561255220323801, + 0.006647374946624041, + 0.003375552361831069, + 0.002716789022088051, + 0.003848566673696041, + 0.00310833309777081, + 0.0028985042590647936, + 0.0033781486563384533, + 0.003369679907336831, + 0.0038911288138478994, + 0.0050866445526480675, + 0.003538968041539192, + 0.025966523215174675, + 0.007547852583229542, + 0.005600577685981989, + 0.009534887038171291, + 0.004921698477119207, + 0.018725194036960602, + 0.00904291681945324, + 0.015176285058259964, + 0.02102612517774105, + 0.037681568413972855, + 0.017848733812570572, + 0.022278480231761932, + 0.009874377399682999, + 0.017326638102531433 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 3 + ], + "selected_reward": -0.00464016804471612 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11290934681892395 + ], + "opposite_logprob_tokens": [ + -2.2379093170166016 + ], + "logprob_reward": -0.11290934681892395, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11290934681892395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08116504549980164, + 0.1410408318042755, + 0.0589587539434433, + 0.04276953637599945, + 0.059056565165519714, + 0.1148759201169014, + 0.04263795539736748, + 0.023409904912114143, + 0.09034372866153717, + 0.34574174880981445 + ], + "credit_map": [ + 0.005758789833635092, + 0.011228011921048164, + 0.01654123142361641, + 0.015306616201996803, + 0.004952226299792528, + 0.010497329756617546, + 0.008403828367590904, + 0.00847702007740736, + 0.00786006823182106, + 0.024459708482027054, + 0.035014864057302475, + 0.030743664130568504, + 0.005392334423959255, + 0.011103048920631409, + 0.012256592512130737, + 0.014210553839802742, + 0.006132739130407572, + 0.005100161302834749, + 0.011659166775643826, + 0.007186232600361109, + 0.006827368400990963, + 0.006162202451378107, + 0.00795059371739626, + 0.00794028677046299, + 0.004333976656198502, + 0.005809484515339136, + 0.006974343676120043, + 0.006933634169399738, + 0.0035920175723731518, + 0.004388757515698671, + 0.004973561968654394, + 0.005763756576925516, + 0.009341389872133732, + 0.0045765964314341545, + 0.008341769687831402, + 0.00943781342357397, + 0.0060863434337079525, + 0.006735509727150202, + 0.006284278817474842, + 0.008252863772213459, + 0.011190778575837612, + 0.014362599700689316, + 0.01557112019509077, + 0.02416408807039261, + 0.012190903536975384, + 0.015944207087159157, + 0.009919863194227219, + 0.01153236161917448, + 0.005892566405236721, + 0.004636065103113651, + 0.006960906088352203, + 0.005680225323885679, + 0.004153831861913204, + 0.004240371752530336, + 0.003669544355943799, + 0.00740444241091609, + 0.0026823384687304497, + 0.0016964101232588291, + 0.0042445543222129345, + 0.0027105906046926975, + 0.001979308668524027, + 0.0031829660292714834, + 0.0021159874740988016, + 0.004797750152647495, + 0.00635330006480217, + 0.005189157091081142, + 0.026567373424768448, + 0.013307114131748676, + 0.006075019016861916, + 0.009458462707698345, + 0.008251609280705452, + 0.015141703188419342, + 0.017451059073209763, + 0.021531052887439728, + 0.11309126764535904, + 0.10101515054702759, + 0.014313401654362679, + 0.02571897767484188, + 0.020003225654363632, + 0.032617613673210144 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.11290934681892395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a single golden ginkgo leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the leaf shown slowly swirling and falling from the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene feature warm afternoon sunlight backlighting the leaf?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera pan forward while zooming in to an extreme close-up of the leaf's veins?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background a blurred autumn forest with a shallow depth of field?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178999780677259 + ], + "opposite_logprob_tokens": [ + -6.751179218292236 + ], + "logprob_reward": -0.001178999780677259, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178999780677259, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14591167867183685, + 0.09530110657215118, + 0.11018642783164978, + 0.09545493870973587, + 0.05655509606003761, + 0.047312647104263306, + 0.05452543497085571, + 0.032799117267131805, + 0.10569585859775543, + 0.25625765323638916 + ], + "credit_map": [ + 0.010732894763350487, + 0.011469339951872826, + 0.021987315267324448, + 0.018114257603883743, + 0.009129653684794903, + 0.03577488660812378, + 0.02186434157192707, + 0.01683899015188217, + 0.0087644187733531, + 0.015410364605486393, + 0.012653468176722527, + 0.013938418589532375, + 0.008086764253675938, + 0.01105421967804432, + 0.009451285935938358, + 0.0159421656280756, + 0.012908142991364002, + 0.009171681478619576, + 0.020976385101675987, + 0.015616822056472301, + 0.010518536902964115, + 0.009900606237351894, + 0.01662154495716095, + 0.014472703449428082, + 0.008963149972259998, + 0.01624997705221176, + 0.013280011713504791, + 0.013673234730958939, + 0.009755278006196022, + 0.010358871892094612, + 0.010372664779424667, + 0.01280174870043993, + 0.008952321484684944, + 0.004223821219056845, + 0.008475132286548615, + 0.008458669297397137, + 0.004497558809816837, + 0.006078630685806274, + 0.005303619429469109, + 0.01056534331291914, + 0.0059661963023245335, + 0.004549240693449974, + 0.009132838808000088, + 0.0061292704194784164, + 0.005644161254167557, + 0.005533008370548487, + 0.00484866090118885, + 0.005509266629815102, + 0.006727681960910559, + 0.005295829847455025, + 0.01064289640635252, + 0.007073423359543085, + 0.004830034915357828, + 0.006913516204804182, + 0.004516418091952801, + 0.00852563139051199, + 0.004948236979544163, + 0.002976824529469013, + 0.005057775415480137, + 0.003331963438540697, + 0.003558114403858781, + 0.004102983046323061, + 0.0032532019540667534, + 0.005570018198341131, + 0.006428740918636322, + 0.005218953359872103, + 0.03246346861124039, + 0.00951777957379818, + 0.007534082513302565, + 0.013460350222885609, + 0.00665191188454628, + 0.02442058175802231, + 0.012282555922865868, + 0.01299796812236309, + 0.08703790605068207, + 0.06244206801056862, + 0.01226987037807703, + 0.01824088953435421, + 0.014420604333281517, + 0.036565788090229034 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 2 + ], + "selected_reward": -0.001178999780677259 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_tensor([6339, 6339])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_tensor([6339, 6339])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..290394ec2d42637d04751a7e25c71e4bfb037651 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_tensor([6339, 6339])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a high-contrast, grainy black-and-white image, the camera begins in a medium shot, focusing on the weathered and cracked carved eaves of a dilapidated Chinese temple. From the deep shadows beneath the eaves, a crow with jet-black, shimmering feathers suddenly flaps its wings, stirring up a cloud of dust and tiny sawdust. It ascends in a powerful, sharp arc toward a gray sky thick with heavy clouds. As the crow leaves its nest, the camera begins an extremely slow and steady pull-back, gradually taking in the entire temple: the abandoned structure still stands, but the roof tiles are broken and covered with moss and weeds, the wooden beams are covered with cobwebs, and the walls are mottled and peeling, revealing the relentless erosion of time. As the camera continues to retreat, the desolate scene around the temple is fully revealed. Several dead trees stretch twistedly towards the sky, the ground is overgrown with weeds, and in the distance are rolling hills with blurred outlines. The entire scene presents rich gray tones under the soft light scattered by the cloudy sky, creating a bleak atmosphere of deep loneliness and time freezing.", + "seed": 7675, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7675_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06450704978778958, + "vlm_logprob": -0.06450704978778958, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.06450704978778958, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2553572952747345 + ], + "opposite_logprob_tokens": [ + -1.505357265472412 + ], + "logprob_reward": -0.2553572952747345, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2553572952747345, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.320011168718338, + 0.05339190363883972, + 0.05200658738613129, + 0.020412443205714226, + 0.28455638885498047, + 0.048842959105968475, + 0.04534178227186203, + 0.01503930613398552, + 0.02654261700809002, + 0.133854940533638 + ], + "credit_map": [ + 0.04381318762898445, + 0.006479834672063589, + 0.05596449226140976, + 0.011150525882840157, + 0.0198318250477314, + 0.07138264924287796, + 0.014313123188912868, + 0.09707550704479218, + 0.007028681691735983, + 0.005041215568780899, + 0.006414154078811407, + 0.005879551637917757, + 0.005155543331056833, + 0.0059916484169662, + 0.012295651249587536, + 0.005585458129644394, + 0.004925135988742113, + 0.010757040232419968, + 0.009483217261731625, + 0.007249860092997551, + 0.004305783659219742, + 0.0046232002787292, + 0.005508220288902521, + 0.005154121667146683, + 0.002299125073477626, + 0.0030933674424886703, + 0.002722513861954212, + 0.0036946828477084637, + 0.0021324208937585354, + 0.0021425646264106035, + 0.0015719834482297301, + 0.0027557832654565573, + 0.04037904366850853, + 0.12105124443769455, + 0.0912458747625351, + 0.009210689924657345, + 0.008333808742463589, + 0.0036950758658349514, + 0.004433044698089361, + 0.006207560189068317, + 0.007303777616471052, + 0.006380049977451563, + 0.00869827251881361, + 0.007506900932639837, + 0.004975369665771723, + 0.0031049279496073723, + 0.0044184415601193905, + 0.0064552174881100655, + 0.008309162221848965, + 0.009406999684870243, + 0.006229056045413017, + 0.00825398787856102, + 0.003618300659582019, + 0.0018215420423075557, + 0.0034250819589942694, + 0.004277648404240608, + 0.002084871055558324, + 0.0025383324827998877, + 0.0027834861539304256, + 0.001541478792205453, + 0.0015115999849513173, + 0.0019975698087364435, + 0.0012239861534908414, + 0.0013579803053289652, + 0.0034636231139302254, + 0.004127900116145611, + 0.0033808820880949497, + 0.002992156893014908, + 0.003695717081427574, + 0.0020493711344897747, + 0.003508754773065448, + 0.0033242071513086557, + 0.028552953153848648, + 0.007276341784745455, + 0.04609183594584465, + 0.011788028292357922, + 0.005210530944168568, + 0.012165886349976063, + 0.01287798024713993, + 0.009891380555927753 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.2553572952747345 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01255913358181715 + ], + "opposite_logprob_tokens": [ + -4.387558937072754 + ], + "logprob_reward": -0.01255913358181715, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.01255913358181715, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25708475708961487, + 0.11258424818515778, + 0.10204914957284927, + 0.03501135855913162, + 0.10994625091552734, + 0.01464598998427391, + 0.04733354598283768, + 0.025519557297229767, + 0.01670962944626808, + 0.2791155278682709 + ], + "credit_map": [ + 0.03413362428545952, + 0.01185943465679884, + 0.03973335400223732, + 0.0164728332310915, + 0.020888933911919594, + 0.04903752729296684, + 0.01715138368308544, + 0.06780768930912018, + 0.014560019597411156, + 0.015389538370072842, + 0.014488515444099903, + 0.017074791714549065, + 0.013802282512187958, + 0.012112618423998356, + 0.014191274531185627, + 0.01096520945429802, + 0.011759782209992409, + 0.024088367819786072, + 0.030451003462076187, + 0.008484088815748692, + 0.0065162694081664085, + 0.006826093420386314, + 0.007557351142168045, + 0.006366200745105743, + 0.0032739832531660795, + 0.00424718065187335, + 0.004054800141602755, + 0.006003618706017733, + 0.0037362128496170044, + 0.004228908568620682, + 0.0036478901747614145, + 0.0058187623508274555, + 0.019890733063220978, + 0.044049959629774094, + 0.02671101875603199, + 0.005232042632997036, + 0.003742717672139406, + 0.003034807974472642, + 0.0027501515578478575, + 0.004534826148301363, + 0.0020640497095882893, + 0.001755966222845018, + 0.0027701908256858587, + 0.0019855902064591646, + 0.0011829249560832977, + 0.0009648934355936944, + 0.001495517324656248, + 0.0024268566630780697, + 0.009071978740394115, + 0.008933653123676777, + 0.007473030127584934, + 0.007681689690798521, + 0.003896780079230666, + 0.002127287443727255, + 0.0035530596505850554, + 0.0045960647985339165, + 0.0029251507949084044, + 0.0029478585347533226, + 0.004980613477528095, + 0.0030285560060292482, + 0.002633425872772932, + 0.002627612790092826, + 0.0023971654009073973, + 0.003979173488914967, + 0.0025414368137717247, + 0.0037607455160468817, + 0.002325638197362423, + 0.0018753472249954939, + 0.001764345564879477, + 0.000973530753981322, + 0.0020153229124844074, + 0.0014532629866153002, + 0.08327840268611908, + 0.020606039091944695, + 0.0180649571120739, + 0.01197720505297184, + 0.014204006642103195, + 0.10593043267726898, + 0.010946611873805523, + 0.014107873663306236 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.01255913358181715 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810866177082062 + ], + "opposite_logprob_tokens": [ + -3.2881085872650146 + ], + "logprob_reward": -0.03810866177082062, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810866177082062, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11444991827011108, + 0.14126616716384888, + 0.06252792477607727, + 0.19869643449783325, + 0.1199529841542244, + 0.022171879187226295, + 0.020120127126574516, + 0.040238287299871445, + 0.03352484852075577, + 0.24705137312412262 + ], + "credit_map": [ + 0.014647546224296093, + 0.003941764123737812, + 0.019151927903294563, + 0.007758613210171461, + 0.01023450493812561, + 0.0237258430570364, + 0.008110441267490387, + 0.026879288256168365, + 0.01765821874141693, + 0.019825827330350876, + 0.017187030985951424, + 0.017822226509451866, + 0.01619979366660118, + 0.017171381041407585, + 0.02211616188287735, + 0.013285532593727112, + 0.009951794520020485, + 0.010576198808848858, + 0.010086591355502605, + 0.007312840782105923, + 0.004936453886330128, + 0.006814782507717609, + 0.006141053978353739, + 0.006708212196826935, + 0.017277590930461884, + 0.030341260135173798, + 0.022948484867811203, + 0.0371074452996254, + 0.021941212937235832, + 0.021095503121614456, + 0.01748623140156269, + 0.030498722568154335, + 0.02063540555536747, + 0.04098795726895332, + 0.02715698629617691, + 0.006400133948773146, + 0.00523340655490756, + 0.005387522745877504, + 0.0057787057012319565, + 0.008372877724468708, + 0.003164610592648387, + 0.0031072464771568775, + 0.004002246540039778, + 0.0032475192565470934, + 0.002556824591010809, + 0.0015932574169710279, + 0.0019772362429648638, + 0.002522940281778574, + 0.004023175686597824, + 0.003926669713109732, + 0.002838842337951064, + 0.0033558590803295374, + 0.001295164693146944, + 0.0011158636771142483, + 0.001507303793914616, + 0.0020572489593178034, + 0.005769716575741768, + 0.00691470131278038, + 0.00949347484856844, + 0.0038306706119328737, + 0.0035066825803369284, + 0.0023792791180312634, + 0.003907728940248489, + 0.004436036571860313, + 0.004987971391528845, + 0.005095337517559528, + 0.007246058434247971, + 0.0036402272526174784, + 0.003529417095705867, + 0.0021453839726746082, + 0.004376208409667015, + 0.0025042498018592596, + 0.028980977833271027, + 0.017206361517310143, + 0.0751536414027214, + 0.02264687977731228, + 0.007667054422199726, + 0.05916616693139076, + 0.020519476383924484, + 0.01571083813905716 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 1 + ], + "selected_reward": -0.03810866177082062 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006735597271472216 + ], + "opposite_logprob_tokens": [ + -5.006735801696777 + ], + "logprob_reward": -0.006735597271472216, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006735597271472216, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2175062894821167, + 0.03295666351914406, + 0.03617607429623604, + 0.033323321491479874, + 0.0852181687951088, + 0.024249140173196793, + 0.02157408557832241, + 0.04248794913291931, + 0.04002003371715546, + 0.4664883017539978 + ], + "credit_map": [ + 0.027486758306622505, + 0.0054153152741491795, + 0.03688786178827286, + 0.007240207865834236, + 0.015793336555361748, + 0.04943098500370979, + 0.00880188588052988, + 0.06644994765520096, + 0.005802031606435776, + 0.004156815353780985, + 0.0048235030844807625, + 0.004875452257692814, + 0.0040735346265137196, + 0.002687641652300954, + 0.003689024830237031, + 0.0028486589435487986, + 0.004796816501766443, + 0.007283062674105167, + 0.0057077533565461636, + 0.004575373604893684, + 0.003408635500818491, + 0.003986623138189316, + 0.0032443914096802473, + 0.003173419740051031, + 0.002813188824802637, + 0.004514126107096672, + 0.004371145274490118, + 0.005143621936440468, + 0.0030997912399470806, + 0.003995553124696016, + 0.0032602197024971247, + 0.006125675514340401, + 0.011491515673696995, + 0.02786804921925068, + 0.019694741815328598, + 0.006077191326767206, + 0.002974993549287319, + 0.004831448197364807, + 0.004149251617491245, + 0.00813097320497036, + 0.0033078037668019533, + 0.004016249440610409, + 0.003764139488339424, + 0.0025587568525224924, + 0.0015580413164570928, + 0.002182895317673683, + 0.0028206519782543182, + 0.004040604457259178, + 0.003578291740268469, + 0.00434207683429122, + 0.002652459079399705, + 0.0037388878408819437, + 0.0015664370730519295, + 0.001434065168723464, + 0.0018069276120513678, + 0.0024549385998398066, + 0.005954225081950426, + 0.005338822491466999, + 0.01065834704786539, + 0.004553294274955988, + 0.004425459075719118, + 0.0029807575047016144, + 0.003633263288065791, + 0.004943780601024628, + 0.005815567914396524, + 0.00652797007933259, + 0.0061899228021502495, + 0.004339179024100304, + 0.004943187348544598, + 0.00299009308218956, + 0.005061449483036995, + 0.004152664449065924, + 0.13938800990581512, + 0.03483140841126442, + 0.02730865404009819, + 0.019872579723596573, + 0.023990176618099213, + 0.17974449694156647, + 0.018078638240695, + 0.02327432855963707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.006735597271472216 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009774561040103436 + ], + "opposite_logprob_tokens": [ + -4.634774684906006 + ], + "logprob_reward": -0.009774561040103436, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009774561040103436, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06386633962392807, + 0.035933639854192734, + 0.02861415594816208, + 0.025661684572696686, + 0.38142091035842896, + 0.027239400893449783, + 0.041599612683057785, + 0.03483707457780838, + 0.07613770663738251, + 0.2846895456314087 + ], + "credit_map": [ + 0.007701519411057234, + 0.00456383777782321, + 0.007045758422464132, + 0.005231503397226334, + 0.0065640415996313095, + 0.011968975886702538, + 0.006958475802093744, + 0.013832218013703823, + 0.00433650566264987, + 0.004012639634311199, + 0.003989615477621555, + 0.0033126301132142544, + 0.0033248828258365393, + 0.00462000397965312, + 0.008548658341169357, + 0.0037887042853981256, + 0.0022519659250974655, + 0.006680153775960207, + 0.006935936398804188, + 0.003439562860876322, + 0.0021665676031261683, + 0.0026529261376708746, + 0.00201281625777483, + 0.002474224893376231, + 0.0025323964655399323, + 0.0037622430827468634, + 0.003848289605230093, + 0.0039220284670591354, + 0.0023919676896184683, + 0.002734371228143573, + 0.002665650798007846, + 0.0038047353737056255, + 0.05620504915714264, + 0.1634775847196579, + 0.1211245059967041, + 0.011866945773363113, + 0.011011694557964802, + 0.004923447500914335, + 0.004530138336122036, + 0.00828150101006031, + 0.004444119520485401, + 0.003648428013548255, + 0.005261022597551346, + 0.004491682164371014, + 0.0020184111781418324, + 0.0021397548262029886, + 0.002279105130583048, + 0.0029568744357675314, + 0.007224305998533964, + 0.008542494848370552, + 0.005448746494948864, + 0.005983722861856222, + 0.0028758442495018244, + 0.002926488406956196, + 0.0032741869799792767, + 0.005323817953467369, + 0.005168659146875143, + 0.0048882849514484406, + 0.009966977871954441, + 0.002733428729698062, + 0.0025785763282328844, + 0.0028023922350257635, + 0.002713155234232545, + 0.003985596355050802, + 0.011947013437747955, + 0.013158814050257206, + 0.014943881891667843, + 0.009607798419892788, + 0.006310204975306988, + 0.004470566753298044, + 0.008022972382605076, + 0.007676443550735712, + 0.03239291533827782, + 0.019362187013030052, + 0.09060055762529373, + 0.02666585147380829, + 0.009127043187618256, + 0.06276410073041916, + 0.024720175191760063, + 0.019056687131524086 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.009774561040103436 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast, grainy black-and-white image, the camera begins in a medium shot, focusing on the weathered and cracked carved eaves of a dilapidated Chinese temple. From the deep shadows beneath the eaves, a crow with jet-black, shimmering feathers suddenly flaps its wings, stirring up a cloud of dust and tiny sawdust. It ascends in a powerful, sharp arc toward a gray sky thick with heavy clouds. As the crow leaves its nest, the camera begins an extremely slow and steady pull-back, gradually taking in the entire temple: the abandoned structure still stands, but the roof tiles are broken and covered with moss and weeds, the wooden beams are covered with cobwebs, and the walls are mottled and peeling, revealing the relentless erosion of time. As the camera continues to retreat, the desolate scene around the temple is fully revealed. Several dead trees stretch twistedly towards the sky, the ground is overgrown with weeds, and in the distance are rolling hills with blurred outlines. The entire scene presents rich gray tones under the soft light scattered by the cloudy sky, creating a bleak atmosphere of deep loneliness and time freezing.", + "seed": 7676, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7676_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.013592545501887798, + "vlm_logprob": -0.013592545501887798, + "vlm_yes_no_margin": 4.675, + "vlm_token_credit": -0.013592545501887798, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0037188436836004257 + ], + "opposite_logprob_tokens": [ + -6.503718852996826 + ], + "logprob_reward": -0.0037188436836004257, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0037188436836004257, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06390572339296341, + 0.035444531589746475, + 0.14274145662784576, + 0.22983157634735107, + 0.04464326053857803, + 0.059584371745586395, + 0.11977356672286987, + 0.025149621069431305, + 0.19767649471759796, + 0.0812494158744812 + ], + "credit_map": [ + 0.005436046048998833, + 0.0071089123375713825, + 0.009636398404836655, + 0.00744714355096221, + 0.00610065134242177, + 0.008358458057045937, + 0.010760600678622723, + 0.009057511575520039, + 0.0035270582884550095, + 0.004819901194423437, + 0.004253271501511335, + 0.004665705841034651, + 0.004612680058926344, + 0.0021742936223745346, + 0.005379396956413984, + 0.006012223660945892, + 0.022704174742102623, + 0.03535844385623932, + 0.023309415206313133, + 0.010954919271171093, + 0.020165253430604935, + 0.006223628763109446, + 0.00978897511959076, + 0.014236635528504848, + 0.019988177344202995, + 0.033836208283901215, + 0.021716000512242317, + 0.03786013647913933, + 0.036834701895713806, + 0.038372334092855453, + 0.013616136275231838, + 0.027607867494225502, + 0.007195254322141409, + 0.007511400617659092, + 0.00463035237044096, + 0.004541419912129641, + 0.006286714691668749, + 0.006104244384914637, + 0.004269161727279425, + 0.0041047073900699615, + 0.00875786505639553, + 0.00521414540708065, + 0.007843303494155407, + 0.011624999344348907, + 0.005296077113598585, + 0.006925765424966812, + 0.006802520714700222, + 0.007119689602404833, + 0.02709888480603695, + 0.011065471917390823, + 0.009457006119191647, + 0.012956022284924984, + 0.00936286523938179, + 0.027614658698439598, + 0.006706805434077978, + 0.015511838719248772, + 0.005085206590592861, + 0.002874892670661211, + 0.0026512101758271456, + 0.003161834552884102, + 0.0024519211146980524, + 0.0030250162817537785, + 0.002148179104551673, + 0.003751358948647976, + 0.030619416385889053, + 0.030081190168857574, + 0.015802698209881783, + 0.018261272460222244, + 0.04021049663424492, + 0.020769262686371803, + 0.01309305988252163, + 0.028839075937867165, + 0.015221485868096352, + 0.006710231304168701, + 0.008142860606312752, + 0.013371576555073261, + 0.008168330416083336, + 0.010828865692019463, + 0.003965236712247133, + 0.014840823598206043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 8, + 2 + ], + "selected_reward": -0.0037188436836004257 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011085941456258297 + ], + "opposite_logprob_tokens": [ + -4.5110859870910645 + ], + "logprob_reward": -0.011085941456258297, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011085941456258297, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07045906037092209, + 0.03860101476311684, + 0.29807671904563904, + 0.04319794103503227, + 0.03378114104270935, + 0.06761988997459412, + 0.21101690828800201, + 0.032389964908361435, + 0.15382595360279083, + 0.05103137716650963 + ], + "credit_map": [ + 0.007646756712347269, + 0.007713456638157368, + 0.010752739384770393, + 0.00811157375574112, + 0.007158830761909485, + 0.008572722785174847, + 0.009616454131901264, + 0.010886527597904205, + 0.003958465531468391, + 0.004995658528059721, + 0.0038155626971274614, + 0.004927271045744419, + 0.004417032469063997, + 0.0031980257481336594, + 0.006597073748707771, + 0.006691925693303347, + 0.020495211705565453, + 0.10871231555938721, + 0.01185154914855957, + 0.023723699152469635, + 0.09665317833423615, + 0.009856724180281162, + 0.01566201262176037, + 0.01112200878560543, + 0.00399071304127574, + 0.0062955948524177074, + 0.004040880128741264, + 0.005461439955979586, + 0.00703965499997139, + 0.008159255608916283, + 0.0037905219942331314, + 0.004419880919158459, + 0.004742880817502737, + 0.008111595176160336, + 0.0035236175172030926, + 0.003578692674636841, + 0.0038030340801924467, + 0.004514369182288647, + 0.0028731978964060545, + 0.002633756259456277, + 0.011635451577603817, + 0.0058717005886137486, + 0.008224040269851685, + 0.013060030527412891, + 0.005854066926985979, + 0.008483625948429108, + 0.0069536310620605946, + 0.007537344936281443, + 0.028454843908548355, + 0.010742790065705776, + 0.008169930428266525, + 0.009782779961824417, + 0.00908597931265831, + 0.07179585844278336, + 0.007220665458589792, + 0.06576406210660934, + 0.006534696090966463, + 0.0025026469957083464, + 0.0049949828535318375, + 0.0031197909265756607, + 0.003041359595954418, + 0.004144942853599787, + 0.0031032427214086056, + 0.0049483017064630985, + 0.022441303357481956, + 0.029775284230709076, + 0.011500387452542782, + 0.01572774350643158, + 0.027958791702985764, + 0.015439450740814209, + 0.009571121074259281, + 0.021411875262856483, + 0.0072611551731824875, + 0.0059393043629825115, + 0.0059408992528915405, + 0.006767691113054752, + 0.0078049274161458015, + 0.007346542552113533, + 0.0035922552924603224, + 0.006378599908202887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 6, + 8 + ], + "selected_reward": -0.011085941456258297 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026345498859882355 + ], + "opposite_logprob_tokens": [ + -3.6513454914093018 + ], + "logprob_reward": -0.026345498859882355, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026345498859882355, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12754006683826447, + 0.032834842801094055, + 0.07043043524026871, + 0.09961007535457611, + 0.22454501688480377, + 0.1718568503856659, + 0.09970682114362717, + 0.021041572093963623, + 0.1131267324090004, + 0.039307549595832825 + ], + "credit_map": [ + 0.011619849130511284, + 0.009137311950325966, + 0.011062657460570335, + 0.01550059113651514, + 0.007339066360145807, + 0.0168012548238039, + 0.01267768070101738, + 0.04340166971087456, + 0.0036288124974817038, + 0.0062563493847846985, + 0.002984057879075408, + 0.003784505883231759, + 0.00493455957621336, + 0.002471870742738247, + 0.004310511518269777, + 0.004464178346097469, + 0.005270233377814293, + 0.018286436796188354, + 0.007329101208597422, + 0.006622387561947107, + 0.016656795516610146, + 0.003559876699000597, + 0.004875758197158575, + 0.007829846814274788, + 0.008817109279334545, + 0.01395860593765974, + 0.009058778174221516, + 0.017231514677405357, + 0.01709040440618992, + 0.015285644680261612, + 0.006362596061080694, + 0.011805429123342037, + 0.03223929554224014, + 0.04250483587384224, + 0.022661607712507248, + 0.022135362029075623, + 0.02568361908197403, + 0.030476056039333344, + 0.01904119923710823, + 0.029803065583109856, + 0.025269601494073868, + 0.012900259345769882, + 0.022491455078125, + 0.040036316961050034, + 0.015128024853765965, + 0.017137372866272926, + 0.022094126790761948, + 0.01679971255362034, + 0.020242983475327492, + 0.010073241777718067, + 0.006778569892048836, + 0.008466510102152824, + 0.008128391578793526, + 0.028046831488609314, + 0.006510753184556961, + 0.011459547095000744, + 0.004755181726068258, + 0.002604530192911625, + 0.0018429774791002274, + 0.0020979775581508875, + 0.0022069092374294996, + 0.0028630367014557123, + 0.0017690971726551652, + 0.0029018630739301443, + 0.01804940402507782, + 0.020315123721957207, + 0.008262032642960548, + 0.013981405645608902, + 0.017923850566148758, + 0.011290464550256729, + 0.005642641335725784, + 0.01766182854771614, + 0.0067831166088581085, + 0.005139779299497604, + 0.004283384885638952, + 0.004537713713943958, + 0.00619554752483964, + 0.004834727384150028, + 0.0033946039620786905, + 0.0041386825032532215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.026345498859882355 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008635558187961578 + ], + "opposite_logprob_tokens": [ + -4.758635520935059 + ], + "logprob_reward": -0.008635558187961578, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008635558187961578, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12560293078422546, + 0.03538793325424194, + 0.1545793116092682, + 0.18176105618476868, + 0.062355343252420425, + 0.19903872907161713, + 0.12499842047691345, + 0.03600401058793068, + 0.044136762619018555, + 0.036135539412498474 + ], + "credit_map": [ + 0.010387323796749115, + 0.007650046143680811, + 0.012135801836848259, + 0.016437292098999023, + 0.007290187757462263, + 0.016396984457969666, + 0.016202449798583984, + 0.039102841168642044, + 0.003268040716648102, + 0.004651952534914017, + 0.004756298381835222, + 0.004467617720365524, + 0.004101782105863094, + 0.002948739565908909, + 0.0057386974804103374, + 0.005454803351312876, + 0.011963287368416786, + 0.0475648008286953, + 0.01588326506316662, + 0.017755959182977676, + 0.03849291801452637, + 0.0060037546791136265, + 0.009456741623580456, + 0.00745859369635582, + 0.01567775197327137, + 0.025452882051467896, + 0.016981216147542, + 0.03116387315094471, + 0.029326582327485085, + 0.03006306290626526, + 0.010913901031017303, + 0.022181790322065353, + 0.01043878123164177, + 0.009061849676072598, + 0.006605265662074089, + 0.005187360569834709, + 0.0067017460241913795, + 0.008064163848757744, + 0.005466820672154427, + 0.01082935743033886, + 0.027406632900238037, + 0.02017676644027233, + 0.029012905433773994, + 0.0444767065346241, + 0.01787714660167694, + 0.017030319198966026, + 0.02303520403802395, + 0.020023060962557793, + 0.027831347659230232, + 0.014724429696798325, + 0.009463757276535034, + 0.012953663244843483, + 0.00885819736868143, + 0.03150312602519989, + 0.0066202618181705475, + 0.013043638318777084, + 0.00635235570371151, + 0.0038273115642368793, + 0.0052479044534265995, + 0.0037630137521773577, + 0.004118903074413538, + 0.0038628526963293552, + 0.0036246604286134243, + 0.005207007750868797, + 0.0059475102461874485, + 0.007191046141088009, + 0.00486905500292778, + 0.006556978914886713, + 0.004120934754610062, + 0.005703417118638754, + 0.004289652220904827, + 0.005458166357129812, + 0.0042142379097640514, + 0.0036879091057926416, + 0.00568512873724103, + 0.005815618205815554, + 0.0035354369319975376, + 0.004409696441143751, + 0.0036564068868756294, + 0.00513110589236021 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.008635558187961578 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018176885321736336 + ], + "opposite_logprob_tokens": [ + -4.018177032470703 + ], + "logprob_reward": -0.018176885321736336, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018176885321736336, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0867965966463089, + 0.03540213406085968, + 0.1605573296546936, + 0.08874345570802689, + 0.028446344658732414, + 0.08012066036462784, + 0.341749370098114, + 0.03185855969786644, + 0.11790039390325546, + 0.028425132855772972 + ], + "credit_map": [ + 0.011394353583455086, + 0.011008123867213726, + 0.009596304967999458, + 0.010293862782418728, + 0.00633376557379961, + 0.00823232065886259, + 0.011250483803451061, + 0.018687384203076363, + 0.004502362571656704, + 0.006276080850511789, + 0.004140973556786776, + 0.004202563315629959, + 0.0033836516086012125, + 0.0030169663950800896, + 0.005207201931625605, + 0.004672338720411062, + 0.0168940220028162, + 0.041471216827631, + 0.018921952694654465, + 0.014947974123060703, + 0.04305649921298027, + 0.006678967271000147, + 0.009456277824938297, + 0.009130432270467281, + 0.007833890616893768, + 0.012349565513432026, + 0.008620282635092735, + 0.014885598793625832, + 0.013444620184600353, + 0.015148306265473366, + 0.005929618142545223, + 0.010531585663557053, + 0.0038372348062694073, + 0.00437192153185606, + 0.0032955314964056015, + 0.002525049028918147, + 0.00289377779699862, + 0.003733712946996093, + 0.0027011088095605373, + 0.0050880094058811665, + 0.011091346852481365, + 0.007680131588131189, + 0.012183469720184803, + 0.018057486042380333, + 0.007414639927446842, + 0.006166817620396614, + 0.009386570192873478, + 0.008140206336975098, + 0.04927469789981842, + 0.008827238343656063, + 0.008010739460587502, + 0.006666835863143206, + 0.0125690633431077, + 0.1424134373664856, + 0.009195087477564812, + 0.10479229688644409, + 0.0062954421155154705, + 0.0023680361919105053, + 0.004033321980386972, + 0.0033192764967679977, + 0.0029871920123696327, + 0.003932994324713945, + 0.002447501989081502, + 0.006474796682596207, + 0.018550852313637733, + 0.020595859736204147, + 0.009087893180549145, + 0.012800744734704494, + 0.020811671391129494, + 0.011957768350839615, + 0.006272580940276384, + 0.017823033034801483, + 0.004271049052476883, + 0.003135930746793747, + 0.002142850076779723, + 0.005167581606656313, + 0.004233364947140217, + 0.0031958671752363443, + 0.0017925366992130876, + 0.0044859545305371284 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 8 + ], + "selected_reward": -0.018176885321736336 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast, grainy black-and-white image, the camera begins in a medium shot, focusing on the weathered and cracked carved eaves of a dilapidated Chinese temple. From the deep shadows beneath the eaves, a crow with jet-black, shimmering feathers suddenly flaps its wings, stirring up a cloud of dust and tiny sawdust. It ascends in a powerful, sharp arc toward a gray sky thick with heavy clouds. As the crow leaves its nest, the camera begins an extremely slow and steady pull-back, gradually taking in the entire temple: the abandoned structure still stands, but the roof tiles are broken and covered with moss and weeds, the wooden beams are covered with cobwebs, and the walls are mottled and peeling, revealing the relentless erosion of time. As the camera continues to retreat, the desolate scene around the temple is fully revealed. Several dead trees stretch twistedly towards the sky, the ground is overgrown with weeds, and in the distance are rolling hills with blurred outlines. The entire scene presents rich gray tones under the soft light scattered by the cloudy sky, creating a bleak atmosphere of deep loneliness and time freezing.", + "seed": 7677, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7677_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.10222484227269887, + "vlm_logprob": -0.10222484227269887, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.10222484227269887, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4765782952308655 + ], + "opposite_logprob_tokens": [ + -0.9765782952308655 + ], + "logprob_reward": -0.4765782952308655, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4765782952308655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5894966125488281, + 0.03991517797112465, + 0.02871934324502945, + 0.04139509052038193, + 0.03968639671802521, + 0.05042751878499985, + 0.1284526139497757, + 0.033023059368133545, + 0.029960691928863525, + 0.018923485651612282 + ], + "credit_map": [ + 0.2031947821378708, + 0.03172960504889488, + 0.0539538711309433, + 0.01522490382194519, + 0.026089690625667572, + 0.2230919897556305, + 0.013822030276060104, + 0.022389741614460945, + 0.004995228257030249, + 0.004692086949944496, + 0.0058458540588617325, + 0.005572436843067408, + 0.0039405375719070435, + 0.0038301385939121246, + 0.003953703213483095, + 0.0070851934142410755, + 0.004052579868584871, + 0.0034631858579814434, + 0.003416200401261449, + 0.0021004374139010906, + 0.004432234913110733, + 0.003560721641406417, + 0.0034945732913911343, + 0.00419941172003746, + 0.005936526693403721, + 0.009799966588616371, + 0.003749338909983635, + 0.003188774921000004, + 0.007400716654956341, + 0.005605760961771011, + 0.002560778521001339, + 0.0031532300636172295, + 0.004567185882478952, + 0.00593275623396039, + 0.005820762366056442, + 0.005962316878139973, + 0.004125430714339018, + 0.0028900906909257174, + 0.0032865528482943773, + 0.007101298775523901, + 0.008073178119957447, + 0.00834877509623766, + 0.005193554796278477, + 0.004636652301996946, + 0.007570565212517977, + 0.006249310914427042, + 0.0031597898341715336, + 0.007195692043751478, + 0.024238957092165947, + 0.025236982852220535, + 0.025197191163897514, + 0.010695305652916431, + 0.01297951303422451, + 0.009431695565581322, + 0.007978839799761772, + 0.012694131582975388, + 0.0039992136880755424, + 0.0061668590642511845, + 0.005219812970608473, + 0.003315265756100416, + 0.006580243352800608, + 0.0029304318595677614, + 0.0027673225849866867, + 0.0020439084619283676, + 0.005157658364623785, + 0.00463824300095439, + 0.003512046532705426, + 0.0029343620408326387, + 0.0062622930854558945, + 0.0032615670934319496, + 0.00253871432505548, + 0.0016558084171265364, + 0.003038024064153433, + 0.0028674595523625612, + 0.002646319568157196, + 0.002220112830400467, + 0.0028705589938908815, + 0.0019364510662853718, + 0.0017507110023871064, + 0.0015938481083139777 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.4765782952308655 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004649423528462648 + ], + "opposite_logprob_tokens": [ + -5.379649639129639 + ], + "logprob_reward": -0.004649423528462648, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004649423528462648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.559869647026062, + 0.06720049679279327, + 0.03886869177222252, + 0.027499860152602196, + 0.06943494081497192, + 0.09530750662088394, + 0.05994997173547745, + 0.029504630714654922, + 0.02370397374033928, + 0.028660304844379425 + ], + "credit_map": [ + 0.20439450442790985, + 0.025334369391202927, + 0.05646723136305809, + 0.029043499380350113, + 0.02754676155745983, + 0.16706271469593048, + 0.022634347900748253, + 0.02738618291914463, + 0.008267487399280071, + 0.010174795985221863, + 0.007070796564221382, + 0.005708317272365093, + 0.010623523034155369, + 0.007937224581837654, + 0.008630305528640747, + 0.008788052015006542, + 0.005668085068464279, + 0.006932549644261599, + 0.003775389399379492, + 0.00267870188690722, + 0.005331249907612801, + 0.003974813036620617, + 0.004214026033878326, + 0.006293874699622393, + 0.0038279143627732992, + 0.005595753435045481, + 0.0030510523356497288, + 0.0021157478913664818, + 0.004316036123782396, + 0.004252607934176922, + 0.001873939880169928, + 0.0024668097030371428, + 0.009131603874266148, + 0.013616974465548992, + 0.007828122936189175, + 0.009043511003255844, + 0.007834622636437416, + 0.006457029841840267, + 0.006415649317204952, + 0.009107422083616257, + 0.016221027821302414, + 0.019681286066770554, + 0.00834150891751051, + 0.007705364841967821, + 0.016442719846963882, + 0.009898334741592407, + 0.007005093153566122, + 0.010012171231210232, + 0.00812904629856348, + 0.007015963084995747, + 0.011842970736324787, + 0.006199679337441921, + 0.005779207684099674, + 0.006988632958382368, + 0.0044514830224215984, + 0.009542984887957573, + 0.004386867396533489, + 0.005153272766619921, + 0.004848000593483448, + 0.003340326715260744, + 0.005193666554987431, + 0.00294895120896399, + 0.0023423738311976194, + 0.001291171065531671, + 0.0023673763498663902, + 0.0028986686374992132, + 0.0037800283171236515, + 0.00343369971960783, + 0.003918088506907225, + 0.0027897909749299288, + 0.0029678121209144592, + 0.0015485073672607541, + 0.00410485127940774, + 0.003877524519339204, + 0.003365962766110897, + 0.002645638305693865, + 0.006351223681122065, + 0.0032443231903016567, + 0.0030162709299474955, + 0.002054510870948434 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.004649423528462648 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419647503644228 + ], + "opposite_logprob_tokens": [ + -4.264196395874023 + ], + "logprob_reward": -0.01419647503644228, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419647503644228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.7702526450157166, + 0.06625210493803024, + 0.01369006559252739, + 0.01273542270064354, + 0.04502676799893379, + 0.017241301015019417, + 0.026638785377144814, + 0.020115572959184647, + 0.01631109230220318, + 0.011736293323338032 + ], + "credit_map": [ + 0.27378055453300476, + 0.038656074553728104, + 0.111205093562603, + 0.03957267478108406, + 0.044647619128227234, + 0.18979769945144653, + 0.034128282219171524, + 0.03846464306116104, + 0.0075049190782010555, + 0.006343204528093338, + 0.007607651874423027, + 0.008531665429472923, + 0.010101822204887867, + 0.006834579631686211, + 0.013328421860933304, + 0.005999839399009943, + 0.0016873001586645842, + 0.0016784887993708253, + 0.0019309560302644968, + 0.001269686734303832, + 0.0015907268971204758, + 0.001541651552543044, + 0.0014836409827694297, + 0.002507612807676196, + 0.001837297691963613, + 0.003634550143033266, + 0.0014432499883696437, + 0.0010937583865597844, + 0.0016223518177866936, + 0.0013152874307706952, + 0.0007640483672730625, + 0.0010248770704492927, + 0.007385301869362593, + 0.012897753156721592, + 0.004067411180585623, + 0.004925093613564968, + 0.006189692299813032, + 0.0029004011303186417, + 0.0025246981531381607, + 0.004136413335800171, + 0.002862880937755108, + 0.003946585115045309, + 0.001834021764807403, + 0.0012648808769881725, + 0.0033156254794448614, + 0.0016250582411885262, + 0.0010205842554569244, + 0.0013716632965952158, + 0.004697094205766916, + 0.004111901856958866, + 0.005445633549243212, + 0.002274305559694767, + 0.002677928190678358, + 0.0027368334122002125, + 0.001928652636706829, + 0.0027664334047585726, + 0.002330247312784195, + 0.0037780930288136005, + 0.0023305814247578382, + 0.0017343776999041438, + 0.004772759508341551, + 0.00198499602265656, + 0.0021296388003975153, + 0.0010548776481300592, + 0.00267430255189538, + 0.0020872822497040033, + 0.0028931819833815098, + 0.0021644087973982096, + 0.002538206521421671, + 0.0016014949651435018, + 0.001396204112097621, + 0.0009560095495544374, + 0.002394451992586255, + 0.0018483122112229466, + 0.0011471437755972147, + 0.0009281547972932458, + 0.0020048306323587894, + 0.001171363634057343, + 0.0011119593400508165, + 0.0011300757760182023 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.01419647503644228 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634353797882795 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634353797882795, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634353797882795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5040497779846191, + 0.055191121995449066, + 0.04279019311070442, + 0.06307904422283173, + 0.057194314897060394, + 0.05633222311735153, + 0.07156568765640259, + 0.05798520892858505, + 0.05535402521491051, + 0.036458417773246765 + ], + "credit_map": [ + 0.192519873380661, + 0.02973276376724243, + 0.03855212777853012, + 0.020653776824474335, + 0.021743085235357285, + 0.16456009447574615, + 0.013546948321163654, + 0.022741137072443962, + 0.007883240468800068, + 0.008750072680413723, + 0.007381913717836142, + 0.003255831776186824, + 0.009406021796166897, + 0.006564044859260321, + 0.007337138056755066, + 0.004612859804183245, + 0.006083714310079813, + 0.006752015557140112, + 0.005452283658087254, + 0.0025653585325926542, + 0.006016043946146965, + 0.004663851577788591, + 0.003876760834828019, + 0.007380164694041014, + 0.008756673894822598, + 0.011479952372610569, + 0.005535789765417576, + 0.006355497520416975, + 0.012066135182976723, + 0.009990430437028408, + 0.004366472829133272, + 0.004528089892119169, + 0.01031282264739275, + 0.017623135820031166, + 0.004346220288425684, + 0.0053757852874696255, + 0.008087637834250927, + 0.004168469924479723, + 0.003199493046849966, + 0.004080752842128277, + 0.009437820874154568, + 0.009407369419932365, + 0.005438520107418299, + 0.004179393406957388, + 0.00876183807849884, + 0.0057919020764529705, + 0.003957472275942564, + 0.009357905946671963, + 0.012022373266518116, + 0.010889757424592972, + 0.013650478795170784, + 0.006814845371991396, + 0.007903805933892727, + 0.008931010030210018, + 0.00587523402646184, + 0.005478181876242161, + 0.009384242817759514, + 0.008897153660655022, + 0.00861250888556242, + 0.00711384043097496, + 0.00918691884726286, + 0.004870607051998377, + 0.005636653862893581, + 0.004283282905817032, + 0.006273719947785139, + 0.008363822475075722, + 0.006880832836031914, + 0.005768573377281427, + 0.012088384479284286, + 0.006045249290764332, + 0.00648812111467123, + 0.0034453195985406637, + 0.005808528512716293, + 0.004654133692383766, + 0.006406093947589397, + 0.0043507968075573444, + 0.00352756236679852, + 0.003949051257222891, + 0.004257761873304844, + 0.003504488617181778 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 3 + ], + "selected_reward": -0.004634353797882795 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011065663769841194 + ], + "opposite_logprob_tokens": [ + -4.511065483093262 + ], + "logprob_reward": -0.011065663769841194, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011065663769841194, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39628297090530396, + 0.05142417550086975, + 0.02689845860004425, + 0.049273531883955, + 0.14563776552677155, + 0.037896547466516495, + 0.08876512944698334, + 0.04177981615066528, + 0.13051146268844604, + 0.03153014928102493 + ], + "credit_map": [ + 0.09924911707639694, + 0.027915704995393753, + 0.05524839833378792, + 0.015776406973600388, + 0.020686525851488113, + 0.13353994488716125, + 0.02564212493598461, + 0.01822475716471672, + 0.007058107294142246, + 0.008016991429030895, + 0.005512090399861336, + 0.004592206794768572, + 0.009424323216080666, + 0.006504017394036055, + 0.0049385675229132175, + 0.005377868190407753, + 0.0037299394607543945, + 0.0034108750987797976, + 0.002906843088567257, + 0.0023776923771947622, + 0.003161763306707144, + 0.0026125446893274784, + 0.003271928057074547, + 0.005426873918622732, + 0.00674145482480526, + 0.008077891543507576, + 0.0059900470077991486, + 0.004748913925141096, + 0.009178354404866695, + 0.007879265584051609, + 0.0033488108310848475, + 0.0033087937626987696, + 0.01996603049337864, + 0.025939999148249626, + 0.016824165359139442, + 0.021311890333890915, + 0.013136647641658783, + 0.012043381109833717, + 0.01498840656131506, + 0.021427234634757042, + 0.0048971823416650295, + 0.004615137819200754, + 0.00532768527045846, + 0.005803920328617096, + 0.0036284751258790493, + 0.003600524738430977, + 0.0024802619591355324, + 0.007543359883129597, + 0.009934543631970882, + 0.012685972265899181, + 0.029087666422128677, + 0.009779468178749084, + 0.00829534325748682, + 0.007867718115448952, + 0.0061639463528990746, + 0.0049504670314490795, + 0.00730176642537117, + 0.005057357717305422, + 0.006499165669083595, + 0.007801671512424946, + 0.003905900986865163, + 0.004044867120683193, + 0.0036851754412055016, + 0.0034839108120650053, + 0.02084840089082718, + 0.015945497900247574, + 0.025757864117622375, + 0.017923973500728607, + 0.017284369096159935, + 0.012989346869289875, + 0.012369932606816292, + 0.007392081897705793, + 0.00643977802246809, + 0.00473828287795186, + 0.0050588748417794704, + 0.0031657423824071884, + 0.0034498339518904686, + 0.0027053034864366055, + 0.003589668544009328, + 0.0023826630786061287 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.011065663769841194 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a high-contrast, grainy black-and-white image, the camera begins in a medium shot, focusing on the weathered and cracked carved eaves of a dilapidated Chinese temple. From the deep shadows beneath the eaves, a crow with jet-black, shimmering feathers suddenly flaps its wings, stirring up a cloud of dust and tiny sawdust. It ascends in a powerful, sharp arc toward a gray sky thick with heavy clouds. As the crow leaves its nest, the camera begins an extremely slow and steady pull-back, gradually taking in the entire temple: the abandoned structure still stands, but the roof tiles are broken and covered with moss and weeds, the wooden beams are covered with cobwebs, and the walls are mottled and peeling, revealing the relentless erosion of time. As the camera continues to retreat, the desolate scene around the temple is fully revealed. Several dead trees stretch twistedly towards the sky, the ground is overgrown with weeds, and in the distance are rolling hills with blurred outlines. The entire scene presents rich gray tones under the soft light scattered by the cloudy sky, creating a bleak atmosphere of deep loneliness and time freezing.", + "seed": 7678, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_6/index_6339_seed_7678_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02210993319749832, + "vlm_logprob": -0.02210993319749832, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.02210993319749832, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01083927508443594 + ], + "opposite_logprob_tokens": [ + -4.885839462280273 + ], + "logprob_reward": -0.01083927508443594, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.01083927508443594, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12574465572834015, + 0.061783671379089355, + 0.03109925054013729, + 0.04192091152071953, + 0.04475721716880798, + 0.046458952128887177, + 0.044836051762104034, + 0.43005383014678955, + 0.11963474750518799, + 0.053710706532001495 + ], + "credit_map": [ + 0.013370323926210403, + 0.01163894310593605, + 0.017548583447933197, + 0.01639997400343418, + 0.01721530221402645, + 0.017970601096749306, + 0.013120822608470917, + 0.018480101600289345, + 0.004634471610188484, + 0.0064993686974048615, + 0.0036476498935371637, + 0.008748264983296394, + 0.0064565446227788925, + 0.005143057554960251, + 0.015108003281056881, + 0.011546311900019646, + 0.0038241343572735786, + 0.004566949792206287, + 0.0030922472942620516, + 0.002962112193927169, + 0.004690111614763737, + 0.002618530998006463, + 0.005646019708365202, + 0.00369914504699409, + 0.006376525387167931, + 0.0040875147096812725, + 0.006244324147701263, + 0.006147149950265884, + 0.005843735300004482, + 0.0026644656900316477, + 0.0054212757386267185, + 0.00513592129573226, + 0.005706208758056164, + 0.005739681422710419, + 0.004799206741154194, + 0.0063081844709813595, + 0.009150822646915913, + 0.004406038671731949, + 0.0045547690242528915, + 0.004092303104698658, + 0.011063046753406525, + 0.0054177287966012955, + 0.004996754694730043, + 0.005818684119731188, + 0.005129410419613123, + 0.007061671931296587, + 0.0032808040268719196, + 0.0036908516194671392, + 0.009211008436977863, + 0.006129271350800991, + 0.007782893721014261, + 0.0031694790814071894, + 0.0050904895178973675, + 0.007295104209333658, + 0.002861813874915242, + 0.00329599273391068, + 0.10067135095596313, + 0.039131708443164825, + 0.07587172836065292, + 0.03568387031555176, + 0.029123641550540924, + 0.07389775663614273, + 0.02933911420404911, + 0.046334657818078995, + 0.023616643622517586, + 0.01675032265484333, + 0.006652959156781435, + 0.004440776538103819, + 0.007939823903143406, + 0.04986710473895073, + 0.003869260661303997, + 0.006497848313301802, + 0.008776978589594364, + 0.00640249066054821, + 0.008056854829192162, + 0.008871418423950672, + 0.004366222769021988, + 0.00826827622950077, + 0.003835381241515279, + 0.005133082624524832 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.01083927508443594 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01607939414680004 + ], + "opposite_logprob_tokens": [ + -4.141079425811768 + ], + "logprob_reward": -0.01607939414680004, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01607939414680004, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26153260469436646, + 0.1315169781446457, + 0.054948367178440094, + 0.07417580485343933, + 0.04691661521792412, + 0.07096216827630997, + 0.09782825410366058, + 0.13400670886039734, + 0.05339592322707176, + 0.07471660524606705 + ], + "credit_map": [ + 0.028358273208141327, + 0.039615120738744736, + 0.04687220975756645, + 0.04294917359948158, + 0.019653039053082466, + 0.015494785271584988, + 0.040812864899635315, + 0.027777140960097313, + 0.011611752212047577, + 0.015436174347996712, + 0.007792093325406313, + 0.02005261741578579, + 0.015454346314072609, + 0.010824463330209255, + 0.029373060911893845, + 0.020972466096282005, + 0.007655829191207886, + 0.008115068078041077, + 0.005412305239588022, + 0.00539200147613883, + 0.006743887905031443, + 0.0035201809369027615, + 0.009427652694284916, + 0.008681442588567734, + 0.009146936237812042, + 0.009504901245236397, + 0.011557294987142086, + 0.014861373230814934, + 0.006665942724794149, + 0.0043874336406588554, + 0.01033541839569807, + 0.00771650206297636, + 0.007146492600440979, + 0.006373945157974958, + 0.006439527962356806, + 0.006451830267906189, + 0.005095573142170906, + 0.004584154114127159, + 0.005857184529304504, + 0.004967906512320042, + 0.014176429249346256, + 0.009958753362298012, + 0.0072402567602694035, + 0.006339464336633682, + 0.010192815214395523, + 0.01035191398113966, + 0.004837931599467993, + 0.007864605635404587, + 0.02579464763402939, + 0.01205455232411623, + 0.014359544962644577, + 0.006357911508530378, + 0.008878996595740318, + 0.018589887768030167, + 0.005480603314936161, + 0.006312109064310789, + 0.033000651746988297, + 0.013491238467395306, + 0.01924930512905121, + 0.0111281368881464, + 0.009247968904674053, + 0.020871799439191818, + 0.00993870384991169, + 0.01707889884710312, + 0.009296806529164314, + 0.007657643873244524, + 0.005224116146564484, + 0.003867322811856866, + 0.004391470458358526, + 0.014537290669977665, + 0.002814855659380555, + 0.005606417544186115, + 0.013030108995735645, + 0.007759723346680403, + 0.012712119147181511, + 0.007635988295078278, + 0.0059111337177455425, + 0.011221823282539845, + 0.004985558334738016, + 0.011460153385996819 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 1 + ], + "selected_reward": -0.01607939414680004 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048659659922122955 + ], + "opposite_logprob_tokens": [ + -3.048659563064575 + ], + "logprob_reward": -0.048659659922122955, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048659659922122955, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09976295381784439, + 0.03704264760017395, + 0.029575124382972717, + 0.03162049874663353, + 0.019052093848586082, + 0.03549477830529213, + 0.042318642139434814, + 0.08157990127801895, + 0.585294783115387, + 0.03825859725475311 + ], + "credit_map": [ + 0.011811506934463978, + 0.016397541388869286, + 0.016952572390437126, + 0.012824011035263538, + 0.009141986258327961, + 0.00690914411097765, + 0.014564180746674538, + 0.011162012815475464, + 0.0039045284502208233, + 0.0047543952241539955, + 0.0024028983898460865, + 0.005257315468043089, + 0.004090236034244299, + 0.002915452467277646, + 0.007627660874277353, + 0.006090158596634865, + 0.004488231614232063, + 0.0034844214096665382, + 0.0038884254172444344, + 0.003278142074123025, + 0.0035403971560299397, + 0.0019426619401201606, + 0.004638523794710636, + 0.004314320161938667, + 0.0028769068885594606, + 0.0050513241440057755, + 0.00456312345340848, + 0.004210704006254673, + 0.004057405982166529, + 0.0027265807148069143, + 0.004176502116024494, + 0.0039579495787620544, + 0.0033526825718581676, + 0.002394844312220812, + 0.0021255617029964924, + 0.0022093181032687426, + 0.0025801039300858974, + 0.002003332832828164, + 0.0021956039126962423, + 0.002190646016970277, + 0.010625860653817654, + 0.003908541984856129, + 0.0030015946831554174, + 0.0031078371684998274, + 0.002273318124935031, + 0.007487145718187094, + 0.0021628187969326973, + 0.002927660010755062, + 0.008140591904520988, + 0.004754970315843821, + 0.010126182809472084, + 0.0025969333946704865, + 0.0042389873415231705, + 0.0060749598778784275, + 0.0022631993051618338, + 0.004122817888855934, + 0.01577869802713394, + 0.0067087397910654545, + 0.016253791749477386, + 0.007496248465031385, + 0.005515689495950937, + 0.013388414867222309, + 0.005762005690485239, + 0.010676320642232895, + 0.12546850740909576, + 0.08804900199174881, + 0.03139618784189224, + 0.022864120081067085, + 0.03540058806538582, + 0.2289511114358902, + 0.020226750522851944, + 0.032938484102487564, + 0.0055769868195056915, + 0.0034923297353088856, + 0.005913779139518738, + 0.004690045490860939, + 0.003984570503234863, + 0.004382847808301449, + 0.004082236438989639, + 0.006135803181678057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.048659659922122955 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637449704110622 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637449704110622, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637449704110622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10509829968214035, + 0.06185464560985565, + 0.06399659067392349, + 0.0974358469247818, + 0.03388825058937073, + 0.06051810458302498, + 0.07678597420454025, + 0.07812945544719696, + 0.2318493127822876, + 0.19044354557991028 + ], + "credit_map": [ + 0.01164553128182888, + 0.013703673146665096, + 0.016481924802064896, + 0.01480211690068245, + 0.009714239276945591, + 0.008321410976350307, + 0.015226791612803936, + 0.015202607028186321, + 0.00632958160713315, + 0.008455537259578705, + 0.0036541444715112448, + 0.009322370402514935, + 0.007405173033475876, + 0.004652093164622784, + 0.012158685363829136, + 0.009877058677375317, + 0.007664041593670845, + 0.009160551242530346, + 0.00535295344889164, + 0.005921582225710154, + 0.008087527938187122, + 0.004577040206640959, + 0.012644928880035877, + 0.010587960481643677, + 0.011057985946536064, + 0.01272769458591938, + 0.011190639808773994, + 0.011860006488859653, + 0.00998245645314455, + 0.006195929367095232, + 0.015162484720349312, + 0.019258644431829453, + 0.00556776812300086, + 0.004133292473852634, + 0.003646241268143058, + 0.004876938182860613, + 0.005107616540044546, + 0.004090439528226852, + 0.004067312926054001, + 0.002398635493591428, + 0.019800405949354172, + 0.006907223723828793, + 0.004385022912174463, + 0.004732639994472265, + 0.004850017838180065, + 0.011716596782207489, + 0.003431893652305007, + 0.004694296512752771, + 0.021254276856780052, + 0.007542197126895189, + 0.01692734844982624, + 0.0037174297031015158, + 0.003523745574057102, + 0.01460007019340992, + 0.002619562903419137, + 0.006601332221180201, + 0.020980950444936752, + 0.010877992026507854, + 0.010226219892501831, + 0.004144702572375536, + 0.004453778266906738, + 0.01741621270775795, + 0.0030449158512055874, + 0.006984673906117678, + 0.043794430792331696, + 0.030797887593507767, + 0.015412730164825916, + 0.014026816934347153, + 0.015346620231866837, + 0.0866873562335968, + 0.010272534564137459, + 0.015510906465351582, + 0.033253856003284454, + 0.01858360879123211, + 0.03534315153956413, + 0.02137720212340355, + 0.013972227461636066, + 0.029265016317367554, + 0.011761859990656376, + 0.026886610314249992 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.008637449704110622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video presented in black-and-white with a grainy texture?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the video begin with a medium shot focusing on weathered and cracked carved eaves of a dilapidated Chinese temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does a crow with jet-black, shimmering feathers flap its wings from beneath the eaves, stirring up dust and sawdust?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera perform an extremely slow and steady pull-back to reveal the entire temple, including broken roof tiles, moss, weeds, cobwebs, and peeling walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background revealed to include dead trees stretching towards the sky, overgrown ground, and distant rolling hills with blurred outlines?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02633388713002205 + ], + "opposite_logprob_tokens": [ + -3.651333808898926 + ], + "logprob_reward": -0.02633388713002205, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02633388713002205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1634238064289093, + 0.07721978425979614, + 0.0449005626142025, + 0.046065181493759155, + 0.022120079025626183, + 0.10068268328905106, + 0.1401512175798416, + 0.09327738732099533, + 0.1346728652715683, + 0.17748644948005676 + ], + "credit_map": [ + 0.017486276105046272, + 0.022313322871923447, + 0.024174978956580162, + 0.022237490862607956, + 0.014851539395749569, + 0.013194163329899311, + 0.029928071424365044, + 0.01923796907067299, + 0.008330692537128925, + 0.00912446714937687, + 0.006094607058912516, + 0.010895549319684505, + 0.007632351480424404, + 0.005455800332129002, + 0.018127834424376488, + 0.01155848614871502, + 0.005015674978494644, + 0.006894515361636877, + 0.003940674010664225, + 0.004946897737681866, + 0.006674824748188257, + 0.0030453610233962536, + 0.007905090227723122, + 0.006477522663772106, + 0.004588646814227104, + 0.006014722865074873, + 0.00689337681978941, + 0.008877233602106571, + 0.0065095010213553905, + 0.0031380127184093, + 0.006046816240996122, + 0.003996872343122959, + 0.0039755068719387054, + 0.0017148852348327637, + 0.002255986910313368, + 0.0039042646531015635, + 0.002458680886775255, + 0.003118466353043914, + 0.00256215943954885, + 0.0021301291417330503, + 0.028469888493418694, + 0.014165421947836876, + 0.0079537034034729, + 0.010380322113633156, + 0.006521700881421566, + 0.01865915022790432, + 0.005915559828281403, + 0.008616930805146694, + 0.0382918156683445, + 0.013978107832372189, + 0.0307147316634655, + 0.006861508823931217, + 0.006764284800738096, + 0.027705108746886253, + 0.00498861214146018, + 0.010847039520740509, + 0.024022264406085014, + 0.012038863264024258, + 0.015632694587111473, + 0.004967071115970612, + 0.005794505123049021, + 0.019161205738782883, + 0.0041432748548686504, + 0.007517512422055006, + 0.02735091559588909, + 0.015733584761619568, + 0.019691146910190582, + 0.008344532921910286, + 0.007524112239480019, + 0.042623598128557205, + 0.005522464402019978, + 0.007882492616772652, + 0.02997477725148201, + 0.02759861946105957, + 0.025186477228999138, + 0.02261686697602272, + 0.013268101960420609, + 0.027442025020718575, + 0.013486884534358978, + 0.017912695184350014 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.02633388713002205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_tensor([4642, 4642])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_tensor([4642, 4642])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4d4362ed1783a30c39eb41c971cec678bf333b5c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_tensor([4642, 4642])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, handheld medium shot struggles to keep pace, capturing an intense and visceral jolt of camera movement as it tracks a jubilant corgi with thick, reddish-brown and white fur and an unusually fluffy tail. The dog bounds across a large expanse of lush green grass, its short legs thrashing frantically in a blur of motion, its entire drooping body thrust forward with boundless energy. Its dark eyes are wide with pure joy, its pink tongue dangling lazily in a blissful grin. The entire scene is drenched in the warm, syrupy golden light of a perfect sunset, which catches the edges of the corgi's fur, causing it to glow and casting long, soft shadows across the lawn. Due to the shallow depth of field and the frantic camera movements, the background, composed of sharply outlined trees and a classic park bench in the distance, melts into a soft, painterly bokeh, creating a powerful sense of motion blur that isolates the dog and emphasizes the overwhelming atmosphere of playful, chaotic excitement.", + "seed": 217, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_217_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.006506382767111063, + "vlm_logprob": -0.006506382767111063, + "vlm_yes_no_margin": 5.325, + "vlm_token_credit": -0.006506382767111063, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0050241886638104916 + ], + "opposite_logprob_tokens": [ + -5.880023956298828 + ], + "logprob_reward": -0.0050241886638104916, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0050241886638104916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14835481345653534, + 0.10361457616090775, + 0.0636889785528183, + 0.020938348025083542, + 0.03305934742093086, + 0.016293132677674294, + 0.3835257887840271, + 0.02343318611383438, + 0.030028309673070908, + 0.17706353962421417 + ], + "credit_map": [ + 0.017329195514321327, + 0.014866804704070091, + 0.012401950545608997, + 0.012439832091331482, + 0.021226510405540466, + 0.03765809163451195, + 0.015057052485644817, + 0.01737535372376442, + 0.01425075437873602, + 0.008110706694424152, + 0.009356855414807796, + 0.01259523630142212, + 0.016818657517433167, + 0.008445671759545803, + 0.017591265961527824, + 0.016445431858301163, + 0.006273196078836918, + 0.006573250982910395, + 0.01349718775600195, + 0.005921098403632641, + 0.006969676818698645, + 0.006757070310413837, + 0.009423253126442432, + 0.008274238556623459, + 0.002887144684791565, + 0.002250988269224763, + 0.004179392009973526, + 0.0018044516909867525, + 0.002149521838873625, + 0.0020286727230995893, + 0.003826566506177187, + 0.001811608555726707, + 0.0052679660730063915, + 0.003080560127273202, + 0.005154011305421591, + 0.0029570695478469133, + 0.004924431908875704, + 0.003390852827578783, + 0.00599506264552474, + 0.0022893890272825956, + 0.002179475035518408, + 0.001202512881718576, + 0.0031012122053653, + 0.002125633880496025, + 0.0019299834966659546, + 0.0016100534703582525, + 0.0029416950419545174, + 0.0012025656178593636, + 0.06355199962854385, + 0.01994108222424984, + 0.05081984028220177, + 0.024352552369236946, + 0.02423069439828396, + 0.05371464416384697, + 0.03446110710501671, + 0.11245384812355042, + 0.0025559021160006523, + 0.0021849360782653093, + 0.0031789953354746103, + 0.0032022607047110796, + 0.003099840134382248, + 0.003021815326064825, + 0.003459310159087181, + 0.0027301234658807516, + 0.002626974368467927, + 0.004432506859302521, + 0.0034759133122861385, + 0.004461697302758694, + 0.004122598562389612, + 0.0035259686410427094, + 0.00458404328674078, + 0.0027986052446067333, + 0.017462050542235374, + 0.014623514376580715, + 0.03343111649155617, + 0.030271777883172035, + 0.016411736607551575, + 0.02469279058277607, + 0.012860388495028019, + 0.027310144156217575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0050241886638104916 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006741517689079046 + ], + "opposite_logprob_tokens": [ + -5.006741523742676 + ], + "logprob_reward": -0.006741517689079046, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006741517689079046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16735830903053284, + 0.10062538087368011, + 0.1167229637503624, + 0.053558193147182465, + 0.15865670144557953, + 0.0429210364818573, + 0.08047014474868774, + 0.060498617589473724, + 0.044527240097522736, + 0.17466147243976593 + ], + "credit_map": [ + 0.01663186214864254, + 0.016355717554688454, + 0.019265059381723404, + 0.018230972811579704, + 0.015855392441153526, + 0.034623607993125916, + 0.019873689860105515, + 0.026522010564804077, + 0.017589375376701355, + 0.005575826391577721, + 0.01648091711103916, + 0.007852141745388508, + 0.007294979877769947, + 0.012389210984110832, + 0.009925955906510353, + 0.023516971617937088, + 0.011185362935066223, + 0.019696343690156937, + 0.02538416162133217, + 0.011587873101234436, + 0.006590834353119135, + 0.014164702035486698, + 0.012011407874524593, + 0.016102280467748642, + 0.00685792276635766, + 0.00422385660931468, + 0.009792355820536613, + 0.007664106786251068, + 0.004587565083056688, + 0.007386159151792526, + 0.008367869071662426, + 0.004678358323872089, + 0.040904752910137177, + 0.009417438879609108, + 0.04019178822636604, + 0.009297143667936325, + 0.009612140245735645, + 0.013674548827111721, + 0.01917152851819992, + 0.016387352719902992, + 0.00690443767234683, + 0.0035141892731189728, + 0.007605561520904303, + 0.004041303880512714, + 0.004653565585613251, + 0.006747217383235693, + 0.006546595133841038, + 0.0029081653337925673, + 0.015150371938943863, + 0.0047372556291520596, + 0.010393333621323109, + 0.005582918878644705, + 0.0037795603275299072, + 0.010801857337355614, + 0.005312647670507431, + 0.02471219375729561, + 0.005980814807116985, + 0.006024785805493593, + 0.0060147191397845745, + 0.008943488821387291, + 0.008682645857334137, + 0.006294243969023228, + 0.011628855019807816, + 0.0069290641695261, + 0.003938097041100264, + 0.0037106992676854134, + 0.006209361366927624, + 0.009333939291536808, + 0.004277489148080349, + 0.004275167360901833, + 0.007943917065858841, + 0.004838570021092892, + 0.01648588851094246, + 0.013344059698283672, + 0.03969239071011543, + 0.024055562913417816, + 0.01854204759001732, + 0.021201495081186295, + 0.015045875683426857, + 0.026294149458408356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.006741517689079046 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004093363881111145 + ], + "opposite_logprob_tokens": [ + -5.504093170166016 + ], + "logprob_reward": -0.004093363881111145, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004093363881111145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12833434343338013, + 0.13654957711696625, + 0.15357574820518494, + 0.03751987963914871, + 0.1886366456747055, + 0.02878200076520443, + 0.0853770300745964, + 0.04205148667097092, + 0.0587538406252861, + 0.14041946828365326 + ], + "credit_map": [ + 0.013985421508550644, + 0.013117793016135693, + 0.01663976162672043, + 0.013852852396667004, + 0.01281148474663496, + 0.023026179522275925, + 0.017200645059347153, + 0.0177001953125, + 0.023060930892825127, + 0.007026056293398142, + 0.021791109815239906, + 0.010528573766350746, + 0.00844618771225214, + 0.011632460169494152, + 0.014689567498862743, + 0.039374690502882004, + 0.012097347527742386, + 0.005616175010800362, + 0.058567341417074203, + 0.00853908434510231, + 0.007261383347213268, + 0.009654088877141476, + 0.014252330176532269, + 0.03758800029754639, + 0.004898101091384888, + 0.004048799630254507, + 0.003777850419282913, + 0.0062213498167693615, + 0.003353959182277322, + 0.002620660001412034, + 0.009973797015845776, + 0.002625360619276762, + 0.02695540338754654, + 0.010834387503564358, + 0.04175623133778572, + 0.011228673160076141, + 0.01380636915564537, + 0.022238057106733322, + 0.019269272685050964, + 0.042548246681690216, + 0.005340998526662588, + 0.001589536084793508, + 0.003429036121815443, + 0.0027675898745656013, + 0.0026969981845468283, + 0.0069154370576143265, + 0.003998107276856899, + 0.0020442993845790625, + 0.014247380197048187, + 0.005977243650704622, + 0.011100249364972115, + 0.007687732577323914, + 0.006445456296205521, + 0.01019949559122324, + 0.007994179613888264, + 0.02172529138624668, + 0.004321510437875986, + 0.0032735911663621664, + 0.0037146429531276226, + 0.0065827276557683945, + 0.006027623079717159, + 0.004777745343744755, + 0.007329876534640789, + 0.006023772060871124, + 0.005899556912481785, + 0.007092087529599667, + 0.005091916304081678, + 0.011776183731853962, + 0.005236704368144274, + 0.007544366642832756, + 0.010012931190431118, + 0.00610009441152215, + 0.013227970339357853, + 0.010857727378606796, + 0.023971326649188995, + 0.027722466737031937, + 0.013225741684436798, + 0.018560098484158516, + 0.011290637776255608, + 0.021563494578003883 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 9 + ], + "selected_reward": -0.004093363881111145 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186132699251175 + ], + "opposite_logprob_tokens": [ + -4.264185905456543 + ], + "logprob_reward": -0.014186132699251175, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186132699251175, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1236523985862732, + 0.10079274326562881, + 0.07176157832145691, + 0.020156633108854294, + 0.053062133491039276, + 0.054169539362192154, + 0.16376139223575592, + 0.07414795458316803, + 0.03761117160320282, + 0.300884485244751 + ], + "credit_map": [ + 0.014188826084136963, + 0.012021836824715137, + 0.023072704672813416, + 0.016444632783532143, + 0.009981397539377213, + 0.013185156509280205, + 0.013694549910724163, + 0.021063288673758507, + 0.012369535863399506, + 0.008052594028413296, + 0.01683097518980503, + 0.013353975489735603, + 0.00953300204128027, + 0.01185582298785448, + 0.016256950795650482, + 0.012539886869490147, + 0.006466485559940338, + 0.004213430918753147, + 0.024002231657505035, + 0.004722229670733213, + 0.003575005568563938, + 0.007554141338914633, + 0.006917082238942385, + 0.014310969039797783, + 0.0030106918420642614, + 0.0018031132640317082, + 0.0039331139996647835, + 0.0015781153924763203, + 0.0012576376320794225, + 0.003011376829817891, + 0.003540770150721073, + 0.0020218128338456154, + 0.011267845518887043, + 0.004390887450426817, + 0.010519130155444145, + 0.003318019211292267, + 0.005484601017087698, + 0.006332576274871826, + 0.00641095545142889, + 0.005338120274245739, + 0.007226265035569668, + 0.0035720239393413067, + 0.013605213724076748, + 0.004704572726041079, + 0.004951746203005314, + 0.007589446846395731, + 0.0072869909927248955, + 0.005233276169747114, + 0.030803920701146126, + 0.007507703732699156, + 0.022678164765238762, + 0.011222045868635178, + 0.005837214644998312, + 0.02222539484500885, + 0.010968402959406376, + 0.05251852422952652, + 0.00782585795968771, + 0.005641425494104624, + 0.014425794593989849, + 0.010255981236696243, + 0.007702153641730547, + 0.008266397751867771, + 0.011236803606152534, + 0.00879353005439043, + 0.0033011247869580984, + 0.0033557198476046324, + 0.006080990191549063, + 0.007940595038235188, + 0.0027337330393493176, + 0.003863090416416526, + 0.0053137619979679585, + 0.005022154189646244, + 0.031925950199365616, + 0.022965090349316597, + 0.055543094873428345, + 0.04789262264966965, + 0.02965189516544342, + 0.046287838369607925, + 0.02057790569961071, + 0.04604005813598633 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.014186132699251175 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0024867109023034573 + ], + "opposite_logprob_tokens": [ + -6.002486705780029 + ], + "logprob_reward": -0.0024867109023034573, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0024867109023034573, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13922671973705292, + 0.16123341023921967, + 0.24018090963363647, + 0.025357432663440704, + 0.05437948554754257, + 0.019158266484737396, + 0.11856972426176071, + 0.08598241955041885, + 0.04180126264691353, + 0.11411043256521225 + ], + "credit_map": [ + 0.01367246825248003, + 0.00996499601751566, + 0.02261585183441639, + 0.013586697168648243, + 0.0129782035946846, + 0.02355439029633999, + 0.01599353924393654, + 0.026860563084483147, + 0.029063068330287933, + 0.007421659305691719, + 0.03010646253824234, + 0.011797726154327393, + 0.008728730492293835, + 0.014530186541378498, + 0.013097059912979603, + 0.04648851230740547, + 0.018786871805787086, + 0.014409106224775314, + 0.0761353075504303, + 0.01615286059677601, + 0.012121068313717842, + 0.026643967255949974, + 0.021255003288388252, + 0.05467672646045685, + 0.0033776925411075354, + 0.0028666446451097727, + 0.005215628072619438, + 0.002354029333218932, + 0.0018592802807688713, + 0.0034710709005594254, + 0.004331509582698345, + 0.0018815777730196714, + 0.012616611085832119, + 0.00324486056342721, + 0.014884118922054768, + 0.004257409833371639, + 0.004692687187343836, + 0.0033874970395118, + 0.006677341181784868, + 0.0046189576387405396, + 0.0028013961855322123, + 0.0018516030395403504, + 0.0030929865315556526, + 0.0022215978242456913, + 0.0027411384508013725, + 0.0025942709762603045, + 0.002303576795384288, + 0.0015516957500949502, + 0.016668301075696945, + 0.011670870706439018, + 0.012846886180341244, + 0.00904450099915266, + 0.014852466993033886, + 0.017413444817066193, + 0.018807366490364075, + 0.017265887930989265, + 0.009709599427878857, + 0.008133279159665108, + 0.010875753127038479, + 0.01248910091817379, + 0.01369090098887682, + 0.009131520055234432, + 0.01323743537068367, + 0.0087148267775774, + 0.003820803016424179, + 0.00330723961815238, + 0.004458144772797823, + 0.009118054993450642, + 0.0032936306670308113, + 0.003433008911088109, + 0.008301431313157082, + 0.006068945396691561, + 0.010935181751847267, + 0.012054342776536942, + 0.019891807809472084, + 0.017213713377714157, + 0.013885090127587318, + 0.01288458239287138, + 0.011381308548152447, + 0.015864398330450058 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0024867109023034573 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld medium shot struggles to keep pace, capturing an intense and visceral jolt of camera movement as it tracks a jubilant corgi with thick, reddish-brown and white fur and an unusually fluffy tail. The dog bounds across a large expanse of lush green grass, its short legs thrashing frantically in a blur of motion, its entire drooping body thrust forward with boundless energy. Its dark eyes are wide with pure joy, its pink tongue dangling lazily in a blissful grin. The entire scene is drenched in the warm, syrupy golden light of a perfect sunset, which catches the edges of the corgi's fur, causing it to glow and casting long, soft shadows across the lawn. Due to the shallow depth of field and the frantic camera movements, the background, composed of sharply outlined trees and a classic park bench in the distance, melts into a soft, painterly bokeh, creating a powerful sense of motion blur that isolates the dog and emphasizes the overwhelming atmosphere of playful, chaotic excitement.", + "seed": 218, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_218_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.004553353833034635, + "vlm_logprob": -0.004553353833034635, + "vlm_yes_no_margin": 5.525, + "vlm_token_credit": -0.004553353833034635, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004764394834637642 + ], + "opposite_logprob_tokens": [ + -5.879764556884766 + ], + "logprob_reward": -0.004764394834637642, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.004764394834637642, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3417869508266449, + 0.03884457051753998, + 0.05883440375328064, + 0.021490955725312233, + 0.05204574391245842, + 0.021673867478966713, + 0.3836793303489685, + 0.034463223069906235, + 0.02032736875116825, + 0.026853615418076515 + ], + "credit_map": [ + 0.03659254312515259, + 0.0340975858271122, + 0.030323892831802368, + 0.051846571266651154, + 0.05982092767953873, + 0.026187529787421227, + 0.06452321261167526, + 0.038394685834646225, + 0.004620617721229792, + 0.002920925384387374, + 0.0045599136501550674, + 0.0039045587182044983, + 0.00482203159481287, + 0.0039694644510746, + 0.005987123120576143, + 0.008059937506914139, + 0.008905008435249329, + 0.006446083541959524, + 0.005259584169834852, + 0.0068684909492731094, + 0.009696700610220432, + 0.005140804685652256, + 0.011915473267436028, + 0.00460225623100996, + 0.005441281478852034, + 0.0018777158111333847, + 0.0029936509672552347, + 0.0017491575563326478, + 0.0016872546402737498, + 0.0017869147704914212, + 0.0018696996849030256, + 0.004085279069840908, + 0.003166012931615114, + 0.010568708181381226, + 0.0032637869007885456, + 0.00999121367931366, + 0.007769772782921791, + 0.003976417705416679, + 0.009760147891938686, + 0.003549681743606925, + 0.0018470559734851122, + 0.003452003002166748, + 0.0019529237179085612, + 0.002421135315671563, + 0.004589158110320568, + 0.0028276056982576847, + 0.0022472701966762543, + 0.002336714416742325, + 0.04779384285211563, + 0.050917237997055054, + 0.0718768760561943, + 0.04593777656555176, + 0.03806734457612038, + 0.041485514491796494, + 0.03374776244163513, + 0.053852956742048264, + 0.0037862504832446575, + 0.004876615479588509, + 0.004161851946264505, + 0.004286861512809992, + 0.003645890625193715, + 0.003785084467381239, + 0.0054220762103796005, + 0.004498592112213373, + 0.0013844214845448732, + 0.002248234348371625, + 0.0022824096959084272, + 0.002271596109494567, + 0.002940951380878687, + 0.0029438105411827564, + 0.002796351909637451, + 0.003459594678133726, + 0.0024510440416634083, + 0.002878726227208972, + 0.0034303031861782074, + 0.0034466467332094908, + 0.00568417739123106, + 0.0026715381536632776, + 0.0037577010225504637, + 0.0025334791280329227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.004764394834637642 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005953910294920206 + ], + "opposite_logprob_tokens": [ + -5.130953788757324 + ], + "logprob_reward": -0.005953910294920206, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005953910294920206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2062923163175583, + 0.04783793166279793, + 0.03963750973343849, + 0.024957535788416862, + 0.0548090785741806, + 0.05420416221022606, + 0.42616140842437744, + 0.057994138449430466, + 0.019396940246224403, + 0.06870899349451065 + ], + "credit_map": [ + 0.024895884096622467, + 0.028126001358032227, + 0.018920818343758583, + 0.027670426294207573, + 0.03501380980014801, + 0.016008751466870308, + 0.03754039108753204, + 0.01811622828245163, + 0.007623941171914339, + 0.0035666590556502342, + 0.006790990941226482, + 0.0042990208603441715, + 0.004178480710834265, + 0.004370440728962421, + 0.00579784857109189, + 0.011210553348064423, + 0.005572204478085041, + 0.003847508691251278, + 0.00451029185205698, + 0.004310538060963154, + 0.006414404138922691, + 0.004474514164030552, + 0.0070410496555268764, + 0.003467000089585781, + 0.004746492952108383, + 0.002744799479842186, + 0.0032370297703891993, + 0.0019501391798257828, + 0.0032499271910637617, + 0.002699119271710515, + 0.002495666267350316, + 0.0038343598134815693, + 0.0037358547560870647, + 0.011283369734883308, + 0.0030793913174420595, + 0.010969234630465508, + 0.007832369767129421, + 0.0043834513053298, + 0.009195812977850437, + 0.004329593386501074, + 0.006578190717846155, + 0.006459397729486227, + 0.009780552238225937, + 0.006761523429304361, + 0.003671911545097828, + 0.008375943638384342, + 0.004631170071661472, + 0.007945475168526173, + 0.05126494541764259, + 0.04782261326909065, + 0.08584267646074295, + 0.05327223241329193, + 0.02958364598453045, + 0.06343536823987961, + 0.045539289712905884, + 0.04940063878893852, + 0.004085101652890444, + 0.00654305424541235, + 0.006916511338204145, + 0.006455583963543177, + 0.009958353824913502, + 0.010232760570943356, + 0.00857904739677906, + 0.0052237240597605705, + 0.0017617929261177778, + 0.0023286514915525913, + 0.00265984027646482, + 0.0027828840538859367, + 0.002292553661391139, + 0.002086671767756343, + 0.002570721320807934, + 0.0029138249810785055, + 0.0070020719431340694, + 0.005910767242312431, + 0.010190878063440323, + 0.01016810443252325, + 0.012322231195867062, + 0.005771024152636528, + 0.008727862499654293, + 0.00861604604870081 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 9 + ], + "selected_reward": -0.005953910294920206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00361420470289886 + ], + "opposite_logprob_tokens": [ + -5.62861442565918 + ], + "logprob_reward": -0.00361420470289886, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00361420470289886, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1992405205965042, + 0.09221233427524567, + 0.08156802505254745, + 0.05641896277666092, + 0.048696160316467285, + 0.04720002040266991, + 0.30223143100738525, + 0.05762296915054321, + 0.04452243074774742, + 0.07028708606958389 + ], + "credit_map": [ + 0.022973859682679176, + 0.024552199989557266, + 0.017595235258340836, + 0.027852632105350494, + 0.036546822637319565, + 0.018784377723932266, + 0.03251458704471588, + 0.018420806154608727, + 0.012130247429013252, + 0.007265559397637844, + 0.012465772219002247, + 0.008063595741987228, + 0.011441478505730629, + 0.009376219473779202, + 0.011959978379309177, + 0.019509484991431236, + 0.015024404972791672, + 0.009844229556620121, + 0.008016320876777172, + 0.007101364899426699, + 0.008710803464055061, + 0.011412409134209156, + 0.013248942792415619, + 0.00820954516530037, + 0.00937728676944971, + 0.008496006950736046, + 0.006336966995149851, + 0.0056609404273331165, + 0.007384733762592077, + 0.005713229067623615, + 0.00776192732155323, + 0.00568787194788456, + 0.003183672670274973, + 0.007433378137648106, + 0.0032095001079142094, + 0.006835834123194218, + 0.008943683467805386, + 0.0047085885889828205, + 0.009831996634602547, + 0.004549502860754728, + 0.0039902422577142715, + 0.0069790007546544075, + 0.005394014064222574, + 0.007424592971801758, + 0.006823970470577478, + 0.005382019095122814, + 0.005559721030294895, + 0.0056464592926204205, + 0.03551676869392395, + 0.035623759031295776, + 0.04805827885866165, + 0.03221261873841286, + 0.035251785069704056, + 0.03511097654700279, + 0.03194743022322655, + 0.04850982129573822, + 0.006402468308806419, + 0.007128359749913216, + 0.0064094094559550285, + 0.006963376421481371, + 0.0072483886033296585, + 0.008310867473483086, + 0.009443291462957859, + 0.005716804880648851, + 0.003798494581133127, + 0.005695731844753027, + 0.007439565844833851, + 0.00627489946782589, + 0.003469429211691022, + 0.007539655547589064, + 0.003909985069185495, + 0.006394668947905302, + 0.007547716144472361, + 0.007004079408943653, + 0.010041504167020321, + 0.011435863561928272, + 0.012048612348735332, + 0.006779268383979797, + 0.009083862416446209, + 0.006346179638057947 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.00361420470289886 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005245612468570471 + ], + "opposite_logprob_tokens": [ + -5.255245685577393 + ], + "logprob_reward": -0.005245612468570471, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005245612468570471, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16192175447940826, + 0.27670133113861084, + 0.05948911979794502, + 0.13541312515735626, + 0.024340448901057243, + 0.1465177834033966, + 0.05154354125261307, + 0.09553642570972443, + 0.02198672853410244, + 0.02654976211488247 + ], + "credit_map": [ + 0.025874720886349678, + 0.023814693093299866, + 0.02480452135205269, + 0.010141092352569103, + 0.012098192237317562, + 0.01855042204260826, + 0.011475577019155025, + 0.0351625420153141, + 0.06178147718310356, + 0.02726723439991474, + 0.0369308777153492, + 0.01244760025292635, + 0.01252749189734459, + 0.025439368560910225, + 0.019771549850702286, + 0.08053571730852127, + 0.009113829582929611, + 0.005404244642704725, + 0.007268986199051142, + 0.005930986255407333, + 0.00501710269600153, + 0.013382119126617908, + 0.0072268713265657425, + 0.006144980899989605, + 0.033297665417194366, + 0.00921043660491705, + 0.020675482228398323, + 0.009507532231509686, + 0.01221948117017746, + 0.012620625086128712, + 0.010900544933974743, + 0.026981353759765625, + 0.0015955453272908926, + 0.004770335275679827, + 0.0019288401817902923, + 0.0043668900616467, + 0.0029003636445850134, + 0.0021500764414668083, + 0.00394859304651618, + 0.00267980364151299, + 0.012283442541956902, + 0.02133217453956604, + 0.013263718225061893, + 0.016579553484916687, + 0.017959224060177803, + 0.033680036664009094, + 0.014837541617453098, + 0.016582094132900238, + 0.0044122072868049145, + 0.006342313718050718, + 0.005271071568131447, + 0.007530814968049526, + 0.007420909591019154, + 0.006933167576789856, + 0.00738060288131237, + 0.006252451799809933, + 0.010751106776297092, + 0.009171969257295132, + 0.010792153887450695, + 0.009799649938941002, + 0.0109053710475564, + 0.01629730872809887, + 0.012288474477827549, + 0.01553038414567709, + 0.0018406468443572521, + 0.0013387621147558093, + 0.006420824211090803, + 0.0028111906722187996, + 0.002097403397783637, + 0.0019355894764885306, + 0.0022496823221445084, + 0.003292629960924387, + 0.0028189406730234623, + 0.0021481665316969156, + 0.004965904168784618, + 0.002789965597912669, + 0.004341918975114822, + 0.0027868638280779123, + 0.003279782133176923, + 0.0034182218369096518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.005245612468570471 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003188646864145994 + ], + "opposite_logprob_tokens": [ + -5.753188610076904 + ], + "logprob_reward": -0.003188646864145994, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003188646864145994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12225153297185898, + 0.053832318633794785, + 0.04663238301873207, + 0.03197769820690155, + 0.051325127482414246, + 0.05230285972356796, + 0.38797277212142944, + 0.15071037411689758, + 0.04336661472916603, + 0.05962829291820526 + ], + "credit_map": [ + 0.018355712294578552, + 0.01445862278342247, + 0.013136985711753368, + 0.015901364386081696, + 0.017743468284606934, + 0.012274041771888733, + 0.01618305593729019, + 0.014198286458849907, + 0.0072313095442950726, + 0.0040075769647955894, + 0.010137940756976604, + 0.004997551906853914, + 0.004888308234512806, + 0.009234174154698849, + 0.0062042647041380405, + 0.007131190970540047, + 0.007616573944687843, + 0.005650546867400408, + 0.0035432002041488886, + 0.005780071951448917, + 0.005107013508677483, + 0.0053170220926404, + 0.007389334961771965, + 0.006228622980415821, + 0.004245620220899582, + 0.0059023043140769005, + 0.0036956898402422667, + 0.003549903631210327, + 0.004209335893392563, + 0.00282790488563478, + 0.005229525733739138, + 0.0023174139205366373, + 0.003668983234092593, + 0.007443350739777088, + 0.004226509016007185, + 0.009198131039738655, + 0.006406495813280344, + 0.004879154730588198, + 0.008413956500589848, + 0.007088546175509691, + 0.0062946449033916, + 0.008159438148140907, + 0.0076240007765591145, + 0.007378363516181707, + 0.005311977118253708, + 0.0055169309489429, + 0.005851836409419775, + 0.00616566464304924, + 0.0440804623067379, + 0.045237429440021515, + 0.07051942497491837, + 0.053415268659591675, + 0.034749384969472885, + 0.05660303309559822, + 0.03101910650730133, + 0.052348680794239044, + 0.017070822417736053, + 0.013201619498431683, + 0.017845172435045242, + 0.014750576578080654, + 0.015899939462542534, + 0.023170458152890205, + 0.02240525186061859, + 0.02636653371155262, + 0.004047904163599014, + 0.005246067885309458, + 0.0058941831812262535, + 0.007337583228945732, + 0.006776781752705574, + 0.003985933493822813, + 0.005452175624668598, + 0.004625988192856312, + 0.0067387530580163, + 0.005429613403975964, + 0.009326937608420849, + 0.006630955263972282, + 0.008182520978152752, + 0.007341461721807718, + 0.0072361682541668415, + 0.008741880767047405 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 0 + ], + "selected_reward": -0.003188646864145994 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld medium shot struggles to keep pace, capturing an intense and visceral jolt of camera movement as it tracks a jubilant corgi with thick, reddish-brown and white fur and an unusually fluffy tail. The dog bounds across a large expanse of lush green grass, its short legs thrashing frantically in a blur of motion, its entire drooping body thrust forward with boundless energy. Its dark eyes are wide with pure joy, its pink tongue dangling lazily in a blissful grin. The entire scene is drenched in the warm, syrupy golden light of a perfect sunset, which catches the edges of the corgi's fur, causing it to glow and casting long, soft shadows across the lawn. Due to the shallow depth of field and the frantic camera movements, the background, composed of sharply outlined trees and a classic park bench in the distance, melts into a soft, painterly bokeh, creating a powerful sense of motion blur that isolates the dog and emphasizes the overwhelming atmosphere of playful, chaotic excitement.", + "seed": 219, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_219_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.012916162703186273, + "vlm_logprob": -0.012916162703186273, + "vlm_yes_no_margin": 4.875, + "vlm_token_credit": -0.012916162703186273, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006114111747592688 + ], + "opposite_logprob_tokens": [ + -6.0061140060424805 + ], + "logprob_reward": -0.006114111747592688, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.006114111747592688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20973864197731018, + 0.02688051015138626, + 0.15376590192317963, + 0.05435067042708397, + 0.11372777074575424, + 0.03333583474159241, + 0.0997093990445137, + 0.12389542162418365, + 0.0648680329322815, + 0.11972784250974655 + ], + "credit_map": [ + 0.01506530400365591, + 0.00789994839578867, + 0.07201658189296722, + 0.019007328897714615, + 0.009978706948459148, + 0.04109318554401398, + 0.014945040456950665, + 0.029732542112469673, + 0.004104249179363251, + 0.003518231911584735, + 0.003777226433157921, + 0.002185428747907281, + 0.002718142233788967, + 0.0025538019835948944, + 0.0037608109414577484, + 0.004262619651854038, + 0.01190302986651659, + 0.010409212671220303, + 0.017819741740822792, + 0.014392643235623837, + 0.0286091398447752, + 0.014311632141470909, + 0.029307255521416664, + 0.02701326459646225, + 0.004686617758125067, + 0.0039134034886956215, + 0.011937274597585201, + 0.00909106433391571, + 0.004695841111242771, + 0.005215891636908054, + 0.004947020206600428, + 0.009863563813269138, + 0.011229502968490124, + 0.006471419241279364, + 0.006694144103676081, + 0.013653596863150597, + 0.020756609737873077, + 0.011138084344565868, + 0.035667747259140015, + 0.008116672746837139, + 0.003033445216715336, + 0.00276700546965003, + 0.0037631269078701735, + 0.004118985962122679, + 0.004573505837470293, + 0.004037816543132067, + 0.007000361569225788, + 0.004041587468236685, + 0.011630124412477016, + 0.011463954113423824, + 0.008000161498785019, + 0.009463113732635975, + 0.01881241239607334, + 0.009318690747022629, + 0.02018412947654724, + 0.010836821049451828, + 0.010803325101733208, + 0.008584241382777691, + 0.01609441451728344, + 0.01671033725142479, + 0.024712206795811653, + 0.010088562965393066, + 0.023535437881946564, + 0.01336690317839384, + 0.00504304887726903, + 0.004795073065906763, + 0.006909527815878391, + 0.005901404656469822, + 0.014481717720627785, + 0.004088557790964842, + 0.01255104597657919, + 0.011097665876150131, + 0.009599484503269196, + 0.005593536421656609, + 0.017058875411748886, + 0.01007128320634365, + 0.013685609214007854, + 0.010374050587415695, + 0.025321342051029205, + 0.028023667633533478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.006114111747592688 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03810625150799751 + ], + "opposite_logprob_tokens": [ + -3.2881062030792236 + ], + "logprob_reward": -0.03810625150799751, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03810625150799751, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31076088547706604, + 0.028933150693774223, + 0.049767449498176575, + 0.08372687548398972, + 0.11898207664489746, + 0.035202283412218094, + 0.07281816750764847, + 0.024649132043123245, + 0.19352620840072632, + 0.0816337913274765 + ], + "credit_map": [ + 0.021827712655067444, + 0.007781424559652805, + 0.12421825528144836, + 0.028731388971209526, + 0.011023788712918758, + 0.0723634585738182, + 0.01873975805938244, + 0.026075104251503944, + 0.004070062655955553, + 0.002629165770485997, + 0.004050104878842831, + 0.0034238710068166256, + 0.003973593935370445, + 0.0028794805984944105, + 0.004064488224685192, + 0.003842382226139307, + 0.004376273602247238, + 0.003742576576769352, + 0.004999435506761074, + 0.005896405782550573, + 0.009784048423171043, + 0.0036630600225180387, + 0.01162093598395586, + 0.005684711970388889, + 0.004858928266912699, + 0.006439965218305588, + 0.02114821784198284, + 0.021242663264274597, + 0.002696123905479908, + 0.009295147843658924, + 0.00459758285433054, + 0.013448238372802734, + 0.012716909870505333, + 0.007093899417668581, + 0.009398485533893108, + 0.013127312995493412, + 0.018397707492113113, + 0.00984915904700756, + 0.03745685890316963, + 0.010941746644675732, + 0.00472285458818078, + 0.0034542283974587917, + 0.004801080096513033, + 0.004635518416762352, + 0.005145795177668333, + 0.0040136706084012985, + 0.004351346753537655, + 0.004077790770679712, + 0.008196202106773853, + 0.010096713900566101, + 0.006642682943493128, + 0.0044034067541360855, + 0.01678956300020218, + 0.007723214104771614, + 0.013688476756215096, + 0.005277904681861401, + 0.002634042175486684, + 0.0018271426670253277, + 0.003116438863798976, + 0.002870141761377454, + 0.004871524404734373, + 0.001713322359137237, + 0.004051687195897102, + 0.0035648320335894823, + 0.022816678509116173, + 0.007109710946679115, + 0.034276124089956284, + 0.013771215453743935, + 0.012581009417772293, + 0.01794613152742386, + 0.01802947372198105, + 0.06699586659669876, + 0.0069868662394583225, + 0.00443659583106637, + 0.011078366078436375, + 0.01604045182466507, + 0.008176978677511215, + 0.0059206546284258366, + 0.016379566863179207, + 0.012614313513040543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 4 + ], + "selected_reward": -0.03810625150799751 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005947985220700502 + ], + "opposite_logprob_tokens": [ + -5.130948066711426 + ], + "logprob_reward": -0.005947985220700502, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005947985220700502, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3079030215740204, + 0.07246091216802597, + 0.11209829151630402, + 0.11345123499631882, + 0.05101614072918892, + 0.03577881678938866, + 0.05700886994600296, + 0.024389291182160378, + 0.09031935781240463, + 0.13557404279708862 + ], + "credit_map": [ + 0.021929137408733368, + 0.008718528784811497, + 0.12332287430763245, + 0.02900772914290428, + 0.011448979377746582, + 0.07446586340665817, + 0.01816844753921032, + 0.020841451361775398, + 0.007084713317453861, + 0.01013891026377678, + 0.012406371533870697, + 0.007686855271458626, + 0.008109797723591328, + 0.006957418285310268, + 0.011964117176830769, + 0.008112728595733643, + 0.009524677880108356, + 0.008280089125037193, + 0.011878084391355515, + 0.011591777205467224, + 0.02260635606944561, + 0.009522910229861736, + 0.02579360269010067, + 0.012900792062282562, + 0.008509921841323376, + 0.0059259794652462006, + 0.02734879031777382, + 0.0196785107254982, + 0.006865644361823797, + 0.014135264791548252, + 0.008718088269233704, + 0.022269032895565033, + 0.006578665226697922, + 0.003860585391521454, + 0.007215936202555895, + 0.003026583231985569, + 0.00698480848222971, + 0.004647800233215094, + 0.010496430099010468, + 0.008205332793295383, + 0.003133931430056691, + 0.003494161181151867, + 0.004988097120076418, + 0.004003910813480616, + 0.005294270813465118, + 0.004534924402832985, + 0.005185279995203018, + 0.005144239403307438, + 0.006188573315739632, + 0.006962121464312077, + 0.004752172157168388, + 0.004417336545884609, + 0.013197774067521095, + 0.006485636346042156, + 0.009175579994916916, + 0.005829679314047098, + 0.002951201982796192, + 0.003280765376985073, + 0.0034358822740614414, + 0.002081802813336253, + 0.004658491350710392, + 0.0020556768868118525, + 0.0032338304445147514, + 0.002691639121621847, + 0.008668890222907066, + 0.0038064823020249605, + 0.013908515684306622, + 0.006956769619137049, + 0.010755575262010098, + 0.007516384590417147, + 0.010936101898550987, + 0.027770640328526497, + 0.011448599398136139, + 0.008294551633298397, + 0.02073066681623459, + 0.03457821160554886, + 0.011169335804879665, + 0.007868200540542603, + 0.024816397577524185, + 0.016668079420924187 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 3 + ], + "selected_reward": -0.005947985220700502 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009776449762284756 + ], + "opposite_logprob_tokens": [ + -4.634776592254639 + ], + "logprob_reward": -0.009776449762284756, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009776449762284756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20390169322490692, + 0.05267399922013283, + 0.10986632853746414, + 0.16340592503547668, + 0.07163497060537338, + 0.15525001287460327, + 0.06874937564134598, + 0.08260583132505417, + 0.021326402202248573, + 0.0705854743719101 + ], + "credit_map": [ + 0.019582334905862808, + 0.020525388419628143, + 0.04545015096664429, + 0.015572892501950264, + 0.014873956330120564, + 0.032352667301893234, + 0.023160379379987717, + 0.03238394111394882, + 0.00820386502891779, + 0.005914165172725916, + 0.011148353107273579, + 0.005314663518220186, + 0.004810676444321871, + 0.004208484664559364, + 0.005928372032940388, + 0.007145421113818884, + 0.01976391300559044, + 0.021442843601107597, + 0.01273021474480629, + 0.016813909634947777, + 0.008249732665717602, + 0.009753258898854256, + 0.00698521314188838, + 0.01412724144756794, + 0.012847852893173695, + 0.01552444975823164, + 0.03365210071206093, + 0.03670762479305267, + 0.007118727546185255, + 0.02042645774781704, + 0.009931573644280434, + 0.027197133749723434, + 0.010764962062239647, + 0.005298301111906767, + 0.012480000965297222, + 0.005710348021239042, + 0.008366082794964314, + 0.00794824119657278, + 0.011642391793429852, + 0.00942464079707861, + 0.021108146756887436, + 0.007753612473607063, + 0.02568173035979271, + 0.01019863598048687, + 0.012641609646379948, + 0.030929455533623695, + 0.013378389179706573, + 0.033558439463377, + 0.01241981890052557, + 0.007313431240618229, + 0.014212662354111671, + 0.006636352278292179, + 0.005876641254872084, + 0.009423759765923023, + 0.005737023428082466, + 0.007129690609872341, + 0.014292391948401928, + 0.007512698881328106, + 0.009623515419661999, + 0.005911892279982567, + 0.013569459319114685, + 0.009343626908957958, + 0.008893272839486599, + 0.013458971865475178, + 0.0027712334413081408, + 0.002205832628533244, + 0.0023983202408999205, + 0.0021206072997301817, + 0.0031648927833884954, + 0.0016578068025410175, + 0.0043808347545564175, + 0.002626874251291156, + 0.005756111815571785, + 0.006602115463465452, + 0.011152569204568863, + 0.018668869510293007, + 0.004096183925867081, + 0.004254861269146204, + 0.00989337358623743, + 0.010161389596760273 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.009776449762284756 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004636015277355909 + ], + "opposite_logprob_tokens": [ + -5.379635810852051 + ], + "logprob_reward": -0.004636015277355909, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004636015277355909, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.419262558221817, + 0.03863126039505005, + 0.0431443527340889, + 0.052813418209552765, + 0.06583914160728455, + 0.04176080599427223, + 0.1622048020362854, + 0.037124332040548325, + 0.048851024359464645, + 0.0903683453798294 + ], + "credit_map": [ + 0.029970882460474968, + 0.010960455052554607, + 0.16998888552188873, + 0.03975911810994148, + 0.014005533419549465, + 0.10085643827915192, + 0.02044776827096939, + 0.03327345848083496, + 0.006906431168317795, + 0.00472905021160841, + 0.005600200966000557, + 0.003494712756946683, + 0.0028404404874891043, + 0.0032505979761481285, + 0.004729247651994228, + 0.0070805782452225685, + 0.003664141520857811, + 0.0028466011863201857, + 0.0052023776806890965, + 0.006304600276052952, + 0.005862123798578978, + 0.0031158672645688057, + 0.006773373577743769, + 0.009375263005495071, + 0.004658610559999943, + 0.004364820662885904, + 0.0073136622086167336, + 0.0076538496650755405, + 0.005154433194547892, + 0.006536567118018866, + 0.006176384165883064, + 0.010955087840557098, + 0.006278498098254204, + 0.004893204662948847, + 0.007117561064660549, + 0.005759427323937416, + 0.009721503593027592, + 0.006102750543504953, + 0.01825683005154133, + 0.007709364872425795, + 0.00397082231938839, + 0.002965807681903243, + 0.004481826908886433, + 0.004341993015259504, + 0.006082719657570124, + 0.006361272651702166, + 0.007301107980310917, + 0.006255252752453089, + 0.01957976073026657, + 0.019479863345623016, + 0.014297009445726871, + 0.011376784183084965, + 0.0374685674905777, + 0.01762397587299347, + 0.024584710597991943, + 0.017794111743569374, + 0.006977067328989506, + 0.004066356457769871, + 0.005520783830434084, + 0.003873276524245739, + 0.0041028778068721294, + 0.003422742709517479, + 0.0032212960068136454, + 0.005939928349107504, + 0.005883429199457169, + 0.003234107978641987, + 0.007372663822025061, + 0.0032440894283354282, + 0.004395205527544022, + 0.004801526665687561, + 0.004774189554154873, + 0.015145807527005672, + 0.010347038507461548, + 0.007703831419348717, + 0.015321440063416958, + 0.02491258643567562, + 0.00518306577578187, + 0.005719470325857401, + 0.009689174592494965, + 0.011491726152598858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 9 + ], + "selected_reward": -0.004636015277355909 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, handheld medium shot struggles to keep pace, capturing an intense and visceral jolt of camera movement as it tracks a jubilant corgi with thick, reddish-brown and white fur and an unusually fluffy tail. The dog bounds across a large expanse of lush green grass, its short legs thrashing frantically in a blur of motion, its entire drooping body thrust forward with boundless energy. Its dark eyes are wide with pure joy, its pink tongue dangling lazily in a blissful grin. The entire scene is drenched in the warm, syrupy golden light of a perfect sunset, which catches the edges of the corgi's fur, causing it to glow and casting long, soft shadows across the lawn. Due to the shallow depth of field and the frantic camera movements, the background, composed of sharply outlined trees and a classic park bench in the distance, melts into a soft, painterly bokeh, creating a powerful sense of motion blur that isolates the dog and emphasizes the overwhelming atmosphere of playful, chaotic excitement.", + "seed": 220, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_7/index_4642_seed_220_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.006576850870624184, + "vlm_logprob": -0.006576850870624184, + "vlm_yes_no_margin": 5.475, + "vlm_token_credit": -0.006576850870624184, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004664255306124687 + ], + "opposite_logprob_tokens": [ + -6.879664421081543 + ], + "logprob_reward": -0.004664255306124687, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.004664255306124687, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15615756809711456, + 0.10277881473302841, + 0.1054350957274437, + 0.04884128272533417, + 0.08379890024662018, + 0.07115621119737625, + 0.04465821385383606, + 0.18589891493320465, + 0.08745226263999939, + 0.11382277309894562 + ], + "credit_map": [ + 0.018615134060382843, + 0.016696002334356308, + 0.010546218603849411, + 0.018625719472765923, + 0.020560165867209435, + 0.0289923045784235, + 0.018542736768722534, + 0.02357928827404976, + 0.007978403009474277, + 0.009777012281119823, + 0.011093913577497005, + 0.014117971993982792, + 0.01201265025883913, + 0.01787681132555008, + 0.009403872303664684, + 0.02051817625761032, + 0.01605343446135521, + 0.0080854631960392, + 0.011281936429440975, + 0.012296779081225395, + 0.009075579233467579, + 0.008113191463053226, + 0.01943403296172619, + 0.021094679832458496, + 0.006269567646086216, + 0.005567896645516157, + 0.00567169114947319, + 0.005980880465358496, + 0.006056061014533043, + 0.006108838599175215, + 0.006046590395271778, + 0.007139755878597498, + 0.00838437955826521, + 0.016410941258072853, + 0.010051898658275604, + 0.008544757962226868, + 0.013095189817249775, + 0.010470480658113956, + 0.00841206032782793, + 0.008429194800555706, + 0.006871468387544155, + 0.01374457124620676, + 0.007464730646461248, + 0.008027367293834686, + 0.007847399450838566, + 0.00883469544351101, + 0.008626092225313187, + 0.009739886969327927, + 0.006170213688164949, + 0.005347151774913073, + 0.004612465389072895, + 0.005598451942205429, + 0.0052718729712069035, + 0.005515063181519508, + 0.007016744464635849, + 0.005126249510794878, + 0.019916467368602753, + 0.022089052945375443, + 0.016847942024469376, + 0.01714550144970417, + 0.01670924574136734, + 0.036867406219244, + 0.01741870678961277, + 0.038904592394828796, + 0.009790086187422276, + 0.00958429928869009, + 0.010244354605674744, + 0.009702078998088837, + 0.009604387916624546, + 0.009356888942420483, + 0.017418794333934784, + 0.011751366779208183, + 0.011604592204093933, + 0.010205848142504692, + 0.025061776861548424, + 0.017581447958946228, + 0.011472773738205433, + 0.018228767439723015, + 0.010217106901109219, + 0.009450460784137249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.004664255306124687 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004106423351913691 + ], + "opposite_logprob_tokens": [ + -5.504106521606445 + ], + "logprob_reward": -0.004106423351913691, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004106423351913691, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09289077669382095, + 0.1813398152589798, + 0.045425623655319214, + 0.02853553369641304, + 0.055958785116672516, + 0.09443346410989761, + 0.02794022113084793, + 0.40375784039497375, + 0.03078365884721279, + 0.03893435746431351 + ], + "credit_map": [ + 0.008968676440417767, + 0.012657573446631432, + 0.007937191985547543, + 0.014403175562620163, + 0.011115516535937786, + 0.009047654457390308, + 0.008972628973424435, + 0.01978834718465805, + 0.017260728403925896, + 0.016550038009881973, + 0.016727788373827934, + 0.015903256833553314, + 0.02381538413465023, + 0.03689572960138321, + 0.01565137505531311, + 0.03853548690676689, + 0.0054382565431296825, + 0.00471260491758585, + 0.005230850540101528, + 0.008228864520788193, + 0.005268332082778215, + 0.0034936454612761736, + 0.005740765482187271, + 0.007312295958399773, + 0.0031779445707798004, + 0.002688421169295907, + 0.004321702755987644, + 0.0036772999446839094, + 0.002775824861600995, + 0.002682683989405632, + 0.0033822187688201666, + 0.0058294362388551235, + 0.005484646186232567, + 0.006878698710352182, + 0.007913390174508095, + 0.00712259067222476, + 0.006606134120374918, + 0.006757806055247784, + 0.008404728956520557, + 0.006790786515921354, + 0.008223196491599083, + 0.012857950292527676, + 0.011134807951748371, + 0.012987563386559486, + 0.010157039389014244, + 0.010548359714448452, + 0.01280275359749794, + 0.01572178490459919, + 0.002336832694709301, + 0.0034696327056735754, + 0.003088802332058549, + 0.0039734807796776295, + 0.0029410417191684246, + 0.0030658207833766937, + 0.005341504234820604, + 0.003723104717209935, + 0.049719639122486115, + 0.01209242083132267, + 0.020113153383135796, + 0.02803417667746544, + 0.014655370265245438, + 0.1251317411661148, + 0.022203831002116203, + 0.13180746138095856, + 0.004081724677234888, + 0.00410262169316411, + 0.004498204682022333, + 0.004029428120702505, + 0.0026339751202613115, + 0.003184839617460966, + 0.00461639603599906, + 0.003636464476585388, + 0.004726751707494259, + 0.004399772733449936, + 0.005789055023342371, + 0.005674649029970169, + 0.004184012766927481, + 0.005266302265226841, + 0.003909169230610132, + 0.004984640050679445 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 1, + 5 + ], + "selected_reward": -0.004106423351913691 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006737846881151199 + ], + "opposite_logprob_tokens": [ + -5.00673770904541 + ], + "logprob_reward": -0.006737846881151199, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006737846881151199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1534925103187561, + 0.18005824089050293, + 0.05071187764406204, + 0.06800538301467896, + 0.1265939623117447, + 0.16165277361869812, + 0.05299389362335205, + 0.1380566507577896, + 0.04360625892877579, + 0.024828368797898293 + ], + "credit_map": [ + 0.018071262165904045, + 0.027548210695385933, + 0.018300380557775497, + 0.025857800617814064, + 0.01744294911623001, + 0.013791807927191257, + 0.014964026398956776, + 0.01751607470214367, + 0.031310807913541794, + 0.014348423108458519, + 0.013805918395519257, + 0.013482153415679932, + 0.026211563497781754, + 0.03138820454478264, + 0.016303740441799164, + 0.033207423985004425, + 0.006006735377013683, + 0.005629854276776314, + 0.006688193418085575, + 0.010026159696280956, + 0.005327295046299696, + 0.005391513463109732, + 0.006003931164741516, + 0.005638195667415857, + 0.00792800635099411, + 0.007844899781048298, + 0.009958028793334961, + 0.007607513573020697, + 0.007430076599121094, + 0.007064328994601965, + 0.009596308693289757, + 0.010576223954558372, + 0.012897513806819916, + 0.019205367192626, + 0.013137185014784336, + 0.013977931812405586, + 0.02120986022055149, + 0.017347605898976326, + 0.014773582108318806, + 0.014044921845197678, + 0.02179788425564766, + 0.031944166868925095, + 0.015177861787378788, + 0.020631935447454453, + 0.016616353765130043, + 0.020191114395856857, + 0.02037840336561203, + 0.014915046282112598, + 0.006941159721463919, + 0.007008896209299564, + 0.005048887804150581, + 0.0058450703509151936, + 0.007382015231996775, + 0.007015030365437269, + 0.008176598697900772, + 0.005576234310865402, + 0.015108606778085232, + 0.012930993922054768, + 0.014628506265580654, + 0.017671266570687294, + 0.018981222063302994, + 0.015876514837145805, + 0.02532018907368183, + 0.017539355903863907, + 0.006077977828681469, + 0.00714652705937624, + 0.00413864478468895, + 0.005635795649141073, + 0.004252232611179352, + 0.005191942676901817, + 0.006644845474511385, + 0.004518295172601938, + 0.003212284529581666, + 0.0028027528896927834, + 0.0027960590086877346, + 0.004358608275651932, + 0.003313853871077299, + 0.003116865176707506, + 0.002893264638260007, + 0.0023346806410700083 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 5, + 0 + ], + "selected_reward": -0.006737846881151199 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014186250045895576 + ], + "opposite_logprob_tokens": [ + -4.264186382293701 + ], + "logprob_reward": -0.014186250045895576, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014186250045895576, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14569967985153198, + 0.24608467519283295, + 0.04921313002705574, + 0.03890125826001167, + 0.10483286529779434, + 0.0667484700679779, + 0.03451784700155258, + 0.2156551480293274, + 0.017753418534994125, + 0.0805935487151146 + ], + "credit_map": [ + 0.0159553661942482, + 0.015909479930996895, + 0.015155320055782795, + 0.02317243069410324, + 0.022104980424046516, + 0.02336551621556282, + 0.016092345118522644, + 0.013944230042397976, + 0.027234183624386787, + 0.016651583835482597, + 0.021870940923690796, + 0.02149362862110138, + 0.03496958687901497, + 0.04925335943698883, + 0.02814849093556404, + 0.046462878584861755, + 0.006814866326749325, + 0.0059035057201981544, + 0.008946417830884457, + 0.007239023223519325, + 0.003491046139970422, + 0.00452579278498888, + 0.005975152365863323, + 0.006317321211099625, + 0.005410370882600546, + 0.003169658361002803, + 0.0058249738067388535, + 0.00524061219766736, + 0.0034014068078249693, + 0.005708691198378801, + 0.004488884937018156, + 0.005656655412167311, + 0.011211362667381763, + 0.014744032174348831, + 0.015004406683146954, + 0.015257788822054863, + 0.012532508000731468, + 0.012178698554635048, + 0.010428676381707191, + 0.0134753892198205, + 0.006815924309194088, + 0.009819938801229, + 0.007527492940425873, + 0.008870823308825493, + 0.008850689977407455, + 0.007961939089000225, + 0.0075906142592430115, + 0.009311041794717312, + 0.0049916463904082775, + 0.00389603222720325, + 0.0035227928310632706, + 0.004502321593463421, + 0.0036747034173458815, + 0.004363141022622585, + 0.004845517221838236, + 0.00472169229760766, + 0.02721877582371235, + 0.007668320555239916, + 0.011723976582288742, + 0.015746764838695526, + 0.009159337729215622, + 0.06384052336215973, + 0.012967283837497234, + 0.06733015924692154, + 0.0026811829302459955, + 0.0029072058387100697, + 0.0015657258918508887, + 0.0022904500365257263, + 0.0017669507069513202, + 0.0019555387552827597, + 0.002756566507741809, + 0.0018297962378710508, + 0.007895873859524727, + 0.007102119270712137, + 0.019998542964458466, + 0.011104543693363667, + 0.007047711871564388, + 0.012069921009242535, + 0.00815624464303255, + 0.007218590471893549 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.014186250045895576 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a corgi with thick reddish-brown and white fur and an unusually fluffy tail?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the corgi bounding across a large expanse of lush green grass with its short legs thrashing in a blur of motion?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the corgi display wide dark eyes, a pink tongue dangling in a grin, and a drooping body thrust forward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the scene illuminated by warm golden sunset light that causes the corgi's fur to glow and casts long soft shadows?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot low-angle and handheld with frantic movement, creating motion blur and a soft painterly bokeh background with trees and a park bench?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031894787680357695 + ], + "opposite_logprob_tokens": [ + -5.753189563751221 + ], + "logprob_reward": -0.0031894787680357695, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031894787680357695, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06719041615724564, + 0.32060760259628296, + 0.04195832461118698, + 0.03278280794620514, + 0.03408932313323021, + 0.051929131150245667, + 0.03157854452729225, + 0.3026067018508911, + 0.06052369251847267, + 0.05673350766301155 + ], + "credit_map": [ + 0.008830013684928417, + 0.007445383816957474, + 0.005890585482120514, + 0.007819118909537792, + 0.00783200841397047, + 0.006322604604065418, + 0.007716817315667868, + 0.015333879739046097, + 0.0354212261736393, + 0.019849572330713272, + 0.02444014698266983, + 0.029806306585669518, + 0.041355498135089874, + 0.08106895536184311, + 0.02990197204053402, + 0.05876390263438225, + 0.006214234046638012, + 0.004302695859223604, + 0.0057675945572555065, + 0.007179876323789358, + 0.00312978052534163, + 0.0030489182099699974, + 0.008164484985172749, + 0.004150733817368746, + 0.00396352494135499, + 0.004213706124573946, + 0.0038923651445657015, + 0.004868665710091591, + 0.004294522572308779, + 0.0029672321397811174, + 0.0049049630761146545, + 0.0036778240464627743, + 0.004644931294023991, + 0.005086670164018869, + 0.004741139244288206, + 0.0038906228728592396, + 0.00400233780965209, + 0.0044108666479587555, + 0.0033050093334168196, + 0.004007745534181595, + 0.0065106539987027645, + 0.006848869379609823, + 0.006053471937775612, + 0.005506026092916727, + 0.004192679189145565, + 0.006608354859054089, + 0.007122687995433807, + 0.00908638071268797, + 0.0037653192412108183, + 0.005101019982248545, + 0.002867138711735606, + 0.003518805606290698, + 0.0037370368372648954, + 0.004136789124459028, + 0.0046629230491817, + 0.003789504524320364, + 0.0374971441924572, + 0.011002643965184689, + 0.017527883872389793, + 0.021807586774230003, + 0.01045145746320486, + 0.09118565171957016, + 0.016731426119804382, + 0.09640289098024368, + 0.008906582370400429, + 0.0091690793633461, + 0.00763779366388917, + 0.006549592595547438, + 0.005534692667424679, + 0.006812840234488249, + 0.009020489640533924, + 0.006892619654536247, + 0.003583375597372651, + 0.003114624647423625, + 0.018324552103877068, + 0.007834847085177898, + 0.0033728694543242455, + 0.01141547504812479, + 0.004815052263438702, + 0.004272706340998411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0031894787680357695 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_tensor([2725, 2725])_group_8_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_tensor([2725, 2725])_group_8_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4f715d347ad70080ef9b134a861be89cf03db48e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_tensor([2725, 2725])_group_8_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a dark anime style, a low-angle, static medium shot establishes the overwhelming presence of the figure in the foreground, their form rendered with sharp, clean lines against the oppressive darkness. They are impeccably dressed in a stylish vintage black tuxedo with a shiny satin lapel paired with a crisp white shirt. The figure's face is completely obscured by a pure white, owl-like porcelain mask, its smooth, cold surface punctuated by two large, round, black, hollow eyeholes. They are situated in a gloomy Gothic room, the background fading into deep, impenetrable shadows. As if delivering a chilling monologue, the figure's head tilts slowly and deliberately to one side. Simultaneously, their right hand, encased in a pristine white glove, rises in a slow, measured, graceful frame. A flickering candlelight source off-camera captures the mask's polished curves, casting a sharp, evocative glint that illuminates a trail of dust motes dancing in the still, heavy air, creating an atmosphere of profound tension and drama.", + "seed": 9503, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_2725_seed_9503_dp_16_rank_32_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3070093670627102, + "vlm_logprob": -0.3070093670627102, + "vlm_yes_no_margin": 2.025, + "vlm_token_credit": -0.3070093670627102, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6304453015327454 + ], + "opposite_logprob_tokens": [ + -0.8804453015327454 + ], + "logprob_reward": -0.6304453015327454, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6304453015327454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3150177001953125, + 0.11565890163183212, + 0.06927934288978577, + 0.04230945184826851, + 0.05370407924056053, + 0.038523752242326736, + 0.09719397127628326, + 0.055456195026636124, + 0.06576438248157501, + 0.14709217846393585 + ], + "credit_map": [ + 0.03157803416252136, + 0.028289010748267174, + 0.04830286651849747, + 0.0369318388402462, + 0.03407999873161316, + 0.05186748504638672, + 0.04323486238718033, + 0.04073360562324524, + 0.013472437858581543, + 0.007955113425850868, + 0.016939107328653336, + 0.017320292070508003, + 0.011931970715522766, + 0.0157929640263319, + 0.017180178314447403, + 0.015066834166646004, + 0.00852399505674839, + 0.0050309100188314915, + 0.013050409965217113, + 0.0070606498047709465, + 0.005869538523256779, + 0.01327497884631157, + 0.008798377588391304, + 0.007670482154935598, + 0.004890336189419031, + 0.002659361343830824, + 0.006072759162634611, + 0.007826918736100197, + 0.0031997482292354107, + 0.0063822208903729916, + 0.005832029040902853, + 0.005446077790111303, + 0.0060319784097373486, + 0.003303907811641693, + 0.008484537713229656, + 0.00889421533793211, + 0.003520869417116046, + 0.009255794808268547, + 0.004875626415014267, + 0.009337148629128933, + 0.0054543353617191315, + 0.00277684535831213, + 0.007484382949769497, + 0.005195781588554382, + 0.003097081556916237, + 0.005637572146952152, + 0.0033323399256914854, + 0.005545411724597216, + 0.0158223919570446, + 0.005226043052971363, + 0.012222004123032093, + 0.019233163446187973, + 0.006577678956091404, + 0.012310165911912918, + 0.01411083247512579, + 0.011691691353917122, + 0.00616971543058753, + 0.004091246053576469, + 0.009650107473134995, + 0.010997829958796501, + 0.0032175760716199875, + 0.009084954857826233, + 0.006482203025370836, + 0.005762564018368721, + 0.00984905194491148, + 0.003622319083660841, + 0.01447264850139618, + 0.009136688895523548, + 0.003748442744836211, + 0.009464168921113014, + 0.004496216308325529, + 0.01097484678030014, + 0.024414626881480217, + 0.004815333988517523, + 0.046052236109972, + 0.012856438755989075, + 0.0054611582309007645, + 0.026809288188815117, + 0.014525895938277245, + 0.012157210148870945 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.6304453015327454 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.523429274559021 + ], + "opposite_logprob_tokens": [ + -0.898429274559021 + ], + "logprob_reward": -0.523429274559021, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.523429274559021, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18586985766887665, + 0.09687376022338867, + 0.05482189357280731, + 0.06658380478620529, + 0.0657973363995552, + 0.07160642743110657, + 0.14592932164669037, + 0.06196998059749603, + 0.16256976127624512, + 0.0879778191447258 + ], + "credit_map": [ + 0.022387836128473282, + 0.01676103286445141, + 0.026139048859477043, + 0.03337317332625389, + 0.014226209372282028, + 0.027378516271710396, + 0.025329791009426117, + 0.020274242386221886, + 0.009611467830836773, + 0.008124676533043385, + 0.013883557170629501, + 0.014783136546611786, + 0.007823032326996326, + 0.016396882012486458, + 0.013363043777644634, + 0.012887959368526936, + 0.006996070500463247, + 0.0049482714384794235, + 0.0077996947802603245, + 0.007378087844699621, + 0.005027512088418007, + 0.007559236139059067, + 0.008727220818400383, + 0.0063858008943498135, + 0.007251235190778971, + 0.004790871404111385, + 0.00931120291352272, + 0.009046460501849651, + 0.007410948630422354, + 0.01109709870070219, + 0.009580071084201336, + 0.00809591542929411, + 0.0069310786202549934, + 0.00445278687402606, + 0.009680151008069515, + 0.007371946237981319, + 0.006891991943120956, + 0.012726973742246628, + 0.009531782940030098, + 0.008210626430809498, + 0.009165898896753788, + 0.005461372435092926, + 0.010582881979644299, + 0.011354572139680386, + 0.00737259816378355, + 0.008824638091027737, + 0.006879900582134724, + 0.01196456141769886, + 0.021543271839618683, + 0.010332794860005379, + 0.02336045540869236, + 0.02321672812104225, + 0.011457115411758423, + 0.018383964896202087, + 0.020246457308530807, + 0.017388543114066124, + 0.007102774456143379, + 0.005581193137913942, + 0.01018217671662569, + 0.009505516849458218, + 0.007129110861569643, + 0.008799874223768711, + 0.008581377565860748, + 0.005087954457849264, + 0.02863391488790512, + 0.0042532277293503284, + 0.04473937302827835, + 0.015315303578972816, + 0.004418153315782547, + 0.024967458099126816, + 0.009846349246799946, + 0.030395979061722755, + 0.012925168499350548, + 0.004291525576263666, + 0.02256481908261776, + 0.012922987341880798, + 0.0033172627445310354, + 0.014519504271447659, + 0.007481683511286974, + 0.009954861365258694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.523429274559021 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2521485388278961 + ], + "opposite_logprob_tokens": [ + -1.5021485090255737 + ], + "logprob_reward": -0.2521485388278961, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2521485388278961, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1515243649482727, + 0.059631817042827606, + 0.06257273256778717, + 0.037092991173267365, + 0.05895201861858368, + 0.04859504848718643, + 0.16870129108428955, + 0.040374673902988434, + 0.05083568021655083, + 0.32171928882598877 + ], + "credit_map": [ + 0.018908079713582993, + 0.010544371791183949, + 0.030252087861299515, + 0.019986111670732498, + 0.008515927940607071, + 0.02084582857787609, + 0.024817707017064095, + 0.017654268071055412, + 0.00607858132570982, + 0.004898285958915949, + 0.009015082381665707, + 0.010009611025452614, + 0.004046402405947447, + 0.008491475135087967, + 0.009518659673631191, + 0.007573728449642658, + 0.007343507371842861, + 0.004795887973159552, + 0.008575648069381714, + 0.012204527854919434, + 0.004751091822981834, + 0.007842281833291054, + 0.007709446828812361, + 0.009350351989269257, + 0.004426612053066492, + 0.0019913199357688427, + 0.00550713250413537, + 0.004854737315326929, + 0.00303683546371758, + 0.006275755353271961, + 0.004361371509730816, + 0.006639230530709028, + 0.006026094313710928, + 0.0035977589432150126, + 0.00992677453905344, + 0.006768168415874243, + 0.004113298375159502, + 0.010994005016982555, + 0.008174802176654339, + 0.009351123124361038, + 0.006571759935468435, + 0.0031361503060907125, + 0.007651550695300102, + 0.007283319719135761, + 0.0032026004046201706, + 0.007485589478164911, + 0.00480277556926012, + 0.008461306802928448, + 0.02248663827776909, + 0.009338921867311, + 0.01649114303290844, + 0.040600329637527466, + 0.007702608592808247, + 0.020929142832756042, + 0.02405860647559166, + 0.027093924582004547, + 0.004399398807436228, + 0.0031881395261734724, + 0.006913002114742994, + 0.00723589351400733, + 0.0025453129783272743, + 0.007486880291253328, + 0.00460785161703825, + 0.003998201806098223, + 0.00700876722112298, + 0.004751225002110004, + 0.01130884699523449, + 0.005634801462292671, + 0.004536810331046581, + 0.007885800674557686, + 0.004845790099352598, + 0.004863644018769264, + 0.05464233458042145, + 0.009944513440132141, + 0.09989891946315765, + 0.030719788745045662, + 0.012888152152299881, + 0.06016295775771141, + 0.027973242104053497, + 0.02548941969871521 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 6, + 0 + ], + "selected_reward": -0.2521485388278961 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019501493079587817 + ], + "opposite_logprob_tokens": [ + -6.251950263977051 + ], + "logprob_reward": -0.0019501493079587817, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019501493079587817, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18827024102210999, + 0.06894609332084656, + 0.06166388466954231, + 0.04131874069571495, + 0.04836307466030121, + 0.06427828222513199, + 0.12640173733234406, + 0.1161956638097763, + 0.08446924388408661, + 0.20009303092956543 + ], + "credit_map": [ + 0.01865803822875023, + 0.01786276511847973, + 0.029728645458817482, + 0.024796223267912865, + 0.01738392934203148, + 0.02726181037724018, + 0.026673175394535065, + 0.025905663147568703, + 0.0071861217729747295, + 0.005086724180728197, + 0.011018531396985054, + 0.008545836433768272, + 0.007293602917343378, + 0.010040097869932652, + 0.009903459809720516, + 0.00987172219902277, + 0.007647852413356304, + 0.005359358619898558, + 0.008051138371229172, + 0.010314589366316795, + 0.006322710774838924, + 0.007880215533077717, + 0.009090526960790157, + 0.006997498217970133, + 0.004919215105473995, + 0.0027321316301822662, + 0.006376876961439848, + 0.0049757761880755424, + 0.003862817771732807, + 0.006811596918851137, + 0.00601614685729146, + 0.005624181590974331, + 0.004870354197919369, + 0.0036211239639669657, + 0.007212616968899965, + 0.007018361706286669, + 0.003978092689067125, + 0.008668548427522182, + 0.005073899868875742, + 0.007920080795884132, + 0.007464184891432524, + 0.005758008453994989, + 0.009750052355229855, + 0.008052016608417034, + 0.0067532905377447605, + 0.00906718336045742, + 0.007765559013932943, + 0.009667993523180485, + 0.01817169040441513, + 0.00706845847889781, + 0.012818180955946445, + 0.030826017260551453, + 0.005791285075247288, + 0.016203980892896652, + 0.016858506947755814, + 0.018663618713617325, + 0.012179647572338581, + 0.007811324670910835, + 0.021615704521536827, + 0.022700810804963112, + 0.010428634472191334, + 0.017546217888593674, + 0.014453324489295483, + 0.009460004977881908, + 0.010264609009027481, + 0.007187555078417063, + 0.017292743548750877, + 0.009482304565608501, + 0.006897041108459234, + 0.014440501108765602, + 0.008965674787759781, + 0.00993881095200777, + 0.033393681049346924, + 0.006412251852452755, + 0.06280948221683502, + 0.017522800713777542, + 0.009844785556197166, + 0.037511084228754044, + 0.018106495961546898, + 0.014492456801235676 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 6 + ], + "selected_reward": -0.0019501493079587817 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.12707357108592987 + ], + "opposite_logprob_tokens": [ + -2.1270735263824463 + ], + "logprob_reward": -0.12707357108592987, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.12707357108592987, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16442357003688812, + 0.08785586059093475, + 0.1053580716252327, + 0.0336327962577343, + 0.0550774447619915, + 0.05148247256875038, + 0.10560393333435059, + 0.1669716089963913, + 0.12896490097045898, + 0.10062932968139648 + ], + "credit_map": [ + 0.017121072858572006, + 0.011964304372668266, + 0.027928052470088005, + 0.024722076952457428, + 0.010094277560710907, + 0.027880139648914337, + 0.027069587260484695, + 0.017644057050347328, + 0.012974542565643787, + 0.006851390935480595, + 0.016376957297325134, + 0.0074715218506753445, + 0.007661900948733091, + 0.01409175805747509, + 0.013160237111151218, + 0.009267550893127918, + 0.013628575950860977, + 0.008084788918495178, + 0.018463080748915672, + 0.012486620806157589, + 0.009285044856369495, + 0.01681625097990036, + 0.014025331474840641, + 0.012568381614983082, + 0.003384609241038561, + 0.002719503827393055, + 0.005035990383476019, + 0.005957948509603739, + 0.0035913519095629454, + 0.004369113594293594, + 0.0041734264232218266, + 0.004400854464620352, + 0.006335487123578787, + 0.005295390263199806, + 0.007483965717256069, + 0.00854510348290205, + 0.005044048186391592, + 0.00674831448122859, + 0.007922576740384102, + 0.007702556438744068, + 0.006571572739630938, + 0.002439865143969655, + 0.0073979622684419155, + 0.00841592438519001, + 0.0033504017628729343, + 0.007975232787430286, + 0.00607939762994647, + 0.009252113290131092, + 0.015509081073105335, + 0.005103715229779482, + 0.013294984586536884, + 0.019349433481693268, + 0.005407365970313549, + 0.015620951540768147, + 0.011869438923895359, + 0.019448963925242424, + 0.02031707391142845, + 0.0063704680651426315, + 0.032808151096105576, + 0.028825124725699425, + 0.011156104505062103, + 0.02019450254738331, + 0.03526884317398071, + 0.01203135121613741, + 0.022372476756572723, + 0.0038587204180657864, + 0.03387891501188278, + 0.01303828600794077, + 0.0038031169679015875, + 0.018940843641757965, + 0.008210400119423866, + 0.024862142279744148, + 0.01610054261982441, + 0.0046145557425916195, + 0.028906535357236862, + 0.010800558142364025, + 0.006464442703872919, + 0.016501370817422867, + 0.009112256579101086, + 0.008129068650305271 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 8 + ], + "selected_reward": -0.12707357108592987 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dark anime style, a low-angle, static medium shot establishes the overwhelming presence of the figure in the foreground, their form rendered with sharp, clean lines against the oppressive darkness. They are impeccably dressed in a stylish vintage black tuxedo with a shiny satin lapel paired with a crisp white shirt. The figure's face is completely obscured by a pure white, owl-like porcelain mask, its smooth, cold surface punctuated by two large, round, black, hollow eyeholes. They are situated in a gloomy Gothic room, the background fading into deep, impenetrable shadows. As if delivering a chilling monologue, the figure's head tilts slowly and deliberately to one side. Simultaneously, their right hand, encased in a pristine white glove, rises in a slow, measured, graceful frame. A flickering candlelight source off-camera captures the mask's polished curves, casting a sharp, evocative glint that illuminates a trail of dust motes dancing in the still, heavy air, creating an atmosphere of profound tension and drama.", + "seed": 9504, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_2725_seed_9504_dp_16_rank_32_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.2965373547747731, + "vlm_logprob": -0.2965373547747731, + "vlm_yes_no_margin": 2.9, + "vlm_token_credit": -0.2965373547747731, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6723003387451172 + ], + "opposite_logprob_tokens": [ + -0.9223003387451172 + ], + "logprob_reward": -0.6723003387451172, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.6723003387451172, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13091407716274261, + 0.1472264528274536, + 0.055253468453884125, + 0.11628808826208115, + 0.04693683236837387, + 0.042095378041267395, + 0.12977589666843414, + 0.07986653596162796, + 0.0986890196800232, + 0.15295416116714478 + ], + "credit_map": [ + 0.021771056577563286, + 0.007920786738395691, + 0.025622336193919182, + 0.01030003186315298, + 0.008941036649048328, + 0.02383972890675068, + 0.017112169414758682, + 0.015406934544444084, + 0.021860726177692413, + 0.013280455023050308, + 0.02189072035253048, + 0.022284148260951042, + 0.013967215083539486, + 0.03103044256567955, + 0.015019252896308899, + 0.00789350550621748, + 0.008145862258970737, + 0.005540443584322929, + 0.008826113305985928, + 0.00504198158159852, + 0.006140829995274544, + 0.010902425274252892, + 0.006228106562048197, + 0.004427707754075527, + 0.016322258859872818, + 0.010426458902657032, + 0.02230621501803398, + 0.014142879284918308, + 0.00895324070006609, + 0.022402623668313026, + 0.014161585830152035, + 0.007572823204100132, + 0.006045559421181679, + 0.003381407121196389, + 0.012249906547367573, + 0.004984090570360422, + 0.003408176824450493, + 0.006149458233267069, + 0.006521728355437517, + 0.0041965036652982235, + 0.0062967706471681595, + 0.0037818443961441517, + 0.007725605741143227, + 0.003520561847835779, + 0.004180992022156715, + 0.0084222462028265, + 0.005069547798484564, + 0.003097811946645379, + 0.01940944977104664, + 0.00902695581316948, + 0.01864524930715561, + 0.014783993363380432, + 0.010113582946360111, + 0.024289574474096298, + 0.025640694424510002, + 0.007866401225328445, + 0.008627768605947495, + 0.00681739067658782, + 0.020946217700839043, + 0.0076852464117109776, + 0.009183211252093315, + 0.010861790738999844, + 0.0098720146343112, + 0.0058728959411382675, + 0.010302131995558739, + 0.011532876640558243, + 0.016885412856936455, + 0.009377209469676018, + 0.015079271979629993, + 0.012863090261816978, + 0.014813002198934555, + 0.007836024276912212, + 0.015635963529348373, + 0.013917463831603527, + 0.021508898586034775, + 0.02202211692929268, + 0.017963411286473274, + 0.019873084500432014, + 0.030964745208621025, + 0.011068476364016533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 0 + ], + "selected_reward": -0.6723003387451172 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579315304756165 + ], + "opposite_logprob_tokens": [ + -0.6329315304756165 + ], + "logprob_reward": -0.7579315304756165, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579315304756165, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04927227273583412, + 0.2004719078540802, + 0.0791507214307785, + 0.0906909704208374, + 0.04833691194653511, + 0.029575204476714134, + 0.17738604545593262, + 0.13161738216876984, + 0.0640498548746109, + 0.1294487863779068 + ], + "credit_map": [ + 0.00668292073532939, + 0.00306363170966506, + 0.01234626118093729, + 0.004776861984282732, + 0.0037461102474480867, + 0.007397301960736513, + 0.005852304864674807, + 0.005406881216913462, + 0.03431867063045502, + 0.016857942566275597, + 0.028248125687241554, + 0.03112276829779148, + 0.016339214518666267, + 0.046403560787439346, + 0.01717660203576088, + 0.010005016811192036, + 0.013170501217246056, + 0.008461251854896545, + 0.01179534662514925, + 0.00625875499099493, + 0.007418959401547909, + 0.017187710851430893, + 0.007762609515339136, + 0.007095587905496359, + 0.012503081001341343, + 0.009486913681030273, + 0.012191998772323132, + 0.009966307319700718, + 0.009757359512150288, + 0.015414089895784855, + 0.013064765371382236, + 0.008306448347866535, + 0.0069189476780593395, + 0.004950720351189375, + 0.007441797759383917, + 0.00501667195931077, + 0.004571481142193079, + 0.008242934942245483, + 0.006085890810936689, + 0.005108468700200319, + 0.003317038295790553, + 0.0019879662431776524, + 0.0073605431243777275, + 0.004052020609378815, + 0.002169728744775057, + 0.00497275497764349, + 0.0027834789361804724, + 0.002931674476712942, + 0.027046987786889076, + 0.0105564771220088, + 0.02605765126645565, + 0.022709263488650322, + 0.01681589148938656, + 0.03654002398252487, + 0.027226319536566734, + 0.010433439165353775, + 0.015444920398294926, + 0.013535699807107449, + 0.02299034409224987, + 0.01128868293017149, + 0.020667601376771927, + 0.017124641686677933, + 0.01829148456454277, + 0.012274019420146942, + 0.007181475404649973, + 0.008170383051037788, + 0.011359772644937038, + 0.004947802517563105, + 0.01074968185275793, + 0.008731851354241371, + 0.00828289333730936, + 0.004625993315130472, + 0.014905419200658798, + 0.014553308486938477, + 0.0297351386398077, + 0.01486120093613863, + 0.018934937193989754, + 0.01702343113720417, + 0.012058889493346214, + 0.007376452907919884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 7 + ], + "selected_reward": -0.7579315304756165 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0076653859578073025 + ], + "opposite_logprob_tokens": [ + -4.882665157318115 + ], + "logprob_reward": -0.0076653859578073025, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0076653859578073025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.341234028339386, + 0.19497831165790558, + 0.07147347927093506, + 0.01703551597893238, + 0.017010901123285294, + 0.03522744029760361, + 0.06761936843395233, + 0.06780287623405457, + 0.07372674345970154, + 0.11389134079217911 + ], + "credit_map": [ + 0.0504884235560894, + 0.02435494400560856, + 0.05105602368712425, + 0.06087110936641693, + 0.027241118252277374, + 0.06655675917863846, + 0.0389835499227047, + 0.02168210782110691, + 0.028747623786330223, + 0.01717737503349781, + 0.0311647467315197, + 0.029346469789743423, + 0.018961291760206223, + 0.03712386637926102, + 0.02239769883453846, + 0.01005923468619585, + 0.00851818174123764, + 0.007049370091408491, + 0.014750991947948933, + 0.008193869143724442, + 0.006439578253775835, + 0.010156204923987389, + 0.009849341586232185, + 0.006515946704894304, + 0.002129651140421629, + 0.001009474159218371, + 0.005132926627993584, + 0.001656207605265081, + 0.0012496982235461473, + 0.002581111853942275, + 0.0014952553901821375, + 0.0017811907455325127, + 0.0022899096366018057, + 0.0009325631544925272, + 0.005679314490407705, + 0.0010435004951432347, + 0.0011684271739795804, + 0.0029117180965840816, + 0.001569782616570592, + 0.001415685866959393, + 0.00493574095889926, + 0.0028605409897863865, + 0.006954801734536886, + 0.004815524909645319, + 0.0030652887653559446, + 0.00676594628021121, + 0.0037151703145354986, + 0.0021144249476492405, + 0.009698293171823025, + 0.004634961020201445, + 0.01555879320949316, + 0.00789786595851183, + 0.005427167750895023, + 0.013059845194220543, + 0.006106082815676928, + 0.005236356984823942, + 0.008062133565545082, + 0.005881943739950657, + 0.014909347519278526, + 0.005644822958856821, + 0.008427094668149948, + 0.00908464752137661, + 0.00948888249695301, + 0.006304000969976187, + 0.0076679084450006485, + 0.00974416732788086, + 0.01157028041779995, + 0.008722759783267975, + 0.012675999663770199, + 0.008744336664676666, + 0.009957981295883656, + 0.004643312189728022, + 0.013548407703638077, + 0.00989449117332697, + 0.018201082944869995, + 0.011464541777968407, + 0.013824858702719212, + 0.015175659209489822, + 0.022025344893336296, + 0.009756949730217457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.0076653859578073025 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017239484004676342 + ], + "opposite_logprob_tokens": [ + -6.376723766326904 + ], + "logprob_reward": -0.0017239484004676342, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017239484004676342, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14433139562606812, + 0.19606222212314606, + 0.13014353811740875, + 0.02882029488682747, + 0.04364955797791481, + 0.021775521337985992, + 0.08697959035634995, + 0.1712590754032135, + 0.12585456669330597, + 0.05112428218126297 + ], + "credit_map": [ + 0.01608744077384472, + 0.012355712242424488, + 0.03273339942097664, + 0.01838424615561962, + 0.014575663954019547, + 0.01957903429865837, + 0.01956643909215927, + 0.011049455031752586, + 0.022575436159968376, + 0.02027079276740551, + 0.03489331528544426, + 0.031180240213871002, + 0.02301192842423916, + 0.02807590737938881, + 0.02407119981944561, + 0.01198340393602848, + 0.020633172243833542, + 0.012043403461575508, + 0.02913801372051239, + 0.010567398741841316, + 0.011106237769126892, + 0.026170402765274048, + 0.013265637680888176, + 0.007219264283776283, + 0.0032371056731790304, + 0.002502794610336423, + 0.006633706856518984, + 0.002606879221275449, + 0.0029374961741268635, + 0.004362011794000864, + 0.00400945357978344, + 0.0025308451149612665, + 0.007403064053505659, + 0.004348069429397583, + 0.00811388622969389, + 0.003512886818498373, + 0.004087225534021854, + 0.007653587032109499, + 0.0057247658260166645, + 0.002806074684485793, + 0.0029571186751127243, + 0.0017926590517163277, + 0.003959384746849537, + 0.0021274022292345762, + 0.0021257579792290926, + 0.0034852931275963783, + 0.0030947071500122547, + 0.0022331983782351017, + 0.013809211552143097, + 0.006581163499504328, + 0.013908767141401768, + 0.008376630954444408, + 0.007141635287553072, + 0.015486417338252068, + 0.014651491306722164, + 0.007024272810667753, + 0.020914945751428604, + 0.017800476402044296, + 0.027938611805438995, + 0.016531027853488922, + 0.026463089510798454, + 0.021079421043395996, + 0.024535922333598137, + 0.015995582565665245, + 0.013371359556913376, + 0.016535207629203796, + 0.017125308513641357, + 0.015393074601888657, + 0.02223959006369114, + 0.016338204964995384, + 0.017969438806176186, + 0.006882386282086372, + 0.0061455294489860535, + 0.00293336040340364, + 0.013792346231639385, + 0.006573403719812632, + 0.004070958588272333, + 0.007429152261465788, + 0.005458553321659565, + 0.004720977041870356 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.0017239484004676342 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043065570294857025 + ], + "opposite_logprob_tokens": [ + -3.1680655479431152 + ], + "logprob_reward": -0.043065570294857025, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043065570294857025, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10587767511606216, + 0.21071988344192505, + 0.044993460178375244, + 0.050644584000110626, + 0.08044829219579697, + 0.018476208671927452, + 0.10793206095695496, + 0.20659838616847992, + 0.060175731778144836, + 0.11413371562957764 + ], + "credit_map": [ + 0.012742758728563786, + 0.008857255801558495, + 0.02121668867766857, + 0.013498538173735142, + 0.012096983380615711, + 0.01425359956920147, + 0.014303811825811863, + 0.008908032439649105, + 0.03473871946334839, + 0.017324404790997505, + 0.026667386293411255, + 0.035861242562532425, + 0.0174326803535223, + 0.046737007796764374, + 0.020694447681307793, + 0.011263998225331306, + 0.005919098854064941, + 0.0027461552526801825, + 0.015882086008787155, + 0.0027281264774501324, + 0.0031497860327363014, + 0.007770161144435406, + 0.003539311233907938, + 0.003258730750530958, + 0.0077332016080617905, + 0.004972823895514011, + 0.008523250930011272, + 0.004547550342977047, + 0.005579835269600153, + 0.008424255065619946, + 0.006875970866531134, + 0.003987693693488836, + 0.013153367675840855, + 0.008829595521092415, + 0.010856526903808117, + 0.008821631781756878, + 0.007986040785908699, + 0.014277379028499126, + 0.010896841995418072, + 0.005626905709505081, + 0.0023990299087017775, + 0.0013831184478476644, + 0.004299999214708805, + 0.0022467223461717367, + 0.0016202230472117662, + 0.0027431847993284464, + 0.0024181483313441277, + 0.001365780015476048, + 0.015584000386297703, + 0.006526563782244921, + 0.017771683633327484, + 0.01499022077769041, + 0.010503102093935013, + 0.021609587594866753, + 0.01376364752650261, + 0.007183248642832041, + 0.024841180071234703, + 0.021531477570533752, + 0.03401018679141998, + 0.01947745867073536, + 0.031490474939346313, + 0.025357462465763092, + 0.029456155374646187, + 0.020433977246284485, + 0.007020754739642143, + 0.008353362791240215, + 0.009353086352348328, + 0.005232490599155426, + 0.010512091219425201, + 0.008252334780991077, + 0.00828373059630394, + 0.0031678760424256325, + 0.01347140222787857, + 0.012203642167150974, + 0.028962668031454086, + 0.012670348398387432, + 0.013699540868401527, + 0.01517916563898325, + 0.01236804760992527, + 0.005578893236815929 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 9 + ], + "selected_reward": -0.043065570294857025 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dark anime style, a low-angle, static medium shot establishes the overwhelming presence of the figure in the foreground, their form rendered with sharp, clean lines against the oppressive darkness. They are impeccably dressed in a stylish vintage black tuxedo with a shiny satin lapel paired with a crisp white shirt. The figure's face is completely obscured by a pure white, owl-like porcelain mask, its smooth, cold surface punctuated by two large, round, black, hollow eyeholes. They are situated in a gloomy Gothic room, the background fading into deep, impenetrable shadows. As if delivering a chilling monologue, the figure's head tilts slowly and deliberately to one side. Simultaneously, their right hand, encased in a pristine white glove, rises in a slow, measured, graceful frame. A flickering candlelight source off-camera captures the mask's polished curves, casting a sharp, evocative glint that illuminates a trail of dust motes dancing in the still, heavy air, creating an atmosphere of profound tension and drama.", + "seed": 9505, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_2725_seed_9505_dp_17_rank_34_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2495291502447799, + "vlm_logprob": -0.2495291502447799, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.2495291502447799, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.46335768699645996 + ], + "opposite_logprob_tokens": [ + -1.21335768699646 + ], + "logprob_reward": -0.46335768699645996, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.46335768699645996, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19602712988853455, + 0.09203383326530457, + 0.04678225517272949, + 0.09148328006267548, + 0.12078426033258438, + 0.06914696097373962, + 0.05309325084090233, + 0.053244736045598984, + 0.06720427423715591, + 0.2102000117301941 + ], + "credit_map": [ + 0.030149007216095924, + 0.032159216701984406, + 0.03580964729189873, + 0.01759842410683632, + 0.014155750162899494, + 0.023650502786040306, + 0.019482847303152084, + 0.023021729663014412, + 0.013750283978879452, + 0.014473604038357735, + 0.014491360634565353, + 0.009099018760025501, + 0.006500097922980785, + 0.01068692747503519, + 0.01146337017416954, + 0.011569171212613583, + 0.006034910213202238, + 0.00737764360383153, + 0.008461643010377884, + 0.004298694897443056, + 0.0036673834547400475, + 0.00530250184237957, + 0.00529630808159709, + 0.006343173328787088, + 0.009884795174002647, + 0.01190746109932661, + 0.01224963553249836, + 0.014464628882706165, + 0.006552288308739662, + 0.011415911838412285, + 0.01306042168289423, + 0.011948136612772942, + 0.012828405946493149, + 0.015497882850468159, + 0.0164837297052145, + 0.01731284335255623, + 0.009323643520474434, + 0.016237886622548103, + 0.017035722732543945, + 0.016064153984189034, + 0.008457153104245663, + 0.010101789608597755, + 0.013099596835672855, + 0.006402790546417236, + 0.00528318015858531, + 0.006707232445478439, + 0.009696788154542446, + 0.009398430585861206, + 0.007047652266919613, + 0.007792230229824781, + 0.007807630579918623, + 0.006121518090367317, + 0.0038869567215442657, + 0.006020704284310341, + 0.006761455908417702, + 0.007655106019228697, + 0.006618164014071226, + 0.00819831807166338, + 0.011736174114048481, + 0.005721614696085453, + 0.004433706868439913, + 0.005652237683534622, + 0.005919327028095722, + 0.004965195897966623, + 0.006917568854987621, + 0.00907159224152565, + 0.010778277181088924, + 0.010176514275372028, + 0.005764876492321491, + 0.00793073046952486, + 0.007717051077634096, + 0.008847666904330254, + 0.028938010334968567, + 0.03008655272424221, + 0.03648519143462181, + 0.024478262290358543, + 0.019069863483309746, + 0.019790729507803917, + 0.023887431249022484, + 0.02746398001909256 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.46335768699645996 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7579164505004883 + ], + "opposite_logprob_tokens": [ + -0.6329164505004883 + ], + "logprob_reward": -0.7579164505004883, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7579164505004883, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.13468369841575623, + 0.06801550090312958, + 0.08414845913648605, + 0.052394699305295944, + 0.09105179458856583, + 0.09763237088918686, + 0.07309499382972717, + 0.17369098961353302, + 0.09813681244850159, + 0.12715066969394684 + ], + "credit_map": [ + 0.018671415746212006, + 0.023055534809827805, + 0.02466527558863163, + 0.01522913295775652, + 0.00787787139415741, + 0.017007533460855484, + 0.01674824394285679, + 0.011428693309426308, + 0.008658817037940025, + 0.012110598385334015, + 0.00990736298263073, + 0.008716065436601639, + 0.004221063572913408, + 0.007996966131031513, + 0.00838214810937643, + 0.008022483438253403, + 0.008779269643127918, + 0.013974452391266823, + 0.01088977325707674, + 0.011575869284570217, + 0.005643558222800493, + 0.009159781970083714, + 0.01285526528954506, + 0.011270495131611824, + 0.006367672700434923, + 0.0063285320065915585, + 0.00981203280389309, + 0.009422882460057735, + 0.0029415879398584366, + 0.005991616751998663, + 0.0051196604035794735, + 0.00641071330755949, + 0.012472432106733322, + 0.01366181019693613, + 0.01946096681058407, + 0.009824068285524845, + 0.007259042467921972, + 0.008000276051461697, + 0.007748145144432783, + 0.012625058181583881, + 0.013591138646006584, + 0.013262041844427586, + 0.0151882478967309, + 0.012873170897364616, + 0.00905914418399334, + 0.008859599940478802, + 0.013711382634937763, + 0.011087654158473015, + 0.010369480587542057, + 0.010934514924883842, + 0.0119984932243824, + 0.01089176069945097, + 0.006012998055666685, + 0.005910343024879694, + 0.008957568556070328, + 0.008019836619496346, + 0.022311782464385033, + 0.026580912992358208, + 0.0395047664642334, + 0.02303209900856018, + 0.016193659976124763, + 0.013624846935272217, + 0.01600874587893486, + 0.01643417961895466, + 0.013636970892548561, + 0.014205670915544033, + 0.016522066667675972, + 0.012214516289532185, + 0.010435874573886395, + 0.009383989498019218, + 0.008025093004107475, + 0.013712625950574875, + 0.024379683658480644, + 0.019401486963033676, + 0.020441589877009392, + 0.01539839431643486, + 0.011060671880841255, + 0.01047323178499937, + 0.01086420752108097, + 0.015131406486034393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.7579164505004883 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011118480004370213 + ], + "opposite_logprob_tokens": [ + -4.511118412017822 + ], + "logprob_reward": -0.011118480004370213, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011118480004370213, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3285372853279114, + 0.1406414806842804, + 0.10599645227193832, + 0.04416605830192566, + 0.06504640728235245, + 0.027479182928800583, + 0.03346726670861244, + 0.08888434618711472, + 0.0684828832745552, + 0.09729860723018646 + ], + "credit_map": [ + 0.026315979659557343, + 0.05629781633615494, + 0.030946407467126846, + 0.04229575768113136, + 0.023221168667078018, + 0.047254905104637146, + 0.06049037352204323, + 0.0417148694396019, + 0.012789212167263031, + 0.02046065405011177, + 0.012997589074075222, + 0.017766298726201057, + 0.009369786828756332, + 0.020332608371973038, + 0.027211016044020653, + 0.019714320078492165, + 0.010868292301893234, + 0.016924072057008743, + 0.011206107214093208, + 0.013249420560896397, + 0.007264154497534037, + 0.011605140753090382, + 0.01934533752501011, + 0.015533922240138054, + 0.004234530031681061, + 0.0062308837659657, + 0.0063182576559484005, + 0.005483846180140972, + 0.0030214949510991573, + 0.005759090185165405, + 0.006000523455440998, + 0.007117428816854954, + 0.006479556206613779, + 0.00844860915094614, + 0.009352670051157475, + 0.011099633760750294, + 0.004278894048184156, + 0.005715430714190006, + 0.010244511999189854, + 0.009427101351320744, + 0.0033761607483029366, + 0.004254045896232128, + 0.005174417980015278, + 0.003188349772244692, + 0.0023933511693030596, + 0.0024882140569388866, + 0.0032609463669359684, + 0.003343698102980852, + 0.005089613609015942, + 0.004747787490487099, + 0.0077567626722157, + 0.00428061094135046, + 0.002544147428125143, + 0.0026832681614905596, + 0.002176585141569376, + 0.004188490565866232, + 0.012779689393937588, + 0.01574561558663845, + 0.014724945649504662, + 0.008696137927472591, + 0.008848879486322403, + 0.0074765621684491634, + 0.012066948227584362, + 0.008545564487576485, + 0.009657511487603188, + 0.008512001484632492, + 0.01100010797381401, + 0.008737606927752495, + 0.00767285143956542, + 0.007489703595638275, + 0.007806319277733564, + 0.0076067796908319, + 0.014557941816747189, + 0.014037655666470528, + 0.015610217116773129, + 0.011526805348694324, + 0.006920108571648598, + 0.008408036082983017, + 0.009907650761306286, + 0.016330188140273094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011118480004370213 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0010528980055823922 + ], + "opposite_logprob_tokens": [ + -6.8760528564453125 + ], + "logprob_reward": -0.0010528980055823922, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0010528980055823922, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16619452834129333, + 0.059516023844480515, + 0.04851214960217476, + 0.026716157793998718, + 0.06568977981805801, + 0.06296960264444351, + 0.03175092861056328, + 0.2632123529911041, + 0.11533454060554504, + 0.1601039618253708 + ], + "credit_map": [ + 0.021323326975107193, + 0.028462721034884453, + 0.023953242227435112, + 0.01628071628510952, + 0.01983281597495079, + 0.021738480776548386, + 0.01630319096148014, + 0.018300022929906845, + 0.0073877181857824326, + 0.008467037230730057, + 0.009045190177857876, + 0.006580038461834192, + 0.005835028365254402, + 0.00802974309772253, + 0.007670767605304718, + 0.006500501185655594, + 0.005120714195072651, + 0.007226807530969381, + 0.006525983102619648, + 0.0061853365041315556, + 0.0037294032517820597, + 0.005210089031606913, + 0.007473865523934364, + 0.007039952091872692, + 0.0037280628457665443, + 0.003590580075979233, + 0.004655771888792515, + 0.003118745284155011, + 0.0020048271398991346, + 0.003115573199465871, + 0.0027817096561193466, + 0.003720886306837201, + 0.007743142545223236, + 0.009312815964221954, + 0.009983261115849018, + 0.011627493426203728, + 0.005737081170082092, + 0.007363999728113413, + 0.0065681771375238895, + 0.007353813387453556, + 0.008705290034413338, + 0.010818911716341972, + 0.010517423041164875, + 0.0073341988027095795, + 0.004588564857840538, + 0.004907050170004368, + 0.006386016495525837, + 0.009712150320410728, + 0.005684679839760065, + 0.0050419834442436695, + 0.003865191014483571, + 0.002439720556139946, + 0.0015197891043499112, + 0.0022761886939406395, + 0.0036027596797794104, + 0.007320613134652376, + 0.02980666235089302, + 0.03708947077393532, + 0.0439283549785614, + 0.029154092073440552, + 0.029741255566477776, + 0.029105564579367638, + 0.029623955488204956, + 0.03476301208138466, + 0.01600939966738224, + 0.015278534963726997, + 0.019219227135181427, + 0.014877408742904663, + 0.01072678528726101, + 0.013261573389172554, + 0.011775532737374306, + 0.014186078682541847, + 0.018380040302872658, + 0.021664056926965714, + 0.018637968227267265, + 0.019698873162269592, + 0.019858885556459427, + 0.019935093820095062, + 0.02125483937561512, + 0.0206742025911808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 9 + ], + "selected_reward": -0.0010528980055823922 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014200235716998577 + ], + "opposite_logprob_tokens": [ + -4.264200210571289 + ], + "logprob_reward": -0.014200235716998577, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014200235716998577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10024680942296982, + 0.05649558827280998, + 0.09413522481918335, + 0.04066060483455658, + 0.11483103036880493, + 0.14689147472381592, + 0.050868477672338486, + 0.09397005289793015, + 0.11801768839359283, + 0.18388304114341736 + ], + "credit_map": [ + 0.013936344534158707, + 0.01464125420898199, + 0.01963711902499199, + 0.010468270629644394, + 0.007571396883577108, + 0.012066872790455818, + 0.00894539337605238, + 0.012980163097381592, + 0.006274016574025154, + 0.009701325558125973, + 0.008144299499690533, + 0.007892988622188568, + 0.0037383597809821367, + 0.006524662021547556, + 0.00722982594743371, + 0.006990112829953432, + 0.008915374055504799, + 0.015773992985486984, + 0.011015081778168678, + 0.013217095285654068, + 0.0062758768908679485, + 0.010444771498441696, + 0.014335514046251774, + 0.01415752712637186, + 0.007004758808761835, + 0.0053675551898777485, + 0.00840733665972948, + 0.004749684128910303, + 0.002981770783662796, + 0.00367338047362864, + 0.004396318923681974, + 0.004079805221408606, + 0.013781819492578506, + 0.015712568536400795, + 0.01567121036350727, + 0.0193563811480999, + 0.010027125477790833, + 0.009695480577647686, + 0.015516901388764381, + 0.015069547109305859, + 0.0183095745742321, + 0.022565744817256927, + 0.0198398157954216, + 0.02034653164446354, + 0.012108998373150826, + 0.013643555343151093, + 0.01730589009821415, + 0.022771373391151428, + 0.00718303956091404, + 0.007164584007114172, + 0.00688308198004961, + 0.005074456334114075, + 0.0036498713307082653, + 0.005148479714989662, + 0.006357467267662287, + 0.009407499805092812, + 0.013007787987589836, + 0.016075782477855682, + 0.024257998913526535, + 0.009337584488093853, + 0.006923273671418428, + 0.006734943948686123, + 0.007877855561673641, + 0.009754836559295654, + 0.017066672444343567, + 0.01617136411368847, + 0.018391849473118782, + 0.017631584778428078, + 0.010287790559232235, + 0.011563023552298546, + 0.011163577437400818, + 0.015741832554340363, + 0.02662188932299614, + 0.026609208434820175, + 0.03412638604640961, + 0.026673318818211555, + 0.011524871923029423, + 0.016651583835482597, + 0.017266035079956055, + 0.024409763514995575 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 8 + ], + "selected_reward": -0.014200235716998577 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a dark anime style, a low-angle, static medium shot establishes the overwhelming presence of the figure in the foreground, their form rendered with sharp, clean lines against the oppressive darkness. They are impeccably dressed in a stylish vintage black tuxedo with a shiny satin lapel paired with a crisp white shirt. The figure's face is completely obscured by a pure white, owl-like porcelain mask, its smooth, cold surface punctuated by two large, round, black, hollow eyeholes. They are situated in a gloomy Gothic room, the background fading into deep, impenetrable shadows. As if delivering a chilling monologue, the figure's head tilts slowly and deliberately to one side. Simultaneously, their right hand, encased in a pristine white glove, rises in a slow, measured, graceful frame. A flickering candlelight source off-camera captures the mask's polished curves, casting a sharp, evocative glint that illuminates a trail of dust motes dancing in the still, heavy air, creating an atmosphere of profound tension and drama.", + "seed": 9506, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_8/index_2725_seed_9506_dp_17_rank_34_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 8, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3409849470714107, + "vlm_logprob": -0.3409849470714107, + "vlm_yes_no_margin": 2.2, + "vlm_token_credit": -0.3409849470714107, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9447050094604492 + ], + "opposite_logprob_tokens": [ + -0.5697050094604492 + ], + "logprob_reward": -0.9447050094604492, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9447050094604492, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.2157469391822815, + 0.049337614327669144, + 0.023868504911661148, + 0.14155973494052887, + 0.19472229480743408, + 0.07655502110719681, + 0.11938115209341049, + 0.043296851217746735, + 0.04423701390624046, + 0.09129495918750763 + ], + "credit_map": [ + 0.016938231885433197, + 0.011818546801805496, + 0.04571596905589104, + 0.026108792051672935, + 0.010809427127242088, + 0.05656443163752556, + 0.02233276329934597, + 0.025458745658397675, + 0.00568718696013093, + 0.0034131966531276703, + 0.008208198472857475, + 0.006803933065384626, + 0.003265052568167448, + 0.007060483563691378, + 0.0066116866655647755, + 0.008287872187793255, + 0.002367720240727067, + 0.002147770021110773, + 0.003333618864417076, + 0.0026177733670920134, + 0.0025847593788057566, + 0.003278173506259918, + 0.0038581977132707834, + 0.0036804890260100365, + 0.011984768323600292, + 0.007063472177833319, + 0.02785111963748932, + 0.010783113539218903, + 0.011901688762009144, + 0.028636939823627472, + 0.01921406388282776, + 0.024124562740325928, + 0.018375439569354057, + 0.007871981710195541, + 0.041323304176330566, + 0.01722135953605175, + 0.013096742331981659, + 0.046425316482782364, + 0.022212186828255653, + 0.028195949271321297, + 0.00658388901501894, + 0.005961321759968996, + 0.011274914257228374, + 0.00849020667374134, + 0.007870646193623543, + 0.013018069788813591, + 0.011732112616300583, + 0.01162385568022728, + 0.010310170240700245, + 0.006010045763105154, + 0.023639418184757233, + 0.017642706632614136, + 0.006433624308556318, + 0.026635384187102318, + 0.010506403632462025, + 0.01820339262485504, + 0.004791411105543375, + 0.0021260285284370184, + 0.0055872355587780476, + 0.0046984367072582245, + 0.0025885947979986668, + 0.0070279366336762905, + 0.007098386529833078, + 0.00937881600111723, + 0.005169061943888664, + 0.0027727195993065834, + 0.005559035111218691, + 0.006935166195034981, + 0.0037821526639163494, + 0.006043922621756792, + 0.006824091542512178, + 0.007150858640670776, + 0.010240261442959309, + 0.006734093185514212, + 0.015502049587666988, + 0.016204139217734337, + 0.005719534121453762, + 0.012485607527196407, + 0.011445480398833752, + 0.01296378392726183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.9447050094604492 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5762235522270203 + ], + "opposite_logprob_tokens": [ + -0.8262235522270203 + ], + "logprob_reward": -0.5762235522270203, + "yes_no_margin_reward": 0.25, + "token_credit_reward": -0.5762235522270203, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13859589397907257, + 0.16250215470790863, + 0.08513040095567703, + 0.06584016978740692, + 0.09560485929250717, + 0.06917621940374374, + 0.10042618215084076, + 0.0577157698571682, + 0.08630009740591049, + 0.138708233833313 + ], + "credit_map": [ + 0.013521452434360981, + 0.006775309331715107, + 0.01585608348250389, + 0.009690443985164165, + 0.0071130068972706795, + 0.024874083697795868, + 0.020810892805457115, + 0.03995461389422417, + 0.008363942615687847, + 0.004319252446293831, + 0.05228850245475769, + 0.006072123069316149, + 0.00707020191475749, + 0.06222182884812355, + 0.008829932659864426, + 0.013336347416043282, + 0.0053491960279643536, + 0.003344494616612792, + 0.02261873334646225, + 0.0032614381052553654, + 0.004783290438354015, + 0.025969576090574265, + 0.006762686651200056, + 0.013040988706052303, + 0.004322227090597153, + 0.0025595524348318577, + 0.016972318291664124, + 0.0056064496748149395, + 0.0036437446251511574, + 0.01997801475226879, + 0.005880315322428942, + 0.006877546664327383, + 0.006339882966130972, + 0.0037509948015213013, + 0.026285098865628242, + 0.007893014699220657, + 0.005442408844828606, + 0.029342079535126686, + 0.006586662959307432, + 0.0099647156894207, + 0.005424169357866049, + 0.003212806535884738, + 0.013567829504609108, + 0.006161884870380163, + 0.004301334265619516, + 0.018431367352604866, + 0.006224935408681631, + 0.011851896531879902, + 0.010254090651869774, + 0.006575863808393478, + 0.015232917852699757, + 0.01536046527326107, + 0.007630310952663422, + 0.014532656408846378, + 0.012920335866510868, + 0.017919542267918587, + 0.005995788611471653, + 0.0033356070052832365, + 0.007979140616953373, + 0.008563563227653503, + 0.004426709376275539, + 0.009681539610028267, + 0.007137879263609648, + 0.010595547035336494, + 0.01021626591682434, + 0.0063811903819441795, + 0.012659537605941296, + 0.011429175734519958, + 0.007244741078466177, + 0.011697721667587757, + 0.01195459347218275, + 0.014716872945427895, + 0.015620367601513863, + 0.009833220392465591, + 0.023647427558898926, + 0.018609177321195602, + 0.012125855311751366, + 0.019369475543498993, + 0.017588669434189796, + 0.021914025768637657 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.5762235522270203 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1270897388458252 + ], + "opposite_logprob_tokens": [ + -2.127089738845825 + ], + "logprob_reward": -0.1270897388458252, + "yes_no_margin_reward": 2.0, + "token_credit_reward": -0.1270897388458252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15860235691070557, + 0.13940531015396118, + 0.03172828629612923, + 0.06734132021665573, + 0.03067140281200409, + 0.12528358399868011, + 0.08820522576570511, + 0.16963335871696472, + 0.026053236797451973, + 0.16307593882083893 + ], + "credit_map": [ + 0.010189719498157501, + 0.007436698768287897, + 0.0317879356443882, + 0.013302243314683437, + 0.010344571433961391, + 0.03346782922744751, + 0.02080305851995945, + 0.03127029538154602, + 0.008203295059502125, + 0.004361588042229414, + 0.03870479017496109, + 0.006263516377657652, + 0.0070666358806192875, + 0.047698818147182465, + 0.010251973755657673, + 0.016854694113135338, + 0.002917221048846841, + 0.0014910033205524087, + 0.004199502058327198, + 0.0026729090604931116, + 0.0020612054504454136, + 0.004884138237684965, + 0.004313274752348661, + 0.009189032018184662, + 0.0063376035541296005, + 0.00415370799601078, + 0.009605387225747108, + 0.006624349858611822, + 0.00617639347910881, + 0.010478734038770199, + 0.008781642653048038, + 0.015183503739535809, + 0.0021515453699976206, + 0.001561581389978528, + 0.005684269592165947, + 0.0027451016940176487, + 0.0021704379469156265, + 0.006112343166023493, + 0.004178585950285196, + 0.006067538168281317, + 0.007931368425488472, + 0.0064131771214306355, + 0.026362091302871704, + 0.010877980850636959, + 0.00834218692034483, + 0.02984972856938839, + 0.012693673372268677, + 0.022813374176621437, + 0.0067305369302630424, + 0.005212624557316303, + 0.016244255006313324, + 0.008001227863132954, + 0.006407792214304209, + 0.019787942990660667, + 0.009544319473206997, + 0.016276521608233452, + 0.014506852254271507, + 0.009079854935407639, + 0.031138397753238678, + 0.016626423224806786, + 0.011015182361006737, + 0.03764921799302101, + 0.02081305719912052, + 0.0288043562322855, + 0.002723562764003873, + 0.0020589111372828484, + 0.004087436944246292, + 0.004133986309170723, + 0.0016973897581920028, + 0.0036664761137217283, + 0.0035932091996073723, + 0.0040922630578279495, + 0.01514870673418045, + 0.01311582513153553, + 0.028885627165436745, + 0.026451053097844124, + 0.011806466616690159, + 0.019607296213507652, + 0.026340825483202934, + 0.021720154210925102 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.1270897388458252 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019471748964861035 + ], + "opposite_logprob_tokens": [ + -6.251947402954102 + ], + "logprob_reward": -0.0019471748964861035, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019471748964861035, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10526096075773239, + 0.3268296718597412, + 0.02774071879684925, + 0.05073942616581917, + 0.04561846703290939, + 0.10457900166511536, + 0.09276366978883743, + 0.09451711922883987, + 0.04324713349342346, + 0.10870371013879776 + ], + "credit_map": [ + 0.008761480450630188, + 0.007747257128357887, + 0.01398465409874916, + 0.01077057421207428, + 0.01003049872815609, + 0.01671569235622883, + 0.01893492043018341, + 0.01831589639186859, + 0.01779117062687874, + 0.00806153379380703, + 0.09885416179895401, + 0.012470525689423084, + 0.01448545791208744, + 0.12206284701824188, + 0.019382573664188385, + 0.03372141718864441, + 0.0022562346421182156, + 0.0013803687179461122, + 0.0037859862204641104, + 0.0022844099439680576, + 0.0018390159821137786, + 0.004495825152844191, + 0.003604327095672488, + 0.00809455569833517, + 0.0037617795169353485, + 0.0022771060466766357, + 0.010280151851475239, + 0.003876931732520461, + 0.003011804074048996, + 0.01332059595733881, + 0.005555444397032261, + 0.008655615150928497, + 0.00519222067669034, + 0.0021814468782395124, + 0.006793971173465252, + 0.0045554316602647305, + 0.002817348577082157, + 0.007573576644062996, + 0.00633343867957592, + 0.01017103623598814, + 0.007711898069828749, + 0.005897012073546648, + 0.01892269402742386, + 0.010078764520585537, + 0.007877162657678127, + 0.020490340888500214, + 0.01168956607580185, + 0.021911580115556717, + 0.008420037105679512, + 0.006023974623531103, + 0.014507975429296494, + 0.009717646986246109, + 0.007171753328293562, + 0.017393939197063446, + 0.011297764256596565, + 0.01823059469461441, + 0.008013743907213211, + 0.005024844314903021, + 0.017880572006106377, + 0.00970698893070221, + 0.005754705984145403, + 0.022495277225971222, + 0.010559247806668282, + 0.015081747435033321, + 0.004013717174530029, + 0.0038141831755638123, + 0.006272953934967518, + 0.006359291262924671, + 0.0037902442272752523, + 0.006118914112448692, + 0.006171293556690216, + 0.006706542801111937, + 0.01286410167813301, + 0.005320089869201183, + 0.014473885297775269, + 0.012164339423179626, + 0.0066872090101242065, + 0.017392128705978394, + 0.016320141032338142, + 0.023481830954551697 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0019471748964861035 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the video shot from a low-angle perspective?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the camera static without any movement?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure wearing a vintage black tuxedo with a shiny satin lapel and a crisp white shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the figure's face obscured by a pure white, owl-like porcelain mask with two large, round, black, hollow eyeholes?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the figure's head tilt slowly to one side while their right hand, in a white glove, rises gracefully?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0549592599272728 + ], + "opposite_logprob_tokens": [ + -2.929959297180176 + ], + "logprob_reward": -0.0549592599272728, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.0549592599272728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12765398621559143, + 0.1926816701889038, + 0.08953601866960526, + 0.06475226581096649, + 0.07633811235427856, + 0.061990078538656235, + 0.07072806358337402, + 0.03327200561761856, + 0.1149694100022316, + 0.1680784374475479 + ], + "credit_map": [ + 0.009392481297254562, + 0.008691202849149704, + 0.016684534028172493, + 0.014706236310303211, + 0.012517673894762993, + 0.018125740811228752, + 0.02163606695830822, + 0.025900039821863174, + 0.010569170117378235, + 0.005264803767204285, + 0.057746514678001404, + 0.008551152423024178, + 0.00811498612165451, + 0.06971468031406403, + 0.01360957883298397, + 0.019110776484012604, + 0.0055153085850179195, + 0.002881113439798355, + 0.026121901348233223, + 0.0051286290399730206, + 0.004073128569871187, + 0.031232891604304314, + 0.00596556207165122, + 0.008617483079433441, + 0.006232379004359245, + 0.0038985605351626873, + 0.00965474359691143, + 0.007933920249342918, + 0.007307274267077446, + 0.00946890190243721, + 0.009997081011533737, + 0.01025940291583538, + 0.005185327492654324, + 0.0031250054016709328, + 0.01882818713784218, + 0.007862204685807228, + 0.005242433864623308, + 0.020238015800714493, + 0.007338637486100197, + 0.008518296293914318, + 0.004479641560465097, + 0.0031186193227767944, + 0.011550232768058777, + 0.006615474820137024, + 0.004042680840939283, + 0.013941311277449131, + 0.006423828657716513, + 0.011818287894129753, + 0.006750300992280245, + 0.003950841259211302, + 0.01285059005022049, + 0.010829362086951733, + 0.004049126990139484, + 0.014281976968050003, + 0.007023273967206478, + 0.010992586612701416, + 0.0029285151977092028, + 0.0017552507342770696, + 0.004797336645424366, + 0.0053570265881717205, + 0.0017000947846099734, + 0.0062552825547754765, + 0.0057401699014008045, + 0.004738329444080591, + 0.014432280324399471, + 0.009081804193556309, + 0.016250312328338623, + 0.015105696395039558, + 0.009972008876502514, + 0.020430583506822586, + 0.014054288156330585, + 0.015642432495951653, + 0.020558806136250496, + 0.00935203768312931, + 0.024897653609514236, + 0.021325288340449333, + 0.011992085725069046, + 0.0243581123650074, + 0.024079320952296257, + 0.03151513263583183 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 0 + ], + "selected_reward": -0.0549592599272728 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_tensor([2444, 2444])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_tensor([2444, 2444])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..5f18b98bd3f15dc622776147e0e82cb23bcf3b3f --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_tensor([2444, 2444])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mesmerizing medium close-up, rendered in ultra-detailed 8K, captures a profoundly serene scene in the abyssal depths of the ocean. The background is a boundless expanse of pure black, broken only by sparse, slowly drifting marine snow, creating a sense of immense depth and isolation. A massive, translucent jellyfish hovers perfectly in the center of the frame, its bell, a delicate glassy dome, pulsating with a slow, hypnotic rhythm, revealing a subtle, intricate network of its internal structure. From beneath the bell, its countless long tentacles begin to calmly unfurl, movements gracefully unfolding with the deliberate care of a ballet. As they unfold, a vibrant, otherworldly bioluminescence flows through them like electricity through a filament, causing them to glow in ever-changing hues of sapphire and ethereal violet. These threadlike tentacles extend downward, trailing into the darkness, their movement creating gentle, swaying patterns in the void. The creature, the sole source of light, casts a soft, cool glow that illuminates the waters around it, creating a stark and beautiful contrast with the abyss. The atmosphere is filled with quiet mystery and profound awe, and the extreme clarity of 8K resolution makes every delicate filament and pulse of light breathtakingly tangible.", + "seed": 31568, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_2444_seed_31568_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06724456259980797, + "vlm_logprob": -0.06724456259980797, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.06724456259980797, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.031549323350191116 + ], + "opposite_logprob_tokens": [ + -3.5315492153167725 + ], + "logprob_reward": -0.031549323350191116, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.031549323350191116, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09690021723508835, + 0.08576864749193192, + 0.02749188244342804, + 0.08706533163785934, + 0.2835264801979065, + 0.03423842415213585, + 0.23642310500144958, + 0.05391213297843933, + 0.025026824325323105, + 0.06964699923992157 + ], + "credit_map": [ + 0.00893046148121357, + 0.008068466559052467, + 0.0170611459761858, + 0.015235992148518562, + 0.006141286808997393, + 0.01268698088824749, + 0.012415917590260506, + 0.0163599643856287, + 0.008037890307605267, + 0.006413209717720747, + 0.009826058521866798, + 0.013795043341815472, + 0.005972904618829489, + 0.012796858325600624, + 0.015784740447998047, + 0.013141937553882599, + 0.0028464244678616524, + 0.002249763812869787, + 0.0032058157958090305, + 0.005348817445337772, + 0.0029309852980077267, + 0.0036331170704215765, + 0.004079464357346296, + 0.003197496524080634, + 0.006051050499081612, + 0.003539719618856907, + 0.008671892806887627, + 0.0067642140202224255, + 0.003776291850954294, + 0.026849400252103806, + 0.006755704991519451, + 0.024657052010297775, + 0.019960710778832436, + 0.019660843536257744, + 0.04435640200972557, + 0.034054700285196304, + 0.02243521809577942, + 0.06343742460012436, + 0.03235992416739464, + 0.04726126417517662, + 0.008307798765599728, + 0.0030959395226091146, + 0.005882980301976204, + 0.004581374581903219, + 0.0024158069863915443, + 0.0025341911241412163, + 0.003447688417509198, + 0.003972644452005625, + 0.013483360409736633, + 0.014561360701918602, + 0.05814201757311821, + 0.017300168052315712, + 0.01819269172847271, + 0.07245051860809326, + 0.017743222415447235, + 0.024549778550863266, + 0.008122129365801811, + 0.005170924589037895, + 0.008367774076759815, + 0.0038668557535856962, + 0.0075605688616633415, + 0.005098328925669193, + 0.005094176158308983, + 0.010631375014781952, + 0.002575376071035862, + 0.002253106329590082, + 0.0037749246694147587, + 0.00321502098813653, + 0.0022107611875981092, + 0.0024668402038514614, + 0.0045188795775175095, + 0.004011915531009436, + 0.011097216047346592, + 0.012528032064437866, + 0.012157348915934563, + 0.005030907690525055, + 0.004858987871557474, + 0.005950528662651777, + 0.005752072669565678, + 0.012271903455257416 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.031549323350191116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028305486775934696 + ], + "opposite_logprob_tokens": [ + -5.877830505371094 + ], + "logprob_reward": -0.0028305486775934696, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028305486775934696, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08259904384613037, + 0.04236462712287903, + 0.055421292781829834, + 0.10096260160207748, + 0.10670099407434464, + 0.17667156457901, + 0.3060634136199951, + 0.028117073699831963, + 0.0484333299100399, + 0.05266613885760307 + ], + "credit_map": [ + 0.008255152963101864, + 0.005803150124847889, + 0.011236698366701603, + 0.011263862252235413, + 0.006895756348967552, + 0.012899959459900856, + 0.008317834697663784, + 0.01792662777006626, + 0.002213660627603531, + 0.0036430649925023317, + 0.006077980622649193, + 0.005794222000986338, + 0.004570723511278629, + 0.006419121753424406, + 0.007804140914231539, + 0.005841708276420832, + 0.0092341098934412, + 0.006232876796275377, + 0.009101186878979206, + 0.006060454994440079, + 0.005152542144060135, + 0.008727838285267353, + 0.004916890989989042, + 0.0059953900054097176, + 0.013361730612814426, + 0.0061644441448152065, + 0.016009436920285225, + 0.008237895555794239, + 0.00780031643807888, + 0.019704723730683327, + 0.012509280815720558, + 0.017174769192934036, + 0.007783535402268171, + 0.00638537434861064, + 0.025154946371912956, + 0.008234457112848759, + 0.007669947575777769, + 0.016526296734809875, + 0.008160995319485664, + 0.026785437017679214, + 0.04057464003562927, + 0.0207827165722847, + 0.02725910022854805, + 0.016255363821983337, + 0.01954677142202854, + 0.015677014365792274, + 0.016137437894940376, + 0.02043852023780346, + 0.017939964309334755, + 0.010555724613368511, + 0.09067000448703766, + 0.01564331166446209, + 0.020613893866539, + 0.11361592262983322, + 0.017047757282853127, + 0.01997680403292179, + 0.0029984035063534975, + 0.0033679059706628323, + 0.004848149139434099, + 0.0023672163952142, + 0.0027647712267935276, + 0.002567804418504238, + 0.002911172341555357, + 0.006291650701314211, + 0.004176795016974211, + 0.004520803689956665, + 0.01182994432747364, + 0.005066162906587124, + 0.00440495228394866, + 0.0033974468242377043, + 0.004297258332371712, + 0.010739962570369244, + 0.007339350413531065, + 0.013324566185474396, + 0.005539111793041229, + 0.003437301842495799, + 0.004112142603844404, + 0.005232820753008127, + 0.00543198361992836, + 0.008248856291174889 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.0028305486775934696 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370397910475731 + ], + "opposite_logprob_tokens": [ + -3.4087040424346924 + ], + "logprob_reward": -0.03370397910475731, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370397910475731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04121068865060806, + 0.032099559903144836, + 0.03855503350496292, + 0.10061398893594742, + 0.04309724643826485, + 0.1042797714471817, + 0.4475208520889282, + 0.04802357405424118, + 0.10425691306591034, + 0.04034237563610077 + ], + "credit_map": [ + 0.0029828003607690334, + 0.0031295225489884615, + 0.009040111675858498, + 0.007737860083580017, + 0.0028172810561954975, + 0.00529724545776844, + 0.00522942328825593, + 0.004976445809006691, + 0.003743199398741126, + 0.0022327147889882326, + 0.0038220451679080725, + 0.0038554989732801914, + 0.0021725078113377094, + 0.005323284305632114, + 0.004965563304722309, + 0.005984748713672161, + 0.009210287593305111, + 0.003111599711701274, + 0.004111809656023979, + 0.00316275074146688, + 0.0022881978657096624, + 0.006387144327163696, + 0.0035678979475051165, + 0.0067153447307646275, + 0.013338325545191765, + 0.005291193723678589, + 0.020100221037864685, + 0.011318397708237171, + 0.006375658791512251, + 0.016230212524533272, + 0.01474444754421711, + 0.013215530663728714, + 0.003385702846571803, + 0.0025771227665245533, + 0.009018050506711006, + 0.003928174264729023, + 0.0027188509702682495, + 0.005679672118276358, + 0.003993486054241657, + 0.011796188540756702, + 0.028590288013219833, + 0.009212258271872997, + 0.015447312034666538, + 0.009599697776138783, + 0.009695244953036308, + 0.011496816761791706, + 0.008468768559396267, + 0.011769385077059269, + 0.025959404185414314, + 0.01612764410674572, + 0.12953057885169983, + 0.028157375752925873, + 0.029300251975655556, + 0.16294120252132416, + 0.025674281641840935, + 0.02983010560274124, + 0.004910917486995459, + 0.0041639250703155994, + 0.008079546503722668, + 0.004529279191046953, + 0.00486328499391675, + 0.004252045415341854, + 0.006457152776420116, + 0.010767422616481781, + 0.013806991279125214, + 0.010203740559518337, + 0.01800977997481823, + 0.018644215539097786, + 0.009490475058555603, + 0.007252799812704325, + 0.011855208314955235, + 0.014993706718087196, + 0.005335626192390919, + 0.010084692388772964, + 0.005466504953801632, + 0.002588655799627304, + 0.0029304816853255033, + 0.003533850656822324, + 0.003321888158097863, + 0.007080676965415478 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 8 + ], + "selected_reward": -0.03370397910475731 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069071367383003 + ], + "opposite_logprob_tokens": [ + -4.141068935394287 + ], + "logprob_reward": -0.016069071367383003, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069071367383003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16316449642181396, + 0.07093405723571777, + 0.04112975299358368, + 0.04058922827243805, + 0.1656123250722885, + 0.018619373440742493, + 0.3842848539352417, + 0.02407638169825077, + 0.06480194628238678, + 0.026787584647536278 + ], + "credit_map": [ + 0.017885003238916397, + 0.01570538617670536, + 0.03250482678413391, + 0.020581956952810287, + 0.00974281970411539, + 0.02136530540883541, + 0.020404240116477013, + 0.02497495897114277, + 0.009064388461411, + 0.007759374100714922, + 0.008861757814884186, + 0.01013493724167347, + 0.006542888469994068, + 0.010173636488616467, + 0.009464814327657223, + 0.008932258933782578, + 0.007056765258312225, + 0.004472644999623299, + 0.004869221244007349, + 0.006634480785578489, + 0.0035844878293573856, + 0.004819002468138933, + 0.005814838223159313, + 0.0038783142808824778, + 0.005508439615368843, + 0.003014872083440423, + 0.005675935186445713, + 0.005061467178165913, + 0.003393175546079874, + 0.005844309460371733, + 0.006861793342977762, + 0.0052292319014668465, + 0.010792510583996773, + 0.009549777954816818, + 0.03448079153895378, + 0.013315724208950996, + 0.010369661264121532, + 0.030138423666357994, + 0.015403174795210361, + 0.04156225174665451, + 0.0029359818436205387, + 0.002066106768324971, + 0.0028474987484514713, + 0.0027183732017874718, + 0.0020875188056379557, + 0.0017607529880478978, + 0.0021094726398587227, + 0.0020936692599207163, + 0.02279719151556492, + 0.016178850084543228, + 0.11088690161705017, + 0.019689850509166718, + 0.028561396524310112, + 0.13941140472888947, + 0.019587628543376923, + 0.027171630412340164, + 0.002315213205292821, + 0.003087275195866823, + 0.003186261048540473, + 0.0027333602774888277, + 0.002385796047747135, + 0.002425275044515729, + 0.0034111919812858105, + 0.004532009828835726, + 0.006288417149335146, + 0.005501577164977789, + 0.01444383803755045, + 0.008134312927722931, + 0.004942628554999828, + 0.005953195504844189, + 0.006974660791456699, + 0.012563318014144897, + 0.002545269438996911, + 0.003420707071200013, + 0.004653343465179205, + 0.0025841842871159315, + 0.0029105767607688904, + 0.002530661877244711, + 0.003041383810341358, + 0.005101459100842476 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 4, + 0 + ], + "selected_reward": -0.016069071367383003 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.252069890499115 + ], + "opposite_logprob_tokens": [ + -1.5020699501037598 + ], + "logprob_reward": -0.252069890499115, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.252069890499115, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0968375951051712, + 0.04497801139950752, + 0.09867555648088455, + 0.059385478496551514, + 0.13084642589092255, + 0.1882733851671219, + 0.1348681002855301, + 0.10549032688140869, + 0.045495085418224335, + 0.09515003114938736 + ], + "credit_map": [ + 0.007629705127328634, + 0.00916903093457222, + 0.017829984426498413, + 0.01589982956647873, + 0.006428051739931107, + 0.012699245475232601, + 0.013872976414859295, + 0.013308770954608917, + 0.004007465671747923, + 0.0040527223609387875, + 0.005293817725032568, + 0.006787989288568497, + 0.0035006359685212374, + 0.006601485423743725, + 0.008052326738834381, + 0.006681566126644611, + 0.023293189704418182, + 0.009023563005030155, + 0.013743101619184017, + 0.01061330083757639, + 0.006874203681945801, + 0.014336623251438141, + 0.00955501664429903, + 0.011236566118896008, + 0.0064475880935788155, + 0.003103467868641019, + 0.008840047754347324, + 0.008823677897453308, + 0.003946537617594004, + 0.011553982272744179, + 0.0062437504529953, + 0.01042642816901207, + 0.009394395165145397, + 0.00917139369994402, + 0.02353731542825699, + 0.00958335679024458, + 0.01044037938117981, + 0.026061037555336952, + 0.010339017026126385, + 0.032319534569978714, + 0.039003174751996994, + 0.017280790954828262, + 0.02815551497042179, + 0.027187054976820946, + 0.019865406677126884, + 0.017575666308403015, + 0.019796568900346756, + 0.01940920762717724, + 0.010816516354680061, + 0.016509125009179115, + 0.017005378380417824, + 0.020539090037345886, + 0.014391036704182625, + 0.01733625866472721, + 0.01822873018682003, + 0.020041966810822487, + 0.013951665721833706, + 0.014491451904177666, + 0.013391830027103424, + 0.010497722774744034, + 0.01495114155113697, + 0.011733132414519787, + 0.012007583864033222, + 0.014465801417827606, + 0.006426693871617317, + 0.0031530316919088364, + 0.008022570051252842, + 0.005056299734860659, + 0.0038434574380517006, + 0.003269999986514449, + 0.004847237374633551, + 0.010875795036554337, + 0.013001803308725357, + 0.02491454966366291, + 0.010667642578482628, + 0.0063807000406086445, + 0.007003881968557835, + 0.008401352912187576, + 0.009488175623118877, + 0.015291922725737095 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 6, + 4 + ], + "selected_reward": -0.252069890499115 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mesmerizing medium close-up, rendered in ultra-detailed 8K, captures a profoundly serene scene in the abyssal depths of the ocean. The background is a boundless expanse of pure black, broken only by sparse, slowly drifting marine snow, creating a sense of immense depth and isolation. A massive, translucent jellyfish hovers perfectly in the center of the frame, its bell, a delicate glassy dome, pulsating with a slow, hypnotic rhythm, revealing a subtle, intricate network of its internal structure. From beneath the bell, its countless long tentacles begin to calmly unfurl, movements gracefully unfolding with the deliberate care of a ballet. As they unfold, a vibrant, otherworldly bioluminescence flows through them like electricity through a filament, causing them to glow in ever-changing hues of sapphire and ethereal violet. These threadlike tentacles extend downward, trailing into the darkness, their movement creating gentle, swaying patterns in the void. The creature, the sole source of light, casts a soft, cool glow that illuminates the waters around it, creating a stark and beautiful contrast with the abyss. The atmosphere is filled with quiet mystery and profound awe, and the extreme clarity of 8K resolution makes every delicate filament and pulse of light breathtakingly tangible.", + "seed": 31569, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_2444_seed_31569_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.052218360966071485, + "vlm_logprob": -0.052218360966071485, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.052218360966071485, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07319571077823639 + ], + "opposite_logprob_tokens": [ + -2.698195695877075 + ], + "logprob_reward": -0.07319571077823639, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07319571077823639, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1250770390033722, + 0.1902552843093872, + 0.10228575766086578, + 0.11158758401870728, + 0.03888291120529175, + 0.20682747662067413, + 0.0230394434183836, + 0.034768760204315186, + 0.08953335136175156, + 0.07774238288402557 + ], + "credit_map": [ + 0.01975250616669655, + 0.010021616704761982, + 0.018738336861133575, + 0.011830955743789673, + 0.01094349566847086, + 0.01885879412293434, + 0.011773084290325642, + 0.023158244788646698, + 0.030530771240592003, + 0.014064652845263481, + 0.016078462824225426, + 0.01716023124754429, + 0.013785096816718578, + 0.034381527453660965, + 0.01322963461279869, + 0.0510249137878418, + 0.020550724118947983, + 0.006188109051436186, + 0.0129699120298028, + 0.01081336010247469, + 0.010370624251663685, + 0.015848049893975258, + 0.009261075407266617, + 0.016283905133605003, + 0.012114427983760834, + 0.0101853646337986, + 0.024823950603604317, + 0.013592250645160675, + 0.012406954541802406, + 0.015058574266731739, + 0.010205660946667194, + 0.013200392946600914, + 0.005724424496293068, + 0.003950418438762426, + 0.007446584291756153, + 0.0033563021570444107, + 0.004055112134665251, + 0.005507446359843016, + 0.003315218258649111, + 0.005527408327907324, + 0.03819404914975166, + 0.014671697281301022, + 0.05151689797639847, + 0.018232271075248718, + 0.01248781755566597, + 0.03241406008601189, + 0.012091533280909061, + 0.02721915952861309, + 0.0043170335702598095, + 0.002239491790533066, + 0.004077471327036619, + 0.00191595742944628, + 0.0018873788649216294, + 0.003223028965294361, + 0.002199492184445262, + 0.0031795892864465714, + 0.006423735059797764, + 0.003085290314629674, + 0.006944494787603617, + 0.003466355614364147, + 0.0020927891600877047, + 0.00435169180855155, + 0.0028373536188155413, + 0.005567049607634544, + 0.01720939576625824, + 0.005483667831867933, + 0.006255988962948322, + 0.005965987220406532, + 0.005991328042000532, + 0.019635150209069252, + 0.010010386817157269, + 0.018981439992785454, + 0.02361522987484932, + 0.0028847469948232174, + 0.004518883302807808, + 0.003465192625299096, + 0.003213130170479417, + 0.016921181231737137, + 0.0037042973563075066, + 0.019419722259044647 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.07319571077823639 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00362905184738338 + ], + "opposite_logprob_tokens": [ + -5.628629207611084 + ], + "logprob_reward": -0.00362905184738338, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.00362905184738338, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05728782340884209, + 0.11288820952177048, + 0.033959902822971344, + 0.2940021753311157, + 0.1087801456451416, + 0.07789010554552078, + 0.032076798379421234, + 0.13998939096927643, + 0.07715396583080292, + 0.06597152352333069 + ], + "credit_map": [ + 0.007947936654090881, + 0.0034780949354171753, + 0.0057326266542077065, + 0.00417641457170248, + 0.005241672974079847, + 0.01021306961774826, + 0.004955265671014786, + 0.015542744658887386, + 0.01732955127954483, + 0.008763290010392666, + 0.00826583057641983, + 0.009526635520160198, + 0.007963771000504494, + 0.020062079653143883, + 0.008966305293142796, + 0.032010745257139206, + 0.0053228153847157955, + 0.002539553912356496, + 0.005614100489765406, + 0.004409837070852518, + 0.0035932271275669336, + 0.004402526188641787, + 0.004078928846865892, + 0.003998912870883942, + 0.033655907958745956, + 0.038831066340208054, + 0.043661195784807205, + 0.03433626890182495, + 0.043590057641267776, + 0.042013395577669144, + 0.03285831958055496, + 0.02505597099661827, + 0.010509386658668518, + 0.021969478577375412, + 0.010482147336006165, + 0.014040978625416756, + 0.014518341049551964, + 0.012869812548160553, + 0.012732247821986675, + 0.01165774930268526, + 0.012903383001685143, + 0.006678733974695206, + 0.008344459347426891, + 0.0069007957354187965, + 0.005963320843875408, + 0.011103738099336624, + 0.007021884433925152, + 0.018973790109157562, + 0.0030059393029659986, + 0.004245249554514885, + 0.0038441403303295374, + 0.003840648103505373, + 0.00497002899646759, + 0.004109797067940235, + 0.0038757871370762587, + 0.0041852109134197235, + 0.01642005331814289, + 0.03572337329387665, + 0.015890315175056458, + 0.01654011942446232, + 0.015791917219758034, + 0.015276598744094372, + 0.012320311740040779, + 0.012026693671941757, + 0.011135688982903957, + 0.006804176140576601, + 0.012612496502697468, + 0.009975651279091835, + 0.007195944897830486, + 0.009732485748827457, + 0.01190035231411457, + 0.007797172758728266, + 0.014987458474934101, + 0.004996504168957472, + 0.007422587368637323, + 0.0055279238149523735, + 0.003965791314840317, + 0.011430810205638409, + 0.005256120581179857, + 0.012384332716464996 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.00362905184738338 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018194211646914482 + ], + "opposite_logprob_tokens": [ + -4.018194198608398 + ], + "logprob_reward": -0.018194211646914482, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018194211646914482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09424901008605957, + 0.3164001405239105, + 0.047718703746795654, + 0.11381179839372635, + 0.0719645544886589, + 0.046606048941612244, + 0.024556126445531845, + 0.08362416177988052, + 0.11365332454442978, + 0.08741604536771774 + ], + "credit_map": [ + 0.011570415459573269, + 0.005746184382587671, + 0.009543359279632568, + 0.009590964764356613, + 0.006782598327845335, + 0.01396203227341175, + 0.012553754262626171, + 0.02449970878660679, + 0.038703553378582, + 0.023148484528064728, + 0.03453315794467926, + 0.03596549108624458, + 0.019735554233193398, + 0.05032087862491608, + 0.03706172853708267, + 0.07693131268024445, + 0.00720735639333725, + 0.004091778304427862, + 0.010229106992483139, + 0.0052050077356398106, + 0.003931084647774696, + 0.006399987265467644, + 0.005028429441154003, + 0.005625957623124123, + 0.014038031920790672, + 0.0073445867747068405, + 0.03149550035595894, + 0.013363199308514595, + 0.011886331252753735, + 0.016781356185674667, + 0.008902850560843945, + 0.009999950416386127, + 0.008723419159650803, + 0.0060415007174015045, + 0.015330159105360508, + 0.00880949106067419, + 0.008984079584479332, + 0.009985839948058128, + 0.006331321317702532, + 0.0077587515115737915, + 0.0054930695332586765, + 0.00320440367795527, + 0.009818664751946926, + 0.0067334105260670185, + 0.004171981010586023, + 0.006006704177707434, + 0.006139360833913088, + 0.005038461647927761, + 0.0022279894910752773, + 0.003771901596337557, + 0.002962350146844983, + 0.00291007780469954, + 0.003568968502804637, + 0.0030479468405246735, + 0.002805279567837715, + 0.0032616148237138987, + 0.014637785032391548, + 0.00596422515809536, + 0.01614103652536869, + 0.008903760462999344, + 0.005734850186854601, + 0.010027335025370121, + 0.007318818010389805, + 0.014896360225975513, + 0.0217143464833498, + 0.01461751852184534, + 0.01306448969990015, + 0.01437662448734045, + 0.009866063483059406, + 0.012070891447365284, + 0.009307284839451313, + 0.01863611489534378, + 0.014127777889370918, + 0.005052107851952314, + 0.02108902484178543, + 0.008056907914578915, + 0.006277656648308039, + 0.01469839084893465, + 0.0064455862157046795, + 0.011668598279356956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 8 + ], + "selected_reward": -0.018194211646914482 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023288190364837646 + ], + "opposite_logprob_tokens": [ + -3.7732882499694824 + ], + "logprob_reward": -0.023288190364837646, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023288190364837646, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05810650810599327, + 0.3698895275592804, + 0.05834750831127167, + 0.10095516592264175, + 0.039177075028419495, + 0.035276345908641815, + 0.022411145269870758, + 0.058115437626838684, + 0.10422132909297943, + 0.1535000056028366 + ], + "credit_map": [ + 0.00936329085379839, + 0.004125726409256458, + 0.006725807674229145, + 0.0056029679253697395, + 0.004784977529197931, + 0.009107647463679314, + 0.005950035527348518, + 0.01244605053216219, + 0.055513896048069, + 0.026543062180280685, + 0.03791450336575508, + 0.03577931225299835, + 0.023380711674690247, + 0.06454797834157944, + 0.02910725399851799, + 0.09710279107093811, + 0.010594310238957405, + 0.0027839220128953457, + 0.013151856139302254, + 0.0048872255720198154, + 0.0033772059250622988, + 0.009108680300414562, + 0.003366026096045971, + 0.011078275740146637, + 0.01190267875790596, + 0.007207090966403484, + 0.0187709778547287, + 0.012015404179692268, + 0.00835306104272604, + 0.015104160644114017, + 0.008751950226724148, + 0.018849827349185944, + 0.005732426885515451, + 0.00455989595502615, + 0.006435762625187635, + 0.003690236248075962, + 0.0041005853563547134, + 0.005483732558786869, + 0.0030846272129565477, + 0.0060898070223629475, + 0.0049342019483447075, + 0.0036025578156113625, + 0.0036471260245889425, + 0.004510752856731415, + 0.004847005940973759, + 0.004637935198843479, + 0.004528870340436697, + 0.0045678927563130856, + 0.0036276360042393208, + 0.0021027366165071726, + 0.0018500470323488116, + 0.0017800132045522332, + 0.002956212731078267, + 0.0030728820711374283, + 0.0021666469983756542, + 0.004854967817664146, + 0.007148846052587032, + 0.013895833864808083, + 0.00512508163228631, + 0.006108167581260204, + 0.007109724450856447, + 0.006572555750608444, + 0.006041150540113449, + 0.0061140707693994045, + 0.014623760245740414, + 0.01022342313081026, + 0.010553475469350815, + 0.011999832466244698, + 0.009493457153439522, + 0.023832477629184723, + 0.008665313944220543, + 0.014829587191343307, + 0.04781517758965492, + 0.005293286871165037, + 0.010678975842893124, + 0.0068215117789804935, + 0.005968357902020216, + 0.03192583844065666, + 0.013302641920745373, + 0.0316942036151886 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.023288190364837646 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14278464019298553 + ], + "opposite_logprob_tokens": [ + -2.017784595489502 + ], + "logprob_reward": -0.14278464019298553, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14278464019298553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4370581805706024, + 0.10601154714822769, + 0.05312546342611313, + 0.08090315759181976, + 0.06231306865811348, + 0.05207633972167969, + 0.02567923069000244, + 0.02189125306904316, + 0.1032000333070755, + 0.05774164944887161 + ], + "credit_map": [ + 0.06668482720851898, + 0.029271699488162994, + 0.03713351860642433, + 0.033952753990888596, + 0.021173425018787384, + 0.08321063220500946, + 0.026088040322065353, + 0.13954328000545502, + 0.015832193195819855, + 0.006079214625060558, + 0.009420391172170639, + 0.00908919982612133, + 0.006821773014962673, + 0.01928934082388878, + 0.011724339798092842, + 0.027755096554756165, + 0.006078517064452171, + 0.004038423299789429, + 0.011121261864900589, + 0.006868628319352865, + 0.003905603429302573, + 0.007992925122380257, + 0.005762808490544558, + 0.007357292342931032, + 0.00865962915122509, + 0.00590544892475009, + 0.0196338202804327, + 0.008932906202971935, + 0.008708166889846325, + 0.011239415965974331, + 0.0072708697989583015, + 0.010552899911999702, + 0.011393588967621326, + 0.0048398664221167564, + 0.009183870628476143, + 0.005295452196151018, + 0.004504032898694277, + 0.009723905473947525, + 0.004761045332998037, + 0.01261130440980196, + 0.007501004729419947, + 0.005030149593949318, + 0.006394345313310623, + 0.006676050368696451, + 0.006393034942448139, + 0.006943993270397186, + 0.0073536466807127, + 0.005784118548035622, + 0.004055623896420002, + 0.0029667166527360678, + 0.005312321241945028, + 0.002013640943914652, + 0.002032925607636571, + 0.003047333564609289, + 0.0023329562973231077, + 0.00391771225258708, + 0.00301803438924253, + 0.002528686309233308, + 0.0025890967808663845, + 0.002549815457314253, + 0.0016071805730462074, + 0.00288452603854239, + 0.002868046285584569, + 0.003845867235213518, + 0.019197886809706688, + 0.010672545060515404, + 0.013503036461770535, + 0.013794251717627048, + 0.01018148846924305, + 0.012296863831579685, + 0.00940143596380949, + 0.014152523130178452, + 0.01107738632708788, + 0.003483871230855584, + 0.012100164778530598, + 0.008502253331243992, + 0.0035524361301213503, + 0.006275173742324114, + 0.004612288903445005, + 0.008138075470924377 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.14278464019298553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mesmerizing medium close-up, rendered in ultra-detailed 8K, captures a profoundly serene scene in the abyssal depths of the ocean. The background is a boundless expanse of pure black, broken only by sparse, slowly drifting marine snow, creating a sense of immense depth and isolation. A massive, translucent jellyfish hovers perfectly in the center of the frame, its bell, a delicate glassy dome, pulsating with a slow, hypnotic rhythm, revealing a subtle, intricate network of its internal structure. From beneath the bell, its countless long tentacles begin to calmly unfurl, movements gracefully unfolding with the deliberate care of a ballet. As they unfold, a vibrant, otherworldly bioluminescence flows through them like electricity through a filament, causing them to glow in ever-changing hues of sapphire and ethereal violet. These threadlike tentacles extend downward, trailing into the darkness, their movement creating gentle, swaying patterns in the void. The creature, the sole source of light, casts a soft, cool glow that illuminates the waters around it, creating a stark and beautiful contrast with the abyss. The atmosphere is filled with quiet mystery and profound awe, and the extreme clarity of 8K resolution makes every delicate filament and pulse of light breathtakingly tangible.", + "seed": 31570, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_2444_seed_31570_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.02770384047180414, + "vlm_logprob": -0.02770384047180414, + "vlm_yes_no_margin": 4.125, + "vlm_token_credit": -0.02770384047180414, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.017016831785440445 + ], + "opposite_logprob_tokens": [ + -4.142016887664795 + ], + "logprob_reward": -0.017016831785440445, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.017016831785440445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23321756720542908, + 0.07551877200603485, + 0.048545513302087784, + 0.07964499294757843, + 0.026402410119771957, + 0.35592564940452576, + 0.029793228954076767, + 0.01431365218013525, + 0.04529261961579323, + 0.09134560078382492 + ], + "credit_map": [ + 0.0654325857758522, + 0.02380860224366188, + 0.045254092663526535, + 0.01914205588400364, + 0.018615927547216415, + 0.019329139962792397, + 0.017920667305588722, + 0.02371448092162609, + 0.019746292382478714, + 0.007021615281701088, + 0.015561528503894806, + 0.008859452791512012, + 0.005228525493294001, + 0.005837245378643274, + 0.005450404714792967, + 0.007813709788024426, + 0.012719666585326195, + 0.0034810202196240425, + 0.010203038342297077, + 0.004662923980504274, + 0.005018896888941526, + 0.004081832244992256, + 0.00417864928022027, + 0.004199486691504717, + 0.02159983478486538, + 0.0074577294290065765, + 0.015263045206665993, + 0.006358693353831768, + 0.008399652317166328, + 0.006253235042095184, + 0.00586222717538476, + 0.008450566790997982, + 0.006016017869114876, + 0.002704265061765909, + 0.004778040107339621, + 0.002905749948695302, + 0.0021984383929520845, + 0.0016172426985576749, + 0.0022974652238190174, + 0.003885191399604082, + 0.0788089781999588, + 0.020837072283029556, + 0.05991499498486519, + 0.05754063278436661, + 0.02295289933681488, + 0.03552835062146187, + 0.03213527798652649, + 0.04820743575692177, + 0.005782862193882465, + 0.0022417546715587378, + 0.004256755579262972, + 0.004734116606414318, + 0.0020622347947210073, + 0.0026177274994552135, + 0.0021262874361127615, + 0.005971490405499935, + 0.0025675545912235975, + 0.0009737843065522611, + 0.0023688157089054585, + 0.0015569308307021856, + 0.0018596870359033346, + 0.0011110001942142844, + 0.001313124899752438, + 0.0025627559516578913, + 0.011721626855432987, + 0.0035347274970263243, + 0.008319801650941372, + 0.006562903523445129, + 0.00319662201218307, + 0.002506287069991231, + 0.003067996818572283, + 0.006382653024047613, + 0.023490795865654945, + 0.009035471826791763, + 0.014840132556855679, + 0.013638856820762157, + 0.007397310808300972, + 0.006554069463163614, + 0.006416382733732462, + 0.00997258722782135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 9 + ], + "selected_reward": -0.017016831785440445 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017220443114638329 + ], + "opposite_logprob_tokens": [ + -6.3767218589782715 + ], + "logprob_reward": -0.0017220443114638329, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017220443114638329, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15003305673599243, + 0.135880708694458, + 0.1086135059595108, + 0.1431448757648468, + 0.04484468325972557, + 0.10278844088315964, + 0.0483587384223938, + 0.11149414628744125, + 0.024324970319867134, + 0.1305168718099594 + ], + "credit_map": [ + 0.035202547907829285, + 0.012071115896105766, + 0.02596900425851345, + 0.012211579829454422, + 0.022200169041752815, + 0.01153391133993864, + 0.01610218733549118, + 0.014742549508810043, + 0.018342340365052223, + 0.012364706955850124, + 0.018861453980207443, + 0.010025566443800926, + 0.03000962734222412, + 0.018021544441580772, + 0.016073059290647507, + 0.012182408943772316, + 0.022200807929039, + 0.006019446533173323, + 0.018950536847114563, + 0.008782860822975636, + 0.012598452158272266, + 0.006167559884488583, + 0.020684199407696724, + 0.013209639117121696, + 0.014465635642409325, + 0.010115639306604862, + 0.016637228429317474, + 0.010188400745391846, + 0.035763662308454514, + 0.012285604141652584, + 0.02394900657236576, + 0.01973969116806984, + 0.008453330025076866, + 0.003046195488423109, + 0.007218037266284227, + 0.004956125281751156, + 0.0060974787920713425, + 0.004733114503324032, + 0.004963883198797703, + 0.005376515444368124, + 0.021202918142080307, + 0.00609651068225503, + 0.01734299585223198, + 0.015213162638247013, + 0.006156934425234795, + 0.006632769946008921, + 0.007776626385748386, + 0.022366531193256378, + 0.007082680705934763, + 0.0034720683470368385, + 0.0064224544912576675, + 0.006104375701397657, + 0.004642900079488754, + 0.005979318171739578, + 0.0034440166782587767, + 0.011210927739739418, + 0.01218079961836338, + 0.011941755190491676, + 0.014704707078635693, + 0.010527273640036583, + 0.01755974441766739, + 0.0117996446788311, + 0.010597508400678635, + 0.02218271791934967, + 0.0044129518792033195, + 0.002034153323620558, + 0.003777212230488658, + 0.0034956305753439665, + 0.0027128520887345076, + 0.0028513246215879917, + 0.002097377786412835, + 0.002943467115983367, + 0.019334658980369568, + 0.01698225550353527, + 0.016319140791893005, + 0.015106643550097942, + 0.015910644084215164, + 0.015114814974367619, + 0.01930864341557026, + 0.012440071441233158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0017220443114638329 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370063751935959 + ], + "opposite_logprob_tokens": [ + -3.408700704574585 + ], + "logprob_reward": -0.03370063751935959, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370063751935959, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2163413017988205, + 0.07364637404680252, + 0.11335407197475433, + 0.11621396243572235, + 0.04508577287197113, + 0.11008945107460022, + 0.08468946069478989, + 0.13782675564289093, + 0.03941832855343819, + 0.06333454698324203 + ], + "credit_map": [ + 0.04506555199623108, + 0.01805436797440052, + 0.03398028761148453, + 0.017349543049931526, + 0.03463131934404373, + 0.017828738316893578, + 0.027492957189679146, + 0.02193855494260788, + 0.020781615749001503, + 0.004957869183272123, + 0.014956588856875896, + 0.007399357855319977, + 0.006862077862024307, + 0.005591304041445255, + 0.005843798629939556, + 0.0072537632659077644, + 0.02404727227985859, + 0.007971302606165409, + 0.01984494924545288, + 0.011628174223005772, + 0.015772314742207527, + 0.008924544788897038, + 0.014751189388334751, + 0.010414321906864643, + 0.016532279551029205, + 0.0093770707026124, + 0.016395414248108864, + 0.011530067771673203, + 0.014247233979403973, + 0.009595101699233055, + 0.017692340537905693, + 0.020844465121626854, + 0.01362144760787487, + 0.003235115436837077, + 0.008722599595785141, + 0.004469006787985563, + 0.0034773037768900394, + 0.003348264843225479, + 0.0037147074472159147, + 0.004497329704463482, + 0.021032191812992096, + 0.005887004546821117, + 0.0172265712171793, + 0.014656828716397285, + 0.0067804413847625256, + 0.009259216487407684, + 0.006862872280180454, + 0.028384322300553322, + 0.012937732972204685, + 0.0057955412194132805, + 0.011049930937588215, + 0.011438626796007156, + 0.009148038923740387, + 0.009348743595182896, + 0.0062030404806137085, + 0.018767807632684708, + 0.018226876854896545, + 0.00985453836619854, + 0.018363868817687035, + 0.01967693492770195, + 0.02554258517920971, + 0.016174521297216415, + 0.014786606654524803, + 0.015200811438262463, + 0.00810174085199833, + 0.0025909394025802612, + 0.005891750101000071, + 0.00512259965762496, + 0.0045919460244476795, + 0.0032881596125662327, + 0.004475145135074854, + 0.00535604590550065, + 0.007918450981378555, + 0.009163156151771545, + 0.009263267740607262, + 0.010500305332243443, + 0.0068609085865318775, + 0.006401026621460915, + 0.007460917811840773, + 0.005766516551375389 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 3 + ], + "selected_reward": -0.03370063751935959 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606954075396061 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.01606954075396061, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606954075396061, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1914912462234497, + 0.162005215883255, + 0.11205322295427322, + 0.16868290305137634, + 0.04355020448565483, + 0.05855224281549454, + 0.0451541468501091, + 0.11930947750806808, + 0.04754819720983505, + 0.05165313556790352 + ], + "credit_map": [ + 0.03608162701129913, + 0.016747744753956795, + 0.02925874851644039, + 0.016364475712180138, + 0.032222628593444824, + 0.022221878170967102, + 0.02057509310543537, + 0.018019042909145355, + 0.016480309888720512, + 0.014639459550380707, + 0.02082124911248684, + 0.0131392115727067, + 0.03752138465642929, + 0.02312120422720909, + 0.021521346643567085, + 0.014761043712496758, + 0.013130954466760159, + 0.006965713109821081, + 0.013590415939688683, + 0.012788306921720505, + 0.019516294822096825, + 0.007486640941351652, + 0.025830930098891258, + 0.012743952684104443, + 0.020329339429736137, + 0.012166055850684643, + 0.021892840042710304, + 0.013537653721868992, + 0.03258379176259041, + 0.01297033205628395, + 0.024057600647211075, + 0.03114529699087143, + 0.007015132810920477, + 0.00411461153998971, + 0.006959995254874229, + 0.00582039263099432, + 0.004561155568808317, + 0.0032704088371247053, + 0.004399115685373545, + 0.007409390062093735, + 0.014802590943872929, + 0.0037120359484106302, + 0.0093538174405694, + 0.008923961780965328, + 0.004852268844842911, + 0.0036071559879928827, + 0.0043215565383434296, + 0.00897885300219059, + 0.00953197106719017, + 0.0038960501551628113, + 0.008044272661209106, + 0.006576129235327244, + 0.0039056933019310236, + 0.00405595637857914, + 0.002849028678610921, + 0.006295050494372845, + 0.012066266499459743, + 0.011730694212019444, + 0.014355740509927273, + 0.010560710914433002, + 0.02686859853565693, + 0.019060028716921806, + 0.011198058724403381, + 0.01346938218921423, + 0.0071461740881204605, + 0.004426931496709585, + 0.0060871499590575695, + 0.008033082820475101, + 0.00497219106182456, + 0.00474992161616683, + 0.005008368287235498, + 0.007124376483261585, + 0.010306044481694698, + 0.003987771458923817, + 0.008372485637664795, + 0.00863380916416645, + 0.005225372966378927, + 0.0035068891011178493, + 0.00675706472247839, + 0.004863696172833443 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.01606954075396061 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001014798879623 + ], + "opposite_logprob_tokens": [ + -2.695010185241699 + ], + "logprob_reward": -0.07001014798879623, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001014798879623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09898760914802551, + 0.058087244629859924, + 0.0759434774518013, + 0.03537215664982796, + 0.034421004354953766, + 0.06558060646057129, + 0.06328172236680984, + 0.2967889606952667, + 0.10114818811416626, + 0.1703890711069107 + ], + "credit_map": [ + 0.020667975768446922, + 0.009010611101984978, + 0.016252608969807625, + 0.010364548303186893, + 0.01042889803647995, + 0.00735206576064229, + 0.011042464524507523, + 0.013868436217308044, + 0.009257495403289795, + 0.004664978012442589, + 0.00866588857024908, + 0.006209454033523798, + 0.008036848157644272, + 0.0070083048194646835, + 0.00764060765504837, + 0.006603667046874762, + 0.010494335554540157, + 0.005216020625084639, + 0.00989819597452879, + 0.00808593351393938, + 0.015821993350982666, + 0.006441863719373941, + 0.012616529129445553, + 0.007368605583906174, + 0.004837913438677788, + 0.0029008365236222744, + 0.005027851089835167, + 0.002836118219420314, + 0.005829345900565386, + 0.0029321101028472185, + 0.003830799600109458, + 0.007177181076258421, + 0.007024943362921476, + 0.0029638735577464104, + 0.005807659588754177, + 0.004585962742567062, + 0.003125963034108281, + 0.0031092155259102583, + 0.0037087886594235897, + 0.004094594623893499, + 0.014146878384053707, + 0.00433642603456974, + 0.011119361035525799, + 0.011437665671110153, + 0.00526414206251502, + 0.004658509511500597, + 0.005717042367905378, + 0.008900581859052181, + 0.008277440443634987, + 0.005964822135865688, + 0.008764993399381638, + 0.009631413966417313, + 0.007320593111217022, + 0.00729198195040226, + 0.005776749923825264, + 0.0102537265047431, + 0.03522935137152672, + 0.026958296075463295, + 0.04066406190395355, + 0.027498971670866013, + 0.056924063712358475, + 0.027588456869125366, + 0.02851719968020916, + 0.05340855196118355, + 0.013461816124618053, + 0.007238440681248903, + 0.014665912836790085, + 0.016561174765229225, + 0.015047122724354267, + 0.009636424481868744, + 0.013481048867106438, + 0.011056244373321533, + 0.025446830317378044, + 0.018905045464634895, + 0.023620229214429855, + 0.02625584602355957, + 0.019586948677897453, + 0.01702704466879368, + 0.01909203827381134, + 0.02045508846640587 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.07001014798879623 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mesmerizing medium close-up, rendered in ultra-detailed 8K, captures a profoundly serene scene in the abyssal depths of the ocean. The background is a boundless expanse of pure black, broken only by sparse, slowly drifting marine snow, creating a sense of immense depth and isolation. A massive, translucent jellyfish hovers perfectly in the center of the frame, its bell, a delicate glassy dome, pulsating with a slow, hypnotic rhythm, revealing a subtle, intricate network of its internal structure. From beneath the bell, its countless long tentacles begin to calmly unfurl, movements gracefully unfolding with the deliberate care of a ballet. As they unfold, a vibrant, otherworldly bioluminescence flows through them like electricity through a filament, causing them to glow in ever-changing hues of sapphire and ethereal violet. These threadlike tentacles extend downward, trailing into the darkness, their movement creating gentle, swaying patterns in the void. The creature, the sole source of light, casts a soft, cool glow that illuminates the waters around it, creating a stark and beautiful contrast with the abyss. The atmosphere is filled with quiet mystery and profound awe, and the extreme clarity of 8K resolution makes every delicate filament and pulse of light breathtakingly tangible.", + "seed": 31571, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000100/group_9/index_2444_seed_31571_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06659292383119464, + "vlm_logprob": -0.06659292383119464, + "vlm_yes_no_margin": 3.575, + "vlm_token_credit": -0.06659292383119464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22761932015419006 + ], + "opposite_logprob_tokens": [ + -1.6026192903518677 + ], + "logprob_reward": -0.22761932015419006, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22761932015419006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16157440841197968, + 0.055360667407512665, + 0.05117795616388321, + 0.04874426871538162, + 0.09036958962678909, + 0.12264003604650497, + 0.06955749541521072, + 0.3488844931125641, + 0.028879331424832344, + 0.022811787202954292 + ], + "credit_map": [ + 0.023843571543693542, + 0.01778530143201351, + 0.023852892220020294, + 0.018943095579743385, + 0.012220105156302452, + 0.02286219410598278, + 0.014630771242082119, + 0.02743646129965782, + 0.007496916223317385, + 0.0058563752099871635, + 0.006821265444159508, + 0.005010960157960653, + 0.00495994184166193, + 0.008437792770564556, + 0.005465404596179724, + 0.011312002316117287, + 0.0078049651347100735, + 0.004414042923599482, + 0.007091105915606022, + 0.006430820561945438, + 0.004319517873227596, + 0.007239294704049826, + 0.008656644262373447, + 0.005221561063081026, + 0.005465819966048002, + 0.004720770753920078, + 0.00549605768173933, + 0.005221773870289326, + 0.005472000688314438, + 0.007239487487822771, + 0.006677960976958275, + 0.008450392633676529, + 0.019158052280545235, + 0.006607098039239645, + 0.014149858616292477, + 0.01025394257158041, + 0.008361858315765858, + 0.009991921484470367, + 0.010717879980802536, + 0.011128973215818405, + 0.020881198346614838, + 0.010273178108036518, + 0.018451806157827377, + 0.016678059473633766, + 0.010015025734901428, + 0.009071271866559982, + 0.00970411766320467, + 0.027565374970436096, + 0.008826246485114098, + 0.007150460034608841, + 0.011800290085375309, + 0.00882771797478199, + 0.007334042806178331, + 0.0069915298372507095, + 0.0069867693819105625, + 0.01164043229073286, + 0.06095684692263603, + 0.02766682766377926, + 0.07407606393098831, + 0.060538869351148605, + 0.018669210374355316, + 0.020274773240089417, + 0.02673427201807499, + 0.059967637062072754, + 0.0036523891612887383, + 0.0034144222736358643, + 0.003450175281614065, + 0.004712570458650589, + 0.002647818997502327, + 0.002829094184562564, + 0.003218131372705102, + 0.004954726900905371, + 0.0029505211859941483, + 0.002012279350310564, + 0.0037845252081751823, + 0.0035655852407217026, + 0.00250770035199821, + 0.001717024715617299, + 0.0020445731934159994, + 0.00422957306727767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.22761932015419006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0025022882036864758 + ], + "opposite_logprob_tokens": [ + -6.00250244140625 + ], + "logprob_reward": -0.0025022882036864758, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.0025022882036864758, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17152689397335052, + 0.07414454966783524, + 0.09245480597019196, + 0.062494486570358276, + 0.035536255687475204, + 0.03431086242198944, + 0.10063257068395615, + 0.10379374772310257, + 0.1956234872341156, + 0.12948234379291534 + ], + "credit_map": [ + 0.029200663790106773, + 0.013880666345357895, + 0.025271907448768616, + 0.019163623452186584, + 0.015993745997548103, + 0.019851969555020332, + 0.024801993742585182, + 0.023362327367067337, + 0.008089026436209679, + 0.009285522624850273, + 0.008408812806010246, + 0.008297758176922798, + 0.010088471695780754, + 0.010219890624284744, + 0.011139706708490849, + 0.008615364320576191, + 0.00809744093567133, + 0.008111795410513878, + 0.013997994363307953, + 0.012966864742338657, + 0.011674625799059868, + 0.01188723649829626, + 0.014609894715249538, + 0.011108956299722195, + 0.0066078584641218185, + 0.006924360059201717, + 0.006426873151212931, + 0.0073544615879654884, + 0.005921025760471821, + 0.009046406485140324, + 0.011104186065495014, + 0.00910931546241045, + 0.006984851323068142, + 0.0033181635662913322, + 0.005428636912256479, + 0.0037694862112402916, + 0.004298269283026457, + 0.0033594018314033747, + 0.0028469657991081476, + 0.0055304826237261295, + 0.005197389982640743, + 0.0035072369500994682, + 0.00494975782930851, + 0.003676084103062749, + 0.00451491167768836, + 0.002988892374560237, + 0.0031168826390057802, + 0.0063597094267606735, + 0.01162767130881548, + 0.007567660417407751, + 0.017610322684049606, + 0.012138002552092075, + 0.0145567050203681, + 0.009862004779279232, + 0.009146906435489655, + 0.01812330074608326, + 0.01594843901693821, + 0.00907063391059637, + 0.01617162674665451, + 0.015216071158647537, + 0.012491696514189243, + 0.009944579564034939, + 0.00958552211523056, + 0.015365181490778923, + 0.022101102396845818, + 0.018087206408381462, + 0.026928508654236794, + 0.026251232251524925, + 0.019465075805783272, + 0.01883768104016781, + 0.024012519046664238, + 0.03994016349315643, + 0.014636479318141937, + 0.014220882207155228, + 0.019166946411132812, + 0.01961613819003105, + 0.011791404336690903, + 0.010984964668750763, + 0.015792377293109894, + 0.023273155093193054 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.0025022882036864758 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03370293974876404 + ], + "opposite_logprob_tokens": [ + -3.408702850341797 + ], + "logprob_reward": -0.03370293974876404, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03370293974876404, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20768457651138306, + 0.22295433282852173, + 0.1758296638727188, + 0.036284033209085464, + 0.05946877971291542, + 0.0755101665854454, + 0.05646197497844696, + 0.11226620525121689, + 0.021810611709952354, + 0.031729649752378464 + ], + "credit_map": [ + 0.028901534155011177, + 0.015180617570877075, + 0.032032791525125504, + 0.021582255139946938, + 0.016955450177192688, + 0.024274561554193497, + 0.032305072993040085, + 0.036452289670705795, + 0.02820422127842903, + 0.026392512023448944, + 0.027675025165081024, + 0.018509969115257263, + 0.028944391757249832, + 0.0348646380007267, + 0.03769959509372711, + 0.020663978531956673, + 0.01886744052171707, + 0.01331857219338417, + 0.032684728503227234, + 0.015732889994978905, + 0.016828052699565887, + 0.02909385971724987, + 0.03304879739880562, + 0.016255322843790054, + 0.006690425798296928, + 0.003132195444777608, + 0.005023962818086147, + 0.003001644043251872, + 0.003462428692728281, + 0.004412834532558918, + 0.004081002436578274, + 0.006479539442807436, + 0.009054017253220081, + 0.00630643405020237, + 0.009505576454102993, + 0.005428222008049488, + 0.005907953716814518, + 0.006145922001451254, + 0.006139892619103193, + 0.010980763472616673, + 0.011807872913777828, + 0.00452783890068531, + 0.013591363094747066, + 0.006904502399265766, + 0.006147152744233608, + 0.006243539042770863, + 0.006647092290222645, + 0.019640810787677765, + 0.006050903350114822, + 0.003621393349021673, + 0.010375842452049255, + 0.004553476348519325, + 0.011908445507287979, + 0.005155026912689209, + 0.0056038107722997665, + 0.009193075820803642, + 0.01814529113471508, + 0.009203064255416393, + 0.01759473979473114, + 0.019157078117132187, + 0.01561096590012312, + 0.007220990490168333, + 0.008673731237649918, + 0.016660349443554878, + 0.0027850251644849777, + 0.00220703543163836, + 0.003396181855350733, + 0.0032010998111218214, + 0.0027908277697861195, + 0.0013242826098576188, + 0.001965020317584276, + 0.0041411384008824825, + 0.003125679213553667, + 0.003212416311725974, + 0.003946473356336355, + 0.005748317111283541, + 0.0048246197402477264, + 0.0026406613178551197, + 0.0037958857137709856, + 0.0044355979189276695 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.03370293974876404 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419694535434246 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.01419694535434246, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419694535434246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16611328721046448, + 0.08237332850694656, + 0.13643372058868408, + 0.0381951741874218, + 0.06549052149057388, + 0.15213236212730408, + 0.08141482621431351, + 0.16496996581554413, + 0.07720554620027542, + 0.035671260207891464 + ], + "credit_map": [ + 0.02290990948677063, + 0.017651837319135666, + 0.029459908604621887, + 0.015397503972053528, + 0.012895496562123299, + 0.02119649574160576, + 0.015228545293211937, + 0.03137359023094177, + 0.010806001722812653, + 0.010100255720317364, + 0.01001712679862976, + 0.01134703028947115, + 0.011967351660132408, + 0.009837368503212929, + 0.009337792173027992, + 0.008960396982729435, + 0.013466338627040386, + 0.012470732443034649, + 0.018513401970267296, + 0.01801338419318199, + 0.019390512257814407, + 0.017192602157592773, + 0.023361319676041603, + 0.014025420881807804, + 0.005606711842119694, + 0.004901216831058264, + 0.005114696454256773, + 0.0037565266247838736, + 0.00459526339545846, + 0.004472535569220781, + 0.004845990799367428, + 0.004902235697954893, + 0.008157527074217796, + 0.004880931694060564, + 0.009791946038603783, + 0.0069686612114310265, + 0.006392318289726973, + 0.007604656275361776, + 0.007814300246536732, + 0.013880176469683647, + 0.02448563650250435, + 0.0138265835121274, + 0.021109404042363167, + 0.026005061343312263, + 0.014046095311641693, + 0.009109742939472198, + 0.013032706454396248, + 0.03051713854074478, + 0.010651038028299809, + 0.0043469760566949844, + 0.019198698922991753, + 0.007004046346992254, + 0.009815417230129242, + 0.0053726802580058575, + 0.005900499410927296, + 0.019125470891594887, + 0.0304503683000803, + 0.0121822040528059, + 0.0327276885509491, + 0.029811503365635872, + 0.014512832276523113, + 0.008700682781636715, + 0.011139540933072567, + 0.025445152074098587, + 0.008327336050570011, + 0.008579320274293423, + 0.010356757789850235, + 0.00988855306059122, + 0.008026516065001488, + 0.007206545677036047, + 0.010013779625296593, + 0.014806734398007393, + 0.0056219156831502914, + 0.0036684111692011356, + 0.005669950041919947, + 0.005667706020176411, + 0.002998399781063199, + 0.0029961569234728813, + 0.0044317953288555145, + 0.0046169268898665905 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 5 + ], + "selected_reward": -0.01419694535434246 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in the abyssal depths of the ocean with a background of pure black?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is a massive, translucent jellyfish with a glassy dome bell visible in the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Do the jellyfish's tentacles glow with bioluminescence in hues of sapphire and ethereal violet?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the jellyfish the sole source of light, casting a soft cool glow that illuminates the surrounding water?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot framed as a static medium close-up?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494312569499016 + ], + "opposite_logprob_tokens": [ + -2.929943084716797 + ], + "logprob_reward": -0.05494312569499016, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494312569499016, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12882569432258606, + 0.05308454856276512, + 0.06028968468308449, + 0.08288203924894333, + 0.027416953817009926, + 0.13786959648132324, + 0.054636478424072266, + 0.09082010388374329, + 0.11477970331907272, + 0.2493952065706253 + ], + "credit_map": [ + 0.01575656794011593, + 0.010902652516961098, + 0.018131576478481293, + 0.01779857464134693, + 0.012854432687163353, + 0.015983913093805313, + 0.018318355083465576, + 0.01907961443066597, + 0.006624658592045307, + 0.008215595968067646, + 0.007165323942899704, + 0.006073845084756613, + 0.006589159369468689, + 0.0067094662226736546, + 0.005671316292136908, + 0.006035184022039175, + 0.0048543838784098625, + 0.00374225783161819, + 0.009776239283382893, + 0.010341455228626728, + 0.005560559686273336, + 0.007920577190816402, + 0.010651936754584312, + 0.00744227459654212, + 0.014415835030376911, + 0.008688476867973804, + 0.01007554866373539, + 0.005803998094052076, + 0.008516381494700909, + 0.012155705131590366, + 0.008290845900774002, + 0.014935251325368881, + 0.00394017156213522, + 0.0033253200817853212, + 0.004114414565265179, + 0.003163642017170787, + 0.0031378723215311766, + 0.002765506040304899, + 0.0021364488638937473, + 0.004833578132092953, + 0.020388495177030563, + 0.011178165674209595, + 0.019087160006165504, + 0.0254862979054451, + 0.013447001576423645, + 0.010523981414735317, + 0.011689899489283562, + 0.026068605482578278, + 0.009208503179252148, + 0.0027641404885798693, + 0.010779641568660736, + 0.003494221717119217, + 0.008250023238360882, + 0.003674930427223444, + 0.0034419987350702286, + 0.013023017905652523, + 0.010852553881704807, + 0.007351601030677557, + 0.013498304411768913, + 0.014253012835979462, + 0.011978053487837315, + 0.007329970132559538, + 0.011247821152210236, + 0.014308786019682884, + 0.01555558294057846, + 0.012155241332948208, + 0.014500565826892853, + 0.016045745462179184, + 0.01240288931876421, + 0.010639743879437447, + 0.014744073152542114, + 0.018735861405730247, + 0.03247430548071861, + 0.022472070530056953, + 0.04245820268988609, + 0.03697669133543968, + 0.022593969479203224, + 0.01980794407427311, + 0.02067749574780464, + 0.05193452909588814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 0 + ], + "selected_reward": -0.05494312569499016 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_tensor([9442, 9442])_group_0_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_tensor([9442, 9442])_group_0_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..c084675f7676d58fa361666b2b12a4bb13eff442 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_tensor([9442, 9442])_group_0_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle shot, the camera slowly tilting upward, follows the rhythmic steps of a female backpacker as she ascends the steep, uneven stone steps of the ancient Great Wall. The backpacker wears a simple, dark brown ponytail, a few strands of hair whipping across her face in the morning breeze. She wears a professional crimson windbreaker, its nylon surface gleaming faintly in the dim light, durable black hiking pants, and a pair of high-top, muddy, dark brown hiking boots. She leans forward slightly to defy gravity, her grip on a carbon fiber trekking pole steadying her on the stone steps with each step. Her breath condenses into a fleeting white mist in the cool air. Behind her is a large, olive-green backpack, a rolled-up moisture-proof mat strapped to its top with a bungee cord, revealing the traces of a long journey. The soft, golden light of morning, just breaking through the horizon, casts a warm luster on the weathered gray masonry of the wall, highlighting its rough texture and the cracks of time. Behind her and in the distance, rolling hills are shrouded in flowing milky white mist, receding layer by layer towards the skyline, leaving only the dark green outlines of the mountain tops exposed, presenting a profound sense of layering like traditional Chinese ink paintings. The entire picture creates an atmosphere of tranquility, magnificence and perseverance.", + "seed": 23321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_9442_seed_23321_dp_0_rank_0_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04740683347918093, + "vlm_logprob": -0.04740683347918093, + "vlm_yes_no_margin": 3.850000023841858, + "vlm_token_credit": -0.04740683347918093, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16775383055210114 + ], + "opposite_logprob_tokens": [ + -2.0427539348602295 + ], + "logprob_reward": -0.16775383055210114, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.16775383055210114, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05905601754784584, + 0.05140375718474388, + 0.028815779834985733, + 0.028039302676916122, + 0.06884308904409409, + 0.05739779770374298, + 0.0470416434109211, + 0.1793709248304367, + 0.4594830870628357, + 0.020548589527606964 + ], + "credit_map": [ + 0.007787593174725771, + 0.005103594623506069, + 0.01165400817990303, + 0.00707941222935915, + 0.006077361758798361, + 0.009324361570179462, + 0.0046781753189861774, + 0.007351511623710394, + 0.006638617254793644, + 0.0033607196528464556, + 0.012981249019503593, + 0.009547870606184006, + 0.0031201145611703396, + 0.005707927979528904, + 0.004959050565958023, + 0.005088206846266985, + 0.004957579541951418, + 0.0019744711462408304, + 0.005668132100254297, + 0.002852948382496834, + 0.0029488150030374527, + 0.0045579345896840096, + 0.003410764504224062, + 0.002445134101435542, + 0.004381360486149788, + 0.0013931241119280457, + 0.006748195271939039, + 0.003477090736851096, + 0.004083013627678156, + 0.0036935009993612766, + 0.00253100274130702, + 0.001732016447931528, + 0.011044885031878948, + 0.007911691442131996, + 0.010331464000046253, + 0.007547892164438963, + 0.009630122222006321, + 0.011418336071074009, + 0.007488574367016554, + 0.0034701237455010414, + 0.009820557199418545, + 0.007987476885318756, + 0.006936599034816027, + 0.004649770446121693, + 0.00812115054577589, + 0.012130402028560638, + 0.004595579579472542, + 0.003156264079734683, + 0.005515620112419128, + 0.006792857311666012, + 0.0071585336700081825, + 0.007729493081569672, + 0.005308323074132204, + 0.004764176905155182, + 0.004933108575642109, + 0.004839526489377022, + 0.05517951026558876, + 0.01704186387360096, + 0.012754997238516808, + 0.011246751993894577, + 0.01109619252383709, + 0.05778740718960762, + 0.007442798465490341, + 0.006821410730481148, + 0.13793957233428955, + 0.020723486319184303, + 0.04779054969549179, + 0.020188428461551666, + 0.025214334949851036, + 0.1777409315109253, + 0.013372302986681461, + 0.016513468697667122, + 0.001878664130344987, + 0.00347125600092113, + 0.002031646901741624, + 0.0032730265520513058, + 0.0024360674433410168, + 0.0021623456850647926, + 0.0024334052577614784, + 0.0028621782548725605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 4 + ], + "selected_reward": -0.16775383055210114 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371930867433548 + ], + "opposite_logprob_tokens": [ + -3.408719301223755 + ], + "logprob_reward": -0.03371930867433548, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371930867433548, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09387243539094925, + 0.05369890481233597, + 0.02647632546722889, + 0.027227485552430153, + 0.03619454428553581, + 0.061373271048069, + 0.06175416335463524, + 0.33384427428245544, + 0.2780703008174896, + 0.027488380670547485 + ], + "credit_map": [ + 0.011344204656779766, + 0.007931879721581936, + 0.018977435305714607, + 0.014664094895124435, + 0.013873433694243431, + 0.009210282936692238, + 0.009757976047694683, + 0.00811312161386013, + 0.006504748482257128, + 0.005856125149875879, + 0.012597287073731422, + 0.00789263192564249, + 0.006643200758844614, + 0.005778906866908073, + 0.004712255205959082, + 0.0037137402687221766, + 0.004603195935487747, + 0.001824956969358027, + 0.005096600390970707, + 0.0027551688253879547, + 0.003335243323817849, + 0.0033319599460810423, + 0.0032759082969278097, + 0.0022532884031534195, + 0.004812788218259811, + 0.0028117725159972906, + 0.005349608603864908, + 0.0036170321982353926, + 0.003134534927085042, + 0.0033823298290371895, + 0.0024166086222976446, + 0.0017028064467012882, + 0.0061266436241567135, + 0.0028208494186401367, + 0.004303471650928259, + 0.0034451629035174847, + 0.006218807306140661, + 0.005030523985624313, + 0.005997940432280302, + 0.0022511424031108618, + 0.008353863842785358, + 0.005873264279216528, + 0.010047800838947296, + 0.005794079042971134, + 0.013534451834857464, + 0.008040975779294968, + 0.004939988721162081, + 0.004788835532963276, + 0.00948341190814972, + 0.006421861704438925, + 0.008057315833866596, + 0.008728939108550549, + 0.015010394155979156, + 0.00691152922809124, + 0.004306113347411156, + 0.002834595739841461, + 0.10497380048036575, + 0.021065903827548027, + 0.026854418218135834, + 0.019447550177574158, + 0.013634962029755116, + 0.12405629456043243, + 0.01112061832100153, + 0.012690670788288116, + 0.08269546180963516, + 0.011308768764138222, + 0.03116966225206852, + 0.010350379161536694, + 0.01582825556397438, + 0.10950127989053726, + 0.007027255836874247, + 0.010189193300902843, + 0.00289377779699862, + 0.0026946617290377617, + 0.00448906235396862, + 0.0038941670209169388, + 0.004818415269255638, + 0.0034479526802897453, + 0.0029191034846007824, + 0.0023312370758503675 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.03371930867433548 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187537789344788 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187537789344788, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187537789344788, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04438406974077225, + 0.09479373693466187, + 0.029454512521624565, + 0.03169424459338188, + 0.023961519822478294, + 0.06404492259025574, + 0.04652085900306702, + 0.21118293702602386, + 0.40362128615379333, + 0.05034198611974716 + ], + "credit_map": [ + 0.006168896332383156, + 0.003626696765422821, + 0.00897684134542942, + 0.005951628088951111, + 0.005392906721681356, + 0.005844894330948591, + 0.003827508771792054, + 0.004594691097736359, + 0.010120227001607418, + 0.006286613177508116, + 0.024782883003354073, + 0.01514535304158926, + 0.008727451786398888, + 0.010536049492657185, + 0.011056004092097282, + 0.008139142766594887, + 0.005298245698213577, + 0.001697215368039906, + 0.005650321487337351, + 0.003748258575797081, + 0.003033689223229885, + 0.004083265084773302, + 0.003453219775110483, + 0.0024902941659092903, + 0.005541171878576279, + 0.0033226008526980877, + 0.005344878416508436, + 0.0034623024985194206, + 0.0044195023365318775, + 0.005311736837029457, + 0.002422173274680972, + 0.0018698755884543061, + 0.0030349374283105135, + 0.0025808389764279127, + 0.003705168841406703, + 0.003158445470035076, + 0.002621142193675041, + 0.0033530546352267265, + 0.0034145605750381947, + 0.002093368908390403, + 0.008641245774924755, + 0.006569750141352415, + 0.009045585989952087, + 0.006348467897623777, + 0.015472881495952606, + 0.008050398901104927, + 0.005278955213725567, + 0.004637625999748707, + 0.005738984793424606, + 0.005486509762704372, + 0.005546583328396082, + 0.006516315974295139, + 0.008031599223613739, + 0.004950554110109806, + 0.006697837263345718, + 0.0035524696577340364, + 0.07030025124549866, + 0.008588471449911594, + 0.016237711533904076, + 0.010355035774409771, + 0.008100838400423527, + 0.08302997052669525, + 0.00589757552370429, + 0.008673051372170448, + 0.1188807487487793, + 0.015030277892947197, + 0.0453866608440876, + 0.016400853171944618, + 0.023015227168798447, + 0.15857364237308502, + 0.010974504984915257, + 0.015359311364591122, + 0.004418369848281145, + 0.0050279549323022366, + 0.006665902677923441, + 0.007981657981872559, + 0.008418886922299862, + 0.00798050407320261, + 0.005248220171779394, + 0.004600482527166605 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.018187537789344788 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031890033278614283 + ], + "opposite_logprob_tokens": [ + -5.7531890869140625 + ], + "logprob_reward": -0.0031890033278614283, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031890033278614283, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11493505537509918, + 0.030391056090593338, + 0.04944315925240517, + 0.048011988401412964, + 0.023021817207336426, + 0.05052744597196579, + 0.0724470242857933, + 0.3075513243675232, + 0.24010495841503143, + 0.06356622278690338 + ], + "credit_map": [ + 0.013122388161718845, + 0.012672120705246925, + 0.03128071874380112, + 0.017349274829030037, + 0.011056064628064632, + 0.009629139676690102, + 0.008104999549686909, + 0.01172034814953804, + 0.0039865244179964066, + 0.001849388936534524, + 0.007338287308812141, + 0.004807817284017801, + 0.002216220600530505, + 0.0040184627287089825, + 0.0031211392488330603, + 0.003053213469684124, + 0.006980197038501501, + 0.002659799763932824, + 0.01036624051630497, + 0.008172761648893356, + 0.003558567725121975, + 0.009534056298434734, + 0.003700096858665347, + 0.004471438471227884, + 0.008472979068756104, + 0.002737234113737941, + 0.010641728527843952, + 0.007355810608714819, + 0.00373647827655077, + 0.008552327752113342, + 0.00290019903331995, + 0.0036152335815131664, + 0.0033794455230236053, + 0.0022416224237531424, + 0.004386804532259703, + 0.004015278071165085, + 0.0025069108232855797, + 0.002968156011775136, + 0.0022216711658984423, + 0.0013019294710829854, + 0.006130222696810961, + 0.006371748633682728, + 0.007038506213575602, + 0.006151687819510698, + 0.008475072681903839, + 0.007261704187840223, + 0.0051865531131625175, + 0.003911952953785658, + 0.01041841134428978, + 0.0046421741135418415, + 0.01472531072795391, + 0.009716385044157505, + 0.009766039438545704, + 0.01134716160595417, + 0.006692332681268454, + 0.005139206536114216, + 0.10936902463436127, + 0.012067636474967003, + 0.023148074746131897, + 0.012268983758985996, + 0.009567645378410816, + 0.12389200925827026, + 0.007326480466872454, + 0.009911474771797657, + 0.06918584555387497, + 0.009067674167454243, + 0.02613060548901558, + 0.010409974493086338, + 0.014765947125852108, + 0.09443346410989761, + 0.0067314282059669495, + 0.009380032308399677, + 0.00496303616091609, + 0.0061807879246771336, + 0.010633053258061409, + 0.006627934519201517, + 0.011366092599928379, + 0.011839833110570908, + 0.006650783121585846, + 0.005304703488945961 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.0031890033278614283 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01418448705226183 + ], + "opposite_logprob_tokens": [ + -4.264184474945068 + ], + "logprob_reward": -0.01418448705226183, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01418448705226183, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15503597259521484, + 0.054247256368398666, + 0.06569433212280273, + 0.03404057398438454, + 0.049452148377895355, + 0.10702434927225113, + 0.06709600239992142, + 0.3511442244052887, + 0.07348166406154633, + 0.04278344660997391 + ], + "credit_map": [ + 0.017672749236226082, + 0.010784278623759747, + 0.03536100685596466, + 0.028295567259192467, + 0.01369921863079071, + 0.019407272338867188, + 0.013124949298799038, + 0.016690930351614952, + 0.006172901950776577, + 0.006079967599362135, + 0.011077498085796833, + 0.00832676887512207, + 0.006533152423799038, + 0.005028606858104467, + 0.005291672423481941, + 0.005736687686294317, + 0.010715458542108536, + 0.0028218403458595276, + 0.01690002717077732, + 0.009045610204339027, + 0.003984319511801004, + 0.009801726788282394, + 0.005519672762602568, + 0.006905677728354931, + 0.004556409548968077, + 0.0020303528290241957, + 0.007201056461781263, + 0.006328993942588568, + 0.003290484193712473, + 0.005959274247288704, + 0.002402930986136198, + 0.002271070843562484, + 0.00759842898696661, + 0.003619645955041051, + 0.008730019442737103, + 0.007623204030096531, + 0.005290366243571043, + 0.009240544401109219, + 0.0048590474762022495, + 0.00249089184217155, + 0.01382699515670538, + 0.011135296896100044, + 0.01907552033662796, + 0.01260640099644661, + 0.016804346814751625, + 0.015866870060563087, + 0.007988163270056248, + 0.009720753878355026, + 0.009183990769088268, + 0.005111438222229481, + 0.013736525550484657, + 0.010791396722197533, + 0.008544787764549255, + 0.008369474671781063, + 0.006987744942307472, + 0.0043706451542675495, + 0.12393435835838318, + 0.016558045521378517, + 0.026627078652381897, + 0.014024357311427593, + 0.010248401202261448, + 0.13875925540924072, + 0.009460685774683952, + 0.011532057076692581, + 0.010792831890285015, + 0.007700777146965265, + 0.00901883840560913, + 0.006870849523693323, + 0.006979715544730425, + 0.014978844672441483, + 0.006793634500354528, + 0.010346169583499432, + 0.0036549055948853493, + 0.003868867177516222, + 0.0053908443078398705, + 0.0045614768750965595, + 0.010541247203946114, + 0.006896250881254673, + 0.00454664696007967, + 0.003323210868984461 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 5 + ], + "selected_reward": -0.01418448705226183 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly tilting upward, follows the rhythmic steps of a female backpacker as she ascends the steep, uneven stone steps of the ancient Great Wall. The backpacker wears a simple, dark brown ponytail, a few strands of hair whipping across her face in the morning breeze. She wears a professional crimson windbreaker, its nylon surface gleaming faintly in the dim light, durable black hiking pants, and a pair of high-top, muddy, dark brown hiking boots. She leans forward slightly to defy gravity, her grip on a carbon fiber trekking pole steadying her on the stone steps with each step. Her breath condenses into a fleeting white mist in the cool air. Behind her is a large, olive-green backpack, a rolled-up moisture-proof mat strapped to its top with a bungee cord, revealing the traces of a long journey. The soft, golden light of morning, just breaking through the horizon, casts a warm luster on the weathered gray masonry of the wall, highlighting its rough texture and the cracks of time. Behind her and in the distance, rolling hills are shrouded in flowing milky white mist, receding layer by layer towards the skyline, leaving only the dark green outlines of the mountain tops exposed, presenting a profound sense of layering like traditional Chinese ink paintings. The entire picture creates an atmosphere of tranquility, magnificence and perseverance.", + "seed": 23322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_9442_seed_23322_dp_0_rank_0_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03600180549547076, + "vlm_logprob": -0.03600180549547076, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.03600180549547076, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.09456899017095566 + ], + "opposite_logprob_tokens": [ + -2.594568967819214 + ], + "logprob_reward": -0.09456899017095566, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.09456899017095566, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10497737675905228, + 0.038599926978349686, + 0.3871697783470154, + 0.11298029124736786, + 0.16380827128887177, + 0.04923972487449646, + 0.042307380586862564, + 0.06250407546758652, + 0.013358653523027897, + 0.025054531171917915 + ], + "credit_map": [ + 0.010644274763762951, + 0.01263552438467741, + 0.012725555337965488, + 0.019560975953936577, + 0.014498918317258358, + 0.007875021547079086, + 0.015071439556777477, + 0.011965668760240078, + 0.0036588734947144985, + 0.004778455942869186, + 0.0037926549557596445, + 0.003404114628210664, + 0.0034754672087728977, + 0.006260695401579142, + 0.006194988265633583, + 0.0070346747525036335, + 0.03324972838163376, + 0.030933205038309097, + 0.12747503817081451, + 0.012068425305187702, + 0.01313257310539484, + 0.1398301124572754, + 0.012016872875392437, + 0.018463818356394768, + 0.014454849995672703, + 0.009771960787475109, + 0.01377350464463234, + 0.009862608276307583, + 0.008290932513773441, + 0.039682988077402115, + 0.006859960965812206, + 0.010283486917614937, + 0.0223824605345726, + 0.018008729442954063, + 0.0255217757076025, + 0.022784708067774773, + 0.011454453691840172, + 0.021291468292474747, + 0.02228742651641369, + 0.020077258348464966, + 0.002663906430825591, + 0.006529098376631737, + 0.005639591254293919, + 0.002995115239173174, + 0.0024740900844335556, + 0.01904747448861599, + 0.002512626815587282, + 0.007377819623798132, + 0.0034255834762007, + 0.007643252145498991, + 0.006705631967633963, + 0.00567572470754385, + 0.005379554815590382, + 0.0033481500577181578, + 0.004722480662167072, + 0.005407002288848162, + 0.0057686250656843185, + 0.005269999150186777, + 0.008224405348300934, + 0.01059538871049881, + 0.0057217120192945, + 0.010356559418141842, + 0.009384282864630222, + 0.007183106150478125, + 0.001214237418025732, + 0.0019443959463387728, + 0.0024524261243641376, + 0.0014541614800691605, + 0.0010410553077235818, + 0.002359944861382246, + 0.0011832374148070812, + 0.0017091960180550814, + 0.0020547693129628897, + 0.0025794452521950006, + 0.003997765947133303, + 0.0045117661356925964, + 0.002265720395371318, + 0.003863890189677477, + 0.002990062814205885, + 0.0027911104261875153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 3 + ], + "selected_reward": -0.09456899017095566 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0160874892026186 + ], + "opposite_logprob_tokens": [ + -4.141087532043457 + ], + "logprob_reward": -0.0160874892026186, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0160874892026186, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0595133900642395, + 0.048892468214035034, + 0.4769704341888428, + 0.030235782265663147, + 0.03028675727546215, + 0.17995509505271912, + 0.028892086818814278, + 0.07746437191963196, + 0.028810273855924606, + 0.038979385048151016 + ], + "credit_map": [ + 0.00595952058210969, + 0.005520861130207777, + 0.005787791218608618, + 0.005798807833343744, + 0.014062655158340931, + 0.007501793559640646, + 0.005826361943036318, + 0.009055599570274353, + 0.005522009916603565, + 0.004745155107229948, + 0.00635885214433074, + 0.0036899999249726534, + 0.005091825034469366, + 0.006036954466253519, + 0.007051898166537285, + 0.010395773686468601, + 0.039957743138074875, + 0.03959925100207329, + 0.15070119500160217, + 0.014345470815896988, + 0.016861112788319588, + 0.17943419516086578, + 0.013086458668112755, + 0.022985011339187622, + 0.005062248092144728, + 0.0030159351881593466, + 0.0063714515417814255, + 0.0021041345316916704, + 0.0029888590797781944, + 0.0033362768590450287, + 0.0027899565175175667, + 0.004566919524222612, + 0.003257348434999585, + 0.00352876796387136, + 0.0031661575194448233, + 0.0033710566349327564, + 0.00296031404286623, + 0.0032857859041541815, + 0.006627829745411873, + 0.004089498892426491, + 0.016338244080543518, + 0.019215472042560577, + 0.026994450017809868, + 0.017098460346460342, + 0.012915120460093021, + 0.02208343707025051, + 0.01829502359032631, + 0.047014880925416946, + 0.0022903585340827703, + 0.0034267432056367397, + 0.0052789379842579365, + 0.003750292584300041, + 0.003860832890495658, + 0.0029080109670758247, + 0.002232389058917761, + 0.005144522059708834, + 0.007556240539997816, + 0.006261269096285105, + 0.011353620328009129, + 0.010960313491523266, + 0.00619780458509922, + 0.015070222318172455, + 0.009413928724825382, + 0.010650972835719585, + 0.00229197321459651, + 0.0019009035313501954, + 0.003681229194626212, + 0.0036432419437915087, + 0.0028487492818385363, + 0.005139417480677366, + 0.004339670296758413, + 0.0049650887958705425, + 0.004730751737952232, + 0.00855774711817503, + 0.005388067569583654, + 0.005524194333702326, + 0.0034930154215544462, + 0.005216515157371759, + 0.0031381358858197927, + 0.0029309578239917755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 7 + ], + "selected_reward": -0.0160874892026186 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009783650748431683 + ], + "opposite_logprob_tokens": [ + -4.634783744812012 + ], + "logprob_reward": -0.009783650748431683, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009783650748431683, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1406225711107254, + 0.08252962678670883, + 0.23610273003578186, + 0.07138992846012115, + 0.11659906804561615, + 0.04840226098895073, + 0.07137615233659744, + 0.034275006502866745, + 0.03150203078985214, + 0.16720068454742432 + ], + "credit_map": [ + 0.013745277188718319, + 0.013866908848285675, + 0.019464049488306046, + 0.016045108437538147, + 0.01573765091598034, + 0.02337736263871193, + 0.02033570036292076, + 0.018050508573651314, + 0.009419061243534088, + 0.007579826284199953, + 0.009431478567421436, + 0.0058443001471459866, + 0.010401027277112007, + 0.01834721304476261, + 0.009523086249828339, + 0.011983634904026985, + 0.02119332179427147, + 0.01925191469490528, + 0.06728158891201019, + 0.005819217301905155, + 0.008797042071819305, + 0.09527077525854111, + 0.005224877968430519, + 0.013263977132737637, + 0.011239484883844852, + 0.0056294724345207214, + 0.008795371279120445, + 0.004410651512444019, + 0.005136795341968536, + 0.024090448394417763, + 0.004663684871047735, + 0.007424020208418369, + 0.014908543787896633, + 0.013179341331124306, + 0.01884087175130844, + 0.01676374301314354, + 0.010498233139514923, + 0.014146960340440273, + 0.013281699270009995, + 0.014979679137468338, + 0.003834450151771307, + 0.005801612511277199, + 0.006993196438997984, + 0.006576498970389366, + 0.005028680898249149, + 0.009236402809619904, + 0.006047187838703394, + 0.004884232301265001, + 0.00796065665781498, + 0.013401946984231472, + 0.014060882851481438, + 0.009713664650917053, + 0.006918168161064386, + 0.005237453617155552, + 0.007018864154815674, + 0.007064509205520153, + 0.005007720552384853, + 0.004848197102546692, + 0.005491908639669418, + 0.0029979327227920294, + 0.0029117227531969547, + 0.0038094683550298214, + 0.00371643272228539, + 0.005491625051945448, + 0.0033893880899995565, + 0.003812543349340558, + 0.0049990504048764706, + 0.0033105413895100355, + 0.0024150165263563395, + 0.004300897475332022, + 0.0036125173792243004, + 0.005662078969180584, + 0.020461468026041985, + 0.031842269003391266, + 0.01829373836517334, + 0.037282101809978485, + 0.013940143398940563, + 0.017545275390148163, + 0.015781037509441376, + 0.012054643593728542 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 0 + ], + "selected_reward": -0.009783650748431683 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004634235519915819 + ], + "opposite_logprob_tokens": [ + -5.379634380340576 + ], + "logprob_reward": -0.004634235519915819, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004634235519915819, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15938372910022736, + 0.08549627661705017, + 0.06456468254327774, + 0.05202924460172653, + 0.10961825400590897, + 0.1786474883556366, + 0.09015680849552155, + 0.09357034415006638, + 0.06414029747247696, + 0.10239290446043015 + ], + "credit_map": [ + 0.018909748643636703, + 0.012798920273780823, + 0.018407313153147697, + 0.013701681047677994, + 0.04196267947554588, + 0.016395842656493187, + 0.014409798197448254, + 0.022797733545303345, + 0.008758184500038624, + 0.008255901746451855, + 0.011304770596325397, + 0.0055589983239769936, + 0.0065009696409106255, + 0.01161758229136467, + 0.010199799202382565, + 0.023300062865018845, + 0.007709100376814604, + 0.0054805707186460495, + 0.022622859105467796, + 0.0029404321685433388, + 0.0032279801089316607, + 0.01587173342704773, + 0.0023939774837344885, + 0.00431802636012435, + 0.009514556266367435, + 0.004460082855075598, + 0.011989961378276348, + 0.0030104275792837143, + 0.0041446080431342125, + 0.0074012805707752705, + 0.0036567591596394777, + 0.007851566188037395, + 0.015264504589140415, + 0.010098091326653957, + 0.02323862537741661, + 0.011410445906221867, + 0.008852409198880196, + 0.015150797553360462, + 0.012375707738101482, + 0.013227669522166252, + 0.010922432877123356, + 0.04594521224498749, + 0.015470198355615139, + 0.014345921576023102, + 0.009320609271526337, + 0.05212640389800072, + 0.008418776094913483, + 0.02209792099893093, + 0.011405753903090954, + 0.011702197603881359, + 0.016690589487552643, + 0.009247537702322006, + 0.008365422487258911, + 0.009003162384033203, + 0.008735228329896927, + 0.01500691007822752, + 0.009963054209947586, + 0.006681521888822317, + 0.017453327775001526, + 0.008660840801894665, + 0.00849201437085867, + 0.019517917186021805, + 0.00922519713640213, + 0.013576459139585495, + 0.005920202471315861, + 0.008803022094070911, + 0.013781829737126827, + 0.00790860503911972, + 0.003509885398671031, + 0.013046392239630222, + 0.005279651377350092, + 0.005890699569135904, + 0.012936622835695744, + 0.01713237166404724, + 0.011439315043389797, + 0.0281797144562006, + 0.007438989356160164, + 0.00778851518407464, + 0.010904032737016678, + 0.006573338992893696 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.004634235519915819 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05493466183543205 + ], + "opposite_logprob_tokens": [ + -2.9299347400665283 + ], + "logprob_reward": -0.05493466183543205, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05493466183543205, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1921018660068512, + 0.06837323307991028, + 0.07714333385229111, + 0.03461473435163498, + 0.07946585863828659, + 0.276822954416275, + 0.07742712646722794, + 0.08492537587881088, + 0.0655834823846817, + 0.043541986495256424 + ], + "credit_map": [ + 0.01977117918431759, + 0.018610557541251183, + 0.024508610367774963, + 0.024022316560149193, + 0.04113695025444031, + 0.016340889036655426, + 0.025459477677941322, + 0.02225188910961151, + 0.008298277854919434, + 0.010180812329053879, + 0.007117999251931906, + 0.005802708212286234, + 0.006226719822734594, + 0.009773715399205685, + 0.010031568817794323, + 0.010941432788968086, + 0.008594480343163013, + 0.006936219520866871, + 0.020755479112267494, + 0.004798340145498514, + 0.005747706163674593, + 0.018431833013892174, + 0.0041016326285898685, + 0.007777645252645016, + 0.004619975574314594, + 0.005102782975882292, + 0.0037242823746055365, + 0.0022037578746676445, + 0.003868278581649065, + 0.008315163664519787, + 0.0031748951878398657, + 0.003605601377785206, + 0.007419995032250881, + 0.007102358620613813, + 0.013088380917906761, + 0.008300437591969967, + 0.005699626170098782, + 0.01406652107834816, + 0.01312912255525589, + 0.010659413412213326, + 0.018211999908089638, + 0.06282198429107666, + 0.03049340844154358, + 0.02064162865281105, + 0.013903169892728329, + 0.08146532624959946, + 0.015391652472317219, + 0.033893782645463943, + 0.009191570803523064, + 0.008925806730985641, + 0.017702514305710793, + 0.0079494658857584, + 0.0055506909266114235, + 0.012012518011033535, + 0.0068952892906963825, + 0.009199278429150581, + 0.009930702857673168, + 0.008313787169754505, + 0.014316924847662449, + 0.011292262002825737, + 0.006191449239850044, + 0.01671929843723774, + 0.008595308288931847, + 0.009565647691488266, + 0.009437345899641514, + 0.005597146227955818, + 0.00878928229212761, + 0.005542765371501446, + 0.0038177277892827988, + 0.011182150803506374, + 0.005137259140610695, + 0.01607980579137802, + 0.004842804279178381, + 0.005623101256787777, + 0.005863908678293228, + 0.009981486946344376, + 0.0031947102397680283, + 0.005055109038949013, + 0.004157701041549444, + 0.004823166877031326 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 7 + ], + "selected_reward": -0.05493466183543205 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly tilting upward, follows the rhythmic steps of a female backpacker as she ascends the steep, uneven stone steps of the ancient Great Wall. The backpacker wears a simple, dark brown ponytail, a few strands of hair whipping across her face in the morning breeze. She wears a professional crimson windbreaker, its nylon surface gleaming faintly in the dim light, durable black hiking pants, and a pair of high-top, muddy, dark brown hiking boots. She leans forward slightly to defy gravity, her grip on a carbon fiber trekking pole steadying her on the stone steps with each step. Her breath condenses into a fleeting white mist in the cool air. Behind her is a large, olive-green backpack, a rolled-up moisture-proof mat strapped to its top with a bungee cord, revealing the traces of a long journey. The soft, golden light of morning, just breaking through the horizon, casts a warm luster on the weathered gray masonry of the wall, highlighting its rough texture and the cracks of time. Behind her and in the distance, rolling hills are shrouded in flowing milky white mist, receding layer by layer towards the skyline, leaving only the dark green outlines of the mountain tops exposed, presenting a profound sense of layering like traditional Chinese ink paintings. The entire picture creates an atmosphere of tranquility, magnificence and perseverance.", + "seed": 23323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_9442_seed_23323_dp_1_rank_2_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.040763987321406604, + "vlm_logprob": -0.040763987321406604, + "vlm_yes_no_margin": 3.875, + "vlm_token_credit": -0.040763987321406604, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10467281937599182 + ], + "opposite_logprob_tokens": [ + -2.479672908782959 + ], + "logprob_reward": -0.10467281937599182, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10467281937599182, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20861703157424927, + 0.14163489639759064, + 0.14799121022224426, + 0.1765463799238205, + 0.04591253399848938, + 0.10991109907627106, + 0.10517583787441254, + 0.021901967003941536, + 0.029926298186182976, + 0.012382746674120426 + ], + "credit_map": [ + 0.026951882988214493, + 0.020771170035004616, + 0.025480935350060463, + 0.024221759289503098, + 0.027301279827952385, + 0.031161069869995117, + 0.023680206388235092, + 0.029048729687929153, + 0.011396397836506367, + 0.006385845132172108, + 0.038513071835041046, + 0.011362825520336628, + 0.011007269844412804, + 0.008539696224033833, + 0.006851661950349808, + 0.047578126192092896, + 0.012409835122525692, + 0.017577936872839928, + 0.02043108455836773, + 0.025660930201411247, + 0.010490535758435726, + 0.020832404494285583, + 0.019880598410964012, + 0.020707886666059494, + 0.03600072115659714, + 0.05702086538076401, + 0.019658192992210388, + 0.00807458721101284, + 0.024407973513007164, + 0.010891949757933617, + 0.010577618144452572, + 0.009914468042552471, + 0.003383907489478588, + 0.003173896111547947, + 0.005948922596871853, + 0.004659583792090416, + 0.01087918784469366, + 0.005856343545019627, + 0.005225088447332382, + 0.006785601377487183, + 0.007837246172130108, + 0.007882454432547092, + 0.010495457798242569, + 0.012418522499501705, + 0.014131490141153336, + 0.021008986979722977, + 0.01491488330066204, + 0.021222054958343506, + 0.014981741085648537, + 0.008782954886555672, + 0.0080975741147995, + 0.008630113676190376, + 0.013768593780696392, + 0.018091898411512375, + 0.014280766248703003, + 0.018542185425758362, + 0.0028083594515919685, + 0.003262102138251066, + 0.002058142563328147, + 0.0030605585779994726, + 0.0029035471379756927, + 0.001967958640307188, + 0.003025196958333254, + 0.0028160992078483105, + 0.0031423738691955805, + 0.003079016227275133, + 0.002673190785571933, + 0.002621863503009081, + 0.0032604963053017855, + 0.006333338096737862, + 0.003371366299688816, + 0.005444653332233429, + 0.0014972887001931667, + 0.0016280574491247535, + 0.0013796811690554023, + 0.001396340667270124, + 0.0019215880893170834, + 0.0012742311228066683, + 0.001304137404076755, + 0.001981422770768404 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.10467281937599182 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04308098554611206 + ], + "opposite_logprob_tokens": [ + -3.168081045150757 + ], + "logprob_reward": -0.04308098554611206, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04308098554611206, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040616389364004135, + 0.7010325193405151, + 0.05748399719595909, + 0.10418084263801575, + 0.028791390359401703, + 0.020882293581962585, + 0.016177384182810783, + 0.012319651432335377, + 0.011678026057779789, + 0.00683754263445735 + ], + "credit_map": [ + 0.006282609887421131, + 0.0039061615243554115, + 0.004644186235964298, + 0.005136032588779926, + 0.005162824876606464, + 0.004763044882565737, + 0.00395673094317317, + 0.006764788180589676, + 0.04083972051739693, + 0.04075255244970322, + 0.2541324198246002, + 0.04642479494214058, + 0.05818983539938927, + 0.032444100826978683, + 0.016948401927947998, + 0.21130062639713287, + 0.0058896103873848915, + 0.010688976384699345, + 0.005312517751008272, + 0.005037092603743076, + 0.01023427490144968, + 0.005628200713545084, + 0.007640955038368702, + 0.007052355445921421, + 0.020929280668497086, + 0.03662865236401558, + 0.009382999502122402, + 0.005132976453751326, + 0.011388287879526615, + 0.005515399854630232, + 0.006331212818622589, + 0.008872021920979023, + 0.0024144805502146482, + 0.002862949389964342, + 0.003126236842945218, + 0.0030422359704971313, + 0.005869122687727213, + 0.0034767542965710163, + 0.0037031355313956738, + 0.004296468570828438, + 0.0016386124771088362, + 0.001394430990330875, + 0.0020078341476619244, + 0.002101678168401122, + 0.0032309284433722496, + 0.0033502080477774143, + 0.0036858750972896814, + 0.003472723765298724, + 0.0023109347093850374, + 0.001488629961386323, + 0.0012094476260244846, + 0.0013278322294354439, + 0.0021787998266518116, + 0.0024082697927951813, + 0.002153864363208413, + 0.0030996014829725027, + 0.0011813101591542363, + 0.0011801699874922633, + 0.0010800494346767664, + 0.002097342861816287, + 0.0013115813490003347, + 0.001533244620077312, + 0.002358843572437763, + 0.0015771066537126899, + 0.0011710880789905787, + 0.0013140605296939611, + 0.000998688512481749, + 0.0013267816975712776, + 0.0012964835623279214, + 0.0020612156949937344, + 0.0011904906714335084, + 0.0023192153312265873, + 0.0008412672905251384, + 0.000800785084720701, + 0.000689809734467417, + 0.0007209594477899373, + 0.0005906248698011041, + 0.0012263586977496743, + 0.0008318090694956481, + 0.0011359266936779022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 2 + ], + "selected_reward": -0.04308098554611206 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011078042909502983 + ], + "opposite_logprob_tokens": [ + -4.511077880859375 + ], + "logprob_reward": -0.011078042909502983, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011078042909502983, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07973058521747589, + 0.28746071457862854, + 0.30467459559440613, + 0.11571155488491058, + 0.03717152774333954, + 0.07016698271036148, + 0.030717646703124046, + 0.02354745753109455, + 0.027765871956944466, + 0.023053040727972984 + ], + "credit_map": [ + 0.009940849617123604, + 0.009498292580246925, + 0.013105472549796104, + 0.010515175759792328, + 0.010350286960601807, + 0.007800211198627949, + 0.009651471860706806, + 0.008868834003806114, + 0.01670219749212265, + 0.015232623554766178, + 0.10062965750694275, + 0.019934482872486115, + 0.02439233660697937, + 0.01385379210114479, + 0.011058460921049118, + 0.08565720915794373, + 0.06203053146600723, + 0.11367781460285187, + 0.035784024745225906, + 0.02307935059070587, + 0.019776852801442146, + 0.016394581645727158, + 0.016494877636432648, + 0.01743657886981964, + 0.01001252792775631, + 0.02287030592560768, + 0.009304272010922432, + 0.011498607695102692, + 0.014064935967326164, + 0.010855121538043022, + 0.009770710952579975, + 0.02733508311212063, + 0.006974782329052687, + 0.00673458119854331, + 0.003261539386585355, + 0.00337971025146544, + 0.0061260065995156765, + 0.0032320960890501738, + 0.0033649711403995752, + 0.004097845405340195, + 0.006054151337593794, + 0.005096002481877804, + 0.007690815720707178, + 0.007022605277597904, + 0.013377049937844276, + 0.012878414243459702, + 0.007825212553143501, + 0.010222739540040493, + 0.0031918780878186226, + 0.003014215501025319, + 0.0034867690410465, + 0.004180457908660173, + 0.0038205059245228767, + 0.003347175195813179, + 0.0039077093824744225, + 0.005768938455730677, + 0.0030741773080080748, + 0.0023000917863100767, + 0.0018752221949398518, + 0.003537122393026948, + 0.0020151601638644934, + 0.0033752836752682924, + 0.0035712560638785362, + 0.0037991467397660017, + 0.0034059016034007072, + 0.0032709650695323944, + 0.00295903324149549, + 0.003432729048654437, + 0.0021228198893368244, + 0.003915074281394482, + 0.0036319675855338573, + 0.005027385428547859, + 0.0020361561328172684, + 0.002318223472684622, + 0.0024434695951640606, + 0.003088552039116621, + 0.0028227618895471096, + 0.003519817953929305, + 0.0027870747726410627, + 0.004036986269056797 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 3 + ], + "selected_reward": -0.011078042909502983 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019388464279472828 + ], + "opposite_logprob_tokens": [ + -6.251938819885254 + ], + "logprob_reward": -0.0019388464279472828, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019388464279472828, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19256459176540375, + 0.22395244240760803, + 0.22353661060333252, + 0.0515916682779789, + 0.060044027864933014, + 0.05337034910917282, + 0.048950131982564926, + 0.07624116539955139, + 0.036802731454372406, + 0.032946210354566574 + ], + "credit_map": [ + 0.02024836465716362, + 0.02013980969786644, + 0.026797503232955933, + 0.027772504836320877, + 0.021133247762918472, + 0.027845468372106552, + 0.01595556177198887, + 0.032672129571437836, + 0.014665449038147926, + 0.012202669866383076, + 0.08080638200044632, + 0.016384536400437355, + 0.019078560173511505, + 0.010343833826482296, + 0.008227847516536713, + 0.06224318593740463, + 0.04895279183983803, + 0.08738230913877487, + 0.02447974495589733, + 0.012276865541934967, + 0.012082710862159729, + 0.014787795022130013, + 0.010795575566589832, + 0.0127788120880723, + 0.004403917584568262, + 0.005431581754237413, + 0.0060845897532999516, + 0.006939548067748547, + 0.006610107608139515, + 0.005149852950125933, + 0.004030060488730669, + 0.012942008674144745, + 0.010079815983772278, + 0.012503710575401783, + 0.0067696045152843, + 0.008603955619037151, + 0.006586285773664713, + 0.003253677161410451, + 0.00575972581282258, + 0.006487250793725252, + 0.005899923853576183, + 0.005859187338501215, + 0.006216084584593773, + 0.007581949699670076, + 0.006082906853407621, + 0.006160354241728783, + 0.010474173352122307, + 0.005095772910863161, + 0.006136486306786537, + 0.005344932898879051, + 0.004339338280260563, + 0.005243314430117607, + 0.007126920390874147, + 0.0067521450109779835, + 0.005204308778047562, + 0.0088026849552989, + 0.008865298703312874, + 0.009986132383346558, + 0.006237969733774662, + 0.012439410202205181, + 0.00771032739430666, + 0.008033775724470615, + 0.014210538938641548, + 0.008757714182138443, + 0.0035858408082276583, + 0.004838111810386181, + 0.003701241919770837, + 0.005223356653004885, + 0.004631298594176769, + 0.004437473602592945, + 0.004441794008016586, + 0.005943613592535257, + 0.003994762431830168, + 0.0035609707701951265, + 0.003612837288528681, + 0.004805347416549921, + 0.0034005618654191494, + 0.00422276183962822, + 0.004327034577727318, + 0.005021935328841209 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.0019388464279472828 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04304924234747887 + ], + "opposite_logprob_tokens": [ + -3.1680493354797363 + ], + "logprob_reward": -0.04304924234747887, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04304924234747887, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13734786212444305, + 0.6111360192298889, + 0.04771967604756355, + 0.030841317027807236, + 0.02670449949800968, + 0.04552005976438522, + 0.027756206691265106, + 0.03212864324450493, + 0.02238629199564457, + 0.018459424376487732 + ], + "credit_map": [ + 0.017139708623290062, + 0.01340420451015234, + 0.01983458921313286, + 0.019373249262571335, + 0.016616955399513245, + 0.017749140039086342, + 0.013052038848400116, + 0.020177962258458138, + 0.03600326180458069, + 0.03340054303407669, + 0.2149929404258728, + 0.04158693179488182, + 0.04879368469119072, + 0.02875709906220436, + 0.014251991175115108, + 0.19334955513477325, + 0.007521057967096567, + 0.011083527468144894, + 0.004569060169160366, + 0.006953428033739328, + 0.003244438674300909, + 0.005234424024820328, + 0.00423758989199996, + 0.00487614655867219, + 0.0054590837098658085, + 0.005488113965839148, + 0.002684095175936818, + 0.002613563323393464, + 0.003254788462072611, + 0.0038309628143906593, + 0.0026485584676265717, + 0.004862149246037006, + 0.0027750066947191954, + 0.0033297045156359673, + 0.004108063876628876, + 0.00441272696480155, + 0.0039964900352060795, + 0.002176837995648384, + 0.002641593338921666, + 0.003264074446633458, + 0.005545975174754858, + 0.00434791948646307, + 0.004979240708053112, + 0.004797987174242735, + 0.00411643274128437, + 0.007307909894734621, + 0.010773388668894768, + 0.0036512031219899654, + 0.004196926020085812, + 0.0023440958466380835, + 0.002156058559194207, + 0.002668926492333412, + 0.003472927026450634, + 0.004751149564981461, + 0.0040086242370307446, + 0.004157496150583029, + 0.0045046741142869, + 0.00444020377472043, + 0.0021713965106755495, + 0.003893436398357153, + 0.0030709155835211277, + 0.005300149321556091, + 0.004216361325234175, + 0.004531505052000284, + 0.0026748899836093187, + 0.0023071072064340115, + 0.0024223746731877327, + 0.0033337196800857782, + 0.001948439166881144, + 0.0026169337797909975, + 0.003251866204664111, + 0.00383095839060843, + 0.0021214967127889395, + 0.002947254339233041, + 0.0016493514413014054, + 0.002080403035506606, + 0.0031928180251270533, + 0.0014175917021930218, + 0.002648402703925967, + 0.0024021058343350887 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.04304924234747887 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle shot, the camera slowly tilting upward, follows the rhythmic steps of a female backpacker as she ascends the steep, uneven stone steps of the ancient Great Wall. The backpacker wears a simple, dark brown ponytail, a few strands of hair whipping across her face in the morning breeze. She wears a professional crimson windbreaker, its nylon surface gleaming faintly in the dim light, durable black hiking pants, and a pair of high-top, muddy, dark brown hiking boots. She leans forward slightly to defy gravity, her grip on a carbon fiber trekking pole steadying her on the stone steps with each step. Her breath condenses into a fleeting white mist in the cool air. Behind her is a large, olive-green backpack, a rolled-up moisture-proof mat strapped to its top with a bungee cord, revealing the traces of a long journey. The soft, golden light of morning, just breaking through the horizon, casts a warm luster on the weathered gray masonry of the wall, highlighting its rough texture and the cracks of time. Behind her and in the distance, rolling hills are shrouded in flowing milky white mist, receding layer by layer towards the skyline, leaving only the dark green outlines of the mountain tops exposed, presenting a profound sense of layering like traditional Chinese ink paintings. The entire picture creates an atmosphere of tranquility, magnificence and perseverance.", + "seed": 23324, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_0/index_9442_seed_23324_dp_1_rank_2_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 0, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04178153220564127, + "vlm_logprob": -0.04178153220564127, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.04178153220564127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10122635960578918 + ], + "opposite_logprob_tokens": [ + -2.476226329803467 + ], + "logprob_reward": -0.10122635960578918, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.10122635960578918, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41925904154777527, + 0.08061709254980087, + 0.04837263748049736, + 0.15591932833194733, + 0.04840726777911186, + 0.05850939452648163, + 0.03422589600086212, + 0.05908377096056938, + 0.03482906520366669, + 0.060776486992836 + ], + "credit_map": [ + 0.13713602721691132, + 0.05102694779634476, + 0.02391774021089077, + 0.014708167873322964, + 0.011054647155106068, + 0.04300285503268242, + 0.02457391284406185, + 0.11383874714374542, + 0.012030025012791157, + 0.0074328649789094925, + 0.012099063955247402, + 0.009048505686223507, + 0.009699000045657158, + 0.009789776057004929, + 0.009376943111419678, + 0.011140916496515274, + 0.006290694233030081, + 0.0045720599591732025, + 0.010464496910572052, + 0.0034674524795264006, + 0.004712631925940514, + 0.010392716154456139, + 0.003349781734868884, + 0.005122806411236525, + 0.017909320071339607, + 0.014974535442888737, + 0.01596192456781864, + 0.0302750151604414, + 0.018100420013070107, + 0.016952527686953545, + 0.02369273081421852, + 0.018052849918603897, + 0.007593227084726095, + 0.008361204527318478, + 0.008972118608653545, + 0.006742723286151886, + 0.0034176355693489313, + 0.00431502889841795, + 0.0040044416673481464, + 0.0050008888356387615, + 0.006726605352014303, + 0.0070313564501702785, + 0.006918012630194426, + 0.010335068218410015, + 0.007959108799695969, + 0.007070076651871204, + 0.007704833522439003, + 0.004764329642057419, + 0.005292439833283424, + 0.0052750906907022, + 0.005891574081033468, + 0.003460171166807413, + 0.002699483186006546, + 0.0037132787983864546, + 0.0040384316816926, + 0.003855427959933877, + 0.014542628079652786, + 0.015613129362463951, + 0.006584825925529003, + 0.006603606045246124, + 0.004718368407338858, + 0.0034484670031815767, + 0.004754341207444668, + 0.0028184016700834036, + 0.004337499383836985, + 0.004191935993731022, + 0.00551461661234498, + 0.005299570970237255, + 0.004507226403802633, + 0.003252935828641057, + 0.0039140027947723866, + 0.0038112737238407135, + 0.014035779051482677, + 0.015863541513681412, + 0.00890229269862175, + 0.00741859246045351, + 0.004621315281838179, + 0.004044367931783199, + 0.0029483367688953876, + 0.00294226105324924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.10122635960578918 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048671357333660126 + ], + "opposite_logprob_tokens": [ + -3.048671245574951 + ], + "logprob_reward": -0.048671357333660126, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048671357333660126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08507641404867172, + 0.23211459815502167, + 0.06974975019693375, + 0.256801038980484, + 0.03574565052986145, + 0.10407309979200363, + 0.07589031010866165, + 0.04275960475206375, + 0.04504639282822609, + 0.052743107080459595 + ], + "credit_map": [ + 0.024527428671717644, + 0.008907349780201912, + 0.008504016324877739, + 0.007809799630194902, + 0.005903280805796385, + 0.011541139334440231, + 0.008795095607638359, + 0.009088302031159401, + 0.031154967844486237, + 0.020370349287986755, + 0.03301781415939331, + 0.01689598150551319, + 0.040107984095811844, + 0.03530601039528847, + 0.018752403557300568, + 0.03650909662246704, + 0.012256567366421223, + 0.00584022980183363, + 0.014246637001633644, + 0.005879380274564028, + 0.005428072530776262, + 0.010544135235249996, + 0.007293292321264744, + 0.00826143566519022, + 0.04532182961702347, + 0.02815791219472885, + 0.01900058053433895, + 0.03436208516359329, + 0.026065345853567123, + 0.028495291247963905, + 0.046911124140024185, + 0.028486885130405426, + 0.003995796199887991, + 0.0035503630060702562, + 0.010213008150458336, + 0.005911576561629772, + 0.0023260870948433876, + 0.0033635857980698347, + 0.00267073349095881, + 0.0037145002279430628, + 0.02285643108189106, + 0.025878215208649635, + 0.019952183589339256, + 0.01421770267188549, + 0.005197308026254177, + 0.007742486894130707, + 0.004559888504445553, + 0.003668884513899684, + 0.01036247331649065, + 0.011779728345572948, + 0.007864193059504032, + 0.009478556923568249, + 0.007917300797998905, + 0.008436824195086956, + 0.011458766646683216, + 0.00859246589243412, + 0.006964174099266529, + 0.0068223001435399055, + 0.004148455336689949, + 0.0033116359263658524, + 0.00799015536904335, + 0.004843004513531923, + 0.004577040672302246, + 0.004102836363017559, + 0.008439346216619015, + 0.008983415551483631, + 0.007339733187109232, + 0.005888692568987608, + 0.0037290584295988083, + 0.0027290189173072577, + 0.00350010278634727, + 0.004437023773789406, + 0.006958750542253256, + 0.006347677670419216, + 0.007055041845887899, + 0.006396031007170677, + 0.007343907840549946, + 0.007270621135830879, + 0.005433006212115288, + 0.005938075948506594 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 5 + ], + "selected_reward": -0.048671357333660126 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020585810765624046 + ], + "opposite_logprob_tokens": [ + -3.895585775375366 + ], + "logprob_reward": -0.020585810765624046, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020585810765624046, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19462452828884125, + 0.12993134558200836, + 0.050759222358465195, + 0.3187545835971832, + 0.03290865570306778, + 0.07961353659629822, + 0.06725180894136429, + 0.04349423572421074, + 0.03787447139620781, + 0.044787682592868805 + ], + "credit_map": [ + 0.05820630118250847, + 0.020921921357512474, + 0.012644247151911259, + 0.013151774182915688, + 0.0073845707811415195, + 0.019221002236008644, + 0.014830810017883778, + 0.04826388508081436, + 0.01633617840707302, + 0.010966086760163307, + 0.018005507066845894, + 0.010309911333024502, + 0.02328428067266941, + 0.017596468329429626, + 0.013419216498732567, + 0.02001369744539261, + 0.009037282317876816, + 0.0040672593750059605, + 0.010467243380844593, + 0.004073846619576216, + 0.0040923007763922215, + 0.00845034047961235, + 0.005458238068968058, + 0.00511271134018898, + 0.0538034550845623, + 0.03019237145781517, + 0.02535141073167324, + 0.05015451833605766, + 0.03042488917708397, + 0.031065281480550766, + 0.061511363834142685, + 0.03625129535794258, + 0.0037262658588588238, + 0.004010005854070187, + 0.006959433667361736, + 0.005021319724619389, + 0.003221520222723484, + 0.0028561679646372795, + 0.0033104964531958103, + 0.003803446190431714, + 0.01322850864380598, + 0.014126488007605076, + 0.0206668172031641, + 0.014593718573451042, + 0.003743600333109498, + 0.00692170113325119, + 0.0037290710024535656, + 0.00260362820699811, + 0.0054746330715715885, + 0.007558357901871204, + 0.0090246070176363, + 0.010079451836645603, + 0.007579007651656866, + 0.006907052826136351, + 0.012592271901667118, + 0.008036426268517971, + 0.00887229386717081, + 0.008701050654053688, + 0.003116670297458768, + 0.0032930367160588503, + 0.008010020479559898, + 0.004685715772211552, + 0.004096126649528742, + 0.0027193224523216486, + 0.0049250777810812, + 0.005563793238252401, + 0.005669646896421909, + 0.005223871674388647, + 0.004328087437897921, + 0.0032227588817477226, + 0.004112015478312969, + 0.004829218611121178, + 0.008983979932963848, + 0.008902420289814472, + 0.00513638649135828, + 0.004710599314421415, + 0.005618995055556297, + 0.004674823023378849, + 0.003959013149142265, + 0.0028014674317091703 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.020585810765624046 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00863425899296999 + ], + "opposite_logprob_tokens": [ + -4.758634090423584 + ], + "logprob_reward": -0.00863425899296999, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00863425899296999, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3299267590045929, + 0.03308115154504776, + 0.04855063930153847, + 0.11878189444541931, + 0.04582661762833595, + 0.10186169296503067, + 0.07455451786518097, + 0.032522767782211304, + 0.17004577815532684, + 0.04484814777970314 + ], + "credit_map": [ + 0.11898193508386612, + 0.04122777283191681, + 0.018912844359874725, + 0.012672808021306992, + 0.0107574462890625, + 0.025944022461771965, + 0.02141844853758812, + 0.08001150190830231, + 0.004867232404649258, + 0.003961856942623854, + 0.0068961544893682, + 0.003563818521797657, + 0.003513928269967437, + 0.0023892992176115513, + 0.004437648691236973, + 0.0034512141719460487, + 0.01131096389144659, + 0.004454974550753832, + 0.010809261351823807, + 0.004454310052096844, + 0.003182243090122938, + 0.004593045450747013, + 0.005276862997561693, + 0.004468975123018026, + 0.00919819250702858, + 0.012215489521622658, + 0.019048770889639854, + 0.02671586349606514, + 0.009226188994944096, + 0.015231749042868614, + 0.01079034898430109, + 0.01635528728365898, + 0.006135969888418913, + 0.008749280124902725, + 0.01181363221257925, + 0.007751620840281248, + 0.00245099444873631, + 0.0030704746022820473, + 0.0033701711799949408, + 0.0024844761937856674, + 0.01457587257027626, + 0.015454902313649654, + 0.024998469278216362, + 0.01786377839744091, + 0.007986913435161114, + 0.008619898930191994, + 0.00786335114389658, + 0.004498517140746117, + 0.010446148924529552, + 0.011779855005443096, + 0.014916480518877506, + 0.011732878163456917, + 0.005627605598419905, + 0.006901624146848917, + 0.007112191990017891, + 0.006037734914571047, + 0.008287507109344006, + 0.008795036002993584, + 0.003991719335317612, + 0.003472881391644478, + 0.0021459900308400393, + 0.0016109084244817495, + 0.0022578134667128325, + 0.00196091178804636, + 0.03387986496090889, + 0.03655557706952095, + 0.027956945821642876, + 0.022653184831142426, + 0.010845246724784374, + 0.010096609592437744, + 0.012893687933683395, + 0.015164661221206188, + 0.010073197074234486, + 0.012475354596972466, + 0.004487084690481424, + 0.0050046141259372234, + 0.0045061251148581505, + 0.0031116828322410583, + 0.0024728209245949984, + 0.002717267954722047 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 3 + ], + "selected_reward": -0.00863425899296999 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot taken from a low angle and slowly tilting upward?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the main subject a female backpacker wearing a crimson windbreaker and dark brown hiking boots?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the backpacker ascending steep, uneven stone steps of the Great Wall while leaning forward and using a carbon fiber trekking pole?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the background environment characterized by weathered gray masonry, rolling hills, and milky white mist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene illuminated by soft, golden morning light breaking through the horizon?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029789874330163002 + ], + "opposite_logprob_tokens": [ + -3.529789924621582 + ], + "logprob_reward": -0.029789874330163002, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029789874330163002, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17706258594989777, + 0.09841544926166534, + 0.03969110548496246, + 0.06691113114356995, + 0.13210800290107727, + 0.13910703361034393, + 0.08501485735177994, + 0.08209998905658722, + 0.06135908141732216, + 0.11823073774576187 + ], + "credit_map": [ + 0.03684455528855324, + 0.022134272381663322, + 0.021053530275821686, + 0.02483174577355385, + 0.012401893734931946, + 0.0190933458507061, + 0.023333944380283356, + 0.017369290813803673, + 0.01252500806003809, + 0.011903153732419014, + 0.014156254939734936, + 0.007682926952838898, + 0.017444387078285217, + 0.013180043548345566, + 0.008869975805282593, + 0.012653695419430733, + 0.007333376910537481, + 0.0053263381123542786, + 0.004836763720959425, + 0.0046173143200576305, + 0.0032257665880024433, + 0.004894152283668518, + 0.0050981249660253525, + 0.004359268583357334, + 0.007640863303095102, + 0.011205678805708885, + 0.009680325165390968, + 0.014281926676630974, + 0.006197272799909115, + 0.006859284359961748, + 0.006619070190936327, + 0.00442670751363039, + 0.017214415594935417, + 0.015646908432245255, + 0.021006405353546143, + 0.016223780810832977, + 0.011529545299708843, + 0.015637101605534554, + 0.015577202662825584, + 0.019272638484835625, + 0.02156923897564411, + 0.021591410040855408, + 0.025954218581318855, + 0.016994595527648926, + 0.014849227853119373, + 0.012778086587786674, + 0.014368386939167976, + 0.01100186724215746, + 0.014245633967220783, + 0.015848897397518158, + 0.009637154638767242, + 0.010538660921156406, + 0.007024951744824648, + 0.007692424580454826, + 0.011340150609612465, + 0.008686983026564121, + 0.02184816263616085, + 0.02282065711915493, + 0.008685767650604248, + 0.009392423555254936, + 0.004658195190131664, + 0.004126285202801228, + 0.006342126987874508, + 0.004226369317620993, + 0.0120963454246521, + 0.01244756206870079, + 0.010008340701460838, + 0.007945164106786251, + 0.004303173162043095, + 0.003468044800683856, + 0.003982712049037218, + 0.007107740733772516, + 0.028587812557816505, + 0.033382777124643326, + 0.011413177475333214, + 0.011125217191874981, + 0.011929094791412354, + 0.00765641126781702, + 0.006491398438811302, + 0.00764485402032733 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 4 + ], + "selected_reward": -0.029789874330163002 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_tensor([4619, 4619])_group_1_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_tensor([4619, 4619])_group_1_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f0cbf3e728da391aca0f7180ff937907ce6c05d8 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_tensor([4619, 4619])_group_1_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "tight close-up shot from a slightly upward angle, perfectly capturing the visual style of nostalgic 1980s animation, locks on a giant king crab, curled up on a dark, mossy volcanic rock beneath the murky seabed. The crab's vibrant orange-red carapace, covered in sharp spines and jagged textures, is rendered with the signature hand-painted cel-like quality of a hard sheen and large, flat, sharp-edged dark shadows. Its disproportionately large claws gently open and close in a slow, rhythmic pattern, revealing a hidden, awe-inspiring power. A faint beam of light from the upper corner of the frame penetrates the gloom of the deep sea, illuminating the top of the crab's shell. The background is a blurry, deep blue water, with wisps of tiny bubbles rising from beneath it. The entire picture is covered with a layer of obvious, analog film-like grain, which contrasts with the cool background of Prussian blue and dark green, creating a mysterious, nostalgic and tense atmosphere.", + "seed": 21850, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21850_dp_2_rank_4_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.14847425762563943, + "vlm_logprob": -0.14847425762563943, + "vlm_yes_no_margin": 3.45, + "vlm_token_credit": -0.14847425762563943, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01199803315103054 + ], + "opposite_logprob_tokens": [ + -4.636998176574707 + ], + "logprob_reward": -0.01199803315103054, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.01199803315103054, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1258256733417511, + 0.13060925900936127, + 0.10806189477443695, + 0.08544208109378815, + 0.07634148001670837, + 0.047518689185380936, + 0.04507138580083847, + 0.03225705027580261, + 0.3079351782798767, + 0.04093730077147484 + ], + "credit_map": [ + 0.024309219792485237, + 0.008586411364376545, + 0.018876567482948303, + 0.0149617288261652, + 0.007522749714553356, + 0.01616995967924595, + 0.010368864051997662, + 0.025030173361301422, + 0.026754969730973244, + 0.009780784137547016, + 0.018964780494570732, + 0.011821966618299484, + 0.010628703981637955, + 0.019917713478207588, + 0.013591607101261616, + 0.019148727878928185, + 0.01973078027367592, + 0.009297854267060757, + 0.015725770965218544, + 0.01522536389529705, + 0.010422464460134506, + 0.011981387622654438, + 0.010536666959524155, + 0.015141601674258709, + 0.01523316279053688, + 0.006902717519551516, + 0.013239008374512196, + 0.011694425716996193, + 0.00553765008226037, + 0.012401637621223927, + 0.007190274074673653, + 0.01324321236461401, + 0.012107418850064278, + 0.005361937452107668, + 0.014391034841537476, + 0.011049757711589336, + 0.006155743729323149, + 0.012632162310183048, + 0.007483256980776787, + 0.007160171400755644, + 0.007966757752001286, + 0.003777973586693406, + 0.00923546589910984, + 0.005467745941132307, + 0.004116884432733059, + 0.005250621121376753, + 0.0057375081814825535, + 0.005965732503682375, + 0.007661525625735521, + 0.0030784131959080696, + 0.007588895037770271, + 0.0049455673433840275, + 0.0034778474364429712, + 0.006589542143046856, + 0.005768324248492718, + 0.0059612710028886795, + 0.006439370568841696, + 0.0028523884247988462, + 0.005012406501919031, + 0.0026964477729052305, + 0.0024144966155290604, + 0.00416230596601963, + 0.0040994747541844845, + 0.004580159671604633, + 0.061381034553050995, + 0.022799205034971237, + 0.03980977088212967, + 0.03489622473716736, + 0.02444949746131897, + 0.0515679270029068, + 0.034462109208106995, + 0.038569431751966476, + 0.0064533911645412445, + 0.0029307473450899124, + 0.00602144468575716, + 0.005030700005590916, + 0.0036380859091877937, + 0.004204503260552883, + 0.007609426975250244, + 0.005049002822488546 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 1, + 0 + ], + "selected_reward": -0.01199803315103054 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062064193189144135 + ], + "opposite_logprob_tokens": [ + -2.8120641708374023 + ], + "logprob_reward": -0.062064193189144135, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062064193189144135, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24863992631435394, + 0.18947985768318176, + 0.07317817211151123, + 0.0462631918489933, + 0.08725766837596893, + 0.11299340426921844, + 0.050371553748846054, + 0.04090841859579086, + 0.1076110452413559, + 0.04329680651426315 + ], + "credit_map": [ + 0.044200509786605835, + 0.019963476806879044, + 0.038350485265254974, + 0.029323959723114967, + 0.019670825451612473, + 0.027924738824367523, + 0.02387097477912903, + 0.045334942638874054, + 0.04057006910443306, + 0.015224985778331757, + 0.030474921688437462, + 0.020839983597397804, + 0.019167548045516014, + 0.019821647554636, + 0.019604887813329697, + 0.023775797337293625, + 0.013241852633655071, + 0.004785885103046894, + 0.01700889691710472, + 0.008011428639292717, + 0.006265909411013126, + 0.009509081020951271, + 0.004698374308645725, + 0.009656742215156555, + 0.007499828469008207, + 0.0037760394625365734, + 0.007755808066576719, + 0.00623071426525712, + 0.002676913980394602, + 0.006048111245036125, + 0.0045069423504173756, + 0.00776883028447628, + 0.019521739333868027, + 0.004280958790332079, + 0.015751026570796967, + 0.006950395181775093, + 0.004479350056499243, + 0.021969521418213844, + 0.005263902246952057, + 0.009040769189596176, + 0.021609753370285034, + 0.008363214321434498, + 0.01779027469456196, + 0.010256607085466385, + 0.007355962879955769, + 0.01971946656703949, + 0.010489027947187424, + 0.017409076914191246, + 0.009086658246815205, + 0.004241518676280975, + 0.006998965982347727, + 0.005825765896588564, + 0.0038436674512922764, + 0.00703298207372427, + 0.005799230188131332, + 0.00754275918006897, + 0.008074397221207619, + 0.0026955497451126575, + 0.006414043251425028, + 0.004828542470932007, + 0.0026797358877956867, + 0.0051651098765432835, + 0.0045969379134476185, + 0.00645409943535924, + 0.030633648857474327, + 0.006004059221595526, + 0.009799872525036335, + 0.007384497206658125, + 0.005143006332218647, + 0.028341392055153847, + 0.010929780080914497, + 0.009374779649078846, + 0.006505044177174568, + 0.0024758477229624987, + 0.006423985119909048, + 0.004139115568250418, + 0.0024619020987302065, + 0.007630244828760624, + 0.00797313079237938, + 0.0056875329464674 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.062064193189144135 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017177602276206017 + ], + "opposite_logprob_tokens": [ + -6.376717567443848 + ], + "logprob_reward": -0.0017177602276206017, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017177602276206017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.30210691690444946, + 0.25204870104789734, + 0.060155708342790604, + 0.04277503117918968, + 0.06885362416505814, + 0.06506288796663284, + 0.05138391628861427, + 0.036883287131786346, + 0.08600135147571564, + 0.034728534519672394 + ], + "credit_map": [ + 0.05739039555191994, + 0.02152886614203453, + 0.05371249467134476, + 0.031305767595767975, + 0.02137252874672413, + 0.035787247121334076, + 0.03162175044417381, + 0.0493878573179245, + 0.04849616065621376, + 0.02094043232500553, + 0.03696053847670555, + 0.033063631504774094, + 0.023468658328056335, + 0.02577507309615612, + 0.02803369238972664, + 0.03531051427125931, + 0.009776798076927662, + 0.004915180616080761, + 0.011950580403208733, + 0.009221065789461136, + 0.004379472695291042, + 0.006409198511391878, + 0.0051619792357087135, + 0.008341435343027115, + 0.006794854998588562, + 0.003484484273940325, + 0.008283905684947968, + 0.007449064403772354, + 0.003066971432417631, + 0.004001423716545105, + 0.00408746674656868, + 0.00560685945674777, + 0.013689091429114342, + 0.003977084532380104, + 0.015164587646722794, + 0.005313341971486807, + 0.0032924185506999493, + 0.015881044790148735, + 0.003617973532527685, + 0.00791808683425188, + 0.012097416445612907, + 0.005077438894659281, + 0.00801053550094366, + 0.006199842318892479, + 0.004389538895338774, + 0.013134324923157692, + 0.00780155323445797, + 0.008352240547537804, + 0.008127509616315365, + 0.00407393416389823, + 0.009764601476490498, + 0.005305589642375708, + 0.0047404952347278595, + 0.006075101438909769, + 0.005898612085729837, + 0.0073980712331831455, + 0.005885769613087177, + 0.0027815774083137512, + 0.006917161867022514, + 0.0034122804645448923, + 0.002507193246856332, + 0.004331869538873434, + 0.003624413628131151, + 0.007423022296279669, + 0.01590302214026451, + 0.0052363816648721695, + 0.008794920518994331, + 0.008256443776190281, + 0.005831711459904909, + 0.02064594253897667, + 0.012350487522780895, + 0.00898244883865118, + 0.0059131477028131485, + 0.0022221484687179327, + 0.005441128741949797, + 0.0038108746521174908, + 0.0021884129382669926, + 0.005980427376925945, + 0.004919083323329687, + 0.004253309685736895 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 8 + ], + "selected_reward": -0.0017177602276206017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328827142715454 + ], + "opposite_logprob_tokens": [ + -0.7578827142715454 + ], + "logprob_reward": -0.6328827142715454, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328827142715454, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17358526587486267, + 0.09767938405275345, + 0.072079136967659, + 0.05396876484155655, + 0.1225489154458046, + 0.08939669281244278, + 0.0953235924243927, + 0.025755666196346283, + 0.2377711683511734, + 0.03189142793416977 + ], + "credit_map": [ + 0.027757830917835236, + 0.01413320004940033, + 0.02175520919263363, + 0.025121640413999557, + 0.01401806902140379, + 0.018468450754880905, + 0.018975354731082916, + 0.033355507999658585, + 0.01886911876499653, + 0.005842963233590126, + 0.016821233555674553, + 0.012821484357118607, + 0.005990492179989815, + 0.013891372829675674, + 0.01001733634620905, + 0.013425379991531372, + 0.01510282326489687, + 0.005815077107399702, + 0.01015794649720192, + 0.011349366046488285, + 0.006527348887175322, + 0.007448074873536825, + 0.006999117322266102, + 0.008679380640387535, + 0.008819661103188992, + 0.003926384262740612, + 0.009958433918654919, + 0.008203684352338314, + 0.0030487023759633303, + 0.00513065978884697, + 0.005458477884531021, + 0.009422753937542439, + 0.0249725840985775, + 0.006164811085909605, + 0.020622966811060905, + 0.01138126477599144, + 0.006755353417247534, + 0.032621804624795914, + 0.007264117244631052, + 0.01276601292192936, + 0.014481858350336552, + 0.006343610119074583, + 0.016382673755288124, + 0.009505496360361576, + 0.006026316434144974, + 0.014373786747455597, + 0.009152742102742195, + 0.01313020195811987, + 0.016980567947030067, + 0.006979205179959536, + 0.0183499064296484, + 0.009961753152310848, + 0.007791203446686268, + 0.013110884465277195, + 0.009875664487481117, + 0.012274395674467087, + 0.004643258638679981, + 0.0019876568112522364, + 0.003933832515031099, + 0.0023995181545615196, + 0.0017390493303537369, + 0.00518119428306818, + 0.0030125111807137728, + 0.0028586427215486765, + 0.040607187896966934, + 0.020175212994217873, + 0.04155317321419716, + 0.03361477330327034, + 0.020445208996534348, + 0.023089751601219177, + 0.023108232766389847, + 0.035177621990442276, + 0.006156228017061949, + 0.0022344370372593403, + 0.0056039514020085335, + 0.002883759094402194, + 0.001980770379304886, + 0.005470383912324905, + 0.0034134024754166603, + 0.004148491192609072 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.6328827142715454 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708587288856506 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708587288856506, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708587288856506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4367358684539795, + 0.09464867413043976, + 0.0449565164744854, + 0.0500488206744194, + 0.14360716938972473, + 0.05219313129782677, + 0.058697376400232315, + 0.020770415663719177, + 0.06326508522033691, + 0.03507692366838455 + ], + "credit_map": [ + 0.07968275249004364, + 0.03382326290011406, + 0.07329971343278885, + 0.04762159287929535, + 0.03339848667383194, + 0.050876595079898834, + 0.04234258085489273, + 0.0756908655166626, + 0.018819672986865044, + 0.00790141336619854, + 0.012992029078304768, + 0.012223837897181511, + 0.008363479748368263, + 0.010135229676961899, + 0.011755228042602539, + 0.012457780539989471, + 0.008295497857034206, + 0.004085111897438765, + 0.008813479915261269, + 0.004549001809209585, + 0.0037158611230552197, + 0.005512565840035677, + 0.0032297098077833652, + 0.006755286827683449, + 0.010104476474225521, + 0.003520037280395627, + 0.009090399369597435, + 0.004596505779772997, + 0.0029492524918168783, + 0.010991849936544895, + 0.002643100218847394, + 0.00615319749340415, + 0.031154654920101166, + 0.008741353638470173, + 0.02557116374373436, + 0.010521822609007359, + 0.006841639056801796, + 0.03594629466533661, + 0.008354558609426022, + 0.016475671902298927, + 0.01006380096077919, + 0.0032801807392388582, + 0.008275806903839111, + 0.004755222704261541, + 0.003406540723517537, + 0.012033098377287388, + 0.004229388199746609, + 0.006149095017462969, + 0.010683701373636723, + 0.004471109248697758, + 0.012312736362218857, + 0.004779758397489786, + 0.00376090663485229, + 0.007880959659814835, + 0.00568121112883091, + 0.009126991964876652, + 0.004442965146154165, + 0.0014017466455698013, + 0.004003520589321852, + 0.0016745455795899034, + 0.0012240654323250055, + 0.002604220760986209, + 0.0016487168613821268, + 0.003770634764805436, + 0.009077704511582851, + 0.00446802843362093, + 0.008927744813263416, + 0.007272553630173206, + 0.00510326586663723, + 0.012656140141189098, + 0.006599243264645338, + 0.009160405024886131, + 0.006310885306447744, + 0.0020697356667369604, + 0.006385394837707281, + 0.003324252786114812, + 0.002188927261158824, + 0.0055399490520358086, + 0.004714410286396742, + 0.004543369635939598 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 1 + ], + "selected_reward": -0.033708587288856506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight close-up shot from a slightly upward angle, perfectly capturing the visual style of nostalgic 1980s animation, locks on a giant king crab, curled up on a dark, mossy volcanic rock beneath the murky seabed. The crab's vibrant orange-red carapace, covered in sharp spines and jagged textures, is rendered with the signature hand-painted cel-like quality of a hard sheen and large, flat, sharp-edged dark shadows. Its disproportionately large claws gently open and close in a slow, rhythmic pattern, revealing a hidden, awe-inspiring power. A faint beam of light from the upper corner of the frame penetrates the gloom of the deep sea, illuminating the top of the crab's shell. The background is a blurry, deep blue water, with wisps of tiny bubbles rising from beneath it. The entire picture is covered with a layer of obvious, analog film-like grain, which contrasts with the cool background of Prussian blue and dark green, creating a mysterious, nostalgic and tense atmosphere.", + "seed": 21851, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21851_dp_2_rank_4_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.172473105462268, + "vlm_logprob": -0.172473105462268, + "vlm_yes_no_margin": 3.4, + "vlm_token_credit": -0.172473105462268, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010134783573448658 + ], + "opposite_logprob_tokens": [ + -4.885134696960449 + ], + "logprob_reward": -0.010134783573448658, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010134783573448658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3336911201477051, + 0.09660107642412186, + 0.08488068729639053, + 0.052276186645030975, + 0.0523674450814724, + 0.1487054079771042, + 0.033636875450611115, + 0.02483956702053547, + 0.12668702006340027, + 0.04631459340453148 + ], + "credit_map": [ + 0.02489880472421646, + 0.028935281559824944, + 0.058833830058574677, + 0.05612565949559212, + 0.03916944935917854, + 0.04144132509827614, + 0.045377109199762344, + 0.038909662514925, + 0.007956779561936855, + 0.007657031994313002, + 0.007373131811618805, + 0.016088591888546944, + 0.009405465796589851, + 0.01599518209695816, + 0.014263834804296494, + 0.017861053347587585, + 0.0058079068548977375, + 0.008629271760582924, + 0.011992905288934708, + 0.01335560530424118, + 0.009460018947720528, + 0.009164477698504925, + 0.012821913696825504, + 0.013648594729602337, + 0.006052679382264614, + 0.004428967833518982, + 0.008121682330965996, + 0.00635951803997159, + 0.005685117561370134, + 0.0040745181031525135, + 0.009398305788636208, + 0.00815539713948965, + 0.004045478999614716, + 0.005407155025750399, + 0.007485808804631233, + 0.006979471538215876, + 0.0058821444399654865, + 0.005875671748071909, + 0.00929046981036663, + 0.007401245180517435, + 0.009101259522140026, + 0.008639181964099407, + 0.009674716740846634, + 0.049100738018751144, + 0.005991731304675341, + 0.007479631341993809, + 0.014407597482204437, + 0.044310543686151505, + 0.003406758652999997, + 0.00385132129304111, + 0.005399416666477919, + 0.005058192647993565, + 0.004138725344091654, + 0.002870467957109213, + 0.004285437520593405, + 0.004626555368304253, + 0.0021155753638595343, + 0.0020022306125611067, + 0.004099464975297451, + 0.004172659944742918, + 0.002809072146192193, + 0.002552436664700508, + 0.0041457475163042545, + 0.0029423802625387907, + 0.012766098603606224, + 0.011915628798305988, + 0.015242844820022583, + 0.01614154502749443, + 0.014797595329582691, + 0.01351343747228384, + 0.023172717541456223, + 0.019137145951390266, + 0.004268839489668608, + 0.0031892242841422558, + 0.009161712601780891, + 0.00757226487621665, + 0.0039859735406935215, + 0.004344260785728693, + 0.007445810362696648, + 0.006346510723233223 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 8 + ], + "selected_reward": -0.010134783573448658 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0620647557079792 + ], + "opposite_logprob_tokens": [ + -2.8120646476745605 + ], + "logprob_reward": -0.0620647557079792, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.0620647557079792, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5361026525497437, + 0.0660531297326088, + 0.09063724428415298, + 0.027310103178024292, + 0.020755477249622345, + 0.12177485227584839, + 0.06605643779039383, + 0.018334107473492622, + 0.03311736136674881, + 0.019858600571751595 + ], + "credit_map": [ + 0.03351733461022377, + 0.030892357230186462, + 0.0458311103284359, + 0.1661522537469864, + 0.02176978811621666, + 0.051811348646879196, + 0.033954307436943054, + 0.15217415988445282, + 0.008315605111420155, + 0.005644702352583408, + 0.012581026181578636, + 0.008450150489807129, + 0.006559029687196016, + 0.008204997517168522, + 0.007782027591019869, + 0.008515587076544762, + 0.014240668155252934, + 0.006812243722379208, + 0.019895771518349648, + 0.009179450571537018, + 0.008636074140667915, + 0.006808519829064608, + 0.012770365923643112, + 0.012294149026274681, + 0.00271616131067276, + 0.002314458368346095, + 0.004544700495898724, + 0.0027333064936101437, + 0.0033799386583268642, + 0.002174684777855873, + 0.005997928325086832, + 0.003448924282565713, + 0.00201131752692163, + 0.002323235385119915, + 0.0033409418538212776, + 0.0031218777876347303, + 0.0022180313244462013, + 0.001602614065632224, + 0.0035665978211909533, + 0.0025708614848554134, + 0.006782287731766701, + 0.00579615356400609, + 0.008604153990745544, + 0.039881665259599686, + 0.003455560654401779, + 0.00576260918751359, + 0.012562636286020279, + 0.038929782807826996, + 0.007148573640733957, + 0.004938928410410881, + 0.011465995572507381, + 0.01138483639806509, + 0.005044019781053066, + 0.004936021752655506, + 0.00903307180851698, + 0.012104997411370277, + 0.002498926827684045, + 0.0014822655357420444, + 0.0038621826097369194, + 0.0020699803717434406, + 0.0020987193565815687, + 0.00180247169919312, + 0.0023600070271641016, + 0.0021595540456473827, + 0.002055536722764373, + 0.002387009095400572, + 0.005122401285916567, + 0.005026673898100853, + 0.0034504218492656946, + 0.0032528007868677378, + 0.005499578081071377, + 0.006322940811514854, + 0.0020470928866416216, + 0.001847802777774632, + 0.0034204975236207247, + 0.003071817336603999, + 0.0025308323092758656, + 0.0016429792158305645, + 0.0024864773731678724, + 0.002811100799590349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.0620647557079792 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002493488835170865 + ], + "opposite_logprob_tokens": [ + -6.002493381500244 + ], + "logprob_reward": -0.002493488835170865, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002493488835170865, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.4563700556755066, + 0.05972835049033165, + 0.0545448437333107, + 0.07118824869394302, + 0.07605545222759247, + 0.0725213810801506, + 0.08042241632938385, + 0.03898042440414429, + 0.06033257022500038, + 0.02985633723437786 + ], + "credit_map": [ + 0.029225774109363556, + 0.023727310821413994, + 0.043011508882045746, + 0.1253271996974945, + 0.016956288367509842, + 0.04784072935581207, + 0.03892160579562187, + 0.13135956227779388, + 0.00542000588029623, + 0.005233872216194868, + 0.0076085785403847694, + 0.01024977583438158, + 0.005052225664258003, + 0.007126708514988422, + 0.006949238013476133, + 0.01208794116973877, + 0.00359003315679729, + 0.004262940492480993, + 0.007780368905514479, + 0.00756351463496685, + 0.005274573341012001, + 0.004342483822256327, + 0.013639681041240692, + 0.008091241121292114, + 0.005694645456969738, + 0.007180873304605484, + 0.011228946968913078, + 0.011112433858215809, + 0.009265699423849583, + 0.005660668946802616, + 0.011177445761859417, + 0.009867524728178978, + 0.010371013544499874, + 0.006912366487085819, + 0.010291717946529388, + 0.009452017955482006, + 0.0072391452267766, + 0.005823063664138317, + 0.013839331455528736, + 0.012126786634325981, + 0.004997896496206522, + 0.003871341235935688, + 0.007073947694152594, + 0.022594274953007698, + 0.0028699671383947134, + 0.0034108166582882404, + 0.007497115526348352, + 0.020206009969115257, + 0.006595614366233349, + 0.004190740175545216, + 0.0135960029438138, + 0.015925709158182144, + 0.0050459960475564, + 0.005464766174554825, + 0.010471017099916935, + 0.019132563844323158, + 0.0032197395339608192, + 0.003547671716660261, + 0.004258440341800451, + 0.006299652624875307, + 0.004746109712868929, + 0.0038771203253418207, + 0.007719056680798531, + 0.005312628112733364, + 0.00645173666998744, + 0.004184212069958448, + 0.007364538498222828, + 0.007833058014512062, + 0.006640071049332619, + 0.006352846976369619, + 0.011975958943367004, + 0.009530133567750454, + 0.003929834347218275, + 0.002800216432660818, + 0.0029569172766059637, + 0.004353080410510302, + 0.004364835564047098, + 0.003003487130627036, + 0.004630568902939558, + 0.00381739204749465 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.002493488835170865 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.7578619122505188 + ], + "opposite_logprob_tokens": [ + -0.6328619122505188 + ], + "logprob_reward": -0.7578619122505188, + "yes_no_margin_reward": -0.125, + "token_credit_reward": -0.7578619122505188, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.1554284244775772, + 0.06696245074272156, + 0.08027888834476471, + 0.06144091114401817, + 0.08824045211076736, + 0.20583486557006836, + 0.07395903766155243, + 0.09041973948478699, + 0.09548591822385788, + 0.08194936066865921 + ], + "credit_map": [ + 0.01707412675023079, + 0.014047146774828434, + 0.019554968923330307, + 0.03350505605340004, + 0.015369454398751259, + 0.011325669474899769, + 0.01862783543765545, + 0.02592415176331997, + 0.006568822078406811, + 0.006346570793539286, + 0.007983803749084473, + 0.009104076772928238, + 0.005943316500633955, + 0.009173611178994179, + 0.009653689339756966, + 0.012188556604087353, + 0.007667754776775837, + 0.008448346517980099, + 0.012222224846482277, + 0.009836708195507526, + 0.009126495569944382, + 0.008765047416090965, + 0.012624125927686691, + 0.01158818043768406, + 0.00600503059104085, + 0.006347560789436102, + 0.008836787194013596, + 0.007273880299180746, + 0.008584776893258095, + 0.0051607717759907246, + 0.009464598260819912, + 0.009767504408955574, + 0.008009002543985844, + 0.008733450435101986, + 0.009484506212174892, + 0.011722748167812824, + 0.009886928834021091, + 0.00929505005478859, + 0.014634336344897747, + 0.016474425792694092, + 0.011826908215880394, + 0.009924172423779964, + 0.012384604662656784, + 0.0646819919347763, + 0.007831399329006672, + 0.01116615254431963, + 0.022574108093976974, + 0.06544550508260727, + 0.006929715164005756, + 0.006240135990083218, + 0.013385898433625698, + 0.010943367145955563, + 0.007681421469897032, + 0.005898443982005119, + 0.012308899313211441, + 0.01057114452123642, + 0.00836658850312233, + 0.007427690085023642, + 0.00920436717569828, + 0.014125771820545197, + 0.010843338444828987, + 0.01116570271551609, + 0.01903052255511284, + 0.01025574654340744, + 0.010115759447216988, + 0.006997884251177311, + 0.01714855432510376, + 0.012800372205674648, + 0.009165163151919842, + 0.006715580355376005, + 0.01838545687496662, + 0.014157145284116268, + 0.005346041638404131, + 0.006014651618897915, + 0.010777557268738747, + 0.011269285343587399, + 0.012592041864991188, + 0.008672776632010937, + 0.01641980931162834, + 0.010857196524739265 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 8 + ], + "selected_reward": -0.7578619122505188 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02981058694422245 + ], + "opposite_logprob_tokens": [ + -3.529810667037964 + ], + "logprob_reward": -0.02981058694422245, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02981058694422245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.541640043258667, + 0.0935027226805687, + 0.03412792831659317, + 0.03292228654026985, + 0.05424563214182854, + 0.032588351517915726, + 0.10278590023517609, + 0.030302532017230988, + 0.048748038709163666, + 0.029136590659618378 + ], + "credit_map": [ + 0.036743443459272385, + 0.036029618233442307, + 0.056346505880355835, + 0.16320577263832092, + 0.02449672482907772, + 0.05300397425889969, + 0.03642325475811958, + 0.1353907287120819, + 0.006778480019420385, + 0.0073774573393166065, + 0.012304064817726612, + 0.021969791501760483, + 0.006576240062713623, + 0.008336006663739681, + 0.011059627868235111, + 0.01910104975104332, + 0.004379868041723967, + 0.00341510190628469, + 0.0035823979414999485, + 0.004027707502245903, + 0.005061768461018801, + 0.0032580397091805935, + 0.005053529981523752, + 0.005349514074623585, + 0.0025862453039735556, + 0.003091466845944524, + 0.0038457801565527916, + 0.003583983052521944, + 0.0038020305801182985, + 0.0023222824092954397, + 0.008636562153697014, + 0.00505393510684371, + 0.007906414568424225, + 0.0050156088545918465, + 0.0069039249792695045, + 0.006256495136767626, + 0.00544453039765358, + 0.004071817733347416, + 0.010161268524825573, + 0.008485565893352032, + 0.0033337522763758898, + 0.0027003467548638582, + 0.00326639786362648, + 0.005167338997125626, + 0.0024142353795468807, + 0.0021060786675661802, + 0.006814389023929834, + 0.006785813253372908, + 0.008883679285645485, + 0.006494384258985519, + 0.017745638266205788, + 0.01853867433965206, + 0.008229736238718033, + 0.008164167404174805, + 0.012978932820260525, + 0.021750684827566147, + 0.0026777933817356825, + 0.002588561037555337, + 0.0053869253024458885, + 0.004950672388076782, + 0.0029591976199299097, + 0.0034146574325859547, + 0.004597838036715984, + 0.0037268870510160923, + 0.004605857189744711, + 0.0037632803432643414, + 0.006925635971128941, + 0.006664264015853405, + 0.004455955699086189, + 0.006655216682702303, + 0.008632334880530834, + 0.007045493461191654, + 0.0038033188320696354, + 0.0024761855602264404, + 0.0036428431048989296, + 0.003845479805022478, + 0.003560759359970689, + 0.003302034456282854, + 0.004711101297289133, + 0.0037948687095195055 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.02981058694422245 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "No", + "Yes" + ] + }, + { + "prompt": "tight close-up shot from a slightly upward angle, perfectly capturing the visual style of nostalgic 1980s animation, locks on a giant king crab, curled up on a dark, mossy volcanic rock beneath the murky seabed. The crab's vibrant orange-red carapace, covered in sharp spines and jagged textures, is rendered with the signature hand-painted cel-like quality of a hard sheen and large, flat, sharp-edged dark shadows. Its disproportionately large claws gently open and close in a slow, rhythmic pattern, revealing a hidden, awe-inspiring power. A faint beam of light from the upper corner of the frame penetrates the gloom of the deep sea, illuminating the top of the crab's shell. The background is a blurry, deep blue water, with wisps of tiny bubbles rising from beneath it. The entire picture is covered with a layer of obvious, analog film-like grain, which contrasts with the cool background of Prussian blue and dark green, creating a mysterious, nostalgic and tense atmosphere.", + "seed": 21852, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21852_dp_3_rank_6_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.09468809759709984, + "vlm_logprob": -0.09468809759709984, + "vlm_yes_no_margin": 3.625, + "vlm_token_credit": -0.09468809759709984, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.010469396598637104 + ], + "opposite_logprob_tokens": [ + -4.885469436645508 + ], + "logprob_reward": -0.010469396598637104, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.010469396598637104, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27815356850624084, + 0.046304043382406235, + 0.022740833461284637, + 0.049798332154750824, + 0.03609536588191986, + 0.07398474216461182, + 0.06134885177016258, + 0.15455415844917297, + 0.058645471930503845, + 0.2183745950460434 + ], + "credit_map": [ + 0.015436816960573196, + 0.011823814362287521, + 0.014140617102384567, + 0.10833598673343658, + 0.007911236956715584, + 0.010889934375882149, + 0.05563552305102348, + 0.053979646414518356, + 0.00476088235154748, + 0.005032055079936981, + 0.005752529948949814, + 0.006959662772715092, + 0.004718853626400232, + 0.004390764515846968, + 0.005576160736382008, + 0.009113132953643799, + 0.0022416028659790754, + 0.003160986350849271, + 0.0026280772872269154, + 0.003862048266455531, + 0.0030292803421616554, + 0.0023563869763165712, + 0.0024752775207161903, + 0.0029871740844100714, + 0.004966430831700563, + 0.004509017802774906, + 0.006394939962774515, + 0.007428196258842945, + 0.005285437218844891, + 0.0058213770389556885, + 0.009730863384902477, + 0.005662071518599987, + 0.004122603219002485, + 0.004828524310141802, + 0.004010024014860392, + 0.004321340937167406, + 0.006064815446734428, + 0.005662277340888977, + 0.003616809379309416, + 0.0034689733292907476, + 0.008683515712618828, + 0.006797483190894127, + 0.0089341439306736, + 0.0069487146101891994, + 0.006809687707573175, + 0.016972368583083153, + 0.011434788815677166, + 0.007404041476547718, + 0.0073387157171964645, + 0.006009106058627367, + 0.007741974201053381, + 0.006976739037781954, + 0.006075878161936998, + 0.008763859048485756, + 0.007584741339087486, + 0.010857841931283474, + 0.01696690171957016, + 0.023236993700265884, + 0.016615403816103935, + 0.017620932310819626, + 0.021399682387709618, + 0.018351566046476364, + 0.01650233380496502, + 0.023860348388552666, + 0.00473769148811698, + 0.006636613514274359, + 0.005899837706238031, + 0.006733588874340057, + 0.007218299899250269, + 0.008178764022886753, + 0.010406079702079296, + 0.0088345967233181, + 0.020311249420046806, + 0.0232223030179739, + 0.02167329005897045, + 0.02379726432263851, + 0.02444952353835106, + 0.02838023193180561, + 0.02373678609728813, + 0.052803944796323776 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 7 + ], + "selected_reward": -0.010469396598637104 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08906956017017365 + ], + "opposite_logprob_tokens": [ + -2.4640696048736572 + ], + "logprob_reward": -0.08906956017017365, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08906956017017365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05464065447449684, + 0.1274404227733612, + 0.07295390963554382, + 0.029924627393484116, + 0.06311056017875671, + 0.11348925530910492, + 0.0434788316488266, + 0.1369708776473999, + 0.12358756363391876, + 0.2344033420085907 + ], + "credit_map": [ + 0.00443505123257637, + 0.004604744259268045, + 0.0060803587548434734, + 0.009281140752136707, + 0.004330539610236883, + 0.004449806641787291, + 0.008012457750737667, + 0.013446551747620106, + 0.012970048934221268, + 0.01556122861802578, + 0.016293982043862343, + 0.019354436546564102, + 0.012581493705511093, + 0.011858748272061348, + 0.014706365764141083, + 0.02411411888897419, + 0.007908905856311321, + 0.009515034034848213, + 0.008094907738268375, + 0.013748486526310444, + 0.009026539511978626, + 0.006696166004985571, + 0.00638428283855319, + 0.011579587124288082, + 0.004170261323451996, + 0.0034068168606609106, + 0.004106989596039057, + 0.0052433982491493225, + 0.0029964474961161613, + 0.0037352528888732195, + 0.002922098385170102, + 0.003343363758176565, + 0.007326969876885414, + 0.006231717765331268, + 0.008780553005635738, + 0.008523018099367619, + 0.006455032620579004, + 0.01277295034378767, + 0.005021926946938038, + 0.007998391054570675, + 0.018266456201672554, + 0.012710470706224442, + 0.012569558806717396, + 0.010191637091338634, + 0.01131946500390768, + 0.016210338100790977, + 0.012184062041342258, + 0.020037271082401276, + 0.005789123475551605, + 0.004355873446911573, + 0.0055748214945197105, + 0.004548795986920595, + 0.0037363525480031967, + 0.005536063574254513, + 0.0057307579554617405, + 0.008207042701542377, + 0.0183732770383358, + 0.01482411939650774, + 0.014605259522795677, + 0.008520616218447685, + 0.01387120597064495, + 0.029028162360191345, + 0.010073467157781124, + 0.027674775570631027, + 0.01423223502933979, + 0.011635774746537209, + 0.013977326452732086, + 0.010978111065924168, + 0.013013788498938084, + 0.013922497630119324, + 0.010182320140302181, + 0.0356455072760582, + 0.02451244182884693, + 0.024942707270383835, + 0.021678483113646507, + 0.024578982964158058, + 0.02778736688196659, + 0.03395909443497658, + 0.0211265180259943, + 0.055817749351263046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.08906956017017365 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001945747178979218 + ], + "opposite_logprob_tokens": [ + -6.251945972442627 + ], + "logprob_reward": -0.001945747178979218, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001945747178979218, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19469845294952393, + 0.059056706726551056, + 0.03315618634223938, + 0.024236872792243958, + 0.04493201896548271, + 0.08621396124362946, + 0.06625185161828995, + 0.07447674125432968, + 0.3098914325237274, + 0.10708587616682053 + ], + "credit_map": [ + 0.011323634535074234, + 0.008040372282266617, + 0.01136658526957035, + 0.06555125117301941, + 0.008212157525122166, + 0.007585523184388876, + 0.06548600643873215, + 0.01713288202881813, + 0.005463290028274059, + 0.008329632692039013, + 0.006062989588826895, + 0.01076051127165556, + 0.007770917844027281, + 0.005939003080129623, + 0.007055841386318207, + 0.007674508262425661, + 0.0025488690007478, + 0.004904880654066801, + 0.004215405322611332, + 0.0045936633832752705, + 0.005466452334076166, + 0.00310739828273654, + 0.003852098947390914, + 0.004467411432415247, + 0.0026485337875783443, + 0.0025839314330369234, + 0.0036204205825924873, + 0.004858647473156452, + 0.0025341068394482136, + 0.002628168324008584, + 0.002938167890533805, + 0.0024248925037682056, + 0.004066200461238623, + 0.005513888783752918, + 0.004776060581207275, + 0.006368396338075399, + 0.006372241768985987, + 0.00879956129938364, + 0.004649965558201075, + 0.00438569625839591, + 0.011613862589001656, + 0.009569073095917702, + 0.010553880594670773, + 0.010850416496396065, + 0.008098148740828037, + 0.012780772522091866, + 0.009138444438576698, + 0.013609351590275764, + 0.009744956158101559, + 0.007947094738483429, + 0.007548995316028595, + 0.0061334106139838696, + 0.007150929421186447, + 0.00863647647202015, + 0.005796284414827824, + 0.013293696567416191, + 0.008435140363872051, + 0.008806047029793262, + 0.007296853233128786, + 0.0067811934277415276, + 0.010378248058259487, + 0.016508284956216812, + 0.007886737585067749, + 0.008384224027395248, + 0.035374779254198074, + 0.032812509685754776, + 0.03460531309247017, + 0.02949550747871399, + 0.032565757632255554, + 0.03611228987574577, + 0.026635190472006798, + 0.08229003846645355, + 0.012383444234728813, + 0.013341174460947514, + 0.009544934146106243, + 0.010329588316380978, + 0.01300001423805952, + 0.015422951430082321, + 0.009998215362429619, + 0.023065531626343727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 9 + ], + "selected_reward": -0.001945747178979218 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3486664295196533 + ], + "opposite_logprob_tokens": [ + -1.2236664295196533 + ], + "logprob_reward": -0.3486664295196533, + "yes_no_margin_reward": 0.875, + "token_credit_reward": -0.3486664295196533, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13957579433918, + 0.05494712293148041, + 0.03367537260055542, + 0.04111411049962044, + 0.013948285020887852, + 0.037299104034900665, + 0.03030339628458023, + 0.10094121098518372, + 0.5047187805175781, + 0.04347684979438782 + ], + "credit_map": [ + 0.007516344543546438, + 0.0054727718234062195, + 0.006753675639629364, + 0.05291926488280296, + 0.00509291235357523, + 0.005246794782578945, + 0.04372430965304375, + 0.012849722057580948, + 0.0065286788158118725, + 0.006424244027584791, + 0.007221081759780645, + 0.006653377786278725, + 0.005635532550513744, + 0.005418782588094473, + 0.005148556549102068, + 0.011916864663362503, + 0.0032439325004816055, + 0.0034422031603753567, + 0.004795688204467297, + 0.005746887996792793, + 0.004860681481659412, + 0.0031674581114202738, + 0.0038055707700550556, + 0.004612952470779419, + 0.004667635075747967, + 0.003285724436864257, + 0.0048570213839411736, + 0.009094258770346642, + 0.004302221816033125, + 0.0036460210103541613, + 0.004521501250565052, + 0.006739722564816475, + 0.0012654847232624888, + 0.0012457219418138266, + 0.002011795062571764, + 0.0024570401292294264, + 0.0015076755080372095, + 0.0028713878709822893, + 0.0011915189679712057, + 0.0013976616319268942, + 0.0038046278059482574, + 0.005345101933926344, + 0.0039733098819851875, + 0.0036872541531920433, + 0.004516163375228643, + 0.006311459466814995, + 0.004875548649579287, + 0.004785637836903334, + 0.004178261384367943, + 0.003682316979393363, + 0.003797838930040598, + 0.0031612867023795843, + 0.00325447297655046, + 0.0033813980408012867, + 0.00308250286616385, + 0.005765318404883146, + 0.01294484082609415, + 0.015433603897690773, + 0.010454262606799603, + 0.00924820825457573, + 0.01302073523402214, + 0.012141288258135319, + 0.00882701389491558, + 0.018871253356337547, + 0.054400842636823654, + 0.055081769824028015, + 0.05480585992336273, + 0.04781363531947136, + 0.05802284926176071, + 0.05810817703604698, + 0.04449029266834259, + 0.13199536502361298, + 0.004402598831802607, + 0.004762963391840458, + 0.004271028563380241, + 0.0050591290928423405, + 0.005578143056482077, + 0.005317323375493288, + 0.004222778137773275, + 0.00986288208514452 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 7 + ], + "selected_reward": -0.3486664295196533 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023289354518055916 + ], + "opposite_logprob_tokens": [ + -3.773289442062378 + ], + "logprob_reward": -0.023289354518055916, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023289354518055916, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19982260465621948, + 0.17470400035381317, + 0.044396061450242996, + 0.05457745119929314, + 0.033340759575366974, + 0.06549026817083359, + 0.06936278194189072, + 0.11470954865217209, + 0.18234936892986298, + 0.06124706566333771 + ], + "credit_map": [ + 0.016664693132042885, + 0.013537595048546791, + 0.018747471272945404, + 0.02603677287697792, + 0.012967370450496674, + 0.01205782126635313, + 0.05445516109466553, + 0.04535573720932007, + 0.01673104800283909, + 0.020259253680706024, + 0.02124760113656521, + 0.027354860678315163, + 0.018883464857935905, + 0.016400128602981567, + 0.02221650816500187, + 0.031611159443855286, + 0.0047648269683122635, + 0.0048920344561338425, + 0.0054739005863666534, + 0.00671798549592495, + 0.005998530425131321, + 0.005669251549988985, + 0.004904173780232668, + 0.005975363310426474, + 0.00691088056191802, + 0.005200993735343218, + 0.007303106132894754, + 0.007280917838215828, + 0.006206714082509279, + 0.007853428833186626, + 0.00730455806478858, + 0.006516858469694853, + 0.003800072008743882, + 0.0038531292229890823, + 0.003641461255028844, + 0.0034490854013711214, + 0.004292580299079418, + 0.007762745022773743, + 0.0028337412513792515, + 0.0037079472094774246, + 0.0077957576140761375, + 0.008433626964688301, + 0.00692273722961545, + 0.007325438316911459, + 0.008647870272397995, + 0.011477325111627579, + 0.0059168036095798016, + 0.008970710448920727, + 0.007261031307280064, + 0.009649396874010563, + 0.007809867616742849, + 0.008481178432703018, + 0.010080228559672832, + 0.006803953088819981, + 0.009629348292946815, + 0.009647788479924202, + 0.014696788042783737, + 0.016006024554371834, + 0.01195183303207159, + 0.00782803725451231, + 0.014714933931827545, + 0.01741017773747444, + 0.008691969327628613, + 0.023409798741340637, + 0.01888512633740902, + 0.020624831318855286, + 0.01913776621222496, + 0.016932113096117973, + 0.023060623556375504, + 0.02006736397743225, + 0.016724422574043274, + 0.046917133033275604, + 0.0057892464101314545, + 0.006437493953853846, + 0.006018166430294514, + 0.0076243942603468895, + 0.007155041676014662, + 0.008505938574671745, + 0.006733267102390528, + 0.012983523309230804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.023289354518055916 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "tight close-up shot from a slightly upward angle, perfectly capturing the visual style of nostalgic 1980s animation, locks on a giant king crab, curled up on a dark, mossy volcanic rock beneath the murky seabed. The crab's vibrant orange-red carapace, covered in sharp spines and jagged textures, is rendered with the signature hand-painted cel-like quality of a hard sheen and large, flat, sharp-edged dark shadows. Its disproportionately large claws gently open and close in a slow, rhythmic pattern, revealing a hidden, awe-inspiring power. A faint beam of light from the upper corner of the frame penetrates the gloom of the deep sea, illuminating the top of the crab's shell. The background is a blurry, deep blue water, with wisps of tiny bubbles rising from beneath it. The entire picture is covered with a layer of obvious, analog film-like grain, which contrasts with the cool background of Prussian blue and dark green, creating a mysterious, nostalgic and tense atmosphere.", + "seed": 21853, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_1/index_4619_seed_21853_dp_3_rank_6_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 1, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06017589399125427, + "vlm_logprob": -0.06017589399125427, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.06017589399125427, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00956275686621666 + ], + "opposite_logprob_tokens": [ + -5.009562969207764 + ], + "logprob_reward": -0.00956275686621666, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00956275686621666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10670769214630127, + 0.09723401814699173, + 0.0862082913517952, + 0.04169546440243721, + 0.1042783185839653, + 0.06697475165128708, + 0.13562601804733276, + 0.2558574080467224, + 0.019218366593122482, + 0.08619966357946396 + ], + "credit_map": [ + 0.00914517417550087, + 0.006180719006806612, + 0.010971828363835812, + 0.014724910259246826, + 0.010717970319092274, + 0.01333114318549633, + 0.022258222103118896, + 0.01937772147357464, + 0.014659504406154156, + 0.00875093974173069, + 0.01082555204629898, + 0.008549551479518414, + 0.010859264992177486, + 0.010007668286561966, + 0.01728983409702778, + 0.016291705891489983, + 0.007798202335834503, + 0.011018417775630951, + 0.010911552235484123, + 0.009802734479308128, + 0.012657633982598782, + 0.00940818339586258, + 0.010873065330088139, + 0.013738501816987991, + 0.004796245601028204, + 0.004543857183307409, + 0.006277936510741711, + 0.005296887364238501, + 0.005153884179890156, + 0.0046704355627298355, + 0.006058984436094761, + 0.004897233098745346, + 0.008665057830512524, + 0.014161649160087109, + 0.016517044976353645, + 0.014777597971260548, + 0.011843587271869183, + 0.01398011390119791, + 0.012319948524236679, + 0.012013314291834831, + 0.007668804842978716, + 0.007559772580862045, + 0.00856513250619173, + 0.007362883538007736, + 0.006698526907712221, + 0.009782246313989162, + 0.01047931332141161, + 0.008858070708811283, + 0.017491592094302177, + 0.013784246519207954, + 0.015037640929222107, + 0.015881560742855072, + 0.012302764691412449, + 0.016209758818149567, + 0.014569004066288471, + 0.03034944273531437, + 0.03465375676751137, + 0.024068981409072876, + 0.031026139855384827, + 0.02697385661303997, + 0.02004721388220787, + 0.03241017833352089, + 0.02660417929291725, + 0.06007309630513191, + 0.0016725544119253755, + 0.0018698475323617458, + 0.002504392759874463, + 0.002184202428907156, + 0.0021632600110024214, + 0.0024599353782832623, + 0.0032333519775420427, + 0.0031308226753026247, + 0.009282237850129604, + 0.009581848978996277, + 0.012797100469470024, + 0.008394407108426094, + 0.008505345322191715, + 0.01099893357604742, + 0.00986502692103386, + 0.01677476242184639 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 0 + ], + "selected_reward": -0.00956275686621666 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11295044422149658 + ], + "opposite_logprob_tokens": [ + -2.237950325012207 + ], + "logprob_reward": -0.11295044422149658, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11295044422149658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2151724100112915, + 0.09987478703260422, + 0.2467065304517746, + 0.17390450835227966, + 0.08666752278804779, + 0.0379067100584507, + 0.03964739665389061, + 0.034121979027986526, + 0.020621534436941147, + 0.04537668824195862 + ], + "credit_map": [ + 0.025043759495019913, + 0.009947672486305237, + 0.015906384214758873, + 0.021649593487381935, + 0.018911922350525856, + 0.029075216501951218, + 0.03954492509365082, + 0.05509290099143982, + 0.012034935876727104, + 0.005071369465440512, + 0.0074503375217318535, + 0.008170738816261292, + 0.008188378997147083, + 0.014503326267004013, + 0.015150821767747402, + 0.02930486388504505, + 0.04835275188088417, + 0.01897748000919819, + 0.024110987782478333, + 0.01934797689318657, + 0.019790593534708023, + 0.028971951454877853, + 0.03628414124250412, + 0.050870630890131, + 0.022877322509884834, + 0.013628823682665825, + 0.02031528577208519, + 0.01924756169319153, + 0.01987709105014801, + 0.024734003469347954, + 0.02909599430859089, + 0.024128392338752747, + 0.00951982568949461, + 0.00714732613414526, + 0.010810207575559616, + 0.01120042521506548, + 0.006198867689818144, + 0.010457580909132957, + 0.01063514780253172, + 0.020698126405477524, + 0.0048506418243050575, + 0.005819143261760473, + 0.0041984617710113525, + 0.005026210565119982, + 0.00371054676361382, + 0.005220515187829733, + 0.004852739162743092, + 0.0042284466326236725, + 0.0045418329536914825, + 0.0044153956696391106, + 0.0044828965328633785, + 0.004438258707523346, + 0.0039236112497746944, + 0.005120856687426567, + 0.004675894044339657, + 0.0080486461520195, + 0.0030481009744107723, + 0.003886106889694929, + 0.005093872081488371, + 0.003389279590919614, + 0.003893355606123805, + 0.004041215870529413, + 0.003807628992944956, + 0.006962416227906942, + 0.002516149776056409, + 0.0017284429632127285, + 0.0025864613708108664, + 0.0017696944996714592, + 0.0016597466310486197, + 0.0031288627069443464, + 0.002807593671604991, + 0.004424579441547394, + 0.0054042283445596695, + 0.0055014705285429955, + 0.006440742872655392, + 0.004004476126283407, + 0.003238935489207506, + 0.006048245821148157, + 0.005117953289300203, + 0.00962062831968069 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.11295044422149658 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0019444384379312396 + ], + "opposite_logprob_tokens": [ + -6.251944541931152 + ], + "logprob_reward": -0.0019444384379312396, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.0019444384379312396, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19732564687728882, + 0.10278499126434326, + 0.35935574769973755, + 0.05063213035464287, + 0.0930531769990921, + 0.05670930817723274, + 0.03860985487699509, + 0.04724302142858505, + 0.019604453817009926, + 0.034681644290685654 + ], + "credit_map": [ + 0.02892127074301243, + 0.009024528786540031, + 0.015019644051790237, + 0.022373786196112633, + 0.016013773158192635, + 0.024745449423789978, + 0.031884294003248215, + 0.049342889338731766, + 0.017745545133948326, + 0.007407106459140778, + 0.01045115478336811, + 0.009452060796320438, + 0.0076575507409870625, + 0.011398479342460632, + 0.019588813185691833, + 0.019084276631474495, + 0.06699410080909729, + 0.03334491327404976, + 0.03573429957032204, + 0.025791697204113007, + 0.02254456840455532, + 0.04300635680556297, + 0.04863758757710457, + 0.08330223709344864, + 0.0050616455264389515, + 0.0041227322071790695, + 0.0058391522616147995, + 0.0060524288564920425, + 0.006060629617422819, + 0.007609979715198278, + 0.007289248518645763, + 0.008596313185989857, + 0.008291316218674183, + 0.008992738090455532, + 0.013535631820559502, + 0.011519866064190865, + 0.010776307433843613, + 0.011782309040427208, + 0.013860740698873997, + 0.014294268563389778, + 0.004712408874183893, + 0.006933712400496006, + 0.0056090871803462505, + 0.0070606982335448265, + 0.006322079803794622, + 0.007294054608792067, + 0.009082961827516556, + 0.009694305248558521, + 0.004839714616537094, + 0.004078676458448172, + 0.0049697537906467915, + 0.0046070232056081295, + 0.0029290132224559784, + 0.004845117684453726, + 0.004859858192503452, + 0.0074806963093578815, + 0.006286109331995249, + 0.005413055419921875, + 0.006806781515479088, + 0.00418637041002512, + 0.00389437866397202, + 0.005013778805732727, + 0.005132168997079134, + 0.010510381311178207, + 0.00205593416467309, + 0.0020555942319333553, + 0.0027642049826681614, + 0.0021074519027024508, + 0.0017345019150525331, + 0.0021743602119386196, + 0.0027460299897938967, + 0.003966376185417175, + 0.004854438826441765, + 0.0036635962314903736, + 0.003777108620852232, + 0.0058624716475605965, + 0.0025049950927495956, + 0.004868954420089722, + 0.004605627618730068, + 0.004544450435787439 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.0019444384379312396 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.16035135090351105 + ], + "opposite_logprob_tokens": [ + -1.9103513956069946 + ], + "logprob_reward": -0.16035135090351105, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.16035135090351105, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13701528310775757, + 0.07306653261184692, + 0.3014858663082123, + 0.08135219663381577, + 0.06309576332569122, + 0.10817164927721024, + 0.0646764487028122, + 0.09342797845602036, + 0.025960510596632957, + 0.05174769461154938 + ], + "credit_map": [ + 0.01899556629359722, + 0.011346286162734032, + 0.014977755956351757, + 0.01676015742123127, + 0.015510710887610912, + 0.011668052524328232, + 0.02065367065370083, + 0.02710309438407421, + 0.011832800693809986, + 0.006064020097255707, + 0.010820094496011734, + 0.008053810335695744, + 0.0055913906544446945, + 0.008291637524962425, + 0.010551766492426395, + 0.011861016973853111, + 0.05111975222826004, + 0.025334758684039116, + 0.03147592395544052, + 0.023461222648620605, + 0.01861698552966118, + 0.036653097718954086, + 0.04042970761656761, + 0.07439445704221725, + 0.00783106591552496, + 0.007579758297652006, + 0.010905119590461254, + 0.008892558515071869, + 0.008379432372748852, + 0.008757372386753559, + 0.0153767429292202, + 0.013630147092044353, + 0.0060880547389388084, + 0.012076780200004578, + 0.007123405579477549, + 0.00602889247238636, + 0.008316408842802048, + 0.00647946959361434, + 0.007749959360808134, + 0.009232793003320694, + 0.0118233198300004, + 0.015580189414322376, + 0.013492308557033539, + 0.012997152283787727, + 0.01081911288201809, + 0.0121511435136199, + 0.014575911685824394, + 0.01673252508044243, + 0.007149335462599993, + 0.007532481569796801, + 0.006523872260004282, + 0.008287233300507069, + 0.0062835123389959335, + 0.0076357354409992695, + 0.007606151979416609, + 0.013658128678798676, + 0.012321516871452332, + 0.009097309783101082, + 0.011292670853435993, + 0.011845104396343231, + 0.006783114746212959, + 0.010799513198435307, + 0.010452577844262123, + 0.020836178213357925, + 0.0022444401402026415, + 0.0032751986291259527, + 0.0036645540967583656, + 0.003218037076294422, + 0.002357310615479946, + 0.003112439764663577, + 0.003308050101622939, + 0.004780480172485113, + 0.0075246780179440975, + 0.006432184483855963, + 0.0067274561151862144, + 0.0055337692610919476, + 0.005061188712716103, + 0.005991884972900152, + 0.005471929907798767, + 0.009004608727991581 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 5 + ], + "selected_reward": -0.16035135090351105 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a tight close-up of a giant king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the crab curled up on a dark, mossy volcanic rock beneath a murky seabed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the crab have a vibrant orange-red carapace with sharp spines and jagged textures?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's large claws gently opening and closing in a slow, rhythmic pattern?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the image rendered with a nostalgic 1980s animation style featuring hand-painted cel-like quality, large flat shadows, and analog film grain?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016070479527115822 + ], + "opposite_logprob_tokens": [ + -4.141070365905762 + ], + "logprob_reward": -0.016070479527115822, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016070479527115822, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21320872008800507, + 0.11658871173858643, + 0.2525435984134674, + 0.03578321263194084, + 0.11339006572961807, + 0.037140682339668274, + 0.06299708038568497, + 0.05369517579674721, + 0.05849390849471092, + 0.05615880712866783 + ], + "credit_map": [ + 0.019531071186065674, + 0.01862824149429798, + 0.019397780299186707, + 0.03185819461941719, + 0.018736040219664574, + 0.028072385117411613, + 0.03327656537294388, + 0.04370848089456558, + 0.01037768367677927, + 0.00602643471211195, + 0.009424800053238869, + 0.008772074244916439, + 0.009665722958743572, + 0.01654091477394104, + 0.020047662779688835, + 0.035733431577682495, + 0.0482170470058918, + 0.01856309175491333, + 0.02428087405860424, + 0.01792833022773266, + 0.014459199272096157, + 0.03161248564720154, + 0.03597920760512352, + 0.061503373086452484, + 0.003378193126991391, + 0.0034415298141539097, + 0.004185829311609268, + 0.005346563644707203, + 0.0036158463917672634, + 0.004559650551527739, + 0.005134031642228365, + 0.006121571641415358, + 0.014526027254760265, + 0.014547029510140419, + 0.01405077800154686, + 0.012328376062214375, + 0.010044701397418976, + 0.013295631855726242, + 0.01616816595196724, + 0.01842937245965004, + 0.00317665608599782, + 0.004216506611555815, + 0.004915738478302956, + 0.005607922095805407, + 0.004055039491504431, + 0.004485582932829857, + 0.005677828565239906, + 0.005005411338061094, + 0.007002519443631172, + 0.006717843469232321, + 0.007969391532242298, + 0.008803602308034897, + 0.0051693120039999485, + 0.006284618284553289, + 0.008646796457469463, + 0.012403002008795738, + 0.005985789000988007, + 0.006546027958393097, + 0.007204832509160042, + 0.004587228409945965, + 0.004749466199427843, + 0.009990903548896313, + 0.006119928322732449, + 0.008511004038155079, + 0.0066812471486628056, + 0.006512834690511227, + 0.007321727462112904, + 0.005944307427853346, + 0.00501552177593112, + 0.0077227698639035225, + 0.007095334120094776, + 0.012200171127915382, + 0.005322285927832127, + 0.00632941909134388, + 0.006734798196703196, + 0.005881530698388815, + 0.003952066879719496, + 0.007884807884693146, + 0.0070580653846263885, + 0.012995840050280094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.016070479527115822 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_tensor([9170, 9170])_group_10_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_tensor([9170, 9170])_group_10_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..1fa993d7e007fb09b8cbab692e985e9ac2098653 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_tensor([9170, 9170])_group_10_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "slow, medium-distance zoom shot, captured with high-definition digital footage, focuses on a towering female martial artist, standing still in the center of an ancient courtyard. Her long, dark hair is pinned high with an exquisite wooden hairpin, her expression stern and focused. She wears a sharply tailored, rich crimson silk robe, delicately embroidered with golden cloud patterns. The robe gleams with a deep luster under the soft light of the late afternoon sun filtering through the clouds. The bluestone floor beneath her feet is strewn with a few fallen leaves, while the backdrop is a rustic wooden corridor and a verdant bamboo forest. She clutches a massive, closed folding fan with black ribs, her legs slightly bent, her center of gravity lowered, poised for action. Suddenly, with a sharp flick of her wrist, and with a crisp click, the enormous fan spun swiftly and smoothly in her hand, forming a perfect black semicircle. At the moment the fan fully unfolded, countless surreal golden light particles suddenly burst forth from the cracks in the ribs and the edges of the fan. As if imbued with life, they traced elegant, intricate spirals through the air, trailing wisps of light that instantly illuminated her resolute face and the flowing sheen of her silk dress. The entire dynamic process was meticulously captured, with the trajectory of each light particle, the texture of the silk, and even the dust particles hanging in the air clearly visible. This created a visual atmosphere imbued with mystery and awe, seamlessly blending the power of classical martial arts with a supernatural aesthetic.", + "seed": 33426, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33426_dp_20_rank_40_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04760441221296787, + "vlm_logprob": -0.04760441221296787, + "vlm_yes_no_margin": 3.15, + "vlm_token_credit": -0.04760441221296787, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07990903407335281 + ], + "opposite_logprob_tokens": [ + -2.829909086227417 + ], + "logprob_reward": -0.07990903407335281, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.07990903407335281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19550825655460358, + 0.04320520535111427, + 0.06218482553958893, + 0.05924553424119949, + 0.1868160367012024, + 0.3196125626564026, + 0.02218978852033615, + 0.023245764896273613, + 0.029997726902365685, + 0.057994283735752106 + ], + "credit_map": [ + 0.01016691979020834, + 0.020921172574162483, + 0.061450064182281494, + 0.010351080447435379, + 0.013465825468301773, + 0.015470833517611027, + 0.011821819469332695, + 0.051860511302948, + 0.0034484986681491137, + 0.004246931057423353, + 0.010471869260072708, + 0.0027865369338542223, + 0.003383214818313718, + 0.004071565344929695, + 0.0033881529234349728, + 0.01140842866152525, + 0.00432160310447216, + 0.0036490957718342543, + 0.011728733777999878, + 0.004505583085119724, + 0.0049606370739638805, + 0.010420436970889568, + 0.0073214503936469555, + 0.015277279540896416, + 0.003944922238588333, + 0.0027987079229205847, + 0.020592592656612396, + 0.004507274366915226, + 0.003852505935356021, + 0.003484395332634449, + 0.003998297266662121, + 0.016066832467913628, + 0.01031626109033823, + 0.009381590411067009, + 0.06031002104282379, + 0.01983439177274704, + 0.009414451196789742, + 0.016450822353363037, + 0.013147728517651558, + 0.047960761934518814, + 0.02429082803428173, + 0.02617638185620308, + 0.025031814351677895, + 0.09725021570920944, + 0.025637125596404076, + 0.050120625644922256, + 0.035066887736320496, + 0.036038681864738464, + 0.002210178878158331, + 0.0033410140313208103, + 0.0035293500404804945, + 0.0023608440533280373, + 0.002922449726611376, + 0.003077663714066148, + 0.002561095869168639, + 0.0021871894132345915, + 0.002517819870263338, + 0.003594063688069582, + 0.0028275770600885153, + 0.0030993276741355658, + 0.002872056793421507, + 0.002575547667220235, + 0.0030798146035522223, + 0.0026795538142323494, + 0.003395466133952141, + 0.0035269800573587418, + 0.0037164667155593634, + 0.004476130474358797, + 0.003295815084129572, + 0.0050619300454854965, + 0.003140572924166918, + 0.003384361742064357, + 0.005415061954408884, + 0.008037327788770199, + 0.008328785188496113, + 0.008182685822248459, + 0.00721953809261322, + 0.010212147608399391, + 0.0063588572666049, + 0.004239877685904503 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.07990903407335281 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043115582317113876 + ], + "opposite_logprob_tokens": [ + -3.1681156158447266 + ], + "logprob_reward": -0.043115582317113876, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043115582317113876, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04826834425330162, + 0.019049786031246185, + 0.07845576107501984, + 0.1310899704694748, + 0.46808305382728577, + 0.1583596169948578, + 0.008461257442831993, + 0.016339929774403572, + 0.03374033421278, + 0.03815191239118576 + ], + "credit_map": [ + 0.003564426675438881, + 0.0054405201226472855, + 0.014292469248175621, + 0.0031620196532458067, + 0.002611038973554969, + 0.004192580468952656, + 0.003706004237756133, + 0.01129928883165121, + 0.00161044392734766, + 0.0019076768076047301, + 0.0034373230300843716, + 0.0021713755559176207, + 0.002473474247381091, + 0.002280666260048747, + 0.0026094233617186546, + 0.002559404820203781, + 0.005137537606060505, + 0.0031384120229631662, + 0.012498860247433186, + 0.0025809414219111204, + 0.002955083269625902, + 0.018857013434171677, + 0.005040308460593224, + 0.028247609734535217, + 0.007619854062795639, + 0.011782287620007992, + 0.01126614585518837, + 0.019319966435432434, + 0.014683888293802738, + 0.025768926367163658, + 0.021665170788764954, + 0.018983734771609306, + 0.025706183165311813, + 0.01207675039768219, + 0.1911756843328476, + 0.019111450761556625, + 0.010677349753677845, + 0.025033462792634964, + 0.01569267176091671, + 0.16860952973365784, + 0.012321054935455322, + 0.012440847232937813, + 0.011659086681902409, + 0.05045203119516373, + 0.009921623393893242, + 0.026287386193871498, + 0.017869798466563225, + 0.017407789826393127, + 0.0007519540376961231, + 0.0013966867700219154, + 0.000959560158662498, + 0.0009999623289331794, + 0.00119023269508034, + 0.0012991591356694698, + 0.0010154745541512966, + 0.0008482288685627282, + 0.0013042972423136234, + 0.0024422253482043743, + 0.0016896837623789907, + 0.0021364318672567606, + 0.0022871829569339752, + 0.002341974526643753, + 0.0023880868684500456, + 0.0017500472022220492, + 0.0029321133624762297, + 0.004907000809907913, + 0.007666533347219229, + 0.005241713020950556, + 0.003072173800319433, + 0.002416739473119378, + 0.004225227516144514, + 0.0032788373064249754, + 0.00392181659117341, + 0.0041668955236673355, + 0.005550468806177378, + 0.005827456247061491, + 0.0030142476316541433, + 0.009117372334003448, + 0.0029926979914307594, + 0.003560960991308093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 3 + ], + "selected_reward": -0.043115582317113876 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0232884231954813 + ], + "opposite_logprob_tokens": [ + -3.7732884883880615 + ], + "logprob_reward": -0.0232884231954813, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.0232884231954813, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.042104072868824005, + 0.013485344126820564, + 0.051400650292634964, + 0.01927361823618412, + 0.6118022799491882, + 0.020793084055185318, + 0.03774765878915787, + 0.012050165794789791, + 0.06606660783290863, + 0.1252765655517578 + ], + "credit_map": [ + 0.002001050394028425, + 0.004865038208663464, + 0.011513376608490944, + 0.0030257459729909897, + 0.0025743625592440367, + 0.004643664695322514, + 0.0024342986289411783, + 0.011046536266803741, + 0.0010187571169808507, + 0.0012309368466958404, + 0.0035600701812654734, + 0.001138704246841371, + 0.0010217841481789947, + 0.0010980168590322137, + 0.0012086511123925447, + 0.0032084232661873102, + 0.0019827212672680616, + 0.002669353736564517, + 0.007202500477433205, + 0.002559704938903451, + 0.003384897019714117, + 0.012762483209371567, + 0.004417974501848221, + 0.016421020030975342, + 0.0019179792143404484, + 0.0027126374188810587, + 0.002362318802624941, + 0.003139577805995941, + 0.002286341041326523, + 0.0018653639126569033, + 0.002607222879305482, + 0.002382177859544754, + 0.03368840366601944, + 0.01462083775550127, + 0.2534159719944, + 0.024218980222940445, + 0.012954019010066986, + 0.03656080365180969, + 0.02076367661356926, + 0.2155795693397522, + 0.0017315492732450366, + 0.0027434814255684614, + 0.0022440124303102493, + 0.0027029679622501135, + 0.003665069816634059, + 0.002827161457389593, + 0.0024215937592089176, + 0.00245724618434906, + 0.0029857615008950233, + 0.007931924425065517, + 0.002613689750432968, + 0.005330702289938927, + 0.005122086964547634, + 0.0048199123702943325, + 0.005222892854362726, + 0.003720684675499797, + 0.0012229534331709146, + 0.0017964221769943833, + 0.000996345654129982, + 0.0017028903821483254, + 0.001892479951493442, + 0.0010988416615873575, + 0.0017256009159609675, + 0.0016146316193044186, + 0.005506270565092564, + 0.010771648027002811, + 0.009999150410294533, + 0.008772427216172218, + 0.007465236820280552, + 0.006584466900676489, + 0.011053982190787792, + 0.005913425236940384, + 0.012445234693586826, + 0.009521187283098698, + 0.010523286648094654, + 0.017649104818701744, + 0.011786999180912971, + 0.03683994710445404, + 0.011642331257462502, + 0.014868473634123802 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 9, + 8 + ], + "selected_reward": -0.0232884231954813 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04865886643528938 + ], + "opposite_logprob_tokens": [ + -3.048658847808838 + ], + "logprob_reward": -0.04865886643528938, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04865886643528938, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05980531871318817, + 0.01898142695426941, + 0.038222409784793854, + 0.029843201860785484, + 0.38862845301628113, + 0.2898622751235962, + 0.018701791763305664, + 0.011745927855372429, + 0.04288564622402191, + 0.10132354497909546 + ], + "credit_map": [ + 0.003856975818052888, + 0.0066414461471140385, + 0.015750084072351456, + 0.0045420946553349495, + 0.00377572700381279, + 0.005621657706797123, + 0.004287715535610914, + 0.015329619869589806, + 0.001937186811119318, + 0.002128368942067027, + 0.0032512068282812834, + 0.0021956933196634054, + 0.0027589520905166864, + 0.002365135820582509, + 0.001996003556996584, + 0.0023488812148571014, + 0.003226085100322962, + 0.0025551733560860157, + 0.004127084277570248, + 0.0022209512535482645, + 0.004196646623313427, + 0.011629467830061913, + 0.003520786063745618, + 0.006746216211467981, + 0.002023223089054227, + 0.0017628977075219154, + 0.005682317074388266, + 0.0029594702646136284, + 0.003413149155676365, + 0.0031319702975451946, + 0.004048703238368034, + 0.006821470335125923, + 0.021626904606819153, + 0.009861982427537441, + 0.15978962182998657, + 0.015566837042570114, + 0.008924615569412708, + 0.021341975778341293, + 0.013542482629418373, + 0.13797403872013092, + 0.019108064472675323, + 0.02559688128530979, + 0.024418748915195465, + 0.08621031045913696, + 0.02826189063489437, + 0.026771750301122665, + 0.0319761261343956, + 0.0475185252726078, + 0.0014925843570381403, + 0.0031972918659448624, + 0.0020975624211132526, + 0.0023568752221763134, + 0.0026028279680758715, + 0.0017437707865610719, + 0.0032890411093831062, + 0.0019218382658436894, + 0.001227142522111535, + 0.0016776318661868572, + 0.0009676773333922029, + 0.0016540667274966836, + 0.0016456044977530837, + 0.0019270633347332478, + 0.0013833775883540511, + 0.001263364334590733, + 0.0033295105677098036, + 0.00713500939309597, + 0.0081633310765028, + 0.0064365556463599205, + 0.0046470933593809605, + 0.004123809281736612, + 0.005499440245330334, + 0.003550897818058729, + 0.01066685002297163, + 0.009427014738321304, + 0.01002393290400505, + 0.014251664280891418, + 0.009140185080468655, + 0.027442993596196175, + 0.009099875576794147, + 0.011271034367382526 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 9 + ], + "selected_reward": -0.04865886643528938 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04305015504360199 + ], + "opposite_logprob_tokens": [ + -3.1680500507354736 + ], + "logprob_reward": -0.04305015504360199, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04305015504360199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0932890847325325, + 0.040771204978227615, + 0.018417589366436005, + 0.014825306832790375, + 0.2757665812969208, + 0.47390395402908325, + 0.027101507410407066, + 0.0075480020605027676, + 0.006330272648483515, + 0.04204653948545456 + ], + "credit_map": [ + 0.0058310856111347675, + 0.013747827149927616, + 0.02552502416074276, + 0.005433899816125631, + 0.0058523621410131454, + 0.01360477413982153, + 0.005073197185993195, + 0.018220912665128708, + 0.003590413834899664, + 0.005325164645910263, + 0.010250779800117016, + 0.0024668981786817312, + 0.002238108543679118, + 0.0059071192517876625, + 0.0021108388900756836, + 0.008881880901753902, + 0.001305782818235457, + 0.001148825860582292, + 0.002816346473991871, + 0.001484101521782577, + 0.0017093587666749954, + 0.0030594717245548964, + 0.001992136472836137, + 0.004901566077023745, + 0.0009294276824221015, + 0.0014489772729575634, + 0.0014499220997095108, + 0.0027890456840395927, + 0.0023866526316851377, + 0.0015655336901545525, + 0.0031436029821634293, + 0.0011121450224891305, + 0.015330477617681026, + 0.00671120872721076, + 0.10881480574607849, + 0.010679429396986961, + 0.005770811811089516, + 0.014268764294683933, + 0.009709091857075691, + 0.10448195785284042, + 0.0352279394865036, + 0.037695687264204025, + 0.039116423577070236, + 0.14690878987312317, + 0.03340575098991394, + 0.06576687842607498, + 0.05277432128787041, + 0.06300818920135498, + 0.0037036153953522444, + 0.002117888303473592, + 0.006280789617449045, + 0.003038531867787242, + 0.00163855473510921, + 0.004543890245258808, + 0.0020488151349127293, + 0.003729424439370632, + 0.0007335393456742167, + 0.0009107575169764459, + 0.0008044674177654088, + 0.0012792246416211128, + 0.0009163696086034179, + 0.0012245040852576494, + 0.0007794810808263719, + 0.0008996583055704832, + 0.0005765909445472062, + 0.0009417994879186153, + 0.0007426925585605204, + 0.0009451406658627093, + 0.0010595604544505477, + 0.0005497389356605709, + 0.0008650231757201254, + 0.000649726134724915, + 0.004655677359551191, + 0.0034917162265628576, + 0.0039042835123836994, + 0.005841120611876249, + 0.0037404641043394804, + 0.01182232890278101, + 0.00370583264157176, + 0.0048851147294044495 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.04305015504360199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-distance zoom shot, captured with high-definition digital footage, focuses on a towering female martial artist, standing still in the center of an ancient courtyard. Her long, dark hair is pinned high with an exquisite wooden hairpin, her expression stern and focused. She wears a sharply tailored, rich crimson silk robe, delicately embroidered with golden cloud patterns. The robe gleams with a deep luster under the soft light of the late afternoon sun filtering through the clouds. The bluestone floor beneath her feet is strewn with a few fallen leaves, while the backdrop is a rustic wooden corridor and a verdant bamboo forest. She clutches a massive, closed folding fan with black ribs, her legs slightly bent, her center of gravity lowered, poised for action. Suddenly, with a sharp flick of her wrist, and with a crisp click, the enormous fan spun swiftly and smoothly in her hand, forming a perfect black semicircle. At the moment the fan fully unfolded, countless surreal golden light particles suddenly burst forth from the cracks in the ribs and the edges of the fan. As if imbued with life, they traced elegant, intricate spirals through the air, trailing wisps of light that instantly illuminated her resolute face and the flowing sheen of her silk dress. The entire dynamic process was meticulously captured, with the trajectory of each light particle, the texture of the silk, and even the dust particles hanging in the air clearly visible. This created a visual atmosphere imbued with mystery and awe, seamlessly blending the power of classical martial arts with a supernatural aesthetic.", + "seed": 33427, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33427_dp_20_rank_40_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03886407986283302, + "vlm_logprob": -0.03886407986283302, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.03886407986283302, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05669010803103447 + ], + "opposite_logprob_tokens": [ + -3.806690216064453 + ], + "logprob_reward": -0.05669010803103447, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.05669010803103447, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.35019803047180176, + 0.06627663224935532, + 0.02287973463535309, + 0.05071315914392471, + 0.1382463574409485, + 0.07146994769573212, + 0.04205316677689552, + 0.1419348120689392, + 0.05063530430197716, + 0.06559287011623383 + ], + "credit_map": [ + 0.03875282406806946, + 0.028765279799699783, + 0.05034756660461426, + 0.06221969053149223, + 0.023979878053069115, + 0.06312147527933121, + 0.025095690041780472, + 0.05791560560464859, + 0.01287857349961996, + 0.007359719835221767, + 0.009052295237779617, + 0.011854683049023151, + 0.006606761831790209, + 0.004981505684554577, + 0.00773193221539259, + 0.005811153911054134, + 0.00352288456633687, + 0.002764122560620308, + 0.0033823270350694656, + 0.003291026223450899, + 0.0023775778245180845, + 0.0018509551882743835, + 0.0033949166536331177, + 0.002295922255143523, + 0.008184986189007759, + 0.004766342230141163, + 0.006936764344573021, + 0.004305270966142416, + 0.006937749218195677, + 0.00957742054015398, + 0.0039800237864255905, + 0.006024593953043222, + 0.04883353039622307, + 0.007935442961752415, + 0.04678390547633171, + 0.006889041513204575, + 0.00908601749688387, + 0.005997934378683567, + 0.0066939485259354115, + 0.006026521325111389, + 0.008199989795684814, + 0.007609666790813208, + 0.01029308419674635, + 0.010865870863199234, + 0.011633619666099548, + 0.007235249504446983, + 0.01081094704568386, + 0.004821509588509798, + 0.003806342137977481, + 0.0031033954583108425, + 0.006397534161806107, + 0.0050135101191699505, + 0.007532006129622459, + 0.0037490413524210453, + 0.006336231250315905, + 0.0061151026748120785, + 0.011486786417663097, + 0.013442736119031906, + 0.02223075181245804, + 0.012733290903270245, + 0.017562996596097946, + 0.033424586057662964, + 0.017697853967547417, + 0.013355796225368977, + 0.005407193209975958, + 0.009028944186866283, + 0.00695490138605237, + 0.00884495209902525, + 0.004163890145719051, + 0.003012876259163022, + 0.008620147593319416, + 0.004602393601089716, + 0.005428377538919449, + 0.0047294399701058865, + 0.02063828334212303, + 0.004771167412400246, + 0.006644822657108307, + 0.005773042794317007, + 0.00956504326313734, + 0.008042683824896812 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 4 + ], + "selected_reward": -0.05669010803103447 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048701558262109756 + ], + "opposite_logprob_tokens": [ + -3.048701524734497 + ], + "logprob_reward": -0.048701558262109756, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048701558262109756, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1224803626537323, + 0.09134294837713242, + 0.05245568975806236, + 0.05199362710118294, + 0.17609287798404694, + 0.03599456697702408, + 0.1596464216709137, + 0.1303732544183731, + 0.10102665424346924, + 0.0785936713218689 + ], + "credit_map": [ + 0.015166581608355045, + 0.014064975082874298, + 0.01706085167825222, + 0.010782336816191673, + 0.010445838794112206, + 0.02681306190788746, + 0.010644925758242607, + 0.01750177890062332, + 0.015066196210682392, + 0.006285740993916988, + 0.01348553691059351, + 0.011025404557585716, + 0.008014057762920856, + 0.01355555560439825, + 0.008496302179992199, + 0.015414138324558735, + 0.006629754323512316, + 0.0053429813124239445, + 0.006265692878514528, + 0.0055298106744885445, + 0.007429345976561308, + 0.008351893164217472, + 0.005587549414485693, + 0.007318657357245684, + 0.007679833564907312, + 0.006522479932755232, + 0.007624817080795765, + 0.005982507020235062, + 0.004901369567960501, + 0.008999600075185299, + 0.0034768106415867805, + 0.0068062045611441135, + 0.05520062521100044, + 0.013233484700322151, + 0.059460338205099106, + 0.007725353818386793, + 0.008787362836301327, + 0.012977015227079391, + 0.006387548986822367, + 0.012321121990680695, + 0.005148858297616243, + 0.006187052465975285, + 0.003487284993752837, + 0.004249236546456814, + 0.004582502879202366, + 0.004834589548408985, + 0.0038678667042404413, + 0.0036371718160808086, + 0.01408097892999649, + 0.006988105829805136, + 0.032367754727602005, + 0.0208249744027853, + 0.01946934685111046, + 0.018369268625974655, + 0.02330605313181877, + 0.02423991449177265, + 0.010437178425490856, + 0.011578782461583614, + 0.018804341554641724, + 0.015343734994530678, + 0.01789921708405018, + 0.02489684335887432, + 0.013250580988824368, + 0.0181625597178936, + 0.010289129801094532, + 0.011605629697442055, + 0.01455089170485735, + 0.019178103655576706, + 0.007184287067502737, + 0.00800029281526804, + 0.01902094855904579, + 0.011197357438504696, + 0.005323502235114574, + 0.006035438738763332, + 0.018272994086146355, + 0.007391742430627346, + 0.008771827444434166, + 0.006291249301284552, + 0.017789684236049652, + 0.008717216551303864 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 7 + ], + "selected_reward": -0.048701558262109756 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016069423407316208 + ], + "opposite_logprob_tokens": [ + -4.141069412231445 + ], + "logprob_reward": -0.016069423407316208, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016069423407316208, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0829416960477829, + 0.038943272083997726, + 0.05935073271393776, + 0.06730199605226517, + 0.09003999084234238, + 0.06041726469993591, + 0.18353188037872314, + 0.3296312689781189, + 0.03228133171796799, + 0.05556059256196022 + ], + "credit_map": [ + 0.01178221870213747, + 0.010068860836327076, + 0.011712630279362202, + 0.012210915796458721, + 0.010862775146961212, + 0.008433797396719456, + 0.007684795185923576, + 0.01018570363521576, + 0.006808358244597912, + 0.003904393408447504, + 0.006015324033796787, + 0.006057396065443754, + 0.005168735980987549, + 0.002238502725958824, + 0.005726637784391642, + 0.0030239219777286053, + 0.009164558723568916, + 0.005522455554455519, + 0.00949085783213377, + 0.004571777768433094, + 0.005250012036412954, + 0.012338939122855663, + 0.004755381029099226, + 0.00825674831867218, + 0.010614186525344849, + 0.007258310914039612, + 0.014149256981909275, + 0.007187023293226957, + 0.0063692424446344376, + 0.0066302758641541, + 0.004553394392132759, + 0.010540306568145752, + 0.03049752302467823, + 0.006265859119594097, + 0.02965247817337513, + 0.004800592083483934, + 0.00347173516638577, + 0.006658537779003382, + 0.003481300314888358, + 0.005211971700191498, + 0.007811021991074085, + 0.005837573204189539, + 0.009581761434674263, + 0.0075927008874714375, + 0.011372052133083344, + 0.00488838879391551, + 0.008228540420532227, + 0.005105225834995508, + 0.014552793465554714, + 0.010956903919577599, + 0.030089035630226135, + 0.027972539886832237, + 0.02737731672823429, + 0.016075875610113144, + 0.027339855208992958, + 0.0291675366461277, + 0.02651512436568737, + 0.031054234132170677, + 0.04958844557404518, + 0.026442671194672585, + 0.04534580931067467, + 0.08054716140031815, + 0.03714686632156372, + 0.0329909510910511, + 0.0035525094717741013, + 0.004805549513548613, + 0.004237893503159285, + 0.004062388092279434, + 0.004896904341876507, + 0.003031230065971613, + 0.004040258005261421, + 0.003654597792774439, + 0.004275789950042963, + 0.0036348362918943167, + 0.016151057556271553, + 0.006359186489135027, + 0.005515593104064465, + 0.0028196689672768116, + 0.012380151078104973, + 0.004424309358000755 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 4 + ], + "selected_reward": -0.016069423407316208 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043070822954177856 + ], + "opposite_logprob_tokens": [ + -3.1680707931518555 + ], + "logprob_reward": -0.043070822954177856, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043070822954177856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14125925302505493, + 0.0496632382273674, + 0.027056854218244553, + 0.0676150992512703, + 0.3771127164363861, + 0.03840670734643936, + 0.1524137556552887, + 0.05901015177369118, + 0.04153748229146004, + 0.04592473804950714 + ], + "credit_map": [ + 0.015922926366329193, + 0.009323948994278908, + 0.01999283954501152, + 0.028500551357865334, + 0.007521732244640589, + 0.023788146674633026, + 0.008563652634620667, + 0.027645455673336983, + 0.009181282483041286, + 0.004671966657042503, + 0.005859851371496916, + 0.008670653216540813, + 0.004876231774687767, + 0.004030476324260235, + 0.006010432727634907, + 0.0063623422756791115, + 0.003644443117082119, + 0.0028253912460058928, + 0.0032149539329111576, + 0.0031018927693367004, + 0.004392667207866907, + 0.0030547636561095715, + 0.0030127547215670347, + 0.0038099894300103188, + 0.011059632524847984, + 0.006197724957019091, + 0.01419305894523859, + 0.0049270568415522575, + 0.00566790672019124, + 0.013059311546385288, + 0.0034984152298420668, + 0.009011993184685707, + 0.1399334967136383, + 0.015326794236898422, + 0.140833780169487, + 0.014486216939985752, + 0.010434640571475029, + 0.024063825607299805, + 0.012681268155574799, + 0.019352711737155914, + 0.004904335830360651, + 0.003747784299775958, + 0.004784789867699146, + 0.004244380164891481, + 0.008845426142215729, + 0.0035181620623916388, + 0.005249022040516138, + 0.003112807171419263, + 0.013709956780076027, + 0.007179240696132183, + 0.03054584003984928, + 0.01940133608877659, + 0.01846766285598278, + 0.017574403434991837, + 0.02234349399805069, + 0.02319183386862278, + 0.004339137580245733, + 0.004970401059836149, + 0.00870268139988184, + 0.004998004529625177, + 0.006762880831956863, + 0.015514718368649483, + 0.007226410321891308, + 0.006495918147265911, + 0.0041428785771131516, + 0.007058874238282442, + 0.0062774112448096275, + 0.007386458572000265, + 0.0031173843890428543, + 0.003364430507645011, + 0.005707297474145889, + 0.004482743795961142, + 0.004741961136460304, + 0.0048867277801036835, + 0.007993756793439388, + 0.0036076856777071953, + 0.007324642967432737, + 0.0034412189852446318, + 0.010583875700831413, + 0.0033448650501668453 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 0 + ], + "selected_reward": -0.043070822954177856 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029788486659526825 + ], + "opposite_logprob_tokens": [ + -3.5297884941101074 + ], + "logprob_reward": -0.029788486659526825, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029788486659526825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3897855281829834, + 0.07780126482248306, + 0.0710984319448471, + 0.0737069696187973, + 0.109646737575531, + 0.04726278781890869, + 0.05028611421585083, + 0.07844959944486618, + 0.023077277466654778, + 0.07888521999120712 + ], + "credit_map": [ + 0.048422522842884064, + 0.03397601470351219, + 0.058642659336328506, + 0.09564092755317688, + 0.02379918284714222, + 0.05711861699819565, + 0.02004738710820675, + 0.05213823541998863, + 0.01535958331078291, + 0.012976004742085934, + 0.014612967148423195, + 0.00867320317775011, + 0.004812463652342558, + 0.0069379983469843864, + 0.004394848365336657, + 0.010034195147454739, + 0.0171798225492239, + 0.008995421230793, + 0.015560058876872063, + 0.005121254827827215, + 0.005464971996843815, + 0.00629831850528717, + 0.0031893241684883833, + 0.00928926095366478, + 0.011244434863328934, + 0.007390128914266825, + 0.014601540751755238, + 0.007736130617558956, + 0.007098184898495674, + 0.007759805768728256, + 0.005568380001932383, + 0.012308362871408463, + 0.04321600869297981, + 0.004166390281170607, + 0.042408257722854614, + 0.004314799793064594, + 0.003362893359735608, + 0.004283385816961527, + 0.0034208407159894705, + 0.004474166315048933, + 0.0068491799756884575, + 0.007305481005460024, + 0.005299443379044533, + 0.005040754098445177, + 0.008837166242301464, + 0.004678715951740742, + 0.004906456917524338, + 0.0043455855920910835, + 0.003858921118080616, + 0.003837931901216507, + 0.009299100376665592, + 0.006546974182128906, + 0.0060976906679570675, + 0.005043332464993, + 0.008114958181977272, + 0.00748720346018672, + 0.006014405284076929, + 0.006954867392778397, + 0.011420954950153828, + 0.007221390958875418, + 0.011630117893218994, + 0.018062828108668327, + 0.009269230999052525, + 0.007875806652009487, + 0.0021026257891207933, + 0.004406095948070288, + 0.0032898199278861284, + 0.003980764653533697, + 0.002195253036916256, + 0.0013362759491428733, + 0.003594988724216819, + 0.002171454718336463, + 0.00756914634257555, + 0.006278717424720526, + 0.014477991499006748, + 0.0067191836424171925, + 0.014768850989639759, + 0.005385393742471933, + 0.01625671796500683, + 0.007429216522723436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 9 + ], + "selected_reward": -0.029788486659526825 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-distance zoom shot, captured with high-definition digital footage, focuses on a towering female martial artist, standing still in the center of an ancient courtyard. Her long, dark hair is pinned high with an exquisite wooden hairpin, her expression stern and focused. She wears a sharply tailored, rich crimson silk robe, delicately embroidered with golden cloud patterns. The robe gleams with a deep luster under the soft light of the late afternoon sun filtering through the clouds. The bluestone floor beneath her feet is strewn with a few fallen leaves, while the backdrop is a rustic wooden corridor and a verdant bamboo forest. She clutches a massive, closed folding fan with black ribs, her legs slightly bent, her center of gravity lowered, poised for action. Suddenly, with a sharp flick of her wrist, and with a crisp click, the enormous fan spun swiftly and smoothly in her hand, forming a perfect black semicircle. At the moment the fan fully unfolded, countless surreal golden light particles suddenly burst forth from the cracks in the ribs and the edges of the fan. As if imbued with life, they traced elegant, intricate spirals through the air, trailing wisps of light that instantly illuminated her resolute face and the flowing sheen of her silk dress. The entire dynamic process was meticulously captured, with the trajectory of each light particle, the texture of the silk, and even the dust particles hanging in the air clearly visible. This created a visual atmosphere imbued with mystery and awe, seamlessly blending the power of classical martial arts with a supernatural aesthetic.", + "seed": 33428, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33428_dp_21_rank_42_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.04268988333642483, + "vlm_logprob": -0.04268988333642483, + "vlm_yes_no_margin": 3.35, + "vlm_token_credit": -0.04268988333642483, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04642288759350777 + ], + "opposite_logprob_tokens": [ + -3.7964229583740234 + ], + "logprob_reward": -0.04642288759350777, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.04642288759350777, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1246141716837883, + 0.05951759219169617, + 0.04811251536011696, + 0.11657506227493286, + 0.11446569859981537, + 0.1476038098335266, + 0.04875839501619339, + 0.11325781047344208, + 0.17214281857013702, + 0.054952170699834824 + ], + "credit_map": [ + 0.0197575893253088, + 0.0194852277636528, + 0.013980558142066002, + 0.009281056933104992, + 0.014029868878424168, + 0.018878862261772156, + 0.00950548890978098, + 0.019695516675710678, + 0.005709935445338488, + 0.006496739108115435, + 0.011224557645618916, + 0.004302246030420065, + 0.005648225545883179, + 0.009329348802566528, + 0.006391808856278658, + 0.010414731688797474, + 0.0042938911356031895, + 0.007906904444098473, + 0.006286923307925463, + 0.005149000324308872, + 0.0068263341672718525, + 0.007647295482456684, + 0.00488204974681139, + 0.005120112560689449, + 0.014311188831925392, + 0.016152793541550636, + 0.015898410230875015, + 0.012870876118540764, + 0.011154155246913433, + 0.018372392281889915, + 0.012255084700882435, + 0.015560157597064972, + 0.008993558585643768, + 0.01483276579529047, + 0.012767321430146694, + 0.012548847123980522, + 0.018285267055034637, + 0.016899216920137405, + 0.0163554847240448, + 0.013783241622149944, + 0.011230949312448502, + 0.019979344680905342, + 0.015728119760751724, + 0.020115891471505165, + 0.02027563378214836, + 0.02040761522948742, + 0.021928904578089714, + 0.017937354743480682, + 0.005570960696786642, + 0.004767109174281359, + 0.010376734659075737, + 0.005315158050507307, + 0.004420051351189613, + 0.0057327705435454845, + 0.006490194238722324, + 0.006085413042455912, + 0.013308757916092873, + 0.012689625844359398, + 0.02265026792883873, + 0.014027073048055172, + 0.013341004028916359, + 0.01210712268948555, + 0.013376343995332718, + 0.011757608503103256, + 0.015553265810012817, + 0.036841318011283875, + 0.027817096561193466, + 0.026785703375935555, + 0.017462968826293945, + 0.01167326606810093, + 0.025110987946391106, + 0.010898218490183353, + 0.004910110030323267, + 0.007305325474590063, + 0.0076432544738054276, + 0.007503442466259003, + 0.006113856099545956, + 0.0062566217966377735, + 0.007662565913051367, + 0.007556992117315531 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 0 + ], + "selected_reward": -0.04642288759350777 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03814641386270523 + ], + "opposite_logprob_tokens": [ + -3.288146495819092 + ], + "logprob_reward": -0.03814641386270523, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03814641386270523, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12706084549427032, + 0.05866922810673714, + 0.04475446045398712, + 0.04034276679158211, + 0.10206085443496704, + 0.16301681101322174, + 0.06316190212965012, + 0.18664953112602234, + 0.08679880201816559, + 0.12748482823371887 + ], + "credit_map": [ + 0.01087096706032753, + 0.010017337277531624, + 0.020208198577165604, + 0.013650971464812756, + 0.015032035298645496, + 0.017866963520646095, + 0.012179373763501644, + 0.02723499946296215, + 0.009075039997696877, + 0.006943147163838148, + 0.007913774810731411, + 0.004298584535717964, + 0.006562609225511551, + 0.009545741602778435, + 0.007010077591985464, + 0.007320243865251541, + 0.003191409632563591, + 0.00814664363861084, + 0.004679450765252113, + 0.004118992947041988, + 0.0067320652306079865, + 0.007594080176204443, + 0.005251770373433828, + 0.005040041171014309, + 0.0030125470366328955, + 0.0036859307438135147, + 0.007327764295041561, + 0.002794882981106639, + 0.0033326128032058477, + 0.007227117195725441, + 0.003855964168906212, + 0.009105944074690342, + 0.009116173721849918, + 0.010282609611749649, + 0.011810371652245522, + 0.007640061434358358, + 0.020930353552103043, + 0.017059488222002983, + 0.008578280918300152, + 0.016643503680825233, + 0.006998617202043533, + 0.010422532446682453, + 0.010297296568751335, + 0.009095278568565845, + 0.010033943690359592, + 0.04989922046661377, + 0.010433580726385117, + 0.055836331099271774, + 0.009056435897946358, + 0.007575459312647581, + 0.012740323320031166, + 0.007749954704195261, + 0.006447546649724245, + 0.00655900314450264, + 0.005657524336129427, + 0.0073756505735218525, + 0.022044964134693146, + 0.019767573103308678, + 0.03519141674041748, + 0.01899874024093151, + 0.023365095257759094, + 0.022278612479567528, + 0.018871920183300972, + 0.026131197810173035, + 0.008272587321698666, + 0.016437342390418053, + 0.01117191556841135, + 0.012352295219898224, + 0.009560794569551945, + 0.006099674850702286, + 0.0156794972717762, + 0.007224692963063717, + 0.009594306349754333, + 0.014244195073843002, + 0.018774643540382385, + 0.018842458724975586, + 0.015142416581511497, + 0.011222977191209793, + 0.02936054952442646, + 0.010303281247615814 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 5, + 9 + ], + "selected_reward": -0.03814641386270523 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018185196444392204 + ], + "opposite_logprob_tokens": [ + -4.018185138702393 + ], + "logprob_reward": -0.018185196444392204, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018185196444392204, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09412303566932678, + 0.09918209165334702, + 0.018260544165968895, + 0.05658328905701637, + 0.10095196217298508, + 0.15841786563396454, + 0.09399961680173874, + 0.12722086906433105, + 0.10440699011087418, + 0.14685367047786713 + ], + "credit_map": [ + 0.013542450964450836, + 0.009915854781866074, + 0.015056799165904522, + 0.007282882928848267, + 0.008517909795045853, + 0.014097388833761215, + 0.005897357128560543, + 0.01981239952147007, + 0.015600639395415783, + 0.012159223668277264, + 0.018489936366677284, + 0.005445186514407396, + 0.008024638518691063, + 0.016867509111762047, + 0.004544503055512905, + 0.018050456419587135, + 0.001675741863436997, + 0.002581964712589979, + 0.0026110829785466194, + 0.001796449301764369, + 0.0029838629998266697, + 0.0029213442467153072, + 0.001645909738726914, + 0.0020441878587007523, + 0.005009811371564865, + 0.008327220566570759, + 0.005670127458870411, + 0.007162869442254305, + 0.006881128530949354, + 0.00797444861382246, + 0.01009825337678194, + 0.005459430627524853, + 0.0073805637657642365, + 0.01287083700299263, + 0.014674920588731766, + 0.0070990510284900665, + 0.00896709319204092, + 0.01539047621190548, + 0.013679773546755314, + 0.020889244973659515, + 0.009057273156940937, + 0.00969701912254095, + 0.009999719448387623, + 0.008228552527725697, + 0.012638235464692116, + 0.04685840383172035, + 0.00839972123503685, + 0.05353894829750061, + 0.01064289826899767, + 0.010446758940815926, + 0.01822754740715027, + 0.010222170501947403, + 0.010590272024273872, + 0.011407067999243736, + 0.010571181774139404, + 0.011891718953847885, + 0.013438712805509567, + 0.012927244417369366, + 0.019801251590251923, + 0.02018573321402073, + 0.01485349889844656, + 0.009533307515084743, + 0.025173749774694443, + 0.011307373642921448, + 0.008659605868160725, + 0.024152841418981552, + 0.017106521874666214, + 0.01548339519649744, + 0.011518935672938824, + 0.007829866372048855, + 0.012691736221313477, + 0.006964083760976791, + 0.012463939376175404, + 0.017771760001778603, + 0.01737336441874504, + 0.020070314407348633, + 0.017709488049149513, + 0.014798934571444988, + 0.031381476670503616, + 0.015284396708011627 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 7 + ], + "selected_reward": -0.018185196444392204 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048657502979040146 + ], + "opposite_logprob_tokens": [ + -3.0486574172973633 + ], + "logprob_reward": -0.048657502979040146, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048657502979040146, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18528343737125397, + 0.061322171241045, + 0.03443099930882454, + 0.09029222279787064, + 0.10431765019893646, + 0.10595542192459106, + 0.046853452920913696, + 0.17015919089317322, + 0.0784904807806015, + 0.12289490550756454 + ], + "credit_map": [ + 0.03435812145471573, + 0.02730564773082733, + 0.03238976374268532, + 0.011427816934883595, + 0.012121654115617275, + 0.028855934739112854, + 0.007889176718890667, + 0.030935335904359818, + 0.00535071874037385, + 0.007617320399731398, + 0.010625098831951618, + 0.006618273910135031, + 0.007479371037334204, + 0.008518524467945099, + 0.008084036409854889, + 0.0070288279093801975, + 0.0024531700182706118, + 0.008443046361207962, + 0.004101657308638096, + 0.0023250968661159277, + 0.003260325873270631, + 0.006036946550011635, + 0.0030064857564866543, + 0.004804270341992378, + 0.007209817413240671, + 0.012218346819281578, + 0.011293305084109306, + 0.00964269507676363, + 0.011045578867197037, + 0.012943729758262634, + 0.014353695325553417, + 0.011585048399865627, + 0.009036236442625523, + 0.010791176930069923, + 0.014752347953617573, + 0.007701900787651539, + 0.01429043710231781, + 0.01552608422935009, + 0.014032799750566483, + 0.01818666234612465, + 0.009115934371948242, + 0.013002772815525532, + 0.010055901482701302, + 0.013875799253582954, + 0.01486125960946083, + 0.01993190310895443, + 0.010228514671325684, + 0.014883333817124367, + 0.006100550293922424, + 0.005144378170371056, + 0.007784523535519838, + 0.004855836275964975, + 0.004619597923010588, + 0.007098631002008915, + 0.006050516851246357, + 0.005199418868869543, + 0.025202658027410507, + 0.021247955039143562, + 0.025163723155856133, + 0.021045856177806854, + 0.020346639677882195, + 0.02279212884604931, + 0.016331173479557037, + 0.01802905835211277, + 0.008216477930545807, + 0.01776518113911152, + 0.010519195348024368, + 0.00912307109683752, + 0.010448841378092766, + 0.0065277377143502235, + 0.009085191413760185, + 0.006804784294217825, + 0.01085257064551115, + 0.013028964400291443, + 0.015615880489349365, + 0.01481372956186533, + 0.01138122845441103, + 0.013784734532237053, + 0.028211647644639015, + 0.015206153504550457 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 9 + ], + "selected_reward": -0.048657502979040146 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06203741580247879 + ], + "opposite_logprob_tokens": [ + -2.812037467956543 + ], + "logprob_reward": -0.06203741580247879, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06203741580247879, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.376643568277359, + 0.05771664157509804, + 0.03040478192269802, + 0.025587918236851692, + 0.041553251445293427, + 0.34266194701194763, + 0.03423398733139038, + 0.019552189856767654, + 0.03144797310233116, + 0.04019770398736 + ], + "credit_map": [ + 0.09594880789518356, + 0.050065599381923676, + 0.0674421563744545, + 0.017132220789790154, + 0.016085777431726456, + 0.05512170493602753, + 0.017854878678917885, + 0.056992437690496445, + 0.005758736282587051, + 0.00831298716366291, + 0.012666342779994011, + 0.005761079955846071, + 0.0047614434733986855, + 0.009532807394862175, + 0.005085534416139126, + 0.005837706383317709, + 0.004993045702576637, + 0.004110557027161121, + 0.004828615114092827, + 0.00231748353689909, + 0.003115575760602951, + 0.004956296179443598, + 0.0025068470276892185, + 0.0035763585474342108, + 0.002207172801718116, + 0.003434249432757497, + 0.0036706062965095043, + 0.002804370131343603, + 0.002873986726626754, + 0.003866378916427493, + 0.0026857240591198206, + 0.004045427311211824, + 0.003171071410179138, + 0.0029659157153218985, + 0.005868907552212477, + 0.0028133532032370567, + 0.003952180501073599, + 0.007664727978408337, + 0.003262704936787486, + 0.011854389682412148, + 0.020766913890838623, + 0.01961538754403591, + 0.023794949054718018, + 0.012005453929305077, + 0.019634295254945755, + 0.10572754591703415, + 0.017908282577991486, + 0.12320913374423981, + 0.004320803564041853, + 0.004397223237901926, + 0.005406395066529512, + 0.004167807288467884, + 0.0030399702955037355, + 0.0048264372162520885, + 0.0037547920364886522, + 0.004320560023188591, + 0.00243048882111907, + 0.00202976749278605, + 0.0025162191595882177, + 0.0021640444174408913, + 0.0018878108821809292, + 0.0027847520541399717, + 0.0025830341037362814, + 0.0031560722272843122, + 0.0034648552536964417, + 0.005042803939431906, + 0.004792742431163788, + 0.004775234963744879, + 0.002795285312458873, + 0.0034636163618415594, + 0.004836601205170155, + 0.0022768343333154917, + 0.003722584340721369, + 0.0047292993403971195, + 0.004592124372720718, + 0.004433175548911095, + 0.003367624245584011, + 0.0046379114501178265, + 0.00832896027714014, + 0.006386022549122572 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.06203741580247879 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "slow, medium-distance zoom shot, captured with high-definition digital footage, focuses on a towering female martial artist, standing still in the center of an ancient courtyard. Her long, dark hair is pinned high with an exquisite wooden hairpin, her expression stern and focused. She wears a sharply tailored, rich crimson silk robe, delicately embroidered with golden cloud patterns. The robe gleams with a deep luster under the soft light of the late afternoon sun filtering through the clouds. The bluestone floor beneath her feet is strewn with a few fallen leaves, while the backdrop is a rustic wooden corridor and a verdant bamboo forest. She clutches a massive, closed folding fan with black ribs, her legs slightly bent, her center of gravity lowered, poised for action. Suddenly, with a sharp flick of her wrist, and with a crisp click, the enormous fan spun swiftly and smoothly in her hand, forming a perfect black semicircle. At the moment the fan fully unfolded, countless surreal golden light particles suddenly burst forth from the cracks in the ribs and the edges of the fan. As if imbued with life, they traced elegant, intricate spirals through the air, trailing wisps of light that instantly illuminated her resolute face and the flowing sheen of her silk dress. The entire dynamic process was meticulously captured, with the trajectory of each light particle, the texture of the silk, and even the dust particles hanging in the air clearly visible. This created a visual atmosphere imbued with mystery and awe, seamlessly blending the power of classical martial arts with a supernatural aesthetic.", + "seed": 33429, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_10/index_9170_seed_33429_dp_21_rank_42_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 10, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04681944400072098, + "vlm_logprob": -0.04681944400072098, + "vlm_yes_no_margin": 3.25, + "vlm_token_credit": -0.04681944400072098, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03886229544878006 + ], + "opposite_logprob_tokens": [ + -3.6638622283935547 + ], + "logprob_reward": -0.03886229544878006, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.03886229544878006, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09617374837398529, + 0.048979662358760834, + 0.13673192262649536, + 0.2070590853691101, + 0.0510844811797142, + 0.11216878145933151, + 0.1858147829771042, + 0.06619186699390411, + 0.0411912240087986, + 0.05460444092750549 + ], + "credit_map": [ + 0.015824830159544945, + 0.00813499465584755, + 0.014652830548584461, + 0.01495790109038353, + 0.010556916706264019, + 0.009859456680715084, + 0.01091549452394247, + 0.011271320283412933, + 0.007196926977485418, + 0.005636339541524649, + 0.00919338408857584, + 0.0074479891918599606, + 0.003952542319893837, + 0.005427188705652952, + 0.004729028325527906, + 0.005396262276917696, + 0.017685292288661003, + 0.017171263694763184, + 0.02095911279320717, + 0.02547060139477253, + 0.014226308092474937, + 0.01246783696115017, + 0.0180729478597641, + 0.010678565129637718, + 0.02515377849340439, + 0.028198685497045517, + 0.020948151126503944, + 0.046887658536434174, + 0.03202115744352341, + 0.01036124862730503, + 0.030683066695928574, + 0.012805335223674774, + 0.00548245245590806, + 0.008398638106882572, + 0.00685147987678647, + 0.008038974367082119, + 0.005485031753778458, + 0.004097194876521826, + 0.008531465195119381, + 0.00419924221932888, + 0.015855753794312477, + 0.013994279317557812, + 0.01697646453976631, + 0.014469895511865616, + 0.010760647244751453, + 0.013537387363612652, + 0.01219587866216898, + 0.014378485269844532, + 0.027731120586395264, + 0.02904698997735977, + 0.024784700945019722, + 0.01643216796219349, + 0.014749656431376934, + 0.03358917310833931, + 0.01259474828839302, + 0.026886237785220146, + 0.010279960930347443, + 0.009216546081006527, + 0.008888917975127697, + 0.010004766285419464, + 0.006176984868943691, + 0.006884511094540358, + 0.0074158827774226665, + 0.007324297912418842, + 0.006557699292898178, + 0.008194363676011562, + 0.006081413011997938, + 0.004156420938670635, + 0.005648429971188307, + 0.002934382064267993, + 0.0046046096831560135, + 0.003013904672116041, + 0.00875797402113676, + 0.007924256846308708, + 0.009194718673825264, + 0.005077239591628313, + 0.005713240709155798, + 0.00609738240018487, + 0.005689536686986685, + 0.006150092463940382 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 2 + ], + "selected_reward": -0.03886229544878006 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06208480894565582 + ], + "opposite_logprob_tokens": [ + -2.812084913253784 + ], + "logprob_reward": -0.06208480894565582, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06208480894565582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09417428076267242, + 0.058764051645994186, + 0.07051385194063187, + 0.037657979875802994, + 0.10502192378044128, + 0.17119576036930084, + 0.0963083952665329, + 0.08524540066719055, + 0.22997911274433136, + 0.05113920569419861 + ], + "credit_map": [ + 0.01446601003408432, + 0.010946819558739662, + 0.01657015085220337, + 0.011476696468889713, + 0.012607577256858349, + 0.006375160999596119, + 0.01068283524364233, + 0.011049030348658562, + 0.0058274115435779095, + 0.006730749271810055, + 0.011119070462882519, + 0.008109753020107746, + 0.008051120676100254, + 0.006300622597336769, + 0.006622087210416794, + 0.006003240123391151, + 0.007650870364159346, + 0.009225212037563324, + 0.012610341422259808, + 0.0147059615701437, + 0.008324850350618362, + 0.005279846489429474, + 0.0072442530654370785, + 0.005472512450069189, + 0.004534157924354076, + 0.005314795766025782, + 0.004937592428177595, + 0.006381070241332054, + 0.005074989516288042, + 0.0028732989449054003, + 0.005019681993871927, + 0.003522387705743313, + 0.013783981092274189, + 0.019656848162412643, + 0.0139735396951437, + 0.013505280949175358, + 0.010126654990017414, + 0.008580854162573814, + 0.016224557533860207, + 0.00917020346969366, + 0.023641914129257202, + 0.02753785438835621, + 0.023872926831245422, + 0.02427065372467041, + 0.024022188037633896, + 0.012317635118961334, + 0.019971318542957306, + 0.015561277978122234, + 0.014209971763193607, + 0.014091688208281994, + 0.010119318030774593, + 0.013340274803340435, + 0.009050996974110603, + 0.01389747578650713, + 0.01223830971866846, + 0.009360363706946373, + 0.013101104646921158, + 0.011125907301902771, + 0.011473032645881176, + 0.015374756418168545, + 0.009407473728060722, + 0.007602686993777752, + 0.007253417745232582, + 0.009907019324600697, + 0.04040054976940155, + 0.029290759935975075, + 0.024814652279019356, + 0.058007992804050446, + 0.033916350454092026, + 0.010506357066333294, + 0.02164100669324398, + 0.011401467025279999, + 0.006706183310598135, + 0.004549100063741207, + 0.008384323678910732, + 0.005278029944747686, + 0.004567925818264484, + 0.007570149842649698, + 0.005176512990146875, + 0.008906979113817215 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 4 + ], + "selected_reward": -0.06208480894565582 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01606731116771698 + ], + "opposite_logprob_tokens": [ + -4.1410675048828125 + ], + "logprob_reward": -0.01606731116771698, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01606731116771698, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0608963742852211, + 0.0687049999833107, + 0.22352133691310883, + 0.22503133118152618, + 0.03120497427880764, + 0.23560774326324463, + 0.040120869874954224, + 0.05661242827773094, + 0.02403685636818409, + 0.03426302596926689 + ], + "credit_map": [ + 0.009929575026035309, + 0.005966914817690849, + 0.007247663103044033, + 0.009252117946743965, + 0.007508233655244112, + 0.005079897586256266, + 0.0091975973919034, + 0.006714378017932177, + 0.008244494907557964, + 0.00693495525047183, + 0.016491148620843887, + 0.0119636757299304, + 0.005951395258307457, + 0.004073827061802149, + 0.0073501537553966045, + 0.007695353124290705, + 0.03168489784002304, + 0.019935429096221924, + 0.047292087227106094, + 0.03970682621002197, + 0.022156303748488426, + 0.015961412340402603, + 0.028078176081180573, + 0.01870618388056755, + 0.027050018310546875, + 0.026426492258906364, + 0.024130994454026222, + 0.04981710761785507, + 0.035397931933403015, + 0.01527082733809948, + 0.029920868575572968, + 0.01701708510518074, + 0.004219697788357735, + 0.005937441252171993, + 0.005663661286234856, + 0.0029813339933753014, + 0.0031879269517958164, + 0.003277527866885066, + 0.003068256424739957, + 0.002869128482416272, + 0.03270190581679344, + 0.0467190146446228, + 0.028882259503006935, + 0.03411214053630829, + 0.033540185540914536, + 0.019915131852030754, + 0.021516192704439163, + 0.018220921978354454, + 0.006511189974844456, + 0.007452759426087141, + 0.005179372616112232, + 0.006923884619027376, + 0.0037508876994252205, + 0.003487372538074851, + 0.0036846366710960865, + 0.0031307667959481478, + 0.008643796667456627, + 0.006600564811378717, + 0.006657641381025314, + 0.008346380665898323, + 0.004851113073527813, + 0.0060237399302423, + 0.007434416096657515, + 0.00805477611720562, + 0.0036370689049363136, + 0.004364843480288982, + 0.0036133246030658484, + 0.002839308464899659, + 0.0038326855283230543, + 0.0020874962210655212, + 0.0022535796742886305, + 0.0014085511211305857, + 0.004217355977743864, + 0.004218460526317358, + 0.004607398062944412, + 0.003796548116952181, + 0.003710664575919509, + 0.00542150903493166, + 0.003803874598816037, + 0.004487217403948307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 3, + 2 + ], + "selected_reward": -0.01606731116771698 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07899043709039688 + ], + "opposite_logprob_tokens": [ + -2.5789904594421387 + ], + "logprob_reward": -0.07899043709039688, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07899043709039688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1202876940369606, + 0.05706193298101425, + 0.24736252427101135, + 0.08824116736650467, + 0.07553410530090332, + 0.08986030519008636, + 0.058657508343458176, + 0.06770196557044983, + 0.12207227945327759, + 0.07322050631046295 + ], + "credit_map": [ + 0.014862388372421265, + 0.011737864464521408, + 0.0169855784624815, + 0.01617823727428913, + 0.013566790148615837, + 0.015571671538054943, + 0.016057105734944344, + 0.015328062698245049, + 0.007243206258863211, + 0.004680303856730461, + 0.00870654359459877, + 0.01107241865247488, + 0.004946737550199032, + 0.00917429756373167, + 0.0051880995742976665, + 0.006050333380699158, + 0.02885596826672554, + 0.02996322698891163, + 0.04279148951172829, + 0.06293296068906784, + 0.02260918915271759, + 0.017079303041100502, + 0.024550514295697212, + 0.018579870462417603, + 0.010481506586074829, + 0.010633709840476513, + 0.013400290161371231, + 0.01832006312906742, + 0.010198368690907955, + 0.006601717788726091, + 0.0107980165630579, + 0.007807496003806591, + 0.009777373634278774, + 0.010384389199316502, + 0.013059391640126705, + 0.00901865679770708, + 0.006222866475582123, + 0.009669297374784946, + 0.009784175083041191, + 0.00761794950813055, + 0.01577826775610447, + 0.010516789741814137, + 0.01553284004330635, + 0.007431416772305965, + 0.007325947284698486, + 0.016092870384454727, + 0.005968117620795965, + 0.01121405977755785, + 0.008958245627582073, + 0.005167166702449322, + 0.009599006734788418, + 0.006951295770704746, + 0.005381647497415543, + 0.007731052581220865, + 0.005405141972005367, + 0.009463952854275703, + 0.009697713889181614, + 0.007775431498885155, + 0.008261729031801224, + 0.01326056383550167, + 0.006803231779485941, + 0.006767227314412594, + 0.009225892834365368, + 0.005910172592848539, + 0.02056390605866909, + 0.015923652797937393, + 0.017274610698223114, + 0.023258861154317856, + 0.013861075975000858, + 0.006626901216804981, + 0.018082624301314354, + 0.00648064911365509, + 0.010895628482103348, + 0.009327162988483906, + 0.009676768444478512, + 0.01121589820832014, + 0.009732471778988838, + 0.008696071803569794, + 0.008868508972227573, + 0.004807994235306978 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.07899043709039688 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot a slow, medium-distance zoom focusing on a female martial artist?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the character wearing a rich crimson silk robe embroidered with golden cloud patterns?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the character perform a sharp wrist flick to unfold a massive black folding fan into a semicircle?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Do countless golden light particles burst from the fan and trace elegant spirals through the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the background composed of a rustic wooden corridor and a verdant bamboo forest?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038092367351055145 + ], + "opposite_logprob_tokens": [ + -3.2880923748016357 + ], + "logprob_reward": -0.038092367351055145, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038092367351055145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33921846747398376, + 0.17794835567474365, + 0.035861920565366745, + 0.04834985360503197, + 0.08334587514400482, + 0.07003233581781387, + 0.10850009322166443, + 0.0566352941095829, + 0.014701896347105503, + 0.065405935049057 + ], + "credit_map": [ + 0.07660888880491257, + 0.04071810841560364, + 0.051539115607738495, + 0.03158273547887802, + 0.022000566124916077, + 0.04901682212948799, + 0.030080322176218033, + 0.037671901285648346, + 0.025580743327736855, + 0.013612634502351284, + 0.038118068128824234, + 0.01796000823378563, + 0.010828700847923756, + 0.025233983993530273, + 0.01238410733640194, + 0.034230101853609085, + 0.005388100165873766, + 0.005208692513406277, + 0.005100547801703215, + 0.005335329100489616, + 0.00471922941505909, + 0.0024260974023491144, + 0.004558472894132137, + 0.0031254517380148172, + 0.007562944665551186, + 0.005276121199131012, + 0.005134479608386755, + 0.005170758813619614, + 0.004955871496349573, + 0.005924122408032417, + 0.004446052480489016, + 0.009879502467811108, + 0.012427009642124176, + 0.008746874518692493, + 0.019894633442163467, + 0.007227438967674971, + 0.006931918673217297, + 0.007101079914718866, + 0.009595183655619621, + 0.011421727016568184, + 0.00896409060806036, + 0.010384712368249893, + 0.009440969675779343, + 0.008215725421905518, + 0.010309460572898388, + 0.008022583089768887, + 0.009025339968502522, + 0.005669447127729654, + 0.013377388939261436, + 0.012355360202491283, + 0.013143593445420265, + 0.008993401192128658, + 0.006711749359965324, + 0.02068459242582321, + 0.016513636335730553, + 0.016720369458198547, + 0.008831494487822056, + 0.006680977996438742, + 0.008146136999130249, + 0.006175345275551081, + 0.004311670083552599, + 0.007584826089441776, + 0.006338339764624834, + 0.00856650061905384, + 0.002240959322080016, + 0.00251679471693933, + 0.0019282664870843291, + 0.0023612508084625006, + 0.0014206088380888104, + 0.0013001278275623918, + 0.001519352663308382, + 0.0014145341701805592, + 0.01195666752755642, + 0.005209243390709162, + 0.01906859129667282, + 0.005614824593067169, + 0.004897722974419594, + 0.0054162247106432915, + 0.006872578524053097, + 0.006370078306645155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.038092367351055145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_tensor([2448, 2448])_group_11_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_tensor([2448, 2448])_group_11_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..4be05d6408532a33aa3c74675a90503af372357a --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_tensor([2448, 2448])_group_11_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic drone shot, captured from a lofty height, captures a breathtaking panoramic view of the majestic mountain ranges as the first rays of morning's golden sun bathe the mountain's rolling hills. The jagged, snow-capped peaks shimmer in the golden light, while the deep valleys remain shrouded in a tranquil, deep blue. A thin, morning mist veils the coniferous forests lining the mountainside. A massive, fully-fledged eagle, its dark brown feathers gleaming metallically in the sunlight, circles steadily. Suddenly, without hesitation, it folds its broad wings back, its body transforming into a streamlined, living arrow, swooping vertically with breathtaking acceleration towards the deep valley below. With equally swift speed, the camera tilts downward and zooms in, closely following the eagle's dive. Its piercing amber eyes are clearly visible, fixed on an unseen target below, and the wind ruffles the feathers on its neck. During this rapid fall, the eagle's tiny and swift figure forms a strong and shocking visual contrast with the eternally still and enormous mountain behind it. The whole picture is full of primitive, wild power and breathtaking tension, creating an epic sense of awe of nature.", + "seed": 10971, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_2448_seed_10971_dp_22_rank_44_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.17332958839833737, + "vlm_logprob": -0.17332958839833737, + "vlm_yes_no_margin": 2.325, + "vlm_token_credit": -0.17332958839833737, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0898013487458229 + ], + "opposite_logprob_tokens": [ + -2.46480131149292 + ], + "logprob_reward": -0.0898013487458229, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0898013487458229, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31098198890686035, + 0.026573646813631058, + 0.03485222905874252, + 0.04664541408419609, + 0.019697001203894615, + 0.12453006953001022, + 0.08113069087266922, + 0.12942282855510712, + 0.12757009267807007, + 0.0985960066318512 + ], + "credit_map": [ + 0.02871137671172619, + 0.0178807582706213, + 0.07451295107603073, + 0.03828330710530281, + 0.02627919800579548, + 0.08388137072324753, + 0.02050596848130226, + 0.020927084609866142, + 0.002847996773198247, + 0.0028059184551239014, + 0.003376713953912258, + 0.0022988838609308004, + 0.00680485088378191, + 0.003879405790939927, + 0.0020518372766673565, + 0.002508042147383094, + 0.00713004358112812, + 0.005605327896773815, + 0.0033156434074044228, + 0.0028876515571027994, + 0.006136962212622166, + 0.003951938822865486, + 0.0025508624967187643, + 0.0032738042064011097, + 0.004328279290348291, + 0.0056040468625724316, + 0.00502257514744997, + 0.004887872375547886, + 0.007703809067606926, + 0.007313156500458717, + 0.004943056497722864, + 0.006842621602118015, + 0.0017154382076114416, + 0.0012770190369337797, + 0.0032165907323360443, + 0.002779513830319047, + 0.0023596088867634535, + 0.0012903311289846897, + 0.0033250150736421347, + 0.003733484772965312, + 0.0083805276080966, + 0.015164580196142197, + 0.018944161012768745, + 0.030030060559511185, + 0.0167358610779047, + 0.01258457638323307, + 0.014162680134177208, + 0.008527631871402264, + 0.008875472471117973, + 0.009867638349533081, + 0.01574026234447956, + 0.010508015751838684, + 0.006252703722566366, + 0.010343283414840698, + 0.010747997090220451, + 0.00879532378166914, + 0.008548036217689514, + 0.009034241549670696, + 0.018548879772424698, + 0.023414907976984978, + 0.013152648694813251, + 0.011513921432197094, + 0.03365696594119072, + 0.011553228832781315, + 0.007911302149295807, + 0.009132903069257736, + 0.03750748187303543, + 0.012024858966469765, + 0.010720341466367245, + 0.006943674758076668, + 0.009471198543906212, + 0.03385835140943527, + 0.0073021529242396355, + 0.010435627773404121, + 0.01720997877418995, + 0.01539644505828619, + 0.011771159246563911, + 0.007262527011334896, + 0.01838402822613716, + 0.01083409134298563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 8 + ], + "selected_reward": -0.0898013487458229 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03371250629425049 + ], + "opposite_logprob_tokens": [ + -3.408712387084961 + ], + "logprob_reward": -0.03371250629425049, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03371250629425049, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36802932620048523, + 0.0767672136425972, + 0.06936290860176086, + 0.04496634006500244, + 0.06144721433520317, + 0.041727639734745026, + 0.058331213891506195, + 0.09885182976722717, + 0.12484198063611984, + 0.055674370378255844 + ], + "credit_map": [ + 0.04512830451130867, + 0.029091939330101013, + 0.05745207145810127, + 0.029594022780656815, + 0.038596514612436295, + 0.12681908905506134, + 0.01823916845023632, + 0.0231082271784544, + 0.0061820014379918575, + 0.007835649885237217, + 0.013523516245186329, + 0.010582841001451015, + 0.01141183078289032, + 0.010501324199140072, + 0.00768637889996171, + 0.009043673053383827, + 0.01756672002375126, + 0.014917107298970222, + 0.007640515919774771, + 0.005751828663051128, + 0.0067452481016516685, + 0.007019215263426304, + 0.004798168782144785, + 0.004924101755023003, + 0.004391324240714312, + 0.006746020168066025, + 0.006711529567837715, + 0.005043703597038984, + 0.008170504122972488, + 0.0037131577264517546, + 0.004717308096587658, + 0.005472792778164148, + 0.003671645186841488, + 0.010515687055885792, + 0.01049803476780653, + 0.009592344053089619, + 0.006405608728528023, + 0.003442048327997327, + 0.007614759728312492, + 0.009707080200314522, + 0.0033495300449430943, + 0.0065276348032057285, + 0.005054542329162359, + 0.007245248183608055, + 0.0042749675922095776, + 0.005426886025816202, + 0.004787612706422806, + 0.0050612203776836395, + 0.005117073189467192, + 0.003719541011378169, + 0.007557517848908901, + 0.008227933198213577, + 0.0038592468481510878, + 0.010598474182188511, + 0.007721960078924894, + 0.011529470793902874, + 0.008964087814092636, + 0.006810122635215521, + 0.012612159363925457, + 0.014440319500863552, + 0.007902846671640873, + 0.010511606931686401, + 0.026642128825187683, + 0.010968560352921486, + 0.009060047566890717, + 0.00924461055546999, + 0.032918497920036316, + 0.019213641062378883, + 0.007227621506899595, + 0.008888895623385906, + 0.008055862039327621, + 0.030232803896069527, + 0.005496140103787184, + 0.006233023479580879, + 0.009479419328272343, + 0.008162885904312134, + 0.007098374422639608, + 0.005236626602709293, + 0.00824409443885088, + 0.005723809357732534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.03371250629425049 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20156370103359222 + ], + "opposite_logprob_tokens": [ + -1.7015637159347534 + ], + "logprob_reward": -0.20156370103359222, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20156370103359222, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11368606239557266, + 0.09313096851110458, + 0.23209978640079498, + 0.10326527804136276, + 0.10725422948598862, + 0.03676685690879822, + 0.04856353625655174, + 0.08679676800966263, + 0.15684352815151215, + 0.02159295417368412 + ], + "credit_map": [ + 0.012737884186208248, + 0.008900054730474949, + 0.010684257373213768, + 0.008829046972095966, + 0.01587666943669319, + 0.03830544650554657, + 0.007399084512144327, + 0.010953621007502079, + 0.009503613226115704, + 0.013433707877993584, + 0.015612419694662094, + 0.013490094803273678, + 0.010127704590559006, + 0.010291104204952717, + 0.009022367186844349, + 0.011649955064058304, + 0.043632376939058304, + 0.01831216737627983, + 0.03129132091999054, + 0.02014201320707798, + 0.020475883036851883, + 0.02867380902171135, + 0.02529999054968357, + 0.04427221789956093, + 0.00994129292666912, + 0.012521994300186634, + 0.013606167398393154, + 0.01129272673279047, + 0.014774390496313572, + 0.019908949732780457, + 0.011386994272470474, + 0.009832763113081455, + 0.008668160997331142, + 0.011955990456044674, + 0.014177443459630013, + 0.012347555719316006, + 0.018681040033698082, + 0.008263613097369671, + 0.013998547568917274, + 0.019161872565746307, + 0.0025289994664490223, + 0.0035862326622009277, + 0.007235066033899784, + 0.008184604346752167, + 0.002620815532281995, + 0.0034193701576441526, + 0.00509215472266078, + 0.004099616315215826, + 0.006407992448657751, + 0.0032658104319125414, + 0.009074504487216473, + 0.009357349947094917, + 0.002990535693243146, + 0.005176825448870659, + 0.006004038266837597, + 0.006286481395363808, + 0.005098952911794186, + 0.005453885067254305, + 0.010846097953617573, + 0.012305321171879768, + 0.007907053455710411, + 0.008736097253859043, + 0.02555229887366295, + 0.010897070169448853, + 0.00817984901368618, + 0.006001262925565243, + 0.05444863438606262, + 0.01691603846848011, + 0.005477326922118664, + 0.006882105953991413, + 0.0075583984144032, + 0.051379911601543427, + 0.002427708590403199, + 0.002378619508817792, + 0.0038371060509234667, + 0.0032162326388061047, + 0.002162114717066288, + 0.0016053441213443875, + 0.0038101121317595243, + 0.002155714901164174 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 8, + 0 + ], + "selected_reward": -0.20156370103359222 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018187304958701134 + ], + "opposite_logprob_tokens": [ + -4.018187522888184 + ], + "logprob_reward": -0.018187304958701134, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018187304958701134, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.38725727796554565, + 0.03274066001176834, + 0.036744218319654465, + 0.04914483800530434, + 0.05347773805260658, + 0.09920712560415268, + 0.058656997978687286, + 0.04997499659657478, + 0.21230098605155945, + 0.02049521543085575 + ], + "credit_map": [ + 0.04572451487183571, + 0.0267291609197855, + 0.06446787714958191, + 0.03219161555171013, + 0.041143421083688736, + 0.135741725564003, + 0.01793297380208969, + 0.023325951769948006, + 0.0031748157925903797, + 0.003877596464008093, + 0.0057952734641730785, + 0.003976569976657629, + 0.0032492922618985176, + 0.004585010930895805, + 0.003655151231214404, + 0.0044269440695643425, + 0.004434499423950911, + 0.002874161349609494, + 0.006134129129350185, + 0.004752218723297119, + 0.005110430996865034, + 0.004577649757266045, + 0.004496969282627106, + 0.0043641505762934685, + 0.0035550303291529417, + 0.007189385127276182, + 0.007389197591692209, + 0.005740975961089134, + 0.006739259231835604, + 0.006002417765557766, + 0.005625796504318714, + 0.006902765482664108, + 0.00357356877066195, + 0.006643937435001135, + 0.009195983409881592, + 0.006832951679825783, + 0.004554859362542629, + 0.0031943500507622957, + 0.009288319386541843, + 0.01019375305622816, + 0.007795958314090967, + 0.00917149055749178, + 0.018631750717759132, + 0.020889664068818092, + 0.010460379533469677, + 0.01072689052671194, + 0.012343091890215874, + 0.009187883697450161, + 0.006530893966555595, + 0.005691432394087315, + 0.008594585582613945, + 0.007210675161331892, + 0.004482487682253122, + 0.010969189926981926, + 0.006113003473728895, + 0.009064718149602413, + 0.004128110129386187, + 0.0034827531781047583, + 0.006578647997230291, + 0.008158487267792225, + 0.004089708440005779, + 0.008480929769575596, + 0.00605651643127203, + 0.008999834768474102, + 0.013877017423510551, + 0.01196452509611845, + 0.05346490070223808, + 0.05016211047768593, + 0.010385374538600445, + 0.009099297225475311, + 0.016361262649297714, + 0.04698646068572998, + 0.002133462578058243, + 0.002375948941335082, + 0.00394135108217597, + 0.003476630663499236, + 0.0017183148302137852, + 0.001873471075668931, + 0.002966362051665783, + 0.002009672811254859 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 5 + ], + "selected_reward": -0.018187304958701134 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5233830809593201 + ], + "opposite_logprob_tokens": [ + -0.8983830809593201 + ], + "logprob_reward": -0.5233830809593201, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5233830809593201, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19173277914524078, + 0.03388199210166931, + 0.02131887525320053, + 0.02522427774965763, + 0.03823310136795044, + 0.08645565062761307, + 0.09595872461795807, + 0.09932015836238861, + 0.2127285748720169, + 0.19514590501785278 + ], + "credit_map": [ + 0.013968001119792461, + 0.00872230064123869, + 0.05071784183382988, + 0.028584333136677742, + 0.01508344616740942, + 0.0387241430580616, + 0.01821400597691536, + 0.017718708142638206, + 0.0038044953253120184, + 0.003747346345335245, + 0.004298587329685688, + 0.0036571077071130276, + 0.007816321216523647, + 0.004691251087933779, + 0.002298818901181221, + 0.003568061627447605, + 0.003560638288035989, + 0.0031864771153777838, + 0.002688758308067918, + 0.002398120705038309, + 0.0037377651315182447, + 0.002463716082274914, + 0.001833636430092156, + 0.0014497633092105389, + 0.0019280023407191038, + 0.0033932728692889214, + 0.003065073397010565, + 0.0027792712207883596, + 0.004871034529060125, + 0.003441399894654751, + 0.0026749144308269024, + 0.003071308834478259, + 0.002945494605228305, + 0.004558941349387169, + 0.007087216712534428, + 0.004898843355476856, + 0.003094538114964962, + 0.002527385950088501, + 0.006017919629812241, + 0.0071027642115950584, + 0.005545229651033878, + 0.008519867435097694, + 0.017501702532172203, + 0.025563599541783333, + 0.007507828529924154, + 0.0058906241320073605, + 0.009737903252243996, + 0.0061888969503343105, + 0.0071340096183121204, + 0.007149428594857454, + 0.015531898476183414, + 0.012204669415950775, + 0.005795541685074568, + 0.019368024542927742, + 0.013596165925264359, + 0.01517898216843605, + 0.005033483263105154, + 0.006942208856344223, + 0.01226661168038845, + 0.013009792193770409, + 0.009590690024197102, + 0.009842448867857456, + 0.028165576979517937, + 0.014469346031546593, + 0.012806028127670288, + 0.011915108188986778, + 0.06639457494020462, + 0.018658293411135674, + 0.016028573736548424, + 0.013017886318266392, + 0.010973216965794563, + 0.06293489038944244, + 0.0171075239777565, + 0.022076310589909554, + 0.030435124412178993, + 0.030176756903529167, + 0.02519041858613491, + 0.017564361914992332, + 0.02946464717388153, + 0.023130757734179497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.5233830809593201 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot, captured from a lofty height, captures a breathtaking panoramic view of the majestic mountain ranges as the first rays of morning's golden sun bathe the mountain's rolling hills. The jagged, snow-capped peaks shimmer in the golden light, while the deep valleys remain shrouded in a tranquil, deep blue. A thin, morning mist veils the coniferous forests lining the mountainside. A massive, fully-fledged eagle, its dark brown feathers gleaming metallically in the sunlight, circles steadily. Suddenly, without hesitation, it folds its broad wings back, its body transforming into a streamlined, living arrow, swooping vertically with breathtaking acceleration towards the deep valley below. With equally swift speed, the camera tilts downward and zooms in, closely following the eagle's dive. Its piercing amber eyes are clearly visible, fixed on an unseen target below, and the wind ruffles the feathers on its neck. During this rapid fall, the eagle's tiny and swift figure forms a strong and shocking visual contrast with the eternally still and enormous mountain behind it. The whole picture is full of primitive, wild power and breathtaking tension, creating an epic sense of awe of nature.", + "seed": 10972, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_2448_seed_10972_dp_22_rank_44_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.06702306251972914, + "vlm_logprob": -0.06702306251972914, + "vlm_yes_no_margin": 3.525, + "vlm_token_credit": -0.06702306251972914, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008262617513537407 + ], + "opposite_logprob_tokens": [ + -4.883262634277344 + ], + "logprob_reward": -0.008262617513537407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.008262617513537407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2906149625778198, + 0.07992280274629593, + 0.047104611992836, + 0.0718633159995079, + 0.05304801091551781, + 0.059638142585754395, + 0.13315626978874207, + 0.061212070286273956, + 0.05383468046784401, + 0.14960519969463348 + ], + "credit_map": [ + 0.031045105308294296, + 0.02787000685930252, + 0.03683377057313919, + 0.03691919893026352, + 0.02150748297572136, + 0.03478242829442024, + 0.040839895606040955, + 0.06081704422831535, + 0.007469616364687681, + 0.013018697500228882, + 0.013321532867848873, + 0.014837916009128094, + 0.0064258575439453125, + 0.0048796930350363255, + 0.015163079835474491, + 0.004806407727301121, + 0.007631607819348574, + 0.0036163788754493, + 0.014291826635599136, + 0.0050242929719388485, + 0.0024487439077347517, + 0.002507095457985997, + 0.008271931670606136, + 0.00331272860057652, + 0.005775789730250835, + 0.005834424402564764, + 0.009665170684456825, + 0.008406540378928185, + 0.01937609352171421, + 0.004193500615656376, + 0.00976988859474659, + 0.008841900154948235, + 0.007124056573957205, + 0.004267789889127016, + 0.005267797503620386, + 0.010225974954664707, + 0.006107032764703035, + 0.004251020960509777, + 0.007893210276961327, + 0.00791112519800663, + 0.005070284474641085, + 0.004753815475851297, + 0.007107197307050228, + 0.007757214363664389, + 0.005254807416349649, + 0.0068139792419970036, + 0.011123503558337688, + 0.011757337488234043, + 0.009458886459469795, + 0.009215258061885834, + 0.031282272189855576, + 0.013344887644052505, + 0.01375182531774044, + 0.009885110892355442, + 0.02304793894290924, + 0.02317008376121521, + 0.006475434638559818, + 0.011149282567203045, + 0.010917973704636097, + 0.003954011481255293, + 0.007074019405990839, + 0.006993296556174755, + 0.0056896205060184, + 0.008958425372838974, + 0.007442953530699015, + 0.009800796397030354, + 0.011692618019878864, + 0.004275094717741013, + 0.004431643057614565, + 0.00357085932046175, + 0.0062919557094573975, + 0.006328755524009466, + 0.025768853724002838, + 0.060809675604104996, + 0.0092097707092762, + 0.006708469707518816, + 0.005625193472951651, + 0.015548255294561386, + 0.00539463572204113, + 0.020540330559015274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 6 + ], + "selected_reward": -0.008262617513537407 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04307344928383827 + ], + "opposite_logprob_tokens": [ + -3.1680734157562256 + ], + "logprob_reward": -0.04307344928383827, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04307344928383827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.32168230414390564, + 0.12004050612449646, + 0.051506854593753815, + 0.07074636965990067, + 0.10869744420051575, + 0.08822997659444809, + 0.04415588825941086, + 0.04830456152558327, + 0.12484993040561676, + 0.021786106750369072 + ], + "credit_map": [ + 0.043809544295072556, + 0.03584195673465729, + 0.037990279495716095, + 0.04200614616274834, + 0.0292224008589983, + 0.0400085374712944, + 0.044070713222026825, + 0.04873272404074669, + 0.00934548955410719, + 0.013514596037566662, + 0.021392906084656715, + 0.03282030671834946, + 0.007292371708899736, + 0.009294167160987854, + 0.01633990928530693, + 0.01004075538367033, + 0.009586727246642113, + 0.004250434692949057, + 0.015645895153284073, + 0.003978402353823185, + 0.002458062255755067, + 0.003645772347226739, + 0.008837146684527397, + 0.003104408970102668, + 0.005404280032962561, + 0.004430470522493124, + 0.011492292396724224, + 0.005903747398406267, + 0.017844464629888535, + 0.0063627539202570915, + 0.010457630269229412, + 0.008850733749568462, + 0.013816483318805695, + 0.007869071327149868, + 0.011253871023654938, + 0.017148654907941818, + 0.0145171033218503, + 0.018825190141797066, + 0.012275452725589275, + 0.012991619296371937, + 0.011732707731425762, + 0.007548831403255463, + 0.011824722401797771, + 0.010782692581415176, + 0.009525849483907223, + 0.009861094877123833, + 0.013867120258510113, + 0.013086958788335323, + 0.006584446877241135, + 0.005959285423159599, + 0.006214790511876345, + 0.00515738083049655, + 0.0030265955720096827, + 0.004719776567071676, + 0.006191170774400234, + 0.006302442401647568, + 0.006661542225629091, + 0.011757207103073597, + 0.005876220762729645, + 0.0031117417383939028, + 0.0036524839233607054, + 0.0054153031669557095, + 0.0044312854297459126, + 0.007398777175694704, + 0.01257353089749813, + 0.020439328625798225, + 0.01804453507065773, + 0.023215726017951965, + 0.010822360403835773, + 0.010104941204190254, + 0.015031609684228897, + 0.014617900364100933, + 0.0038322540931403637, + 0.0035354080609977245, + 0.0018171007977798581, + 0.0009336713119409978, + 0.001410129596479237, + 0.002113261492922902, + 0.001171205542050302, + 0.006973076146095991 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.04307344928383827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02057494968175888 + ], + "opposite_logprob_tokens": [ + -3.8955750465393066 + ], + "logprob_reward": -0.02057494968175888, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02057494968175888, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08330269902944565, + 0.1694149672985077, + 0.1683541089296341, + 0.08625826239585876, + 0.04988129809498787, + 0.346350759267807, + 0.027642948552966118, + 0.01826992630958557, + 0.022625019773840904, + 0.027899980545043945 + ], + "credit_map": [ + 0.01260532159358263, + 0.00798586755990982, + 0.009012730792164803, + 0.01247312780469656, + 0.00990977417677641, + 0.010535611771047115, + 0.00840461440384388, + 0.012375645339488983, + 0.014573679305613041, + 0.011051127687096596, + 0.04062778502702713, + 0.011805412359535694, + 0.008066130802035332, + 0.027273178100585938, + 0.03338470682501793, + 0.02263294719159603, + 0.028705749660730362, + 0.007951924577355385, + 0.06088979169726372, + 0.014254650101065636, + 0.0052843717858195305, + 0.008959842845797539, + 0.033357731997966766, + 0.008950051851570606, + 0.010071803815662861, + 0.005653680302202702, + 0.012900080531835556, + 0.0069281128235161304, + 0.013672051019966602, + 0.01409810408949852, + 0.01247256901115179, + 0.010461870580911636, + 0.005414052866399288, + 0.003546570660546422, + 0.0056505403481423855, + 0.006262633483856916, + 0.00480810459703207, + 0.010032902471721172, + 0.007129393517971039, + 0.00703709851950407, + 0.03745532035827637, + 0.031223440542817116, + 0.04231724143028259, + 0.04394203796982765, + 0.03222127631306648, + 0.054850559681653976, + 0.05132457613945007, + 0.053016290068626404, + 0.0032776654697954655, + 0.0031552293803542852, + 0.005602193996310234, + 0.003120309440419078, + 0.002795883221551776, + 0.0021131534595042467, + 0.003942553885281086, + 0.0036359592340886593, + 0.0024067354388535023, + 0.0038079586811363697, + 0.0030676699243485928, + 0.0013716969406232238, + 0.0018024579621851444, + 0.0018222633516415954, + 0.001739204628393054, + 0.0022519391495734453, + 0.002655661664903164, + 0.0038111817557364702, + 0.0033378980588167906, + 0.003658399684354663, + 0.001812367932870984, + 0.0016446012305095792, + 0.0032757797744125128, + 0.002429130021482706, + 0.0022778960410505533, + 0.009302234277129173, + 0.0014380017528310418, + 0.0009327618172392249, + 0.0011853412725031376, + 0.0027454597875475883, + 0.001029395847581327, + 0.008988888934254646 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.02057494968175888 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011074741370975971 + ], + "opposite_logprob_tokens": [ + -4.511074542999268 + ], + "logprob_reward": -0.011074741370975971, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011074741370975971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24764639139175415, + 0.07707389444112778, + 0.13803628087043762, + 0.053654350340366364, + 0.060162514448165894, + 0.1616637408733368, + 0.07386253029108047, + 0.0690939873456955, + 0.09348172694444656, + 0.025324607267975807 + ], + "credit_map": [ + 0.03533337265253067, + 0.0291385967284441, + 0.028708912432193756, + 0.03074255958199501, + 0.02337912656366825, + 0.03146028146147728, + 0.032569531351327896, + 0.036314014345407486, + 0.00807784590870142, + 0.00768490182235837, + 0.015107233077287674, + 0.011379055678844452, + 0.005080589093267918, + 0.011750834994018078, + 0.008875423111021519, + 0.00911801215261221, + 0.017329620197415352, + 0.00674149114638567, + 0.050671741366386414, + 0.01160325575619936, + 0.005928359925746918, + 0.006184612866491079, + 0.03214532136917114, + 0.007431878708302975, + 0.005271542351692915, + 0.003736106213182211, + 0.007275479380041361, + 0.006244316231459379, + 0.010983528569340706, + 0.004163553472608328, + 0.0084080183878541, + 0.007571806199848652, + 0.005522529594600201, + 0.006040843669325113, + 0.005498276092112064, + 0.009682601317763329, + 0.007918639108538628, + 0.007109927013516426, + 0.00937519408762455, + 0.009014503099024296, + 0.0199226476252079, + 0.013208581134676933, + 0.025182433426380157, + 0.0245126374065876, + 0.015898242592811584, + 0.018599294126033783, + 0.022969556972384453, + 0.021370330825448036, + 0.011322415433824062, + 0.010404514148831367, + 0.011448594741523266, + 0.006561114452779293, + 0.006664699409157038, + 0.007797780446708202, + 0.008739215321838856, + 0.010924188420176506, + 0.007849814370274544, + 0.01594318263232708, + 0.00885713379830122, + 0.004347286652773619, + 0.008232689462602139, + 0.007495887111872435, + 0.0061911316588521, + 0.010176870040595531, + 0.006802240386605263, + 0.012162517756223679, + 0.01275723334401846, + 0.018624303862452507, + 0.009300361387431622, + 0.008829471655189991, + 0.013696705922484398, + 0.011308894492685795, + 0.0038953821640461683, + 0.00983611773699522, + 0.0016717336839064956, + 0.001286897575482726, + 0.0012790836626663804, + 0.003646333934739232, + 0.00100723747164011, + 0.002701823366805911 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.011074741370975971 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.25212955474853516 + ], + "opposite_logprob_tokens": [ + -1.5021295547485352 + ], + "logprob_reward": -0.25212955474853516, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.25212955474853516, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12570340931415558, + 0.11744470149278641, + 0.1029801219701767, + 0.057471536099910736, + 0.06990603357553482, + 0.08085787296295166, + 0.0873822271823883, + 0.12653225660324097, + 0.0844220519065857, + 0.14729979634284973 + ], + "credit_map": [ + 0.014290525577962399, + 0.013283244334161282, + 0.01648334227502346, + 0.021833011880517006, + 0.00994729995727539, + 0.012845364399254322, + 0.01856953278183937, + 0.01845109462738037, + 0.007817533798515797, + 0.012735757976770401, + 0.030442766845226288, + 0.008666936308145523, + 0.007050172425806522, + 0.008772499859333038, + 0.03399960324168205, + 0.007959438487887383, + 0.014820443466305733, + 0.005027688108384609, + 0.03796021267771721, + 0.00879100151360035, + 0.004811273422092199, + 0.003587243380025029, + 0.022335823625326157, + 0.00564642995595932, + 0.004428656306117773, + 0.007462673354893923, + 0.007614866364747286, + 0.011527450755238533, + 0.009352810680866241, + 0.004172303713858128, + 0.006226913072168827, + 0.0066858637146651745, + 0.00936875119805336, + 0.007038153242319822, + 0.0061860294081270695, + 0.005815418437123299, + 0.008284217678010464, + 0.014936285093426704, + 0.00756800314411521, + 0.01070917397737503, + 0.006934648379683495, + 0.00517614372074604, + 0.012158024124801159, + 0.010465972125530243, + 0.009755154140293598, + 0.009001207537949085, + 0.01329877506941557, + 0.01406794972717762, + 0.010687383823096752, + 0.011860828846693039, + 0.016640298068523407, + 0.008821594528853893, + 0.008341746404767036, + 0.007762749679386616, + 0.010195452719926834, + 0.013072173111140728, + 0.011735516600310802, + 0.02364993467926979, + 0.022565297782421112, + 0.007314042188227177, + 0.01418481208384037, + 0.01611989364027977, + 0.01206911914050579, + 0.018893634900450706, + 0.014533411711454391, + 0.01778782345354557, + 0.008440665900707245, + 0.008821644820272923, + 0.008025423623621464, + 0.006475455593317747, + 0.009454555809497833, + 0.010883073322474957, + 0.011561153456568718, + 0.03867418318986893, + 0.009024159982800484, + 0.006515068933367729, + 0.006965648382902145, + 0.015552185475826263, + 0.0058798105455935, + 0.053127583116292953 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.25212955474853516 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot, captured from a lofty height, captures a breathtaking panoramic view of the majestic mountain ranges as the first rays of morning's golden sun bathe the mountain's rolling hills. The jagged, snow-capped peaks shimmer in the golden light, while the deep valleys remain shrouded in a tranquil, deep blue. A thin, morning mist veils the coniferous forests lining the mountainside. A massive, fully-fledged eagle, its dark brown feathers gleaming metallically in the sunlight, circles steadily. Suddenly, without hesitation, it folds its broad wings back, its body transforming into a streamlined, living arrow, swooping vertically with breathtaking acceleration towards the deep valley below. With equally swift speed, the camera tilts downward and zooms in, closely following the eagle's dive. Its piercing amber eyes are clearly visible, fixed on an unseen target below, and the wind ruffles the feathers on its neck. During this rapid fall, the eagle's tiny and swift figure forms a strong and shocking visual contrast with the eternally still and enormous mountain behind it. The whole picture is full of primitive, wild power and breathtaking tension, creating an epic sense of awe of nature.", + "seed": 10973, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_2448_seed_10973_dp_23_rank_46_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0214330873452127, + "vlm_logprob": -0.0214330873452127, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.0214330873452127, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01463382225483656 + ], + "opposite_logprob_tokens": [ + -4.264633655548096 + ], + "logprob_reward": -0.01463382225483656, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01463382225483656, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24387218058109283, + 0.11506573110818863, + 0.03863270580768585, + 0.07091432064771652, + 0.03754190355539322, + 0.03368932381272316, + 0.2427458018064499, + 0.10347620397806168, + 0.06884120404720306, + 0.045220617204904556 + ], + "credit_map": [ + 0.029369324445724487, + 0.07119474560022354, + 0.020970143377780914, + 0.015632526949048042, + 0.018075158819556236, + 0.013587726280093193, + 0.05824332311749458, + 0.016799243167042732, + 0.012306377291679382, + 0.028139818459749222, + 0.00849644374102354, + 0.007602517027407885, + 0.018465355038642883, + 0.013072814792394638, + 0.01875675469636917, + 0.00822564959526062, + 0.004205306060612202, + 0.0036315678153187037, + 0.004163368139415979, + 0.008078795857727528, + 0.004668490961194038, + 0.00362953869625926, + 0.0060305180959403515, + 0.004225123208016157, + 0.009707028977572918, + 0.00431057158857584, + 0.006371073424816132, + 0.011354242451488972, + 0.004110001493245363, + 0.004863659851253033, + 0.010630358941853046, + 0.01956738717854023, + 0.0029806215316057205, + 0.003197343787178397, + 0.004853499121963978, + 0.0031477652955800295, + 0.00406640162691474, + 0.006863167509436607, + 0.0066747162491083145, + 0.0057583884336054325, + 0.002411746885627508, + 0.0034176548942923546, + 0.0043223947286605835, + 0.0028980569913983345, + 0.0047004008665680885, + 0.005556249059736729, + 0.005081037525087595, + 0.005301782861351967, + 0.015764135867357254, + 0.018298732116818428, + 0.02632189355790615, + 0.02643848955631256, + 0.015175500884652138, + 0.06080949679017067, + 0.024328524246811867, + 0.05560903623700142, + 0.01556654367595911, + 0.010955371893942356, + 0.016029763966798782, + 0.01420100498944521, + 0.010327309370040894, + 0.011862625367939472, + 0.0124201700091362, + 0.012113414704799652, + 0.006995093543082476, + 0.0072052921168506145, + 0.009889516979455948, + 0.009632070548832417, + 0.008115655742585659, + 0.007749136537313461, + 0.00911833718419075, + 0.010136102326214314, + 0.005094410385936499, + 0.0045830849558115005, + 0.005529443267732859, + 0.005257982295006514, + 0.00416537607088685, + 0.004517796915024519, + 0.0075096506625413895, + 0.008562875911593437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.01463382225483656 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026370229199528694 + ], + "opposite_logprob_tokens": [ + -3.6513702869415283 + ], + "logprob_reward": -0.026370229199528694, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026370229199528694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17872492969036102, + 0.08036547154188156, + 0.04875081405043602, + 0.03561842814087868, + 0.11743699759244919, + 0.13315115869045258, + 0.13701796531677246, + 0.1710222214460373, + 0.04003223776817322, + 0.05787977576255798 + ], + "credit_map": [ + 0.020268477499485016, + 0.0485728457570076, + 0.015345785766839981, + 0.012888839468359947, + 0.015916408970952034, + 0.01881476491689682, + 0.03611697256565094, + 0.01080083567649126, + 0.009459310211241245, + 0.016937682405114174, + 0.006211176980286837, + 0.007379849907010794, + 0.012121778912842274, + 0.009470071643590927, + 0.009911141358315945, + 0.008874458260834217, + 0.004271605052053928, + 0.005521527491509914, + 0.007098976522684097, + 0.007619335316121578, + 0.005810067988932133, + 0.005844836123287678, + 0.007136396132409573, + 0.005448071286082268, + 0.003792425151914358, + 0.0030124583281576633, + 0.005712105892598629, + 0.0055074067786335945, + 0.004889755509793758, + 0.003284822916612029, + 0.005067357327789068, + 0.004352097399532795, + 0.008688841015100479, + 0.011586913838982582, + 0.018258290365338326, + 0.01191877480596304, + 0.009443962946534157, + 0.010419156402349472, + 0.026292376220226288, + 0.02082868106663227, + 0.010633977130055428, + 0.013829745352268219, + 0.015388429164886475, + 0.010007305070757866, + 0.01827162876725197, + 0.022629207000136375, + 0.020331911742687225, + 0.022058948874473572, + 0.011213966645300388, + 0.019093550741672516, + 0.02012048289179802, + 0.01765187457203865, + 0.011264879256486893, + 0.023425355553627014, + 0.01611434482038021, + 0.018133513629436493, + 0.01833527535200119, + 0.02823982946574688, + 0.01314332615584135, + 0.009983747266232967, + 0.01684163324534893, + 0.016867784783244133, + 0.02986534684896469, + 0.03774527832865715, + 0.0039694057777523994, + 0.004106555134057999, + 0.006282770074903965, + 0.0068811397068202496, + 0.002965961815789342, + 0.0042112600058317184, + 0.005657893139868975, + 0.005957248155027628, + 0.0042114946991205215, + 0.0033167367801070213, + 0.007436185143887997, + 0.0071357241831719875, + 0.0068966238759458065, + 0.008590588346123695, + 0.008344789035618305, + 0.0119476318359375 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 6 + ], + "selected_reward": -0.026370229199528694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253083072602749 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253083072602749, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253083072602749, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2962576448917389, + 0.10637319087982178, + 0.0967218205332756, + 0.09975773841142654, + 0.0760514959692955, + 0.04968426376581192, + 0.1476539820432663, + 0.05203679949045181, + 0.02888793870806694, + 0.04657517001032829 + ], + "credit_map": [ + 0.0355120450258255, + 0.08062218129634857, + 0.022707417607307434, + 0.020972300320863724, + 0.03805084899067879, + 0.04205353185534477, + 0.03414226695895195, + 0.02219703607261181, + 0.007876522839069366, + 0.0180559903383255, + 0.008313399739563465, + 0.009776589460670948, + 0.013411851599812508, + 0.011610104702413082, + 0.023242613300681114, + 0.014086115173995495, + 0.016972268000245094, + 0.008904245682060719, + 0.008670969866216183, + 0.01098970789462328, + 0.019427737221121788, + 0.013784070499241352, + 0.009758411906659603, + 0.008214402012526989, + 0.017498351633548737, + 0.005933911073952913, + 0.0038454828318208456, + 0.012563945725560188, + 0.008841104805469513, + 0.008561944589018822, + 0.014586416073143482, + 0.027926580980420113, + 0.005528636276721954, + 0.006990508176386356, + 0.01129799336194992, + 0.007828038185834885, + 0.005861629731953144, + 0.006458665244281292, + 0.017696170136332512, + 0.014389852993190289, + 0.003994123078882694, + 0.004486461635679007, + 0.004347150679677725, + 0.0030960666481405497, + 0.0075428965501487255, + 0.005348685197532177, + 0.010353855788707733, + 0.010515021160244942, + 0.012085122987627983, + 0.015057317912578583, + 0.028219541534781456, + 0.026001758873462677, + 0.008779889903962612, + 0.02062978781759739, + 0.019478540867567062, + 0.017402034252882004, + 0.005975435022264719, + 0.009300688281655312, + 0.005009811371564865, + 0.00584055669605732, + 0.006257317960262299, + 0.005777647718787193, + 0.007154479622840881, + 0.006720861420035362, + 0.004465410020202398, + 0.003431869437918067, + 0.003667257959023118, + 0.004505959805101156, + 0.0015663173981010914, + 0.002494495827704668, + 0.004003387410193682, + 0.004753239452838898, + 0.003592483000829816, + 0.0031178537756204605, + 0.00852823443710804, + 0.00546813290566206, + 0.005617129150778055, + 0.00757426954805851, + 0.0048427642323076725, + 0.007834300398826599 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005253083072602749 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005948815029114485 + ], + "opposite_logprob_tokens": [ + -5.130949020385742 + ], + "logprob_reward": -0.005948815029114485, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005948815029114485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0782741978764534, + 0.09061773121356964, + 0.0468081459403038, + 0.036280494183301926, + 0.0577092245221138, + 0.06723885238170624, + 0.35906586050987244, + 0.1460525095462799, + 0.04536232724785805, + 0.07259067893028259 + ], + "credit_map": [ + 0.008833425119519234, + 0.018577154725790024, + 0.008250998333096504, + 0.005844280123710632, + 0.009178309701383114, + 0.008226802572607994, + 0.012846274301409721, + 0.006516946479678154, + 0.011191774159669876, + 0.020516611635684967, + 0.006971281953155994, + 0.006603078916668892, + 0.01645345613360405, + 0.012647421099245548, + 0.010076506063342094, + 0.0061575984582304955, + 0.00436019990593195, + 0.004563672933727503, + 0.004961459897458553, + 0.007126184646040201, + 0.0065637193620204926, + 0.006454386282712221, + 0.0062935324385762215, + 0.006484988611191511, + 0.004314302001148462, + 0.003171134041622281, + 0.004854752216488123, + 0.005119585897773504, + 0.005045440047979355, + 0.004204245284199715, + 0.005129074677824974, + 0.004441962111741304, + 0.004785893484950066, + 0.0058779665268957615, + 0.007607809267938137, + 0.007330622058361769, + 0.004141986835747957, + 0.00909155048429966, + 0.009967478923499584, + 0.008905918337404728, + 0.00499931164085865, + 0.006407227832823992, + 0.010072492994368076, + 0.005961824674159288, + 0.008852428756654263, + 0.009298558346927166, + 0.010590252466499805, + 0.011056755669414997, + 0.0225442536175251, + 0.026406001299619675, + 0.04702706262469292, + 0.04260547459125519, + 0.018919680267572403, + 0.0895993709564209, + 0.03965736925601959, + 0.07230666279792786, + 0.021668938919901848, + 0.022823650389909744, + 0.012070940807461739, + 0.011181517504155636, + 0.018250074237585068, + 0.01561721321195364, + 0.02880811132490635, + 0.01563206873834133, + 0.005856896284967661, + 0.005717113614082336, + 0.007142951712012291, + 0.0071904840879142284, + 0.0048048365861177444, + 0.0043557630851864815, + 0.004815222229808569, + 0.005479058716446161, + 0.0054271407425403595, + 0.006018443964421749, + 0.010463540442287922, + 0.009033992886543274, + 0.009701109491288662, + 0.012701678089797497, + 0.009366204962134361, + 0.009878569282591343 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 1 + ], + "selected_reward": -0.005948815029114485 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054959487169981 + ], + "opposite_logprob_tokens": [ + -2.929959535598755 + ], + "logprob_reward": -0.054959487169981, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054959487169981, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18858174979686737, + 0.035114794969558716, + 0.055181145668029785, + 0.041269369423389435, + 0.10656624287366867, + 0.03895978629589081, + 0.1647886484861374, + 0.09535098075866699, + 0.18100552260875702, + 0.0931818038225174 + ], + "credit_map": [ + 0.020659975707530975, + 0.05147466063499451, + 0.01719735935330391, + 0.01494729332625866, + 0.01345873810350895, + 0.009466213174164295, + 0.046733930706977844, + 0.014643566682934761, + 0.003965830430388451, + 0.005911481101065874, + 0.00297390460036695, + 0.003526626154780388, + 0.003953680396080017, + 0.0031316920649260283, + 0.007373852655291557, + 0.004277726169675589, + 0.00963850412517786, + 0.004042271990329027, + 0.005828267429023981, + 0.006286701187491417, + 0.012469783425331116, + 0.0062789092771708965, + 0.005789618473500013, + 0.004847091156989336, + 0.008168323896825314, + 0.003118576016277075, + 0.004421938676387072, + 0.00673283077776432, + 0.0028718069661408663, + 0.0030573986005038023, + 0.005249569658190012, + 0.00764892203733325, + 0.008280791342258453, + 0.009683225303888321, + 0.014235805720090866, + 0.009682368487119675, + 0.00864112377166748, + 0.013738103210926056, + 0.023087574169039726, + 0.01921725831925869, + 0.0026522844564169645, + 0.0030115230474621058, + 0.0032100272364914417, + 0.0038810870610177517, + 0.004678354598581791, + 0.003776747500523925, + 0.009835917502641678, + 0.007913846522569656, + 0.01329764723777771, + 0.016337063163518906, + 0.03141180798411369, + 0.028068969026207924, + 0.00921130832284689, + 0.024610236287117004, + 0.022242793813347816, + 0.019608817994594574, + 0.010237644426524639, + 0.010394759476184845, + 0.012584799900650978, + 0.014601963572204113, + 0.012217089533805847, + 0.013273876160383224, + 0.011494804173707962, + 0.010546036064624786, + 0.03163043409585953, + 0.02561080828309059, + 0.019928786903619766, + 0.028380153700709343, + 0.013796684332191944, + 0.016123680397868156, + 0.022136583924293518, + 0.023398397490382195, + 0.011245462112128735, + 0.011589853093028069, + 0.01858382858335972, + 0.011691950261592865, + 0.008770832791924477, + 0.008070851676166058, + 0.011057617142796516, + 0.012171415612101555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 6 + ], + "selected_reward": -0.054959487169981 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic drone shot, captured from a lofty height, captures a breathtaking panoramic view of the majestic mountain ranges as the first rays of morning's golden sun bathe the mountain's rolling hills. The jagged, snow-capped peaks shimmer in the golden light, while the deep valleys remain shrouded in a tranquil, deep blue. A thin, morning mist veils the coniferous forests lining the mountainside. A massive, fully-fledged eagle, its dark brown feathers gleaming metallically in the sunlight, circles steadily. Suddenly, without hesitation, it folds its broad wings back, its body transforming into a streamlined, living arrow, swooping vertically with breathtaking acceleration towards the deep valley below. With equally swift speed, the camera tilts downward and zooms in, closely following the eagle's dive. Its piercing amber eyes are clearly visible, fixed on an unseen target below, and the wind ruffles the feathers on its neck. During this rapid fall, the eagle's tiny and swift figure forms a strong and shocking visual contrast with the eternally still and enormous mountain behind it. The whole picture is full of primitive, wild power and breathtaking tension, creating an epic sense of awe of nature.", + "seed": 10974, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_11/index_2448_seed_10974_dp_23_rank_46_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 11, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.10418436080217361, + "vlm_logprob": -0.10418436080217361, + "vlm_yes_no_margin": 2.675, + "vlm_token_credit": -0.10418436080217361, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10077755153179169 + ], + "opposite_logprob_tokens": [ + -2.3507776260375977 + ], + "logprob_reward": -0.10077755153179169, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10077755153179169, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09832023084163666, + 0.05023174732923508, + 0.11586268991231918, + 0.06397131085395813, + 0.026804078370332718, + 0.03585313633084297, + 0.18882805109024048, + 0.14914768934249878, + 0.16644389927387238, + 0.10453715920448303 + ], + "credit_map": [ + 0.008273380808532238, + 0.007177354767918587, + 0.017182227224111557, + 0.011455582454800606, + 0.010749763809144497, + 0.015537306666374207, + 0.012447194196283817, + 0.01549740880727768, + 0.004026459529995918, + 0.004958962090313435, + 0.0041336375288665295, + 0.007316789589822292, + 0.007524255663156509, + 0.006157731171697378, + 0.00813037808984518, + 0.007983529940247536, + 0.010140540078282356, + 0.010389504954218864, + 0.013517217710614204, + 0.013073330745100975, + 0.01708710938692093, + 0.020519467070698738, + 0.01164537388831377, + 0.019490152597427368, + 0.0056733060628175735, + 0.005456722341477871, + 0.00549068721011281, + 0.011168992146849632, + 0.006736730691045523, + 0.004300136584788561, + 0.013551663607358932, + 0.011593067087233067, + 0.002481235656887293, + 0.002543980022892356, + 0.0033334088511765003, + 0.003972445148974657, + 0.004030496813356876, + 0.0025694197975099087, + 0.004212766420096159, + 0.0036603272892534733, + 0.002676868811249733, + 0.003816901706159115, + 0.0035951167810708284, + 0.006091665476560593, + 0.005210124421864748, + 0.00480706337839365, + 0.004238645546138287, + 0.005416752304881811, + 0.0178429763764143, + 0.023890625685453415, + 0.0184661652892828, + 0.01948048174381256, + 0.02190427854657173, + 0.03417644649744034, + 0.028905505314469337, + 0.024161556735634804, + 0.01659329980611801, + 0.021077968180179596, + 0.0251767635345459, + 0.02171754278242588, + 0.012778904289007187, + 0.014642239548265934, + 0.015192371793091297, + 0.021968601271510124, + 0.01898530311882496, + 0.01914418302476406, + 0.04124557971954346, + 0.012816681526601315, + 0.011833281256258488, + 0.014140686951577663, + 0.012691649608314037, + 0.03558654338121414, + 0.011347743682563305, + 0.010406664572656155, + 0.011828032322227955, + 0.015727806836366653, + 0.007708428893238306, + 0.01110328733921051, + 0.015729369595646858, + 0.02068583108484745 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 7 + ], + "selected_reward": -0.10077755153179169 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430864654481411 + ], + "opposite_logprob_tokens": [ + -3.168086528778076 + ], + "logprob_reward": -0.0430864654481411, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430864654481411, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12753772735595703, + 0.15792059898376465, + 0.09339514374732971, + 0.05131278932094574, + 0.04345649108290672, + 0.04294872283935547, + 0.14961779117584229, + 0.12627370655536652, + 0.1447959691286087, + 0.06274111568927765 + ], + "credit_map": [ + 0.01371146459132433, + 0.008194583468139172, + 0.017597652971744537, + 0.020690802484750748, + 0.014286715537309647, + 0.01881834864616394, + 0.01767711341381073, + 0.01656104438006878, + 0.02079114504158497, + 0.011590673588216305, + 0.01298836525529623, + 0.023725111037492752, + 0.022431617602705956, + 0.017897946760058403, + 0.025469642132520676, + 0.023026086390018463, + 0.008934401907026768, + 0.010048051364719868, + 0.012410491704940796, + 0.013816837221384048, + 0.010118288919329643, + 0.012509934604167938, + 0.010252553969621658, + 0.015304573811590672, + 0.004293559119105339, + 0.0037121749483048916, + 0.00694128917530179, + 0.007028070278465748, + 0.005467959214001894, + 0.005482162348926067, + 0.009239655919373035, + 0.009147911332547665, + 0.004309150390326977, + 0.0035437762271612883, + 0.006189520936459303, + 0.007404564879834652, + 0.006605787668377161, + 0.004812419414520264, + 0.005088085308670998, + 0.005503182765096426, + 0.00442741671577096, + 0.006140814628452063, + 0.004446112550795078, + 0.007514127064496279, + 0.004130829591304064, + 0.0035101270768791437, + 0.0060189152136445045, + 0.006760377436876297, + 0.014029907062649727, + 0.020139120519161224, + 0.01316810492426157, + 0.02387867495417595, + 0.017239557579159737, + 0.020949462428689003, + 0.021200215443968773, + 0.019012732431292534, + 0.01337353978306055, + 0.012348388321697712, + 0.02370086871087551, + 0.018015269190073013, + 0.010632204823195934, + 0.015015929937362671, + 0.012903805822134018, + 0.020283691585063934, + 0.01580427587032318, + 0.01736118644475937, + 0.016906900331377983, + 0.012368897907435894, + 0.01561592984944582, + 0.016816210001707077, + 0.02280501089990139, + 0.027117550373077393, + 0.006134769879281521, + 0.005152955651283264, + 0.012160140089690685, + 0.010831248946487904, + 0.005005340091884136, + 0.005187923554331064, + 0.008093278855085373, + 0.010175454430282116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 8 + ], + "selected_reward": -0.0430864654481411 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280107617378235 + ], + "opposite_logprob_tokens": [ + -2.01780104637146 + ], + "logprob_reward": -0.14280107617378235, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.14280107617378235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41646698117256165, + 0.11511446535587311, + 0.12943114340305328, + 0.048826996237039566, + 0.03952370956540108, + 0.036189790815114975, + 0.054726582020521164, + 0.07630500942468643, + 0.032943256199359894, + 0.05047208443284035 + ], + "credit_map": [ + 0.051020026206970215, + 0.03886728733778, + 0.04579879343509674, + 0.051156360656023026, + 0.052159179002046585, + 0.06661780923604965, + 0.049290403723716736, + 0.0615571103990078, + 0.010192141868174076, + 0.012144333682954311, + 0.008866696618497372, + 0.016151782125234604, + 0.018120845779776573, + 0.0176621675491333, + 0.015216367319226265, + 0.01676012948155403, + 0.012304645963013172, + 0.010379822924733162, + 0.015579485334455967, + 0.012040751986205578, + 0.022322364151477814, + 0.01923523098230362, + 0.01230707298964262, + 0.025261763483285904, + 0.003897579852491617, + 0.0036324092652648687, + 0.005093216896057129, + 0.007481347769498825, + 0.007834888994693756, + 0.0056292032822966576, + 0.0075714318081736565, + 0.007686918135732412, + 0.0039049864280968904, + 0.003877092618495226, + 0.003576706163585186, + 0.006339855957776308, + 0.0055562290363013744, + 0.005262315738946199, + 0.005446332041174173, + 0.005560189485549927, + 0.003566032275557518, + 0.004625123925507069, + 0.0032687396742403507, + 0.006323218811303377, + 0.005580615252256393, + 0.0036914406809955835, + 0.004374471493065357, + 0.00476014893501997, + 0.005176817066967487, + 0.00917903147637844, + 0.006299959030002356, + 0.007820937782526016, + 0.004735617432743311, + 0.007142421789467335, + 0.006934599485248327, + 0.007437197957187891, + 0.007140485569834709, + 0.008174946531653404, + 0.014974483288824558, + 0.013081532903015614, + 0.005383613985031843, + 0.00786785501986742, + 0.008569389581680298, + 0.011112702079117298, + 0.0036108940839767456, + 0.004099049139767885, + 0.007847780361771584, + 0.002601982094347477, + 0.002648446010425687, + 0.0034066534135490656, + 0.0033370989840477705, + 0.005391351878643036, + 0.005091926548629999, + 0.006363958586007357, + 0.0059161619283258915, + 0.009623672813177109, + 0.004663220141083002, + 0.003642757423222065, + 0.00865880586206913, + 0.006511578802019358 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.14280107617378235 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008640285581350327 + ], + "opposite_logprob_tokens": [ + -4.758640289306641 + ], + "logprob_reward": -0.008640285581350327, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008640285581350327, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08466258645057678, + 0.06120643764734268, + 0.054125502705574036, + 0.06969199329614639, + 0.02578924037516117, + 0.0768723264336586, + 0.11938314139842987, + 0.12373393028974533, + 0.2602642774581909, + 0.12427059561014175 + ], + "credit_map": [ + 0.010556385852396488, + 0.006111779250204563, + 0.012292005121707916, + 0.01021656021475792, + 0.009740370325744152, + 0.010100454092025757, + 0.0112714609131217, + 0.014373572543263435, + 0.0064265914261341095, + 0.006090749055147171, + 0.005576412659138441, + 0.008730484172701836, + 0.009314266964793205, + 0.00871029868721962, + 0.008304600603878498, + 0.008053036406636238, + 0.005060558207333088, + 0.004764888435602188, + 0.0052606514655053616, + 0.005342706572264433, + 0.010506724938750267, + 0.011739368550479412, + 0.004694046918302774, + 0.006756552495062351, + 0.005980404559522867, + 0.006132600363343954, + 0.0053621744737029076, + 0.013042252510786057, + 0.00821383111178875, + 0.004743405617773533, + 0.014006673358380795, + 0.012210655026137829, + 0.0027745135594159365, + 0.002856664126738906, + 0.0025284301955252886, + 0.0036048886831849813, + 0.0042027863673865795, + 0.0030831105541437864, + 0.003214901313185692, + 0.003523945575580001, + 0.0047947415150702, + 0.008110052905976772, + 0.010518255643546581, + 0.011427471414208412, + 0.011263000778853893, + 0.012895847670733929, + 0.007586834952235222, + 0.010276124812662601, + 0.010976891033351421, + 0.016601432114839554, + 0.008480971679091454, + 0.01618538610637188, + 0.013913974165916443, + 0.017708513885736465, + 0.020101116970181465, + 0.015414867550134659, + 0.009761680848896503, + 0.013036922551691532, + 0.018337691202759743, + 0.016257410869002342, + 0.017353516072034836, + 0.014889407902956009, + 0.014231055974960327, + 0.019866235554218292, + 0.025475075468420982, + 0.03460915386676788, + 0.04072345793247223, + 0.027217645198106766, + 0.024680882692337036, + 0.02767917700111866, + 0.037997737526893616, + 0.04188114404678345, + 0.01164610031992197, + 0.009263110347092152, + 0.015282735228538513, + 0.01821940392255783, + 0.010418042540550232, + 0.016372591257095337, + 0.01691225729882717, + 0.02615634724497795 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 7 + ], + "selected_reward": -0.008640285581350327 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is a massive eagle with dark brown feathers visible in the video?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the eagle perform a vertical dive with acceleration towards a deep valley?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Are snow-capped mountain peaks visible shimmering in golden sunlight?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera tilt downward and zoom in to follow the eagle's dive?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Are the eagle's amber eyes clearly visible and fixed on a target?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2256164252758026 + ], + "opposite_logprob_tokens": [ + -1.600616455078125 + ], + "logprob_reward": -0.2256164252758026, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.2256164252758026, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.053167570382356644, + 0.04886888340115547, + 0.09190624207258224, + 0.0341782346367836, + 0.015499170869588852, + 0.05865449458360672, + 0.15049947798252106, + 0.1634530872106552, + 0.2336711436510086, + 0.1501016914844513 + ], + "credit_map": [ + 0.00661521265283227, + 0.0044318451546132565, + 0.007590772118419409, + 0.008610976859927177, + 0.004931892734020948, + 0.00646241195499897, + 0.007443228270858526, + 0.0070812334306538105, + 0.004599347710609436, + 0.004171410109847784, + 0.008107083849608898, + 0.00677194818854332, + 0.004202994983643293, + 0.005921864416450262, + 0.006431365851312876, + 0.008662872947752476, + 0.009679151698946953, + 0.009007422253489494, + 0.009118522517383099, + 0.01060619205236435, + 0.013655449263751507, + 0.01668342761695385, + 0.00954504869878292, + 0.013611024245619774, + 0.0032179856207221746, + 0.003083626041188836, + 0.004120636265724897, + 0.003999981097877026, + 0.005505699664354324, + 0.005843874532729387, + 0.0035775681026279926, + 0.004828860517591238, + 0.00281366310082376, + 0.0018008166225627065, + 0.0015188725665211678, + 0.0021066712215542793, + 0.002267664996907115, + 0.001623277086764574, + 0.001595280016772449, + 0.0017729259561747313, + 0.00417822552844882, + 0.007026633247733116, + 0.005194263998419046, + 0.012729447335004807, + 0.008341881446540356, + 0.0062050651758909225, + 0.005964950192719698, + 0.009014023467898369, + 0.013082016259431839, + 0.015590447932481766, + 0.01255010161548853, + 0.024313045665621758, + 0.01528005488216877, + 0.018925439566373825, + 0.03271953761577606, + 0.018038829788565636, + 0.01841498538851738, + 0.021791154518723488, + 0.029383277520537376, + 0.02117755077779293, + 0.014292718842625618, + 0.017230574041604996, + 0.017119985073804855, + 0.024042844772338867, + 0.0365314707159996, + 0.02885686792433262, + 0.037493184208869934, + 0.016827471554279327, + 0.021429790183901787, + 0.019112708047032356, + 0.024435242637991905, + 0.04898442327976227, + 0.015152611769735813, + 0.011634149588644505, + 0.017464620992541313, + 0.02116728015244007, + 0.011480080895125866, + 0.018872732296586037, + 0.021866895258426666, + 0.03246331959962845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 6 + ], + "selected_reward": -0.2256164252758026 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_tensor([5642, 5642])_group_12_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_tensor([5642, 5642])_group_12_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..b7011eb6e4820b41a75de1459042ab7fa45bc3e9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_tensor([5642, 5642])_group_12_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium shot, shot from a slightly low angle and sideways, steadily follows an agitated figure as they pace relentlessly back and forth within the narrow confines of a dimly lit nighttime alley. The figure has messy short hair plastered to a pale, sweaty forehead and wears a simple dark wool coat over an unbuttoned gray T-shirt. As they take three steps on the damp cobblestones, their bodies tense, their shoulders hunched, suddenly rotating, and then repeating the motion. Their frantic gaze darts across the grimy brick wall before darting to their right hand, which clutches a gleaming gold coin and shakes uncontrollably. Their knuckles white with pressure, their thumbs obsessively rubbing the intricate, worn insignia on the coin's surface. A flickering street lamp overhead casts long, dancing shadows, creating a stark contrast of light and dark, accentuating the sharp planes of their faces and the metallic sheen of the coin. The overall color palette is cold and desaturated, dominated by the dark, damp tones of the alley, with the only warmth coming from sickly yellow light and brilliant gold, amplifying the scene's palpable atmosphere of intense anxiety and desperate indecision.", + "seed": 7590, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_5642_seed_7590_dp_24_rank_48_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.027648237813264132, + "vlm_logprob": -0.027648237813264132, + "vlm_yes_no_margin": 3.925, + "vlm_token_credit": -0.027648237813264132, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.040379200130701065 + ], + "opposite_logprob_tokens": [ + -3.415379285812378 + ], + "logprob_reward": -0.040379200130701065, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.040379200130701065, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21471871435642242, + 0.14597150683403015, + 0.026575593277812004, + 0.03600158914923668, + 0.389676570892334, + 0.03310853987932205, + 0.013208496384322643, + 0.057636626064777374, + 0.03758859634399414, + 0.04551377892494202 + ], + "credit_map": [ + 0.023166870698332787, + 0.020103003829717636, + 0.023586032912135124, + 0.023724252358078957, + 0.03160073980689049, + 0.02198745869100094, + 0.04442385956645012, + 0.026126502081751823, + 0.010040704160928726, + 0.010226808488368988, + 0.013254809193313122, + 0.016066227108240128, + 0.034179769456386566, + 0.02182832546532154, + 0.027250155806541443, + 0.013124704360961914, + 0.002071324735879898, + 0.00276946066878736, + 0.002942347200587392, + 0.0036776913329958916, + 0.003293492365628481, + 0.003967804368585348, + 0.003323404584079981, + 0.0045300694182515144, + 0.004062136169523001, + 0.0029463712126016617, + 0.003893532557412982, + 0.005065855570137501, + 0.0031369065400213003, + 0.005892424378544092, + 0.005472453776746988, + 0.0055319094099104404, + 0.007635029964148998, + 0.011321748606860638, + 0.005323979537934065, + 0.014133201912045479, + 0.006351286545395851, + 0.16635973751544952, + 0.008621735498309135, + 0.1699298471212387, + 0.0036184939090162516, + 0.0038938457146286964, + 0.005576451309025288, + 0.006388403940945864, + 0.003711162833496928, + 0.003576684510335326, + 0.002928301226347685, + 0.003415194572880864, + 0.001397541956976056, + 0.0023253117688000202, + 0.001567260012961924, + 0.002020427957177162, + 0.0018514853436499834, + 0.0010487569961696863, + 0.0017154283123090863, + 0.0012822842691093683, + 0.00515262084081769, + 0.004730951506644487, + 0.004026200622320175, + 0.005631685256958008, + 0.00741511769592762, + 0.015216095373034477, + 0.007605075370520353, + 0.007858873344957829, + 0.003654759842902422, + 0.0033244576770812273, + 0.003324532648548484, + 0.005965951830148697, + 0.0059835477732121944, + 0.00500150304287672, + 0.005558479111641645, + 0.00477536441758275, + 0.004484688397496939, + 0.0033968158531934023, + 0.0060212318785488605, + 0.0036529910285025835, + 0.006434816401451826, + 0.009295385330915451, + 0.005243557505309582, + 0.006984292529523373 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.040379200130701065 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060239389538765 + ], + "opposite_logprob_tokens": [ + -3.8956024646759033 + ], + "logprob_reward": -0.02060239389538765, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060239389538765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11019149422645569, + 0.21587897837162018, + 0.04361354559659958, + 0.07550637423992157, + 0.23260876536369324, + 0.06479209661483765, + 0.08816508948802948, + 0.09809604287147522, + 0.033714696764945984, + 0.037432923913002014 + ], + "credit_map": [ + 0.012730210088193417, + 0.008848435245454311, + 0.012261025607585907, + 0.012605051510035992, + 0.016297221183776855, + 0.01063358411192894, + 0.02085714228451252, + 0.015958819538354874, + 0.013149049133062363, + 0.01492390874773264, + 0.019287915900349617, + 0.025214223191142082, + 0.05455795302987099, + 0.0321793407201767, + 0.03867148235440254, + 0.017895100638270378, + 0.00368169485591352, + 0.003180606523528695, + 0.005542841274291277, + 0.005281719844788313, + 0.006907924078404903, + 0.007232938427478075, + 0.00664088549092412, + 0.005144933238625526, + 0.009477658197283745, + 0.006613234523683786, + 0.013290401548147202, + 0.010102766565978527, + 0.006040044594556093, + 0.012384424917399883, + 0.008526702411472797, + 0.009071141481399536, + 0.006525106728076935, + 0.007210375741124153, + 0.0069088502787053585, + 0.013920743018388748, + 0.011614734306931496, + 0.0858873650431633, + 0.010994907468557358, + 0.08954667299985886, + 0.005629341583698988, + 0.006409598048776388, + 0.007917426526546478, + 0.009234431199729443, + 0.012293263338506222, + 0.007937958464026451, + 0.0065925149247050285, + 0.00877755880355835, + 0.01039136666804552, + 0.011947551742196083, + 0.009057460352778435, + 0.011209000833332539, + 0.013480283319950104, + 0.008211487904191017, + 0.011530605144798756, + 0.012337331660091877, + 0.016909485682845116, + 0.009719221852719784, + 0.016089504584670067, + 0.011919558048248291, + 0.013514338061213493, + 0.006991440895944834, + 0.011979877017438412, + 0.010972613468766212, + 0.002719976007938385, + 0.0016870484687387943, + 0.0020823609083890915, + 0.0043357671238482, + 0.006173346191644669, + 0.0034900095779448748, + 0.006842567585408688, + 0.006383619736880064, + 0.004751007538288832, + 0.0021236848551779985, + 0.004421154502779245, + 0.002695600502192974, + 0.0037127446848899126, + 0.008589520119130611, + 0.00520746735855937, + 0.005931743420660496 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.02060239389538765 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05497167259454727 + ], + "opposite_logprob_tokens": [ + -2.929971694946289 + ], + "logprob_reward": -0.05497167259454727, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05497167259454727, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03855021670460701, + 0.06567645072937012, + 0.0550905279815197, + 0.05117073282599449, + 0.7006208300590515, + 0.022564949467778206, + 0.029522908851504326, + 0.014318786561489105, + 0.010348417796194553, + 0.012136191129684448 + ], + "credit_map": [ + 0.004211382009088993, + 0.0031374285463243723, + 0.004558319691568613, + 0.004444326274096966, + 0.005584282800555229, + 0.00381990778259933, + 0.007745586801320314, + 0.005048978608101606, + 0.00423264829441905, + 0.004544273484498262, + 0.0062729003839194775, + 0.007971293292939663, + 0.01591923087835312, + 0.00962918158620596, + 0.011312640272080898, + 0.005794278811663389, + 0.003898437600582838, + 0.0038131289184093475, + 0.005944109056144953, + 0.006633122451603413, + 0.010044257156550884, + 0.008516345173120499, + 0.009566963650286198, + 0.00667415838688612, + 0.004403210710734129, + 0.004278333391994238, + 0.008155972696840763, + 0.006871615536510944, + 0.006248363293707371, + 0.008287995122373104, + 0.00861371960490942, + 0.004311514087021351, + 0.013457371853291988, + 0.020023131743073463, + 0.009293639101088047, + 0.02518703043460846, + 0.009576963260769844, + 0.30155226588249207, + 0.01401046197861433, + 0.30751997232437134, + 0.001873454893939197, + 0.0022177950013428926, + 0.0023738080635666847, + 0.0035034921020269394, + 0.0032056730706244707, + 0.0033050444908440113, + 0.002094455761834979, + 0.003991225268691778, + 0.0025034707505255938, + 0.005299036856740713, + 0.0023890677839517593, + 0.003673186060041189, + 0.0044657220132648945, + 0.00432328088209033, + 0.004354796372354031, + 0.0025143467355519533, + 0.0009609540575183928, + 0.0011104036821052432, + 0.0007979981601238251, + 0.0014037677319720387, + 0.0014033546904101968, + 0.002277301624417305, + 0.0025903754867613316, + 0.003774629905819893, + 0.0006615622551180422, + 0.0006227372796274722, + 0.000582995533477515, + 0.0018609047401696444, + 0.0017889668233692646, + 0.0007743599126115441, + 0.0019457227317616343, + 0.0021111671812832355, + 0.001124182133935392, + 0.0008680103928782046, + 0.0010749840876087546, + 0.0010613849153742194, + 0.0013749065110459924, + 0.0031896948348730803, + 0.0015498583670705557, + 0.0018931690137833357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 2 + ], + "selected_reward": -0.05497167259454727 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018186485394835472 + ], + "opposite_logprob_tokens": [ + -4.018186569213867 + ], + "logprob_reward": -0.018186485394835472, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018186485394835472, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08755678683519363, + 0.13300500810146332, + 0.0655844584107399, + 0.03858845680952072, + 0.4697301685810089, + 0.0663137212395668, + 0.03157510235905647, + 0.032636094838380814, + 0.03205224871635437, + 0.042957961559295654 + ], + "credit_map": [ + 0.010697009041905403, + 0.008434277959167957, + 0.009856200776994228, + 0.011490731500089169, + 0.012801066040992737, + 0.007626216858625412, + 0.011702769435942173, + 0.01494851615279913, + 0.0086511904373765, + 0.00956400204449892, + 0.013022441416978836, + 0.01653899811208248, + 0.02957593835890293, + 0.018555544316768646, + 0.026705285534262657, + 0.010391606017947197, + 0.003988952841609716, + 0.004524959251284599, + 0.006411746609956026, + 0.007372615858912468, + 0.010645091533660889, + 0.01065014861524105, + 0.011634395457804203, + 0.010356544516980648, + 0.0033468964975327253, + 0.003922669682651758, + 0.005975817330181599, + 0.005705746356397867, + 0.003811846487224102, + 0.005458848550915718, + 0.006660353392362595, + 0.003706277348101139, + 0.011714320629835129, + 0.015067487955093384, + 0.013209367170929909, + 0.019784605130553246, + 0.010085733607411385, + 0.19250060617923737, + 0.01287844218313694, + 0.19448961317539215, + 0.005175325553864241, + 0.006190586369484663, + 0.011361388489603996, + 0.01395346224308014, + 0.00803026370704174, + 0.008854001760482788, + 0.006015345919877291, + 0.006733345799148083, + 0.0030161633621901274, + 0.004880208987742662, + 0.003298981813713908, + 0.004113299306482077, + 0.006322025787085295, + 0.0025681359693408012, + 0.004604100715368986, + 0.0027721847873181105, + 0.0023742124903947115, + 0.0021868203766644, + 0.0017640432342886925, + 0.004540642723441124, + 0.0051236990839242935, + 0.004449749831110239, + 0.006105905398726463, + 0.00609102426096797, + 0.0031213199254125357, + 0.0021825688891112804, + 0.0029049820732325315, + 0.0038834090810269117, + 0.006470057647675276, + 0.003736113663762808, + 0.004722995217889547, + 0.005030802451074123, + 0.005627013277262449, + 0.0024342271499335766, + 0.003500442486256361, + 0.0029251466039568186, + 0.007675426080822945, + 0.007576070260256529, + 0.007001189049333334, + 0.006218451540917158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.018186485394835472 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004101437050849199 + ], + "opposite_logprob_tokens": [ + -5.504101276397705 + ], + "logprob_reward": -0.004101437050849199, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004101437050849199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1983335316181183, + 0.052447784692049026, + 0.05109955370426178, + 0.043370090425014496, + 0.3975078761577606, + 0.0818156898021698, + 0.060957297682762146, + 0.025106552988290787, + 0.03175778687000275, + 0.05760372430086136 + ], + "credit_map": [ + 0.016708776354789734, + 0.019856665283441544, + 0.019571825861930847, + 0.030074313282966614, + 0.024736162275075912, + 0.02461155131459236, + 0.03905045613646507, + 0.02372380904853344, + 0.0037670983001589775, + 0.0058554112911224365, + 0.0064057717099785805, + 0.006083578336983919, + 0.007546000182628632, + 0.00709396181628108, + 0.010305476374924183, + 0.0053904918022453785, + 0.003732302226126194, + 0.0049128043465316296, + 0.005108306184411049, + 0.007021142635494471, + 0.008107233792543411, + 0.007409696467220783, + 0.006237422581762075, + 0.008570652455091476, + 0.004669023212045431, + 0.0037961204070597887, + 0.006272135302424431, + 0.006010363809764385, + 0.0041595627553761005, + 0.00809948705136776, + 0.0050896285101771355, + 0.005273775197565556, + 0.007713469676673412, + 0.011860454455018044, + 0.006013517268002033, + 0.014438066631555557, + 0.007175614591687918, + 0.17017437517642975, + 0.008580807596445084, + 0.17155160009860992, + 0.007398999761790037, + 0.007604952901601791, + 0.01198042742908001, + 0.015913918614387512, + 0.007721155881881714, + 0.011483640410006046, + 0.007182308007031679, + 0.012530297040939331, + 0.006008592434227467, + 0.010831598192453384, + 0.0049668424762785435, + 0.009042835794389248, + 0.008556708693504333, + 0.005498841404914856, + 0.008111025206744671, + 0.007940863259136677, + 0.0023285383358597755, + 0.0023748513776808977, + 0.002284885849803686, + 0.0034421016462147236, + 0.004267967771738768, + 0.004296682309359312, + 0.003460423555225134, + 0.0026511054020375013, + 0.0028211823664605618, + 0.00262548984028399, + 0.002140452153980732, + 0.005076697561889887, + 0.005372789688408375, + 0.002780264476314187, + 0.0051925042644143105, + 0.005748409312218428, + 0.007037489674985409, + 0.004181069787591696, + 0.008655976504087448, + 0.005919359624385834, + 0.00800081342458725, + 0.008073575794696808, + 0.008014766499400139, + 0.007720677647739649 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.004101437050849199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly low angle and sideways, steadily follows an agitated figure as they pace relentlessly back and forth within the narrow confines of a dimly lit nighttime alley. The figure has messy short hair plastered to a pale, sweaty forehead and wears a simple dark wool coat over an unbuttoned gray T-shirt. As they take three steps on the damp cobblestones, their bodies tense, their shoulders hunched, suddenly rotating, and then repeating the motion. Their frantic gaze darts across the grimy brick wall before darting to their right hand, which clutches a gleaming gold coin and shakes uncontrollably. Their knuckles white with pressure, their thumbs obsessively rubbing the intricate, worn insignia on the coin's surface. A flickering street lamp overhead casts long, dancing shadows, creating a stark contrast of light and dark, accentuating the sharp planes of their faces and the metallic sheen of the coin. The overall color palette is cold and desaturated, dominated by the dark, damp tones of the alley, with the only warmth coming from sickly yellow light and brilliant gold, amplifying the scene's palpable atmosphere of intense anxiety and desperate indecision.", + "seed": 7591, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_5642_seed_7591_dp_24_rank_48_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0228257437935099, + "vlm_logprob": -0.0228257437935099, + "vlm_yes_no_margin": 4.249999952316284, + "vlm_token_credit": -0.0228257437935099, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.021378668025135994 + ], + "opposite_logprob_tokens": [ + -4.021378517150879 + ], + "logprob_reward": -0.021378668025135994, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.021378668025135994, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17276982963085175, + 0.07512922585010529, + 0.05516752600669861, + 0.040551818907260895, + 0.08268918097019196, + 0.0816834419965744, + 0.08692632615566254, + 0.12363287061452866, + 0.04719329625368118, + 0.23425650596618652 + ], + "credit_map": [ + 0.04142830893397331, + 0.01610749401152134, + 0.021273178979754448, + 0.020814714953303337, + 0.013980118557810783, + 0.02095329388976097, + 0.017695680260658264, + 0.02051704004406929, + 0.0056532323360443115, + 0.004187943879514933, + 0.008595073595643044, + 0.008214178495109081, + 0.00703840795904398, + 0.01588793843984604, + 0.008123235777020454, + 0.01742921769618988, + 0.006877496372908354, + 0.0027507836930453777, + 0.005285298451781273, + 0.008412033319473267, + 0.004422660917043686, + 0.011791834607720375, + 0.00658691581338644, + 0.009040501900017262, + 0.0033021620474755764, + 0.003870552871376276, + 0.005669502075761557, + 0.00510328309610486, + 0.006573760882019997, + 0.005649044178426266, + 0.004737451206892729, + 0.005646060686558485, + 0.003082424169406295, + 0.00708751380443573, + 0.004423707723617554, + 0.009525690227746964, + 0.02244332805275917, + 0.01592635177075863, + 0.011382414028048515, + 0.008817750960588455, + 0.006257874425500631, + 0.008362052962183952, + 0.011494236998260021, + 0.011677002534270287, + 0.01448978390544653, + 0.009050311520695686, + 0.010442285798490047, + 0.009909896180033684, + 0.005357825197279453, + 0.0095573291182518, + 0.0090866694226861, + 0.011498057283461094, + 0.015271198935806751, + 0.009272979572415352, + 0.016524722799658775, + 0.010357537306845188, + 0.01167806051671505, + 0.017301227897405624, + 0.009907213039696217, + 0.019039899110794067, + 0.022800054401159286, + 0.012189283035695553, + 0.018809830769896507, + 0.011907306499779224, + 0.003666613018140197, + 0.003175122430548072, + 0.006942322477698326, + 0.007908972911536694, + 0.006075352896004915, + 0.0075461603701114655, + 0.006675648037344217, + 0.005203107371926308, + 0.011014887131750584, + 0.013003521598875523, + 0.0821436271071434, + 0.09318512678146362, + 0.008265957236289978, + 0.011497276835143566, + 0.009499708190560341, + 0.005646406207233667 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 7 + ], + "selected_reward": -0.021378668025135994 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023305660113692284 + ], + "opposite_logprob_tokens": [ + -3.773305654525757 + ], + "logprob_reward": -0.023305660113692284, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023305660113692284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0806054100394249, + 0.06724374741315842, + 0.2813895046710968, + 0.037132613360881805, + 0.08773232251405716, + 0.04024815931916237, + 0.028301771730184555, + 0.04942147061228752, + 0.06047965586185455, + 0.2674453854560852 + ], + "credit_map": [ + 0.008077672682702541, + 0.0077065229415893555, + 0.006378798745572567, + 0.009086701087653637, + 0.012399650178849697, + 0.014166521839797497, + 0.011516191996634007, + 0.011273347772657871, + 0.006138595286756754, + 0.004165360238403082, + 0.009248794056475163, + 0.007378916721791029, + 0.00670261075720191, + 0.013901880010962486, + 0.007882985286414623, + 0.011824601329863071, + 0.023218689486384392, + 0.03233250975608826, + 0.025219742208719254, + 0.04981929436326027, + 0.03602277487516403, + 0.0416826531291008, + 0.03160391375422478, + 0.04148992523550987, + 0.0022432170808315277, + 0.003754645586013794, + 0.0029033152386546135, + 0.0060967509634792805, + 0.007064747624099255, + 0.004632504191249609, + 0.005414893850684166, + 0.005022536031901836, + 0.0032204941380769014, + 0.007974151521921158, + 0.006053206045180559, + 0.014052871614694595, + 0.016574891284108162, + 0.014971667900681496, + 0.015876565128564835, + 0.00900847278535366, + 0.002430807100608945, + 0.0035870647989213467, + 0.0032794352155178785, + 0.00497814454138279, + 0.009799112565815449, + 0.005535626318305731, + 0.0060263448394834995, + 0.004611624870449305, + 0.0028495436999946833, + 0.003824521554633975, + 0.005024294834583998, + 0.005316664464771748, + 0.0032210666686296463, + 0.0022881648037582636, + 0.003627511439844966, + 0.002150001237168908, + 0.004779148381203413, + 0.007972035557031631, + 0.004762666299939156, + 0.008603454567492008, + 0.006075837649405003, + 0.005338041111826897, + 0.0055930763483047485, + 0.006297206040471792, + 0.0038791107945144176, + 0.004059819970279932, + 0.008699862286448479, + 0.011219432577490807, + 0.008537408895790577, + 0.010825392790138721, + 0.007771026808768511, + 0.005487600341439247, + 0.013253501616418362, + 0.016593659296631813, + 0.07956314086914062, + 0.09565186500549316, + 0.014733782969415188, + 0.01695585623383522, + 0.019836457446217537, + 0.010857103392481804 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 9, + 4 + ], + "selected_reward": -0.023305660113692284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.054966144263744354 + ], + "opposite_logprob_tokens": [ + -2.9299662113189697 + ], + "logprob_reward": -0.054966144263744354, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.054966144263744354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.040912285447120667, + 0.0486639030277729, + 0.06305187195539474, + 0.0749075710773468, + 0.1705860197544098, + 0.06079027056694031, + 0.08665084838867188, + 0.0884527638554573, + 0.1955154985189438, + 0.1704690307378769 + ], + "credit_map": [ + 0.004054591991007328, + 0.005619218107312918, + 0.00484816636890173, + 0.005354320164769888, + 0.00499323196709156, + 0.00516694039106369, + 0.006477002985775471, + 0.004398808348923922, + 0.006886996328830719, + 0.0037730184849351645, + 0.006819835398346186, + 0.007324183825403452, + 0.00370943546295166, + 0.01031205989420414, + 0.0047881631180644035, + 0.0050502042286098, + 0.007154625840485096, + 0.0055016339756548405, + 0.006091184448450804, + 0.009217333048582077, + 0.008238282985985279, + 0.0108015863224864, + 0.00909255351871252, + 0.0069546690210700035, + 0.004175219219177961, + 0.00886776763945818, + 0.0053501734510064125, + 0.013610501773655415, + 0.012214659713208675, + 0.015336912125349045, + 0.00899050384759903, + 0.006361827719956636, + 0.0062147448770701885, + 0.013949096202850342, + 0.011235660873353481, + 0.027125779539346695, + 0.03532155975699425, + 0.02802126109600067, + 0.03284715488553047, + 0.015870749950408936, + 0.005136109888553619, + 0.005069604143500328, + 0.007154275197535753, + 0.006515339016914368, + 0.010561477392911911, + 0.007470917422324419, + 0.010280395857989788, + 0.008602144196629524, + 0.007142957299947739, + 0.009125233627855778, + 0.01072641834616661, + 0.01214123610407114, + 0.013434897176921368, + 0.010032408870756626, + 0.012687599286437035, + 0.011360094882547855, + 0.006908285431563854, + 0.00946167204529047, + 0.006444631144404411, + 0.013511200435459614, + 0.013469903729856014, + 0.009715531952679157, + 0.019075801596045494, + 0.009865731932222843, + 0.012123535387217999, + 0.007522300351411104, + 0.05361226201057434, + 0.050627466291189194, + 0.01963622123003006, + 0.017987100407481194, + 0.019219057634472847, + 0.014787556603550911, + 0.009042017161846161, + 0.011964251287281513, + 0.04472450911998749, + 0.053307585418224335, + 0.01463803555816412, + 0.013171481899917126, + 0.012478659860789776, + 0.011142461560666561 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 4, + 9 + ], + "selected_reward": -0.054966144263744354 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012535235844552517 + ], + "opposite_logprob_tokens": [ + -4.387535095214844 + ], + "logprob_reward": -0.012535235844552517, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012535235844552517, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13364772498607635, + 0.08879604935646057, + 0.14671236276626587, + 0.06582733243703842, + 0.08070630580186844, + 0.08354949951171875, + 0.06171538308262825, + 0.06137049198150635, + 0.1213880106806755, + 0.15628686547279358 + ], + "credit_map": [ + 0.016703233122825623, + 0.012158102355897427, + 0.014011799357831478, + 0.015807408839464188, + 0.016112715005874634, + 0.021666890010237694, + 0.01421471405774355, + 0.022972844541072845, + 0.006503494922071695, + 0.005873971618711948, + 0.010590393096208572, + 0.009119691327214241, + 0.00847341027110815, + 0.01780972070991993, + 0.00945753138512373, + 0.020967818796634674, + 0.010888426564633846, + 0.016983801499009132, + 0.01281149871647358, + 0.02719329670071602, + 0.021208081394433975, + 0.022166507318615913, + 0.020704569295048714, + 0.014756173826754093, + 0.005223162937909365, + 0.008829394355416298, + 0.008212058804929256, + 0.009392466396093369, + 0.009627437219023705, + 0.008613130077719688, + 0.009192217141389847, + 0.006737460847944021, + 0.0036559116560965776, + 0.009704875759780407, + 0.0061733778566122055, + 0.007595772854983807, + 0.017116889357566833, + 0.012818352319300175, + 0.01558671984821558, + 0.008054403588175774, + 0.003742507193237543, + 0.006914746016263962, + 0.006884890142828226, + 0.011202273890376091, + 0.01971234194934368, + 0.01015044841915369, + 0.010677829384803772, + 0.014264457859098911, + 0.005691655911505222, + 0.008966516703367233, + 0.0064978874288499355, + 0.013631322421133518, + 0.007945786230266094, + 0.006869122385978699, + 0.0070836544036865234, + 0.005029434338212013, + 0.007260926067829132, + 0.00644714618101716, + 0.005473727826029062, + 0.00901411660015583, + 0.008116092532873154, + 0.008192130364477634, + 0.00813268031924963, + 0.008733664639294147, + 0.008285066112875938, + 0.007365596015006304, + 0.035650771111249924, + 0.03189726546406746, + 0.009985403157770634, + 0.00954420492053032, + 0.010039214044809341, + 0.00862047728151083, + 0.007433827500790358, + 0.008741374127566814, + 0.0509127676486969, + 0.05847460776567459, + 0.007036578841507435, + 0.008267858996987343, + 0.009668281301856041, + 0.005751558113843203 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 0 + ], + "selected_reward": -0.012535235844552517 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001943010720424354 + ], + "opposite_logprob_tokens": [ + -6.251943111419678 + ], + "logprob_reward": -0.001943010720424354, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001943010720424354, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12174461781978607, + 0.06995248794555664, + 0.0881742388010025, + 0.14982786774635315, + 0.05193061754107475, + 0.10126279294490814, + 0.06305801868438721, + 0.060432545840740204, + 0.0866989940404892, + 0.20691782236099243 + ], + "credit_map": [ + 0.019768817350268364, + 0.011327840387821198, + 0.01267937570810318, + 0.014702511951327324, + 0.017099132761359215, + 0.020670507103204727, + 0.013007939793169498, + 0.012488475069403648, + 0.005073639564216137, + 0.0037886237259954214, + 0.007697834633290768, + 0.007296248804777861, + 0.00733644375577569, + 0.013410480692982674, + 0.008730880916118622, + 0.016618330031633377, + 0.007337284740060568, + 0.008772761560976505, + 0.008271101862192154, + 0.016099827364087105, + 0.008782506920397282, + 0.01604599505662918, + 0.010849971324205399, + 0.012014784850180149, + 0.015220404602587223, + 0.025554386898875237, + 0.016179686412215233, + 0.018448926508426666, + 0.017985602840781212, + 0.021928878501057625, + 0.017550738528370857, + 0.01695922762155533, + 0.0024709689896553755, + 0.008162234909832478, + 0.003805967280641198, + 0.006461146287620068, + 0.012208309024572372, + 0.007478006649762392, + 0.006921026855707169, + 0.004422951489686966, + 0.007502485997974873, + 0.007386056240648031, + 0.008163389749825, + 0.01180487684905529, + 0.01899128407239914, + 0.016913525760173798, + 0.016269708052277565, + 0.014231452718377113, + 0.006869475357234478, + 0.009716874919831753, + 0.009944647550582886, + 0.013685277663171291, + 0.006548355799168348, + 0.005601366050541401, + 0.0062787276692688465, + 0.004413289483636618, + 0.006211015861481428, + 0.00718541257083416, + 0.005952271167188883, + 0.008766346611082554, + 0.010405262932181358, + 0.006315560545772314, + 0.009169843047857285, + 0.006426826119422913, + 0.005778017919510603, + 0.005587870255112648, + 0.017196174710989, + 0.018483469262719154, + 0.01066701952368021, + 0.010165358893573284, + 0.010478905402123928, + 0.00834217295050621, + 0.010187949053943157, + 0.01095644012093544, + 0.07154978066682816, + 0.08124743402004242, + 0.0074567836709320545, + 0.010332639329135418, + 0.009817478246986866, + 0.0053693074733018875 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 3, + 0 + ], + "selected_reward": -0.001943010720424354 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly low angle and sideways, steadily follows an agitated figure as they pace relentlessly back and forth within the narrow confines of a dimly lit nighttime alley. The figure has messy short hair plastered to a pale, sweaty forehead and wears a simple dark wool coat over an unbuttoned gray T-shirt. As they take three steps on the damp cobblestones, their bodies tense, their shoulders hunched, suddenly rotating, and then repeating the motion. Their frantic gaze darts across the grimy brick wall before darting to their right hand, which clutches a gleaming gold coin and shakes uncontrollably. Their knuckles white with pressure, their thumbs obsessively rubbing the intricate, worn insignia on the coin's surface. A flickering street lamp overhead casts long, dancing shadows, creating a stark contrast of light and dark, accentuating the sharp planes of their faces and the metallic sheen of the coin. The overall color palette is cold and desaturated, dominated by the dark, damp tones of the alley, with the only warmth coming from sickly yellow light and brilliant gold, amplifying the scene's palpable atmosphere of intense anxiety and desperate indecision.", + "seed": 7592, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_5642_seed_7592_dp_25_rank_50_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.044318783283233645, + "vlm_logprob": -0.044318783283233645, + "vlm_yes_no_margin": 3.799999952316284, + "vlm_token_credit": -0.044318783283233645, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04226904362440109 + ], + "opposite_logprob_tokens": [ + -3.417268991470337 + ], + "logprob_reward": -0.04226904362440109, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.04226904362440109, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24263784289360046, + 0.09100303798913956, + 0.21253296732902527, + 0.08820835500955582, + 0.07378260791301727, + 0.0423811636865139, + 0.051632948219776154, + 0.04582168161869049, + 0.08051399886608124, + 0.07148543000221252 + ], + "credit_map": [ + 0.04060632735490799, + 0.04106550291180611, + 0.02575477585196495, + 0.030271785333752632, + 0.035348668694496155, + 0.02114192396402359, + 0.025277091190218925, + 0.023171760141849518, + 0.009982735849916935, + 0.01430647261440754, + 0.013473103754222393, + 0.016848670318722725, + 0.011521748267114162, + 0.007834970019757748, + 0.010087169706821442, + 0.006948165129870176, + 0.02645566314458847, + 0.03794877603650093, + 0.02387065440416336, + 0.030939752236008644, + 0.02872641198337078, + 0.02043122425675392, + 0.028025781735777855, + 0.01613469608128071, + 0.011324265040457249, + 0.015380706638097763, + 0.013919373974204063, + 0.017445668578147888, + 0.010890602134168148, + 0.004502926953136921, + 0.008964687585830688, + 0.005780123174190521, + 0.0064116730354726315, + 0.010177768766880035, + 0.00788862630724907, + 0.009539252147078514, + 0.013177447952330112, + 0.005787211004644632, + 0.012737245298922062, + 0.008063379675149918, + 0.0034602757077664137, + 0.0074539524503052235, + 0.0047485074028372765, + 0.0047585065476596355, + 0.004508094862103462, + 0.007143572438508272, + 0.004556530155241489, + 0.005751726683229208, + 0.004381358157843351, + 0.008339104242622852, + 0.006492051761597395, + 0.009562857449054718, + 0.00708056939765811, + 0.005586417391896248, + 0.004939805716276169, + 0.005250783637166023, + 0.004636617377400398, + 0.007655024062842131, + 0.006794450804591179, + 0.008515160530805588, + 0.0047488948330283165, + 0.00344103598035872, + 0.005423606373369694, + 0.004606893286108971, + 0.007849800400435925, + 0.01141154021024704, + 0.015619986690580845, + 0.012086101807653904, + 0.008366063237190247, + 0.012267795391380787, + 0.005863694939762354, + 0.007049011066555977, + 0.007441990543156862, + 0.008699030615389347, + 0.007099972106516361, + 0.009181946516036987, + 0.009257255122065544, + 0.01195061206817627, + 0.009206162765622139, + 0.008648460730910301 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.04226904362440109 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821153610944748 + ], + "opposite_logprob_tokens": [ + -4.018211364746094 + ], + "logprob_reward": -0.01821153610944748, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01821153610944748, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12111351639032364, + 0.07399536669254303, + 0.05953134596347809, + 0.07717491686344147, + 0.08458904176950455, + 0.0677589625120163, + 0.041294410824775696, + 0.14904223382472992, + 0.095980204641819, + 0.2295200377702713 + ], + "credit_map": [ + 0.010608972981572151, + 0.015619457699358463, + 0.01608387380838394, + 0.020823048427700996, + 0.010453667491674423, + 0.011772560887038708, + 0.01822296716272831, + 0.0175289548933506, + 0.006135707721114159, + 0.011513592675328255, + 0.01099316868931055, + 0.010144351981580257, + 0.01136403251439333, + 0.009094499051570892, + 0.007882065139710903, + 0.0068679493851959705, + 0.005786013323813677, + 0.008318568579852581, + 0.007399654481559992, + 0.007372416090220213, + 0.00920503307133913, + 0.006654776632785797, + 0.009672925807535648, + 0.005121954716742039, + 0.009058219380676746, + 0.006959307007491589, + 0.010984287597239017, + 0.015948262065649033, + 0.011938708834350109, + 0.00496828556060791, + 0.011802966706454754, + 0.00551487086340785, + 0.006447210442274809, + 0.010691244155168533, + 0.010289386846125126, + 0.013731028884649277, + 0.012113929726183414, + 0.006803981494158506, + 0.0156398955732584, + 0.008872359059751034, + 0.006123432889580727, + 0.009587102569639683, + 0.007208668161183596, + 0.008593626320362091, + 0.00613680062815547, + 0.010171429254114628, + 0.009807583875954151, + 0.01013030856847763, + 0.004392257425934076, + 0.004388394299894571, + 0.005776872858405113, + 0.007029505912214518, + 0.002967421431094408, + 0.005569050554186106, + 0.0051091136410832405, + 0.006061793304979801, + 0.013109206221997738, + 0.01366860419511795, + 0.019040875136852264, + 0.024099363014101982, + 0.020443039014935493, + 0.01472932193428278, + 0.016713259741663933, + 0.02723855897784233, + 0.004805755335837603, + 0.005529081914573908, + 0.012576773762702942, + 0.026300789788365364, + 0.007601480931043625, + 0.009791078045964241, + 0.006178933195769787, + 0.023196304216980934, + 0.017128106206655502, + 0.023655837401747704, + 0.024582093581557274, + 0.03461199253797531, + 0.02793898619711399, + 0.03331639617681503, + 0.02756373956799507, + 0.040722865611314774 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 0 + ], + "selected_reward": -0.01821153610944748 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1428213268518448 + ], + "opposite_logprob_tokens": [ + -2.0178213119506836 + ], + "logprob_reward": -0.1428213268518448, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.1428213268518448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10741729289293289, + 0.09508033841848373, + 0.05685315653681755, + 0.06971185654401779, + 0.10722678154706955, + 0.2878076136112213, + 0.08874779939651489, + 0.07137072086334229, + 0.06256775557994843, + 0.05321667343378067 + ], + "credit_map": [ + 0.00798714067786932, + 0.00961239542812109, + 0.013337691314518452, + 0.01570959948003292, + 0.012487093918025494, + 0.01207820139825344, + 0.013727166689932346, + 0.022478006780147552, + 0.007282909471541643, + 0.011475570499897003, + 0.013747664168477058, + 0.009929549880325794, + 0.017809884622693062, + 0.014344463124871254, + 0.011855743825435638, + 0.008634560741484165, + 0.005079838912934065, + 0.006424414925277233, + 0.007113033905625343, + 0.007031586486846209, + 0.010279522277414799, + 0.007503286935389042, + 0.008761066012084484, + 0.004660408943891525, + 0.008823499083518982, + 0.006862688343971968, + 0.010681087151169777, + 0.014438576996326447, + 0.006856206804513931, + 0.004429037682712078, + 0.012204689905047417, + 0.0054160719737410545, + 0.007734688464552164, + 0.016142748296260834, + 0.00969691202044487, + 0.011661643162369728, + 0.01931552216410637, + 0.01224406249821186, + 0.01909865066409111, + 0.011332545429468155, + 0.026294676586985588, + 0.03807031363248825, + 0.029477618634700775, + 0.035935208201408386, + 0.02079205960035324, + 0.03765582665801048, + 0.04766319319605827, + 0.05191871151328087, + 0.007870792411267757, + 0.013145322911441326, + 0.009518668055534363, + 0.011575142852962017, + 0.013103856705129147, + 0.008403938263654709, + 0.014878667891025543, + 0.010251405648887157, + 0.00775420106947422, + 0.011530669406056404, + 0.009969018399715424, + 0.015256800688803196, + 0.008750328794121742, + 0.0048499503172934055, + 0.00761995417997241, + 0.0056397998705506325, + 0.00269713276065886, + 0.004097919445484877, + 0.014918558299541473, + 0.02076469175517559, + 0.007337005343288183, + 0.0032185660675168037, + 0.0039214156568050385, + 0.005612468812614679, + 0.0030024286825209856, + 0.0038996038492769003, + 0.005297859199345112, + 0.007868167944252491, + 0.007895310409367085, + 0.009108473546802998, + 0.006170159671455622, + 0.009974666871130466 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.1428213268518448 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193184673786163 + ], + "opposite_logprob_tokens": [ + -4.264193058013916 + ], + "logprob_reward": -0.014193184673786163, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193184673786163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1163489818572998, + 0.0931919664144516, + 0.0862952470779419, + 0.07725834846496582, + 0.15321654081344604, + 0.2228982299566269, + 0.04721362516283989, + 0.06731504946947098, + 0.09111496061086655, + 0.04514702782034874 + ], + "credit_map": [ + 0.008703302592039108, + 0.011668589897453785, + 0.014443882741034031, + 0.018481258302927017, + 0.009414225816726685, + 0.010027970187366009, + 0.018465092405676842, + 0.025144658982753754, + 0.008335036225616932, + 0.016246279701590538, + 0.012166357599198818, + 0.015040900558233261, + 0.013103400357067585, + 0.011447342112660408, + 0.009614442475140095, + 0.0072382059879601, + 0.012330555357038975, + 0.014427674934267998, + 0.008624055422842503, + 0.010494734160602093, + 0.011959771625697613, + 0.007220174185931683, + 0.012372992001473904, + 0.008865293115377426, + 0.007843942381441593, + 0.007518639788031578, + 0.009745161049067974, + 0.011324098333716393, + 0.010500856675207615, + 0.0052221426740288734, + 0.01575484871864319, + 0.009348656982183456, + 0.011859826743602753, + 0.01613391563296318, + 0.01579836942255497, + 0.024876968935132027, + 0.029659796506166458, + 0.01187048852443695, + 0.03018219769001007, + 0.012834982946515083, + 0.01966283656656742, + 0.0283389650285244, + 0.023717792704701424, + 0.0279778391122818, + 0.01635902374982834, + 0.030007055029273033, + 0.03729238733649254, + 0.03954235091805458, + 0.004312316421419382, + 0.005992045626044273, + 0.005537895951420069, + 0.006513986270874739, + 0.005457789171487093, + 0.005843448918312788, + 0.006748117506504059, + 0.006808028090745211, + 0.006102477200329304, + 0.008913835510611534, + 0.008627444505691528, + 0.012210133485496044, + 0.009900964796543121, + 0.006666109897196293, + 0.008146381936967373, + 0.006747703533619642, + 0.004499961622059345, + 0.006394980940967798, + 0.01687442511320114, + 0.021141931414604187, + 0.010263226926326752, + 0.00787096843123436, + 0.010503713972866535, + 0.01356575544923544, + 0.002988581545650959, + 0.004703182727098465, + 0.004795058164745569, + 0.0067516230046749115, + 0.005509325303137302, + 0.006846051663160324, + 0.005050006788223982, + 0.008503198623657227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 0 + ], + "selected_reward": -0.014193184673786163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409882515668869 + ], + "opposite_logprob_tokens": [ + -5.504098892211914 + ], + "logprob_reward": -0.00409882515668869, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409882515668869, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1217629611492157, + 0.16425718367099762, + 0.1716286689043045, + 0.07340316474437714, + 0.10580526292324066, + 0.07290458679199219, + 0.061236415058374405, + 0.06284479796886444, + 0.09442205727100372, + 0.07173489779233932 + ], + "credit_map": [ + 0.014658876694738865, + 0.01580134779214859, + 0.014131960459053516, + 0.013715043663978577, + 0.01880105771124363, + 0.014928249642252922, + 0.014563988894224167, + 0.015162433497607708, + 0.018779462203383446, + 0.030155126005411148, + 0.02738254703581333, + 0.03467536345124245, + 0.01591527834534645, + 0.00927732978016138, + 0.016239074990153313, + 0.011833005584776402, + 0.016809683293104172, + 0.027504412457346916, + 0.01902453973889351, + 0.028439410030841827, + 0.02967114932835102, + 0.02020481787621975, + 0.016070617362856865, + 0.01390402764081955, + 0.00880417600274086, + 0.011210315860807896, + 0.011726939119398594, + 0.01333554182201624, + 0.006784636992961168, + 0.0055805533193051815, + 0.01016529742628336, + 0.0057957107201218605, + 0.008244884200394154, + 0.013757270760834217, + 0.010544153861701488, + 0.012397190555930138, + 0.019724316895008087, + 0.011398819275200367, + 0.020205233246088028, + 0.009533397853374481, + 0.005932365544140339, + 0.011193862184882164, + 0.007336628157645464, + 0.009280784986913204, + 0.006405370309948921, + 0.011641990393400192, + 0.0081477714702487, + 0.012965814210474491, + 0.005519377999007702, + 0.005459228064864874, + 0.008449077606201172, + 0.011111305095255375, + 0.005207535810768604, + 0.009037217125296593, + 0.007848166860640049, + 0.0086045041680336, + 0.006380132865160704, + 0.00849216803908348, + 0.008255847729742527, + 0.012419858016073704, + 0.008238377049565315, + 0.004528869409114122, + 0.008769831620156765, + 0.00575971370562911, + 0.008705956861376762, + 0.013533016666769981, + 0.023511331528425217, + 0.013744055293500423, + 0.012673291377723217, + 0.00709620350971818, + 0.007288945373147726, + 0.00786925945430994, + 0.004726334009319544, + 0.005888388957828283, + 0.009078488685190678, + 0.011995748616755009, + 0.013060146942734718, + 0.009515082463622093, + 0.008302170783281326, + 0.009168539196252823 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.00409882515668869 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium shot, shot from a slightly low angle and sideways, steadily follows an agitated figure as they pace relentlessly back and forth within the narrow confines of a dimly lit nighttime alley. The figure has messy short hair plastered to a pale, sweaty forehead and wears a simple dark wool coat over an unbuttoned gray T-shirt. As they take three steps on the damp cobblestones, their bodies tense, their shoulders hunched, suddenly rotating, and then repeating the motion. Their frantic gaze darts across the grimy brick wall before darting to their right hand, which clutches a gleaming gold coin and shakes uncontrollably. Their knuckles white with pressure, their thumbs obsessively rubbing the intricate, worn insignia on the coin's surface. A flickering street lamp overhead casts long, dancing shadows, creating a stark contrast of light and dark, accentuating the sharp planes of their faces and the metallic sheen of the coin. The overall color palette is cold and desaturated, dominated by the dark, damp tones of the alley, with the only warmth coming from sickly yellow light and brilliant gold, amplifying the scene's palpable atmosphere of intense anxiety and desperate indecision.", + "seed": 7593, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_12/index_5642_seed_7593_dp_25_rank_50_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 12, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.02485112939029932, + "vlm_logprob": -0.02485112939029932, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.02485112939029932, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.019266724586486816 + ], + "opposite_logprob_tokens": [ + -4.144266605377197 + ], + "logprob_reward": -0.019266724586486816, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.019266724586486816, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04503009095788002, + 0.061798058450222015, + 0.029105912894010544, + 0.0939955934882164, + 0.08046597987413406, + 0.17929379642009735, + 0.0769386738538742, + 0.034058138728141785, + 0.30696311593055725, + 0.09235064685344696 + ], + "credit_map": [ + 0.0037657725624740124, + 0.005647454410791397, + 0.005850557703524828, + 0.005787416361272335, + 0.007012267131358385, + 0.003490869188681245, + 0.00706882681697607, + 0.006406924221664667, + 0.004566418472677469, + 0.014032822102308273, + 0.005446544382721186, + 0.00897024292498827, + 0.007028715685009956, + 0.005460888613015413, + 0.007108781486749649, + 0.009183639660477638, + 0.0031170176807790995, + 0.0048770359717309475, + 0.0019385715713724494, + 0.004103147890418768, + 0.0029942465480417013, + 0.005376058630645275, + 0.0032280990853905678, + 0.0034717342350631952, + 0.005692643579095602, + 0.01715695671737194, + 0.007850193418562412, + 0.013528300449252129, + 0.0100798849016428, + 0.011698191985487938, + 0.01932903565466404, + 0.00866038165986538, + 0.00511540099978447, + 0.01169936265796423, + 0.007067116443067789, + 0.010442987084388733, + 0.009356481023132801, + 0.01381008792668581, + 0.012996782548725605, + 0.009977759793400764, + 0.019353652372956276, + 0.02227064222097397, + 0.02513415925204754, + 0.016857856884598732, + 0.02068413980305195, + 0.024991510435938835, + 0.027336500585079193, + 0.022665323689579964, + 0.007501351181417704, + 0.011774725280702114, + 0.00866189505904913, + 0.007502841297537088, + 0.009348083287477493, + 0.011423470452427864, + 0.00873604416847229, + 0.01199025847017765, + 0.003946707118302584, + 0.0050721485167741776, + 0.006499566603451967, + 0.0049301935359835625, + 0.0032674872782081366, + 0.0033362729009240866, + 0.004330586642026901, + 0.0026751745026558638, + 0.02584240213036537, + 0.023333067074418068, + 0.02874651923775673, + 0.047968845814466476, + 0.04005435109138489, + 0.04667620360851288, + 0.046791281551122665, + 0.04755042865872383, + 0.007907135412096977, + 0.006010373588651419, + 0.010153569281101227, + 0.006039471831172705, + 0.016966061666607857, + 0.015358341857790947, + 0.017296582460403442, + 0.012619106099009514 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 5, + 3 + ], + "selected_reward": -0.019266724586486816 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04309776797890663 + ], + "opposite_logprob_tokens": [ + -3.168097734451294 + ], + "logprob_reward": -0.04309776797890663, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04309776797890663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07260989397764206, + 0.1004374548792839, + 0.09635891020298004, + 0.08561986684799194, + 0.19347341358661652, + 0.06190474331378937, + 0.12725308537483215, + 0.04887247085571289, + 0.15544700622558594, + 0.05802313983440399 + ], + "credit_map": [ + 0.009327619336545467, + 0.009705057367682457, + 0.009344390593469143, + 0.009554184973239899, + 0.007823626510798931, + 0.00643144641071558, + 0.010609576478600502, + 0.009813995100557804, + 0.009194951504468918, + 0.02721862681210041, + 0.010903297923505306, + 0.01564077101647854, + 0.010327083989977837, + 0.007560397498309612, + 0.013490662910044193, + 0.0061016688123345375, + 0.006920539308339357, + 0.01884915493428707, + 0.008186587132513523, + 0.017104381695389748, + 0.007204785477370024, + 0.01388489082455635, + 0.013546377420425415, + 0.010662190616130829, + 0.0055702016688883305, + 0.013970181345939636, + 0.0051931049674749374, + 0.011555285193026066, + 0.007699871435761452, + 0.014096111990511417, + 0.016207218170166016, + 0.01132789347320795, + 0.008030897006392479, + 0.03559516370296478, + 0.009020432829856873, + 0.03161133453249931, + 0.025626778602600098, + 0.016806965693831444, + 0.04402631148695946, + 0.022755518555641174, + 0.005508051253855228, + 0.00675062695518136, + 0.006968130823224783, + 0.0077120838686823845, + 0.007940088398754597, + 0.008774089626967907, + 0.00789982732385397, + 0.01035184133797884, + 0.012729918584227562, + 0.021590624004602432, + 0.010377967730164528, + 0.01092509925365448, + 0.017221558839082718, + 0.024868391454219818, + 0.011659475974738598, + 0.017880044877529144, + 0.006704255472868681, + 0.006436663679778576, + 0.007414231542497873, + 0.0077284169383347034, + 0.006640868727117777, + 0.0026913487818092108, + 0.007226034067571163, + 0.0040306514129042625, + 0.013708771206438541, + 0.01214736606925726, + 0.02288375049829483, + 0.01928623951971531, + 0.014942818321287632, + 0.0341130830347538, + 0.020344721153378487, + 0.01802026852965355, + 0.004706729203462601, + 0.003918853588402271, + 0.006424444727599621, + 0.005497746169567108, + 0.010205687023699284, + 0.008813584223389626, + 0.012379354797303677, + 0.0060767377726733685 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 6 + ], + "selected_reward": -0.04309776797890663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.038116924464702606 + ], + "opposite_logprob_tokens": [ + -3.288116931915283 + ], + "logprob_reward": -0.038116924464702606, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.038116924464702606, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05764634162187576, + 0.031609151512384415, + 0.03741876780986786, + 0.05814224109053612, + 0.1073388159275055, + 0.060233622789382935, + 0.1787455677986145, + 0.28169140219688416, + 0.05641871318221092, + 0.13075532019138336 + ], + "credit_map": [ + 0.0066490513272583485, + 0.014628924429416656, + 0.005977678578346968, + 0.00910860300064087, + 0.007204713299870491, + 0.0040396424010396, + 0.0049654035829007626, + 0.005072331987321377, + 0.003290321910753846, + 0.006034186575561762, + 0.0022157125640660524, + 0.00431513786315918, + 0.003267649095505476, + 0.004206547979265451, + 0.0043971724808216095, + 0.003882426768541336, + 0.0038255497347563505, + 0.006573152262717485, + 0.0033968754578381777, + 0.006224229000508785, + 0.002856443403288722, + 0.004455443471670151, + 0.004729441832751036, + 0.005357632413506508, + 0.004329240880906582, + 0.005577564239501953, + 0.005622366908937693, + 0.008834325708448887, + 0.004225729499012232, + 0.011600403115153313, + 0.01112739834934473, + 0.0068252189084887505, + 0.006730279885232449, + 0.01930033601820469, + 0.007212477270513773, + 0.016759315505623817, + 0.014200752601027489, + 0.01167986448854208, + 0.01717834360897541, + 0.014277460984885693, + 0.004408657550811768, + 0.007736128754913807, + 0.004431677516549826, + 0.00628945417702198, + 0.009620037861168385, + 0.009652643464505672, + 0.009231055155396461, + 0.008863978087902069, + 0.01645728573203087, + 0.02805164083838463, + 0.014444581232964993, + 0.01773880608379841, + 0.02168542705476284, + 0.03679604455828667, + 0.017134109511971474, + 0.02643769420683384, + 0.04407272860407829, + 0.05263174697756767, + 0.05307954177260399, + 0.04184652492403984, + 0.022249920293688774, + 0.019237125292420387, + 0.031949155032634735, + 0.016624681651592255, + 0.004467886406928301, + 0.003983920905739069, + 0.007595760282129049, + 0.008855905383825302, + 0.006956230849027634, + 0.004704830702394247, + 0.010547852143645287, + 0.009306333027780056, + 0.014122464694082737, + 0.010782268829643726, + 0.014408664777874947, + 0.010909332893788815, + 0.03306282311677933, + 0.0179936233907938, + 0.017244823276996613, + 0.012231336906552315 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 6, + 9 + ], + "selected_reward": -0.038116924464702606 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020578453317284584 + ], + "opposite_logprob_tokens": [ + -3.895578384399414 + ], + "logprob_reward": -0.020578453317284584, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020578453317284584, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06055385246872902, + 0.06437233835458755, + 0.11359977722167969, + 0.08083454519510269, + 0.07665068656206131, + 0.11120052635669708, + 0.16126103699207306, + 0.12084045261144638, + 0.11642489582300186, + 0.09426193684339523 + ], + "credit_map": [ + 0.0058687347918748856, + 0.009690308943390846, + 0.006955188699066639, + 0.0070563266053795815, + 0.01087070070207119, + 0.005325585138052702, + 0.007338977884501219, + 0.007448028307408094, + 0.005767186172306538, + 0.010041088797152042, + 0.007994302548468113, + 0.010351560078561306, + 0.0067418478429317474, + 0.006657688412815332, + 0.010557339526712894, + 0.006261325906962156, + 0.015549954026937485, + 0.021983005106449127, + 0.012205752544105053, + 0.017409127205610275, + 0.008218147791922092, + 0.013116385787725449, + 0.01474686712026596, + 0.010370536707341671, + 0.0043130237609148026, + 0.011916710063815117, + 0.004975567571818829, + 0.011682446114718914, + 0.006918124854564667, + 0.01709667406976223, + 0.012963375076651573, + 0.01096862182021141, + 0.005349182058125734, + 0.012481038458645344, + 0.0074998438358306885, + 0.011979150585830212, + 0.008275269530713558, + 0.009151759557425976, + 0.010750760324299335, + 0.01116368267685175, + 0.008874310180544853, + 0.011252284049987793, + 0.01205473393201828, + 0.015008017420768738, + 0.019659094512462616, + 0.008825558237731457, + 0.018912503495812416, + 0.016614019870758057, + 0.015224236994981766, + 0.02575812116265297, + 0.01364039070904255, + 0.014035798609256744, + 0.021622750908136368, + 0.028670743107795715, + 0.01823379285633564, + 0.024075202643871307, + 0.02071407623589039, + 0.022853564471006393, + 0.02331126108765602, + 0.01821020245552063, + 0.008090275339782238, + 0.009439592249691486, + 0.011868569068610668, + 0.00635290564969182, + 0.010617032647132874, + 0.00949157029390335, + 0.010414152406156063, + 0.019110901281237602, + 0.015242356806993484, + 0.01217155996710062, + 0.017187748104333878, + 0.022189568728208542, + 0.005447670817375183, + 0.006389289628714323, + 0.008647724986076355, + 0.008607246913015842, + 0.019898584112524986, + 0.013416304253041744, + 0.02058713324368, + 0.011267977766692638 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 8 + ], + "selected_reward": -0.020578453317284584 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set in a dimly lit nighttime alley with damp cobblestones and grimy brick walls?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the figure wearing a dark wool coat over an unbuttoned gray T-shirt with messy short hair plastered to a pale, sweaty forehead?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the figure clutching a gleaming gold coin and shaking it uncontrollably while rubbing its worn insignia?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera maintain a medium shot from a slightly low angle and sideways position, steadily following the figure?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the lighting dominated by a flickering street lamp casting long shadows, with a cold, desaturated palette featuring sickly yellow and brilliant gold highlights?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003195776604115963 + ], + "opposite_logprob_tokens": [ + -5.753195762634277 + ], + "logprob_reward": -0.003195776604115963, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003195776604115963, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08603546023368835, + 0.07218541949987411, + 0.07805860042572021, + 0.17434139549732208, + 0.06760551035404205, + 0.06458991765975952, + 0.19321857392787933, + 0.09976793080568314, + 0.05914118140935898, + 0.10505596548318863 + ], + "credit_map": [ + 0.009524905122816563, + 0.015780119225382805, + 0.009188471361994743, + 0.011590396985411644, + 0.011326988227665424, + 0.00629004929214716, + 0.011388114653527737, + 0.010946430265903473, + 0.005369876977056265, + 0.020244069397449493, + 0.004621326923370361, + 0.01297434139996767, + 0.0059738666750490665, + 0.008633822202682495, + 0.008917160332202911, + 0.005450970493257046, + 0.005109114106744528, + 0.013384430669248104, + 0.0065809329971671104, + 0.012869239784777164, + 0.008718971163034439, + 0.013513730838894844, + 0.009405993856489658, + 0.008476199582219124, + 0.009418278932571411, + 0.026994718238711357, + 0.009338145144283772, + 0.02707645110785961, + 0.018193915486335754, + 0.02863471955060959, + 0.032980650663375854, + 0.021704530343413353, + 0.004240505397319794, + 0.007970262318849564, + 0.005698822438716888, + 0.009492600336670876, + 0.008713352493941784, + 0.009221158921718597, + 0.014273461885750294, + 0.007995354011654854, + 0.005328062921762466, + 0.007612762041389942, + 0.007206104230135679, + 0.007777518127113581, + 0.010956103913486004, + 0.006057984661310911, + 0.010155819356441498, + 0.009495565667748451, + 0.01890549436211586, + 0.036392468959093094, + 0.019316071644425392, + 0.01491217315196991, + 0.02179558388888836, + 0.036581918597221375, + 0.016353536397218704, + 0.02896135114133358, + 0.01792147569358349, + 0.017328478395938873, + 0.019075024873018265, + 0.015677087008953094, + 0.007981211878359318, + 0.006694718264043331, + 0.00908276904374361, + 0.006007174029946327, + 0.0061005051247775555, + 0.006908352952450514, + 0.007784306537359953, + 0.008549311198294163, + 0.005200580228120089, + 0.008692026138305664, + 0.008084696717560291, + 0.007821410894393921, + 0.007997321896255016, + 0.00881759449839592, + 0.010443915612995625, + 0.011172764003276825, + 0.02378246933221817, + 0.014770066365599632, + 0.014162548817694187, + 0.013909297063946724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 9 + ], + "selected_reward": -0.003195776604115963 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_tensor([9683, 9683])_group_13_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_tensor([9683, 9683])_group_13_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9ee83a567d82bf4f29677f56249553b6e45bf74d --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_tensor([9683, 9683])_group_13_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, eye-level medium shot, rendered in a lush, hand-drawn, Ghibli-inspired art style, focuses on a clump of tall cattail plants at the center of the frame. Their fibrous green stems, with their soft, painterly texture, are topped with distinctive spikes of velvety flowers a rich chocolate brown. The entire scene is bathed in the warm, soft glow of the afternoon sun, illuminating the tranquil marsh setting. Crystal-clear water rises steadily and deliberately around the green stems, its calm surface like a mirror, reflecting the soft hues of pink, pale gold, and lavender against the sky. Tiny, shimmering dragonflies, their iridescent wings, dance and hover in the air between the reeds, their delicate movements adding a touch of life and magic. The combination of the impending tide, the warm, bright lighting, and the classic animation aesthetic creates an overwhelming atmosphere of tranquility and magic.", + "seed": 27120, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27120_dp_26_rank_52_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.06917876317165791, + "vlm_logprob": -0.06917876317165791, + "vlm_yes_no_margin": 3.874999976158142, + "vlm_token_credit": -0.06917876317165791, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.18371379375457764 + ], + "opposite_logprob_tokens": [ + -1.8087137937545776 + ], + "logprob_reward": -0.18371379375457764, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.18371379375457764, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1198958232998848, + 0.13712140917778015, + 0.100238136947155, + 0.14548376202583313, + 0.030011020600795746, + 0.06967141479253769, + 0.14408700168132782, + 0.026371702551841736, + 0.1401369571685791, + 0.08698278665542603 + ], + "credit_map": [ + 0.016413899138569832, + 0.008968157693743706, + 0.01845710538327694, + 0.03101409412920475, + 0.008106905035674572, + 0.011849567294120789, + 0.012655734084546566, + 0.012430357746779919, + 0.017023123800754547, + 0.008440669625997543, + 0.015009509399533272, + 0.02102273888885975, + 0.010517195798456669, + 0.011208257637917995, + 0.03823890537023544, + 0.015661006793379784, + 0.012436040677130222, + 0.007927081547677517, + 0.01418614573776722, + 0.017061656340956688, + 0.007240026257932186, + 0.012981359846889973, + 0.010928978212177753, + 0.017476845532655716, + 0.024197420105338097, + 0.007671893574297428, + 0.013636043295264244, + 0.013238389976322651, + 0.0071825492195785046, + 0.006162210833281279, + 0.0562436543405056, + 0.0171516016125679, + 0.00396130234003067, + 0.0025011971592903137, + 0.0052221170626580715, + 0.005830479320138693, + 0.00166160031221807, + 0.0031100932974368334, + 0.004015026614069939, + 0.0037092023994773626, + 0.009311477653682232, + 0.00426354119554162, + 0.02125570736825466, + 0.009355254471302032, + 0.0026769640389829874, + 0.004828518722206354, + 0.011829657480120659, + 0.0061502885073423386, + 0.017939738929271698, + 0.006468576844781637, + 0.045304324477910995, + 0.014461149461567402, + 0.006400591693818569, + 0.010014144703745842, + 0.03182448446750641, + 0.011673999018967152, + 0.004003830719739199, + 0.001402197522111237, + 0.007945092394948006, + 0.003699005115777254, + 0.0011017388897016644, + 0.002522665774449706, + 0.003486116649582982, + 0.0022110554855316877, + 0.01670985482633114, + 0.006511410232633352, + 0.04567454382777214, + 0.017874641343951225, + 0.0050942562520504, + 0.008581167086958885, + 0.021096564829349518, + 0.018594538792967796, + 0.013415194116532803, + 0.0034460341557860374, + 0.036719225347042084, + 0.007234652992337942, + 0.0024879479315131903, + 0.003136763582006097, + 0.013220788910984993, + 0.007322174031287432 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 6, + 8 + ], + "selected_reward": -0.18371379375457764 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002209009835496545 + ], + "opposite_logprob_tokens": [ + -6.127209186553955 + ], + "logprob_reward": -0.002209009835496545, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.002209009835496545, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16490060091018677, + 0.06454543769359589, + 0.06753379106521606, + 0.30295172333717346, + 0.03450585529208183, + 0.03149442374706268, + 0.09002076089382172, + 0.1299332082271576, + 0.08925849944353104, + 0.02485569752752781 + ], + "credit_map": [ + 0.01810123771429062, + 0.011028791777789593, + 0.0271275844424963, + 0.018640965223312378, + 0.013318028301000595, + 0.03668450191617012, + 0.017910871654748917, + 0.022088631987571716, + 0.004893438424915075, + 0.0050991009920835495, + 0.016819335520267487, + 0.005896834656596184, + 0.005610000807791948, + 0.0052845049649477005, + 0.014734411612153053, + 0.006207816768437624, + 0.006626517046242952, + 0.0059953187592327595, + 0.008562224917113781, + 0.01367797888815403, + 0.0056886677630245686, + 0.004884003661572933, + 0.014257155358791351, + 0.00784192606806755, + 0.045273154973983765, + 0.015316839329898357, + 0.03510137274861336, + 0.027448777109384537, + 0.01472235657274723, + 0.012215940281748772, + 0.11904572695493698, + 0.03382755443453789, + 0.004434036556631327, + 0.0018687539268285036, + 0.004028173629194498, + 0.005142515059560537, + 0.0017255882266908884, + 0.0035523499827831984, + 0.008374370634555817, + 0.0053800721652805805, + 0.0029283580370247364, + 0.0029433309100568295, + 0.006050468422472477, + 0.006317677441984415, + 0.0018624758813530207, + 0.0030030354391783476, + 0.005799727980047464, + 0.0025893505662679672, + 0.012576532550156116, + 0.00368884252384305, + 0.028763534501194954, + 0.010942711494863033, + 0.00337416329421103, + 0.005509178154170513, + 0.017396876588463783, + 0.007768930401653051, + 0.017946450039744377, + 0.005320627707988024, + 0.049408335238695145, + 0.010219354182481766, + 0.00448575709015131, + 0.0031992467120289803, + 0.030361216515302658, + 0.008992232382297516, + 0.011034208349883556, + 0.003907414153218269, + 0.031025629490613937, + 0.01029094960540533, + 0.003322417614981532, + 0.004483536351472139, + 0.014052164740860462, + 0.011142185889184475, + 0.0035915542393922806, + 0.0013830061070621014, + 0.009491312317550182, + 0.0022104857489466667, + 0.0009661215590313077, + 0.0012158234603703022, + 0.003921236377209425, + 0.0020761610940098763 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 7 + ], + "selected_reward": -0.002209009835496545 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012532763183116913 + ], + "opposite_logprob_tokens": [ + -4.387532711029053 + ], + "logprob_reward": -0.012532763183116913, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012532763183116913, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12786483764648438, + 0.09760651737451553, + 0.09720180183649063, + 0.4657444357872009, + 0.03255458176136017, + 0.009274246171116829, + 0.040588878095149994, + 0.011511627584695816, + 0.06086215749382973, + 0.056790951639413834 + ], + "credit_map": [ + 0.010930203832685947, + 0.009057826362550259, + 0.02385810762643814, + 0.02286878228187561, + 0.009095694869756699, + 0.014348061755299568, + 0.026329251006245613, + 0.011376901529729366, + 0.008194182999432087, + 0.005049103870987892, + 0.031924087554216385, + 0.0066364905796945095, + 0.005274306051433086, + 0.004721084143966436, + 0.028447484597563744, + 0.007359758485108614, + 0.015680644661188126, + 0.005864296108484268, + 0.010917537845671177, + 0.01276268158107996, + 0.005922201555222273, + 0.00736205093562603, + 0.03018302284181118, + 0.008509357459843159, + 0.0828881487250328, + 0.025221969932317734, + 0.04051731154322624, + 0.04331282898783684, + 0.02273380011320114, + 0.018810516223311424, + 0.1788642257452011, + 0.0533955842256546, + 0.003567424835637212, + 0.002004259964451194, + 0.006070213858038187, + 0.005399209912866354, + 0.0015335422940552235, + 0.003620811039581895, + 0.004828702658414841, + 0.005530411843210459, + 0.0008254878921434283, + 0.0009029690991155803, + 0.0016540432116016746, + 0.0015347690787166357, + 0.000676770112477243, + 0.0011494193458929658, + 0.001645512180402875, + 0.0008852735627442598, + 0.004471396096050739, + 0.002358170924708247, + 0.01084340550005436, + 0.003996092360466719, + 0.0022215722128748894, + 0.0041091530583798885, + 0.008215491659939289, + 0.00437359232455492, + 0.0019036943558603525, + 0.0012780696852132678, + 0.0016574274050071836, + 0.001904972130432725, + 0.0007444248185493052, + 0.0015301628736779094, + 0.0012870975770056248, + 0.0012057783314958215, + 0.0069291298277676105, + 0.0026555578224360943, + 0.018560590222477913, + 0.00784964207559824, + 0.0024883474688977003, + 0.004932919517159462, + 0.007975687272846699, + 0.009470274671912193, + 0.00782094057649374, + 0.002665960695594549, + 0.02188369631767273, + 0.005251592490822077, + 0.0017840700456872582, + 0.0031566936522722244, + 0.009490261785686016, + 0.004737728275358677 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.012532763183116913 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280448853969574 + ], + "opposite_logprob_tokens": [ + -2.0178043842315674 + ], + "logprob_reward": -0.14280448853969574, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280448853969574, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13317416608333588, + 0.13699977099895477, + 0.07319819927215576, + 0.2733355164527893, + 0.030013807117938995, + 0.03050476871430874, + 0.06931232661008835, + 0.16093501448631287, + 0.07619427889585495, + 0.016332093626260757 + ], + "credit_map": [ + 0.015535401180386543, + 0.009236728772521019, + 0.019291626289486885, + 0.01760556921362877, + 0.010037435218691826, + 0.02986805886030197, + 0.01267717219889164, + 0.018922166898846626, + 0.011644499376416206, + 0.008200494572520256, + 0.04431480914354324, + 0.009920845739543438, + 0.008372016251087189, + 0.007471680641174316, + 0.03672512248158455, + 0.010350299067795277, + 0.008134212344884872, + 0.005590024869889021, + 0.009032836183905602, + 0.011227704584598541, + 0.005746207665652037, + 0.0057359663769602776, + 0.01882973499596119, + 0.008901515044271946, + 0.04898323118686676, + 0.01470895390957594, + 0.023306911811232567, + 0.0254986472427845, + 0.013268661685287952, + 0.01124501321464777, + 0.10505147278308868, + 0.031272612512111664, + 0.003549972316250205, + 0.0018741354579105973, + 0.005865815095603466, + 0.004766115918755531, + 0.0012977662263438106, + 0.0024440616834908724, + 0.007112288847565651, + 0.0031036504078656435, + 0.00345097784884274, + 0.0022941422648727894, + 0.008141588419675827, + 0.0038639537524431944, + 0.001664973795413971, + 0.002596208592876792, + 0.005602770950645208, + 0.002890155417844653, + 0.009806702844798565, + 0.0026646703481674194, + 0.022586088627576828, + 0.007470839191228151, + 0.002608272712677717, + 0.004655907861888409, + 0.014034652151167393, + 0.005485194269567728, + 0.022548893466591835, + 0.006386613007634878, + 0.06336423754692078, + 0.012047206982970238, + 0.005072948522865772, + 0.003651681588962674, + 0.03671688213944435, + 0.011146563105285168, + 0.00889773853123188, + 0.0034615378826856613, + 0.024604177102446556, + 0.009420572780072689, + 0.0029579910915344954, + 0.004978645592927933, + 0.0112867821007967, + 0.01058683730661869, + 0.002324674278497696, + 0.0009869628120213747, + 0.005940030328929424, + 0.0015593082644045353, + 0.0009849218185991049, + 0.0011303031351417303, + 0.0018697138875722885, + 0.0015361799160018563 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 1 + ], + "selected_reward": -0.14280448853969574 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004633760545402765 + ], + "opposite_logprob_tokens": [ + -5.379633903503418 + ], + "logprob_reward": -0.004633760545402765, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004633760545402765, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.046548329293727875, + 0.04646081104874611, + 0.054976243525743484, + 0.3146189749240875, + 0.0493583083152771, + 0.02344398945569992, + 0.12482541054487228, + 0.18220128118991852, + 0.12928199768066406, + 0.028284728527069092 + ], + "credit_map": [ + 0.007535069715231657, + 0.003609235165640712, + 0.005921519361436367, + 0.007619922980666161, + 0.004136941861361265, + 0.006143270526081324, + 0.007414332125335932, + 0.004168037325143814, + 0.006267417222261429, + 0.0036758084315806627, + 0.012767053209245205, + 0.007178818807005882, + 0.003838466014713049, + 0.0037761463318020105, + 0.004702726844698191, + 0.00425437418743968, + 0.007996408268809319, + 0.0039284201338887215, + 0.00801671389490366, + 0.011072611436247826, + 0.0038556582294404507, + 0.005062888842076063, + 0.006815765518695116, + 0.008227778598666191, + 0.06963784247636795, + 0.016830939799547195, + 0.045630648732185364, + 0.029157627373933792, + 0.01343231275677681, + 0.011598058976233006, + 0.0951664075255394, + 0.03316512331366539, + 0.005400144029408693, + 0.0032396838068962097, + 0.009497815743088722, + 0.008657174184918404, + 0.0024441105779260397, + 0.004741983488202095, + 0.006756248418241739, + 0.008621150627732277, + 0.002659868448972702, + 0.0016813889378681779, + 0.004559221211820841, + 0.005926281679421663, + 0.0010802332544699311, + 0.0017238399013876915, + 0.0029786776285618544, + 0.0028344786260277033, + 0.01661306619644165, + 0.006093545816838741, + 0.03545233607292175, + 0.012535633519291878, + 0.006964571308344603, + 0.016125859692692757, + 0.016107652336359024, + 0.014932742342352867, + 0.02473282814025879, + 0.007722044829279184, + 0.06648777425289154, + 0.013904736377298832, + 0.006730638910084963, + 0.00830315612256527, + 0.04191789776086807, + 0.012402208521962166, + 0.016264235600829124, + 0.0059175752103328705, + 0.04329542815685272, + 0.017097484320402145, + 0.004642861895263195, + 0.006636402569711208, + 0.02003432996571064, + 0.015393677167594433, + 0.005079582333564758, + 0.001282649114727974, + 0.010461735539138317, + 0.003821418620646, + 0.0010520205833017826, + 0.0016704249428585172, + 0.002398395212367177, + 0.002518503926694393 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 7, + 8 + ], + "selected_reward": -0.004633760545402765 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, rendered in a lush, hand-drawn, Ghibli-inspired art style, focuses on a clump of tall cattail plants at the center of the frame. Their fibrous green stems, with their soft, painterly texture, are topped with distinctive spikes of velvety flowers a rich chocolate brown. The entire scene is bathed in the warm, soft glow of the afternoon sun, illuminating the tranquil marsh setting. Crystal-clear water rises steadily and deliberately around the green stems, its calm surface like a mirror, reflecting the soft hues of pink, pale gold, and lavender against the sky. Tiny, shimmering dragonflies, their iridescent wings, dance and hover in the air between the reeds, their delicate movements adding a touch of life and magic. The combination of the impending tide, the warm, bright lighting, and the classic animation aesthetic creates an overwhelming atmosphere of tranquility and magic.", + "seed": 27121, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27121_dp_26_rank_52_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1043999871239066, + "vlm_logprob": -0.1043999871239066, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.1043999871239066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.32018616795539856 + ], + "opposite_logprob_tokens": [ + -1.3201861381530762 + ], + "logprob_reward": -0.32018616795539856, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.32018616795539856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.41214972734451294, + 0.122316874563694, + 0.07777243852615356, + 0.19113010168075562, + 0.024161946028470993, + 0.02412104234099388, + 0.0239022895693779, + 0.08099202811717987, + 0.019404035061597824, + 0.024049455299973488 + ], + "credit_map": [ + 0.03558151051402092, + 0.031929902732372284, + 0.05829087644815445, + 0.049500834196805954, + 0.026841888204216957, + 0.0475301668047905, + 0.05713832005858421, + 0.1053362786769867, + 0.0163656547665596, + 0.01592961698770523, + 0.012244981713593006, + 0.01816575974225998, + 0.012773365713655949, + 0.015606450848281384, + 0.018153328448534012, + 0.01307772658765316, + 0.007902740500867367, + 0.010071001946926117, + 0.010359983891248703, + 0.0075377910397946835, + 0.0077524809166789055, + 0.00845203548669815, + 0.013117659837007523, + 0.012578751891851425, + 0.023118041455745697, + 0.0291350856423378, + 0.021483447402715683, + 0.023956092074513435, + 0.023510850965976715, + 0.029122676700353622, + 0.021971572190523148, + 0.018832355737686157, + 0.0025043548084795475, + 0.0031829429790377617, + 0.0029616327956318855, + 0.0024298601783812046, + 0.0025676328223198652, + 0.0034159766510128975, + 0.0030338726937770844, + 0.0040656765922904015, + 0.0029976211953908205, + 0.003305442165583372, + 0.0035993210040032864, + 0.0020864924881607294, + 0.002463581971824169, + 0.004153733141720295, + 0.0017260119784623384, + 0.0037888435181230307, + 0.0027926929760724306, + 0.001975474413484335, + 0.0032179916743189096, + 0.0024595260620117188, + 0.0019166379934176803, + 0.003324806224554777, + 0.0029656372498720884, + 0.005249525420367718, + 0.012626192532479763, + 0.006345899775624275, + 0.011380650103092194, + 0.009050137363374233, + 0.007385140750557184, + 0.009402874857187271, + 0.008614545688033104, + 0.01618659682571888, + 0.0022159635554999113, + 0.0023996976669877768, + 0.002610264578834176, + 0.0017358921468257904, + 0.002247678115963936, + 0.0026001902297139168, + 0.002277892315760255, + 0.003316457150503993, + 0.0027888575568795204, + 0.0030397861264646053, + 0.00302905379794538, + 0.0022604623809456825, + 0.002849856624379754, + 0.003047246253117919, + 0.0031288338359445333, + 0.0039053603541105986 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.32018616795539856 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036326153203845024 + ], + "opposite_logprob_tokens": [ + -5.628632545471191 + ], + "logprob_reward": -0.0036326153203845024, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036326153203845024, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3544887602329254, + 0.07599148154258728, + 0.07081341743469238, + 0.10633226484060287, + 0.19956748187541962, + 0.010906921699643135, + 0.02579472027719021, + 0.10427785664796829, + 0.027891449630260468, + 0.023935653269290924 + ], + "credit_map": [ + 0.02923029102385044, + 0.022530680522322655, + 0.05124979838728905, + 0.049658000469207764, + 0.017950724810361862, + 0.04326643794775009, + 0.0371963195502758, + 0.10340650379657745, + 0.011436432600021362, + 0.006226895842701197, + 0.009583772160112858, + 0.007027085404843092, + 0.006374198943376541, + 0.010235203430056572, + 0.00778190465644002, + 0.0173259936273098, + 0.006707001011818647, + 0.0066687664948403835, + 0.009857900440692902, + 0.008471266366541386, + 0.005972817540168762, + 0.0069360192865133286, + 0.01377643458545208, + 0.012423216365277767, + 0.012352481484413147, + 0.01645074225962162, + 0.013026297092437744, + 0.01203961856663227, + 0.013090427033603191, + 0.01540264580398798, + 0.012466882355511189, + 0.011503173969686031, + 0.025624457746744156, + 0.02762553095817566, + 0.02160041034221649, + 0.024455178529024124, + 0.023603888228535652, + 0.0367455780506134, + 0.016087913885712624, + 0.02382454089820385, + 0.0011610786896198988, + 0.0012140744365751743, + 0.0013161487877368927, + 0.001444858149625361, + 0.0011338831391185522, + 0.001253925496712327, + 0.0018253347370773554, + 0.0015576188452541828, + 0.0026823515072464943, + 0.0020508195739239454, + 0.002899979939684272, + 0.0024134593550115824, + 0.002354878233745694, + 0.0029585734009742737, + 0.005356168374419212, + 0.005078491754829884, + 0.012125346809625626, + 0.008071338757872581, + 0.02007252350449562, + 0.012157824821770191, + 0.007985284551978111, + 0.009360541589558125, + 0.010617024265229702, + 0.023887980729341507, + 0.002798687666654587, + 0.0026669478975236416, + 0.003658178960904479, + 0.003284669714048505, + 0.0026364452205598354, + 0.003524856409057975, + 0.004327376373112202, + 0.004994290415197611, + 0.00263608293607831, + 0.002517030108720064, + 0.0028129438869655132, + 0.003340336726978421, + 0.0020604776218533516, + 0.002772842301055789, + 0.004631206393241882, + 0.0031647358555346727 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.0036326153203845024 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014191891998052597 + ], + "opposite_logprob_tokens": [ + -4.2641921043396 + ], + "logprob_reward": -0.014191891998052597, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014191891998052597, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3552533686161041, + 0.2298499196767807, + 0.03502792492508888, + 0.1124836727976799, + 0.042840376496315, + 0.046572376042604446, + 0.06854309141635895, + 0.053699128329753876, + 0.022560179233551025, + 0.033169958740472794 + ], + "credit_map": [ + 0.027965102344751358, + 0.033044785261154175, + 0.05932622775435448, + 0.0416044220328331, + 0.025856947526335716, + 0.03155340999364853, + 0.03471633419394493, + 0.10118615627288818, + 0.02848026342689991, + 0.029956594109535217, + 0.025624621659517288, + 0.03300970420241356, + 0.020069118589162827, + 0.03420979157090187, + 0.0287864338606596, + 0.029713386669754982, + 0.004016184248030186, + 0.003742340486496687, + 0.005637092515826225, + 0.00386302312836051, + 0.003048862796276808, + 0.004470710176974535, + 0.0038747789803892374, + 0.006374933756887913, + 0.012466798536479473, + 0.01616171933710575, + 0.012603126466274261, + 0.015114338137209415, + 0.012503975071012974, + 0.01701737940311432, + 0.015140064060688019, + 0.011476277373731136, + 0.005841317120939493, + 0.005860449280589819, + 0.005093312356621027, + 0.0046562207862734795, + 0.004305788315832615, + 0.009560572914779186, + 0.003628325881436467, + 0.0038943919353187084, + 0.0051010483875870705, + 0.005982077680528164, + 0.0055162664502859116, + 0.00787974800914526, + 0.004726790823042393, + 0.006393508519977331, + 0.005677056033164263, + 0.005295883398503065, + 0.006366989575326443, + 0.004355980549007654, + 0.0086883045732975, + 0.012339371256530285, + 0.004196773283183575, + 0.006512789521366358, + 0.017824484035372734, + 0.008258399553596973, + 0.0050679282285273075, + 0.003195260651409626, + 0.007866875268518925, + 0.0089173074811697, + 0.0031068981625139713, + 0.0038006927352398634, + 0.008919927291572094, + 0.012824244797229767, + 0.0026017585769295692, + 0.001991257071495056, + 0.0025405657943338156, + 0.002853423822671175, + 0.0025123567320406437, + 0.0034816605038940907, + 0.003693428821861744, + 0.002885731402784586, + 0.003439183346927166, + 0.003715855535119772, + 0.003944827243685722, + 0.005102104973047972, + 0.002982150763273239, + 0.003009604290127754, + 0.006643928121775389, + 0.004332308191806078 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.014191891998052597 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17989672720432281 + ], + "opposite_logprob_tokens": [ + -1.8048967123031616 + ], + "logprob_reward": -0.17989672720432281, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17989672720432281, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.466924250125885, + 0.09758090227842331, + 0.041460782289505005, + 0.15025921165943146, + 0.08139035850763321, + 0.022533375769853592, + 0.027613449841737747, + 0.07021039724349976, + 0.021742120385169983, + 0.02028520777821541 + ], + "credit_map": [ + 0.038839105516672134, + 0.03229549527168274, + 0.07005445659160614, + 0.0601637028157711, + 0.024331673979759216, + 0.05129887908697128, + 0.04740658774971962, + 0.14253433048725128, + 0.012691443786025047, + 0.010350062511861324, + 0.009482553228735924, + 0.017573723569512367, + 0.008476157672703266, + 0.014295077882707119, + 0.013537969440221786, + 0.0111739132553339, + 0.0041506593115627766, + 0.004110647365450859, + 0.005455339793115854, + 0.005048390012234449, + 0.00347812008112669, + 0.00497414730489254, + 0.006842239294201136, + 0.007401238661259413, + 0.01822686195373535, + 0.023045137524604797, + 0.016921063885092735, + 0.018134744837880135, + 0.018738416954874992, + 0.022748885676264763, + 0.01770811155438423, + 0.01473599299788475, + 0.011296023614704609, + 0.010713019408285618, + 0.009780620224773884, + 0.008506998419761658, + 0.008279554545879364, + 0.01867886632680893, + 0.00606307340785861, + 0.008072193711996078, + 0.002597708022221923, + 0.0031643828842788935, + 0.002508215606212616, + 0.0037980233319103718, + 0.0023836472537368536, + 0.0027754856273531914, + 0.003028361825272441, + 0.0022775507532060146, + 0.0027458665426820517, + 0.002374840434640646, + 0.004030426032841206, + 0.0033521500881761312, + 0.0019468235550448298, + 0.0034089372493326664, + 0.005405821371823549, + 0.004348585847765207, + 0.0069824219681322575, + 0.007782428991049528, + 0.010180477984249592, + 0.007857725024223328, + 0.006892017088830471, + 0.008409370668232441, + 0.009839065372943878, + 0.012266883626580238, + 0.0024501713924109936, + 0.0019410207169130445, + 0.0027029516641050577, + 0.002619642298668623, + 0.0019582107197493315, + 0.003381500020623207, + 0.0037329564802348614, + 0.002955667208880186, + 0.0021478014532476664, + 0.002743086079135537, + 0.002614751225337386, + 0.0023314349818974733, + 0.0022469277027994394, + 0.002721938770264387, + 0.0030352247413247824, + 0.0024440425913780928 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.17989672720432281 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004092533141374588 + ], + "opposite_logprob_tokens": [ + -5.504092693328857 + ], + "logprob_reward": -0.004092533141374588, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004092533141374588, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.36770448088645935, + 0.15660938620567322, + 0.0537574402987957, + 0.03346257656812668, + 0.08751284331083298, + 0.07266482710838318, + 0.04447627812623978, + 0.13070720434188843, + 0.028505923226475716, + 0.024599093943834305 + ], + "credit_map": [ + 0.0350847952067852, + 0.04137110337615013, + 0.05789186805486679, + 0.03640580177307129, + 0.028729131445288658, + 0.03606034070253372, + 0.0339953750371933, + 0.09816606342792511, + 0.019611725583672523, + 0.02368594892323017, + 0.021792098879814148, + 0.01939672976732254, + 0.016569098457694054, + 0.02440837025642395, + 0.013976626098155975, + 0.01716878078877926, + 0.006337335333228111, + 0.007360049989074469, + 0.010442166589200497, + 0.005605045706033707, + 0.004814283922314644, + 0.007933787070214748, + 0.004086246248334646, + 0.007178522180765867, + 0.003978032618761063, + 0.004385475534945726, + 0.0037477787118405104, + 0.003587845480069518, + 0.003581282449886203, + 0.004569609183818102, + 0.005329757928848267, + 0.004282794427126646, + 0.011998958885669708, + 0.012196526862680912, + 0.010068136267364025, + 0.009370961226522923, + 0.009038018994033337, + 0.019531965255737305, + 0.006601974368095398, + 0.008706293068826199, + 0.00887687224894762, + 0.01035058218985796, + 0.008365457877516747, + 0.01125005166977644, + 0.007714024744927883, + 0.010522310622036457, + 0.007379377260804176, + 0.008206148631870747, + 0.0048064421862363815, + 0.0039016008377075195, + 0.004863044247031212, + 0.007387327495962381, + 0.003056582063436508, + 0.005425793118774891, + 0.009391849860548973, + 0.005643639247864485, + 0.015826711431145668, + 0.008465828374028206, + 0.017761170864105225, + 0.020792247727513313, + 0.009769786149263382, + 0.011218823492527008, + 0.01967342384159565, + 0.027199197560548782, + 0.0028282550629228354, + 0.002745451871305704, + 0.003565546590834856, + 0.0034549611154943705, + 0.0027530731167644262, + 0.004359041340649128, + 0.00451406417414546, + 0.004285529721528292, + 0.002498166635632515, + 0.0020226568449288607, + 0.0036770415026694536, + 0.0031738635152578354, + 0.001878321054391563, + 0.0022101211361587048, + 0.0046098316088318825, + 0.0045290919952094555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.004092533141374588 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, rendered in a lush, hand-drawn, Ghibli-inspired art style, focuses on a clump of tall cattail plants at the center of the frame. Their fibrous green stems, with their soft, painterly texture, are topped with distinctive spikes of velvety flowers a rich chocolate brown. The entire scene is bathed in the warm, soft glow of the afternoon sun, illuminating the tranquil marsh setting. Crystal-clear water rises steadily and deliberately around the green stems, its calm surface like a mirror, reflecting the soft hues of pink, pale gold, and lavender against the sky. Tiny, shimmering dragonflies, their iridescent wings, dance and hover in the air between the reeds, their delicate movements adding a touch of life and magic. The combination of the impending tide, the warm, bright lighting, and the classic animation aesthetic creates an overwhelming atmosphere of tranquility and magic.", + "seed": 27122, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27122_dp_27_rank_54_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.1820785873569548, + "vlm_logprob": -0.1820785873569548, + "vlm_yes_no_margin": 2.825000011920929, + "vlm_token_credit": -0.1820785873569548, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3904838263988495 + ], + "opposite_logprob_tokens": [ + -1.1404838562011719 + ], + "logprob_reward": -0.3904838263988495, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.3904838263988495, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33969104290008545, + 0.1360185742378235, + 0.06736551970243454, + 0.051871318370103836, + 0.05817078799009323, + 0.12475854158401489, + 0.04143385961651802, + 0.0563892126083374, + 0.08023001253604889, + 0.04407119005918503 + ], + "credit_map": [ + 0.05357898026704788, + 0.02717297151684761, + 0.08033953607082367, + 0.02422333136200905, + 0.021271081641316414, + 0.027351975440979004, + 0.029403476044535637, + 0.07634969055652618, + 0.019716648384928703, + 0.011432455852627754, + 0.02783784456551075, + 0.013894502073526382, + 0.010913386940956116, + 0.013192391954362392, + 0.013470130041241646, + 0.02556120604276657, + 0.01778181828558445, + 0.006900815758854151, + 0.012009745463728905, + 0.0055343881249427795, + 0.0047957091592252254, + 0.0057401810772717, + 0.007618582807481289, + 0.006984272040426731, + 0.009478217922151089, + 0.004696512594819069, + 0.007726483047008514, + 0.006465643644332886, + 0.00405553774908185, + 0.005753881298005581, + 0.00592006603255868, + 0.007774971425533295, + 0.0105085838586092, + 0.006209935527294874, + 0.006930577103048563, + 0.005192782264202833, + 0.006520546041429043, + 0.00835779681801796, + 0.006649197544902563, + 0.007801367435604334, + 0.01058342307806015, + 0.009596221148967743, + 0.02744109183549881, + 0.010517474263906479, + 0.009626233950257301, + 0.013711826875805855, + 0.006363012827932835, + 0.0369192510843277, + 0.006259771529585123, + 0.004720926750451326, + 0.006102867890149355, + 0.004013054072856903, + 0.005474297795444727, + 0.004390384536236525, + 0.004133036360144615, + 0.006339523009955883, + 0.006957796402275562, + 0.005401912610977888, + 0.008312619291245937, + 0.006974996533244848, + 0.007320143282413483, + 0.006952093914151192, + 0.005902303382754326, + 0.008567343465983868, + 0.005784562323242426, + 0.0022590127773582935, + 0.02933807298541069, + 0.004830024670809507, + 0.0031291621271520853, + 0.02468826249241829, + 0.0030424343422055244, + 0.00715847359970212, + 0.003622719319537282, + 0.0020490684546530247, + 0.012808267958462238, + 0.0035699785221368074, + 0.003286222228780389, + 0.012181330472230911, + 0.0029391476418823004, + 0.0036144540645182133 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.3904838263988495 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005268024746328592 + ], + "opposite_logprob_tokens": [ + -5.255268096923828 + ], + "logprob_reward": -0.005268024746328592, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005268024746328592, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1902131736278534, + 0.055957648903131485, + 0.03596208244562149, + 0.054409611970186234, + 0.17627575993537903, + 0.09875843673944473, + 0.049840789288282394, + 0.15574605762958527, + 0.13068629801273346, + 0.05215011537075043 + ], + "credit_map": [ + 0.0196510162204504, + 0.013103587552905083, + 0.06021353602409363, + 0.01480848528444767, + 0.007396978326141834, + 0.014031284488737583, + 0.01333248894661665, + 0.04767579212784767, + 0.013277147896587849, + 0.004564101807773113, + 0.01073792390525341, + 0.005166410934180021, + 0.0034698545932769775, + 0.0049437410198152065, + 0.004383108112961054, + 0.009415357373654842, + 0.00427780207246542, + 0.004886156879365444, + 0.0055425348691642284, + 0.00336452410556376, + 0.006420404650270939, + 0.004379463382065296, + 0.0037663804832845926, + 0.0033248139079660177, + 0.016039233654737473, + 0.005094620399177074, + 0.008282804861664772, + 0.0045000058598816395, + 0.003898404072970152, + 0.005650844890624285, + 0.004426282364875078, + 0.00651741586625576, + 0.01555558480322361, + 0.005705635994672775, + 0.048077069222927094, + 0.013806424103677273, + 0.005665381904691458, + 0.015556962229311466, + 0.009203343652188778, + 0.06270534545183182, + 0.006291108671575785, + 0.0022624353878200054, + 0.03548632934689522, + 0.008528684265911579, + 0.0032124361023306847, + 0.010745086707174778, + 0.0035601267591118813, + 0.028672223910689354, + 0.007783448789268732, + 0.006131048314273357, + 0.007912321016192436, + 0.004132145084440708, + 0.006738157942891121, + 0.005520117934793234, + 0.004119350574910641, + 0.007504199631512165, + 0.015035375021398067, + 0.020503172650933266, + 0.024854712188243866, + 0.015762344002723694, + 0.028912410140037537, + 0.020372411236166954, + 0.016612110659480095, + 0.013693507760763168, + 0.00936293788254261, + 0.005004459992051125, + 0.047308191657066345, + 0.00807313434779644, + 0.0061197709292173386, + 0.037740856409072876, + 0.00539891142398119, + 0.011678045615553856, + 0.005733060650527477, + 0.0032007102854549885, + 0.014406541362404823, + 0.0055668712593615055, + 0.003556953277438879, + 0.008008009754121304, + 0.0040765549056231976, + 0.007601412944495678 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.005268024746328592 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897798717021942 + ], + "opposite_logprob_tokens": [ + -2.5789780616760254 + ], + "logprob_reward": -0.07897798717021942, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897798717021942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18561750650405884, + 0.10738672316074371, + 0.09119195491075516, + 0.05767039582133293, + 0.2478601336479187, + 0.09043896943330765, + 0.04639052599668503, + 0.07729233801364899, + 0.07203076779842377, + 0.024120694026350975 + ], + "credit_map": [ + 0.026683486998081207, + 0.015986809507012367, + 0.03448488563299179, + 0.03374798595905304, + 0.017483608797192574, + 0.016999227926135063, + 0.01637471839785576, + 0.023856788873672485, + 0.02124318666756153, + 0.008244430646300316, + 0.02350037358701229, + 0.010615737177431583, + 0.007107215467840433, + 0.009694586507976055, + 0.00804920680820942, + 0.01893199235200882, + 0.02043440192937851, + 0.008656435646116734, + 0.013311224058270454, + 0.013685150071978569, + 0.009207343682646751, + 0.008655440993607044, + 0.007737499661743641, + 0.00950445793569088, + 0.008338145911693573, + 0.004664369393140078, + 0.01076758187264204, + 0.006391385570168495, + 0.004245141986757517, + 0.005050570238381624, + 0.0042252312414348125, + 0.013987970538437366, + 0.026828309521079063, + 0.007007440086454153, + 0.07265740633010864, + 0.021031927317380905, + 0.005322701297700405, + 0.01784469373524189, + 0.009978177025914192, + 0.08718948811292648, + 0.005074342247098684, + 0.002986227162182331, + 0.025040065869688988, + 0.007303200662136078, + 0.0032765951473265886, + 0.010044426657259464, + 0.0034834127873182297, + 0.033230699598789215, + 0.005929272621870041, + 0.006969270762056112, + 0.006481044460088015, + 0.004688716493546963, + 0.00865667499601841, + 0.006110140588134527, + 0.003909477032721043, + 0.003645928343757987, + 0.011408987455070019, + 0.004122809041291475, + 0.01893250085413456, + 0.006367845926433802, + 0.004574886057525873, + 0.013862308114767075, + 0.00599939888343215, + 0.012023595161736012, + 0.005124736577272415, + 0.002374319126829505, + 0.026637673377990723, + 0.004319797735661268, + 0.0029305664356797934, + 0.021400941535830498, + 0.0028266070876270533, + 0.006416121497750282, + 0.0026695572305470705, + 0.0013117945054545999, + 0.007142947055399418, + 0.001691476907581091, + 0.0020501415710896254, + 0.005351871717721224, + 0.0015390446642413735, + 0.0023638594429939985 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.07897798717021942 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42893046140670776 + ], + "opposite_logprob_tokens": [ + -1.0539305210113525 + ], + "logprob_reward": -0.42893046140670776, + "yes_no_margin_reward": 0.6250000596046448, + "token_credit_reward": -0.42893046140670776, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12907446920871735, + 0.07798229157924652, + 0.05799836665391922, + 0.09325063228607178, + 0.19224661588668823, + 0.1426367163658142, + 0.062436483800411224, + 0.1204366683959961, + 0.07582499831914902, + 0.04811272770166397 + ], + "credit_map": [ + 0.01763744279742241, + 0.013320495374500751, + 0.01842743530869484, + 0.015446568839251995, + 0.0130012147128582, + 0.01499455887824297, + 0.013238923624157906, + 0.023007817566394806, + 0.016399312764406204, + 0.0056760190054774284, + 0.018547743558883667, + 0.00729141803458333, + 0.003899473464116454, + 0.00684628589078784, + 0.006028483156114817, + 0.013293556869029999, + 0.014265895821154118, + 0.00625237263739109, + 0.010739855468273163, + 0.0049853576347231865, + 0.005787968169897795, + 0.005131764803081751, + 0.0048014880158007145, + 0.006033659912645817, + 0.02706371247768402, + 0.010852193459868431, + 0.012146510183811188, + 0.008015117608010769, + 0.007662210147827864, + 0.009130021557211876, + 0.007808832451701164, + 0.010572032071650028, + 0.015462415292859077, + 0.006099279969930649, + 0.05662688612937927, + 0.017403412610292435, + 0.006978018209338188, + 0.013895989395678043, + 0.010092557407915592, + 0.06568805873394012, + 0.01356072910130024, + 0.004032338969409466, + 0.03504889830946922, + 0.009351154789328575, + 0.004679701756685972, + 0.04978116601705551, + 0.005632785148918629, + 0.020549939945340157, + 0.01004024874418974, + 0.005578638520091772, + 0.010048916563391685, + 0.006501981057226658, + 0.005801595281809568, + 0.0064148795790970325, + 0.006712105590850115, + 0.011338124051690102, + 0.015231562778353691, + 0.009492781944572926, + 0.020333338528871536, + 0.016211602836847305, + 0.011231492273509502, + 0.016035767272114754, + 0.013214104808866978, + 0.018686018884181976, + 0.005457388237118721, + 0.0051663946360349655, + 0.02306712418794632, + 0.004946096334606409, + 0.005403433460742235, + 0.02175978198647499, + 0.00372429215349257, + 0.006300485227257013, + 0.004362448584288359, + 0.003638981143012643, + 0.01245022565126419, + 0.0034635344054549932, + 0.004391808062791824, + 0.01216331496834755, + 0.0035963451955467463, + 0.004046069923788309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.42893046140670776 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0067326370626688 + ], + "opposite_logprob_tokens": [ + -5.00673246383667 + ], + "logprob_reward": -0.0067326370626688, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.0067326370626688, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18149255216121674, + 0.03668571636080742, + 0.03603582829236984, + 0.06727929413318634, + 0.16269180178642273, + 0.042732611298561096, + 0.06510655581951141, + 0.143097385764122, + 0.2169550508260727, + 0.04792317748069763 + ], + "credit_map": [ + 0.020821649581193924, + 0.01626969873905182, + 0.02907465398311615, + 0.03467787429690361, + 0.014161115512251854, + 0.01732020080089569, + 0.018374819308519363, + 0.030792539939284325, + 0.004115891642868519, + 0.0028966290410608053, + 0.009139779023826122, + 0.0028117047622799873, + 0.0025174475740641356, + 0.0034396981354802847, + 0.0035667556803673506, + 0.008197806775569916, + 0.008598213084042072, + 0.002743312856182456, + 0.004828914999961853, + 0.004201716277748346, + 0.002623265376314521, + 0.0042297798208892345, + 0.003823325037956238, + 0.004987293854355812, + 0.007143781520426273, + 0.004709663800895214, + 0.013171953149139881, + 0.008557667024433613, + 0.005489182658493519, + 0.005801843944936991, + 0.005310028791427612, + 0.01709516905248165, + 0.013437601737678051, + 0.004534316714853048, + 0.047243744134902954, + 0.015386709943413734, + 0.004437583032995462, + 0.012258801609277725, + 0.009027274325489998, + 0.05636576563119888, + 0.0048841494135558605, + 0.002174602821469307, + 0.011385266669094563, + 0.0037628507707268, + 0.002563084941357374, + 0.009029045701026917, + 0.002715044654905796, + 0.006218560039997101, + 0.010049717500805855, + 0.008353746496140957, + 0.009844709187746048, + 0.006386674474924803, + 0.009742574766278267, + 0.007603703532367945, + 0.005413510370999575, + 0.007711919490247965, + 0.02099120430648327, + 0.006992664188146591, + 0.035580944269895554, + 0.011539829894900322, + 0.008080721832811832, + 0.025706738233566284, + 0.011522951535880566, + 0.022682320326566696, + 0.015529202297329903, + 0.006496177986264229, + 0.08034273982048035, + 0.013042982667684555, + 0.008825989440083504, + 0.0634574294090271, + 0.009299132972955704, + 0.019961383193731308, + 0.0038606340531259775, + 0.0020737359300255775, + 0.014726515859365463, + 0.002764681354165077, + 0.0038264160975813866, + 0.015500068664550781, + 0.0021891773212701082, + 0.002981943544000387 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 4 + ], + "selected_reward": -0.0067326370626688 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, eye-level medium shot, rendered in a lush, hand-drawn, Ghibli-inspired art style, focuses on a clump of tall cattail plants at the center of the frame. Their fibrous green stems, with their soft, painterly texture, are topped with distinctive spikes of velvety flowers a rich chocolate brown. The entire scene is bathed in the warm, soft glow of the afternoon sun, illuminating the tranquil marsh setting. Crystal-clear water rises steadily and deliberately around the green stems, its calm surface like a mirror, reflecting the soft hues of pink, pale gold, and lavender against the sky. Tiny, shimmering dragonflies, their iridescent wings, dance and hover in the air between the reeds, their delicate movements adding a touch of life and magic. The combination of the impending tide, the warm, bright lighting, and the classic animation aesthetic creates an overwhelming atmosphere of tranquility and magic.", + "seed": 27123, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_13/index_9683_seed_27123_dp_27_rank_54_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 13, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.13508756505325437, + "vlm_logprob": -0.13508756505325437, + "vlm_yes_no_margin": 3.075, + "vlm_token_credit": -0.13508756505325437, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20646576583385468 + ], + "opposite_logprob_tokens": [ + -1.706465721130371 + ], + "logprob_reward": -0.20646576583385468, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20646576583385468, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3782414197921753, + 0.09095276147127151, + 0.09564312547445297, + 0.042277637869119644, + 0.07998614013195038, + 0.09726960957050323, + 0.13342241942882538, + 0.02663259766995907, + 0.031818289309740067, + 0.023756036534905434 + ], + "credit_map": [ + 0.04787139967083931, + 0.01855849288403988, + 0.047550175338983536, + 0.05161178112030029, + 0.030306454747915268, + 0.024092519655823708, + 0.09811226278543472, + 0.060138337314128876, + 0.011515654623508453, + 0.00643069576472044, + 0.012413544580340385, + 0.013488473370671272, + 0.007686609867960215, + 0.008507126942276955, + 0.016797136515378952, + 0.014113517478108406, + 0.01161185372620821, + 0.004661434795707464, + 0.011637435294687748, + 0.015726329758763313, + 0.00820079818367958, + 0.006156620103865862, + 0.022903939709067345, + 0.014744716696441174, + 0.0072870212607085705, + 0.002800272312015295, + 0.0073838247917592525, + 0.005130619276314974, + 0.0032168624456971884, + 0.0045195999555289745, + 0.007454520557075739, + 0.0044849165715277195, + 0.015415173955261707, + 0.003941353876143694, + 0.01676429808139801, + 0.007765789981931448, + 0.005161399021744728, + 0.008552977815270424, + 0.009802945889532566, + 0.012582198716700077, + 0.012255149893462658, + 0.00649745250120759, + 0.013022537343204021, + 0.017410753294825554, + 0.00922652892768383, + 0.006120127160102129, + 0.020404089242219925, + 0.012332966551184654, + 0.01589842699468136, + 0.009083232842385769, + 0.02090013585984707, + 0.026446526870131493, + 0.012725181877613068, + 0.01075247023254633, + 0.02462805062532425, + 0.01298838946968317, + 0.00318029779009521, + 0.0018459113780409098, + 0.004657533951103687, + 0.0033466648310422897, + 0.0026067972648888826, + 0.003312993561848998, + 0.0037545168306678534, + 0.003927882760763168, + 0.00441637821495533, + 0.0019538355991244316, + 0.006202122196555138, + 0.00406888360157609, + 0.0028714004438370466, + 0.003514245618134737, + 0.0035426439717411995, + 0.005248784087598324, + 0.00286328443326056, + 0.0017715352587401867, + 0.0040228781290352345, + 0.0035787164233624935, + 0.002705964492633939, + 0.0030590591486543417, + 0.0028063443023711443, + 0.002948254579678178 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 5 + ], + "selected_reward": -0.20646576583385468 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00865730457007885 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.00865730457007885, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.00865730457007885, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.200839102268219, + 0.16532860696315765, + 0.04478323087096214, + 0.07230985909700394, + 0.1018093079328537, + 0.05236909165978432, + 0.0767861008644104, + 0.07054978609085083, + 0.18409305810928345, + 0.031131824478507042 + ], + "credit_map": [ + 0.0236051082611084, + 0.015127675607800484, + 0.030728086829185486, + 0.03054358810186386, + 0.016945499926805496, + 0.023709725588560104, + 0.022287964820861816, + 0.03789146617054939, + 0.020431464537978172, + 0.010086850263178349, + 0.02092006616294384, + 0.03414323925971985, + 0.013914359733462334, + 0.011360840871930122, + 0.030112147331237793, + 0.024359652772545815, + 0.006597703322768211, + 0.0030718266498297453, + 0.006024679634720087, + 0.006137610413134098, + 0.004056887235492468, + 0.004897368606179953, + 0.008398438803851604, + 0.005598717834800482, + 0.012552405707538128, + 0.004059286322444677, + 0.011106574907898903, + 0.009665464051067829, + 0.005035864654928446, + 0.006237787194550037, + 0.015361079946160316, + 0.008291401900351048, + 0.015165531076490879, + 0.0065590497106313705, + 0.013942687772214413, + 0.016313083469867706, + 0.01050048228353262, + 0.008647652342915535, + 0.01997222937643528, + 0.010708591900765896, + 0.006700597237795591, + 0.0036225614603608847, + 0.006570015102624893, + 0.008580721914768219, + 0.0053922077640891075, + 0.004160434473305941, + 0.011274989694356918, + 0.006067566107958555, + 0.008489045314490795, + 0.0051466720178723335, + 0.01168100070208311, + 0.014381084591150284, + 0.00823319423943758, + 0.005579137708991766, + 0.013164716772735119, + 0.010111252777278423, + 0.008875438012182713, + 0.005464302841573954, + 0.010004514828324318, + 0.011353861540555954, + 0.0075414152815938, + 0.006487096659839153, + 0.013928967528045177, + 0.006894188467413187, + 0.02176501229405403, + 0.013909469358623028, + 0.026483405381441116, + 0.03385986387729645, + 0.02110408805310726, + 0.014740688726305962, + 0.03483569249510765, + 0.01739484630525112, + 0.0035063044633716345, + 0.0019759477581828833, + 0.004991963971406221, + 0.00524884182959795, + 0.002805748488754034, + 0.00262964959256351, + 0.005803342442959547, + 0.004170027561485767 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.00865730457007885 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07002470642328262 + ], + "opposite_logprob_tokens": [ + -2.6950247287750244 + ], + "logprob_reward": -0.07002470642328262, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07002470642328262, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21684223413467407, + 0.19790978729724884, + 0.05019020661711693, + 0.23822931945323944, + 0.07713950425386429, + 0.06541500240564346, + 0.03135623782873154, + 0.04248800873756409, + 0.03608887642621994, + 0.044340863823890686 + ], + "credit_map": [ + 0.027808139100670815, + 0.014590387232601643, + 0.03210766986012459, + 0.0367739200592041, + 0.015410852618515491, + 0.024293968454003334, + 0.02580149471759796, + 0.040055785328149796, + 0.029297973960638046, + 0.010817007161676884, + 0.02306257374584675, + 0.03176179528236389, + 0.014798290096223354, + 0.012604502961039543, + 0.04565977677702904, + 0.02990787662565708, + 0.008140699937939644, + 0.0038817094173282385, + 0.008617207407951355, + 0.004847949370741844, + 0.004423214588314295, + 0.007298574782907963, + 0.004935771692544222, + 0.008045081049203873, + 0.038969941437244415, + 0.012607607990503311, + 0.03202129155397415, + 0.02788008190691471, + 0.017773255705833435, + 0.017947331070899963, + 0.05895093083381653, + 0.03207887336611748, + 0.00956866517663002, + 0.005330004263669252, + 0.0107701001688838, + 0.013115650042891502, + 0.006764882244169712, + 0.00554141029715538, + 0.014478050172328949, + 0.011570742353796959, + 0.01008890476077795, + 0.004052326548844576, + 0.0121436333283782, + 0.01020434033125639, + 0.004769491031765938, + 0.005221240688115358, + 0.010341021232306957, + 0.00859404169023037, + 0.005322705954313278, + 0.003034335793927312, + 0.004256499465554953, + 0.004194892477244139, + 0.004239281173795462, + 0.003400839166715741, + 0.0033982028253376484, + 0.0035094774793833494, + 0.004312288947403431, + 0.0027837750967592, + 0.006202294025570154, + 0.009157541207969189, + 0.004719514399766922, + 0.0030719018541276455, + 0.007085367571562529, + 0.0051553272642195225, + 0.004308838862925768, + 0.002711719134822488, + 0.005182814318686724, + 0.006530932150781155, + 0.004254015162587166, + 0.00283659715205431, + 0.006530213635414839, + 0.0037337469402700663, + 0.005413869861513376, + 0.002778155729174614, + 0.007271475624293089, + 0.006926375906914473, + 0.003806931432336569, + 0.002955304691568017, + 0.009967570193111897, + 0.005221182014793158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.07002470642328262 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38709819316864014 + ], + "opposite_logprob_tokens": [ + -1.1370981931686401 + ], + "logprob_reward": -0.38709819316864014, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38709819316864014, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2983739376068115, + 0.15646305680274963, + 0.0728377178311348, + 0.05935453623533249, + 0.10746025294065475, + 0.05830679461359978, + 0.06386369466781616, + 0.09337692707777023, + 0.05368591472506523, + 0.036277156323194504 + ], + "credit_map": [ + 0.03655051440000534, + 0.020839514210820198, + 0.04609978571534157, + 0.04884624108672142, + 0.020591825246810913, + 0.03288774564862251, + 0.03647872060537338, + 0.05607958883047104, + 0.02176784910261631, + 0.010281319729983807, + 0.018902955576777458, + 0.026819545775651932, + 0.012222062796354294, + 0.013042010366916656, + 0.02939569763839245, + 0.024031613022089005, + 0.008108758367598057, + 0.0047693075612187386, + 0.009030649438500404, + 0.014457435347139835, + 0.006818459834903479, + 0.006228333339095116, + 0.011827267706394196, + 0.011597500182688236, + 0.009741125628352165, + 0.0031397445127367973, + 0.008689483627676964, + 0.006003627087920904, + 0.005221892613917589, + 0.0044072531163692474, + 0.013001936487853527, + 0.009149466641247272, + 0.016394464299082756, + 0.005779491737484932, + 0.0239866990596056, + 0.01196242030709982, + 0.006794820073992014, + 0.010867685079574585, + 0.013132174499332905, + 0.018542489036917686, + 0.00723614264279604, + 0.004104523453861475, + 0.0086720185354352, + 0.00932662095874548, + 0.005312093999236822, + 0.004301280248910189, + 0.011350681073963642, + 0.008003431372344494, + 0.007118039298802614, + 0.0062998030334711075, + 0.012696775607764721, + 0.007316072937101126, + 0.005626785568892956, + 0.00664026103913784, + 0.006937232334166765, + 0.01122872345149517, + 0.013220608234405518, + 0.006853620056062937, + 0.013106366619467735, + 0.014055426232516766, + 0.00840433407574892, + 0.008052225224673748, + 0.01997794583439827, + 0.009706400334835052, + 0.007293162401765585, + 0.0034801620058715343, + 0.009771485812962055, + 0.007545933593064547, + 0.0036411830224096775, + 0.004473543725907803, + 0.008720358833670616, + 0.008760078810155392, + 0.004765884019434452, + 0.0029465060215443373, + 0.006100537721067667, + 0.005055942106992006, + 0.003916424233466387, + 0.004402756225317717, + 0.005043049808591604, + 0.004046052694320679 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.38709819316864014 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the art style hand-drawn and Ghibli-inspired?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are tall cattail plants with fibrous green stems and chocolate brown flower spikes visible at the center of the frame?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the scene set in a tranquil marsh with crystal-clear water rising around the stems?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are tiny, shimmering dragonflies with iridescent wings visible dancing and hovering between the reeds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the scene bathed in the warm, soft glow of afternoon sun with a calm water surface reflecting pink, pale gold, and lavender hues?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0031918552704155445 + ], + "opposite_logprob_tokens": [ + -5.753191947937012 + ], + "logprob_reward": -0.0031918552704155445, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.0031918552704155445, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23979781568050385, + 0.18792076408863068, + 0.0734536275267601, + 0.09933344274759293, + 0.07309512048959732, + 0.03487791121006012, + 0.17033517360687256, + 0.030612995848059654, + 0.07163812965154648, + 0.018935056403279305 + ], + "credit_map": [ + 0.03540825843811035, + 0.015897417441010475, + 0.03071541339159012, + 0.03308737650513649, + 0.015964575111865997, + 0.0280309971421957, + 0.04585346579551697, + 0.03484028950333595, + 0.017876744270324707, + 0.0141080217435956, + 0.029742185026407242, + 0.039254240691661835, + 0.014206320978701115, + 0.019198495894670486, + 0.0239842738956213, + 0.029550477862358093, + 0.007980511523783207, + 0.004284210037440062, + 0.009336560033261776, + 0.013386714272201061, + 0.0052914791740477085, + 0.005974120460450649, + 0.014814701862633228, + 0.012385326437652111, + 0.017477253451943398, + 0.005668682511895895, + 0.011692788451910019, + 0.011353600770235062, + 0.008120141923427582, + 0.008004655130207539, + 0.023330112919211388, + 0.013686198741197586, + 0.012930496595799923, + 0.004903874360024929, + 0.013795153237879276, + 0.008072871714830399, + 0.007277630269527435, + 0.008440004661679268, + 0.006786685436964035, + 0.010888399556279182, + 0.0058920253068208694, + 0.0026686107739806175, + 0.0051920535042881966, + 0.004152083769440651, + 0.0037641962990164757, + 0.0037513759452849627, + 0.005571478977799416, + 0.003886084072291851, + 0.02183661423623562, + 0.011238407343626022, + 0.02316812053322792, + 0.030295684933662415, + 0.015596139244735241, + 0.013618778437376022, + 0.03506818041205406, + 0.019513234496116638, + 0.0045371693558990955, + 0.0022297825198620558, + 0.004839868284761906, + 0.004764295183122158, + 0.0020230135414749384, + 0.004399703815579414, + 0.0036067890468984842, + 0.004212372004985809, + 0.00875123217701912, + 0.005196296144276857, + 0.011090428568422794, + 0.012817940674722195, + 0.00821826234459877, + 0.005724926944822073, + 0.012790239416062832, + 0.007048801984637976, + 0.002459675772115588, + 0.001150545198470354, + 0.003838571021333337, + 0.0023100252728909254, + 0.0014807378174737096, + 0.0026095479261130095, + 0.0020513904746621847, + 0.003034563036635518 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.0031918552704155445 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_tensor([4311, 4311])_group_14_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_tensor([4311, 4311])_group_14_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..09b014a6ba036fb67a7aa8dca01e2ecb4fef2915 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_tensor([4311, 4311])_group_14_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "stunning low-angle, wide-angle shot captures a scene of pure fantasy. The camera slowly and steadily pans to the right, tracking a massive crocodile soaring majestically across a vast, cloudy twilight sky. The horizon, shimmering with deep purple and soft, fiery orange, casts a warm, ethereal glow upward onto the underbelly of billowing, cotton-like clouds. The crocodile, an ancient and immense creature, moves with remarkable grace, its massive body silhouetted against the fading light. Its scales, a mosaic of deep moss green and obsidian black, possess an iridescent quality, shimmering with reflected hues of purple and gold as it catches the last traces of sunlight. Its powerful legs are neatly tucked at its sides, while its mighty, ridged tail ripples in a slow, deliberate, and powerful rhythm, propelling it forward as if swimming through dense air. A wisp of cloud drifts past its long snout and ancient, wise eyes, momentarily blurring before revealing its immense size, creating an atmosphere of tranquility, awe, and surreal majesty.", + "seed": 15914, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_4311_seed_15914_dp_28_rank_56_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03844629205996171, + "vlm_logprob": -0.03844629205996171, + "vlm_yes_no_margin": 4.05, + "vlm_token_credit": -0.03844629205996171, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02827511914074421 + ], + "opposite_logprob_tokens": [ + -3.6532750129699707 + ], + "logprob_reward": -0.02827511914074421, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02827511914074421, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20202450454235077, + 0.11627291142940521, + 0.2351665496826172, + 0.08342092484235764, + 0.052863240242004395, + 0.1103452518582344, + 0.06100383400917053, + 0.04648013412952423, + 0.03370623290538788, + 0.05871644988656044 + ], + "credit_map": [ + 0.01191745325922966, + 0.02487063780426979, + 0.021480252966284752, + 0.036420371383428574, + 0.016551107168197632, + 0.024573035538196564, + 0.04597928747534752, + 0.020232323557138443, + 0.010936531238257885, + 0.014796878211200237, + 0.013260411098599434, + 0.013102402910590172, + 0.01758481375873089, + 0.014359261840581894, + 0.015665069222450256, + 0.016567526385188103, + 0.02447514608502388, + 0.02317379042506218, + 0.026521775871515274, + 0.031014790758490562, + 0.03176487982273102, + 0.03409373387694359, + 0.026803547516465187, + 0.03731884807348251, + 0.010456046089529991, + 0.008881916292011738, + 0.008206536993384361, + 0.01140026468783617, + 0.012327141128480434, + 0.012562310323119164, + 0.01011962816119194, + 0.009467068128287792, + 0.007529630325734615, + 0.0068457904271781445, + 0.005277823191136122, + 0.005194299854338169, + 0.00848122127354145, + 0.007932241074740887, + 0.005551367066800594, + 0.006050861906260252, + 0.013997804373502731, + 0.011908608488738537, + 0.013258569873869419, + 0.0093294158577919, + 0.01731059141457081, + 0.014520745724439621, + 0.013326493091881275, + 0.016693009063601494, + 0.019654151052236557, + 0.003118169726803899, + 0.002461842494085431, + 0.021651459857821465, + 0.0028129974380135536, + 0.0021331782918423414, + 0.003933645319193602, + 0.0052383835427463055, + 0.006815443746745586, + 0.0034177196212112904, + 0.0023813864681869745, + 0.013827341608703136, + 0.004162231460213661, + 0.0051291389390826225, + 0.006206831894814968, + 0.004540036432445049, + 0.003934125415980816, + 0.0020921535324305296, + 0.004504749551415443, + 0.0045099686831235886, + 0.003183667780831456, + 0.004801766946911812, + 0.005311333574354649, + 0.005368463229387999, + 0.0048706382513046265, + 0.0033649115357548, + 0.010747787542641163, + 0.016701715067029, + 0.0029097888618707657, + 0.003372224746271968, + 0.00970617588609457, + 0.0070432028733193874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.02827511914074421 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04866885766386986 + ], + "opposite_logprob_tokens": [ + -3.04866886138916 + ], + "logprob_reward": -0.04866885766386986, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04866885766386986, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08999063074588776, + 0.04580340534448624, + 0.1444561928510666, + 0.026710517704486847, + 0.029875079169869423, + 0.08669307827949524, + 0.42901262640953064, + 0.027167223393917084, + 0.017976928502321243, + 0.10231433063745499 + ], + "credit_map": [ + 0.007020650897175074, + 0.015577748417854309, + 0.007832527160644531, + 0.011408773250877857, + 0.00796233769506216, + 0.010503144934773445, + 0.019950799643993378, + 0.009734641760587692, + 0.006119837053120136, + 0.005499124526977539, + 0.004563287366181612, + 0.004836220759898424, + 0.007970231585204601, + 0.005069708451628685, + 0.004539088811725378, + 0.00720590865239501, + 0.01259496808052063, + 0.014726074412465096, + 0.015265430323779583, + 0.017801886424422264, + 0.016054999083280563, + 0.02145862951874733, + 0.020544007420539856, + 0.02601020410656929, + 0.00342936092056334, + 0.0027256004977971315, + 0.0019431195687502623, + 0.0034437633585184813, + 0.003495413577184081, + 0.005053041037172079, + 0.0035219574347138405, + 0.0030982608441263437, + 0.0037836753763258457, + 0.0045632063411176205, + 0.002935109892860055, + 0.0037239810917526484, + 0.004942768719047308, + 0.0032940900418907404, + 0.003543474245816469, + 0.0030887729953974485, + 0.013215572573244572, + 0.008896803483366966, + 0.010588458739221096, + 0.008760332129895687, + 0.011620661243796349, + 0.009354504756629467, + 0.013297670520842075, + 0.010959071107208729, + 0.1427234560251236, + 0.0212535560131073, + 0.01659453846514225, + 0.15845203399658203, + 0.01476387120783329, + 0.012967889197170734, + 0.02628047578036785, + 0.035976797342300415, + 0.003120583947747946, + 0.002691972069442272, + 0.002269330434501171, + 0.0037130468990653753, + 0.003961860667914152, + 0.0039405133575201035, + 0.0035642057191580534, + 0.003905709134414792, + 0.0023995235096663237, + 0.0014423165703192353, + 0.0017845176625996828, + 0.002322000917047262, + 0.001781898783519864, + 0.003100736765190959, + 0.0022676431108266115, + 0.002878290368244052, + 0.011398587375879288, + 0.008715278469026089, + 0.024000780656933784, + 0.017310084775090218, + 0.006703370716422796, + 0.005138275679200888, + 0.018749987706542015, + 0.010297971777617931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 2, + 9 + ], + "selected_reward": -0.04866885766386986 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008110094931907952 + ], + "opposite_logprob_tokens": [ + -7.1258111000061035 + ], + "logprob_reward": -0.0008110094931907952, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008110094931907952, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13176262378692627, + 0.12978294491767883, + 0.0904146283864975, + 0.04049522057175636, + 0.05730900913476944, + 0.06778108328580856, + 0.22214822471141815, + 0.09411586076021194, + 0.02554117701947689, + 0.14064927399158478 + ], + "credit_map": [ + 0.008713053539395332, + 0.015293248929083347, + 0.010715410113334656, + 0.017494168132543564, + 0.009350399486720562, + 0.01782662607729435, + 0.03209070488810539, + 0.02027900330722332, + 0.010000944137573242, + 0.02001800574362278, + 0.01603463478386402, + 0.016505390405654907, + 0.019543034955859184, + 0.01639789342880249, + 0.013829139992594719, + 0.01745389960706234, + 0.01206991821527481, + 0.01217760518193245, + 0.007771011907607317, + 0.015333874151110649, + 0.009554759599268436, + 0.007055522873997688, + 0.013431399129331112, + 0.0130205312743783, + 0.009309700690209866, + 0.005244727712124586, + 0.0037549196276813745, + 0.0035543746780604124, + 0.005778426304459572, + 0.004177117254585028, + 0.003898760536685586, + 0.0047771926037967205, + 0.007812423165887594, + 0.008566910400986671, + 0.005619381554424763, + 0.006491397507488728, + 0.008342950604856014, + 0.0077440557070076466, + 0.0065628415904939175, + 0.0061690472066402435, + 0.010175492614507675, + 0.008417804725468159, + 0.005447898525744677, + 0.00664431182667613, + 0.006844521500170231, + 0.008495017886161804, + 0.010177385061979294, + 0.011578644625842571, + 0.07385927438735962, + 0.011640336364507675, + 0.00852910801768303, + 0.08232872188091278, + 0.007407975848764181, + 0.006103005260229111, + 0.01376388594508171, + 0.018515896052122116, + 0.01851281151175499, + 0.00599292665719986, + 0.0047211111523211, + 0.03358820825815201, + 0.004763854201883078, + 0.008340609259903431, + 0.011468620970845222, + 0.0067277103662490845, + 0.0038412935100495815, + 0.0019963120575994253, + 0.002749609760940075, + 0.0032968102023005486, + 0.003366601187735796, + 0.002761449432000518, + 0.004486489109694958, + 0.003042608965188265, + 0.018495194613933563, + 0.011467729695141315, + 0.031501688063144684, + 0.02121894620358944, + 0.008806001394987106, + 0.007988437078893185, + 0.027249738574028015, + 0.013921529054641724 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.0008110094931907952 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028552263975143 + ], + "opposite_logprob_tokens": [ + -2.350285530090332 + ], + "logprob_reward": -0.10028552263975143, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028552263975143, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12367913872003555, + 0.07321107387542725, + 0.2091154158115387, + 0.06202970817685127, + 0.15893572568893433, + 0.01904425024986267, + 0.13972990214824677, + 0.1343529373407364, + 0.04638609290122986, + 0.03351578116416931 + ], + "credit_map": [ + 0.008641492575407028, + 0.015671079978346825, + 0.014478459022939205, + 0.01750866137444973, + 0.009710297919809818, + 0.017786556854844093, + 0.027225252240896225, + 0.012657334096729755, + 0.005800112150609493, + 0.012400093488395214, + 0.010452765971422195, + 0.009073606692254543, + 0.010441514663398266, + 0.009331625886261463, + 0.006763999816030264, + 0.008947350084781647, + 0.02019030787050724, + 0.022860797122120857, + 0.021330663934350014, + 0.023193176835775375, + 0.02667609602212906, + 0.0277855321764946, + 0.028800858184695244, + 0.03827795386314392, + 0.010645014233887196, + 0.006271475460380316, + 0.00602260697633028, + 0.00789534393697977, + 0.007935517467558384, + 0.008485399186611176, + 0.007479825057089329, + 0.007294518873095512, + 0.022081462666392326, + 0.02227761223912239, + 0.015931982547044754, + 0.018671490252017975, + 0.0269386675208807, + 0.019598837941884995, + 0.018928619101643562, + 0.014507032930850983, + 0.002874711761251092, + 0.0029606407042592764, + 0.0014244427438825369, + 0.0017852864693850279, + 0.002420806558802724, + 0.002344471635296941, + 0.002375468146055937, + 0.0028584194369614124, + 0.044999077916145325, + 0.00730852922424674, + 0.005512901581823826, + 0.05013005807995796, + 0.0060041300021111965, + 0.00519520603120327, + 0.00874591339379549, + 0.011834080331027508, + 0.024768264964222908, + 0.014065436087548733, + 0.009888972155749798, + 0.029766812920570374, + 0.00950116291642189, + 0.01590886525809765, + 0.013945875689387321, + 0.016507534310221672, + 0.004379954654723406, + 0.004515201319009066, + 0.006366453133523464, + 0.006042201537638903, + 0.006566903088241816, + 0.006271988619118929, + 0.00570230046287179, + 0.006541087757796049, + 0.0030628798995167017, + 0.0023884742986410856, + 0.004901609383523464, + 0.0064401328563690186, + 0.0023867276031523943, + 0.0025513602886348963, + 0.006149083375930786, + 0.00563551252707839 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 6 + ], + "selected_reward": -0.10028552263975143 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014190951362252235 + ], + "opposite_logprob_tokens": [ + -4.264191150665283 + ], + "logprob_reward": -0.014190951362252235, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014190951362252235, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1027553528547287, + 0.07966267317533493, + 0.10167459398508072, + 0.06441342085599899, + 0.0656176209449768, + 0.05678676813840866, + 0.2724144756793976, + 0.11304239928722382, + 0.03922441229224205, + 0.1044083684682846 + ], + "credit_map": [ + 0.006795456632971764, + 0.01149600651115179, + 0.010286539793014526, + 0.017819544300436974, + 0.008722447790205479, + 0.012963579036295414, + 0.024060742929577827, + 0.010611042380332947, + 0.007008751388639212, + 0.011495093815028667, + 0.010749472305178642, + 0.00898565910756588, + 0.012851542793214321, + 0.009983568452298641, + 0.008515252731740475, + 0.01007333304733038, + 0.011447129771113396, + 0.011719384230673313, + 0.008145119063556194, + 0.01264953799545765, + 0.014359141699969769, + 0.011216810904443264, + 0.01481796894222498, + 0.01731949858367443, + 0.007277223747223616, + 0.007014904636889696, + 0.006758519448339939, + 0.010963505133986473, + 0.009321318008005619, + 0.009230362251400948, + 0.006765850819647312, + 0.0070817358791828156, + 0.010739893652498722, + 0.010347816161811352, + 0.0061025070026516914, + 0.008344870992004871, + 0.00724724680185318, + 0.005376516841351986, + 0.007926537655293941, + 0.009532232768833637, + 0.0083433473482728, + 0.006783406715840101, + 0.004708013031631708, + 0.004773745778948069, + 0.00954695139080286, + 0.005530711729079485, + 0.006493272725492716, + 0.010607317090034485, + 0.0899476706981659, + 0.01353172305971384, + 0.010737113654613495, + 0.1002119928598404, + 0.009354537352919579, + 0.007903849706053734, + 0.01729266718029976, + 0.023434899747371674, + 0.019989242777228355, + 0.006063640117645264, + 0.006666585337370634, + 0.037681806832551956, + 0.009212136268615723, + 0.008745672181248665, + 0.014407902956008911, + 0.010275415144860744, + 0.005363793577998877, + 0.003970805089920759, + 0.0040106684900820255, + 0.0035664516035467386, + 0.006283423397690058, + 0.004779868293553591, + 0.005525720305740833, + 0.005723683163523674, + 0.011864832602441311, + 0.01162331085652113, + 0.020310450345277786, + 0.016936173662543297, + 0.0074895028956234455, + 0.007018616423010826, + 0.01930634118616581, + 0.009859145618975163 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 7, + 9 + ], + "selected_reward": -0.014190951362252235 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle, wide-angle shot captures a scene of pure fantasy. The camera slowly and steadily pans to the right, tracking a massive crocodile soaring majestically across a vast, cloudy twilight sky. The horizon, shimmering with deep purple and soft, fiery orange, casts a warm, ethereal glow upward onto the underbelly of billowing, cotton-like clouds. The crocodile, an ancient and immense creature, moves with remarkable grace, its massive body silhouetted against the fading light. Its scales, a mosaic of deep moss green and obsidian black, possess an iridescent quality, shimmering with reflected hues of purple and gold as it catches the last traces of sunlight. Its powerful legs are neatly tucked at its sides, while its mighty, ridged tail ripples in a slow, deliberate, and powerful rhythm, propelling it forward as if swimming through dense air. A wisp of cloud drifts past its long snout and ancient, wise eyes, momentarily blurring before revealing its immense size, creating an atmosphere of tranquility, awe, and surreal majesty.", + "seed": 15915, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_4311_seed_15915_dp_28_rank_56_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.0558923308737576, + "vlm_logprob": -0.0558923308737576, + "vlm_yes_no_margin": 3.6, + "vlm_token_credit": -0.0558923308737576, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.11457546055316925 + ], + "opposite_logprob_tokens": [ + -2.2395753860473633 + ], + "logprob_reward": -0.11457546055316925, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.11457546055316925, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26071757078170776, + 0.1173214241862297, + 0.061953626573085785, + 0.07304447889328003, + 0.055562447756528854, + 0.08604023605585098, + 0.143428772687912, + 0.07664172351360321, + 0.07151545584201813, + 0.053774233907461166 + ], + "credit_map": [ + 0.03605429455637932, + 0.03335551545023918, + 0.027000874280929565, + 0.03919452056288719, + 0.02267124317586422, + 0.03356160596013069, + 0.0256804209202528, + 0.043199095875024796, + 0.010663180612027645, + 0.01061038114130497, + 0.006813804153352976, + 0.010780646465718746, + 0.01428050547838211, + 0.03073592111468315, + 0.018712425604462624, + 0.014724570326507092, + 0.01125106867402792, + 0.009555891156196594, + 0.005640005227178335, + 0.00721137085929513, + 0.006175949703902006, + 0.0058965315110981464, + 0.007320414762943983, + 0.008902400732040405, + 0.011211195029318333, + 0.01037736888974905, + 0.008288955315947533, + 0.009197200648486614, + 0.01204637996852398, + 0.008135173469781876, + 0.006212437991052866, + 0.007575778756290674, + 0.007592543959617615, + 0.00738696800544858, + 0.005092830862849951, + 0.005298718810081482, + 0.008405073545873165, + 0.006379255559295416, + 0.005574145819991827, + 0.009832915849983692, + 0.011891072615981102, + 0.01551717147231102, + 0.010960865765810013, + 0.009703140705823898, + 0.010811698623001575, + 0.007293435279279947, + 0.008422810584306717, + 0.01144004613161087, + 0.012065094895660877, + 0.019706837832927704, + 0.013579522259533405, + 0.015248189680278301, + 0.02330981194972992, + 0.015681922435760498, + 0.017662661150097847, + 0.02617473714053631, + 0.007050079759210348, + 0.009132620878517628, + 0.009173855185508728, + 0.008448394015431404, + 0.009975136257708073, + 0.007509417366236448, + 0.013927450403571129, + 0.011424773372709751, + 0.006656191311776638, + 0.0077052428387105465, + 0.01331025268882513, + 0.0071740346029400826, + 0.008047685027122498, + 0.011694638058543205, + 0.006495176814496517, + 0.010432234965264797, + 0.005074270535260439, + 0.005007547792047262, + 0.00975488219410181, + 0.006127186585217714, + 0.007821144536137581, + 0.007323523983359337, + 0.00561664579436183, + 0.007049034349620342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.11457546055316925 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048688389360904694 + ], + "opposite_logprob_tokens": [ + -3.0486884117126465 + ], + "logprob_reward": -0.048688389360904694, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048688389360904694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24802391231060028, + 0.18376114964485168, + 0.06043549254536629, + 0.056742459535598755, + 0.05510127544403076, + 0.09116275608539581, + 0.10012021660804749, + 0.04805000498890877, + 0.06403326243162155, + 0.09256953746080399 + ], + "credit_map": [ + 0.033157240599393845, + 0.03625280410051346, + 0.01690089888870716, + 0.03726404532790184, + 0.024968979880213737, + 0.03390657901763916, + 0.02474776655435562, + 0.04082556813955307, + 0.016699638217687607, + 0.01565093919634819, + 0.00919433031231165, + 0.016392091289162636, + 0.027189262211322784, + 0.04352688416838646, + 0.027493951842188835, + 0.027614040300250053, + 0.007903377525508404, + 0.008335345424711704, + 0.00399261387065053, + 0.008895193226635456, + 0.00736673129722476, + 0.007182024419307709, + 0.009232883341610432, + 0.0075273169204592705, + 0.005216320510953665, + 0.008074314333498478, + 0.004582077730447054, + 0.007416234817355871, + 0.006001245696097612, + 0.006012918893247843, + 0.007474612444639206, + 0.011964728124439716, + 0.007198118604719639, + 0.007793644908815622, + 0.0037744459696114063, + 0.004678557626903057, + 0.007538000121712685, + 0.010382206179201603, + 0.003956761211156845, + 0.009779537096619606, + 0.009596970863640308, + 0.016004890203475952, + 0.00759320892393589, + 0.009098975919187069, + 0.012153537012636662, + 0.01416888926178217, + 0.008539741858839989, + 0.014006534591317177, + 0.006791995372623205, + 0.008842631243169308, + 0.01289679016917944, + 0.011874510906636715, + 0.012483199127018452, + 0.010587358847260475, + 0.018300946801900864, + 0.018342776224017143, + 0.003963542636483908, + 0.007658103480935097, + 0.005993128754198551, + 0.006283643189817667, + 0.006838975939899683, + 0.005273426417261362, + 0.005460925400257111, + 0.006578252650797367, + 0.007178948726505041, + 0.005998607259243727, + 0.010544166900217533, + 0.0064710406586527824, + 0.00619289418682456, + 0.011820907704532146, + 0.006130233872681856, + 0.009696458466351032, + 0.009528841823339462, + 0.00978588405996561, + 0.01399949286133051, + 0.00857630930840969, + 0.011714911088347435, + 0.016774306073784828, + 0.0090213967487216, + 0.013168399222195148 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.048688389360904694 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017120479606091976 + ], + "opposite_logprob_tokens": [ + -6.376711845397949 + ], + "logprob_reward": -0.0017120479606091976, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017120479606091976, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16523292660713196, + 0.20531320571899414, + 0.033955659717321396, + 0.054864317178726196, + 0.0278641227632761, + 0.10482917726039886, + 0.15580615401268005, + 0.11370047181844711, + 0.07954517006874084, + 0.05888878554105759 + ], + "credit_map": [ + 0.017382584512233734, + 0.01515730656683445, + 0.015899894759058952, + 0.021941982209682465, + 0.02028738521039486, + 0.03061269037425518, + 0.021591782569885254, + 0.022359300404787064, + 0.0259296502918005, + 0.01875867322087288, + 0.015031146816909313, + 0.022774284705519676, + 0.019740771502256393, + 0.04225178062915802, + 0.03314734995365143, + 0.02767954207956791, + 0.0058755818754434586, + 0.0036207784432917833, + 0.0026731460820883512, + 0.0037997602485120296, + 0.003990111406892538, + 0.00513335969299078, + 0.004955087788403034, + 0.003907833714038134, + 0.006986637599766254, + 0.008520102128386497, + 0.005456401035189629, + 0.00632586469873786, + 0.008960099890828133, + 0.006909514777362347, + 0.004664960782974958, + 0.007040735334157944, + 0.0019040217157453299, + 0.0027345982380211353, + 0.0018074684776365757, + 0.0019601637031883, + 0.0030090820509940386, + 0.007565762847661972, + 0.002123927464708686, + 0.006759100127965212, + 0.011056174524128437, + 0.00759909488260746, + 0.008377961814403534, + 0.007521799765527248, + 0.011485644616186619, + 0.02309343032538891, + 0.014673151075839996, + 0.021021919324994087, + 0.008875325322151184, + 0.01973215863108635, + 0.010994243435561657, + 0.012035390362143517, + 0.022041214630007744, + 0.022767039015889168, + 0.020877845585346222, + 0.03848293796181679, + 0.009905032813549042, + 0.008225714787840843, + 0.008825846947729588, + 0.01141934935003519, + 0.013375622220337391, + 0.01796651817858219, + 0.021652784198522568, + 0.022329606115818024, + 0.007793428376317024, + 0.006469199433922768, + 0.02015216462314129, + 0.007341177202761173, + 0.005992727819830179, + 0.012316443957388401, + 0.00744094280526042, + 0.012039083987474442, + 0.0071221343241631985, + 0.00524023687466979, + 0.007110649719834328, + 0.006209114566445351, + 0.008623397909104824, + 0.00804367195814848, + 0.007191563490778208, + 0.009348020888864994 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0017120479606091976 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10029210150241852 + ], + "opposite_logprob_tokens": [ + -2.350292205810547 + ], + "logprob_reward": -0.10029210150241852, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10029210150241852, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13158346712589264, + 0.14090986549854279, + 0.04063892737030983, + 0.035528991371393204, + 0.02815937250852585, + 0.04040944203734398, + 0.07000596821308136, + 0.22096951305866241, + 0.23405970633029938, + 0.057734668254852295 + ], + "credit_map": [ + 0.024647697806358337, + 0.015127795748412609, + 0.012436515651643276, + 0.019403306767344475, + 0.011583239771425724, + 0.014222036115825176, + 0.015159541741013527, + 0.019003333523869514, + 0.02145293727517128, + 0.014429530128836632, + 0.010191514156758785, + 0.014875038526952267, + 0.0218518003821373, + 0.022385308519005775, + 0.017796950414776802, + 0.01792677864432335, + 0.008765527047216892, + 0.005304075311869383, + 0.0030334959737956524, + 0.005179052706807852, + 0.0032993683125823736, + 0.004306982737034559, + 0.005475358106195927, + 0.005275066010653973, + 0.0031906231306493282, + 0.0050653875805437565, + 0.0024707906413823366, + 0.0041173784993588924, + 0.0075268554501235485, + 0.004925597924739122, + 0.0038326382637023926, + 0.004399718716740608, + 0.0029062568210065365, + 0.002990334527567029, + 0.0022963506635278463, + 0.002797172637656331, + 0.003456280566751957, + 0.004370477516204119, + 0.0029571251943707466, + 0.006385372951626778, + 0.0056228176690638065, + 0.004577561281621456, + 0.0038990082684904337, + 0.004982212092727423, + 0.004158162046223879, + 0.005246833432465792, + 0.00575865525752306, + 0.006164193619042635, + 0.005751258693635464, + 0.007313875015825033, + 0.0060361223295331, + 0.00757821649312973, + 0.005784592591226101, + 0.01018693670630455, + 0.01176914107054472, + 0.015585822984576225, + 0.02181082032620907, + 0.02679324336349964, + 0.02992105856537819, + 0.02691657282412052, + 0.027819596230983734, + 0.02400710992515087, + 0.03153000771999359, + 0.0321710966527462, + 0.02078881673514843, + 0.017548153176903725, + 0.038058262318372726, + 0.02058357745409012, + 0.026979811489582062, + 0.04534032940864563, + 0.028696779161691666, + 0.03606397286057472, + 0.004446579143404961, + 0.005452714394778013, + 0.014082598499953747, + 0.0070390175096690655, + 0.0047084572724998, + 0.005871372297406197, + 0.007193269673734903, + 0.008940660394728184 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 7, + 1 + ], + "selected_reward": -0.10029210150241852 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014193654991686344 + ], + "opposite_logprob_tokens": [ + -4.264193534851074 + ], + "logprob_reward": -0.014193654991686344, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014193654991686344, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13114510476589203, + 0.08962184935808182, + 0.08791212737560272, + 0.05896570533514023, + 0.03952721506357193, + 0.06400247663259506, + 0.042669329792261124, + 0.03155044838786125, + 0.07408443093299866, + 0.380521297454834 + ], + "credit_map": [ + 0.023579256609082222, + 0.013204344548285007, + 0.01268482580780983, + 0.021010365337133408, + 0.010802274569869041, + 0.017674146220088005, + 0.01610148698091507, + 0.01608840748667717, + 0.01079021766781807, + 0.008703308179974556, + 0.006166230887174606, + 0.010459068231284618, + 0.01364356093108654, + 0.015738433226943016, + 0.012317958287894726, + 0.01180307101458311, + 0.017057953402400017, + 0.012179054319858551, + 0.007255787495523691, + 0.009955773130059242, + 0.009678005240857601, + 0.008497712202370167, + 0.0108034061267972, + 0.01248443964868784, + 0.007190716918557882, + 0.005833953153342009, + 0.005321066360920668, + 0.006461544428020716, + 0.008871112018823624, + 0.00580618716776371, + 0.008022492751479149, + 0.011458639986813068, + 0.003469410352408886, + 0.006252385210245848, + 0.0025227125734090805, + 0.0028759906999766827, + 0.005602074787020683, + 0.007930119521915913, + 0.0020504286512732506, + 0.008824094198644161, + 0.006750069558620453, + 0.009977664798498154, + 0.004778832662850618, + 0.004858509637415409, + 0.007173175923526287, + 0.010688254609704018, + 0.005573727190494537, + 0.014202244579792023, + 0.003294551046565175, + 0.005206637550145388, + 0.0029175083618611097, + 0.004213655833154917, + 0.007307983934879303, + 0.00511213019490242, + 0.005736095830798149, + 0.00888077076524496, + 0.0034133829176425934, + 0.0035122970584779978, + 0.002607212634757161, + 0.0036474871449172497, + 0.004789912607520819, + 0.0036398686934262514, + 0.004398740362375975, + 0.005541547201573849, + 0.005692158360034227, + 0.00773610221222043, + 0.014095704071223736, + 0.009277445264160633, + 0.008101696148514748, + 0.011887319386005402, + 0.0065801190212368965, + 0.010713892057538033, + 0.02842933125793934, + 0.0402098149061203, + 0.055948324501514435, + 0.03877090662717819, + 0.058415453881025314, + 0.05357460305094719, + 0.05091771110892296, + 0.054255161434412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -0.014193654991686344 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle, wide-angle shot captures a scene of pure fantasy. The camera slowly and steadily pans to the right, tracking a massive crocodile soaring majestically across a vast, cloudy twilight sky. The horizon, shimmering with deep purple and soft, fiery orange, casts a warm, ethereal glow upward onto the underbelly of billowing, cotton-like clouds. The crocodile, an ancient and immense creature, moves with remarkable grace, its massive body silhouetted against the fading light. Its scales, a mosaic of deep moss green and obsidian black, possess an iridescent quality, shimmering with reflected hues of purple and gold as it catches the last traces of sunlight. Its powerful legs are neatly tucked at its sides, while its mighty, ridged tail ripples in a slow, deliberate, and powerful rhythm, propelling it forward as if swimming through dense air. A wisp of cloud drifts past its long snout and ancient, wise eyes, momentarily blurring before revealing its immense size, creating an atmosphere of tranquility, awe, and surreal majesty.", + "seed": 15916, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_4311_seed_15916_dp_29_rank_58_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2855199329555035, + "vlm_logprob": -0.2855199329555035, + "vlm_yes_no_margin": 2.974999952316284, + "vlm_token_credit": -0.2855199329555035, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.3141506910324097 + ], + "opposite_logprob_tokens": [ + -0.31415069103240967 + ], + "logprob_reward": -1.3141506910324097, + "yes_no_margin_reward": -1.0, + "token_credit_reward": -1.3141506910324097, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.41185638308525085, + 0.3753264844417572, + 0.05913211405277252, + 0.03455154225230217, + 0.01670435070991516, + 0.00970681942999363, + 0.014235301874577999, + 0.005416734144091606, + 0.007212599739432335, + 0.06585769355297089 + ], + "credit_map": [ + 0.051093343645334244, + 0.03702384978532791, + 0.040047939866781235, + 0.03909335657954216, + 0.038173917680978775, + 0.03982704132795334, + 0.060391563922166824, + 0.10620536655187607, + 0.059738509356975555, + 0.0553940050303936, + 0.03618473932147026, + 0.031994663178920746, + 0.03662805259227753, + 0.06797442585229874, + 0.04533329978585243, + 0.04207876697182655, + 0.008507450111210346, + 0.005626056808978319, + 0.00445179408416152, + 0.00646609952673316, + 0.008949415758252144, + 0.008982832543551922, + 0.008349969983100891, + 0.007798496168106794, + 0.006757188588380814, + 0.0034971318673342466, + 0.0022446601651608944, + 0.002999644260853529, + 0.0035412912257015705, + 0.003799589117988944, + 0.00627320958301425, + 0.005438825581222773, + 0.0023170970380306244, + 0.0016615622444078326, + 0.0024937535636126995, + 0.002032721880823374, + 0.002110597910359502, + 0.001693270867690444, + 0.0024528740905225277, + 0.001942471950314939, + 0.0014556687092408538, + 0.0009881507139652967, + 0.001301006879657507, + 0.001088238786906004, + 0.0012196846073493361, + 0.001010889420285821, + 0.0013263046275824308, + 0.00131687568500638, + 0.0014263119082897902, + 0.0013195124920457602, + 0.0012500965967774391, + 0.003004638710990548, + 0.0013597042998299003, + 0.0012791284825652838, + 0.0027065176982432604, + 0.0018893913365900517, + 0.0007114988402463496, + 0.000601748179178685, + 0.0005714779836125672, + 0.0009330098982900381, + 0.00046309849130921066, + 0.0005671469843946397, + 0.0008556987741030753, + 0.0007130551966838539, + 0.0008792247390374541, + 0.0010850591352209449, + 0.001076195389032364, + 0.0010166444117203355, + 0.0006494535482488573, + 0.0005389883299358189, + 0.0008153988746926188, + 0.0011516354279592633, + 0.007616846822202206, + 0.005916332360357046, + 0.012509731575846672, + 0.007140514440834522, + 0.0056344070471823215, + 0.006086548790335655, + 0.011805543676018715, + 0.009147771634161472 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -1.3141506910324097 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020611735060811043 + ], + "opposite_logprob_tokens": [ + -3.8956117630004883 + ], + "logprob_reward": -0.020611735060811043, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020611735060811043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3996393084526062, + 0.3764638602733612, + 0.03737394139170647, + 0.035865042358636856, + 0.02314329333603382, + 0.018884193152189255, + 0.015993714332580566, + 0.008603458292782307, + 0.015765732154250145, + 0.06826745718717575 + ], + "credit_map": [ + 0.050381068140268326, + 0.03585461527109146, + 0.03921627998352051, + 0.03791426494717598, + 0.036537349224090576, + 0.038672786206007004, + 0.05871457979083061, + 0.10234833508729935, + 0.060567911714315414, + 0.05505429580807686, + 0.036395493894815445, + 0.0316927433013916, + 0.038068901747465134, + 0.06763250380754471, + 0.04539065062999725, + 0.04166138544678688, + 0.0045696753077209, + 0.002791224978864193, + 0.0028591598384082317, + 0.003659905167296529, + 0.0057334983721375465, + 0.0071422141045331955, + 0.005516104865819216, + 0.005102163180708885, + 0.0069566043093800545, + 0.003883203724399209, + 0.0024686595425009727, + 0.003464331617578864, + 0.003711580764502287, + 0.004164956044405699, + 0.005898418370634317, + 0.005317289847880602, + 0.0031891455873847008, + 0.002246920485049486, + 0.0034647793509066105, + 0.0034578933846205473, + 0.0032110942993313074, + 0.0026312123518437147, + 0.002966703148558736, + 0.0019755454268306494, + 0.0016886808443814516, + 0.0019838165026158094, + 0.0021196245215833187, + 0.002188794082030654, + 0.0027070697396993637, + 0.00216305791400373, + 0.0032998735550791025, + 0.0027332762256264687, + 0.001501271384768188, + 0.0012949148658663034, + 0.0014321719063445926, + 0.0027067800983786583, + 0.0019690042827278376, + 0.00194340490270406, + 0.0031686085276305676, + 0.0019775573164224625, + 0.001106312032788992, + 0.0007963576936163008, + 0.0008650369709357619, + 0.0018580796895548701, + 0.0006111838156357408, + 0.0008359347702935338, + 0.0014506627339869738, + 0.0010798906441777945, + 0.0015915868571028113, + 0.001601513708010316, + 0.0016687437891960144, + 0.0021720859222114086, + 0.0011999732814729214, + 0.0010566398268565536, + 0.0022332780063152313, + 0.0042419093661010265, + 0.008339812979102135, + 0.006890767719596624, + 0.011426903307437897, + 0.0078102173283696175, + 0.005839481949806213, + 0.00559283047914505, + 0.012965325266122818, + 0.009402123279869556 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.020611735060811043 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004637913778424263 + ], + "opposite_logprob_tokens": [ + -5.379637718200684 + ], + "logprob_reward": -0.004637913778424263, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004637913778424263, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.39894527196884155, + 0.37114009261131287, + 0.05650228261947632, + 0.037690747529268265, + 0.01861598715186119, + 0.013649477623403072, + 0.015474226325750351, + 0.020827043801546097, + 0.007114688400179148, + 0.060040183365345 + ], + "credit_map": [ + 0.050247129052877426, + 0.035950466990470886, + 0.03800499066710472, + 0.037445683032274246, + 0.03600688278675079, + 0.03925009444355965, + 0.058720167726278305, + 0.10331983864307404, + 0.06019432097673416, + 0.05473092943429947, + 0.036416567862033844, + 0.031624685972929, + 0.0361684150993824, + 0.06557907909154892, + 0.04518380016088486, + 0.041242264211177826, + 0.008777898736298084, + 0.005295191891491413, + 0.004112713970243931, + 0.005751632619649172, + 0.008622872643172741, + 0.008411332033574581, + 0.007795719429850578, + 0.007734912447631359, + 0.007802486885339022, + 0.0037786313332617283, + 0.0022273187059909105, + 0.003094483632594347, + 0.0038916775956749916, + 0.00375291146337986, + 0.007316768169403076, + 0.005826467182487249, + 0.0025610330048948526, + 0.0017592201475054026, + 0.003016735427081585, + 0.002599585335701704, + 0.00252679456025362, + 0.001930175581946969, + 0.0023806181270629168, + 0.0018418229883536696, + 0.001898498390801251, + 0.0013938258634880185, + 0.001787229790352285, + 0.0013977994676679373, + 0.0017440860392525792, + 0.0016306376783177257, + 0.0020177687983959913, + 0.001779629965312779, + 0.0016962637891992927, + 0.0012655244208872318, + 0.0014272368280217052, + 0.0034094108268618584, + 0.00155793281737715, + 0.00136606115847826, + 0.002713845344260335, + 0.0020379505585879087, + 0.002382125938311219, + 0.0018946926575154066, + 0.0018150961259379983, + 0.005001068115234375, + 0.001176429446786642, + 0.001468199072405696, + 0.0032807285897433758, + 0.003808703040704131, + 0.0012946234783157706, + 0.0009235550533048809, + 0.00104500912129879, + 0.0008521572453901172, + 0.0006727568688802421, + 0.000666382780764252, + 0.0007255900418385863, + 0.0009346128208562732, + 0.008108079433441162, + 0.004454092122614384, + 0.005630244500935078, + 0.008235628716647625, + 0.005006494466215372, + 0.00434927549213171, + 0.014897875487804413, + 0.009358486160635948 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 9 + ], + "selected_reward": -0.004637913778424263 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001014798879623 + ], + "opposite_logprob_tokens": [ + -2.695010185241699 + ], + "logprob_reward": -0.07001014798879623, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001014798879623, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1961641162633896, + 0.33828410506248474, + 0.040827441960573196, + 0.03183796629309654, + 0.016280794516205788, + 0.03343087434768677, + 0.1303645521402359, + 0.017714746296405792, + 0.04966426268219948, + 0.14543116092681885 + ], + "credit_map": [ + 0.024032503366470337, + 0.01560670230537653, + 0.01998971775174141, + 0.017233021557331085, + 0.01933538354933262, + 0.019628968089818954, + 0.029834864661097527, + 0.0505029521882534, + 0.05142349749803543, + 0.05021018162369728, + 0.03284081071615219, + 0.029509825631976128, + 0.031065845862030983, + 0.06387827545404434, + 0.041033025830984116, + 0.038322627544403076, + 0.006824476644396782, + 0.004265394993126392, + 0.0032303587067872286, + 0.00539487274363637, + 0.005694622173905373, + 0.004224963951855898, + 0.0067468793131411076, + 0.004445874132215977, + 0.0051185935735702515, + 0.0031583155505359173, + 0.0022319809067994356, + 0.002894981298595667, + 0.0035094365011900663, + 0.004133402369916439, + 0.005342849995940924, + 0.005448406096547842, + 0.0022106196265667677, + 0.0014483697013929486, + 0.002336654346436262, + 0.0018580943578854203, + 0.00247146375477314, + 0.0027294440660625696, + 0.0016013684216886759, + 0.0016247790772467852, + 0.005206911358982325, + 0.0037459346931427717, + 0.0037728429306298494, + 0.0032608825713396072, + 0.003940149210393429, + 0.003223836189135909, + 0.005333309061825275, + 0.004947008565068245, + 0.009796940721571445, + 0.012081126682460308, + 0.011025415733456612, + 0.020550964400172234, + 0.014739357866346836, + 0.015175249427556992, + 0.030924171209335327, + 0.016071315854787827, + 0.002038542414084077, + 0.0017659299774095416, + 0.0017689740052446723, + 0.0031220526434481144, + 0.0013191481120884418, + 0.0016333376988768578, + 0.003165960544720292, + 0.002900799037888646, + 0.005846649408340454, + 0.006454152520745993, + 0.008565494790673256, + 0.006218727212399244, + 0.004673774354159832, + 0.003930556587874889, + 0.006601525470614433, + 0.007373384665697813, + 0.019947310909628868, + 0.016700146719813347, + 0.018608910962939262, + 0.03067099116742611, + 0.010823633521795273, + 0.011660445481538773, + 0.021142199635505676, + 0.01587751880288124 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 9 + ], + "selected_reward": -0.07001014798879623 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818917691707611 + ], + "opposite_logprob_tokens": [ + -4.018188953399658 + ], + "logprob_reward": -0.01818917691707611, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.01818917691707611, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33599162101745605, + 0.4161463677883148, + 0.05183497443795204, + 0.03024054318666458, + 0.017664384096860886, + 0.02157398872077465, + 0.049311406910419464, + 0.014305786229670048, + 0.017638670280575752, + 0.04529227688908577 + ], + "credit_map": [ + 0.04268290102481842, + 0.030251363292336464, + 0.031759537756443024, + 0.032036203891038895, + 0.030224554240703583, + 0.033640071749687195, + 0.048594165593385696, + 0.08680283278226852, + 0.0665786936879158, + 0.06118801236152649, + 0.041427914053201675, + 0.03552231565117836, + 0.039962150156497955, + 0.07501835376024246, + 0.0502229668200016, + 0.04622596502304077, + 0.006127229426056147, + 0.005642220377922058, + 0.0033648721873760223, + 0.006138427648693323, + 0.007411088794469833, + 0.008510949090123177, + 0.007191690616309643, + 0.007448498159646988, + 0.005731448996812105, + 0.003354528220370412, + 0.002133795525878668, + 0.0029544455464929342, + 0.003124901559203863, + 0.003449534997344017, + 0.0053323679603636265, + 0.004159518983215094, + 0.002431373344734311, + 0.0017725243233144283, + 0.0020970613695681095, + 0.0022616058122366667, + 0.002799596404656768, + 0.002093362854793668, + 0.0024931728839874268, + 0.001715688849799335, + 0.002686752937734127, + 0.002713264198973775, + 0.003036605892702937, + 0.0024691547732800245, + 0.002801027148962021, + 0.0027345591224730015, + 0.0028833237010985613, + 0.0022493002470582724, + 0.0032783409114927053, + 0.004010084550827742, + 0.004593828693032265, + 0.008021908812224865, + 0.004164506681263447, + 0.006173974368721247, + 0.01256492268294096, + 0.0065038385801017284, + 0.0018951304955407977, + 0.0013894543517380953, + 0.0010872877901419997, + 0.002824644558131695, + 0.001175292651169002, + 0.0012816826347261667, + 0.002334575867280364, + 0.002317718230187893, + 0.0020942860282957554, + 0.0019426709040999413, + 0.0022782788146287203, + 0.0017265543574467301, + 0.0014540571719408035, + 0.001772445160895586, + 0.002075138036161661, + 0.004295239225029945, + 0.005767519120126963, + 0.0030862330459058285, + 0.006180709693580866, + 0.006929471157491207, + 0.0034897516015917063, + 0.0035628818441182375, + 0.009424572810530663, + 0.0068511394783854485 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.01818917691707611 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "stunning low-angle, wide-angle shot captures a scene of pure fantasy. The camera slowly and steadily pans to the right, tracking a massive crocodile soaring majestically across a vast, cloudy twilight sky. The horizon, shimmering with deep purple and soft, fiery orange, casts a warm, ethereal glow upward onto the underbelly of billowing, cotton-like clouds. The crocodile, an ancient and immense creature, moves with remarkable grace, its massive body silhouetted against the fading light. Its scales, a mosaic of deep moss green and obsidian black, possess an iridescent quality, shimmering with reflected hues of purple and gold as it catches the last traces of sunlight. Its powerful legs are neatly tucked at its sides, while its mighty, ridged tail ripples in a slow, deliberate, and powerful rhythm, propelling it forward as if swimming through dense air. A wisp of cloud drifts past its long snout and ancient, wise eyes, momentarily blurring before revealing its immense size, creating an atmosphere of tranquility, awe, and surreal majesty.", + "seed": 15917, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_14/index_4311_seed_15917_dp_29_rank_58_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 14, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.05942994842771441, + "vlm_logprob": -0.05942994842771441, + "vlm_yes_no_margin": 4.2, + "vlm_token_credit": -0.05942994842771441, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.015710627660155296 + ], + "opposite_logprob_tokens": [ + -4.265710830688477 + ], + "logprob_reward": -0.015710627660155296, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.015710627660155296, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.8318668007850647, + 0.050343506038188934, + 0.017730440944433212, + 0.00812153797596693, + 0.014703105203807354, + 0.007292331196367741, + 0.044482141733169556, + 0.00914648175239563, + 0.011874180287122726, + 0.004439455457031727 + ], + "credit_map": [ + 0.06323575228452682, + 0.07925423979759216, + 0.035584498196840286, + 0.03346315771341324, + 0.3586381673812866, + 0.19895507395267487, + 0.03287332504987717, + 0.029862554743885994, + 0.00944200623780489, + 0.0025391310919076204, + 0.022236449643969536, + 0.0031251832842826843, + 0.005550289526581764, + 0.00299970805644989, + 0.0020538093522191048, + 0.0023969290778040886, + 0.0026705837808549404, + 0.0018422267166897655, + 0.0023736925795674324, + 0.002542313188314438, + 0.0015712665626779199, + 0.0018995690625160933, + 0.0016737430123612285, + 0.0031570461578667164, + 0.0008629983058199286, + 0.0009492186945863068, + 0.0018732603639364243, + 0.001333664869889617, + 0.0007554059266112745, + 0.0009457163396291435, + 0.0007430596160702407, + 0.0006582134519703686, + 0.002457789611071348, + 0.00238855485804379, + 0.002364563988521695, + 0.0023684073239564896, + 0.0012435449752956629, + 0.0011765449307858944, + 0.0015012469375506043, + 0.0012024511815980077, + 0.0007681841962039471, + 0.0012253663735464215, + 0.0013825668720528483, + 0.0012665652902796865, + 0.0006094275740906596, + 0.00039294600719586015, + 0.0007750887889415026, + 0.0008721862686797976, + 0.009894877672195435, + 0.005070190876722336, + 0.006734258495271206, + 0.00751569913700223, + 0.002723198151215911, + 0.0023936787620186806, + 0.006081307306885719, + 0.004068932496011257, + 0.0009265885455533862, + 0.0019758837297558784, + 0.0011420020600780845, + 0.0012177280150353909, + 0.0012782772537320852, + 0.0006486832862719893, + 0.0011486347066238523, + 0.0008086839807219803, + 0.0012322727125138044, + 0.0014797531766816974, + 0.0016502694925293326, + 0.0013228171737864614, + 0.001144828856922686, + 0.0009932384127750993, + 0.001751045580022037, + 0.002299955114722252, + 0.0005335651221685112, + 0.0005435294006019831, + 0.00048085409798659384, + 0.000745305442251265, + 0.000457645277492702, + 0.00036496343091130257, + 0.0008156500989571214, + 0.0004979430814273655 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.015710627660155296 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020588846877217293 + ], + "opposite_logprob_tokens": [ + -3.8955888748168945 + ], + "logprob_reward": -0.020588846877217293, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020588846877217293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5134965777397156, + 0.26330357789993286, + 0.03071707673370838, + 0.0358850322663784, + 0.010659645311534405, + 0.007465210743248463, + 0.0898929238319397, + 0.021187497302889824, + 0.01548211369663477, + 0.011910396628081799 + ], + "credit_map": [ + 0.03816600888967514, + 0.049648527055978775, + 0.022609522566199303, + 0.021338481456041336, + 0.21809439361095428, + 0.12341495603322983, + 0.02070918120443821, + 0.01951543428003788, + 0.037883657962083817, + 0.013885878957808018, + 0.11457280814647675, + 0.0187296774238348, + 0.018696196377277374, + 0.03242569416761398, + 0.010716980323195457, + 0.016392679885029793, + 0.0036093369126319885, + 0.002763498807325959, + 0.004482882097363472, + 0.00526861846446991, + 0.002435932867228985, + 0.003261141711845994, + 0.002720150165259838, + 0.006175509188324213, + 0.00511149363592267, + 0.005396721418946981, + 0.00531628867611289, + 0.0068162232637405396, + 0.0028853672556579113, + 0.0033489747438579798, + 0.003772669704630971, + 0.0032372884452342987, + 0.001258997479453683, + 0.0018635570304468274, + 0.0017428171122446656, + 0.0017288012895733118, + 0.0009830717463046312, + 0.0007046495447866619, + 0.0010156257776543498, + 0.0013621235266327858, + 0.0011304891668260098, + 0.001244380953721702, + 0.001011862768791616, + 0.0010794034460559487, + 0.0009366602171212435, + 0.0004128425207454711, + 0.0008046702132560313, + 0.000844900612719357, + 0.014374449849128723, + 0.006938226521015167, + 0.011833724565804005, + 0.017491808161139488, + 0.006120417267084122, + 0.004211453255265951, + 0.022842930629849434, + 0.006079905200749636, + 0.0018877455731853843, + 0.0037622645031660795, + 0.0024522230960428715, + 0.00332955876365304, + 0.002899756422266364, + 0.0018244409002363682, + 0.0023740562610328197, + 0.0026574493385851383, + 0.0017188799101859331, + 0.0018779693637043238, + 0.0021652637515217066, + 0.0030958561692386866, + 0.002091265982016921, + 0.0009758209343999624, + 0.0018124254420399666, + 0.0017446308629587293, + 0.0012621319619938731, + 0.0017545429291203618, + 0.0013862993801012635, + 0.0017915416974574327, + 0.001819817814975977, + 0.0010248133912682533, + 0.001867836806923151, + 0.0010034112492576241 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.020588846877217293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001178047270514071 + ], + "opposite_logprob_tokens": [ + -6.75117826461792 + ], + "logprob_reward": -0.001178047270514071, + "yes_no_margin_reward": 6.75, + "token_credit_reward": -0.001178047270514071, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.639901876449585, + 0.17554011940956116, + 0.033569030463695526, + 0.031047150492668152, + 0.010779591277241707, + 0.011332210153341293, + 0.05254187807440758, + 0.02128784917294979, + 0.013077287003397942, + 0.010922981426119804 + ], + "credit_map": [ + 0.045757558196783066, + 0.060964349657297134, + 0.026936938986182213, + 0.026918567717075348, + 0.2702081501483917, + 0.1589786857366562, + 0.025694040581583977, + 0.024443604052066803, + 0.02360973320901394, + 0.009820829145610332, + 0.07917853444814682, + 0.012846771627664566, + 0.012665915302932262, + 0.023260129615664482, + 0.005678145680576563, + 0.008480047807097435, + 0.0049582901410758495, + 0.0035534012131392956, + 0.009611230343580246, + 0.0037148932460695505, + 0.002185867168009281, + 0.0040320442058146, + 0.002784731099382043, + 0.0027285709511488676, + 0.0031612615566700697, + 0.007223495747894049, + 0.005668409168720245, + 0.004075393080711365, + 0.0025364519096910954, + 0.0026753884740173817, + 0.0032043291721493006, + 0.002502419985830784, + 0.0019714473746716976, + 0.0016170963644981384, + 0.0018726831767708063, + 0.0017101347912102938, + 0.0010040417546406388, + 0.0009047610219568014, + 0.0008831906015984714, + 0.0008162345038726926, + 0.0011216506827622652, + 0.002164852572605014, + 0.001908057020045817, + 0.0017628668574616313, + 0.0014799721539020538, + 0.0007283536833710968, + 0.0010729263303801417, + 0.0010935293976217508, + 0.0065613980405032635, + 0.0045371511951088905, + 0.007361625321209431, + 0.010471174493432045, + 0.003532606177031994, + 0.0025726386811584234, + 0.012679795734584332, + 0.004825487732887268, + 0.0016444248612970114, + 0.003980774898082018, + 0.0026166390161961317, + 0.0037984985392540693, + 0.0025473625864833593, + 0.0014778495533391833, + 0.002828299766406417, + 0.0023940003011375666, + 0.0015303128166124225, + 0.0016169934533536434, + 0.0021859577391296625, + 0.0020721114706248045, + 0.0012967728544026613, + 0.0013090481515973806, + 0.0015285544795915484, + 0.0015375349903479218, + 0.0011947795283049345, + 0.001121102599427104, + 0.0014625475741922855, + 0.0016449015820398927, + 0.001292807748541236, + 0.0011218684958294034, + 0.0015686911065131426, + 0.0015162817435339093 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.001178047270514071 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520463466644287 + ], + "opposite_logprob_tokens": [ + -1.5020463466644287 + ], + "logprob_reward": -0.2520463466644287, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520463466644287, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6640161275863647, + 0.12406610697507858, + 0.05467434599995613, + 0.02342929132282734, + 0.029316183179616928, + 0.010354937054216862, + 0.01840863563120365, + 0.03224891051650047, + 0.02658964693546295, + 0.01689586415886879 + ], + "credit_map": [ + 0.04554356262087822, + 0.06340494751930237, + 0.034804053604602814, + 0.027711445465683937, + 0.2675875127315521, + 0.16943904757499695, + 0.02568216808140278, + 0.029843349009752274, + 0.0285758376121521, + 0.004937986843287945, + 0.041609615087509155, + 0.006618422456085682, + 0.016690904274582863, + 0.010450109839439392, + 0.005836009047925472, + 0.009347212500870228, + 0.006521590519696474, + 0.005286698695272207, + 0.014579283073544502, + 0.008739710785448551, + 0.004159325268119574, + 0.004605486057698727, + 0.005110828205943108, + 0.005671418737620115, + 0.00319559033960104, + 0.0021763283293694258, + 0.006565258372575045, + 0.004115202929824591, + 0.0016841990873217583, + 0.002496647648513317, + 0.0017558374674990773, + 0.001440227497369051, + 0.004376496188342571, + 0.006634836550801992, + 0.004493875429034233, + 0.003188869683071971, + 0.0036713650915771723, + 0.002178862923756242, + 0.0025132973678410053, + 0.0022585755214095116, + 0.001598495990037918, + 0.0018504421459510922, + 0.0014299077447503805, + 0.001299991738051176, + 0.0010337646817788482, + 0.0007698462577536702, + 0.001207850407809019, + 0.0011646372731775045, + 0.00279616960324347, + 0.0019524359377101064, + 0.0026013399474322796, + 0.0033521486911922693, + 0.0012429263442754745, + 0.0012161819031462073, + 0.0036137260030955076, + 0.0016337046399712563, + 0.002587398746982217, + 0.006105148699134588, + 0.004204052500426769, + 0.006192989181727171, + 0.004425297025591135, + 0.0019482284551486373, + 0.0036912374198436737, + 0.003094554878771305, + 0.0025698801036924124, + 0.0032026164699345827, + 0.004772122483700514, + 0.003945052623748779, + 0.003390156663954258, + 0.002041900297626853, + 0.0038069547154009342, + 0.002860961016267538, + 0.0017119062831625342, + 0.001959382789209485, + 0.0024829318281263113, + 0.003018669318407774, + 0.0017579469131305814, + 0.0011951979249715805, + 0.0025113821029663086, + 0.0022584456019103527 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.2520463466644287 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive crocodile soaring through the sky?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the crocodile's scales a mosaic of deep moss green and obsidian black with iridescent purple and gold reflections?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast cloudy twilight sky with a horizon shimmering in deep purple and soft fiery orange?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the camera slowly and steadily pan to the right while tracking the crocodile?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the crocodile's body silhouetted against the fading light with its legs tucked and tail rippling in a slow rhythm?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007625873666256666 + ], + "opposite_logprob_tokens": [ + -4.882626056671143 + ], + "logprob_reward": -0.007625873666256666, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007625873666256666, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.709599494934082, + 0.08795277774333954, + 0.013467355631291866, + 0.02735288254916668, + 0.009633788838982582, + 0.013402914628386497, + 0.06861010938882828, + 0.03306170925498009, + 0.028887687250971794, + 0.008031352423131466 + ], + "credit_map": [ + 0.049833592027425766, + 0.06663568317890167, + 0.03301867097616196, + 0.028407888486981392, + 0.30191898345947266, + 0.17477135360240936, + 0.02769896760582924, + 0.027314385399222374, + 0.017253300175070763, + 0.0043182517401874065, + 0.037877775728702545, + 0.005417836830019951, + 0.00950750894844532, + 0.005110032390803099, + 0.003699116175994277, + 0.004768957383930683, + 0.0018348207231611013, + 0.0015213171718642116, + 0.0017362931976094842, + 0.0016854563727974892, + 0.0011468231678009033, + 0.0018990188837051392, + 0.0015511451056227088, + 0.0020924804266542196, + 0.0036011431366205215, + 0.004328841809183359, + 0.0076012578792870045, + 0.0033540453296154737, + 0.002233704784885049, + 0.0020595237147063017, + 0.0021667794790118933, + 0.0020075845532119274, + 0.0017619464779272676, + 0.0019205262651666999, + 0.001523862243629992, + 0.0013874220894649625, + 0.000836273655295372, + 0.000759426096919924, + 0.0007954049506224692, + 0.0006489266525022686, + 0.0013975793262943625, + 0.0028222822584211826, + 0.0020789855625480413, + 0.00208844430744648, + 0.0013873602729290724, + 0.0008701428887434304, + 0.00130972929764539, + 0.0014483905397355556, + 0.010935410857200623, + 0.006616543047130108, + 0.00981744285672903, + 0.012358330190181732, + 0.005046416539698839, + 0.003699292428791523, + 0.013617615215480328, + 0.00651905732229352, + 0.0024762952234596014, + 0.006053324788808823, + 0.0043304478749632835, + 0.006653930991888046, + 0.003928066231310368, + 0.0022804858162999153, + 0.004028563853353262, + 0.003310598200187087, + 0.0037666456773877144, + 0.004228604957461357, + 0.00361255113966763, + 0.005351576954126358, + 0.003141916124150157, + 0.00165925535839051, + 0.002997961826622486, + 0.004129176493734121, + 0.0008557779947295785, + 0.0008917119121178985, + 0.0010009771212935448, + 0.0012243740493431687, + 0.000987463048659265, + 0.0005943345022387803, + 0.0013524579117074609, + 0.0011242554755881429 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.007625873666256666 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_tensor([4835, 4835])_group_15_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_tensor([4835, 4835])_group_15_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..e8d60c698aa76b4a9f8717473fbab1d34b8e074c --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_tensor([4835, 4835])_group_15_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "In a single panoramic shot, the camera maintains an eye-level perspective and pans smoothly to the right, perfectly tracking a lone, majestic giraffe across a vast savannah landscape. The giraffe's coat is a stunning mosaic of large, polygonal patches of brown and orange, separated by intricate cream-colored lines. It moves from the left to the right side of the frame with a deliberate and steady grace. Its long legs rise and fall with a unique, rhythmic cadence, creating a fluid, continuous stroll that feels both powerful and tranquil. The setting is a vast expanse of dry, golden-brown grassland, sparsely dotted with low-lying green shrubs and the iconic flat-topped acacia trees. In the distance, a range of hazy, blue mountains rises against the horizon, and a vast, overcast sky is filled with soft, white clouds. The entire scene is bathed in a soft, diffuse light that eliminates harsh shadows and enhances the earthy color palette, creating an atmosphere of profound tranquility and naturalness, highlighting the creature's peaceful, solitary journey through its vast habitat.", + "seed": 11889, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_4835_seed_11889_dp_30_rank_60_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.1306413757847622, + "vlm_logprob": -0.1306413757847622, + "vlm_yes_no_margin": 4.025, + "vlm_token_credit": -0.1306413757847622, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00465986505150795 + ], + "opposite_logprob_tokens": [ + -6.879659652709961 + ], + "logprob_reward": -0.00465986505150795, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.00465986505150795, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19104348123073578, + 0.07786200195550919, + 0.044523686170578, + 0.035847652703523636, + 0.09205469489097595, + 0.0703582689166069, + 0.10082080215215683, + 0.12571898102760315, + 0.17460916936397552, + 0.08716118335723877 + ], + "credit_map": [ + 0.03188519924879074, + 0.012019936926662922, + 0.03133254125714302, + 0.016985559836030006, + 0.015535527840256691, + 0.04281722381711006, + 0.024600675329566002, + 0.015866853296756744, + 0.00974329374730587, + 0.010714776813983917, + 0.008742368780076504, + 0.008730566129088402, + 0.00788871943950653, + 0.012333608232438564, + 0.009694753214716911, + 0.010013923980295658, + 0.006626847665756941, + 0.0032906520646065474, + 0.006189865991473198, + 0.004590065684169531, + 0.003658401546999812, + 0.006239669397473335, + 0.006197876762598753, + 0.007730310782790184, + 0.004821178503334522, + 0.0031149378046393394, + 0.00416967086493969, + 0.005704707466065884, + 0.004251911770552397, + 0.0051320274360477924, + 0.004083176143467426, + 0.00457004364579916, + 0.012745682150125504, + 0.006767896469682455, + 0.017188865691423416, + 0.01165145542472601, + 0.009282851591706276, + 0.017232606187462807, + 0.00783559586852789, + 0.009349743835628033, + 0.01004976499825716, + 0.006850685924291611, + 0.012566708028316498, + 0.010869350284337997, + 0.009560671634972095, + 0.007568479515612125, + 0.00558658828958869, + 0.007306030951440334, + 0.012586405500769615, + 0.0069043198600411415, + 0.013872412033379078, + 0.013696849346160889, + 0.007488996256142855, + 0.010168630629777908, + 0.015268071554601192, + 0.020835118368268013, + 0.016817225143313408, + 0.010974445380270481, + 0.014474762603640556, + 0.01514685619622469, + 0.013764946721494198, + 0.013695360161364079, + 0.01748502068221569, + 0.02336038649082184, + 0.017663566395640373, + 0.008712338283658028, + 0.029099134728312492, + 0.013832177966833115, + 0.01632920280098915, + 0.023841198533773422, + 0.031382061541080475, + 0.03374951332807541, + 0.009566864930093288, + 0.0058660320937633514, + 0.012266959995031357, + 0.0077625056728720665, + 0.008799353614449501, + 0.010387079790234566, + 0.011383376084268093, + 0.021129021421074867 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.00465986505150795 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.38712871074676514 + ], + "opposite_logprob_tokens": [ + -1.1371287107467651 + ], + "logprob_reward": -0.38712871074676514, + "yes_no_margin_reward": 0.75, + "token_credit_reward": -0.38712871074676514, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05982667952775955, + 0.04256123676896095, + 0.04387890174984932, + 0.06062978133559227, + 0.17864558100700378, + 0.047630585730075836, + 0.13712741434574127, + 0.06672653555870056, + 0.1694597601890564, + 0.19351348280906677 + ], + "credit_map": [ + 0.012086194939911366, + 0.004016355611383915, + 0.005802684463560581, + 0.006527918856590986, + 0.00650940602645278, + 0.011395241133868694, + 0.005890293046832085, + 0.007598583120852709, + 0.005361513700336218, + 0.004771215841174126, + 0.003901130985468626, + 0.0063759624026715755, + 0.005847974680364132, + 0.007497479673475027, + 0.0038520884700119495, + 0.004953868221491575, + 0.005553871858865023, + 0.003512590890750289, + 0.008796083740890026, + 0.00673306081444025, + 0.005991219077259302, + 0.0039512403309345245, + 0.00439539784565568, + 0.004945435095578432, + 0.009487682022154331, + 0.005072320345789194, + 0.0068412600085139275, + 0.010126166976988316, + 0.008350729010999203, + 0.006892667151987553, + 0.006512303836643696, + 0.007346651051193476, + 0.022792551666498184, + 0.01127159409224987, + 0.03498706594109535, + 0.02570478618144989, + 0.018129825592041016, + 0.034314919263124466, + 0.014879165217280388, + 0.016565673053264618, + 0.005187785718590021, + 0.0039039929397404194, + 0.009000319056212902, + 0.00867913756519556, + 0.0061500235460698605, + 0.005978757049888372, + 0.0038724890910089016, + 0.004858081229031086, + 0.015030712820589542, + 0.013314181007444859, + 0.02066868543624878, + 0.027386823669075966, + 0.013629128225147724, + 0.016385188326239586, + 0.01373964175581932, + 0.016973057761788368, + 0.007875010371208191, + 0.006393202114850283, + 0.012955832295119762, + 0.012451614253222942, + 0.007094938773661852, + 0.005136820022016764, + 0.006662276573479176, + 0.008156838826835155, + 0.01983022689819336, + 0.010355092585086823, + 0.025359777733683586, + 0.012938665226101875, + 0.017786089330911636, + 0.017570821568369865, + 0.0311457309871912, + 0.03447335585951805, + 0.02351303957402706, + 0.011311942711472511, + 0.028902968391776085, + 0.018208712339401245, + 0.017325324937701225, + 0.028698744252324104, + 0.024867799133062363, + 0.040684960782527924 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 4, + 8 + ], + "selected_reward": -0.38712871074676514 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017106198938563466 + ], + "opposite_logprob_tokens": [ + -6.376710414886475 + ], + "logprob_reward": -0.0017106198938563466, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017106198938563466, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27527284622192383, + 0.10465486347675323, + 0.10037992894649506, + 0.0738287940621376, + 0.04800800234079361, + 0.05429806187748909, + 0.06962036341428757, + 0.10277453064918518, + 0.12090256065130234, + 0.05026007071137428 + ], + "credit_map": [ + 0.026907585561275482, + 0.013689335435628891, + 0.029079874977469444, + 0.01839853636920452, + 0.021196749061346054, + 0.055791400372982025, + 0.034784018993377686, + 0.07542532682418823, + 0.013263157568871975, + 0.007840655744075775, + 0.015991825610399246, + 0.011092981323599815, + 0.00878816470503807, + 0.015378016978502274, + 0.013882231898605824, + 0.01841782219707966, + 0.012171050533652306, + 0.009352153167128563, + 0.014800751581788063, + 0.010103132575750351, + 0.008797296322882175, + 0.013774862512946129, + 0.011762099340558052, + 0.019618572667241096, + 0.015069524757564068, + 0.0042676301673054695, + 0.006651025265455246, + 0.005441707093268633, + 0.006712424103170633, + 0.020270273089408875, + 0.005373260471969843, + 0.010042952373623848, + 0.005940363742411137, + 0.004016004968434572, + 0.006021294742822647, + 0.00743188988417387, + 0.0049821725115180016, + 0.006751230452209711, + 0.006142179016023874, + 0.006722862366586924, + 0.009847018867731094, + 0.004897221922874451, + 0.00560436537489295, + 0.006761310622096062, + 0.00679562846198678, + 0.007104234304279089, + 0.006067471578717232, + 0.0072208065539598465, + 0.008290194906294346, + 0.005003141239285469, + 0.008566438220441341, + 0.008924133144319057, + 0.006575196515768766, + 0.007128371857106686, + 0.010924974456429482, + 0.01420790795236826, + 0.013801701366901398, + 0.008587242104113102, + 0.012494787573814392, + 0.013180733658373356, + 0.009972806088626385, + 0.012786457315087318, + 0.013460775837302208, + 0.018490025773644447, + 0.01486140489578247, + 0.005974780768156052, + 0.024733522906899452, + 0.00905111525207758, + 0.011024359613656998, + 0.016880253329873085, + 0.019950922578573227, + 0.018426191061735153, + 0.008266512304544449, + 0.002917732112109661, + 0.006511367857456207, + 0.003799827303737402, + 0.00495719350874424, + 0.007314363028854132, + 0.005835217889398336, + 0.010657853446900845 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 1 + ], + "selected_reward": -0.0017106198938563466 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007630014326423407 + ], + "opposite_logprob_tokens": [ + -4.882629871368408 + ], + "logprob_reward": -0.007630014326423407, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007630014326423407, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.741646945476532, + 0.02567131817340851, + 0.01803913153707981, + 0.014012696221470833, + 0.030141513794660568, + 0.02624235488474369, + 0.02580304630100727, + 0.03467147797346115, + 0.06053415313363075, + 0.023237405344843864 + ], + "credit_map": [ + 0.19342289865016937, + 0.018979277461767197, + 0.04003924876451492, + 0.022640347480773926, + 0.05170697346329689, + 0.2572384774684906, + 0.04276926815509796, + 0.11485043913125992, + 0.004155331756919622, + 0.0021246876567602158, + 0.0027665910311043262, + 0.0035422404762357473, + 0.0028998705092817545, + 0.004646406043320894, + 0.002159836934879422, + 0.0033763526007533073, + 0.002939564175903797, + 0.0016147713176906109, + 0.0023714948911219835, + 0.0017876250203698874, + 0.0023057209327816963, + 0.00371812772937119, + 0.0014126176247373223, + 0.0018892110092565417, + 0.002412477508187294, + 0.0010741336736828089, + 0.0013455491280183196, + 0.002612331649288535, + 0.001689337776042521, + 0.0019801543094217777, + 0.0012795481598004699, + 0.001619164482690394, + 0.0032267433125525713, + 0.0023245823103934526, + 0.0049270219169557095, + 0.005526866298168898, + 0.0034753228537738323, + 0.004387252498418093, + 0.003127851989120245, + 0.0031458730809390545, + 0.0031613600440323353, + 0.002271419856697321, + 0.005256117321550846, + 0.004251509439200163, + 0.003557513700798154, + 0.0027799010276794434, + 0.00207439623773098, + 0.002890137257054448, + 0.0032329282257705927, + 0.0025012765545397997, + 0.0032763569615781307, + 0.00413645850494504, + 0.002402369398623705, + 0.002952948911115527, + 0.003249059896916151, + 0.004051646683365107, + 0.00516594760119915, + 0.0028200368396937847, + 0.004253581631928682, + 0.004645848646759987, + 0.003250639420002699, + 0.003917755093425512, + 0.004522520583122969, + 0.006095148157328367, + 0.008272198028862476, + 0.0029612749349325895, + 0.012399522587656975, + 0.004279688932001591, + 0.005004224833101034, + 0.009611871093511581, + 0.009087852202355862, + 0.008917523548007011, + 0.004347227979451418, + 0.001189221628010273, + 0.003163349349051714, + 0.0016593148466199636, + 0.0017880502855405211, + 0.0041213915683329105, + 0.0025803386233747005, + 0.004388509783893824 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.007630014326423407 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.2520776689052582 + ], + "opposite_logprob_tokens": [ + -1.5020776987075806 + ], + "logprob_reward": -0.2520776689052582, + "yes_no_margin_reward": 1.25, + "token_credit_reward": -0.2520776689052582, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16520695388317108, + 0.04339046776294708, + 0.04717401787638664, + 0.02381223812699318, + 0.053278256207704544, + 0.053562190383672714, + 0.04812511056661606, + 0.022621523588895798, + 0.27802130579948425, + 0.26480793952941895 + ], + "credit_map": [ + 0.045732516795396805, + 0.006525249220430851, + 0.008563116192817688, + 0.007370667532086372, + 0.013373174704611301, + 0.05627380311489105, + 0.01064490806311369, + 0.016723519191145897, + 0.0064879716373980045, + 0.003811076283454895, + 0.005472779739648104, + 0.006706297397613525, + 0.006034627091139555, + 0.005778362508863211, + 0.004440987017005682, + 0.004658369813114405, + 0.006477843504399061, + 0.004263060633093119, + 0.00798644870519638, + 0.007218522951006889, + 0.00615826016291976, + 0.005139467306435108, + 0.004427946172654629, + 0.005502467975020409, + 0.002983209677040577, + 0.0021802783012390137, + 0.0032253125682473183, + 0.0035758616868406534, + 0.0029454058967530727, + 0.0026442501693964005, + 0.00265842885710299, + 0.003599491436034441, + 0.005304406862705946, + 0.003817034186795354, + 0.00781456008553505, + 0.01019157562404871, + 0.006382551044225693, + 0.008760006166994572, + 0.005751647520810366, + 0.005256476812064648, + 0.008009010925889015, + 0.004766177851706743, + 0.00993282813578844, + 0.006789581850171089, + 0.007004222366958857, + 0.005850682500749826, + 0.004454387817531824, + 0.006755300797522068, + 0.007129242643713951, + 0.003137354040518403, + 0.006451720837503672, + 0.006385385990142822, + 0.003541673067957163, + 0.007386127486824989, + 0.0042982762679457664, + 0.00979533325880766, + 0.003930836915969849, + 0.0017891483148559928, + 0.0029707318171858788, + 0.0026254584081470966, + 0.0018683035159483552, + 0.0036371524911373854, + 0.0018325577257201076, + 0.003967334516346455, + 0.033095233142375946, + 0.01509865839034319, + 0.040363479405641556, + 0.0213069599121809, + 0.027025612071156502, + 0.0278348159044981, + 0.05108044296503067, + 0.062216103076934814, + 0.027969740331172943, + 0.016345622017979622, + 0.04044300690293312, + 0.02025962993502617, + 0.021565183997154236, + 0.03703773766756058, + 0.03414071351289749, + 0.06704631447792053 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 9, + 0 + ], + "selected_reward": -0.2520776689052582 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single panoramic shot, the camera maintains an eye-level perspective and pans smoothly to the right, perfectly tracking a lone, majestic giraffe across a vast savannah landscape. The giraffe's coat is a stunning mosaic of large, polygonal patches of brown and orange, separated by intricate cream-colored lines. It moves from the left to the right side of the frame with a deliberate and steady grace. Its long legs rise and fall with a unique, rhythmic cadence, creating a fluid, continuous stroll that feels both powerful and tranquil. The setting is a vast expanse of dry, golden-brown grassland, sparsely dotted with low-lying green shrubs and the iconic flat-topped acacia trees. In the distance, a range of hazy, blue mountains rises against the horizon, and a vast, overcast sky is filled with soft, white clouds. The entire scene is bathed in a soft, diffuse light that eliminates harsh shadows and enhances the earthy color palette, creating an atmosphere of profound tranquility and naturalness, highlighting the creature's peaceful, solitary journey through its vast habitat.", + "seed": 11890, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_4835_seed_11890_dp_30_rank_60_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.16543402071110905, + "vlm_logprob": -0.16543402071110905, + "vlm_yes_no_margin": 3.849999988079071, + "vlm_token_credit": -0.16543402071110905, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003712667850777507 + ], + "opposite_logprob_tokens": [ + -6.5037126541137695 + ], + "logprob_reward": -0.003712667850777507, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.003712667850777507, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13993088901042938, + 0.06271528452634811, + 0.2145603895187378, + 0.09266050904989243, + 0.05225779488682747, + 0.1533348709344864, + 0.09039580821990967, + 0.04780272766947746, + 0.08651335537433624, + 0.059828341007232666 + ], + "credit_map": [ + 0.01568559929728508, + 0.018308119848370552, + 0.018530061468482018, + 0.016725247725844383, + 0.015352251008152962, + 0.0196697898209095, + 0.015717893838882446, + 0.01994192972779274, + 0.005307159386575222, + 0.007510331925004721, + 0.007309526205062866, + 0.011424848809838295, + 0.007956626825034618, + 0.007826442830264568, + 0.007833296433091164, + 0.007547062821686268, + 0.007847495377063751, + 0.006803735624998808, + 0.08204468339681625, + 0.013538723811507225, + 0.005436242092400789, + 0.00874068308621645, + 0.01382607314735651, + 0.0763227641582489, + 0.007846272550523281, + 0.009290377609431744, + 0.028413891792297363, + 0.007975323125720024, + 0.005186658352613449, + 0.005502705462276936, + 0.006321525201201439, + 0.02212376520037651, + 0.006229670252650976, + 0.008644730783998966, + 0.010440384969115257, + 0.005261546932160854, + 0.003005434526130557, + 0.003796696662902832, + 0.004502631723880768, + 0.010376705788075924, + 0.01578524149954319, + 0.014955212362110615, + 0.01700088568031788, + 0.025476599112153053, + 0.02328329160809517, + 0.019199693575501442, + 0.019030818715691566, + 0.0186031274497509, + 0.006532795261591673, + 0.004993395414203405, + 0.015642251819372177, + 0.014445006847381592, + 0.009170956909656525, + 0.007081411778926849, + 0.014949253760278225, + 0.01758074015378952, + 0.006580994464457035, + 0.007004863116890192, + 0.00909158494323492, + 0.004127354361116886, + 0.0034546046517789364, + 0.0043877544812858105, + 0.005177112761884928, + 0.007978466339409351, + 0.006939281709492207, + 0.007486334070563316, + 0.01081186905503273, + 0.015973174944519997, + 0.007384775672107935, + 0.007318209856748581, + 0.017728213220834732, + 0.012871497310698032, + 0.007121731527149677, + 0.006064507644623518, + 0.009282073006033897, + 0.008255095221102238, + 0.006184106692671776, + 0.008516739122569561, + 0.00735119916498661, + 0.00705289700999856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.003712667850777507 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.42895498871803284 + ], + "opposite_logprob_tokens": [ + -1.0539549589157104 + ], + "logprob_reward": -0.42895498871803284, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.42895498871803284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11105573922395706, + 0.04539179056882858, + 0.07769877463579178, + 0.03295136243104935, + 0.3192288875579834, + 0.07835052162408829, + 0.09686511754989624, + 0.12849804759025574, + 0.06759828329086304, + 0.04236145690083504 + ], + "credit_map": [ + 0.010700001381337643, + 0.012151654809713364, + 0.012770802713930607, + 0.023019053041934967, + 0.012934381142258644, + 0.011794873513281345, + 0.013542226515710354, + 0.014142753556370735, + 0.004010812845081091, + 0.004981531761586666, + 0.006870816927403212, + 0.008485386148095131, + 0.004606745671480894, + 0.0052496204152703285, + 0.004909413401037455, + 0.006277461536228657, + 0.003546494757756591, + 0.004346481524407864, + 0.026858076453208923, + 0.0052847652696073055, + 0.00316171208396554, + 0.004003888461738825, + 0.005339357536286116, + 0.025157997384667397, + 0.004483907483518124, + 0.0036992542445659637, + 0.004358039703220129, + 0.004898509476333857, + 0.003169652307406068, + 0.002888038521632552, + 0.00465385103598237, + 0.004800112918019295, + 0.014206087216734886, + 0.015331678092479706, + 0.10017167031764984, + 0.023779038339853287, + 0.014878927730023861, + 0.016802795231342316, + 0.03180231153964996, + 0.10225637257099152, + 0.008792933076620102, + 0.008849888108670712, + 0.010915318503975868, + 0.011451211757957935, + 0.009260471910238266, + 0.00833505392074585, + 0.00844858679920435, + 0.012297055684030056, + 0.005436894949525595, + 0.004814873915165663, + 0.01803717389702797, + 0.01611887291073799, + 0.00925646536052227, + 0.008329990319907665, + 0.014160373248159885, + 0.020710477605462074, + 0.007558756973594427, + 0.007766717113554478, + 0.019420618191361427, + 0.027723658829927444, + 0.009071910753846169, + 0.007922964170575142, + 0.02650168538093567, + 0.022531721740961075, + 0.006656404584646225, + 0.011434173211455345, + 0.006513865664601326, + 0.007330981083214283, + 0.006508358754217625, + 0.010267076082527637, + 0.011073779314756393, + 0.007813646458089352, + 0.0044583724811673164, + 0.004418550059199333, + 0.004790367558598518, + 0.008579978719353676, + 0.004914404358714819, + 0.005402192007750273, + 0.0059403483755886555, + 0.003857243340462446 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.42895498871803284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0021947361528873444 + ], + "opposite_logprob_tokens": [ + -6.127194881439209 + ], + "logprob_reward": -0.0021947361528873444, + "yes_no_margin_reward": 6.125, + "token_credit_reward": -0.0021947361528873444, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10554345697164536, + 0.04709925875067711, + 0.42141422629356384, + 0.031605128198862076, + 0.07772145420312881, + 0.07622222602367401, + 0.04689798504114151, + 0.06700597703456879, + 0.09131314605474472, + 0.035177070647478104 + ], + "credit_map": [ + 0.01321159116923809, + 0.01381540484726429, + 0.019624363631010056, + 0.014108426868915558, + 0.007810359355062246, + 0.010914458893239498, + 0.011698906309902668, + 0.014359951950609684, + 0.00511660473421216, + 0.005498871207237244, + 0.0057994225062429905, + 0.0076974863186478615, + 0.006541689392179251, + 0.004846431314945221, + 0.006074178498238325, + 0.005524585489183664, + 0.014494460076093674, + 0.012309846468269825, + 0.16388064622879028, + 0.03449724614620209, + 0.00963007565587759, + 0.011782020330429077, + 0.02780991978943348, + 0.14701002836227417, + 0.00382889062166214, + 0.003909549210220575, + 0.005896974820643663, + 0.004525740630924702, + 0.0026667865458875895, + 0.0022398510482162237, + 0.0032446894329041243, + 0.005292650777846575, + 0.005791791714727879, + 0.006328553892672062, + 0.021946759894490242, + 0.008005714043974876, + 0.004446603357791901, + 0.003752519143745303, + 0.0071990652941167355, + 0.020250461995601654, + 0.010707350447773933, + 0.012978018261492252, + 0.011978079564869404, + 0.008011776953935623, + 0.0062580471858382225, + 0.005805084016174078, + 0.00898662582039833, + 0.0114972535520792, + 0.005517460405826569, + 0.0068082972429692745, + 0.0100545110180974, + 0.005133945494890213, + 0.0030677984468638897, + 0.0023029239382594824, + 0.00557352602481842, + 0.00843952875584364, + 0.00588605273514986, + 0.006624297704547644, + 0.01445616316050291, + 0.010199819691479206, + 0.004768664017319679, + 0.005434711929410696, + 0.010528169572353363, + 0.009108106605708599, + 0.009413216263055801, + 0.015179182402789593, + 0.01677376590669155, + 0.007168336305767298, + 0.003954561427235603, + 0.0070428927429020405, + 0.008645103313028812, + 0.02313610166311264, + 0.004909691866487265, + 0.004121575504541397, + 0.006864951457828283, + 0.004129235632717609, + 0.002460444113239646, + 0.00285096513107419, + 0.003456628881394863, + 0.006383584812283516 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 8 + ], + "selected_reward": -0.0021947361528873444 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052520157769322395 + ], + "opposite_logprob_tokens": [ + -5.255251884460449 + ], + "logprob_reward": -0.0052520157769322395, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052520157769322395, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13032883405685425, + 0.12217855453491211, + 0.22792235016822815, + 0.09467906504869461, + 0.17961999773979187, + 0.04200633987784386, + 0.0551750585436821, + 0.06433293223381042, + 0.02966013178229332, + 0.05409669503569603 + ], + "credit_map": [ + 0.016899554058909416, + 0.018481846898794174, + 0.014141194522380829, + 0.023673607036471367, + 0.012709507718682289, + 0.012950076721608639, + 0.01665922999382019, + 0.014813819900155067, + 0.015619347803294659, + 0.01976826786994934, + 0.02080177702009678, + 0.016256118193268776, + 0.011646327562630177, + 0.010763267986476421, + 0.011586673557758331, + 0.015736769884824753, + 0.00790038425475359, + 0.007252597715705633, + 0.08676977455615997, + 0.019839707762002945, + 0.005682250019162893, + 0.007812002208083868, + 0.01582975499331951, + 0.07683587074279785, + 0.007836863398551941, + 0.01268185954540968, + 0.027985617518424988, + 0.007035989314317703, + 0.003940161317586899, + 0.006483857985585928, + 0.006219122093170881, + 0.022495592013001442, + 0.019345296546816826, + 0.01622163876891136, + 0.05248049274086952, + 0.011379217728972435, + 0.006371759343892336, + 0.006904841400682926, + 0.013046539388597012, + 0.05387020483613014, + 0.0051366169936954975, + 0.007952774874866009, + 0.006615789141505957, + 0.0036342497915029526, + 0.0040739416144788265, + 0.00352979707531631, + 0.004663900006562471, + 0.006399272941052914, + 0.004265655297785997, + 0.0039011731278151274, + 0.012490387074649334, + 0.007831256836652756, + 0.004242271184921265, + 0.003481297753751278, + 0.007586330641061068, + 0.01137668639421463, + 0.008216365240514278, + 0.006378549616783857, + 0.010953143239021301, + 0.01134481467306614, + 0.004502582363784313, + 0.0044667888432741165, + 0.009849440306425095, + 0.008621248416602612, + 0.003920049872249365, + 0.002491972642019391, + 0.005746104288846254, + 0.0044111218303442, + 0.0023986073210835457, + 0.0022407385986298323, + 0.004563967231661081, + 0.003887570695951581, + 0.004706051200628281, + 0.006354851182550192, + 0.013036456890404224, + 0.00547528313472867, + 0.0028982243966311216, + 0.004413762595504522, + 0.005206027999520302, + 0.01200603786855936 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 4, + 0 + ], + "selected_reward": -0.0052520157769322395 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3870556950569153 + ], + "opposite_logprob_tokens": [ + -1.1370556354522705 + ], + "logprob_reward": -0.3870556950569153, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3870556950569153, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14722612500190735, + 0.061890456825494766, + 0.12175726145505905, + 0.07878780364990234, + 0.11437106877565384, + 0.07995492964982986, + 0.15405400097370148, + 0.08965862542390823, + 0.06322290003299713, + 0.08907683193683624 + ], + "credit_map": [ + 0.024338005110621452, + 0.019527826458215714, + 0.01570478454232216, + 0.024556320160627365, + 0.01605585776269436, + 0.014365448616445065, + 0.01446537859737873, + 0.018212513998150826, + 0.006945109460502863, + 0.006880618166178465, + 0.009329430758953094, + 0.008970977738499641, + 0.006265167146921158, + 0.0058974046260118484, + 0.008857321925461292, + 0.008744425140321255, + 0.006321823690086603, + 0.007336074486374855, + 0.039715006947517395, + 0.013065633364021778, + 0.004057761747390032, + 0.006251644343137741, + 0.010445504449307919, + 0.034563805907964706, + 0.009448015131056309, + 0.010046769864857197, + 0.01637641340494156, + 0.007443086244165897, + 0.0071845813654363155, + 0.005233043804764748, + 0.008049732074141502, + 0.01500615756958723, + 0.005312247667461634, + 0.005256927106529474, + 0.03644116595387459, + 0.009342759847640991, + 0.005573850125074387, + 0.005385058466345072, + 0.009996692650020123, + 0.03706236183643341, + 0.008407066576182842, + 0.014771162532269955, + 0.011759884655475616, + 0.01066980417817831, + 0.007885907776653767, + 0.007468771655112505, + 0.010267788544297218, + 0.008724541403353214, + 0.0096352007240057, + 0.009859347715973854, + 0.03302100673317909, + 0.016264380887150764, + 0.015851324424147606, + 0.013841342180967331, + 0.020914345979690552, + 0.03466704860329628, + 0.007434196770191193, + 0.010469505563378334, + 0.01432349719107151, + 0.009691724553704262, + 0.007089701946824789, + 0.008622650988399982, + 0.01107809692621231, + 0.020949246361851692, + 0.008013863116502762, + 0.006665906868875027, + 0.010497305542230606, + 0.007334314286708832, + 0.005772861186414957, + 0.009649529121816158, + 0.008527402766048908, + 0.006761717144399881, + 0.013565702363848686, + 0.015594521537423134, + 0.012112293392419815, + 0.008920945227146149, + 0.008266628719866276, + 0.011759888380765915, + 0.006908157374709845, + 0.011948696337640285 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 2 + ], + "selected_reward": -0.3870556950569153 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single panoramic shot, the camera maintains an eye-level perspective and pans smoothly to the right, perfectly tracking a lone, majestic giraffe across a vast savannah landscape. The giraffe's coat is a stunning mosaic of large, polygonal patches of brown and orange, separated by intricate cream-colored lines. It moves from the left to the right side of the frame with a deliberate and steady grace. Its long legs rise and fall with a unique, rhythmic cadence, creating a fluid, continuous stroll that feels both powerful and tranquil. The setting is a vast expanse of dry, golden-brown grassland, sparsely dotted with low-lying green shrubs and the iconic flat-topped acacia trees. In the distance, a range of hazy, blue mountains rises against the horizon, and a vast, overcast sky is filled with soft, white clouds. The entire scene is bathed in a soft, diffuse light that eliminates harsh shadows and enhances the earthy color palette, creating an atmosphere of profound tranquility and naturalness, highlighting the creature's peaceful, solitary journey through its vast habitat.", + "seed": 11891, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_4835_seed_11891_dp_31_rank_62_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.07911816644482314, + "vlm_logprob": -0.07911816644482314, + "vlm_yes_no_margin": 4.3, + "vlm_token_credit": -0.07911816644482314, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032400989439338446 + ], + "opposite_logprob_tokens": [ + -6.87824010848999 + ], + "logprob_reward": -0.0032400989439338446, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0032400989439338446, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10044530034065247, + 0.16932199895381927, + 0.045383013784885406, + 0.04957561567425728, + 0.0940406545996666, + 0.04156693071126938, + 0.0660507082939148, + 0.202106773853302, + 0.17752887308597565, + 0.053980108350515366 + ], + "credit_map": [ + 0.012057045474648476, + 0.01205581333488226, + 0.013598368503153324, + 0.016658565029501915, + 0.010980811901390553, + 0.011911958456039429, + 0.012864391319453716, + 0.010318346321582794, + 0.01643519103527069, + 0.013892637565732002, + 0.012929631397128105, + 0.034368112683296204, + 0.029163163155317307, + 0.01634804531931877, + 0.029647883027791977, + 0.016537345945835114, + 0.0053527820855379105, + 0.004302825778722763, + 0.003637222573161125, + 0.006986803375184536, + 0.0074764434248209, + 0.0054762838408350945, + 0.00791782047599554, + 0.004232834093272686, + 0.00567092327401042, + 0.006072324234992266, + 0.006522466894239187, + 0.007210903335362673, + 0.006318660452961922, + 0.005126670468598604, + 0.008524056524038315, + 0.0041296095587313175, + 0.00817768182605505, + 0.012712832540273666, + 0.009632661938667297, + 0.017789553850889206, + 0.011734810657799244, + 0.011762822046875954, + 0.014132239855825901, + 0.0080980584025383, + 0.0036519167479127645, + 0.00489543192088604, + 0.004085463937371969, + 0.007752047386020422, + 0.004036693833768368, + 0.0045794215984642506, + 0.004998272750526667, + 0.007567680440843105, + 0.0042954618111252785, + 0.004281593952327967, + 0.0077667515724897385, + 0.01078704558312893, + 0.011242837645113468, + 0.007887844927608967, + 0.009532908909022808, + 0.010256258770823479, + 0.016466809436678886, + 0.022290987893939018, + 0.022817743942141533, + 0.03266707807779312, + 0.020703010261058807, + 0.019380055367946625, + 0.02932613715529442, + 0.03845495730638504, + 0.01566098816692829, + 0.030404452234506607, + 0.025080343708395958, + 0.01771998591721058, + 0.020387599244713783, + 0.028147641569375992, + 0.0198972225189209, + 0.02023063413798809, + 0.00648987153545022, + 0.005849604029208422, + 0.007677929010242224, + 0.010683939792215824, + 0.005157994106411934, + 0.005972216371446848, + 0.0069665610790252686, + 0.005181995220482349 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 1 + ], + "selected_reward": -0.0032400989439338446 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1603771448135376 + ], + "opposite_logprob_tokens": [ + -1.9103771448135376 + ], + "logprob_reward": -0.1603771448135376, + "yes_no_margin_reward": 1.75, + "token_credit_reward": -0.1603771448135376, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06902986019849777, + 0.1275467574596405, + 0.028644001111388206, + 0.04646923393011093, + 0.04577259346842766, + 0.08555832505226135, + 0.0641331747174263, + 0.1356797218322754, + 0.19671036303043365, + 0.20045599341392517 + ], + "credit_map": [ + 0.0067393290810287, + 0.007910559885203838, + 0.007039302960038185, + 0.014497737400233746, + 0.007543926127254963, + 0.006190796382725239, + 0.011916188523173332, + 0.007192022632807493, + 0.011542563326656818, + 0.01451406255364418, + 0.010118567384779453, + 0.021618904545903206, + 0.022510932758450508, + 0.011910212226212025, + 0.023374060168862343, + 0.01195745449513197, + 0.0030791766475886106, + 0.0026422718074172735, + 0.001984732924029231, + 0.0037157076876610518, + 0.005977684631943703, + 0.003775737015530467, + 0.005041156895458698, + 0.0024275355972349644, + 0.005022638943046331, + 0.004784994293004274, + 0.004857767838984728, + 0.009221568703651428, + 0.004950899165123701, + 0.004584003705531359, + 0.00710588414222002, + 0.005941473413258791, + 0.0033064980525523424, + 0.004801705479621887, + 0.00521277217194438, + 0.010303467512130737, + 0.004293111618608236, + 0.0039010546170175076, + 0.00850536860525608, + 0.005448615178465843, + 0.006490690167993307, + 0.011584465391933918, + 0.008366030640900135, + 0.018031056970357895, + 0.009087230078876019, + 0.009154291823506355, + 0.013928614556789398, + 0.00891595147550106, + 0.005109598860144615, + 0.007233827840536833, + 0.005214768927544355, + 0.008838897570967674, + 0.012915472500026226, + 0.009549987502396107, + 0.008996398188173771, + 0.006274224258959293, + 0.009992596693336964, + 0.015101427212357521, + 0.015969740226864815, + 0.022677576169371605, + 0.009757624007761478, + 0.01607441157102585, + 0.01742551475763321, + 0.028680823743343353, + 0.012216172181069851, + 0.02062557265162468, + 0.0164616908878088, + 0.03279974311590195, + 0.029320938512682915, + 0.03245461732149124, + 0.03252147510647774, + 0.020310157909989357, + 0.016152914613485336, + 0.021854214370250702, + 0.04504948481917381, + 0.03378450870513916, + 0.015900785103440285, + 0.021198244765400887, + 0.026901038363575935, + 0.019614804536104202 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 8, + 7 + ], + "selected_reward": -0.1603771448135376 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002814263105392456 + ], + "opposite_logprob_tokens": [ + -5.877814292907715 + ], + "logprob_reward": -0.002814263105392456, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002814263105392456, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14675623178482056, + 0.14069423079490662, + 0.048782482743263245, + 0.03748985752463341, + 0.3553241491317749, + 0.05527957156300545, + 0.025917014107108116, + 0.03561388701200485, + 0.051525961607694626, + 0.10261660069227219 + ], + "credit_map": [ + 0.021579954773187637, + 0.019029704853892326, + 0.013667927123606205, + 0.020714566111564636, + 0.02351420558989048, + 0.015077975578606129, + 0.01710023730993271, + 0.01607166975736618, + 0.01442705374211073, + 0.01590416394174099, + 0.012128670699894428, + 0.02541298419237137, + 0.022757166996598244, + 0.012676429003477097, + 0.023519648239016533, + 0.013868104666471481, + 0.007108308374881744, + 0.0070283194072544575, + 0.006912598852068186, + 0.006107211112976074, + 0.005174348130822182, + 0.005015191622078419, + 0.0049976930022239685, + 0.00643880944699049, + 0.0041516488417983055, + 0.004391923546791077, + 0.004119283054023981, + 0.006337639410048723, + 0.0050749508664011955, + 0.003965425305068493, + 0.005105361342430115, + 0.004343624226748943, + 0.02609233930706978, + 0.0485188253223896, + 0.04570121318101883, + 0.05532500147819519, + 0.044913243502378464, + 0.03959415853023529, + 0.050347309559583664, + 0.04483208432793617, + 0.005208216607570648, + 0.00755456043407321, + 0.006587572395801544, + 0.006784255150705576, + 0.006727974861860275, + 0.0059816669672727585, + 0.006420936435461044, + 0.010014391504228115, + 0.0020686062052845955, + 0.0021597447339445353, + 0.0024490649811923504, + 0.006098649464547634, + 0.0039406828582286835, + 0.0027809981256723404, + 0.0039311065338552, + 0.002488161902874708, + 0.0038197373505681753, + 0.004227625206112862, + 0.0037483577616512775, + 0.004780641756951809, + 0.004498098976910114, + 0.0038820155896246433, + 0.00520760752260685, + 0.005449802614748478, + 0.0036951573565602303, + 0.009199387393891811, + 0.005656531546264887, + 0.007010613102465868, + 0.004132591653615236, + 0.008045622147619724, + 0.008229895494878292, + 0.00555616058409214, + 0.009052237495779991, + 0.017615970224142075, + 0.016503741964697838, + 0.012426032684743404, + 0.00985134206712246, + 0.019044024869799614, + 0.008644851855933666, + 0.009478398598730564 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.002814263105392456 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0036149173974990845 + ], + "opposite_logprob_tokens": [ + -5.628614902496338 + ], + "logprob_reward": -0.0036149173974990845, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.0036149173974990845, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11094577610492706, + 0.098148874938488, + 0.06079069525003433, + 0.04528845101594925, + 0.07112045586109161, + 0.034349966794252396, + 0.01957116834819317, + 0.11815453320741653, + 0.11275682598352432, + 0.32887327671051025 + ], + "credit_map": [ + 0.021070877090096474, + 0.014033091254532337, + 0.012186262756586075, + 0.01251088548451662, + 0.008537034504115582, + 0.013698029331862926, + 0.01127682626247406, + 0.017632771283388138, + 0.011654546484351158, + 0.010112641379237175, + 0.007937589660286903, + 0.018026085570454597, + 0.015167463570833206, + 0.009217236191034317, + 0.017039764672517776, + 0.008993545547127724, + 0.012406042777001858, + 0.006828582379966974, + 0.0049174996092915535, + 0.007327769882977009, + 0.008519197814166546, + 0.007664496544748545, + 0.007689737249165773, + 0.005437369924038649, + 0.005488850176334381, + 0.005534029565751553, + 0.00438192393630743, + 0.009364087134599686, + 0.005446494556963444, + 0.004494146443903446, + 0.007365122437477112, + 0.003213797463104129, + 0.005921541713178158, + 0.01199430413544178, + 0.00944384653121233, + 0.012922722846269608, + 0.0062430426478385925, + 0.008219144307076931, + 0.008886906318366528, + 0.007488951552659273, + 0.0026319576427340508, + 0.0036099248100072145, + 0.00347230676561594, + 0.007115988992154598, + 0.004040014464408159, + 0.003575299633666873, + 0.005371301434934139, + 0.004533171188086271, + 0.0015696436166763306, + 0.001963791437447071, + 0.0018284434918314219, + 0.0047126831486821175, + 0.002119069453328848, + 0.0020150181371718645, + 0.003111724741756916, + 0.0022507947869598866, + 0.009330407716333866, + 0.013528671115636826, + 0.01265545655041933, + 0.019726945087313652, + 0.014118637889623642, + 0.012021638453006744, + 0.01644003577530384, + 0.020332742482423782, + 0.00757615827023983, + 0.015277236700057983, + 0.011462494730949402, + 0.01524779386818409, + 0.012936476618051529, + 0.022112522274255753, + 0.014117664657533169, + 0.014026463963091373, + 0.022686617448925972, + 0.044588737189769745, + 0.037441786378622055, + 0.04727718234062195, + 0.031853754073381424, + 0.06928078830242157, + 0.040967199951410294, + 0.034777216613292694 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.0036149173974990845 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554440796375275 + ], + "opposite_logprob_tokens": [ + -1.6005444526672363 + ], + "logprob_reward": -0.22554440796375275, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554440796375275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08810234069824219, + 0.07595273107290268, + 0.022895552217960358, + 0.03490400314331055, + 0.046589985489845276, + 0.02733197622001171, + 0.024085531011223793, + 0.433736115694046, + 0.09857453405857086, + 0.14782726764678955 + ], + "credit_map": [ + 0.008801326155662537, + 0.00955201592296362, + 0.008801908232271671, + 0.017092902213335037, + 0.01130310446023941, + 0.008436220698058605, + 0.015109623782336712, + 0.009005242958664894, + 0.007371323648840189, + 0.006534340791404247, + 0.00658701965585351, + 0.014290634542703629, + 0.010628968477249146, + 0.006988999433815479, + 0.01517496444284916, + 0.008376484736800194, + 0.0027735040057450533, + 0.0028139392379671335, + 0.0021724679972976446, + 0.0036175609566271305, + 0.0030161335598677397, + 0.0022437269799411297, + 0.003971477504819632, + 0.0022867422085255384, + 0.005248234141618013, + 0.005236750468611717, + 0.0037223934195935726, + 0.00603615865111351, + 0.0028052404522895813, + 0.003981537185609341, + 0.00534907216206193, + 0.002524618525058031, + 0.0036787225399166346, + 0.005709812045097351, + 0.005430440418422222, + 0.007101211231201887, + 0.006291009020060301, + 0.005106776487082243, + 0.007193530444055796, + 0.0060784826055169106, + 0.0025676812510937452, + 0.0034281511325389147, + 0.0027880221605300903, + 0.005376472603529692, + 0.002586359390988946, + 0.002717948518693447, + 0.0037690557073801756, + 0.004098284989595413, + 0.002466953359544277, + 0.0030834993813186884, + 0.002241587033495307, + 0.003502335399389267, + 0.0032848436385393143, + 0.0027529567014425993, + 0.0029701869934797287, + 0.003783167339861393, + 0.03206000477075577, + 0.043714363127946854, + 0.048358336091041565, + 0.07002997398376465, + 0.04554353281855583, + 0.04545178636908531, + 0.06051789969205856, + 0.08806021511554718, + 0.005545952823013067, + 0.009657657705247402, + 0.00940637942403555, + 0.01593460515141487, + 0.01482036616653204, + 0.015412520617246628, + 0.016281625255942345, + 0.011515416204929352, + 0.010961368680000305, + 0.017527755349874496, + 0.022059187293052673, + 0.020976196974515915, + 0.0170921441167593, + 0.02590670995414257, + 0.019291579723358154, + 0.014012332074344158 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 8 + ], + "selected_reward": -0.22554440796375275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "In a single panoramic shot, the camera maintains an eye-level perspective and pans smoothly to the right, perfectly tracking a lone, majestic giraffe across a vast savannah landscape. The giraffe's coat is a stunning mosaic of large, polygonal patches of brown and orange, separated by intricate cream-colored lines. It moves from the left to the right side of the frame with a deliberate and steady grace. Its long legs rise and fall with a unique, rhythmic cadence, creating a fluid, continuous stroll that feels both powerful and tranquil. The setting is a vast expanse of dry, golden-brown grassland, sparsely dotted with low-lying green shrubs and the iconic flat-topped acacia trees. In the distance, a range of hazy, blue mountains rises against the horizon, and a vast, overcast sky is filled with soft, white clouds. The entire scene is bathed in a soft, diffuse light that eliminates harsh shadows and enhances the earthy color palette, creating an atmosphere of profound tranquility and naturalness, highlighting the creature's peaceful, solitary journey through its vast habitat.", + "seed": 11892, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_15/index_4835_seed_11892_dp_31_rank_62_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 15, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.23031975333578886, + "vlm_logprob": -0.23031975333578886, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.23031975333578886, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0032436635810881853 + ], + "opposite_logprob_tokens": [ + -6.878243446350098 + ], + "logprob_reward": -0.0032436635810881853, + "yes_no_margin_reward": 6.875, + "token_credit_reward": -0.0032436635810881853, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15319852530956268, + 0.14897753298282623, + 0.21841099858283997, + 0.05454089120030403, + 0.058001771569252014, + 0.1538044959306717, + 0.06616795808076859, + 0.059329308569431305, + 0.04438132792711258, + 0.043187204748392105 + ], + "credit_map": [ + 0.015916040167212486, + 0.009283397346735, + 0.02048998326063156, + 0.023643316701054573, + 0.01324908621609211, + 0.013998947106301785, + 0.027811024338006973, + 0.02880672551691532, + 0.018095698207616806, + 0.014178507961332798, + 0.020138205960392952, + 0.02278916724026203, + 0.020508598536252975, + 0.018505806103348732, + 0.020959213376045227, + 0.013802318833768368, + 0.02371978759765625, + 0.01702303998172283, + 0.03679763898253441, + 0.03403870388865471, + 0.020404992625117302, + 0.023212803527712822, + 0.03808752819895744, + 0.02512650564312935, + 0.007179219741374254, + 0.009969854727387428, + 0.005642857868224382, + 0.0066014002077281475, + 0.005210334900766611, + 0.007984954863786697, + 0.004591297823935747, + 0.007360970601439476, + 0.0076803783886134624, + 0.007579200901091099, + 0.007452106568962336, + 0.012514866888523102, + 0.003929492086172104, + 0.006151092238724232, + 0.00485824141651392, + 0.007836391218006611, + 0.014960257336497307, + 0.015636321157217026, + 0.03450433164834976, + 0.01803537830710411, + 0.0070823682472109795, + 0.01434847991913557, + 0.01523035578429699, + 0.034007005393505096, + 0.008855227380990982, + 0.007729418110102415, + 0.008541551418602467, + 0.012801537290215492, + 0.006778153590857983, + 0.00907053891569376, + 0.006670318078249693, + 0.005721209570765495, + 0.00692441500723362, + 0.005873083136975765, + 0.00894233863800764, + 0.009952492080628872, + 0.006353477947413921, + 0.00603995006531477, + 0.006596677005290985, + 0.008646873757243156, + 0.006984461098909378, + 0.006471911445260048, + 0.006041798274964094, + 0.005280112847685814, + 0.0021798626985400915, + 0.006146631669253111, + 0.003306847531348467, + 0.007969699800014496, + 0.005478811450302601, + 0.005163997877389193, + 0.007362206932157278, + 0.0056663332507014275, + 0.003486791392788291, + 0.0065492806024849415, + 0.00410829484462738, + 0.005371489562094212 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 0 + ], + "selected_reward": -0.0032436635810881853 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.8261996507644653 + ], + "opposite_logprob_tokens": [ + -0.5761996507644653 + ], + "logprob_reward": -0.8261996507644653, + "yes_no_margin_reward": -0.25, + "token_credit_reward": -0.8261996507644653, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.04579360783100128, + 0.03089396469295025, + 0.4307178556919098, + 0.15198858082294464, + 0.06252560764551163, + 0.040150050073862076, + 0.0834493413567543, + 0.06967195123434067, + 0.034442104399204254, + 0.050366900861263275 + ], + "credit_map": [ + 0.004490342456847429, + 0.002818644978106022, + 0.006970415823161602, + 0.007029147818684578, + 0.003741167951375246, + 0.0039216820150613785, + 0.008868806064128876, + 0.007953405380249023, + 0.0033126710914075375, + 0.0022543608210980892, + 0.004360316786915064, + 0.00458500673994422, + 0.003667373675853014, + 0.0034416024573147297, + 0.0051278006285429, + 0.004144834354519844, + 0.04762699827551842, + 0.03653087839484215, + 0.06959801912307739, + 0.06572356820106506, + 0.04572271183133125, + 0.04355259984731674, + 0.07104355841875076, + 0.0509195514023304, + 0.011901004239916801, + 0.015732137486338615, + 0.017644334584474564, + 0.022644398733973503, + 0.018997758626937866, + 0.01834181509912014, + 0.015527020208537579, + 0.0312001071870327, + 0.006148734595626593, + 0.00829743966460228, + 0.01000190805643797, + 0.008842965587973595, + 0.003450461896136403, + 0.007049825973808765, + 0.006304314825683832, + 0.012429960072040558, + 0.004639343358576298, + 0.005083797499537468, + 0.004451185930520296, + 0.00841678399592638, + 0.003348870435729623, + 0.004411068744957447, + 0.004627496004104614, + 0.005171504803001881, + 0.006204177625477314, + 0.00889277458190918, + 0.009307357482612133, + 0.018848344683647156, + 0.008378610946238041, + 0.01262653712183237, + 0.010828407481312752, + 0.008363137021660805, + 0.007543592713773251, + 0.006974783260375261, + 0.008755055256187916, + 0.013625862076878548, + 0.006659836042672396, + 0.007725195027887821, + 0.0091461893171072, + 0.00924143847078085, + 0.002850800286978483, + 0.0026929927989840508, + 0.004645183682441711, + 0.007914775051176548, + 0.0032741103786975145, + 0.0036484396550804377, + 0.004101266153156757, + 0.005314539186656475, + 0.005869033746421337, + 0.006577561609447002, + 0.007661032024770975, + 0.007039602380245924, + 0.0037915243301540613, + 0.007890905253589153, + 0.0047806911170482635, + 0.006756552495062351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 6 + ], + "selected_reward": -0.8261996507644653 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004091108217835426 + ], + "opposite_logprob_tokens": [ + -5.504091262817383 + ], + "logprob_reward": -0.004091108217835426, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004091108217835426, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13816344738006592, + 0.11949460208415985, + 0.24649770557880402, + 0.158296599984169, + 0.052177343517541885, + 0.1208217591047287, + 0.042695656418800354, + 0.08662552386522293, + 0.013640419580042362, + 0.021586976945400238 + ], + "credit_map": [ + 0.017488522455096245, + 0.009273859672248363, + 0.021670518442988396, + 0.017745355144143105, + 0.007989304140210152, + 0.016458388417959213, + 0.01974603347480297, + 0.027791466563940048, + 0.020706184208393097, + 0.013571270741522312, + 0.016393499448895454, + 0.01488662138581276, + 0.01118214800953865, + 0.017804250121116638, + 0.013758279383182526, + 0.01119234785437584, + 0.03188033401966095, + 0.021412834525108337, + 0.03314987197518349, + 0.039173200726509094, + 0.024231910705566406, + 0.03194937855005264, + 0.038607556372880936, + 0.026092620566487312, + 0.020212344825267792, + 0.022177334874868393, + 0.01701946370303631, + 0.020107608288526535, + 0.016990434378385544, + 0.018944550305604935, + 0.01521452609449625, + 0.027630336582660675, + 0.008162569254636765, + 0.008113890886306763, + 0.0074319844134151936, + 0.006636720150709152, + 0.00300647527910769, + 0.007021049503237009, + 0.004014459904283285, + 0.007790192496031523, + 0.014165912754833698, + 0.012479203753173351, + 0.02467789687216282, + 0.012695860117673874, + 0.0040269033052027225, + 0.011476396583020687, + 0.01012156717479229, + 0.031178023666143417, + 0.010079141706228256, + 0.00652583921328187, + 0.004313479643315077, + 0.004368226509541273, + 0.00216534617356956, + 0.008145293220877647, + 0.003117318032309413, + 0.00398101145401597, + 0.008602689020335674, + 0.01004360057413578, + 0.015918172895908356, + 0.012480058707296848, + 0.005932760424911976, + 0.010992974042892456, + 0.010846294462680817, + 0.011808970011770725, + 0.0015753827756270766, + 0.0012678757775574923, + 0.002325954381376505, + 0.0019204707350581884, + 0.0007219320978038013, + 0.001442423672415316, + 0.0013844905188307166, + 0.0030018899124115705, + 0.0031387393828481436, + 0.0025680235121399164, + 0.004199614282697439, + 0.0018535698764026165, + 0.0008757790201343596, + 0.00295063853263855, + 0.0019791785161942244, + 0.0040214345790445805 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.004091108217835426 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638862796127796 + ], + "opposite_logprob_tokens": [ + -5.379638671875 + ], + "logprob_reward": -0.004638862796127796, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638862796127796, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09070435166358948, + 0.2583049535751343, + 0.06874968111515045, + 0.20356610417366028, + 0.07156993448734283, + 0.12493929266929626, + 0.0470927432179451, + 0.05331428721547127, + 0.041646745055913925, + 0.04011189565062523 + ], + "credit_map": [ + 0.01083377841860056, + 0.00649053230881691, + 0.01540741790086031, + 0.016077514737844467, + 0.006735004484653473, + 0.011418242938816547, + 0.011687506921589375, + 0.012054353021085262, + 0.04098398983478546, + 0.02830701693892479, + 0.03613349422812462, + 0.03249381110072136, + 0.0287117101252079, + 0.03820217400789261, + 0.03208722174167633, + 0.021385518833994865, + 0.011045252904295921, + 0.006930618081241846, + 0.009818099439144135, + 0.009440507739782333, + 0.00765256630256772, + 0.0085174310952425, + 0.007947796024382114, + 0.007397414185106754, + 0.05290655046701431, + 0.04457392171025276, + 0.016645438969135284, + 0.013986370526254177, + 0.012458797544240952, + 0.035273756831884384, + 0.015379187650978565, + 0.012342100031673908, + 0.007181265391409397, + 0.008535216562449932, + 0.009953778237104416, + 0.011362935416400433, + 0.005927735939621925, + 0.008967200294137001, + 0.008656025864183903, + 0.010985770262777805, + 0.012686317786574364, + 0.010186709463596344, + 0.02832091972231865, + 0.013299530372023582, + 0.0036723280791193247, + 0.011774467304348946, + 0.010528156533837318, + 0.03447085618972778, + 0.011471422389149666, + 0.00849861279129982, + 0.004437432158738375, + 0.004701769445091486, + 0.0030694352462887764, + 0.008353238925337791, + 0.003089158795773983, + 0.0034716734662652016, + 0.005067570600658655, + 0.00603978568688035, + 0.008561928756535053, + 0.008851995691657066, + 0.005632804706692696, + 0.006092548370361328, + 0.006316881626844406, + 0.006750775966793299, + 0.006958670914173126, + 0.006731262896209955, + 0.005367091856896877, + 0.005391663871705532, + 0.002545237308368087, + 0.005193576216697693, + 0.003310266649350524, + 0.00614897208288312, + 0.005479427985846996, + 0.00606740964576602, + 0.006676658987998962, + 0.0027611020486801863, + 0.0018609951948747039, + 0.008245051838457584, + 0.002809777855873108, + 0.006211474537849426 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 5 + ], + "selected_reward": -0.004638862796127796 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a lone giraffe with a coat of brown and orange polygonal patches separated by cream-colored lines?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Does the giraffe move from the left side of the frame to the right side?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the background a vast savannah with dry golden-brown grass, green shrubs, and flat-topped acacia trees?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are hazy blue mountains visible in the distance against an overcast sky with soft white clouds?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera pan smoothly to the right while maintaining an eye-level perspective?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3134254813194275 + ], + "opposite_logprob_tokens": [ + -1.3134255409240723 + ], + "logprob_reward": -0.3134254813194275, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.3134254813194275, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.28063201904296875, + 0.09101514518260956, + 0.060089290142059326, + 0.12034709006547928, + 0.05731724202632904, + 0.14314202964305878, + 0.05720217153429985, + 0.06276169419288635, + 0.05409162864089012, + 0.07340165972709656 + ], + "credit_map": [ + 0.04444575682282448, + 0.020665185526013374, + 0.04210187867283821, + 0.039816346019506454, + 0.030026694759726524, + 0.037274494767189026, + 0.022871429100632668, + 0.04343026131391525, + 0.012697698548436165, + 0.008110846392810345, + 0.013694701716303825, + 0.011559892445802689, + 0.009708937257528305, + 0.010389472357928753, + 0.013880997896194458, + 0.010972606018185616, + 0.008057249709963799, + 0.00480226194486022, + 0.00887519121170044, + 0.007018922362476587, + 0.0058040558360517025, + 0.007151905447244644, + 0.007357504218816757, + 0.011022207327187061, + 0.01342635229229927, + 0.015031206421554089, + 0.015269046649336815, + 0.015519281849265099, + 0.014510435983538628, + 0.016087360680103302, + 0.009905930608510971, + 0.020597489550709724, + 0.0058752670884132385, + 0.004607717040926218, + 0.00965960044413805, + 0.008031598292291164, + 0.003194665303453803, + 0.0042494614608585835, + 0.006783795077353716, + 0.014915144070982933, + 0.014974063262343407, + 0.016301309689879417, + 0.0312650203704834, + 0.015426555648446083, + 0.004847180098295212, + 0.01306584756821394, + 0.012822442688047886, + 0.03443961963057518, + 0.007336587645113468, + 0.007296376395970583, + 0.00843636505305767, + 0.011029841378331184, + 0.004049531649798155, + 0.006334830075502396, + 0.006075525190681219, + 0.0066431183367967606, + 0.005850036628544331, + 0.00843388307839632, + 0.010062141343951225, + 0.013195150531828403, + 0.004294955637305975, + 0.006667492911219597, + 0.00758977560326457, + 0.006668264977633953, + 0.007004904560744762, + 0.0048907906748354435, + 0.008510543033480644, + 0.010452923364937305, + 0.004349942784756422, + 0.005793919786810875, + 0.005708273500204086, + 0.007380339782685041, + 0.010723250918090343, + 0.013758013024926186, + 0.009581957943737507, + 0.005519560072571039, + 0.0039615388959646225, + 0.017312973737716675, + 0.004600860178470612, + 0.007943508215248585 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 3 + ], + "selected_reward": -0.3134254813194275 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "No", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_tensor([9075, 9075])_group_2_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_tensor([9075, 9075])_group_2_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..9186abeec5048e396dd39b847511c21b79b942ef --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_tensor([9075, 9075])_group_2_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "steady, close-up slowly and steadily approaches, plunging the viewer into a deep-sea microcosm dominated by a chilling blue-green hue. The absolute focal point is the Portuguese Man-of-War jellyfish's nearly transparent, deadly tentacles, emitting an eerie, biological glow. Like deadly threads studded with countless tiny, pearl-like sacs of stinging cells, they slowly burrow into the body of a small, silvery fish with a steady, relentless rhythm. Under the faint, cold light of the tentacles themselves, the fish's scales reflect a desperate metallic luster, its wide-open eyes frozen in the terror and vacillation of its final moments. The camera meticulously captures the moment the tentacle tip pierces the skin, pushing aside the scales. The slender stinging threads penetrate the flesh, leaving a faint, necrotic white stain on the tissue surrounding the point of contact, a foreboding sign of death. The surroundings are a profound, silent darkness, with only countless suspended particles drifting silently in the dim light, creating a microscopic atmosphere that is both magnificent and extremely dangerous, deeply unsettling, and eerie.", + "seed": 13442, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13442_dp_4_rank_8_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.020525194285437465, + "vlm_logprob": -0.020525194285437465, + "vlm_yes_no_margin": 4.5, + "vlm_token_credit": -0.020525194285437465, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03942583128809929 + ], + "opposite_logprob_tokens": [ + -3.289425849914551 + ], + "logprob_reward": -0.03942583128809929, + "yes_no_margin_reward": 3.25, + "token_credit_reward": -0.03942583128809929, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09829267114400864, + 0.1004071980714798, + 0.0950358510017395, + 0.04913514107465744, + 0.11610877513885498, + 0.11336667835712433, + 0.17693153023719788, + 0.07232239097356796, + 0.1361285001039505, + 0.042271193116903305 + ], + "credit_map": [ + 0.006539034657180309, + 0.007469139061868191, + 0.02080378495156765, + 0.009964656084775925, + 0.014145485125482082, + 0.012517863884568214, + 0.0067439004778862, + 0.02010880969464779, + 0.00856224074959755, + 0.009418741799890995, + 0.017123503610491753, + 0.00943700224161148, + 0.00991210900247097, + 0.016313642263412476, + 0.00950382649898529, + 0.020136136561632156, + 0.00954767968505621, + 0.007891925983130932, + 0.017044350504875183, + 0.01081711146980524, + 0.008008548058569431, + 0.01474769413471222, + 0.014291860163211823, + 0.012686684727668762, + 0.005098363384604454, + 0.005808709189295769, + 0.007700988557189703, + 0.0035231327638030052, + 0.006247332319617271, + 0.0045729344710707664, + 0.00529523566365242, + 0.01088844332844019, + 0.015471604652702808, + 0.01166424248367548, + 0.014149284921586514, + 0.010707770474255085, + 0.011660778895020485, + 0.01720878668129444, + 0.016820618882775307, + 0.018425695598125458, + 0.006901252083480358, + 0.008752056397497654, + 0.017737841233611107, + 0.015518414787948132, + 0.01415558997541666, + 0.010731477290391922, + 0.02170090563595295, + 0.017869139090180397, + 0.014324390329420567, + 0.029820045456290245, + 0.03051992505788803, + 0.02681485377252102, + 0.022922338917851448, + 0.014760218560695648, + 0.015315507538616657, + 0.022454265505075455, + 0.008869513869285583, + 0.006224042735993862, + 0.01138787716627121, + 0.0065879360772669315, + 0.007157350890338421, + 0.009112955071032047, + 0.011719142086803913, + 0.011263572610914707, + 0.009695847518742085, + 0.010559930466115475, + 0.015878036618232727, + 0.017824361100792885, + 0.009253584779798985, + 0.012518121860921383, + 0.041658610105514526, + 0.01874002255499363, + 0.0037337173707783222, + 0.006010210607200861, + 0.0050055827014148235, + 0.0036176370922476053, + 0.007511222269386053, + 0.004687073174864054, + 0.005741241853684187, + 0.00596450874581933 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 8, + 4 + ], + "selected_reward": -0.03942583128809929 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004107254557311535 + ], + "opposite_logprob_tokens": [ + -5.504107475280762 + ], + "logprob_reward": -0.004107254557311535, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004107254557311535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13278383016586304, + 0.07612410932779312, + 0.1176203191280365, + 0.08139785379171371, + 0.10597683489322662, + 0.14357328414916992, + 0.1724294275045395, + 0.03161892294883728, + 0.08040782809257507, + 0.05806756019592285 + ], + "credit_map": [ + 0.009618590585887432, + 0.011676792986690998, + 0.02338780276477337, + 0.016875170171260834, + 0.016367483884096146, + 0.013884034007787704, + 0.01564755104482174, + 0.025326404720544815, + 0.006908078212291002, + 0.008560597896575928, + 0.01193440705537796, + 0.006762173026800156, + 0.007489682175219059, + 0.008163991384208202, + 0.011317871510982513, + 0.014987312257289886, + 0.015463506802916527, + 0.014798112213611603, + 0.019317949190735817, + 0.011928037740290165, + 0.0082158874720335, + 0.015940526500344276, + 0.016405340284109116, + 0.015550956130027771, + 0.005625209771096706, + 0.005319551564753056, + 0.017587633803486824, + 0.008325882256031036, + 0.007101395633071661, + 0.008143470622599125, + 0.007105944212526083, + 0.02218877151608467, + 0.015477572567760944, + 0.01225935760885477, + 0.012827781029045582, + 0.01306767575442791, + 0.010222177952528, + 0.014001574367284775, + 0.011910440400242805, + 0.016210244968533516, + 0.007411882281303406, + 0.015212604776024818, + 0.022445542737841606, + 0.015411587432026863, + 0.017867233604192734, + 0.014132369309663773, + 0.018614863976836205, + 0.03247721120715141, + 0.014156516641378403, + 0.012883809395134449, + 0.030157005414366722, + 0.031013045459985733, + 0.012530879117548466, + 0.02052498236298561, + 0.025705814361572266, + 0.02545737288892269, + 0.0027344352565705776, + 0.0038105822168290615, + 0.004880601540207863, + 0.004153165966272354, + 0.003371977247297764, + 0.003771006828173995, + 0.003407518845051527, + 0.005489637143909931, + 0.006442346144467592, + 0.008057637140154839, + 0.00901961512863636, + 0.008611532859504223, + 0.008483844809234142, + 0.006720235571265221, + 0.021845167502760887, + 0.011227448470890522, + 0.007679598871618509, + 0.0058964635245501995, + 0.006676734425127506, + 0.005261098966002464, + 0.010069053620100021, + 0.006758301984518766, + 0.007211191579699516, + 0.008515117689967155 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.004107254557311535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007632498629391193 + ], + "opposite_logprob_tokens": [ + -4.882632732391357 + ], + "logprob_reward": -0.007632498629391193, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007632498629391193, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.099459707736969, + 0.07224225252866745, + 0.15692970156669617, + 0.0604916475713253, + 0.18391279876232147, + 0.10485672950744629, + 0.15236018598079681, + 0.05094089359045029, + 0.05342373996973038, + 0.06538234651088715 + ], + "credit_map": [ + 0.008366791531443596, + 0.008400872349739075, + 0.016989709809422493, + 0.013248959556221962, + 0.012920384295284748, + 0.01082624215632677, + 0.010799586772918701, + 0.017907164990901947, + 0.006340339779853821, + 0.010482325218617916, + 0.010106517933309078, + 0.00817115604877472, + 0.008999345824122429, + 0.0076980902813375, + 0.00903309416025877, + 0.01141138095408678, + 0.017531678080558777, + 0.012806008569896221, + 0.02822277322411537, + 0.018917864188551903, + 0.011609631590545177, + 0.024411821737885475, + 0.019244594499468803, + 0.024185335263609886, + 0.005821895319968462, + 0.006059608887881041, + 0.010873349383473396, + 0.005102573428303003, + 0.006175316870212555, + 0.007030817214399576, + 0.008013084530830383, + 0.011414999142289162, + 0.030591173097491264, + 0.020158320665359497, + 0.019275063648819923, + 0.020941119641065598, + 0.01568194106221199, + 0.027307484298944473, + 0.020148130133748055, + 0.029809579253196716, + 0.005679164081811905, + 0.010602951049804688, + 0.01628858968615532, + 0.011503911577165127, + 0.012041162699460983, + 0.008442913182079792, + 0.014624197967350483, + 0.025673840194940567, + 0.011219869367778301, + 0.024908622726798058, + 0.019277920946478844, + 0.01507079042494297, + 0.018995381891727448, + 0.018178487196564674, + 0.017629701644182205, + 0.027079420164227486, + 0.005627261940389872, + 0.007229638751596212, + 0.008531168103218079, + 0.0052312565967440605, + 0.00635811360552907, + 0.004949222318828106, + 0.004595709498971701, + 0.008418526500463486, + 0.003177422797307372, + 0.007818731479346752, + 0.006355698220431805, + 0.004118343815207481, + 0.009856794029474258, + 0.004124104976654053, + 0.007721438072621822, + 0.01025120634585619, + 0.007945614866912365, + 0.007756566163152456, + 0.0068738930858671665, + 0.005579502787441015, + 0.012966881506145, + 0.006984050385653973, + 0.008964529260993004, + 0.008311307989060879 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 6 + ], + "selected_reward": -0.007632498629391193 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04864410310983658 + ], + "opposite_logprob_tokens": [ + -3.0486440658569336 + ], + "logprob_reward": -0.04864410310983658, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.04864410310983658, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08661077916622162, + 0.06028638407588005, + 0.10969090461730957, + 0.05253198742866516, + 0.2759830355644226, + 0.05791337043046951, + 0.16395443677902222, + 0.05025402083992958, + 0.036467261612415314, + 0.10630784183740616 + ], + "credit_map": [ + 0.006228390149772167, + 0.006987682543694973, + 0.016371993348002434, + 0.011120806448161602, + 0.011931871995329857, + 0.010436099953949451, + 0.0075202505104243755, + 0.016013674437999725, + 0.005744090769439936, + 0.01165201235562563, + 0.008298343978822231, + 0.006044118199497461, + 0.006331689655780792, + 0.005560240242630243, + 0.007873238064348698, + 0.008782649412751198, + 0.012366102077066898, + 0.00985005684196949, + 0.019045613706111908, + 0.012002774514257908, + 0.0077371965162456036, + 0.016884882003068924, + 0.013658605515956879, + 0.018145663663744926, + 0.004721552599221468, + 0.007492452394217253, + 0.007074700202792883, + 0.0040642861276865005, + 0.007346435450017452, + 0.00633895443752408, + 0.005717321764677763, + 0.00977628119289875, + 0.04003363102674484, + 0.027806513011455536, + 0.03258723393082619, + 0.02680373564362526, + 0.027634020894765854, + 0.04024053364992142, + 0.03780904412269592, + 0.0430683009326458, + 0.004062006715685129, + 0.006291063968092203, + 0.008035320788621902, + 0.00679793581366539, + 0.006295097526162863, + 0.006262700539082289, + 0.01036739069968462, + 0.009801848791539669, + 0.011379825882613659, + 0.02142120897769928, + 0.025428295135498047, + 0.023532699793577194, + 0.017942342907190323, + 0.014406745322048664, + 0.027375560253858566, + 0.022467754781246185, + 0.004662737715989351, + 0.00823463685810566, + 0.007445260416716337, + 0.0038429738488048315, + 0.006735618691891432, + 0.004637371748685837, + 0.007674301043152809, + 0.007021117024123669, + 0.003636007895693183, + 0.003901198273524642, + 0.0036851062905043364, + 0.0030940950382500887, + 0.005789193324744701, + 0.0040043131448328495, + 0.007755360100418329, + 0.0046019842848181725, + 0.007951959036290646, + 0.011298472061753273, + 0.010667524300515652, + 0.008837596513330936, + 0.022544559091329575, + 0.015968121588230133, + 0.013764971867203712, + 0.015274635516107082 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 6, + 2 + ], + "selected_reward": -0.04864410310983658 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002816283842548728 + ], + "opposite_logprob_tokens": [ + -5.877816200256348 + ], + "logprob_reward": -0.002816283842548728, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.002816283842548728, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16651630401611328, + 0.09776217490434647, + 0.04707799106836319, + 0.05449793487787247, + 0.12875859439373016, + 0.2002149522304535, + 0.07786662131547928, + 0.10442647337913513, + 0.09253250062465668, + 0.030346430838108063 + ], + "credit_map": [ + 0.017280178144574165, + 0.014725936576724052, + 0.035709552466869354, + 0.01564950682222843, + 0.014984463341534138, + 0.019371822476387024, + 0.01269553229212761, + 0.03609931096434593, + 0.007656083442270756, + 0.007536110933870077, + 0.021225010976195335, + 0.011164735071361065, + 0.009203251451253891, + 0.009426519274711609, + 0.010900922119617462, + 0.020649539306759834, + 0.004734010435640812, + 0.004274036735296249, + 0.0070893242955207825, + 0.005115324631333351, + 0.0046180374920368195, + 0.006445067934691906, + 0.007007219363003969, + 0.0077949706465005875, + 0.003752203891053796, + 0.005257405806332827, + 0.010262267664074898, + 0.004832438658922911, + 0.0068509806878864765, + 0.005061469506472349, + 0.005367861595004797, + 0.013113305903971195, + 0.019768118858337402, + 0.012236414477229118, + 0.017207108438014984, + 0.012213330715894699, + 0.010434894822537899, + 0.020389847457408905, + 0.015967005863785744, + 0.020541878417134285, + 0.009359078481793404, + 0.013286014087498188, + 0.04081586003303528, + 0.023113977164030075, + 0.018335696309804916, + 0.012482492253184319, + 0.02846132591366768, + 0.0543605200946331, + 0.005889364518225193, + 0.010019561275839806, + 0.009996865876019001, + 0.011917374096810818, + 0.008894262835383415, + 0.009557344019412994, + 0.010205705650150776, + 0.011386139318346977, + 0.011091788299381733, + 0.009538104757666588, + 0.01872939057648182, + 0.009969172067940235, + 0.011655150912702084, + 0.011580496095120907, + 0.012766815721988678, + 0.019095556810498238, + 0.007612272165715694, + 0.013344384729862213, + 0.010564657859504223, + 0.009802058339118958, + 0.010749823413789272, + 0.011306564323604107, + 0.016515811905264854, + 0.012636927887797356, + 0.0033437495585530996, + 0.003371353028342128, + 0.0037699732929468155, + 0.0028761867433786392, + 0.00662956153973937, + 0.0033954691607505083, + 0.0034592237789183855, + 0.003500913269817829 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 4 + ], + "selected_reward": -0.002816283842548728 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, close-up slowly and steadily approaches, plunging the viewer into a deep-sea microcosm dominated by a chilling blue-green hue. The absolute focal point is the Portuguese Man-of-War jellyfish's nearly transparent, deadly tentacles, emitting an eerie, biological glow. Like deadly threads studded with countless tiny, pearl-like sacs of stinging cells, they slowly burrow into the body of a small, silvery fish with a steady, relentless rhythm. Under the faint, cold light of the tentacles themselves, the fish's scales reflect a desperate metallic luster, its wide-open eyes frozen in the terror and vacillation of its final moments. The camera meticulously captures the moment the tentacle tip pierces the skin, pushing aside the scales. The slender stinging threads penetrate the flesh, leaving a faint, necrotic white stain on the tissue surrounding the point of contact, a foreboding sign of death. The surroundings are a profound, silent darkness, with only countless suspended particles drifting silently in the dim light, creating a microscopic atmosphere that is both magnificent and extremely dangerous, deeply unsettling, and eerie.", + "seed": 13443, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13443_dp_4_rank_8_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04239282477647066, + "vlm_logprob": -0.04239282477647066, + "vlm_yes_no_margin": 3.55, + "vlm_token_credit": -0.04239282477647066, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05016881600022316 + ], + "opposite_logprob_tokens": [ + -3.050168752670288 + ], + "logprob_reward": -0.05016881600022316, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.05016881600022316, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.24185308814048767, + 0.12396141141653061, + 0.054177213460206985, + 0.081321582198143, + 0.05726390331983566, + 0.03159891068935394, + 0.07421249896287918, + 0.10663988441228867, + 0.0775907039642334, + 0.15138083696365356 + ], + "credit_map": [ + 0.019974635913968086, + 0.033357687294483185, + 0.030060064047574997, + 0.014422887936234474, + 0.03186427801847458, + 0.037796031683683395, + 0.021235479041934013, + 0.053142011165618896, + 0.00902624987065792, + 0.012843688018620014, + 0.029056234285235405, + 0.01046605035662651, + 0.0077095660381019115, + 0.01093495637178421, + 0.012733959592878819, + 0.031190715730190277, + 0.003450701478868723, + 0.006508917547762394, + 0.008667570538818836, + 0.004699293058365583, + 0.004639210179448128, + 0.007170617580413818, + 0.007553866133093834, + 0.011487037874758244, + 0.01085607148706913, + 0.01345926895737648, + 0.007851802743971348, + 0.007118576671928167, + 0.010167769156396389, + 0.0067286561243236065, + 0.01213369145989418, + 0.01300574280321598, + 0.006607268936932087, + 0.008090444840490818, + 0.004384052008390427, + 0.0070284511893987656, + 0.007033320143818855, + 0.010149944573640823, + 0.0071434215642511845, + 0.006826999597251415, + 0.002491707680746913, + 0.0032890192233026028, + 0.004420334938913584, + 0.0033232998102903366, + 0.005201020743697882, + 0.00448856595903635, + 0.004797779023647308, + 0.0035871812142431736, + 0.00639504985883832, + 0.01174869667738676, + 0.004749923478811979, + 0.00832098163664341, + 0.010803616605699062, + 0.007881000638008118, + 0.01541995257139206, + 0.008893278427422047, + 0.0070327152498066425, + 0.016399450600147247, + 0.004056776873767376, + 0.00801707524806261, + 0.008583886548876762, + 0.01657603308558464, + 0.03638911992311478, + 0.00958482176065445, + 0.005156330298632383, + 0.012441285885870457, + 0.008243639022111893, + 0.007462115492671728, + 0.010215939953923225, + 0.0059947967529296875, + 0.010340836830437183, + 0.017735756933689117, + 0.021742084994912148, + 0.025433219969272614, + 0.014208871871232986, + 0.015521235764026642, + 0.023420505225658417, + 0.01691761240363121, + 0.019044548273086548, + 0.015092749148607254 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 1 + ], + "selected_reward": -0.05016881600022316 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008651158772408962 + ], + "opposite_logprob_tokens": [ + -4.758651256561279 + ], + "logprob_reward": -0.008651158772408962, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008651158772408962, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26765844225883484, + 0.15359674394130707, + 0.10979315638542175, + 0.09604164212942123, + 0.07537656277418137, + 0.07340682297945023, + 0.018330499529838562, + 0.11180141568183899, + 0.024340374395251274, + 0.06965427845716476 + ], + "credit_map": [ + 0.022642508149147034, + 0.038128409534692764, + 0.032054029405117035, + 0.02064111828804016, + 0.041100893169641495, + 0.03955277428030968, + 0.023243669420480728, + 0.05029505863785744, + 0.011855058372020721, + 0.017642792314291, + 0.03385872021317482, + 0.00861587468534708, + 0.011959538795053959, + 0.016979284584522247, + 0.011770844459533691, + 0.040914636105298996, + 0.005455770995467901, + 0.01106688380241394, + 0.01872781477868557, + 0.006427492946386337, + 0.012066752649843693, + 0.014223203994333744, + 0.01326159667223692, + 0.028563642874360085, + 0.008274471387267113, + 0.010039889253675938, + 0.012995699420571327, + 0.011571134440600872, + 0.009992741048336029, + 0.00870048813521862, + 0.018929392099380493, + 0.015537833794951439, + 0.010957595892250538, + 0.010082816705107689, + 0.007861025631427765, + 0.007105992641299963, + 0.008980709128081799, + 0.013918290846049786, + 0.009967070072889328, + 0.006503061391413212, + 0.004362135659903288, + 0.007963079027831554, + 0.00984534528106451, + 0.008681966923177242, + 0.009601671248674393, + 0.010459573939442635, + 0.01464952901005745, + 0.007843519560992718, + 0.0017421533120796084, + 0.002139471471309662, + 0.0012984604109078646, + 0.0019810509402304888, + 0.002819194458425045, + 0.0030968869104981422, + 0.0024364045821130276, + 0.0028168768621981144, + 0.007391580380499363, + 0.00974305160343647, + 0.0035999128594994545, + 0.005898626055568457, + 0.007273965980857611, + 0.022545386105775833, + 0.04501572996377945, + 0.010333163663744926, + 0.0019524479284882545, + 0.002620684914290905, + 0.002857266692444682, + 0.003217339515686035, + 0.002253512851893902, + 0.004607332404702902, + 0.0034429950173944235, + 0.0033887948375195265, + 0.004960882943123579, + 0.009246702305972576, + 0.004487480968236923, + 0.0043688262812793255, + 0.011359824799001217, + 0.014105782844126225, + 0.013233575038611889, + 0.007891200482845306 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.008651158772408962 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00978483073413372 + ], + "opposite_logprob_tokens": [ + -4.634784698486328 + ], + "logprob_reward": -0.00978483073413372, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.00978483073413372, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15767961740493774, + 0.04349755868315697, + 0.05670514702796936, + 0.0835423469543457, + 0.08219823986291885, + 0.06179369240999222, + 0.029185455292463303, + 0.21433408558368683, + 0.06808071583509445, + 0.20298317074775696 + ], + "credit_map": [ + 0.013879923149943352, + 0.02242421731352806, + 0.022163035348057747, + 0.014164939522743225, + 0.019228601828217506, + 0.02222665399312973, + 0.0160357803106308, + 0.02755645476281643, + 0.004585140384733677, + 0.005473226308822632, + 0.007171332836151123, + 0.0037648382131010294, + 0.0033691031858325005, + 0.005078798159956932, + 0.006058951374143362, + 0.007996163330972195, + 0.003178509185090661, + 0.004812678787857294, + 0.008931885473430157, + 0.00470188120380044, + 0.004304713103920221, + 0.009850543923676014, + 0.007410725113004446, + 0.013514203019440174, + 0.007807320915162563, + 0.010839461348950863, + 0.010037959553301334, + 0.007383720949292183, + 0.009650102816522121, + 0.0069062658585608006, + 0.017774712294340134, + 0.013142799027264118, + 0.010353181511163712, + 0.012012086808681488, + 0.009003986604511738, + 0.006164050195366144, + 0.010376119054853916, + 0.014525639824569225, + 0.007000577170401812, + 0.012762594036757946, + 0.006007459945976734, + 0.004508033394813538, + 0.016672585159540176, + 0.005850335117429495, + 0.004294363781809807, + 0.008654222823679447, + 0.00700268242508173, + 0.008804005570709705, + 0.002603371161967516, + 0.005217909347265959, + 0.0020610534120351076, + 0.0038918182253837585, + 0.004246464464813471, + 0.0036676903255283833, + 0.003706899704411626, + 0.003790246555581689, + 0.012763693928718567, + 0.02239234745502472, + 0.0053606415167450905, + 0.012379659339785576, + 0.01382464449852705, + 0.0430300347507, + 0.08532024174928665, + 0.01926279440522194, + 0.005457448773086071, + 0.009232675656676292, + 0.007961057126522064, + 0.005541996099054813, + 0.007782287430018187, + 0.006401744671165943, + 0.009618747048079967, + 0.016084754839539528, + 0.026545757427811623, + 0.03162189573049545, + 0.01674485392868519, + 0.015899086371064186, + 0.02938057854771614, + 0.017366468906402588, + 0.04975993186235428, + 0.015664583072066307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 9, + 0 + ], + "selected_reward": -0.00978483073413372 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10031053423881531 + ], + "opposite_logprob_tokens": [ + -2.3503105640411377 + ], + "logprob_reward": -0.10031053423881531, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10031053423881531, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3477741479873657, + 0.06217025965452194, + 0.04571668803691864, + 0.05015506222844124, + 0.15721386671066284, + 0.061605729162693024, + 0.03664599359035492, + 0.11689367890357971, + 0.052054621279239655, + 0.06977003067731857 + ], + "credit_map": [ + 0.03014417737722397, + 0.05264453962445259, + 0.03917581960558891, + 0.022574253380298615, + 0.05073359236121178, + 0.051039181649684906, + 0.033771079033613205, + 0.06769149005413055, + 0.0065231285989284515, + 0.00940688792616129, + 0.00939656887203455, + 0.005778050050139427, + 0.006907131988555193, + 0.005987001117318869, + 0.008427251130342484, + 0.009744237177073956, + 0.0027500037103891373, + 0.0058431727811694145, + 0.0069192396476864815, + 0.0041863746009767056, + 0.004110903013497591, + 0.0058117397129535675, + 0.006588263437151909, + 0.00950698647648096, + 0.005500746425241232, + 0.005671500228345394, + 0.006814956199377775, + 0.006011119578033686, + 0.005147415678948164, + 0.004144458565860987, + 0.008409718982875347, + 0.008455137722194195, + 0.01640329696238041, + 0.02345733530819416, + 0.013410842046141624, + 0.01530909352004528, + 0.020240837708115578, + 0.022927064448595047, + 0.017600687220692635, + 0.027864698320627213, + 0.004986517131328583, + 0.00821528397500515, + 0.008624398149549961, + 0.006071955431252718, + 0.008012603037059307, + 0.008662622421979904, + 0.00855335034430027, + 0.008478990755975246, + 0.003257337724789977, + 0.004948257002979517, + 0.0024037533439695835, + 0.004534389358013868, + 0.005559117998927832, + 0.004291963297873735, + 0.005555022973567247, + 0.006096146535128355, + 0.007250610738992691, + 0.009838680736720562, + 0.005129109136760235, + 0.00854466576129198, + 0.008492693305015564, + 0.023222707211971283, + 0.04246674105525017, + 0.01194846536964178, + 0.004533515777438879, + 0.00854812003672123, + 0.007268194574862719, + 0.005518433637917042, + 0.005736884195357561, + 0.005254690069705248, + 0.005950598046183586, + 0.009244181215763092, + 0.007045576348900795, + 0.011090553365647793, + 0.006363419350236654, + 0.004194417968392372, + 0.010502263903617859, + 0.014006908982992172, + 0.008043683134019375, + 0.008523198775947094 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 7 + ], + "selected_reward": -0.10031053423881531 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043048784136772156 + ], + "opposite_logprob_tokens": [ + -3.168048858642578 + ], + "logprob_reward": -0.043048784136772156, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043048784136772156, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11909060925245285, + 0.07393182069063187, + 0.05945449694991112, + 0.05267445743083954, + 0.2680644392967224, + 0.1121702492237091, + 0.07751595973968506, + 0.10377248376607895, + 0.032468102872371674, + 0.10085739195346832 + ], + "credit_map": [ + 0.01221007201820612, + 0.014050616882741451, + 0.015059324912726879, + 0.010904556140303612, + 0.01312121283262968, + 0.01945945993065834, + 0.016091585159301758, + 0.018193764612078667, + 0.006728684529662132, + 0.005789873190224171, + 0.016454283148050308, + 0.005142162553966045, + 0.003988174721598625, + 0.00788621511310339, + 0.009292486123740673, + 0.018649937584996223, + 0.00416514091193676, + 0.006655529141426086, + 0.009776162914931774, + 0.0048549482598900795, + 0.00477186031639576, + 0.006936338730156422, + 0.008826063945889473, + 0.013468445278704166, + 0.0064042783342301846, + 0.008562030270695686, + 0.007046036422252655, + 0.005438921041786671, + 0.007116704247891903, + 0.0044072517193853855, + 0.004746320191770792, + 0.008952910080552101, + 0.03215865045785904, + 0.04002966731786728, + 0.019227322190999985, + 0.021995436400175095, + 0.04224041849374771, + 0.042045384645462036, + 0.03424053266644478, + 0.03612698242068291, + 0.008541104383766651, + 0.010364016517996788, + 0.019141295924782753, + 0.011475210078060627, + 0.014222527854144573, + 0.016495216637849808, + 0.02067313715815544, + 0.011257720179855824, + 0.00919017568230629, + 0.011511769145727158, + 0.004325052723288536, + 0.0068571558222174644, + 0.011738575994968414, + 0.010997487232089043, + 0.012950042262673378, + 0.00994569156318903, + 0.006087759509682655, + 0.007865389809012413, + 0.004358092788606882, + 0.006024311762303114, + 0.006508887745440006, + 0.02273433655500412, + 0.03999628126621246, + 0.010197412222623825, + 0.0024500908330082893, + 0.003117782063782215, + 0.0027156230062246323, + 0.0018792353803291917, + 0.0046107773669064045, + 0.008215452544391155, + 0.003504202701151371, + 0.005974933970719576, + 0.01265168096870184, + 0.01326481718569994, + 0.00764439394697547, + 0.007160192355513573, + 0.01762469857931137, + 0.015161728486418724, + 0.016552455723285675, + 0.010797410272061825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 5 + ], + "selected_reward": -0.043048784136772156 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, close-up slowly and steadily approaches, plunging the viewer into a deep-sea microcosm dominated by a chilling blue-green hue. The absolute focal point is the Portuguese Man-of-War jellyfish's nearly transparent, deadly tentacles, emitting an eerie, biological glow. Like deadly threads studded with countless tiny, pearl-like sacs of stinging cells, they slowly burrow into the body of a small, silvery fish with a steady, relentless rhythm. Under the faint, cold light of the tentacles themselves, the fish's scales reflect a desperate metallic luster, its wide-open eyes frozen in the terror and vacillation of its final moments. The camera meticulously captures the moment the tentacle tip pierces the skin, pushing aside the scales. The slender stinging threads penetrate the flesh, leaving a faint, necrotic white stain on the tissue surrounding the point of contact, a foreboding sign of death. The surroundings are a profound, silent darkness, with only countless suspended particles drifting silently in the dim light, creating a microscopic atmosphere that is both magnificent and extremely dangerous, deeply unsettling, and eerie.", + "seed": 13444, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13444_dp_5_rank_10_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.14167374335229396, + "vlm_logprob": -0.14167374335229396, + "vlm_yes_no_margin": 2.8499999761581423, + "vlm_token_credit": -0.14167374335229396, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4758312404155731 + ], + "opposite_logprob_tokens": [ + -0.9758312702178955 + ], + "logprob_reward": -0.4758312404155731, + "yes_no_margin_reward": 0.5, + "token_credit_reward": -0.4758312404155731, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09374885261058807, + 0.05396644398570061, + 0.14204290509223938, + 0.05257004126906395, + 0.05561132729053497, + 0.3067047595977783, + 0.0788276344537735, + 0.1334128975868225, + 0.026468927040696144, + 0.056646183133125305 + ], + "credit_map": [ + 0.0123138427734375, + 0.011792158707976341, + 0.008608490228652954, + 0.007034232374280691, + 0.017324186861515045, + 0.016822613775730133, + 0.007558008655905724, + 0.012295314110815525, + 0.003952853847295046, + 0.004397707059979439, + 0.006300170905888081, + 0.006212922744452953, + 0.0060605392791330814, + 0.007426199968904257, + 0.011399946175515652, + 0.008216104470193386, + 0.010865568183362484, + 0.015608802437782288, + 0.017157454043626785, + 0.017099913209676743, + 0.01411767303943634, + 0.023222001269459724, + 0.023786919191479683, + 0.02018456719815731, + 0.0046258470974862576, + 0.007159735541790724, + 0.004490652587264776, + 0.008804450742900372, + 0.006165463477373123, + 0.006168385036289692, + 0.00887876097112894, + 0.006276744417846203, + 0.006414446048438549, + 0.005668445024639368, + 0.010523786768317223, + 0.0066943042911589146, + 0.004744526464492083, + 0.00755486823618412, + 0.005036858841776848, + 0.008974091149866581, + 0.04725959897041321, + 0.03536035865545273, + 0.051948241889476776, + 0.03864302113652229, + 0.02398986741900444, + 0.03696807846426964, + 0.03264710307121277, + 0.03988848626613617, + 0.005543410778045654, + 0.007480314467102289, + 0.007875896990299225, + 0.012875097803771496, + 0.012854264117777348, + 0.010999339632689953, + 0.008335667662322521, + 0.012863640673458576, + 0.011905081570148468, + 0.016883157193660736, + 0.014750142581760883, + 0.028264250606298447, + 0.011073226109147072, + 0.01823679730296135, + 0.01817299611866474, + 0.014127247035503387, + 0.0026448690332472324, + 0.00201621581800282, + 0.002072085626423359, + 0.005181970540434122, + 0.0017765633529052138, + 0.005026018712669611, + 0.0025265133008360863, + 0.005224688909947872, + 0.005315973423421383, + 0.005928292404860258, + 0.006373514886945486, + 0.00872118305414915, + 0.004303323104977608, + 0.010352292098104954, + 0.005740735679864883, + 0.009910868480801582 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 7 + ], + "selected_reward": -0.4758312404155731 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657068014144897 + ], + "opposite_logprob_tokens": [ + -4.758656978607178 + ], + "logprob_reward": -0.008657068014144897, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657068014144897, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11265230178833008, + 0.1107107400894165, + 0.08205115795135498, + 0.0651475191116333, + 0.02945174276828766, + 0.24993644654750824, + 0.07329245656728745, + 0.07020803540945053, + 0.04751782491803169, + 0.15903180837631226 + ], + "credit_map": [ + 0.015378192998468876, + 0.013418327085673809, + 0.009038509801030159, + 0.007985814474523067, + 0.026837455108761787, + 0.015593362040817738, + 0.011012519709765911, + 0.013388117775321007, + 0.006572471000254154, + 0.009107286110520363, + 0.012950482778251171, + 0.012999716214835644, + 0.013014565221965313, + 0.014196270145475864, + 0.022131912410259247, + 0.019738031551241875, + 0.005344281438738108, + 0.00975004956126213, + 0.010063166730105877, + 0.012006743811070919, + 0.009822740219533443, + 0.014512976631522179, + 0.009743625298142433, + 0.01080756913870573, + 0.005544685758650303, + 0.00624049361795187, + 0.007573965471237898, + 0.007149756420403719, + 0.005928033497184515, + 0.015205081552267075, + 0.006514259148389101, + 0.010991246439516544, + 0.00304452795535326, + 0.0027772923931479454, + 0.0043847160413861275, + 0.003651834325864911, + 0.002759830793365836, + 0.004566274117678404, + 0.003765364410355687, + 0.004501901101320982, + 0.04048163443803787, + 0.02518279477953911, + 0.03309301286935806, + 0.0311258677393198, + 0.02097325026988983, + 0.03502177819609642, + 0.028268126770853996, + 0.035789988934993744, + 0.005793298128992319, + 0.009608552791178226, + 0.008589452132582664, + 0.011951453983783722, + 0.008007938042283058, + 0.011490242555737495, + 0.0063967108726501465, + 0.011454805731773376, + 0.006989589426666498, + 0.008266380056738853, + 0.008701418526470661, + 0.013822624459862709, + 0.006479318719357252, + 0.008579898625612259, + 0.010237399488687515, + 0.007131401915103197, + 0.009952615946531296, + 0.004082502331584692, + 0.005327408667653799, + 0.0055958060547709465, + 0.0026799910701811314, + 0.006567453965544701, + 0.0058609917759895325, + 0.007451054640114307, + 0.02030080370604992, + 0.012449466623365879, + 0.011021961458027363, + 0.02065664529800415, + 0.008506549522280693, + 0.03727623447775841, + 0.013373296707868576, + 0.03544685244560242 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 9, + 0 + ], + "selected_reward": -0.008657068014144897 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075330898165703 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075330898165703, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075330898165703, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06391798704862595, + 0.03858624026179314, + 0.09108559042215347, + 0.056776512414216995, + 0.052036698907613754, + 0.24398373067378998, + 0.04746798425912857, + 0.04606309533119202, + 0.03769441321492195, + 0.3223877549171448 + ], + "credit_map": [ + 0.006504693999886513, + 0.006965266540646553, + 0.004473258275538683, + 0.0046684336848556995, + 0.013066265732049942, + 0.014871596358716488, + 0.005485665984451771, + 0.007882801815867424, + 0.0025333960074931383, + 0.0046019661240279675, + 0.005230247043073177, + 0.00336682191118598, + 0.005211597308516502, + 0.0056024882942438126, + 0.00503239082172513, + 0.007007329259067774, + 0.006504573859274387, + 0.009737626649439335, + 0.009873145259916782, + 0.012453977949917316, + 0.008339764550328255, + 0.014795464463531971, + 0.015963612124323845, + 0.013417422771453857, + 0.004538477398455143, + 0.006544134579598904, + 0.006105055101215839, + 0.00961743202060461, + 0.005229681730270386, + 0.008126700296998024, + 0.0090398658066988, + 0.007575168274343014, + 0.005585437174886465, + 0.004306089598685503, + 0.007540049962699413, + 0.007757934276014566, + 0.003218222176656127, + 0.0074655343778431416, + 0.006907007656991482, + 0.009256421588361263, + 0.044158369302749634, + 0.026552481576800346, + 0.03762434795498848, + 0.030598416924476624, + 0.018925562500953674, + 0.02780986949801445, + 0.026740120723843575, + 0.031574562191963196, + 0.00314992293715477, + 0.005053554195910692, + 0.005568633787333965, + 0.009075134992599487, + 0.005718432366847992, + 0.006767034064978361, + 0.005093788728117943, + 0.0070414855144917965, + 0.0033193309791386127, + 0.006777070928364992, + 0.004905302543193102, + 0.005467163398861885, + 0.00577453151345253, + 0.007456415798515081, + 0.007167457137256861, + 0.005195823032408953, + 0.008247479796409607, + 0.003231261856853962, + 0.004463863093405962, + 0.004241107031702995, + 0.0021385704167187214, + 0.00516144186258316, + 0.004574934486299753, + 0.005635752808302641, + 0.03406187891960144, + 0.02013477310538292, + 0.01462758332490921, + 0.03566838055849075, + 0.012216306291520596, + 0.09237570315599442, + 0.019748549908399582, + 0.09355459362268448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 5, + 2 + ], + "selected_reward": -0.011075330898165703 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14280159771442413 + ], + "opposite_logprob_tokens": [ + -2.017801523208618 + ], + "logprob_reward": -0.14280159771442413, + "yes_no_margin_reward": 1.8749998807907104, + "token_credit_reward": -0.14280159771442413, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.045375723391771317, + 0.10455897450447083, + 0.14523851871490479, + 0.04936075955629349, + 0.06982148438692093, + 0.3058057129383087, + 0.09997369349002838, + 0.06539799273014069, + 0.03733639791607857, + 0.0771307423710823 + ], + "credit_map": [ + 0.00589331891387701, + 0.0049967472441494465, + 0.004129831679165363, + 0.004873707424849272, + 0.009583525359630585, + 0.005893490742892027, + 0.005440876353532076, + 0.004564224276691675, + 0.007068349048495293, + 0.007622912991791964, + 0.009683812037110329, + 0.011722990311682224, + 0.01238294318318367, + 0.015375399962067604, + 0.019923051819205284, + 0.020779509097337723, + 0.010082133114337921, + 0.014608532190322876, + 0.014446995221078396, + 0.019550666213035583, + 0.013456684537231922, + 0.02579347789287567, + 0.02616085857152939, + 0.02113915979862213, + 0.004827541299164295, + 0.005609579384326935, + 0.006321574095636606, + 0.007690232712775469, + 0.0038384220097213984, + 0.00706274900585413, + 0.007184420712292194, + 0.006826236844062805, + 0.010878367349505424, + 0.005727165378630161, + 0.010711880400776863, + 0.00798423495143652, + 0.005533514078706503, + 0.010094777680933475, + 0.0063763950020074844, + 0.012515142560005188, + 0.05775096267461777, + 0.03147473931312561, + 0.048542845994234085, + 0.034617554396390915, + 0.026210999116301537, + 0.03676743432879448, + 0.0309331975877285, + 0.039507970213890076, + 0.007767559494823217, + 0.008770604617893696, + 0.010493428446352482, + 0.021632647141814232, + 0.007965074852108955, + 0.017094505950808525, + 0.009515620768070221, + 0.016734246164560318, + 0.004905082751065493, + 0.007103088777512312, + 0.008553643710911274, + 0.011327496729791164, + 0.005732789635658264, + 0.011770535260438919, + 0.007458236068487167, + 0.00854711327701807, + 0.00744842691347003, + 0.003152957884594798, + 0.004764361307024956, + 0.004207902122288942, + 0.00226423772983253, + 0.0053856465965509415, + 0.0047240909188985825, + 0.005388773512095213, + 0.010005615651607513, + 0.005385478492826223, + 0.009011225774884224, + 0.00818570051342249, + 0.005639835260808468, + 0.015374054200947285, + 0.006578243803232908, + 0.016950583085417747 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 1 + ], + "selected_reward": -0.14280159771442413 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000347971916199 + ], + "opposite_logprob_tokens": [ + -2.6950035095214844 + ], + "logprob_reward": -0.07000347971916199, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000347971916199, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2603912949562073, + 0.025939621031284332, + 0.061880238354206085, + 0.03205399587750435, + 0.022954363375902176, + 0.18586018681526184, + 0.0753333643078804, + 0.030023621395230293, + 0.0514252595603466, + 0.25413796305656433 + ], + "credit_map": [ + 0.04383070766925812, + 0.03562542796134949, + 0.015475533902645111, + 0.025266611948609352, + 0.06145905703306198, + 0.038259632885456085, + 0.020487355068325996, + 0.019986988976597786, + 0.0019107174593955278, + 0.0031441666651517153, + 0.003723598551005125, + 0.002678066259250045, + 0.0031211236491799355, + 0.003728981362655759, + 0.003134956816211343, + 0.004498014226555824, + 0.0043001919984817505, + 0.007110437843948603, + 0.007249778136610985, + 0.007020922377705574, + 0.006021591834723949, + 0.013076696544885635, + 0.008614622056484222, + 0.008486003614962101, + 0.0026677886489778757, + 0.0030773847829550505, + 0.003581192810088396, + 0.0037973078433424234, + 0.003006039885804057, + 0.0063593508675694466, + 0.004759772680699825, + 0.004805160686373711, + 0.0025418568402528763, + 0.0019936643075197935, + 0.002719397656619549, + 0.0032816596794873476, + 0.001916198874823749, + 0.004150074440985918, + 0.00282888556830585, + 0.003522627055644989, + 0.02033333294093609, + 0.020129991695284843, + 0.017457595095038414, + 0.015606852248311043, + 0.01821478083729744, + 0.0441870242357254, + 0.01888975501060486, + 0.031040871515870094, + 0.008021792396903038, + 0.007873584516346455, + 0.00916248932480812, + 0.009988918900489807, + 0.008355654776096344, + 0.013485204428434372, + 0.008784866891801357, + 0.00966085959225893, + 0.0023750485852360725, + 0.003607455873861909, + 0.0034330454654991627, + 0.0037185880355536938, + 0.0030301494989544153, + 0.00529340049251914, + 0.004610314965248108, + 0.003955620341002941, + 0.00716425571590662, + 0.004404924809932709, + 0.005038711708039045, + 0.008003459312021732, + 0.003596656024456024, + 0.00810337346047163, + 0.006439367774873972, + 0.008674515411257744, + 0.027253642678260803, + 0.01729872077703476, + 0.010753985494375229, + 0.025268860161304474, + 0.009464267641305923, + 0.07399440556764603, + 0.015782853588461876, + 0.07432126253843307 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 9, + 5 + ], + "selected_reward": -0.07000347971916199 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "steady, close-up slowly and steadily approaches, plunging the viewer into a deep-sea microcosm dominated by a chilling blue-green hue. The absolute focal point is the Portuguese Man-of-War jellyfish's nearly transparent, deadly tentacles, emitting an eerie, biological glow. Like deadly threads studded with countless tiny, pearl-like sacs of stinging cells, they slowly burrow into the body of a small, silvery fish with a steady, relentless rhythm. Under the faint, cold light of the tentacles themselves, the fish's scales reflect a desperate metallic luster, its wide-open eyes frozen in the terror and vacillation of its final moments. The camera meticulously captures the moment the tentacle tip pierces the skin, pushing aside the scales. The slender stinging threads penetrate the flesh, leaving a faint, necrotic white stain on the tissue surrounding the point of contact, a foreboding sign of death. The surroundings are a profound, silent darkness, with only countless suspended particles drifting silently in the dim light, creating a microscopic atmosphere that is both magnificent and extremely dangerous, deeply unsettling, and eerie.", + "seed": 13445, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_2/index_9075_seed_13445_dp_5_rank_10_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 2, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.03218635553494096, + "vlm_logprob": -0.03218635553494096, + "vlm_yes_no_margin": 4.0, + "vlm_token_credit": -0.03218635553494096, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.044596295803785324 + ], + "opposite_logprob_tokens": [ + -3.1695961952209473 + ], + "logprob_reward": -0.044596295803785324, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.044596295803785324, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12429648637771606, + 0.08067912608385086, + 0.08002501726150513, + 0.138633131980896, + 0.20832210779190063, + 0.04352755472064018, + 0.04044971615076065, + 0.09799935668706894, + 0.13447657227516174, + 0.0515909418463707 + ], + "credit_map": [ + 0.010024120099842548, + 0.01218312606215477, + 0.012516568414866924, + 0.01493985764682293, + 0.011474424973130226, + 0.023696016520261765, + 0.011576448567211628, + 0.027885936200618744, + 0.007255973294377327, + 0.014007954858243465, + 0.005273019894957542, + 0.009796732105314732, + 0.007834442891180515, + 0.013226611539721489, + 0.01013402733951807, + 0.013150367885828018, + 0.0058356341905891895, + 0.004878573585301638, + 0.0072733499109745026, + 0.011753180995583534, + 0.007637204136699438, + 0.015703920274972916, + 0.009913972578942776, + 0.017029186710715294, + 0.009978190064430237, + 0.01157309114933014, + 0.00869393814355135, + 0.0174262672662735, + 0.014903396368026733, + 0.02754427120089531, + 0.027418412268161774, + 0.021095559000968933, + 0.0197321604937315, + 0.029948413372039795, + 0.015648361295461655, + 0.017451032996177673, + 0.02431529015302658, + 0.033330999314785004, + 0.036899272352457047, + 0.03099658340215683, + 0.0050179921090602875, + 0.006595936603844166, + 0.004559704568237066, + 0.004286917392164469, + 0.0045667593367397785, + 0.006849211174994707, + 0.0065895733423531055, + 0.005061458796262741, + 0.0030361576937139034, + 0.004035753197968006, + 0.00524570606648922, + 0.0054605333134531975, + 0.005132964346557856, + 0.006255360785871744, + 0.006158966105431318, + 0.005124276969581842, + 0.010398613288998604, + 0.009526271373033524, + 0.008861672133207321, + 0.017161786556243896, + 0.010791327804327011, + 0.012411974370479584, + 0.019151171669363976, + 0.00969653856009245, + 0.01107898447662592, + 0.011298257857561111, + 0.01806017756462097, + 0.018188657239079475, + 0.024539019912481308, + 0.019192533567547798, + 0.013718256726861, + 0.018400689586997032, + 0.00582388648763299, + 0.00856451690196991, + 0.008251645602285862, + 0.007095469627529383, + 0.005628480575978756, + 0.005845061969012022, + 0.003654049476608634, + 0.006727830972522497 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 3, + 8 + ], + "selected_reward": -0.044596295803785324 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0052668387070298195 + ], + "opposite_logprob_tokens": [ + -5.2552666664123535 + ], + "logprob_reward": -0.0052668387070298195, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.0052668387070298195, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09693324565887451, + 0.0635235607624054, + 0.12917780876159668, + 0.03970944136381149, + 0.05875325947999954, + 0.03711828589439392, + 0.07730339467525482, + 0.3217749297618866, + 0.09569580107927322, + 0.08001027256250381 + ], + "credit_map": [ + 0.009653301909565926, + 0.01058162096887827, + 0.009265697561204433, + 0.008610299788415432, + 0.00901311170309782, + 0.01871887966990471, + 0.009520742110908031, + 0.02156957983970642, + 0.006651552394032478, + 0.008826415054500103, + 0.004182996228337288, + 0.006512385793030262, + 0.005865310784429312, + 0.010832090862095356, + 0.006446640007197857, + 0.01420616451650858, + 0.012585380114614964, + 0.009472494944930077, + 0.01966542936861515, + 0.018867429345846176, + 0.01377976592630148, + 0.02291366085410118, + 0.011656652204692364, + 0.020236987620592117, + 0.005138051230460405, + 0.006330272648483515, + 0.005021786782890558, + 0.0027727163396775723, + 0.004041797947138548, + 0.006563607603311539, + 0.004309640731662512, + 0.005531564820557833, + 0.00591916823759675, + 0.007340000476688147, + 0.008077401667833328, + 0.004873488564044237, + 0.006373964250087738, + 0.008491172455251217, + 0.009020945988595486, + 0.008657113648951054, + 0.003785129636526108, + 0.0030133824329823256, + 0.004379500634968281, + 0.005217443685978651, + 0.0022630293387919664, + 0.00592762790620327, + 0.007893737405538559, + 0.004638430662453175, + 0.006623845547437668, + 0.0063671027310192585, + 0.00836406834423542, + 0.009070871397852898, + 0.009574235416948795, + 0.012994077987968922, + 0.011662044562399387, + 0.012647145427763462, + 0.033621687442064285, + 0.03542694076895714, + 0.03289339318871498, + 0.03790317475795746, + 0.05420161783695221, + 0.04822567477822304, + 0.043941568583250046, + 0.035560864955186844, + 0.008093715645372868, + 0.010033270344138145, + 0.013445583172142506, + 0.009841901250183582, + 0.017237026244401932, + 0.01727421209216118, + 0.010584701783955097, + 0.009185380302369595, + 0.014843166805803776, + 0.012068675830960274, + 0.009200366213917732, + 0.008262007497251034, + 0.009349379688501358, + 0.010691499337553978, + 0.006338956765830517, + 0.009256220422685146 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 2, + 0 + ], + "selected_reward": -0.0052668387070298195 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.016080450266599655 + ], + "opposite_logprob_tokens": [ + -4.141080379486084 + ], + "logprob_reward": -0.016080450266599655, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.016080450266599655, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10163116455078125, + 0.05340217426419258, + 0.060379739850759506, + 0.06869920343160629, + 0.10581459105014801, + 0.07340124994516373, + 0.13754212856292725, + 0.1757531464099884, + 0.1461358517408371, + 0.0772407054901123 + ], + "credit_map": [ + 0.009095925837755203, + 0.008734691888093948, + 0.01081202831119299, + 0.00942565593868494, + 0.010020346380770206, + 0.019524330273270607, + 0.009988821111619473, + 0.024029366672039032, + 0.006613938603550196, + 0.006089942064136267, + 0.006070360075682402, + 0.005697113461792469, + 0.004393618553876877, + 0.008881666697561741, + 0.005404334980994463, + 0.010251201689243317, + 0.005131334997713566, + 0.004298517946153879, + 0.00590201560407877, + 0.00905293133109808, + 0.005256631411612034, + 0.01177056971937418, + 0.006167674902826548, + 0.01280006393790245, + 0.005806155037134886, + 0.00680193305015564, + 0.005859140772372484, + 0.007943511940538883, + 0.0069964611902832985, + 0.013424667529761791, + 0.011610133573412895, + 0.010257198475301266, + 0.007347084581851959, + 0.012785506434738636, + 0.011083263903856277, + 0.008336571976542473, + 0.008632272481918335, + 0.01888171024620533, + 0.014864158816635609, + 0.02388402447104454, + 0.0059532057493925095, + 0.007096779998391867, + 0.005908344406634569, + 0.010279516689479351, + 0.004241408314555883, + 0.011652424931526184, + 0.018960144370794296, + 0.00930942315608263, + 0.016501793637871742, + 0.012450897134840488, + 0.017380274832248688, + 0.015268239192664623, + 0.015490387566387653, + 0.021220020949840546, + 0.019968951120972633, + 0.01926155388355255, + 0.020833611488342285, + 0.019693301990628242, + 0.015246073715388775, + 0.021554261445999146, + 0.03249168395996094, + 0.02584577538073063, + 0.024309253320097923, + 0.015779174864292145, + 0.01063545886427164, + 0.011413206346333027, + 0.018469607457518578, + 0.017175495624542236, + 0.021473905071616173, + 0.030346522107720375, + 0.018594011664390564, + 0.01802765019237995, + 0.010360737331211567, + 0.015638526529073715, + 0.01007319986820221, + 0.009501175954937935, + 0.007139597553759813, + 0.008359267376363277, + 0.006012453231960535, + 0.010155744850635529 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 6 + ], + "selected_reward": -0.016080450266599655 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08904317021369934 + ], + "opposite_logprob_tokens": [ + -2.464043140411377 + ], + "logprob_reward": -0.08904317021369934, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.08904317021369934, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14992573857307434, + 0.08646807074546814, + 0.0565306693315506, + 0.06022972613573074, + 0.07368530333042145, + 0.06956986337900162, + 0.1917482614517212, + 0.12163855135440826, + 0.09972068667411804, + 0.0904831737279892 + ], + "credit_map": [ + 0.012365972623229027, + 0.01928688958287239, + 0.013961175456643105, + 0.01967213675379753, + 0.015741704031825066, + 0.025953048840165138, + 0.016556616872549057, + 0.026388194411993027, + 0.008473824709653854, + 0.009565887041389942, + 0.00790376029908657, + 0.0111264418810606, + 0.008421371690928936, + 0.014786496758460999, + 0.008897884748876095, + 0.01729239709675312, + 0.005854313727468252, + 0.0046625458635389805, + 0.005931516643613577, + 0.004727728199213743, + 0.006254660431295633, + 0.011585553176701069, + 0.004905341193079948, + 0.012609007768332958, + 0.004898322280496359, + 0.008300541900098324, + 0.005622625816613436, + 0.005575723946094513, + 0.006104030646383762, + 0.010194933973252773, + 0.008589230477809906, + 0.010944319888949394, + 0.005255257245153189, + 0.009943695738911629, + 0.011095508001744747, + 0.005945697892457247, + 0.005069356877356768, + 0.01076939795166254, + 0.013742029666900635, + 0.011864356696605682, + 0.006910531781613827, + 0.007575624156743288, + 0.007493131794035435, + 0.007064953446388245, + 0.006076070945709944, + 0.0108583252876997, + 0.016508933156728745, + 0.00708229374140501, + 0.013237034901976585, + 0.016042904928326607, + 0.018096189945936203, + 0.025251857936382294, + 0.017809266224503517, + 0.030874961987137794, + 0.041182536631822586, + 0.029253516346216202, + 0.012478887103497982, + 0.009876695461571217, + 0.01622295007109642, + 0.016822565346956253, + 0.01846950873732567, + 0.0160240288823843, + 0.019204173237085342, + 0.012539739720523357, + 0.007795424200594425, + 0.008635601960122585, + 0.011566739529371262, + 0.011073027737438679, + 0.01766129769384861, + 0.01587962545454502, + 0.012888789176940918, + 0.01422017440199852, + 0.015422730706632137, + 0.019581547006964684, + 0.009470061399042606, + 0.008145959116518497, + 0.011212698183953762, + 0.007036997936666012, + 0.007533687632530928, + 0.012079484760761261 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 7 + ], + "selected_reward": -0.08904317021369934 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a Portuguese Man-of-War jellyfish attacking a small silvery fish?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Are the jellyfish's tentacles nearly transparent with a biological glow and pearl-like sacs?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the fish display wide-open eyes and scales reflecting a metallic luster?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is there a visible faint necrotic white stain on the fish's tissue where the tentacle penetrates?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the environment a deep-sea setting with profound darkness and suspended particles?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059450226835906506 + ], + "opposite_logprob_tokens": [ + -5.130945205688477 + ], + "logprob_reward": -0.0059450226835906506, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059450226835906506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23570075631141663, + 0.22907479107379913, + 0.08368083834648132, + 0.047964926809072495, + 0.03396500274538994, + 0.016475196927785873, + 0.07182422280311584, + 0.13630518317222595, + 0.06975464522838593, + 0.07525436580181122 + ], + "credit_map": [ + 0.023434294387698174, + 0.0167391374707222, + 0.02113981544971466, + 0.024395186454057693, + 0.02000156044960022, + 0.04272260144352913, + 0.029929306358098984, + 0.05733887478709221, + 0.022973090410232544, + 0.016618400812149048, + 0.01858670637011528, + 0.015077703632414341, + 0.017817089334130287, + 0.04431574046611786, + 0.027582885697484016, + 0.06610319018363953, + 0.0077676367945969105, + 0.0078065968118608, + 0.006846147123724222, + 0.007849691435694695, + 0.008333271369338036, + 0.016181690618395805, + 0.008949542418122292, + 0.019946271553635597, + 0.004636711440980434, + 0.0072322022169828415, + 0.004430430941283703, + 0.0041308216750621796, + 0.005272601265460253, + 0.008001776412129402, + 0.0075322785414755344, + 0.006728108040988445, + 0.0032367161475121975, + 0.0042429217137396336, + 0.0035068513825535774, + 0.0033347175922244787, + 0.004049083683639765, + 0.005443417001515627, + 0.0047181458212435246, + 0.00543315289542079, + 0.0013455178122967482, + 0.002415651921182871, + 0.002945787040516734, + 0.0013685058802366257, + 0.0017361858626827598, + 0.0024427049793303013, + 0.00205208919942379, + 0.002168754581362009, + 0.009797363542020321, + 0.007421411573886871, + 0.006600138731300831, + 0.0076912641525268555, + 0.009175900369882584, + 0.011037603951990604, + 0.010814808309078217, + 0.009285741485655308, + 0.018291043117642403, + 0.013181372545659542, + 0.012465555220842361, + 0.019535919651389122, + 0.02259357087314129, + 0.01797787845134735, + 0.02120297960937023, + 0.011056878603994846, + 0.007393077947199345, + 0.008800359442830086, + 0.013075396418571472, + 0.010455570183694363, + 0.010746209882199764, + 0.005960410460829735, + 0.0050733149982988834, + 0.008250313811004162, + 0.011615609750151634, + 0.00887005403637886, + 0.014373165555298328, + 0.00944325141608715, + 0.008506402373313904, + 0.005056825466454029, + 0.006046845577657223, + 0.011342211626470089 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.0059450226835906506 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_tensor([2934, 2934])_group_3_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_tensor([2934, 2934])_group_3_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..49b42650d307be5321c2d7df650e00ee7f556897 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_tensor([2934, 2934])_group_3_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "shot shot from a level angle begins in medium shot, then slowly pulls back to reveal a wider scene. The frame focuses on a silver-haired elderly man, clad in a loose, white, traditional cotton and linen training suit, nestled in a tranquil bamboo forest. The faint morning sun, the sole source of light, penetrates the towering, verdant bamboo poles and dense leaves, creating soft, voluminous beams that are visible through the misty air and cast shifting, dappled shadows on the damp, leaf-covered ground. The elderly man's hair is meticulously combed, his face deeply wrinkled, weathered. His eyes are slightly closed, his expression focused and peaceful. He is performing Tai Chi with incredible slowness and stability, his arms tracing smooth arcs at a speed barely perceptible to the naked eye, his weight shifting effortlessly between his legs. Each movement is imbued with inner, composed strength. His wide white sleeves fluttered gently as his body turned, and their trajectory seemed to merge with the flowing air and light. There was no pause between his movements, perfectly interpreting the artistic conception of \"flowing like clouds and running water\". The whole scene presented an extremely quiet, solemn and Zen-like meditative atmosphere.", + "seed": 6320, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6320_dp_6_rank_12_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.03714155936613679, + "vlm_logprob": -0.03714155936613679, + "vlm_yes_no_margin": 4.175000023841858, + "vlm_token_credit": -0.03714155936613679, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0095902681350708 + ], + "opposite_logprob_tokens": [ + -4.884590148925781 + ], + "logprob_reward": -0.0095902681350708, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.0095902681350708, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2895842492580414, + 0.13420023024082184, + 0.01916227862238884, + 0.02744845487177372, + 0.3670598566532135, + 0.02323492243885994, + 0.03062807023525238, + 0.02867637202143669, + 0.043690312653779984, + 0.0363151840865612 + ], + "credit_map": [ + 0.045044057071208954, + 0.052287306636571884, + 0.022900735959410667, + 0.0277544017881155, + 0.04542455077171326, + 0.0578623004257679, + 0.02151261270046234, + 0.01679827645421028, + 0.016502901911735535, + 0.023749182000756264, + 0.01263775397092104, + 0.009812543168663979, + 0.0308324433863163, + 0.020317882299423218, + 0.01053172443062067, + 0.009815800003707409, + 0.002198971575126052, + 0.002763447817414999, + 0.003139261621981859, + 0.0022084799129515886, + 0.003246734384447336, + 0.0026002235244959593, + 0.0017075676005333662, + 0.0012975928839296103, + 0.0028693685308098793, + 0.002144578378647566, + 0.0055690426379442215, + 0.002640452701598406, + 0.0029522294644266367, + 0.003213725285604596, + 0.006072390824556351, + 0.001986668910831213, + 0.033661138266325, + 0.011508656665682793, + 0.13412737846374512, + 0.018162071704864502, + 0.015293476171791553, + 0.07183703780174255, + 0.06622883677482605, + 0.01624128594994545, + 0.003941322211176157, + 0.003685107221826911, + 0.0036793926265090704, + 0.001752824173308909, + 0.002798134693875909, + 0.004350721836090088, + 0.001652433886192739, + 0.0013749880017712712, + 0.006173933856189251, + 0.0022081504575908184, + 0.006399819161742926, + 0.001550283981487155, + 0.0044516585767269135, + 0.0024945572949945927, + 0.004958422388881445, + 0.002391249407082796, + 0.006024885922670364, + 0.003634765511378646, + 0.00775641854852438, + 0.0015692162560299039, + 0.0023294477723538876, + 0.0025821994058787823, + 0.002801844384521246, + 0.001977597363293171, + 0.00917486846446991, + 0.004787680692970753, + 0.005022597033530474, + 0.005592309404164553, + 0.006586420349776745, + 0.005432851612567902, + 0.003922590054571629, + 0.003170999465510249, + 0.004025546833872795, + 0.003600733121857047, + 0.004306627903133631, + 0.0045465691946446896, + 0.006719807628542185, + 0.004642857704311609, + 0.005371371749788523, + 0.003101675072684884 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.0095902681350708 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008657422848045826 + ], + "opposite_logprob_tokens": [ + -4.758657455444336 + ], + "logprob_reward": -0.008657422848045826, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008657422848045826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11763367056846619, + 0.27289119362831116, + 0.031378984451293945, + 0.03500829264521599, + 0.3149602711200714, + 0.01552089024335146, + 0.050277285277843475, + 0.07930997759103775, + 0.045769985765218735, + 0.037249427288770676 + ], + "credit_map": [ + 0.012349428609013557, + 0.013464033603668213, + 0.015593661926686764, + 0.01365671120584011, + 0.026241125538945198, + 0.01456039771437645, + 0.013562621548771858, + 0.008205701597034931, + 0.03811989724636078, + 0.05212737247347832, + 0.026661835610866547, + 0.028249913826584816, + 0.040310975164175034, + 0.03776794299483299, + 0.029056912288069725, + 0.020596345886588097, + 0.0038220358546823263, + 0.003948302008211613, + 0.005913496948778629, + 0.003827176056802273, + 0.003927323967218399, + 0.004648200236260891, + 0.0030129882507026196, + 0.0022794653195887804, + 0.0031374297104775906, + 0.0031416441779583693, + 0.006921462714672089, + 0.005976636428385973, + 0.003013867884874344, + 0.005635528825223446, + 0.004771782550960779, + 0.0024099445436149836, + 0.029639629647135735, + 0.00996160414069891, + 0.11690567433834076, + 0.015393304638564587, + 0.012922008521854877, + 0.06257046014070511, + 0.054142825305461884, + 0.013424756936728954, + 0.0017579043051227927, + 0.0019676925148814917, + 0.0020006957929581404, + 0.0013222047127783298, + 0.0021141653414815664, + 0.001965882023796439, + 0.002629917813464999, + 0.0017624285537749529, + 0.01038589607924223, + 0.007021381985396147, + 0.011787423864006996, + 0.0032946846913546324, + 0.00527229206636548, + 0.004580556880682707, + 0.004025319591164589, + 0.003909735009074211, + 0.012816126458346844, + 0.012441840022802353, + 0.014381825923919678, + 0.005278438795357943, + 0.0113037945702672, + 0.009094681590795517, + 0.006872160360217094, + 0.007121115457266569, + 0.008578172884881496, + 0.006858757697045803, + 0.0061869993805885315, + 0.008161056786775589, + 0.004042095970362425, + 0.004486000165343285, + 0.003896987298503518, + 0.0035599221009761095, + 0.005943788215517998, + 0.005046264734119177, + 0.005183560308068991, + 0.005155768245458603, + 0.004679067060351372, + 0.0035916434135288, + 0.004863969516009092, + 0.0027853669598698616 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 1, + 0 + ], + "selected_reward": -0.008657422848045826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00409490754827857 + ], + "opposite_logprob_tokens": [ + -5.504095077514648 + ], + "logprob_reward": -0.00409490754827857, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.00409490754827857, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0889899730682373, + 0.05484265089035034, + 0.09650697559118271, + 0.04480547085404396, + 0.26902034878730774, + 0.08529067784547806, + 0.025453170761466026, + 0.06978476047515869, + 0.16512952744960785, + 0.10017646104097366 + ], + "credit_map": [ + 0.008830925449728966, + 0.015087373554706573, + 0.008074773475527763, + 0.010516948997974396, + 0.015611603856086731, + 0.01720314286649227, + 0.007826784625649452, + 0.005838421173393726, + 0.00877594854682684, + 0.013120465911924839, + 0.005058028269559145, + 0.0045851003378629684, + 0.007570337038487196, + 0.008587905205786228, + 0.0038475203327834606, + 0.003297345247119665, + 0.011293475516140461, + 0.011840607039630413, + 0.027665412053465843, + 0.01178183313459158, + 0.005807290319353342, + 0.016468849033117294, + 0.00687767518684268, + 0.004771828185766935, + 0.004973436705768108, + 0.00387245649471879, + 0.008184153586626053, + 0.0041877757757902145, + 0.00468312157317996, + 0.0051689366810023785, + 0.010390915907919407, + 0.003344674361869693, + 0.023891516029834747, + 0.010065644048154354, + 0.09449949860572815, + 0.013029206544160843, + 0.011207172647118568, + 0.051100485026836395, + 0.052297335118055344, + 0.012929454445838928, + 0.015335985459387302, + 0.012964884750545025, + 0.013805578462779522, + 0.004924853332340717, + 0.005740261171013117, + 0.02114098332822323, + 0.005983713548630476, + 0.005394417326897383, + 0.0047617885284125805, + 0.0036698721814900637, + 0.004629930946975946, + 0.002324141561985016, + 0.0023586731404066086, + 0.002363076899200678, + 0.003011211520060897, + 0.0023344752844423056, + 0.012971784919500351, + 0.012689249590039253, + 0.017280902713537216, + 0.004072634968906641, + 0.004536631051450968, + 0.009993467479944229, + 0.004575985949486494, + 0.0036641026381403208, + 0.04686564579606056, + 0.02228786237537861, + 0.027615178376436234, + 0.013664892874658108, + 0.012147748842835426, + 0.017733385786414146, + 0.0097412234172225, + 0.015073596499860287, + 0.023755600675940514, + 0.018417922779917717, + 0.015654893592000008, + 0.008966969326138496, + 0.005510951858013868, + 0.017388170585036278, + 0.004365282598882914, + 0.006116668228060007 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 9 + ], + "selected_reward": -0.00409490754827857 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020590482279658318 + ], + "opposite_logprob_tokens": [ + -3.8955905437469482 + ], + "logprob_reward": -0.020590482279658318, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020590482279658318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13295364379882812, + 0.045691631734371185, + 0.03109772317111492, + 0.03732142969965935, + 0.30156809091567993, + 0.024834241718053818, + 0.034399084746837616, + 0.17288906872272491, + 0.15923704206943512, + 0.06000807136297226 + ], + "credit_map": [ + 0.022347401827573776, + 0.025377629324793816, + 0.013094304129481316, + 0.010404347442090511, + 0.018593383952975273, + 0.024616854265332222, + 0.008957481943070889, + 0.009562256745994091, + 0.004226603079587221, + 0.007208430208265781, + 0.004940222483128309, + 0.004041511565446854, + 0.009373251348733902, + 0.006656016688793898, + 0.0038100217934697866, + 0.00543557433411479, + 0.00339834694750607, + 0.0031570312567055225, + 0.0076330313459038734, + 0.0034793808590620756, + 0.0035707675851881504, + 0.004903484135866165, + 0.0028360991273075342, + 0.0021195828448981047, + 0.004276600666344166, + 0.0029720792081207037, + 0.006430601701140404, + 0.004291844554245472, + 0.004727636929601431, + 0.0039490810595452785, + 0.007999313063919544, + 0.0026742718182504177, + 0.026566121727228165, + 0.011343640275299549, + 0.10330755263566971, + 0.01450811792165041, + 0.013446762226521969, + 0.05511912330985069, + 0.06210379675030708, + 0.015172985382378101, + 0.00247294083237648, + 0.0026722298935055733, + 0.004071839153766632, + 0.002380738500505686, + 0.0025347175542265177, + 0.005131045822054148, + 0.003255629213526845, + 0.0023151005152612925, + 0.0062298704870045185, + 0.006029204465448856, + 0.007554218638688326, + 0.0029808718245476484, + 0.002957363147288561, + 0.0030834211502224207, + 0.0034721221309155226, + 0.002092010574415326, + 0.032221756875514984, + 0.03280523419380188, + 0.042724449187517166, + 0.011387236416339874, + 0.011359790340065956, + 0.020161723718047142, + 0.01163062360137701, + 0.010598256252706051, + 0.03331752493977547, + 0.02711823396384716, + 0.021279947832226753, + 0.01753557100892067, + 0.018772941082715988, + 0.01758841797709465, + 0.010465158149600029, + 0.013159248046576977, + 0.013579999096691608, + 0.0116496030241251, + 0.009177744388580322, + 0.006768832914531231, + 0.004463264252990484, + 0.006952442228794098, + 0.003599760355427861, + 0.0038164269644767046 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 8 + ], + "selected_reward": -0.020590482279658318 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14277471601963043 + ], + "opposite_logprob_tokens": [ + -2.017774820327759 + ], + "logprob_reward": -0.14277471601963043, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14277471601963043, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04556140676140785, + 0.04218103364109993, + 0.023998964577913284, + 0.022974135354161263, + 0.5247201919555664, + 0.020449543371796608, + 0.025294803082942963, + 0.09740018844604492, + 0.17892320454120636, + 0.018496518954634666 + ], + "credit_map": [ + 0.00681786984205246, + 0.007488344796001911, + 0.003716610372066498, + 0.004899810533970594, + 0.006698257755488157, + 0.008576472289860249, + 0.004291236866265535, + 0.003072798717767, + 0.006409462075680494, + 0.007796552963554859, + 0.00453612394630909, + 0.0034871050156652927, + 0.00531019875779748, + 0.0060813515447080135, + 0.0031037568114697933, + 0.005456477869302034, + 0.002799602225422859, + 0.0031720190308988094, + 0.005044741090387106, + 0.002471701707690954, + 0.0029419793281704187, + 0.003453504294157028, + 0.0021779362577944994, + 0.0019374804105609655, + 0.002601602813228965, + 0.00291177025064826, + 0.003746182657778263, + 0.0026450941804796457, + 0.0036158463917672634, + 0.0023390427231788635, + 0.003230500500649214, + 0.0018840960692614317, + 0.046674322336912155, + 0.017646344378590584, + 0.18670368194580078, + 0.02533469907939434, + 0.020714081823825836, + 0.09975839406251907, + 0.10315509885549545, + 0.024733591824769974, + 0.0021418570540845394, + 0.0027156209107488394, + 0.0032244501635432243, + 0.0019699796102941036, + 0.0022835403215140104, + 0.004010475240647793, + 0.0019976815674453974, + 0.0021059352438896894, + 0.0040696351788938046, + 0.0030236910097301006, + 0.004398474935442209, + 0.0018665946554392576, + 0.003997118677943945, + 0.002336698118597269, + 0.0035208689514547586, + 0.0020817196927964687, + 0.018616221845149994, + 0.008875689469277859, + 0.016015931963920593, + 0.006033045705407858, + 0.01344713382422924, + 0.010816924273967743, + 0.013595408760011196, + 0.0099998340010643, + 0.03320583328604698, + 0.020211821421980858, + 0.023552851751446724, + 0.0215388722717762, + 0.028258394449949265, + 0.022667666897177696, + 0.013256941922008991, + 0.016230808570981026, + 0.003980178385972977, + 0.0025629636365920305, + 0.0026898961514234543, + 0.0019462928175926208, + 0.0014710808172821999, + 0.0034343514125794172, + 0.0012752556940540671, + 0.0011364973615854979 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 8, + 7 + ], + "selected_reward": -0.14277471601963043 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level angle begins in medium shot, then slowly pulls back to reveal a wider scene. The frame focuses on a silver-haired elderly man, clad in a loose, white, traditional cotton and linen training suit, nestled in a tranquil bamboo forest. The faint morning sun, the sole source of light, penetrates the towering, verdant bamboo poles and dense leaves, creating soft, voluminous beams that are visible through the misty air and cast shifting, dappled shadows on the damp, leaf-covered ground. The elderly man's hair is meticulously combed, his face deeply wrinkled, weathered. His eyes are slightly closed, his expression focused and peaceful. He is performing Tai Chi with incredible slowness and stability, his arms tracing smooth arcs at a speed barely perceptible to the naked eye, his weight shifting effortlessly between his legs. Each movement is imbued with inner, composed strength. His wide white sleeves fluttered gently as his body turned, and their trajectory seemed to merge with the flowing air and light. There was no pause between his movements, perfectly interpreting the artistic conception of \"flowing like clouds and running water\". The whole scene presented an extremely quiet, solemn and Zen-like meditative atmosphere.", + "seed": 6321, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6321_dp_6_rank_12_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.04746798346750438, + "vlm_logprob": -0.04746798346750438, + "vlm_yes_no_margin": 4.075, + "vlm_token_credit": -0.04746798346750438, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011472680605947971 + ], + "opposite_logprob_tokens": [ + -4.761472702026367 + ], + "logprob_reward": -0.011472680605947971, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.011472680605947971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.44454872608184814, + 0.1089661717414856, + 0.029256895184516907, + 0.03676320239901543, + 0.07846382260322571, + 0.051140375435352325, + 0.09431345760822296, + 0.037392955273389816, + 0.06761160492897034, + 0.05154271796345711 + ], + "credit_map": [ + 0.04904256761074066, + 0.057187359780073166, + 0.06064091622829437, + 0.03991881012916565, + 0.07682163268327713, + 0.08192567527294159, + 0.043703410774469376, + 0.03530842065811157, + 0.01385914720594883, + 0.00935329869389534, + 0.017851263284683228, + 0.01022325363010168, + 0.023132650181651115, + 0.018816404044628143, + 0.0080670565366745, + 0.007663115859031677, + 0.0058027165941894054, + 0.004371936433017254, + 0.004309939220547676, + 0.0040563070215284824, + 0.003656756831333041, + 0.0030632433481514454, + 0.0021053843665868044, + 0.001890616724267602, + 0.005083509720861912, + 0.002173332031816244, + 0.005814432632178068, + 0.003083673771470785, + 0.00767095759510994, + 0.006496082060039043, + 0.0037328109610825777, + 0.0027084071189165115, + 0.009744027629494667, + 0.010203110054135323, + 0.009607258252799511, + 0.007460892200469971, + 0.014403585344552994, + 0.010991816408932209, + 0.00598178431391716, + 0.01007136795669794, + 0.007382554467767477, + 0.009427886456251144, + 0.006342902313917875, + 0.005189551040530205, + 0.0043979003094136715, + 0.007327063474804163, + 0.004492374137043953, + 0.006580148823559284, + 0.015759287402033806, + 0.01751096546649933, + 0.01146770641207695, + 0.007928484119474888, + 0.0050265127792954445, + 0.015610567294061184, + 0.0064061470329761505, + 0.014603803865611553, + 0.008233180269598961, + 0.005130409263074398, + 0.004215293563902378, + 0.0022906791418790817, + 0.003348916070535779, + 0.00642409548163414, + 0.0030299746431410313, + 0.0047204140573740005, + 0.007522457279264927, + 0.006626948248594999, + 0.01171569712460041, + 0.005839185323566198, + 0.010132513009011745, + 0.00827124621719122, + 0.005075326655060053, + 0.012428242713212967, + 0.009468692354857922, + 0.00862487219274044, + 0.006944118533283472, + 0.00907148141413927, + 0.0045771473087370396, + 0.003429301315918565, + 0.004971298389136791, + 0.004455814603716135 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.011472680605947971 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804781526327133 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.009804781526327133, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804781526327133, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.355500191450119, + 0.07931281626224518, + 0.049694158136844635, + 0.05086183920502663, + 0.060811690986156464, + 0.060196422040462494, + 0.05214635282754898, + 0.09626327455043793, + 0.1142207458615303, + 0.08099254220724106 + ], + "credit_map": [ + 0.03944474458694458, + 0.06799764931201935, + 0.047908883541822433, + 0.04782837629318237, + 0.052378300577402115, + 0.047957442700862885, + 0.031577400863170624, + 0.02040736936032772, + 0.007667697966098785, + 0.011660035699605942, + 0.012845831923186779, + 0.007539799436926842, + 0.017132893204689026, + 0.010941481217741966, + 0.00596665171906352, + 0.00555841252207756, + 0.007785778492689133, + 0.008089983835816383, + 0.007219159044325352, + 0.007223098538815975, + 0.005240970756858587, + 0.006619608961045742, + 0.003722232999280095, + 0.003793322714045644, + 0.008702804334461689, + 0.004468042869120836, + 0.008299440145492554, + 0.0029682570602744818, + 0.008758332580327988, + 0.01063539832830429, + 0.003806248540058732, + 0.0032233090605586767, + 0.00805848278105259, + 0.011299227364361286, + 0.007089519873261452, + 0.00433967774733901, + 0.00960632972419262, + 0.008519647642970085, + 0.0044644479639828205, + 0.007434348575770855, + 0.009718098677694798, + 0.014880298636853695, + 0.0055856057442724705, + 0.0035588270984590054, + 0.005163864232599735, + 0.009524527005851269, + 0.004292522557079792, + 0.007472672499716282, + 0.006683968473225832, + 0.008671394549310207, + 0.005968819838017225, + 0.00585015257820487, + 0.0052703325636684895, + 0.007732164114713669, + 0.004417454358190298, + 0.007552062161266804, + 0.01327791903167963, + 0.00949945766478777, + 0.010888654738664627, + 0.014510229229927063, + 0.013116443529725075, + 0.012061569839715958, + 0.012769966386258602, + 0.010139016434550285, + 0.012820775620639324, + 0.010624045506119728, + 0.01610349491238594, + 0.010953882709145546, + 0.021313505247235298, + 0.012255185283720493, + 0.010316439904272556, + 0.019833410158753395, + 0.011381885036826134, + 0.008913715369999409, + 0.014764554798603058, + 0.01613534428179264, + 0.008164361119270325, + 0.005688758101314306, + 0.008290741592645645, + 0.007653175387531519 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.009804781526327133 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.002489208010956645 + ], + "opposite_logprob_tokens": [ + -6.00248908996582 + ], + "logprob_reward": -0.002489208010956645, + "yes_no_margin_reward": 6.0, + "token_credit_reward": -0.002489208010956645, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21193833649158478, + 0.11773667484521866, + 0.1224963590502739, + 0.0761396735906601, + 0.11738470941781998, + 0.08756609261035919, + 0.06007813662290573, + 0.049070850014686584, + 0.09031924605369568, + 0.06726992130279541 + ], + "credit_map": [ + 0.025206468999385834, + 0.03402364253997803, + 0.038301147520542145, + 0.03413783013820648, + 0.029335135594010353, + 0.021275624632835388, + 0.01573914662003517, + 0.013919349759817123, + 0.02077505923807621, + 0.021405842155218124, + 0.01833845116198063, + 0.015511786565184593, + 0.016320722177624702, + 0.009008552879095078, + 0.007162527181208134, + 0.009213740937411785, + 0.030438685789704323, + 0.02124134823679924, + 0.02203715406358242, + 0.014157624915242195, + 0.010302882641553879, + 0.012681807391345501, + 0.00615534046664834, + 0.005481522064656019, + 0.01690581813454628, + 0.012030563317239285, + 0.011693534441292286, + 0.005485403351485729, + 0.006165092810988426, + 0.01341183390468359, + 0.0048818085342645645, + 0.005565618630498648, + 0.01907292753458023, + 0.022060267627239227, + 0.011775022372603416, + 0.00921459961682558, + 0.016840912401676178, + 0.017557157203555107, + 0.008467663079500198, + 0.01239616610109806, + 0.014307999983429909, + 0.023214325308799744, + 0.007279888726770878, + 0.004708264954388142, + 0.007998835295438766, + 0.013788105919957161, + 0.005312368739396334, + 0.01095630880445242, + 0.012473762966692448, + 0.010055821388959885, + 0.008534583263099194, + 0.0036796359345316887, + 0.0032594110816717148, + 0.009646925143897533, + 0.003914792090654373, + 0.008513211272656918, + 0.0060116625390946865, + 0.00593121163547039, + 0.0034183242823928595, + 0.003742985427379608, + 0.0027807217556983232, + 0.015045246109366417, + 0.003491800744086504, + 0.008648899383842945, + 0.019470777362585068, + 0.010269664227962494, + 0.023676523938775063, + 0.005679979920387268, + 0.006369801703840494, + 0.0057906112633645535, + 0.007262771017849445, + 0.011799117550253868, + 0.011667706072330475, + 0.01041843555867672, + 0.010474617592990398, + 0.013391361571848392, + 0.005656317807734013, + 0.0046166712418198586, + 0.006241361144930124, + 0.004803453106433153 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.002489208010956645 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033708587288856506 + ], + "opposite_logprob_tokens": [ + -3.4087085723876953 + ], + "logprob_reward": -0.033708587288856506, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033708587288856506, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.20927861332893372, + 0.14205047488212585, + 0.030417393893003464, + 0.04541286826133728, + 0.1725165694952011, + 0.11626914143562317, + 0.04912445694208145, + 0.03852442651987076, + 0.14957275986671448, + 0.04683329537510872 + ], + "credit_map": [ + 0.049024008214473724, + 0.03401639312505722, + 0.03623735532164574, + 0.024801598861813545, + 0.02397475391626358, + 0.016720501706004143, + 0.012390932999551296, + 0.012113077566027641, + 0.019662611186504364, + 0.03564353659749031, + 0.024040766060352325, + 0.01764208823442459, + 0.018085576593875885, + 0.010117908008396626, + 0.007135189604014158, + 0.009722798131406307, + 0.004043627995997667, + 0.006131681613624096, + 0.004433987196534872, + 0.0044494555331766605, + 0.002860061591491103, + 0.003992740996181965, + 0.0024905717000365257, + 0.0020152658689767122, + 0.008010645397007465, + 0.007865545339882374, + 0.0073296804912388325, + 0.003221338614821434, + 0.003940000664442778, + 0.008428110741078854, + 0.0036049336194992065, + 0.0030126161873340607, + 0.02609560266137123, + 0.037912558764219284, + 0.014250071719288826, + 0.013617862947285175, + 0.02078496664762497, + 0.03115885518491268, + 0.011484025046229362, + 0.01721261255443096, + 0.019092313945293427, + 0.033662814646959305, + 0.007529656868427992, + 0.006963911931961775, + 0.011011479422450066, + 0.018936308100819588, + 0.007141264155507088, + 0.011931388638913631, + 0.012014145962893963, + 0.00763334147632122, + 0.006492251064628363, + 0.0030290225986391306, + 0.002707602921873331, + 0.007981602102518082, + 0.0025827346835285425, + 0.00668375426903367, + 0.005472525488585234, + 0.005104061681777239, + 0.006910506170243025, + 0.0042596664279699326, + 0.002665010979399085, + 0.0055954172275960445, + 0.002713046735152602, + 0.005804192274808884, + 0.030477682128548622, + 0.01621476374566555, + 0.042334720492362976, + 0.008854063227772713, + 0.008246131241321564, + 0.010787450708448887, + 0.009574184194207191, + 0.02308376505970955, + 0.006596943363547325, + 0.00568041019141674, + 0.00687083974480629, + 0.011685637757182121, + 0.003310062689706683, + 0.003500087186694145, + 0.005057984031736851, + 0.004131334368139505 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 8 + ], + "selected_reward": -0.033708587288856506 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.17986465990543365 + ], + "opposite_logprob_tokens": [ + -1.8048646450042725 + ], + "logprob_reward": -0.17986465990543365, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.17986465990543365, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1905827671289444, + 0.20045053958892822, + 0.04798058047890663, + 0.09669002145528793, + 0.10217751562595367, + 0.1686650514602661, + 0.0451454296708107, + 0.04184143617749214, + 0.0662805587053299, + 0.040186088532209396 + ], + "credit_map": [ + 0.02612837217748165, + 0.03424646705389023, + 0.036201171576976776, + 0.02943972684442997, + 0.022662049159407616, + 0.015361831523478031, + 0.014608575962483883, + 0.011934571899473667, + 0.0390603132545948, + 0.04508721083402634, + 0.03422585874795914, + 0.025731010362505913, + 0.020195428282022476, + 0.01306435838341713, + 0.009797058999538422, + 0.013289286755025387, + 0.010092376731336117, + 0.00788007490336895, + 0.007892818190157413, + 0.0045184544287621975, + 0.006836751475930214, + 0.0048730941489338875, + 0.0032061110250651836, + 0.002680903533473611, + 0.012729003094136715, + 0.014327510260045528, + 0.016109494492411613, + 0.008691716939210892, + 0.007882772944867611, + 0.021973907947540283, + 0.008157985284924507, + 0.006817633286118507, + 0.013624287210404873, + 0.01297845784574747, + 0.012544486671686172, + 0.012401174753904343, + 0.015685152262449265, + 0.01245120633393526, + 0.01249641366302967, + 0.009996339678764343, + 0.020709829404950142, + 0.02112571895122528, + 0.01707119308412075, + 0.01737228035926819, + 0.026054630056023598, + 0.02328699827194214, + 0.017494359984993935, + 0.025550037622451782, + 0.00972942914813757, + 0.00544913811609149, + 0.006259260233491659, + 0.003029228188097477, + 0.005896164570003748, + 0.0061995177529752254, + 0.003447841154411435, + 0.005134849343448877, + 0.007392808794975281, + 0.005980644840747118, + 0.005671380553394556, + 0.003938098903745413, + 0.003676152788102627, + 0.007265632972121239, + 0.003213253803551197, + 0.004703464452177286, + 0.010069243609905243, + 0.005064683500677347, + 0.016390547156333923, + 0.0040542553178966045, + 0.0034215757623314857, + 0.005964125040918589, + 0.005513722077012062, + 0.015802402049303055, + 0.006475062575191259, + 0.005568215157836676, + 0.005372494924813509, + 0.009183757938444614, + 0.0031423650216311216, + 0.002860761247575283, + 0.0037751614581793547, + 0.0038082723040133715 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.17986465990543365 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level angle begins in medium shot, then slowly pulls back to reveal a wider scene. The frame focuses on a silver-haired elderly man, clad in a loose, white, traditional cotton and linen training suit, nestled in a tranquil bamboo forest. The faint morning sun, the sole source of light, penetrates the towering, verdant bamboo poles and dense leaves, creating soft, voluminous beams that are visible through the misty air and cast shifting, dappled shadows on the damp, leaf-covered ground. The elderly man's hair is meticulously combed, his face deeply wrinkled, weathered. His eyes are slightly closed, his expression focused and peaceful. He is performing Tai Chi with incredible slowness and stability, his arms tracing smooth arcs at a speed barely perceptible to the naked eye, his weight shifting effortlessly between his legs. Each movement is imbued with inner, composed strength. His wide white sleeves fluttered gently as his body turned, and their trajectory seemed to merge with the flowing air and light. There was no pause between his movements, perfectly interpreting the artistic conception of \"flowing like clouds and running water\". The whole scene presented an extremely quiet, solemn and Zen-like meditative atmosphere.", + "seed": 6322, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6322_dp_7_rank_14_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.05963534601032734, + "vlm_logprob": -0.05963534601032734, + "vlm_yes_no_margin": 3.9, + "vlm_token_credit": -0.05963534601032734, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00889434851706028 + ], + "opposite_logprob_tokens": [ + -5.258894443511963 + ], + "logprob_reward": -0.00889434851706028, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.00889434851706028, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.5213862657546997, + 0.051791202276945114, + 0.050343114882707596, + 0.0483647845685482, + 0.04186403006315231, + 0.02944389544427395, + 0.14173966646194458, + 0.04230188578367233, + 0.03752411901950836, + 0.03524107486009598 + ], + "credit_map": [ + 0.08104931563138962, + 0.0493396557867527, + 0.027213245630264282, + 0.04541270062327385, + 0.03588765859603882, + 0.12270621210336685, + 0.016584906727075577, + 0.1431925743818283, + 0.005464108195155859, + 0.008640903979539871, + 0.004659916739910841, + 0.007812167052179575, + 0.006710244342684746, + 0.006462601944804192, + 0.0046829720959067345, + 0.007358286064118147, + 0.00395910395309329, + 0.01000542938709259, + 0.004615494981408119, + 0.009013889357447624, + 0.006175458896905184, + 0.006073832046240568, + 0.0065833949483931065, + 0.003916509915143251, + 0.010940620675683022, + 0.004285803996026516, + 0.007626351900398731, + 0.005785146728157997, + 0.004553128965198994, + 0.007247754372656345, + 0.003235763870179653, + 0.004690213594585657, + 0.006309723015874624, + 0.007614905945956707, + 0.006402766332030296, + 0.004660395905375481, + 0.0037722880952060223, + 0.00745623791590333, + 0.0030209918040782213, + 0.002626719418913126, + 0.005203330423682928, + 0.004009278025478125, + 0.003946582321077585, + 0.0038107563741505146, + 0.00427884329110384, + 0.003393697552382946, + 0.0023722578771412373, + 0.002429148880764842, + 0.018256468698382378, + 0.01675344817340374, + 0.018577996641397476, + 0.016387220472097397, + 0.011687266640365124, + 0.01975220814347267, + 0.02349160425364971, + 0.016833443194627762, + 0.005764858331531286, + 0.004553318489342928, + 0.00522012310102582, + 0.005579608026891947, + 0.005290145054459572, + 0.007575749885290861, + 0.003385738003998995, + 0.0049323453567922115, + 0.002789938123896718, + 0.004291159100830555, + 0.006767018232494593, + 0.008304345421493053, + 0.004119522869586945, + 0.004238041117787361, + 0.0025353101082146168, + 0.004478782881051302, + 0.0038611479103565216, + 0.005844607949256897, + 0.0028564073145389557, + 0.004383348394185305, + 0.005687393248081207, + 0.0034389779902994633, + 0.0021707091946154833, + 0.006998484488576651 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.00889434851706028 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009804898872971535 + ], + "opposite_logprob_tokens": [ + -4.634804725646973 + ], + "logprob_reward": -0.009804898872971535, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009804898872971535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.285094290971756, + 0.09232265502214432, + 0.1014835312962532, + 0.04759315028786659, + 0.06660719960927963, + 0.03192032501101494, + 0.17830556631088257, + 0.05640092119574547, + 0.0766223669052124, + 0.06364991515874863 + ], + "credit_map": [ + 0.032336946576833725, + 0.028020784258842468, + 0.012785372324287891, + 0.01783280074596405, + 0.021587098017334938, + 0.0753231942653656, + 0.010394811630249023, + 0.08681333065032959, + 0.008532348088920116, + 0.017543256282806396, + 0.012162134982645512, + 0.011625823564827442, + 0.009423971176147461, + 0.011873476207256317, + 0.008830910548567772, + 0.012330744415521622, + 0.011215437203645706, + 0.016745101660490036, + 0.009174935519695282, + 0.013879709877073765, + 0.015994055196642876, + 0.013715147040784359, + 0.011936023831367493, + 0.00882312934845686, + 0.01158047653734684, + 0.005718700122088194, + 0.007429051212966442, + 0.004042725544422865, + 0.006043720990419388, + 0.005636621732264757, + 0.002801261143758893, + 0.004340597428381443, + 0.009944973513484001, + 0.005851346999406815, + 0.010639221407473087, + 0.009678792208433151, + 0.0068605812266469, + 0.014401586726307869, + 0.0038250924553722143, + 0.005405612289905548, + 0.005311429034918547, + 0.0038397989701479673, + 0.005144116468727589, + 0.002971053821966052, + 0.004285010974854231, + 0.004353619180619717, + 0.0030111016239970922, + 0.0030041951686143875, + 0.01996457949280739, + 0.018568944185972214, + 0.024675676599144936, + 0.024139225482940674, + 0.013024763204157352, + 0.030464379116892815, + 0.028406154364347458, + 0.019061850383877754, + 0.005671415477991104, + 0.00841983687132597, + 0.008675492368638515, + 0.006282074376940727, + 0.005600833799690008, + 0.010870121419429779, + 0.0038317558355629444, + 0.00704939616844058, + 0.007952610962092876, + 0.00825242605060339, + 0.017382916063070297, + 0.011119227856397629, + 0.010222637094557285, + 0.007636267226189375, + 0.006253567058593035, + 0.007802717853337526, + 0.007372637279331684, + 0.012563045136630535, + 0.00854315236210823, + 0.008482271805405617, + 0.008049906231462955, + 0.004500240553170443, + 0.005044063087552786, + 0.00909460335969925 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 2 + ], + "selected_reward": -0.009804898872971535 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005253320559859276 + ], + "opposite_logprob_tokens": [ + -5.255253314971924 + ], + "logprob_reward": -0.005253320559859276, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005253320559859276, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33439967036247253, + 0.09991209954023361, + 0.06233338639140129, + 0.027599362656474113, + 0.04078160598874092, + 0.03594384714961052, + 0.28787893056869507, + 0.04460315778851509, + 0.03925304859876633, + 0.027294952422380447 + ], + "credit_map": [ + 0.031299784779548645, + 0.021666329354047775, + 0.011021682992577553, + 0.0167622622102499, + 0.022959653288125992, + 0.10754546523094177, + 0.00776330754160881, + 0.11538115888834, + 0.012847472913563251, + 0.011108915321528912, + 0.008471209555864334, + 0.012098891660571098, + 0.015635821968317032, + 0.015758207067847252, + 0.006536638364195824, + 0.017454946413636208, + 0.005319306161254644, + 0.005496102850884199, + 0.005944015923887491, + 0.014744296669960022, + 0.0050376104190945625, + 0.013058682903647423, + 0.00711471401154995, + 0.005618650931864977, + 0.005990031640976667, + 0.00343048176728189, + 0.003846216481178999, + 0.0034266712609678507, + 0.0029843582306057215, + 0.0036173986736685038, + 0.00189712003339082, + 0.002407082822173834, + 0.005612394772469997, + 0.005551021080464125, + 0.006819671951234341, + 0.004890813957899809, + 0.003907411824911833, + 0.008473731577396393, + 0.0024675794411450624, + 0.0030589804518967867, + 0.007437896449118853, + 0.0063953218050301075, + 0.004875322803854942, + 0.0034526223316788673, + 0.0033486546017229557, + 0.005890222266316414, + 0.0018480331636965275, + 0.002695772796869278, + 0.030961966142058372, + 0.02287033200263977, + 0.042637381702661514, + 0.0401763990521431, + 0.017214056104421616, + 0.057304490357637405, + 0.04249672591686249, + 0.03421758860349655, + 0.007260976359248161, + 0.003132224315777421, + 0.0059500643983483315, + 0.002910128328949213, + 0.006248395424336195, + 0.011132881045341492, + 0.0024451299104839563, + 0.005523354280740023, + 0.0043699368834495544, + 0.0029886667616665363, + 0.007222886197268963, + 0.0030662359204143286, + 0.0064209140837192535, + 0.0073592024855315685, + 0.0034827578347176313, + 0.0043424516916275024, + 0.0024045000318437815, + 0.002976826624944806, + 0.0014912447659298778, + 0.0020022825337946415, + 0.004908757284283638, + 0.006668925751000643, + 0.002370404312387109, + 0.004472009837627411 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.005253320559859276 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048678621649742126 + ], + "opposite_logprob_tokens": [ + -3.0486786365509033 + ], + "logprob_reward": -0.048678621649742126, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048678621649742126, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.25910910964012146, + 0.10830318182706833, + 0.026832340285182, + 0.05312497541308403, + 0.03584430739283562, + 0.04724741354584694, + 0.3669934868812561, + 0.05951691418886185, + 0.02525199204683304, + 0.017776308581233025 + ], + "credit_map": [ + 0.024698063731193542, + 0.02154599502682686, + 0.010034034959971905, + 0.014139124192297459, + 0.021710222586989403, + 0.07713573426008224, + 0.007233778014779091, + 0.08261216431856155, + 0.013853983022272587, + 0.01466086320579052, + 0.009305995889008045, + 0.013267965987324715, + 0.01685398258268833, + 0.01655229926109314, + 0.006325330585241318, + 0.017482757568359375, + 0.003707269672304392, + 0.0033645357470959425, + 0.0038245536852627993, + 0.0037064829375594854, + 0.0027076625265181065, + 0.0044203284196555614, + 0.0026338291354477406, + 0.0024676774628460407, + 0.008954343385994434, + 0.005324691068381071, + 0.010535289533436298, + 0.00417563458904624, + 0.005231526680290699, + 0.010899808257818222, + 0.0032478959765285254, + 0.004755785688757896, + 0.003665381809696555, + 0.004068837966769934, + 0.006310644559562206, + 0.005023157689720392, + 0.0036153008695691824, + 0.008204196579754353, + 0.0024514743126928806, + 0.002505312440916896, + 0.010150108486413956, + 0.0077563985250890255, + 0.006937981583178043, + 0.004319190047681332, + 0.004135967697948217, + 0.0084806177765131, + 0.0025279303081333637, + 0.0029392160940915346, + 0.04267537221312523, + 0.03536754101514816, + 0.052222613245248795, + 0.04948065057396889, + 0.025568248704075813, + 0.06417392939329147, + 0.05808057263493538, + 0.03942454606294632, + 0.009362130425870419, + 0.003956587053835392, + 0.0070158797316253185, + 0.004419635981321335, + 0.007863430306315422, + 0.014208323322236538, + 0.005114282015711069, + 0.007576642092317343, + 0.002535552019253373, + 0.0026694771368056536, + 0.006111506372690201, + 0.0025511602871119976, + 0.003969191107898951, + 0.0032359943725168705, + 0.0016532321460545063, + 0.002525879768654704, + 0.002255517290905118, + 0.0020010899752378464, + 0.0015351600013673306, + 0.0016941482899710536, + 0.002796603599563241, + 0.0036435609217733145, + 0.0014631933299824595, + 0.0023870361037552357 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 1 + ], + "selected_reward": -0.048678621649742126 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.22554554045200348 + ], + "opposite_logprob_tokens": [ + -1.6005455255508423 + ], + "logprob_reward": -0.22554554045200348, + "yes_no_margin_reward": 1.375, + "token_credit_reward": -0.22554554045200348, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10431213676929474, + 0.07790611684322357, + 0.06269988417625427, + 0.09810099005699158, + 0.07258893549442291, + 0.02800094150006771, + 0.4751398265361786, + 0.03659398853778839, + 0.02360454946756363, + 0.021052682772278786 + ], + "credit_map": [ + 0.021787624806165695, + 0.016887491568922997, + 0.008865044452250004, + 0.016684383153915405, + 0.011097058653831482, + 0.012268121354281902, + 0.006101369857788086, + 0.010621036402881145, + 0.009948878549039364, + 0.01147967018187046, + 0.006057475693523884, + 0.0086790407076478, + 0.014093161560595036, + 0.011173917911946774, + 0.004872583784162998, + 0.011601381935179234, + 0.006416722666472197, + 0.011266413144767284, + 0.006735665258020163, + 0.008405501022934914, + 0.009294711984694004, + 0.006223035044968128, + 0.007956093177199364, + 0.0064017376862466335, + 0.015737151727080345, + 0.010205971077084541, + 0.019708100706338882, + 0.00786901917308569, + 0.00905653927475214, + 0.01946522481739521, + 0.006200178526341915, + 0.009858808480203152, + 0.010282941162586212, + 0.009132757782936096, + 0.011759222485125065, + 0.00755040580406785, + 0.006342476233839989, + 0.01493862085044384, + 0.004993359558284283, + 0.007589150685817003, + 0.0030842453707009554, + 0.0032515351194888353, + 0.005225092172622681, + 0.003009482752531767, + 0.0034848684445023537, + 0.0051895854994654655, + 0.0021464333403855562, + 0.002609696937724948, + 0.05720493569970131, + 0.04393745958805084, + 0.06645282357931137, + 0.06682093441486359, + 0.029720911756157875, + 0.0845620334148407, + 0.07196029275655746, + 0.0544804148375988, + 0.002943394472822547, + 0.004906358662992716, + 0.005719677545130253, + 0.004990129731595516, + 0.0027816439978778362, + 0.0059846132062375546, + 0.004852773156017065, + 0.004415397997945547, + 0.002318106358870864, + 0.003074726788327098, + 0.0036784999538213015, + 0.003826287342235446, + 0.002792512299492955, + 0.003146883100271225, + 0.0016596924979239702, + 0.003107842057943344, + 0.0016831440152600408, + 0.0027754735201597214, + 0.0030200304463505745, + 0.0036722435615956783, + 0.002336174016818404, + 0.0021441320423036814, + 0.0019284626469016075, + 0.003493022872135043 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 0, + 3 + ], + "selected_reward": -0.22554554045200348 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "shot shot from a level angle begins in medium shot, then slowly pulls back to reveal a wider scene. The frame focuses on a silver-haired elderly man, clad in a loose, white, traditional cotton and linen training suit, nestled in a tranquil bamboo forest. The faint morning sun, the sole source of light, penetrates the towering, verdant bamboo poles and dense leaves, creating soft, voluminous beams that are visible through the misty air and cast shifting, dappled shadows on the damp, leaf-covered ground. The elderly man's hair is meticulously combed, his face deeply wrinkled, weathered. His eyes are slightly closed, his expression focused and peaceful. He is performing Tai Chi with incredible slowness and stability, his arms tracing smooth arcs at a speed barely perceptible to the naked eye, his weight shifting effortlessly between his legs. Each movement is imbued with inner, composed strength. His wide white sleeves fluttered gently as his body turned, and their trajectory seemed to merge with the flowing air and light. There was no pause between his movements, perfectly interpreting the artistic conception of \"flowing like clouds and running water\". The whole scene presented an extremely quiet, solemn and Zen-like meditative atmosphere.", + "seed": 6323, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_3/index_2934_seed_6323_dp_7_rank_14_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 3, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.06487950952723623, + "vlm_logprob": -0.06487950952723623, + "vlm_yes_no_margin": 4.225, + "vlm_token_credit": -0.06487950952723623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006820137612521648 + ], + "opposite_logprob_tokens": [ + -5.631820201873779 + ], + "logprob_reward": -0.006820137612521648, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.006820137612521648, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3264426589012146, + 0.06530611962080002, + 0.05221449211239815, + 0.16841691732406616, + 0.07518699020147324, + 0.08703847974538803, + 0.0384080670773983, + 0.04920642077922821, + 0.059340950101614, + 0.07843893766403198 + ], + "credit_map": [ + 0.030871711671352386, + 0.043629977852106094, + 0.02983560599386692, + 0.02898288518190384, + 0.07464123517274857, + 0.05262766405940056, + 0.03219061344861984, + 0.033662959933280945, + 0.00843766238540411, + 0.009722943417727947, + 0.007824870757758617, + 0.008693370036780834, + 0.006986138876527548, + 0.007622214965522289, + 0.009685064665973186, + 0.006333854515105486, + 0.005020370706915855, + 0.006149898748844862, + 0.007556175347417593, + 0.004286251496523619, + 0.00505274161696434, + 0.008639698848128319, + 0.005809352267533541, + 0.009700005874037743, + 0.022994214668869972, + 0.011997655965387821, + 0.029655195772647858, + 0.009626633487641811, + 0.012631298042833805, + 0.03300148993730545, + 0.009479394182562828, + 0.03903103992342949, + 0.00877644307911396, + 0.005331780295819044, + 0.019594749435782433, + 0.004401859361678362, + 0.006570339668542147, + 0.012363968417048454, + 0.004976923577487469, + 0.013170926831662655, + 0.014906315132975578, + 0.016558894887566566, + 0.012560881674289703, + 0.006484217941761017, + 0.00957318115979433, + 0.01140685100108385, + 0.00608953507617116, + 0.009458609856665134, + 0.005873008631169796, + 0.005673285108059645, + 0.005776194855570793, + 0.00381092494353652, + 0.003879892872646451, + 0.005199680104851723, + 0.004597960971295834, + 0.0035971198230981827, + 0.007547841407358646, + 0.005001815035939217, + 0.008140201680362225, + 0.004407504107803106, + 0.006137820426374674, + 0.006349723786115646, + 0.005662901327013969, + 0.005958614405244589, + 0.007256091572344303, + 0.005359265021979809, + 0.013024857267737389, + 0.004154680296778679, + 0.005728530697524548, + 0.008476059883832932, + 0.006645629182457924, + 0.008695835247635841, + 0.00977272354066372, + 0.007608492858707905, + 0.007729466538876295, + 0.008201350457966328, + 0.010430552996695042, + 0.00909158680588007, + 0.013470510020852089, + 0.012134245596826077 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 5 + ], + "selected_reward": -0.006820137612521648 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006752174347639084 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006752174347639084, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006752174347639084, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17289915680885315, + 0.07986778765916824, + 0.20529086887836456, + 0.12154710292816162, + 0.05356401950120926, + 0.04675612226128578, + 0.05518766865134239, + 0.07485491037368774, + 0.04701574891805649, + 0.14301660656929016 + ], + "credit_map": [ + 0.025436466559767723, + 0.021571870893239975, + 0.01741715520620346, + 0.015699883922934532, + 0.027791395783424377, + 0.027485959231853485, + 0.021684180945158005, + 0.015812266618013382, + 0.009240355342626572, + 0.009112393483519554, + 0.011676201596856117, + 0.012305894866585732, + 0.009563351050019264, + 0.009643311612308025, + 0.010504035279154778, + 0.007822244428098202, + 0.017594140022993088, + 0.023163745179772377, + 0.030684277415275574, + 0.021601565182209015, + 0.014115390367805958, + 0.03710135817527771, + 0.023879768326878548, + 0.03715064376592636, + 0.021116062998771667, + 0.010122956708073616, + 0.02588794194161892, + 0.00770513666793704, + 0.008063883520662785, + 0.01895195059478283, + 0.007356526330113411, + 0.022342665120959282, + 0.006214895751327276, + 0.004157254938036203, + 0.012677239254117012, + 0.0030839720275253057, + 0.006136086769402027, + 0.008312917314469814, + 0.003912035375833511, + 0.009069626219570637, + 0.0036636872682720423, + 0.0045836325734853745, + 0.00747613376006484, + 0.005900491029024124, + 0.008303262293338776, + 0.006928637158125639, + 0.004932830110192299, + 0.004967452958226204, + 0.008418895304203033, + 0.005849113222211599, + 0.01171767059713602, + 0.004805008880794048, + 0.005761719308793545, + 0.008413015864789486, + 0.005916086956858635, + 0.004306164104491472, + 0.008440746925771236, + 0.0064685712568461895, + 0.010065254755318165, + 0.008807082660496235, + 0.006467175669968128, + 0.009708077646791935, + 0.014743556268513203, + 0.010154453106224537, + 0.005822336301207542, + 0.0046041361056268215, + 0.011168860830366611, + 0.0035986355505883694, + 0.0036633575800806284, + 0.006567453965544701, + 0.0040465835481882095, + 0.007544388994574547, + 0.0169419813901186, + 0.013676649890840054, + 0.019739437848329544, + 0.015390624292194843, + 0.020224064588546753, + 0.017448609694838524, + 0.023347051814198494, + 0.016248192638158798 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 9 + ], + "selected_reward": -0.006752174347639084 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.003190429415553808 + ], + "opposite_logprob_tokens": [ + -5.753190517425537 + ], + "logprob_reward": -0.003190429415553808, + "yes_no_margin_reward": 5.75, + "token_credit_reward": -0.003190429415553808, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11897522211074829, + 0.09097147732973099, + 0.16322247684001923, + 0.15805694460868835, + 0.08599144220352173, + 0.09889864176511765, + 0.04174257069826126, + 0.029805893078446388, + 0.03411593288183212, + 0.17821943759918213 + ], + "credit_map": [ + 0.0248970128595829, + 0.01505418960005045, + 0.01467402558773756, + 0.009940382093191147, + 0.0137168625369668, + 0.016473393887281418, + 0.011765378527343273, + 0.01245398074388504, + 0.018520038574934006, + 0.0085102254524827, + 0.013241979293525219, + 0.010341005399823189, + 0.009466620162129402, + 0.01346904132515192, + 0.008168819360435009, + 0.009253747761249542, + 0.017631836235523224, + 0.011412731371819973, + 0.014660927467048168, + 0.01139108743518591, + 0.012733153067529202, + 0.04239645600318909, + 0.009251383133232594, + 0.04374488815665245, + 0.0275631882250309, + 0.011865600012242794, + 0.03323695808649063, + 0.009495160542428493, + 0.01019134558737278, + 0.025919681414961815, + 0.00914648175239563, + 0.030638523399829865, + 0.015153895132243633, + 0.005987835582345724, + 0.020644376054406166, + 0.00607419665902853, + 0.0061680325306952, + 0.01259132381528616, + 0.005191613920032978, + 0.014180174097418785, + 0.01664673537015915, + 0.017836207523941994, + 0.013887583278119564, + 0.00802786834537983, + 0.011705602519214153, + 0.012783865444362164, + 0.007533145137131214, + 0.010477634146809578, + 0.006726366933435202, + 0.005308723077178001, + 0.008815725333988667, + 0.0029224655590951443, + 0.0042554051615297794, + 0.006626081187278032, + 0.0036463148426264524, + 0.003441489301621914, + 0.004120963625609875, + 0.0032992358319461346, + 0.005804372951388359, + 0.002392665483057499, + 0.0031038853339850903, + 0.003774776589125395, + 0.003134352620691061, + 0.0041756401769816875, + 0.0037015890702605247, + 0.002214132808148861, + 0.009020608849823475, + 0.0026364580262452364, + 0.0030612542759627104, + 0.005153208039700985, + 0.0036573223769664764, + 0.004671359434723854, + 0.0354401059448719, + 0.023843223229050636, + 0.04291439801454544, + 0.010843590833246708, + 0.011501477099955082, + 0.024644019082188606, + 0.012877939268946648, + 0.016154689714312553 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 2, + 3 + ], + "selected_reward": -0.003190429415553808 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026353394612669945 + ], + "opposite_logprob_tokens": [ + -3.651353359222412 + ], + "logprob_reward": -0.026353394612669945, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026353394612669945, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1723639816045761, + 0.055572282522916794, + 0.10882720351219177, + 0.33020535111427307, + 0.11990994215011597, + 0.05128444358706474, + 0.03938741236925125, + 0.024960830807685852, + 0.05607437342405319, + 0.041414178907871246 + ], + "credit_map": [ + 0.03140212222933769, + 0.024788130074739456, + 0.01457980927079916, + 0.014293375425040722, + 0.03152379021048546, + 0.023952672258019447, + 0.016285832971334457, + 0.015538248233497143, + 0.008830804377794266, + 0.007337363436818123, + 0.006685273256152868, + 0.007151911500841379, + 0.004645247478038073, + 0.006821489427238703, + 0.007374799810349941, + 0.006725396495312452, + 0.013867072761058807, + 0.008489113301038742, + 0.009681057184934616, + 0.006788074504584074, + 0.008954480290412903, + 0.027959296479821205, + 0.005795920267701149, + 0.027292191982269287, + 0.05153587833046913, + 0.030416831374168396, + 0.049089834094047546, + 0.015247557312250137, + 0.03145968168973923, + 0.059610508382320404, + 0.012738906778395176, + 0.08010617643594742, + 0.023504212498664856, + 0.010010954923927784, + 0.02678435854613781, + 0.006013466045260429, + 0.010534310713410378, + 0.0173125471919775, + 0.006205452606081963, + 0.019544653594493866, + 0.006419827230274677, + 0.006816016975790262, + 0.00901766587048769, + 0.004699250217527151, + 0.006453198380768299, + 0.006767468061298132, + 0.0054840161465108395, + 0.005627007223665714, + 0.007688959129154682, + 0.004253209102898836, + 0.006780092138797045, + 0.0032570853363722563, + 0.0038964124396443367, + 0.005690857768058777, + 0.005043675657361746, + 0.0027771219611167908, + 0.00322408857755363, + 0.0021277409978210926, + 0.00361104984767735, + 0.0024214484728872776, + 0.0027120437007397413, + 0.0034015115816146135, + 0.0031817678827792406, + 0.004281180445104837, + 0.0074426885694265366, + 0.004908240865916014, + 0.01524383295327425, + 0.004027045797556639, + 0.004351483657956123, + 0.008353021927177906, + 0.006286256015300751, + 0.005461809225380421, + 0.0037834870163351297, + 0.003971404861658812, + 0.005655912682414055, + 0.004756774753332138, + 0.004618355073034763, + 0.006015575025230646, + 0.0069367061369121075, + 0.005675963591784239 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.026353394612669945 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject an elderly man with silver hair and deeply wrinkled, weathered skin?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a loose, white traditional cotton and linen training suit with wide sleeves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the setting a tranquil bamboo forest with towering poles and dense leaves?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the scene feature soft, voluminous beams of light penetrating the mist and casting dappled shadows on the ground?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera start in a medium shot and slowly pull back to reveal a wider scene?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.28128141164779663 + ], + "opposite_logprob_tokens": [ + -1.4062814712524414 + ], + "logprob_reward": -0.28128141164779663, + "yes_no_margin_reward": 1.125, + "token_credit_reward": -0.28128141164779663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14523810148239136, + 0.08738411962985992, + 0.08770797401666641, + 0.18606902658939362, + 0.04985986649990082, + 0.06449079513549805, + 0.11273594945669174, + 0.05376609042286873, + 0.10085860639810562, + 0.11188945174217224 + ], + "credit_map": [ + 0.02365499921143055, + 0.02123376727104187, + 0.017226722091436386, + 0.011097575537860394, + 0.0244422759860754, + 0.02098434790968895, + 0.015228901989758015, + 0.011369523592293262, + 0.006828548852354288, + 0.01133472379297018, + 0.012720925733447075, + 0.012783234938979149, + 0.006059941370040178, + 0.009548299014568329, + 0.018207410350441933, + 0.009901034645736217, + 0.010233588516712189, + 0.007468161638826132, + 0.008607462048530579, + 0.0060646566562354565, + 0.008592560887336731, + 0.018370812758803368, + 0.00671809259802103, + 0.02165263332426548, + 0.023311596363782883, + 0.012360893189907074, + 0.0266389399766922, + 0.01037738099694252, + 0.023033421486616135, + 0.032663676887750626, + 0.01023129466921091, + 0.04745182767510414, + 0.005200735293328762, + 0.005786015186458826, + 0.00781822856515646, + 0.004098741803318262, + 0.007674147840589285, + 0.006233753636479378, + 0.005531735252588987, + 0.00751651031896472, + 0.00482999486848712, + 0.006312606390565634, + 0.008813057094812393, + 0.009048919193446636, + 0.008984401822090149, + 0.008368071168661118, + 0.010676104575395584, + 0.007457636762410402, + 0.019177736714482307, + 0.011148365214467049, + 0.020844953134655952, + 0.010858946479856968, + 0.012237579561769962, + 0.01602533832192421, + 0.013364949263632298, + 0.009078086353838444, + 0.005967914126813412, + 0.0044301520101726055, + 0.006399831734597683, + 0.004816447850316763, + 0.006595965474843979, + 0.00634430069476366, + 0.009630708023905754, + 0.009580769576132298, + 0.007580544333904982, + 0.0069679152220487595, + 0.01696009188890457, + 0.00983168464154005, + 0.013903706334531307, + 0.01303590927273035, + 0.01606900990009308, + 0.01650974154472351, + 0.01541938353329897, + 0.01319647952914238, + 0.018242686986923218, + 0.010623067617416382, + 0.011751393787562847, + 0.01430091168731451, + 0.01593519188463688, + 0.012420336715877056 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 6 + ], + "selected_reward": -0.28128141164779663 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_tensor([9150, 9150])_group_4_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_tensor([9150, 9150])_group_4_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..25024e5121c6569916e7000a997da9d3677c7fb9 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_tensor([9150, 9150])_group_4_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "low-angle, static shot depicts a massive red king crab, its form defined by bold black line art, emphasizing its immense size and power. The creature strikes a dramatic, energetic pose atop a massive, jagged slab of volcanic rock, its fractured, obsidian-like surface gleaming with light. Its fiery red shell, painted in high-contrast shading, carved into deep, stark shadows, is covered in sharp, dagger-like spikes. The crab's long, spiny legs are spread wide, its pointed tips gripping the rock for stability. Its massive claws—one a powerful crusher, the other a razor-sharp pincer—are raised high and extended in defensive formation, poised to strike. The entire scene is imbued with a tense, confrontational atmosphere, capturing a frozen moment of heightened alertness, as the crab's tiny, stalk-like eyes scan its surroundings for unseen enemies.", + "seed": 15270, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_9150_seed_15270_dp_8_rank_16_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.021013447525911035, + "vlm_logprob": -0.021013447525911035, + "vlm_yes_no_margin": 4.474999952316284, + "vlm_token_credit": -0.021013447525911035, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00893700122833252 + ], + "opposite_logprob_tokens": [ + -5.008936882019043 + ], + "logprob_reward": -0.00893700122833252, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00893700122833252, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26748451590538025, + 0.12022145837545395, + 0.04276605695486069, + 0.05630134046077728, + 0.04748708754777908, + 0.18171082437038422, + 0.10086775571107864, + 0.06383412331342697, + 0.0667281225323677, + 0.05259872227907181 + ], + "credit_map": [ + 0.02843727543950081, + 0.05393990874290466, + 0.028587833046913147, + 0.022287512198090553, + 0.03981860354542732, + 0.04595041647553444, + 0.02510189078748226, + 0.023361077532172203, + 0.013153553009033203, + 0.020654169842600822, + 0.01814919151365757, + 0.01082887314260006, + 0.02097574807703495, + 0.013485660776495934, + 0.012793843634426594, + 0.01018042303621769, + 0.0043094223365187645, + 0.007424267008900642, + 0.005299652460962534, + 0.0037394177634269, + 0.007765501271933317, + 0.006138534750789404, + 0.004111896734684706, + 0.0039773606695234776, + 0.00543049443513155, + 0.010260708630084991, + 0.005781815852969885, + 0.006232151295989752, + 0.00755377858877182, + 0.008734378963708878, + 0.005928653758019209, + 0.006379358470439911, + 0.004342994652688503, + 0.007072328124195337, + 0.006395427044481039, + 0.0029059776570647955, + 0.010179287753999233, + 0.008421247825026512, + 0.0035629114136099815, + 0.004606910049915314, + 0.014351795427501202, + 0.0338808074593544, + 0.014794551767408848, + 0.010976208373904228, + 0.048232223838567734, + 0.02878091111779213, + 0.018130652606487274, + 0.012563672848045826, + 0.007982242852449417, + 0.016637414693832397, + 0.009646655060350895, + 0.00674164155498147, + 0.0214036013931036, + 0.01664276048541069, + 0.010643018409609795, + 0.011170418001711369, + 0.005063229240477085, + 0.010210598818957806, + 0.007486152928322554, + 0.005097974091768265, + 0.013090246357023716, + 0.010111610405147076, + 0.006001575384289026, + 0.006772737950086594, + 0.006712320260703564, + 0.012051616795361042, + 0.006042649503797293, + 0.006230519153177738, + 0.008644754067063332, + 0.007816548459231853, + 0.008953648619353771, + 0.010276068933308125, + 0.0038051074370741844, + 0.010413864627480507, + 0.003149457508698106, + 0.0036226236261427402, + 0.014982466585934162, + 0.008456360548734665, + 0.005188890732824802, + 0.0029799523763358593 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 1 + ], + "selected_reward": -0.00893700122833252 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0017315646400675178 + ], + "opposite_logprob_tokens": [ + -6.3767313957214355 + ], + "logprob_reward": -0.0017315646400675178, + "yes_no_margin_reward": 6.375, + "token_credit_reward": -0.0017315646400675178, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31215178966522217, + 0.1960178166627884, + 0.05922091007232666, + 0.05547729879617691, + 0.051433954387903214, + 0.09698686003684998, + 0.04766319692134857, + 0.04104813188314438, + 0.051960572600364685, + 0.08803939074277878 + ], + "credit_map": [ + 0.029097653925418854, + 0.05883284658193588, + 0.02061922289431095, + 0.03969058394432068, + 0.045617230236530304, + 0.05646421015262604, + 0.029771365225315094, + 0.03205869719386101, + 0.016933390870690346, + 0.03257443755865097, + 0.020863046869635582, + 0.017880136147141457, + 0.043401144444942474, + 0.024266842752695084, + 0.020319726318120956, + 0.019779091700911522, + 0.005243388004601002, + 0.010257994756102562, + 0.007179074455052614, + 0.006525995675474405, + 0.008556538261473179, + 0.007031889632344246, + 0.006557951681315899, + 0.007868077605962753, + 0.005013547837734222, + 0.009861255064606667, + 0.00822228193283081, + 0.0039480458945035934, + 0.010738998651504517, + 0.007206338923424482, + 0.005674643442034721, + 0.004812189377844334, + 0.005572568159550428, + 0.009343594312667847, + 0.006457324139773846, + 0.004313453566282988, + 0.008706735447049141, + 0.008562496863305569, + 0.003951104823499918, + 0.004526676144450903, + 0.007840859703719616, + 0.015280107967555523, + 0.011045455001294613, + 0.006279815454035997, + 0.02445579506456852, + 0.014529189094901085, + 0.009816248901188374, + 0.007739380467683077, + 0.005051220767199993, + 0.008874936029314995, + 0.006870853714644909, + 0.0052798776887357235, + 0.005695418920367956, + 0.005403551273047924, + 0.004531147889792919, + 0.005956187844276428, + 0.004347162786871195, + 0.0064736357890069485, + 0.004175282083451748, + 0.003993903752416372, + 0.00675071869045496, + 0.006969128269702196, + 0.003478292841464281, + 0.00486000906676054, + 0.005269212182611227, + 0.007479209918528795, + 0.004549958277493715, + 0.004942173138260841, + 0.009123668074607849, + 0.007070113904774189, + 0.007082017604261637, + 0.006444221828132868, + 0.006313687190413475, + 0.015036724507808685, + 0.007625199854373932, + 0.0070604002103209496, + 0.023950044065713882, + 0.012660738080739975, + 0.00986834242939949, + 0.005524259060621262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 5 + ], + "selected_reward": -0.0017315646400675178 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0181814506649971 + ], + "opposite_logprob_tokens": [ + -4.018181324005127 + ], + "logprob_reward": -0.0181814506649971, + "yes_no_margin_reward": 3.999999761581421, + "token_credit_reward": -0.0181814506649971, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2919245660305023, + 0.07680217176675797, + 0.07668397575616837, + 0.043490178883075714, + 0.1648562103509903, + 0.12385591119527817, + 0.055172327905893326, + 0.06126495823264122, + 0.0546078160405159, + 0.051341839134693146 + ], + "credit_map": [ + 0.027232667431235313, + 0.04715878888964653, + 0.02256634086370468, + 0.03723897784948349, + 0.04331877455115318, + 0.04992247745394707, + 0.031056508421897888, + 0.03343004360795021, + 0.008442348800599575, + 0.011223065666854382, + 0.009159470908343792, + 0.008571977727115154, + 0.013271831907331944, + 0.013145474717020988, + 0.0059041366912424564, + 0.007083863485604525, + 0.007468305993825197, + 0.012371714226901531, + 0.009054269641637802, + 0.005342971533536911, + 0.01719076745212078, + 0.010260283015668392, + 0.008057746104896069, + 0.0069379196502268314, + 0.005362129304558039, + 0.007752818986773491, + 0.004967384971678257, + 0.0031469326931983232, + 0.008175588212907314, + 0.005940149538218975, + 0.004061530809849501, + 0.004083648789674044, + 0.013937645591795444, + 0.02780492976307869, + 0.018730968236923218, + 0.013143092393875122, + 0.04114527627825737, + 0.021150674670934677, + 0.018616102635860443, + 0.01032752450555563, + 0.013599839992821217, + 0.016896402463316917, + 0.02074463851749897, + 0.013624598272144794, + 0.015976885333657265, + 0.016570722684264183, + 0.013207840733230114, + 0.013234988786280155, + 0.0036228562239557505, + 0.008472474291920662, + 0.0036471427883952856, + 0.004829155746847391, + 0.014723383821547031, + 0.00953042134642601, + 0.006550296675413847, + 0.003796602366492152, + 0.004583002999424934, + 0.010917393490672112, + 0.006517019588500261, + 0.005270897410809994, + 0.01588420569896698, + 0.008411113172769547, + 0.006373715586960316, + 0.0033076119143515825, + 0.004359962418675423, + 0.009328273124992847, + 0.0059012011624872684, + 0.004304856527596712, + 0.012546064332127571, + 0.007522829808294773, + 0.0063931928016245365, + 0.004251439590007067, + 0.00523997750133276, + 0.00866605993360281, + 0.008521659299731255, + 0.004955623764544725, + 0.009279160760343075, + 0.0066971080377697945, + 0.004506741650402546, + 0.0034755088854581118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 5 + ], + "selected_reward": -0.0181814506649971 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06202027574181557 + ], + "opposite_logprob_tokens": [ + -2.8120203018188477 + ], + "logprob_reward": -0.06202027574181557, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06202027574181557, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19151708483695984, + 0.19255994260311127, + 0.10289259254932404, + 0.10027838498353958, + 0.08313267678022385, + 0.09080780297517776, + 0.055184315890073776, + 0.05052199214696884, + 0.04828088730573654, + 0.08482421934604645 + ], + "credit_map": [ + 0.01860007271170616, + 0.04522322863340378, + 0.012937761843204498, + 0.020010055974125862, + 0.03111453540623188, + 0.027769066393375397, + 0.015739548951387405, + 0.020122801885008812, + 0.014325344003736973, + 0.028139201924204826, + 0.013375875540077686, + 0.016065876930952072, + 0.049112074077129364, + 0.03693583235144615, + 0.020409010350704193, + 0.01419672742486, + 0.0069799902848899364, + 0.016608763486146927, + 0.008335013873875141, + 0.010828324593603611, + 0.02557075396180153, + 0.016620978713035583, + 0.010097960941493511, + 0.007850809022784233, + 0.006712195463478565, + 0.017354967072606087, + 0.009176918305456638, + 0.007970111444592476, + 0.023760179057717323, + 0.01660020835697651, + 0.009617705829441547, + 0.009086094796657562, + 0.006402559112757444, + 0.013865598477423191, + 0.011948201805353165, + 0.006497625727206469, + 0.014863747172057629, + 0.01313918549567461, + 0.007724767550826073, + 0.008690995164215565, + 0.007416208274662495, + 0.014897455461323261, + 0.008325944654643536, + 0.007416033651679754, + 0.022478023543953896, + 0.012250997126102448, + 0.009062943048775196, + 0.008960197679698467, + 0.0033038498368114233, + 0.008423548191785812, + 0.0045670378021895885, + 0.005140071269124746, + 0.013327207416296005, + 0.010167647153139114, + 0.005651498679071665, + 0.0046034567058086395, + 0.0042927321046590805, + 0.007623780053108931, + 0.00761052779853344, + 0.003696725470945239, + 0.008862627670168877, + 0.008186799474060535, + 0.004716746509075165, + 0.00553205655887723, + 0.003931776154786348, + 0.007691069971770048, + 0.004000022076070309, + 0.003741833847016096, + 0.012149628251791, + 0.00721387006342411, + 0.0052424268797039986, + 0.0043102591298520565, + 0.006409069523215294, + 0.013780824840068817, + 0.0065009151585400105, + 0.007542563136667013, + 0.023425573483109474, + 0.012024120427668095, + 0.008496508002281189, + 0.0066446359269320965 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 2 + ], + "selected_reward": -0.06202027574181557 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01419694535434246 + ], + "opposite_logprob_tokens": [ + -4.264196872711182 + ], + "logprob_reward": -0.01419694535434246, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.01419694535434246, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1526862382888794, + 0.2080165296792984, + 0.11001352965831757, + 0.07919628918170929, + 0.08414594829082489, + 0.11313752830028534, + 0.07024595886468887, + 0.040889985859394073, + 0.0991416946053505, + 0.04252628609538078 + ], + "credit_map": [ + 0.01727142184972763, + 0.026143981143832207, + 0.018339699134230614, + 0.016405794769525528, + 0.02139316499233246, + 0.021341804414987564, + 0.013803866691887379, + 0.01798650063574314, + 0.018592841923236847, + 0.03735964745283127, + 0.020496848970651627, + 0.018136218190193176, + 0.0460001677274704, + 0.02733251079916954, + 0.023570876568555832, + 0.016527412459254265, + 0.009916385635733604, + 0.01781681552529335, + 0.009773527272045612, + 0.010158192366361618, + 0.024906231090426445, + 0.0156885776668787, + 0.01130429096519947, + 0.01044951006770134, + 0.009273814968764782, + 0.013225642964243889, + 0.009196634404361248, + 0.007223064545542002, + 0.014222861267626286, + 0.009818573482334614, + 0.0070783402770757675, + 0.009157354012131691, + 0.007015526760369539, + 0.01284542866051197, + 0.007116177584975958, + 0.011287183500826359, + 0.016581391915678978, + 0.010502174496650696, + 0.009520945139229298, + 0.00927711557596922, + 0.008474566042423248, + 0.01973716914653778, + 0.009862878359854221, + 0.008012903854250908, + 0.029502755030989647, + 0.018101830035448074, + 0.0119257140904665, + 0.007519707549363375, + 0.006366194691509008, + 0.010466675274074078, + 0.005051629617810249, + 0.004946747794747353, + 0.01721494272351265, + 0.013099721632897854, + 0.007619173731654882, + 0.005480873864144087, + 0.004907588008791208, + 0.005304160062223673, + 0.004513605963438749, + 0.006142009515315294, + 0.006674207281321287, + 0.004822356626391411, + 0.004672763403505087, + 0.0038532947655767202, + 0.007134751416742802, + 0.016812479123473167, + 0.005306258797645569, + 0.008286098949611187, + 0.027171852067112923, + 0.01660769246518612, + 0.012076851911842823, + 0.0057457201182842255, + 0.004652445204555988, + 0.00764778396114707, + 0.0034758327528834343, + 0.00489088287577033, + 0.007995480671525002, + 0.005797616671770811, + 0.0038699486758559942, + 0.004196296911686659 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 5 + ], + "selected_reward": -0.01419694535434246 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts a massive red king crab, its form defined by bold black line art, emphasizing its immense size and power. The creature strikes a dramatic, energetic pose atop a massive, jagged slab of volcanic rock, its fractured, obsidian-like surface gleaming with light. Its fiery red shell, painted in high-contrast shading, carved into deep, stark shadows, is covered in sharp, dagger-like spikes. The crab's long, spiny legs are spread wide, its pointed tips gripping the rock for stability. Its massive claws—one a powerful crusher, the other a razor-sharp pincer—are raised high and extended in defensive formation, poised to strike. The entire scene is imbued with a tense, confrontational atmosphere, capturing a frozen moment of heightened alertness, as the crab's tiny, stalk-like eyes scan its surroundings for unseen enemies.", + "seed": 15271, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_9150_seed_15271_dp_8_rank_16_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.03311073146760464, + "vlm_logprob": -0.03311073146760464, + "vlm_yes_no_margin": 3.85, + "vlm_token_credit": -0.03311073146760464, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018295930698513985 + ], + "opposite_logprob_tokens": [ + -4.268295764923096 + ], + "logprob_reward": -0.018295930698513985, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.018295930698513985, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05311744287610054, + 0.5963717103004456, + 0.039770592004060745, + 0.03640715032815933, + 0.027282001450657845, + 0.01900465227663517, + 0.07948832958936691, + 0.01807488314807415, + 0.011411765590310097, + 0.11907145380973816 + ], + "credit_map": [ + 0.008924909867346287, + 0.005450226366519928, + 0.005817415192723274, + 0.003976708743721247, + 0.007905503734946251, + 0.008654926903545856, + 0.008020846173167229, + 0.004366907756775618, + 0.17679141461849213, + 0.0793670117855072, + 0.03637159243226051, + 0.01884552650153637, + 0.030868351459503174, + 0.1900627464056015, + 0.026653548702597618, + 0.03741154074668884, + 0.007239647209644318, + 0.011207793839275837, + 0.0035249730572104454, + 0.0030258558690547943, + 0.005711084697395563, + 0.0033747476991266012, + 0.0028798291459679604, + 0.0028066665399819613, + 0.005819284822791815, + 0.0036330088041722775, + 0.006009761244058609, + 0.005305446684360504, + 0.0037769065238535404, + 0.002605660120025277, + 0.005003849510103464, + 0.0042532361112535, + 0.0048272875137627125, + 0.0027203515637665987, + 0.004387869033962488, + 0.003721205983310938, + 0.0028466416988521814, + 0.002163595287129283, + 0.003289109328761697, + 0.0033259426709264517, + 0.003622356802225113, + 0.0024370921310037374, + 0.002792789600789547, + 0.0024243551306426525, + 0.0018357568187639117, + 0.0016654031351208687, + 0.002070910530164838, + 0.002155988710001111, + 0.010396970435976982, + 0.009047887288033962, + 0.014231841079890728, + 0.01298301387578249, + 0.0077676367945969105, + 0.0076558953151106834, + 0.007891188375651836, + 0.009513898752629757, + 0.002992326393723488, + 0.0019436534494161606, + 0.002504515927284956, + 0.0022222287952899933, + 0.001682397210970521, + 0.0026224227622151375, + 0.002144142519682646, + 0.0019631979521363974, + 0.0015532171819359064, + 0.0017144849989563227, + 0.0017336936434730887, + 0.0014598775887861848, + 0.001435946673154831, + 0.0014232508838176727, + 0.0010625878348946571, + 0.0010287086479365826, + 0.024474389851093292, + 0.010880580171942711, + 0.006352870259433985, + 0.005328197963535786, + 0.04712816700339317, + 0.010016633197665215, + 0.008364603854715824, + 0.006526024546474218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 6 + ], + "selected_reward": -0.018295930698513985 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004121026024222374 + ], + "opposite_logprob_tokens": [ + -5.504120826721191 + ], + "logprob_reward": -0.004121026024222374, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004121026024222374, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.051585592329502106, + 0.43488702178001404, + 0.03966161236166954, + 0.04087177664041519, + 0.012210801243782043, + 0.026691975072026253, + 0.09046171605587006, + 0.10296095907688141, + 0.04030030965805054, + 0.16036824882030487 + ], + "credit_map": [ + 0.010315815918147564, + 0.005986999720335007, + 0.006116455886512995, + 0.004898945335298777, + 0.006848027929663658, + 0.009077939204871655, + 0.00410690251737833, + 0.00423450767993927, + 0.1275535225868225, + 0.06220683455467224, + 0.02674766257405281, + 0.013701953925192356, + 0.023323647677898407, + 0.13161781430244446, + 0.022697415202856064, + 0.02703815884888172, + 0.007258289959281683, + 0.011237379163503647, + 0.004102705977857113, + 0.003989934455603361, + 0.0037358945701271296, + 0.0036268301773816347, + 0.0031166053377091885, + 0.002593970624729991, + 0.008996441960334778, + 0.00793801061809063, + 0.004094170406460762, + 0.0026570677291601896, + 0.0036655752919614315, + 0.007029480766505003, + 0.0031166679691523314, + 0.003374361200258136, + 0.0021002504508942366, + 0.0014028531732037663, + 0.0015978667652234435, + 0.0009706965065561235, + 0.0014991395873948932, + 0.002314751734957099, + 0.0011616058181971312, + 0.001163635984994471, + 0.004320535343140364, + 0.003150111995637417, + 0.0030032687354832888, + 0.0023437985219061375, + 0.0036474009975790977, + 0.005248899571597576, + 0.0021571412216871977, + 0.0028208200819790363, + 0.027812793850898743, + 0.008544797077775002, + 0.006746947765350342, + 0.006084307562559843, + 0.00419720821082592, + 0.026362400501966476, + 0.004030072130262852, + 0.00668318709358573, + 0.021632138639688492, + 0.011764127761125565, + 0.007983213290572166, + 0.00794442929327488, + 0.01293366402387619, + 0.0260796919465065, + 0.006402981001883745, + 0.008220717310905457, + 0.005621446296572685, + 0.004407550673931837, + 0.007687781471759081, + 0.0050836969166994095, + 0.0038490048609673977, + 0.0036096659023314714, + 0.004306894261389971, + 0.005734268110245466, + 0.02952638827264309, + 0.014986472204327583, + 0.007561247795820236, + 0.007279439829289913, + 0.06292040646076202, + 0.01767773926258087, + 0.011562841944396496, + 0.008853708393871784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 7 + ], + "selected_reward": -0.004121026024222374 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000792026519775 + ], + "opposite_logprob_tokens": [ + -2.695007801055908 + ], + "logprob_reward": -0.07000792026519775, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000792026519775, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15661504864692688, + 0.1959218680858612, + 0.08536312729120255, + 0.06829506903886795, + 0.04768461361527443, + 0.025170467793941498, + 0.1075492799282074, + 0.1671488881111145, + 0.07342859357595444, + 0.07282300293445587 + ], + "credit_map": [ + 0.03291034698486328, + 0.013892276212573051, + 0.014762192964553833, + 0.011534377001225948, + 0.02026752382516861, + 0.03422856703400612, + 0.014966585673391819, + 0.014053195714950562, + 0.05410098657011986, + 0.028473403304815292, + 0.013243005611002445, + 0.008547937497496605, + 0.010659716092050076, + 0.056960832327604294, + 0.011449664831161499, + 0.012486325576901436, + 0.0135857705026865, + 0.017049387097358704, + 0.009120749309659004, + 0.008341158740222454, + 0.009334320202469826, + 0.010480755940079689, + 0.009508850052952766, + 0.007942148484289646, + 0.014581282623112202, + 0.008270247839391232, + 0.007889660075306892, + 0.005947403144091368, + 0.006872262340039015, + 0.013246887363493443, + 0.005367704667150974, + 0.006119625177234411, + 0.007929019629955292, + 0.005182162392884493, + 0.005621465854346752, + 0.004832359030842781, + 0.00659391051158309, + 0.008080154657363892, + 0.004291758872568607, + 0.005153785925358534, + 0.003463944885879755, + 0.0032686626072973013, + 0.003614622401073575, + 0.0030056906398385763, + 0.0036227048840373755, + 0.0029706174973398447, + 0.0026510534808039665, + 0.0025731734931468964, + 0.02831200137734413, + 0.00965102855116129, + 0.010497136041522026, + 0.00970896054059267, + 0.005820469465106726, + 0.029531406238675117, + 0.005329966079443693, + 0.008698317222297192, + 0.031768154352903366, + 0.015270846895873547, + 0.019610684365034103, + 0.012798192910850048, + 0.01847865991294384, + 0.04141296446323395, + 0.012506679631769657, + 0.015302727930247784, + 0.008509128354489803, + 0.015378672629594803, + 0.007012358400970697, + 0.008008132688701153, + 0.01140343677252531, + 0.009480957873165607, + 0.0061903223395347595, + 0.007445588242262602, + 0.008671039715409279, + 0.008518348447978497, + 0.008979394100606441, + 0.008914186619222164, + 0.012913492508232594, + 0.008881824091076851, + 0.0064097517170012, + 0.009534972719848156 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 0 + ], + "selected_reward": -0.07000792026519775 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05494030565023422 + ], + "opposite_logprob_tokens": [ + -2.9299402236938477 + ], + "logprob_reward": -0.05494030565023422, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05494030565023422, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05484230816364288, + 0.0863599181175232, + 0.10221538692712784, + 0.06413768976926804, + 0.15527912974357605, + 0.037044916301965714, + 0.06786387413740158, + 0.15654844045639038, + 0.07750213146209717, + 0.19820615649223328 + ], + "credit_map": [ + 0.009124387055635452, + 0.0075267646461725235, + 0.006938963197171688, + 0.004990763496607542, + 0.006145922001451254, + 0.008403385989367962, + 0.006431553978472948, + 0.005280568730086088, + 0.011867844499647617, + 0.012752608396112919, + 0.010073243640363216, + 0.007672321051359177, + 0.005272991489619017, + 0.0177273191511631, + 0.012060423381626606, + 0.008933164179325104, + 0.023034753277897835, + 0.03516696020960808, + 0.006984981242567301, + 0.004964016377925873, + 0.00908439327031374, + 0.00918244756758213, + 0.00791540089994669, + 0.005882437340915203, + 0.01046477910131216, + 0.00680356752127409, + 0.00916079618036747, + 0.008135777898132801, + 0.007114902604371309, + 0.006995409727096558, + 0.008470454253256321, + 0.006992001552134752, + 0.026618370786309242, + 0.016588710248470306, + 0.02425922453403473, + 0.022804055362939835, + 0.018116187304258347, + 0.011353334411978722, + 0.017021892592310905, + 0.01851736195385456, + 0.005270420573651791, + 0.004996231757104397, + 0.005130980163812637, + 0.005047022830694914, + 0.004940001759678125, + 0.0033873338252305984, + 0.00439591845497489, + 0.0038770055398344994, + 0.02001524344086647, + 0.006017354782670736, + 0.005164791829884052, + 0.0033605219796299934, + 0.003501313738524914, + 0.021780414506793022, + 0.002894621342420578, + 0.005129612050950527, + 0.03460441157221794, + 0.016153153032064438, + 0.013609875924885273, + 0.009814503602683544, + 0.018395235762000084, + 0.04089457169175148, + 0.009942175820469856, + 0.013134516775608063, + 0.010142499580979347, + 0.01057348120957613, + 0.012935549952089787, + 0.00922650657594204, + 0.009217067621648312, + 0.009506496600806713, + 0.006765871308743954, + 0.009134652093052864, + 0.04246466979384422, + 0.0175595935434103, + 0.010448961518704891, + 0.009755026549100876, + 0.07653044164180756, + 0.015708869323134422, + 0.014540615491569042, + 0.011197972111403942 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 4 + ], + "selected_reward": -0.05494030565023422 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01818847469985485 + ], + "opposite_logprob_tokens": [ + -4.0181884765625 + ], + "logprob_reward": -0.01818847469985485, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01818847469985485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0726708173751831, + 0.026873674243688583, + 0.06028135120868683, + 0.017139814794063568, + 0.017979200929403305, + 0.015081952325999737, + 0.4915652573108673, + 0.024918092414736748, + 0.07016243040561676, + 0.20332741737365723 + ], + "credit_map": [ + 0.014943377114832401, + 0.007578575052320957, + 0.006400397978723049, + 0.005473119206726551, + 0.009434114210307598, + 0.016520820558071136, + 0.005550750531256199, + 0.006769665516912937, + 0.0035761818289756775, + 0.0029086763970553875, + 0.003724545706063509, + 0.0030227068345993757, + 0.0023891667369753122, + 0.0034673227928578854, + 0.004750633612275124, + 0.0030344396363943815, + 0.013754738494753838, + 0.021623317152261734, + 0.0038926247507333755, + 0.002398149576038122, + 0.005593790207058191, + 0.006476442329585552, + 0.0032845111563801765, + 0.0032577798701822758, + 0.002985097700729966, + 0.0031294040381908417, + 0.00207585864700377, + 0.001237428979948163, + 0.0017240411834791303, + 0.003162228735163808, + 0.0012816369999200106, + 0.0015441174618899822, + 0.0030093842651695013, + 0.0016917483881115913, + 0.0024908212944865227, + 0.0020370413549244404, + 0.00198048772290349, + 0.003248111344873905, + 0.001606074394658208, + 0.0019155306508764625, + 0.0021094607654958963, + 0.0015796332154422998, + 0.002474589040502906, + 0.0015107165090739727, + 0.0014891870087012649, + 0.002667091554030776, + 0.0015294482000172138, + 0.0017218259163200855, + 0.15972378849983215, + 0.05445406213402748, + 0.031790345907211304, + 0.016217021271586418, + 0.022982195019721985, + 0.15739360451698303, + 0.014553052373230457, + 0.03445117920637131, + 0.005005059763789177, + 0.0025291990023106337, + 0.0023407896514981985, + 0.0021215963643044233, + 0.0023767154198139906, + 0.006071956362575293, + 0.0020836880430579185, + 0.0023890864104032516, + 0.010739600285887718, + 0.008427956141531467, + 0.012928186915814877, + 0.009009656496345997, + 0.006816740147769451, + 0.0066438643261790276, + 0.0069182575680315495, + 0.008678169921040535, + 0.041512854397296906, + 0.01916310004889965, + 0.00839114747941494, + 0.007931563071906567, + 0.08297491073608398, + 0.019910741597414017, + 0.013768134638667107, + 0.009674967266619205 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 9, + 0 + ], + "selected_reward": -0.01818847469985485 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts a massive red king crab, its form defined by bold black line art, emphasizing its immense size and power. The creature strikes a dramatic, energetic pose atop a massive, jagged slab of volcanic rock, its fractured, obsidian-like surface gleaming with light. Its fiery red shell, painted in high-contrast shading, carved into deep, stark shadows, is covered in sharp, dagger-like spikes. The crab's long, spiny legs are spread wide, its pointed tips gripping the rock for stability. Its massive claws—one a powerful crusher, the other a razor-sharp pincer—are raised high and extended in defensive formation, poised to strike. The entire scene is imbued with a tense, confrontational atmosphere, capturing a frozen moment of heightened alertness, as the crab's tiny, stalk-like eyes scan its surroundings for unseen enemies.", + "seed": 15272, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_9150_seed_15272_dp_9_rank_18_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.0950024351477623, + "vlm_logprob": -0.0950024351477623, + "vlm_yes_no_margin": 2.75, + "vlm_token_credit": -0.0950024351477623, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04666149988770485 + ], + "opposite_logprob_tokens": [ + -3.171661615371704 + ], + "logprob_reward": -0.04666149988770485, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04666149988770485, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.40439698100090027, + 0.13688704371452332, + 0.13661180436611176, + 0.054126888513565063, + 0.0782470628619194, + 0.11742699891328812, + 0.01567360945045948, + 0.011734111234545708, + 0.0187921691685915, + 0.026103412732481956 + ], + "credit_map": [ + 0.050692468881607056, + 0.03975171223282814, + 0.014971388503909111, + 0.03734209015965462, + 0.15499818325042725, + 0.0504029206931591, + 0.026382317766547203, + 0.02985585480928421, + 0.010459556244313717, + 0.026453306898474693, + 0.013489028438925743, + 0.019056124612689018, + 0.02062496729195118, + 0.01666027121245861, + 0.018714291974902153, + 0.011429494246840477, + 0.017261957749724388, + 0.01108784694224596, + 0.009836842305958271, + 0.009072289802134037, + 0.021968306973576546, + 0.04932265728712082, + 0.009129340760409832, + 0.008932542987167835, + 0.0050788940861821175, + 0.009067081846296787, + 0.0028468018863350153, + 0.005400652997195721, + 0.00921710580587387, + 0.01486784778535366, + 0.004436565563082695, + 0.0032119357492774725, + 0.00768157048150897, + 0.0090108010917902, + 0.0070899855345487595, + 0.006515050772577524, + 0.014873510226607323, + 0.00971249584108591, + 0.01297074370086193, + 0.010392896831035614, + 0.012927149422466755, + 0.018886862322688103, + 0.00719480263069272, + 0.011031411588191986, + 0.02933664619922638, + 0.015658000484108925, + 0.012249561958014965, + 0.010142545215785503, + 0.0014304423239082098, + 0.0026136841624975204, + 0.001474202610552311, + 0.001965450355783105, + 0.002329536247998476, + 0.0024204363580793142, + 0.0017930917674675584, + 0.0016467637615278363, + 0.0012353170895949006, + 0.0019076841417700052, + 0.001430135453119874, + 0.0011251618852838874, + 0.0018441140418872237, + 0.001470447750762105, + 0.0015942325117066503, + 0.0011270176619291306, + 0.0018521430902183056, + 0.0030054005328565836, + 0.0017720587784424424, + 0.0014965144218876958, + 0.0025790866930037737, + 0.003225199179723859, + 0.002562320791184902, + 0.002299443818628788, + 0.0033819691743701696, + 0.0042335279285907745, + 0.0030459000263363123, + 0.002202700125053525, + 0.003280447795987129, + 0.003924602642655373, + 0.002134998794645071, + 0.0038992646150290966 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.04666149988770485 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02060951665043831 + ], + "opposite_logprob_tokens": [ + -3.8956096172332764 + ], + "logprob_reward": -0.02060951665043831, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.02060951665043831, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07575760781764984, + 0.06350386887788773, + 0.08526475727558136, + 0.050460223108530045, + 0.18436157703399658, + 0.4243459701538086, + 0.055478449910879135, + 0.015536523424088955, + 0.03437229245901108, + 0.010918705724179745 + ], + "credit_map": [ + 0.010059114545583725, + 0.007460004650056362, + 0.004166683182120323, + 0.007194250822067261, + 0.029229765757918358, + 0.007593631278723478, + 0.0059453570283949375, + 0.004108802881091833, + 0.005605185870081186, + 0.010654890909790993, + 0.0074800304137170315, + 0.008166537620127201, + 0.009434697218239307, + 0.007072713226079941, + 0.009935068897902966, + 0.005154745653271675, + 0.011108564212918282, + 0.007389588747173548, + 0.005745292641222477, + 0.005644146353006363, + 0.013551194220781326, + 0.03190239518880844, + 0.004596848972141743, + 0.005326730664819479, + 0.005175644066184759, + 0.005766607820987701, + 0.003881293348968029, + 0.005819680169224739, + 0.007058234419673681, + 0.011409048922359943, + 0.0065843164920806885, + 0.004765400197356939, + 0.020561881363391876, + 0.01683085411787033, + 0.012171832844614983, + 0.010796123184263706, + 0.02810066007077694, + 0.055993467569351196, + 0.015774130821228027, + 0.024132629856467247, + 0.0345434844493866, + 0.042352344840765, + 0.026500580832362175, + 0.0383281484246254, + 0.08995925635099411, + 0.11015908420085907, + 0.046064138412475586, + 0.03643893077969551, + 0.0055861943401396275, + 0.00999470055103302, + 0.00389857217669487, + 0.004319431260228157, + 0.013846569694578648, + 0.008334263227880001, + 0.004350791685283184, + 0.005147929303348064, + 0.0013006356311962008, + 0.0022699127439409494, + 0.0021052793599665165, + 0.001139103784225881, + 0.002466225530952215, + 0.0021930974908173084, + 0.0018434006487950683, + 0.0022188688162714243, + 0.004001286346465349, + 0.005963001865893602, + 0.0027494539972394705, + 0.004082792438566685, + 0.004652054514735937, + 0.004640616476535797, + 0.005091638769954443, + 0.003191448748111725, + 0.0013339124852791429, + 0.001563293393701315, + 0.0010036814492195845, + 0.0009188010590150952, + 0.001578918774612248, + 0.0014774490846320987, + 0.001665815943852067, + 0.001376832602545619 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.02060951665043831 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1798633635044098 + ], + "opposite_logprob_tokens": [ + -1.8048633337020874 + ], + "logprob_reward": -0.1798633635044098, + "yes_no_margin_reward": 1.625, + "token_credit_reward": -0.1798633635044098, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1587618887424469, + 0.07632050663232803, + 0.032154761254787445, + 0.03781973943114281, + 0.067843496799469, + 0.4316747784614563, + 0.09232152253389359, + 0.02970351278781891, + 0.04937310516834259, + 0.024026677012443542 + ], + "credit_map": [ + 0.023913593962788582, + 0.015658726915717125, + 0.005921792704612017, + 0.013011354953050613, + 0.06501719355583191, + 0.016271760687232018, + 0.009710044600069523, + 0.00925744790583849, + 0.006643611937761307, + 0.014109964482486248, + 0.008289087563753128, + 0.009807223454117775, + 0.010527466423809528, + 0.009847388602793217, + 0.010320414789021015, + 0.00677535543218255, + 0.0032513001933693886, + 0.004608790390193462, + 0.003401907393708825, + 0.003771240822970867, + 0.004704662598669529, + 0.0038023146335035563, + 0.005106403026729822, + 0.003508144523948431, + 0.0037524865474551916, + 0.006348983384668827, + 0.002298150910064578, + 0.004139885772019625, + 0.005178379826247692, + 0.00896171573549509, + 0.004173263907432556, + 0.002966879401355982, + 0.008552173152565956, + 0.007212970871478319, + 0.004545721225440502, + 0.004203725606203079, + 0.011535558849573135, + 0.01854485645890236, + 0.006520635448396206, + 0.006727860774844885, + 0.03184165433049202, + 0.036574166268110275, + 0.02738935500383377, + 0.0325799435377121, + 0.09547358006238937, + 0.13393259048461914, + 0.039512768387794495, + 0.03437072038650513, + 0.008770328015089035, + 0.01740916632115841, + 0.006327946670353413, + 0.007175264414399862, + 0.022277649492025375, + 0.014003812335431576, + 0.007714715786278248, + 0.008642642758786678, + 0.00350197427906096, + 0.004160505253821611, + 0.003678039414808154, + 0.002445015124976635, + 0.00434727082028985, + 0.00436151959002018, + 0.0034265234135091305, + 0.003782668150961399, + 0.0054954891093075275, + 0.008731236681342125, + 0.0040566809475421906, + 0.005942207761108875, + 0.006600674241781235, + 0.006850356236100197, + 0.007077243644744158, + 0.004619220271706581, + 0.001947105978615582, + 0.004292520694434643, + 0.002287124050781131, + 0.0018699042266234756, + 0.0033062025904655457, + 0.0039946651086211205, + 0.0030999998562037945, + 0.003229158464819193 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 0, + 6 + ], + "selected_reward": -0.1798633635044098 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.20153427124023438 + ], + "opposite_logprob_tokens": [ + -1.7015342712402344 + ], + "logprob_reward": -0.20153427124023438, + "yes_no_margin_reward": 1.5, + "token_credit_reward": -0.20153427124023438, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16701993346214294, + 0.09668964892625809, + 0.027095649391412735, + 0.029934324324131012, + 0.1480594128370285, + 0.08696956932544708, + 0.044404659420251846, + 0.023706767708063126, + 0.043694738298654556, + 0.3324253261089325 + ], + "credit_map": [ + 0.021593932062387466, + 0.014992736279964447, + 0.007754059042781591, + 0.013725722208619118, + 0.05858652666211128, + 0.020420484244823456, + 0.011785769835114479, + 0.01816071756184101, + 0.008601916022598743, + 0.01730104349553585, + 0.010444470681250095, + 0.012801378034055233, + 0.013258683495223522, + 0.013664149679243565, + 0.012189419940114021, + 0.008428594097495079, + 0.0027167368680238724, + 0.0022073988802731037, + 0.002429983578622341, + 0.0031261593103408813, + 0.003711487166583538, + 0.007391586434096098, + 0.0034020820166915655, + 0.0021102149039506912, + 0.002724271733313799, + 0.0036073399242013693, + 0.002548073185607791, + 0.004098555073142052, + 0.003865580540150404, + 0.005072749685496092, + 0.004167141858488321, + 0.003850613720715046, + 0.017226558178663254, + 0.018301652744412422, + 0.007763030473142862, + 0.010263738222420216, + 0.029493208974599838, + 0.038980040699243546, + 0.015860378742218018, + 0.01017079595476389, + 0.010296267457306385, + 0.008084467612206936, + 0.006376044824719429, + 0.006407233886420727, + 0.012891445308923721, + 0.027585409581661224, + 0.00807097740471363, + 0.007257722783833742, + 0.004225416574627161, + 0.007666300050914288, + 0.0023247224744409323, + 0.0027677644975483418, + 0.013116001151502132, + 0.006977497600018978, + 0.0038200183771550655, + 0.0035069386940449476, + 0.0025650006718933582, + 0.003819303587079048, + 0.0022569468710571527, + 0.002590469317510724, + 0.004160990007221699, + 0.003623961005359888, + 0.002607596106827259, + 0.002082499908283353, + 0.003611036343500018, + 0.006700201891362667, + 0.005093009676784277, + 0.00345876207575202, + 0.007976287975907326, + 0.005581732373684645, + 0.006477044895291328, + 0.0047966595739126205, + 0.03142567723989487, + 0.05217070132493973, + 0.04052230715751648, + 0.027262216433882713, + 0.045130837708711624, + 0.057020239531993866, + 0.03415276110172272, + 0.04474059119820595 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 4 + ], + "selected_reward": -0.20153427124023438 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634352445602417 + ], + "opposite_logprob_tokens": [ + -3.651343584060669 + ], + "logprob_reward": -0.02634352445602417, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634352445602417, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22724854946136475, + 0.07392514497041702, + 0.030476855114102364, + 0.0396069698035717, + 0.1476077139377594, + 0.14256878197193146, + 0.16666501760482788, + 0.03442469239234924, + 0.06099509820342064, + 0.07648118585348129 + ], + "credit_map": [ + 0.027520129457116127, + 0.024698294699192047, + 0.00889304094016552, + 0.021746717393398285, + 0.08300697058439255, + 0.028178513050079346, + 0.015066808089613914, + 0.018138060346245766, + 0.005973284598439932, + 0.009569979272782803, + 0.009379200637340546, + 0.008469175547361374, + 0.011397479102015495, + 0.010616800747811794, + 0.01130003109574318, + 0.0072191921062767506, + 0.0030740287620574236, + 0.0027040832210332155, + 0.0023174944799393415, + 0.002825002418830991, + 0.005384538322687149, + 0.009760579094290733, + 0.0019771018996834755, + 0.002434027148410678, + 0.0031735370866954327, + 0.009497251361608505, + 0.002921082777902484, + 0.004816051106899977, + 0.006476258859038353, + 0.00515126995742321, + 0.004380745813250542, + 0.0031907702796161175, + 0.01939111016690731, + 0.010589654557406902, + 0.007298617623746395, + 0.008884910494089127, + 0.02438150905072689, + 0.06069985777139664, + 0.007568096276372671, + 0.008793960325419903, + 0.011141974478960037, + 0.016304781660437584, + 0.009402409195899963, + 0.014177551493048668, + 0.03227207064628601, + 0.03079216554760933, + 0.016450034454464912, + 0.012027792632579803, + 0.017108816653490067, + 0.028719164431095123, + 0.010931737720966339, + 0.011800989508628845, + 0.044717200100421906, + 0.025347115471959114, + 0.013054072856903076, + 0.014985926449298859, + 0.0026794690638780594, + 0.004929900169372559, + 0.005100649781525135, + 0.002446216531097889, + 0.004764959216117859, + 0.005077691748738289, + 0.004363612737506628, + 0.0050621917471289635, + 0.006672043353319168, + 0.010659098625183105, + 0.00516856973990798, + 0.0073332772590219975, + 0.008020853623747826, + 0.008161707781255245, + 0.009067029692232609, + 0.005912515800446272, + 0.007259472273290157, + 0.015204083174467087, + 0.00913385208696127, + 0.005743303336203098, + 0.009867490269243717, + 0.011774254031479359, + 0.007327659521251917, + 0.010171078145503998 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 4 + ], + "selected_reward": -0.02634352445602417 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "low-angle, static shot depicts a massive red king crab, its form defined by bold black line art, emphasizing its immense size and power. The creature strikes a dramatic, energetic pose atop a massive, jagged slab of volcanic rock, its fractured, obsidian-like surface gleaming with light. Its fiery red shell, painted in high-contrast shading, carved into deep, stark shadows, is covered in sharp, dagger-like spikes. The crab's long, spiny legs are spread wide, its pointed tips gripping the rock for stability. Its massive claws—one a powerful crusher, the other a razor-sharp pincer—are raised high and extended in defensive formation, poised to strike. The entire scene is imbued with a tense, confrontational atmosphere, capturing a frozen moment of heightened alertness, as the crab's tiny, stalk-like eyes scan its surroundings for unseen enemies.", + "seed": 15273, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_4/index_9150_seed_15273_dp_9_rank_18_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 4, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.046253875736147164, + "vlm_logprob": -0.046253875736147164, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.046253875736147164, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0740433782339096 + ], + "opposite_logprob_tokens": [ + -2.6990432739257812 + ], + "logprob_reward": -0.0740433782339096, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.0740433782339096, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09358015656471252, + 0.04849971458315849, + 0.14348606765270233, + 0.0806216225028038, + 0.0330447182059288, + 0.43016350269317627, + 0.09118708223104477, + 0.02599213272333145, + 0.016701454296708107, + 0.03672349452972412 + ], + "credit_map": [ + 0.016099873930215836, + 0.012005333788692951, + 0.012343965470790863, + 0.009509287774562836, + 0.007729611825197935, + 0.014849565923213959, + 0.010173789225518703, + 0.010868735611438751, + 0.012473422102630138, + 0.00424411753192544, + 0.0037028815131634474, + 0.0029501765966415405, + 0.004794709384441376, + 0.011900644749403, + 0.002474256092682481, + 0.005959507543593645, + 0.022432606667280197, + 0.0138543751090765, + 0.012557313777506351, + 0.016277821734547615, + 0.012389296665787697, + 0.0332641527056694, + 0.011587143875658512, + 0.02112337201833725, + 0.011567417532205582, + 0.008152893744409084, + 0.014218753203749657, + 0.021194351837038994, + 0.006143159233033657, + 0.005048743914812803, + 0.007116889581084251, + 0.007179417181760073, + 0.006012758705765009, + 0.003820429788902402, + 0.002520490437746048, + 0.0029533267952501774, + 0.0036209782119840384, + 0.00791869591921568, + 0.0018439595587551594, + 0.004354081582278013, + 0.07155860215425491, + 0.04365253075957298, + 0.03801027685403824, + 0.04153057560324669, + 0.04072391614317894, + 0.10884853452444077, + 0.03294971585273743, + 0.0528893768787384, + 0.017192259430885315, + 0.010685352608561516, + 0.006983563769608736, + 0.009226717986166477, + 0.008957761339843273, + 0.011880329810082912, + 0.011350706219673157, + 0.014910398051142693, + 0.009980370290577412, + 0.0018666359828785062, + 0.0018424356821924448, + 0.0015464704483747482, + 0.00405114097520709, + 0.003062272211536765, + 0.0016075997846201062, + 0.002035209210589528, + 0.001987307332456112, + 0.0027357153594493866, + 0.0018099828157573938, + 0.00206795078702271, + 0.0026668275240808725, + 0.0019459149334579706, + 0.0016913266154006124, + 0.001796429743990302, + 0.005076862871646881, + 0.005003584548830986, + 0.003374938853085041, + 0.005230332259088755, + 0.004699363373219967, + 0.00540182227268815, + 0.004655224271118641, + 0.003281369572505355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 0 + ], + "selected_reward": -0.0740433782339096 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0013552061282098293 + ], + "opposite_logprob_tokens": [ + -6.626355171203613 + ], + "logprob_reward": -0.0013552061282098293, + "yes_no_margin_reward": 6.625, + "token_credit_reward": -0.0013552061282098293, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.03825119510293007, + 0.29257091879844666, + 0.0428638719022274, + 0.12065854668617249, + 0.2395590990781784, + 0.06134575605392456, + 0.05458703264594078, + 0.043315108865499496, + 0.05130176991224289, + 0.05554675683379173 + ], + "credit_map": [ + 0.005970823112875223, + 0.005378497764468193, + 0.005897160153836012, + 0.005235145799815655, + 0.002854390535503626, + 0.004558902699500322, + 0.003653047839179635, + 0.00470322510227561, + 0.051703691482543945, + 0.02727421745657921, + 0.022682910785079002, + 0.01950949989259243, + 0.031590692698955536, + 0.0882972925901413, + 0.016550244763493538, + 0.034962356090545654, + 0.0057080951519310474, + 0.0040057869628071785, + 0.004255236592143774, + 0.0058299582451581955, + 0.0034860132727771997, + 0.009546796791255474, + 0.003584976540878415, + 0.006447004619985819, + 0.036591965705156326, + 0.009571367874741554, + 0.007240517530590296, + 0.037824176251888275, + 0.006438687909394503, + 0.008206860162317753, + 0.0069954534992575645, + 0.007789507508277893, + 0.04111171513795853, + 0.025663606822490692, + 0.01823711395263672, + 0.021087054163217545, + 0.02708517760038376, + 0.060177452862262726, + 0.012995561584830284, + 0.033201418817043304, + 0.009855405427515507, + 0.006407748442143202, + 0.00548628531396389, + 0.006061133928596973, + 0.005819813814014196, + 0.015523646026849747, + 0.004685717169195414, + 0.007505999878048897, + 0.01099738385528326, + 0.005357852205634117, + 0.0037301809061318636, + 0.006508887279778719, + 0.004411735106259584, + 0.0065918429754674435, + 0.005427259486168623, + 0.01156188640743494, + 0.013787862844765186, + 0.0044305576011538506, + 0.003416171995922923, + 0.0027873259969055653, + 0.0069735064171254635, + 0.004947292152792215, + 0.0030345323029905558, + 0.0039378562942147255, + 0.012013622559607029, + 0.004996932111680508, + 0.0025755998212844133, + 0.014836468733847141, + 0.004670714493840933, + 0.004189027938991785, + 0.005821152124553919, + 0.002198250265792012, + 0.007429168093949556, + 0.009771604090929031, + 0.005526663735508919, + 0.006149803288280964, + 0.006644976790994406, + 0.008261797949671745, + 0.006600379478186369, + 0.005162358749657869 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 3 + ], + "selected_reward": -0.0013552061282098293 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.10028864443302155 + ], + "opposite_logprob_tokens": [ + -2.3502886295318604 + ], + "logprob_reward": -0.10028864443302155, + "yes_no_margin_reward": 2.25, + "token_credit_reward": -0.10028864443302155, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08588745445013046, + 0.052911411970853806, + 0.197246253490448, + 0.07169350981712341, + 0.06709668040275574, + 0.26559311151504517, + 0.02794414944946766, + 0.08519072830677032, + 0.04888957738876343, + 0.09754710644483566 + ], + "credit_map": [ + 0.014358904212713242, + 0.012603153474628925, + 0.009533844888210297, + 0.00910135731101036, + 0.00728054903447628, + 0.013597135432064533, + 0.010136719793081284, + 0.00927579402923584, + 0.01754283532500267, + 0.00708579458296299, + 0.005444999318569899, + 0.004252279177308083, + 0.004948351066559553, + 0.005294710397720337, + 0.0030021348502486944, + 0.005340305157005787, + 0.031247319653630257, + 0.018878312781453133, + 0.017370395362377167, + 0.021889738738536835, + 0.017339905723929405, + 0.04669841006398201, + 0.01565409265458584, + 0.028168099001049995, + 0.018982095643877983, + 0.0068191587924957275, + 0.004725220147520304, + 0.02033773995935917, + 0.004389130976051092, + 0.006643061060458422, + 0.005234254989773035, + 0.004562842659652233, + 0.009480596520006657, + 0.007898643612861633, + 0.006178568582981825, + 0.007004525978118181, + 0.008144297637045383, + 0.016161678358912468, + 0.0051809255965054035, + 0.007047439459711313, + 0.04319820925593376, + 0.02718532271683216, + 0.02393987774848938, + 0.025974946096539497, + 0.024761775508522987, + 0.0666174441576004, + 0.020671004429459572, + 0.0332445353269577, + 0.005061198025941849, + 0.0034403777681291103, + 0.0024865344166755676, + 0.003184234257787466, + 0.002819020999595523, + 0.003735616570338607, + 0.0035051910672336817, + 0.0037119747139513493, + 0.015749439597129822, + 0.013596609234809875, + 0.00795009359717369, + 0.007787307724356651, + 0.011105995625257492, + 0.008677295409142971, + 0.008986732922494411, + 0.011337250471115112, + 0.011333970353007317, + 0.005127632990479469, + 0.002387134125456214, + 0.014439152553677559, + 0.005235118791460991, + 0.0031175503972917795, + 0.005205125082284212, + 0.0020438884384930134, + 0.014801260083913803, + 0.011928319931030273, + 0.01097055058926344, + 0.014862176030874252, + 0.01056027039885521, + 0.01067445334047079, + 0.015039361082017422, + 0.008710712194442749 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 9 + ], + "selected_reward": -0.10028864443302155 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0430445596575737 + ], + "opposite_logprob_tokens": [ + -3.1680445671081543 + ], + "logprob_reward": -0.0430445596575737, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0430445596575737, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06775133311748505, + 0.05160804092884064, + 0.19662660360336304, + 0.26614388823509216, + 0.06196601316332817, + 0.01666739583015442, + 0.13458074629306793, + 0.04690710827708244, + 0.050928037613630295, + 0.10682083666324615 + ], + "credit_map": [ + 0.010110341012477875, + 0.010194478556513786, + 0.007761686574667692, + 0.006168785504996777, + 0.006133370567113161, + 0.013187170028686523, + 0.007781838532537222, + 0.006413663271814585, + 0.008769123815000057, + 0.006644920911639929, + 0.0055764117278158665, + 0.004745019134134054, + 0.006499649956822395, + 0.010957622900605202, + 0.0032246559858322144, + 0.005190637893974781, + 0.03523428365588188, + 0.018779106438159943, + 0.0158398300409317, + 0.01782171241939068, + 0.018036793917417526, + 0.05000706762075424, + 0.013882783241569996, + 0.02702503278851509, + 0.056985050439834595, + 0.01781223528087139, + 0.0323868989944458, + 0.09567319601774216, + 0.016511347144842148, + 0.01400725170969963, + 0.017524026334285736, + 0.015243874862790108, + 0.011366273276507854, + 0.007836504839360714, + 0.005753436125814915, + 0.006582110654562712, + 0.006401458289474249, + 0.011500709690153599, + 0.004494310822337866, + 0.008031210862100124, + 0.0035066590644419193, + 0.00242994399741292, + 0.001976451138034463, + 0.0020485396962612867, + 0.0018486533081158996, + 0.001742248423397541, + 0.0013159014051780105, + 0.0017989985644817352, + 0.028759844601154327, + 0.008983413688838482, + 0.008999115787446499, + 0.02219283953309059, + 0.008435636758804321, + 0.012364744208753109, + 0.00821472518146038, + 0.03663042187690735, + 0.014723656699061394, + 0.0041166795417666435, + 0.003954913001507521, + 0.0033761723898351192, + 0.007504416164010763, + 0.004911488853394985, + 0.003252185881137848, + 0.005067594815045595, + 0.013161823153495789, + 0.004874418955296278, + 0.0032725201454013586, + 0.010141738690435886, + 0.004910474177449942, + 0.006182318553328514, + 0.004938678350299597, + 0.003446064190939069, + 0.011310910806059837, + 0.01634286344051361, + 0.00995546579360962, + 0.016936570405960083, + 0.018164344131946564, + 0.013258003629744053, + 0.011715305969119072, + 0.009137372486293316 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 6 + ], + "selected_reward": -0.0430445596575737 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a massive red king crab?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the visual style defined by bold black line art with high-contrast shading?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the crab positioned atop a massive, jagged slab of volcanic rock with an obsidian-like surface?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Are the crab's claws raised high and extended in a defensive formation?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the shot a low-angle, static shot?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.012537590228021145 + ], + "opposite_logprob_tokens": [ + -4.387537479400635 + ], + "logprob_reward": -0.012537590228021145, + "yes_no_margin_reward": 4.375, + "token_credit_reward": -0.012537590228021145, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06210268661379814, + 0.03360108286142349, + 0.02893046848475933, + 0.037837520241737366, + 0.10434029251337051, + 0.24553780257701874, + 0.3920819163322449, + 0.038883745670318604, + 0.0258695799857378, + 0.03081483766436577 + ], + "credit_map": [ + 0.008713017217814922, + 0.006438245065510273, + 0.009840317070484161, + 0.012104659341275692, + 0.00538694066926837, + 0.006261047907173634, + 0.0066775609739124775, + 0.006680906750261784, + 0.008306083269417286, + 0.002956165699288249, + 0.0026035974733531475, + 0.0025102393701672554, + 0.002990293549373746, + 0.008222965523600578, + 0.002058885060250759, + 0.0039528547786176205, + 0.003689577104523778, + 0.004126219544559717, + 0.002965671243146062, + 0.0032286474015563726, + 0.002965651685371995, + 0.00555396918207407, + 0.002459812443703413, + 0.00394092034548521, + 0.008598758839070797, + 0.00310034048743546, + 0.0037826288025826216, + 0.012447978369891644, + 0.002148261060938239, + 0.002433565678074956, + 0.0027079940773546696, + 0.002617995021864772, + 0.015017982572317123, + 0.0125742731615901, + 0.008875065483152866, + 0.00965688843280077, + 0.01317367423325777, + 0.025263356044888496, + 0.0062639592215418816, + 0.013515107333660126, + 0.04057898372411728, + 0.02532682567834854, + 0.021609721705317497, + 0.02365630865097046, + 0.023402463644742966, + 0.06146298348903656, + 0.01926734670996666, + 0.03023318201303482, + 0.04863544926047325, + 0.03444167599081993, + 0.029130786657333374, + 0.06438914686441422, + 0.039418451488018036, + 0.09799908101558685, + 0.023784678429365158, + 0.05428269878029823, + 0.01407722570002079, + 0.0033391746692359447, + 0.002915413584560156, + 0.0022031718399375677, + 0.006355988793075085, + 0.004051214549690485, + 0.0025273291394114494, + 0.003414230654016137, + 0.004516722168773413, + 0.0031824701000005007, + 0.0016119256615638733, + 0.006055038422346115, + 0.0028001265600323677, + 0.002622813917696476, + 0.003182023297995329, + 0.001898462767712772, + 0.004423364996910095, + 0.0035095091443508863, + 0.0042383503168821335, + 0.005285651423037052, + 0.0030573674011975527, + 0.003013746114447713, + 0.004538459703326225, + 0.0027483913581818342 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.012537590228021145 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_tensor([8392, 8392])_group_5_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_tensor([8392, 8392])_group_5_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..f9ac4249ce2002ef937107945703951d432a667e --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_tensor([8392, 8392])_group_5_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "medium close-up shot, shot from a slightly upward angle, with an almost imperceptible slow push-in, gradually approaches a man in intense agony. He wears a charcoal-gray, crew-neck cotton T-shirt, his collar and temples soaked with sweat, causing his short, dark brown hair to stick to his forehead. He raises his hands in despair, pressing the heels of his palms against his temples, his fingers splayed out, digging deep into his hair, as if attempting to physically stifle the invisible torment raging in his mind. His eyes are tightly closed, his eyelids trembling slightly from the effort, and his brows are knitted into a painful knot, creating deep vertical lines between them. A single, warm-toned desk light from the right side of the frame casts a stark contrast of light and dark on his face, highlighting the tense jawline and the ridges of muscle in his cheeks from clenching his teeth. The background is a dim and blurry interior, with only the vague outline of the out-of-focus bookshelf able to be discerned. The entire scene is shrouded in a repressive, tense and extremely private atmosphere, conveying an indescribable physical or psychological pain.", + "seed": 26473, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_8392_seed_26473_dp_10_rank_20_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.04870626945048571, + "vlm_logprob": -0.04870626945048571, + "vlm_yes_no_margin": 3.65, + "vlm_token_credit": -0.04870626945048571, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01261846348643303 + ], + "opposite_logprob_tokens": [ + -5.887618541717529 + ], + "logprob_reward": -0.01261846348643303, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.01261846348643303, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.306159108877182, + 0.07912437617778778, + 0.06379731744527817, + 0.11988939344882965, + 0.040449753403663635, + 0.06207796931266785, + 0.07871963828802109, + 0.09602494537830353, + 0.03823549672961235, + 0.11552194505929947 + ], + "credit_map": [ + 0.03807908669114113, + 0.04190576821565628, + 0.04572630673646927, + 0.045445751398801804, + 0.03839505836367607, + 0.020671892911195755, + 0.03984517231583595, + 0.03609008342027664, + 0.010034655220806599, + 0.007140418980270624, + 0.012866110540926456, + 0.01648678630590439, + 0.010596039704978466, + 0.003917654976248741, + 0.011197750456631184, + 0.00688497070223093, + 0.008797370828688145, + 0.006274539977312088, + 0.01504601165652275, + 0.0056772902607917786, + 0.009534447453916073, + 0.004600724205374718, + 0.008507299236953259, + 0.005359645001590252, + 0.01663001999258995, + 0.009625009261071682, + 0.026270415633916855, + 0.012418456375598907, + 0.01581679843366146, + 0.0054288893006742, + 0.02381785772740841, + 0.009881953708827496, + 0.003118386259302497, + 0.0036687941756099463, + 0.003719642525538802, + 0.005274684634059668, + 0.007252419833093882, + 0.0027877080719918013, + 0.006600576918572187, + 0.008027544245123863, + 0.008529993705451488, + 0.006582078989595175, + 0.012185543775558472, + 0.006042430177330971, + 0.0070630633272230625, + 0.0034670347813516855, + 0.011069740168750286, + 0.007138086948543787, + 0.01193835400044918, + 0.008342625573277473, + 0.013194823637604713, + 0.008260906673967838, + 0.011375688016414642, + 0.005152111407369375, + 0.011483646929264069, + 0.008971485309302807, + 0.0133468983694911, + 0.009251470677554607, + 0.015199135057628155, + 0.013071761466562748, + 0.013295810669660568, + 0.007791445124894381, + 0.015519526787102222, + 0.008548899553716183, + 0.005297886673361063, + 0.0030997341964393854, + 0.005037464201450348, + 0.005106980446726084, + 0.0072657219134271145, + 0.0032157725654542446, + 0.006369624752551317, + 0.002842317335307598, + 0.013156511820852757, + 0.010428767651319504, + 0.008648409508168697, + 0.01872936449944973, + 0.02840775065124035, + 0.010642295703291893, + 0.014397995546460152, + 0.011110861785709858 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.01261846348643303 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890641063451767 + ], + "opposite_logprob_tokens": [ + -2.464064121246338 + ], + "logprob_reward": -0.0890641063451767, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890641063451767, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09752041846513748, + 0.09057150781154633, + 0.07057803124189377, + 0.0677405446767807, + 0.03710287809371948, + 0.12228833138942719, + 0.3238862454891205, + 0.06446868181228638, + 0.05759645625948906, + 0.06824694573879242 + ], + "credit_map": [ + 0.019691329449415207, + 0.009148071520030499, + 0.015032489784061909, + 0.008391440846025944, + 0.012320631183683872, + 0.006272871978580952, + 0.015785537660121918, + 0.010878038592636585, + 0.013067353516817093, + 0.008861959911882877, + 0.014043849892914295, + 0.021027686074376106, + 0.009437378495931625, + 0.003920719958841801, + 0.012404145672917366, + 0.007808412425220013, + 0.007765960413962603, + 0.0066587068140506744, + 0.016984444111585617, + 0.007741523906588554, + 0.01365304458886385, + 0.003969032317399979, + 0.009204613976180553, + 0.004600700456649065, + 0.009480617009103298, + 0.005067262798547745, + 0.014116190373897552, + 0.006814321503043175, + 0.009840899147093296, + 0.005865567829459906, + 0.011328864842653275, + 0.005226814188063145, + 0.0033014381770044565, + 0.003145083086565137, + 0.0043235355988144875, + 0.004476944915950298, + 0.007872011512517929, + 0.002720694290474057, + 0.0064773899503052235, + 0.004785778000950813, + 0.013099548406898975, + 0.007239393889904022, + 0.026556873694062233, + 0.013126974925398827, + 0.013447229750454426, + 0.006126714404672384, + 0.02312321774661541, + 0.019568372517824173, + 0.034201670438051224, + 0.03261526674032211, + 0.039788082242012024, + 0.02833104319870472, + 0.08437096327543259, + 0.026091979816555977, + 0.04502200335264206, + 0.03346520662307739, + 0.006038793362677097, + 0.0067529696971178055, + 0.006930344272404909, + 0.007150182034820318, + 0.013874332420527935, + 0.005496452562510967, + 0.009712225757539272, + 0.00851337518543005, + 0.005843016318976879, + 0.005746098700910807, + 0.005028450395911932, + 0.007214444223791361, + 0.015230313874781132, + 0.0048187957145273685, + 0.008252168074250221, + 0.005463162902742624, + 0.00812396127730608, + 0.005825890693813562, + 0.0048103914596140385, + 0.01057499460875988, + 0.016115859150886536, + 0.0064393593929708, + 0.00825614482164383, + 0.008100335486233234 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.0890641063451767 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.009786837734282017 + ], + "opposite_logprob_tokens": [ + -4.634786605834961 + ], + "logprob_reward": -0.009786837734282017, + "yes_no_margin_reward": 4.625, + "token_credit_reward": -0.009786837734282017, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27389779686927795, + 0.12879808247089386, + 0.08149480819702148, + 0.10208301246166229, + 0.032117653638124466, + 0.0854300782084465, + 0.12721461057662964, + 0.06866808980703354, + 0.02594953216612339, + 0.07434629648923874 + ], + "credit_map": [ + 0.0441213957965374, + 0.04203807935118675, + 0.03772907331585884, + 0.025421882048249245, + 0.03749056160449982, + 0.025435861200094223, + 0.03521798551082611, + 0.026442978531122208, + 0.012758645229041576, + 0.014425625093281269, + 0.019258996471762657, + 0.018503965809941292, + 0.02884138934314251, + 0.008470991626381874, + 0.016598407179117203, + 0.009940069168806076, + 0.009604335762560368, + 0.009844041429460049, + 0.012762644328176975, + 0.0066118608228862286, + 0.01772158034145832, + 0.005798760801553726, + 0.010204763151705265, + 0.008946813642978668, + 0.012959356419742107, + 0.009590987116098404, + 0.019779717549681664, + 0.011637313291430473, + 0.01938512548804283, + 0.0059798420406877995, + 0.015427825972437859, + 0.007322841789573431, + 0.005559643730521202, + 0.002899276092648506, + 0.004106380045413971, + 0.003288849489763379, + 0.005313153378665447, + 0.002124018734320998, + 0.005189598072320223, + 0.0036367347929626703, + 0.013682495802640915, + 0.011239642277359962, + 0.018917139619588852, + 0.008356685750186443, + 0.007059674710035324, + 0.005469291936606169, + 0.012067063711583614, + 0.008638083003461361, + 0.0134992441162467, + 0.016955696046352386, + 0.013387615792453289, + 0.010959979146718979, + 0.03281757980585098, + 0.009184347465634346, + 0.018249565735459328, + 0.01216058898717165, + 0.010168366134166718, + 0.006599463988095522, + 0.009535286575555801, + 0.007302061654627323, + 0.011510313488543034, + 0.005833004601299763, + 0.010978661477565765, + 0.006740927696228027, + 0.0029431094881147146, + 0.0022630931343883276, + 0.003569129155948758, + 0.004160415381193161, + 0.004797945264726877, + 0.0028379515279084444, + 0.003434831276535988, + 0.0019430581014603376, + 0.009137425571680069, + 0.006991172209382057, + 0.007535616867244244, + 0.009742150083184242, + 0.01406457182019949, + 0.007689418271183968, + 0.010101934894919395, + 0.009084003046154976 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 6 + ], + "selected_reward": -0.009786837734282017 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07001993060112 + ], + "opposite_logprob_tokens": [ + -2.6950199604034424 + ], + "logprob_reward": -0.07001993060112, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07001993060112, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3544977009296417, + 0.14101125299930573, + 0.1243809163570404, + 0.09888675808906555, + 0.07646454125642776, + 0.04410577565431595, + 0.05486709251999855, + 0.026346275582909584, + 0.02732519619166851, + 0.052114471793174744 + ], + "credit_map": [ + 0.05204758048057556, + 0.050262946635484695, + 0.05877075344324112, + 0.04821731150150299, + 0.04314839094877243, + 0.027348995208740234, + 0.04272891581058502, + 0.031972840428352356, + 0.015901217237114906, + 0.01740420050919056, + 0.01843396946787834, + 0.025282757356762886, + 0.0191334281116724, + 0.007293807342648506, + 0.022600525990128517, + 0.014961348846554756, + 0.012235489673912525, + 0.009956730529665947, + 0.03724262863397598, + 0.015711085870862007, + 0.0197211354970932, + 0.005529578309506178, + 0.016872074455022812, + 0.007112183608114719, + 0.0134694529697299, + 0.008199942298233509, + 0.021923143416643143, + 0.009973105974495411, + 0.012378821149468422, + 0.007766375318169594, + 0.016375310719013214, + 0.008800608105957508, + 0.008022351190447807, + 0.004947901237756014, + 0.007975362241268158, + 0.007610226050019264, + 0.01027649361640215, + 0.006790216080844402, + 0.012898121029138565, + 0.01794387400150299, + 0.004614529199898243, + 0.002849131589755416, + 0.0077954442240297794, + 0.003286130027845502, + 0.004093350376933813, + 0.002515555592253804, + 0.008242198266088963, + 0.010709433816373348, + 0.007004231214523315, + 0.005792959593236446, + 0.0056394292041659355, + 0.004348649177700281, + 0.013807162642478943, + 0.0051571824587881565, + 0.008195933885872364, + 0.004921541549265385, + 0.0038285828195512295, + 0.0024216510355472565, + 0.0037929946556687355, + 0.003105351235717535, + 0.004285809583961964, + 0.002162558725103736, + 0.003911591600626707, + 0.0028377361595630646, + 0.0026534933131188154, + 0.002637050813063979, + 0.0024606306105852127, + 0.004039233084768057, + 0.006908662151545286, + 0.0022603452671319246, + 0.0037488113157451153, + 0.002616969170048833, + 0.006342846900224686, + 0.004397474229335785, + 0.0037531754933297634, + 0.008864366449415684, + 0.011677112430334091, + 0.004889579955488443, + 0.006695917807519436, + 0.0054939985275268555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.07001993060112 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062042009085416794 + ], + "opposite_logprob_tokens": [ + -2.812041997909546 + ], + "logprob_reward": -0.062042009085416794, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062042009085416794, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10420477390289307, + 0.14561490714550018, + 0.2752317786216736, + 0.10686569660902023, + 0.06339841336011887, + 0.17203576862812042, + 0.030872469767928123, + 0.04208441078662872, + 0.022903675213456154, + 0.03678809478878975 + ], + "credit_map": [ + 0.01838870719075203, + 0.01365779060870409, + 0.015169963240623474, + 0.009767941199243069, + 0.018665499985218048, + 0.007466995157301426, + 0.014160187914967537, + 0.006927682552486658, + 0.01918363757431507, + 0.015605098567903042, + 0.017611639574170113, + 0.015099550597369671, + 0.03294681757688522, + 0.008060041815042496, + 0.02417820878326893, + 0.012929911725223064, + 0.02571030706167221, + 0.024412481114268303, + 0.0774860829114914, + 0.033750101923942566, + 0.04273873195052147, + 0.015414906665682793, + 0.03615863621234894, + 0.01956051215529442, + 0.01139629352837801, + 0.010226160287857056, + 0.02745630592107773, + 0.01053173653781414, + 0.012212773784995079, + 0.008270911872386932, + 0.018262628465890884, + 0.008508888073265553, + 0.004635295830667019, + 0.0039855013601481915, + 0.00861882884055376, + 0.005957418121397495, + 0.007357527036219835, + 0.005912066902965307, + 0.013421657495200634, + 0.01351012010127306, + 0.018719952553510666, + 0.023645061999559402, + 0.02595728635787964, + 0.017088549211621284, + 0.019941430538892746, + 0.01268917415291071, + 0.029949475079774857, + 0.024044841527938843, + 0.004823707975447178, + 0.003139793872833252, + 0.003282166551798582, + 0.0027458483818918467, + 0.005524537991732359, + 0.0030432480853050947, + 0.004850288387387991, + 0.0034628796856850386, + 0.004924275446683168, + 0.0035262331366539, + 0.005931542720645666, + 0.005429241806268692, + 0.007722217123955488, + 0.0034266647417098284, + 0.006496216636151075, + 0.004628018941730261, + 0.003493453375995159, + 0.0017062707338482141, + 0.0024022976867854595, + 0.00462574465200305, + 0.003534948220476508, + 0.0015541277825832367, + 0.0038494286127388477, + 0.001737403217703104, + 0.004087657202035189, + 0.0023320300970226526, + 0.004525862634181976, + 0.009862094186246395, + 0.00405602902173996, + 0.002434356603771448, + 0.005586378276348114, + 0.0039036860689520836 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 5, + 1 + ], + "selected_reward": -0.062042009085416794 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, with an almost imperceptible slow push-in, gradually approaches a man in intense agony. He wears a charcoal-gray, crew-neck cotton T-shirt, his collar and temples soaked with sweat, causing his short, dark brown hair to stick to his forehead. He raises his hands in despair, pressing the heels of his palms against his temples, his fingers splayed out, digging deep into his hair, as if attempting to physically stifle the invisible torment raging in his mind. His eyes are tightly closed, his eyelids trembling slightly from the effort, and his brows are knitted into a painful knot, creating deep vertical lines between them. A single, warm-toned desk light from the right side of the frame casts a stark contrast of light and dark on his face, highlighting the tense jawline and the ridges of muscle in his cheeks from clenching his teeth. The background is a dim and blurry interior, with only the vague outline of the out-of-focus bookshelf able to be discerned. The entire scene is shrouded in a repressive, tense and extremely private atmosphere, conveying an indescribable physical or psychological pain.", + "seed": 26474, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_8392_seed_26474_dp_10_rank_20_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.047135360166430475, + "vlm_logprob": -0.047135360166430475, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.047135360166430475, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01783933863043785 + ], + "opposite_logprob_tokens": [ + -5.642839431762695 + ], + "logprob_reward": -0.01783933863043785, + "yes_no_margin_reward": 5.625, + "token_credit_reward": -0.01783933863043785, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10150202363729477, + 0.34454432129859924, + 0.030942803248763084, + 0.03259958326816559, + 0.07825914025306702, + 0.16722159087657928, + 0.16767898201942444, + 0.03964943066239357, + 0.021210305392742157, + 0.016391785815358162 + ], + "credit_map": [ + 0.013413961976766586, + 0.015012283809483051, + 0.017112350091338158, + 0.01472166832536459, + 0.013405108824372292, + 0.006940558087080717, + 0.01400829665362835, + 0.006887800060212612, + 0.10066499561071396, + 0.019976705312728882, + 0.13052481412887573, + 0.020993346348404884, + 0.0157904215157032, + 0.008850586600601673, + 0.02990448847413063, + 0.017838969826698303, + 0.004466407001018524, + 0.004432136192917824, + 0.004239819478243589, + 0.004057666752487421, + 0.003637253772467375, + 0.003357752924785018, + 0.004041554406285286, + 0.0027102152816951275, + 0.0051518091931939125, + 0.0049985600635409355, + 0.003182316431775689, + 0.00445189606398344, + 0.004911152180284262, + 0.002413522219285369, + 0.005552374757826328, + 0.0019379541045054793, + 0.0107722831889987, + 0.00962035171687603, + 0.012983533553779125, + 0.008031862787902355, + 0.008853381499648094, + 0.006736848037689924, + 0.014489096589386463, + 0.0067717828787863255, + 0.040528129786252975, + 0.0075589134357869625, + 0.06069580093026161, + 0.014145308174192905, + 0.01019713282585144, + 0.007709723897278309, + 0.014355680905282497, + 0.012030908837914467, + 0.035603832453489304, + 0.0112013453617692, + 0.05547982081770897, + 0.01595357619225979, + 0.014270229265093803, + 0.008300520479679108, + 0.01708226464688778, + 0.009787390939891338, + 0.006814007181674242, + 0.004412120208144188, + 0.0035752514377236366, + 0.004800126887857914, + 0.0063277133740484715, + 0.002537291496992111, + 0.009081835858523846, + 0.0021010818891227245, + 0.003087823512032628, + 0.0029086964204907417, + 0.002663371618837118, + 0.0019578037317842245, + 0.004652684088796377, + 0.001584265730343759, + 0.00310290209017694, + 0.0012527592480182648, + 0.002483401447534561, + 0.002108196960762143, + 0.0014811246655881405, + 0.0017192922532558441, + 0.002215231768786907, + 0.0015535021666437387, + 0.0038040729705244303, + 0.0010269630001857877 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.01783933863043785 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.062110576778650284 + ], + "opposite_logprob_tokens": [ + -2.812110662460327 + ], + "logprob_reward": -0.062110576778650284, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.062110576778650284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0990319475531578, + 0.1440799981355667, + 0.016006048768758774, + 0.012167868204414845, + 0.06907996535301208, + 0.23946455121040344, + 0.32537558674812317, + 0.035372473299503326, + 0.037642717361450195, + 0.021778898313641548 + ], + "credit_map": [ + 0.012107928283512592, + 0.010007829405367374, + 0.012641122564673424, + 0.019191546365618706, + 0.009927400387823582, + 0.005224976222962141, + 0.02394215203821659, + 0.00598899507895112, + 0.039298903197050095, + 0.009903493337333202, + 0.05085811764001846, + 0.01148228719830513, + 0.00786822009831667, + 0.005349550861865282, + 0.012568924576044083, + 0.006750514265149832, + 0.0021366174332797527, + 0.002868850016966462, + 0.0022624440025538206, + 0.002099253237247467, + 0.002282114466652274, + 0.0007934794411994517, + 0.0023549189791083336, + 0.0012083722976967692, + 0.001652865088544786, + 0.0018596055451780558, + 0.0011717831948772073, + 0.0018536433344706893, + 0.0014690496027469635, + 0.0011587539920583367, + 0.0021527698263525963, + 0.000849397387355566, + 0.009615899994969368, + 0.007724476978182793, + 0.009872235357761383, + 0.0077322558499872684, + 0.00819291453808546, + 0.006225425284355879, + 0.013946916908025742, + 0.005769833922386169, + 0.06799719482660294, + 0.009619559161365032, + 0.09299859404563904, + 0.018409298732876778, + 0.009057402610778809, + 0.008648836985230446, + 0.014315766282379627, + 0.01841788925230503, + 0.05994677543640137, + 0.019941914826631546, + 0.09133323282003403, + 0.0663704127073288, + 0.020982054993510246, + 0.015655042603611946, + 0.02939504384994507, + 0.021751083433628082, + 0.00477344635874033, + 0.006937570404261351, + 0.004750181920826435, + 0.0040308162569999695, + 0.004464745055884123, + 0.004114879760891199, + 0.0040940395556390285, + 0.0022067963145673275, + 0.005452541168779135, + 0.006398163270205259, + 0.0048581273294985294, + 0.0040790848433971405, + 0.005992549005895853, + 0.0026277611032128334, + 0.005882901139557362, + 0.0023515913635492325, + 0.00356482551433146, + 0.0023965071886777878, + 0.0020265274215489626, + 0.002628485905006528, + 0.003049868158996105, + 0.0018772257026284933, + 0.004946432542055845, + 0.0012890270445495844 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 1 + ], + "selected_reward": -0.062110576778650284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006743056699633598 + ], + "opposite_logprob_tokens": [ + -5.00674295425415 + ], + "logprob_reward": -0.006743056699633598, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006743056699633598, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11975342035293579, + 0.18195396661758423, + 0.03885256499052048, + 0.025800757110118866, + 0.1613548994064331, + 0.12911401689052582, + 0.10669159889221191, + 0.1696031540632248, + 0.0315374955534935, + 0.035338129848241806 + ], + "credit_map": [ + 0.01237990241497755, + 0.018874315544962883, + 0.0193860474973917, + 0.017654960975050926, + 0.015261518768966198, + 0.007013918831944466, + 0.018826445564627647, + 0.01035631075501442, + 0.048728737980127335, + 0.010061603970825672, + 0.06369150429964066, + 0.012037504464387894, + 0.013497469946742058, + 0.007614274974912405, + 0.016779696568846703, + 0.009543190710246563, + 0.004670321475714445, + 0.006672346964478493, + 0.004582019057124853, + 0.006152858026325703, + 0.005097521003335714, + 0.0032781062182039022, + 0.005502195563167334, + 0.002897196216508746, + 0.0033967436756938696, + 0.004854152910411358, + 0.002761784940958023, + 0.0033072535879909992, + 0.0036245277151465416, + 0.0014405339024960995, + 0.004772310610860586, + 0.0016434505814686418, + 0.022530287504196167, + 0.013944548554718494, + 0.022383932024240494, + 0.019286856055259705, + 0.019237130880355835, + 0.012809334322810173, + 0.03673749044537544, + 0.014425325207412243, + 0.016081852838397026, + 0.02713930606842041, + 0.018363866955041885, + 0.014668978750705719, + 0.020317861810326576, + 0.007419946603477001, + 0.013379471376538277, + 0.011742747388780117, + 0.023634787648916245, + 0.006859845481812954, + 0.037667855620384216, + 0.009541255421936512, + 0.006921884138137102, + 0.006892919074743986, + 0.008966295048594475, + 0.006206766236573458, + 0.022927088662981987, + 0.03207620978355408, + 0.023375676944851875, + 0.015336559154093266, + 0.022694680839776993, + 0.01978762075304985, + 0.022715453058481216, + 0.010689866729080677, + 0.0046038152649998665, + 0.005369312595576048, + 0.003841015277430415, + 0.0026198928244411945, + 0.004116509575396776, + 0.00249545369297266, + 0.006693226285278797, + 0.0017982751596719027, + 0.005901910364627838, + 0.004474049899727106, + 0.0026286481879651546, + 0.004700021352618933, + 0.0036652155686169863, + 0.003483335953205824, + 0.008677025325596333, + 0.0018079255241900682 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 7, + 4 + ], + "selected_reward": -0.006743056699633598 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897479087114334 + ], + "opposite_logprob_tokens": [ + -2.578974723815918 + ], + "logprob_reward": -0.07897479087114334, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897479087114334, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11692459136247635, + 0.5801903605461121, + 0.013739294372498989, + 0.0076746693812310696, + 0.06331253051757812, + 0.017790086567401886, + 0.1390526294708252, + 0.017604779452085495, + 0.0234660804271698, + 0.02024492435157299 + ], + "credit_map": [ + 0.011768124997615814, + 0.013162268325686455, + 0.021485179662704468, + 0.018886219710111618, + 0.014136197976768017, + 0.009934998117387295, + 0.017962487414479256, + 0.009589132852852345, + 0.16021038591861725, + 0.034419137984514236, + 0.19813022017478943, + 0.03552792966365814, + 0.04462118446826935, + 0.023514725267887115, + 0.051227811723947525, + 0.0325389988720417, + 0.0017323247157037258, + 0.0021767897997051477, + 0.0016714544035494328, + 0.0020370737183839083, + 0.0016843652119860053, + 0.0014323329087346792, + 0.0019597525242716074, + 0.0010452030692249537, + 0.0011377554619684815, + 0.0012165206717327237, + 0.0009115682332776487, + 0.0007916100439615548, + 0.0013659653486683965, + 0.0007277423865161836, + 0.0007536137127317488, + 0.0007698945701122284, + 0.009011441841721535, + 0.00996081717312336, + 0.009638825431466103, + 0.007678644265979528, + 0.0066863009706139565, + 0.005250588059425354, + 0.009190935641527176, + 0.005894986446946859, + 0.0027938203420490026, + 0.002185035962611437, + 0.0038498619105666876, + 0.0020199576392769814, + 0.0019674492068588734, + 0.0014196770498529077, + 0.0021066570188850164, + 0.0014476299984380603, + 0.03369688242673874, + 0.008107165805995464, + 0.053757283836603165, + 0.010321099311113358, + 0.00801644753664732, + 0.00723773380741477, + 0.009041802026331425, + 0.008874229155480862, + 0.002778170397505164, + 0.0028208952862769365, + 0.002243643393740058, + 0.0023043728433549404, + 0.002490903250873089, + 0.002175360918045044, + 0.0016147196292877197, + 0.0011767162941396236, + 0.0033923746086657047, + 0.005329855252057314, + 0.0027035751845687628, + 0.002762122079730034, + 0.00340284570120275, + 0.0015653730370104313, + 0.0030436404049396515, + 0.0012662982335314155, + 0.0032165437005460262, + 0.0024258363991975784, + 0.0018990968819707632, + 0.0016166489804163575, + 0.0030477691907435656, + 0.0013175521744415164, + 0.005591727793216705, + 0.001129753072746098 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 0 + ], + "selected_reward": -0.07897479087114334 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07000903785228729 + ], + "opposite_logprob_tokens": [ + -2.6950089931488037 + ], + "logprob_reward": -0.07000903785228729, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07000903785228729, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13595709204673767, + 0.04402028024196625, + 0.044018615037202835, + 0.03253568336367607, + 0.050933029502630234, + 0.2758493721485138, + 0.3504808843135834, + 0.029065420851111412, + 0.019697049632668495, + 0.01744253933429718 + ], + "credit_map": [ + 0.016461864113807678, + 0.01632452756166458, + 0.029393669217824936, + 0.020464029163122177, + 0.016908956691622734, + 0.0082350242882967, + 0.014672243036329746, + 0.013496779836714268, + 0.00613451236858964, + 0.006352693308144808, + 0.004657970741391182, + 0.005609601736068726, + 0.006209108512848616, + 0.003724742215126753, + 0.0064486293122172356, + 0.004883021581918001, + 0.006583760026842356, + 0.007796651218086481, + 0.004573958460241556, + 0.006719653028994799, + 0.005310703068971634, + 0.003796417498961091, + 0.0066665164195001125, + 0.002570957178249955, + 0.004827925469726324, + 0.005027223378419876, + 0.0028455685824155807, + 0.0050419713370501995, + 0.004295899532735348, + 0.0025781842414289713, + 0.00602504750713706, + 0.0018938627326861024, + 0.006978573277592659, + 0.005446093622595072, + 0.004566705785691738, + 0.005343092139810324, + 0.0067682163789868355, + 0.008033115416765213, + 0.009143801406025887, + 0.004653431009501219, + 0.07572826743125916, + 0.013540873304009438, + 0.10690779983997345, + 0.022291917353868484, + 0.009570736438035965, + 0.009872511960566044, + 0.01638290099799633, + 0.02155437134206295, + 0.08441583067178726, + 0.021147532388567924, + 0.1347724348306656, + 0.02675192430615425, + 0.02067406103014946, + 0.017997149378061295, + 0.02277768775820732, + 0.021944275125861168, + 0.004464026540517807, + 0.005836945492774248, + 0.004798687994480133, + 0.002373358001932502, + 0.003476192941889167, + 0.002984813880175352, + 0.0030525748152285814, + 0.0020788235124200583, + 0.002983483951538801, + 0.0021743313409388065, + 0.0023765850346535444, + 0.0018504861509427428, + 0.00390235404483974, + 0.0012350192992016673, + 0.0036682975478470325, + 0.0015064921462908387, + 0.0025564238894730806, + 0.0017401985824108124, + 0.0017877272330224514, + 0.001886827521957457, + 0.002787379315122962, + 0.0016754346434026957, + 0.0034685900900512934, + 0.0015399584081023932 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 0 + ], + "selected_reward": -0.07000903785228729 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, with an almost imperceptible slow push-in, gradually approaches a man in intense agony. He wears a charcoal-gray, crew-neck cotton T-shirt, his collar and temples soaked with sweat, causing his short, dark brown hair to stick to his forehead. He raises his hands in despair, pressing the heels of his palms against his temples, his fingers splayed out, digging deep into his hair, as if attempting to physically stifle the invisible torment raging in his mind. His eyes are tightly closed, his eyelids trembling slightly from the effort, and his brows are knitted into a painful knot, creating deep vertical lines between them. A single, warm-toned desk light from the right side of the frame casts a stark contrast of light and dark on his face, highlighting the tense jawline and the ridges of muscle in his cheeks from clenching his teeth. The background is a dim and blurry interior, with only the vague outline of the out-of-focus bookshelf able to be discerned. The entire scene is shrouded in a repressive, tense and extremely private atmosphere, conveying an indescribable physical or psychological pain.", + "seed": 26475, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_8392_seed_26475_dp_11_rank_22_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.047768013644963504, + "vlm_logprob": -0.047768013644963504, + "vlm_yes_no_margin": 3.8, + "vlm_token_credit": -0.047768013644963504, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01450165081769228 + ], + "opposite_logprob_tokens": [ + -6.264501571655273 + ], + "logprob_reward": -0.01450165081769228, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.01450165081769228, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11056405305862427, + 0.056660447269678116, + 0.12838751077651978, + 0.17262868583202362, + 0.0221463181078434, + 0.11715740710496902, + 0.2748603820800781, + 0.0380854494869709, + 0.04285764694213867, + 0.03665217384696007 + ], + "credit_map": [ + 0.018218329176306725, + 0.010754836723208427, + 0.015259038656949997, + 0.013498096726834774, + 0.0141000896692276, + 0.008113091811537743, + 0.017904553562402725, + 0.012716027908027172, + 0.009276178665459156, + 0.006100011989474297, + 0.010017591528594494, + 0.007926045916974545, + 0.007259990554302931, + 0.0038273704703897238, + 0.007800370920449495, + 0.004452884197235107, + 0.022260554134845734, + 0.010218617506325245, + 0.012750781141221523, + 0.008729550987482071, + 0.02627699263393879, + 0.011644872836768627, + 0.013755048625171185, + 0.0227510929107666, + 0.02991531416773796, + 0.0214170403778553, + 0.012557117268443108, + 0.012058356776833534, + 0.04225000739097595, + 0.013380002230405807, + 0.020674580708146095, + 0.020376253873109818, + 0.003908692393451929, + 0.0034640226513147354, + 0.0022889713291078806, + 0.0029724163468927145, + 0.0027787494473159313, + 0.0022365746553987265, + 0.0031451412942260504, + 0.0013517492916435003, + 0.023481562733650208, + 0.01152342464774847, + 0.006321204360574484, + 0.010301897302269936, + 0.027231520041823387, + 0.011855055578052998, + 0.020259831100702286, + 0.006182910408824682, + 0.04337893798947334, + 0.023979485034942627, + 0.01723303645849228, + 0.05705675482749939, + 0.04059983789920807, + 0.03184721991419792, + 0.03787621110677719, + 0.02288891188800335, + 0.006403681356459856, + 0.004067919682711363, + 0.0024425238370895386, + 0.003549079643562436, + 0.005235415883362293, + 0.008741962723433971, + 0.0049062566831707954, + 0.0027386092115193605, + 0.004724089056253433, + 0.005014842841774225, + 0.002035868354141712, + 0.0037382645532488823, + 0.00726029509678483, + 0.00813568476587534, + 0.008751683868467808, + 0.0031969179399311543, + 0.0037436296697705984, + 0.004227757453918457, + 0.0016047832323238254, + 0.004381027538329363, + 0.0038309625815600157, + 0.009840698912739754, + 0.006580620538443327, + 0.0024426961317658424 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 3, + 2 + ], + "selected_reward": -0.01450165081769228 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0890740305185318 + ], + "opposite_logprob_tokens": [ + -2.46407413482666 + ], + "logprob_reward": -0.0890740305185318, + "yes_no_margin_reward": 2.375, + "token_credit_reward": -0.0890740305185318, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.07791624963283539, + 0.055393658578395844, + 0.09974142163991928, + 0.0374375618994236, + 0.19364383816719055, + 0.4204849898815155, + 0.03094490058720112, + 0.018952907994389534, + 0.03421171009540558, + 0.031272709369659424 + ], + "credit_map": [ + 0.013546898029744625, + 0.007414117455482483, + 0.009179155342280865, + 0.012351909652352333, + 0.010512420907616615, + 0.006516632158309221, + 0.008041941560804844, + 0.010353181511163712, + 0.009332200512290001, + 0.0039039684925228357, + 0.004546475131064653, + 0.008026503026485443, + 0.007374839391559362, + 0.004108516499400139, + 0.013074038550257683, + 0.005027121864259243, + 0.017008380964398384, + 0.007678291294723749, + 0.010189391672611237, + 0.00639371620491147, + 0.020863328129053116, + 0.00929146260023117, + 0.011175891384482384, + 0.017140965908765793, + 0.006486381404101849, + 0.003850127337500453, + 0.003517074976116419, + 0.005495963152498007, + 0.006117633543908596, + 0.003501370083540678, + 0.0054208566434681416, + 0.003048158250749111, + 0.033158499747514725, + 0.026840651407837868, + 0.017358938232064247, + 0.028167257085442543, + 0.02486070804297924, + 0.02154679037630558, + 0.025432109832763672, + 0.016278890892863274, + 0.08809344470500946, + 0.03744204342365265, + 0.02142379991710186, + 0.03968357667326927, + 0.09329060465097427, + 0.04809419810771942, + 0.05950668826699257, + 0.03295066952705383, + 0.0039880163967609406, + 0.0025817693676799536, + 0.002173457993194461, + 0.007429983001202345, + 0.004050698131322861, + 0.0031940527260303497, + 0.00466295937076211, + 0.002863965928554535, + 0.003506086068227887, + 0.00223844638094306, + 0.0009949598461389542, + 0.002129962434992194, + 0.002386635635048151, + 0.0033959972206503153, + 0.00263240747153759, + 0.0016684142174199224, + 0.0018547314684838057, + 0.001670493045821786, + 0.0008875990752130747, + 0.0015916069969534874, + 0.002035176381468773, + 0.013081667013466358, + 0.011428109370172024, + 0.0016623273259028792, + 0.0021184091456234455, + 0.005256193690001965, + 0.0019169773440808058, + 0.005545454099774361, + 0.0031229909509420395, + 0.005997197236865759, + 0.00497572124004364, + 0.0023397670593112707 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 4, + 2 + ], + "selected_reward": -0.0890740305185318 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.007639714982360601 + ], + "opposite_logprob_tokens": [ + -4.8826398849487305 + ], + "logprob_reward": -0.007639714982360601, + "yes_no_margin_reward": 4.875, + "token_credit_reward": -0.007639714982360601, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08685095608234406, + 0.07607200741767883, + 0.0681338980793953, + 0.05585739389061928, + 0.08829911053180695, + 0.1096973717212677, + 0.36890119314193726, + 0.05464709550142288, + 0.06097574159502983, + 0.030565306544303894 + ], + "credit_map": [ + 0.01184159703552723, + 0.007317855954170227, + 0.00847010686993599, + 0.006814595311880112, + 0.010300610214471817, + 0.010349217802286148, + 0.0123139638453722, + 0.01944301277399063, + 0.011672718450427055, + 0.007214928511530161, + 0.014399080537259579, + 0.009990951977670193, + 0.008634798228740692, + 0.006201427895575762, + 0.010518812574446201, + 0.0074392883107066154, + 0.010749706998467445, + 0.004837661050260067, + 0.008638010360300541, + 0.005567208863794804, + 0.014571614563465118, + 0.0057344576343894005, + 0.00927292462438345, + 0.008762315846979618, + 0.009227225556969643, + 0.0069299801252782345, + 0.004148735199123621, + 0.003722691209986806, + 0.013907549902796745, + 0.005153458099812269, + 0.007166917901486158, + 0.005600834731012583, + 0.015270091593265533, + 0.012695660814642906, + 0.004778728820383549, + 0.009677733294665813, + 0.02288581244647503, + 0.008653617464005947, + 0.009468642994761467, + 0.004868817515671253, + 0.018526162952184677, + 0.012313433922827244, + 0.007168812211602926, + 0.01601298525929451, + 0.01951522007584572, + 0.01361776888370514, + 0.011549036018550396, + 0.010993951931595802, + 0.058621086180210114, + 0.03217443823814392, + 0.022848650813102722, + 0.07859497517347336, + 0.05245235562324524, + 0.043527279049158096, + 0.05046166479587555, + 0.03022071160376072, + 0.016202665865421295, + 0.005177940707653761, + 0.002251893514767289, + 0.0030318191275000572, + 0.00473823444917798, + 0.01601317897439003, + 0.004355786833912134, + 0.0028755769599229097, + 0.005639538634568453, + 0.005576526280492544, + 0.001975621795281768, + 0.005045745521783829, + 0.004608664195984602, + 0.019078493118286133, + 0.0140412338078022, + 0.005009920801967382, + 0.0031183643732219934, + 0.004397669807076454, + 0.0022870872635394335, + 0.0035121850669384003, + 0.003190805437043309, + 0.006012244615703821, + 0.00471071619540453, + 0.0033362330868840218 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 6, + 5, + 4 + ], + "selected_reward": -0.007639714982360601 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.048649441450834274 + ], + "opposite_logprob_tokens": [ + -3.048649549484253 + ], + "logprob_reward": -0.048649441450834274, + "yes_no_margin_reward": 3.0, + "token_credit_reward": -0.048649441450834274, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2427782118320465, + 0.2035750150680542, + 0.16737999022006989, + 0.05082479119300842, + 0.02745019644498825, + 0.06418964266777039, + 0.06367617845535278, + 0.05975435674190521, + 0.08944353461265564, + 0.030928125604987144 + ], + "credit_map": [ + 0.044155292212963104, + 0.027389951050281525, + 0.020418856292963028, + 0.027513382956385612, + 0.03322194889187813, + 0.02850796841084957, + 0.030820004642009735, + 0.03075079619884491, + 0.033055007457733154, + 0.015008799731731415, + 0.017663951963186264, + 0.034689996391534805, + 0.023365307599306107, + 0.0206504687666893, + 0.03900475054979324, + 0.020136723294854164, + 0.02857094630599022, + 0.013266402296721935, + 0.01743045076727867, + 0.010132169350981712, + 0.03664816543459892, + 0.014678161591291428, + 0.018893949687480927, + 0.027759727090597153, + 0.008497621864080429, + 0.00634808512404561, + 0.003627614350989461, + 0.0038487310521304607, + 0.01192170474678278, + 0.004237849730998278, + 0.006627275608479977, + 0.0057159047573804855, + 0.004799490328878164, + 0.003388012293726206, + 0.0019168431172147393, + 0.003697312204167247, + 0.004671431612223387, + 0.003122492227703333, + 0.003854219103232026, + 0.002000390086323023, + 0.013915492221713066, + 0.006350422278046608, + 0.003170958487316966, + 0.004471673164516687, + 0.015037169679999352, + 0.0069289919920265675, + 0.010329162701964378, + 0.00398576445877552, + 0.01073093805462122, + 0.005538535304367542, + 0.003770221024751663, + 0.013985507190227509, + 0.008831995539367199, + 0.00798863172531128, + 0.008723264560103416, + 0.004107076209038496, + 0.01034819707274437, + 0.0069602965377271175, + 0.0036154750268906355, + 0.0042652846314013, + 0.009973671287298203, + 0.008643937297165394, + 0.011692972853779793, + 0.004254522267729044, + 0.010240216739475727, + 0.005628252401947975, + 0.0027524344623088837, + 0.004160341341048479, + 0.006751099601387978, + 0.03392978385090828, + 0.02224288508296013, + 0.003738506929948926, + 0.002279119798913598, + 0.0044310870580375195, + 0.001718267914839089, + 0.004519585520029068, + 0.003776099532842636, + 0.0070747495628893375, + 0.005040271673351526, + 0.0020889420993626118 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.048649441450834274 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07897523045539856 + ], + "opposite_logprob_tokens": [ + -2.578975200653076 + ], + "logprob_reward": -0.07897523045539856, + "yes_no_margin_reward": 2.5, + "token_credit_reward": -0.07897523045539856, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11034359782934189, + 0.07490743696689606, + 0.15260037779808044, + 0.04210883751511574, + 0.058757178485393524, + 0.23923063278198242, + 0.13536958396434784, + 0.064766526222229, + 0.07963763177394867, + 0.04227819666266441 + ], + "credit_map": [ + 0.018262263387441635, + 0.011707485653460026, + 0.017772026360034943, + 0.014446473680436611, + 0.01860172301530838, + 0.00881950557231903, + 0.011705709621310234, + 0.009028407745063305, + 0.010710540227591991, + 0.00933114718645811, + 0.009501037187874317, + 0.01455040741711855, + 0.010828325524926186, + 0.004747606813907623, + 0.009146873839199543, + 0.006091505289077759, + 0.021040964871644974, + 0.011802054941654205, + 0.019419541582465172, + 0.016730831936001778, + 0.032015424221754074, + 0.016430402174592018, + 0.023999864235520363, + 0.011161303147673607, + 0.00720954081043601, + 0.0059349993243813515, + 0.0031310287304222584, + 0.003605624893680215, + 0.009913499467074871, + 0.0031522002536803484, + 0.0049921623431146145, + 0.004169783089309931, + 0.010544081218540668, + 0.007632406894117594, + 0.0034488432575017214, + 0.006550862453877926, + 0.01251656748354435, + 0.005379925947636366, + 0.00766754848882556, + 0.005016939714550972, + 0.05084676295518875, + 0.023112226277589798, + 0.010812375694513321, + 0.016882920637726784, + 0.05670291557908058, + 0.025407658889889717, + 0.0405612476170063, + 0.014904530718922615, + 0.022502297535538673, + 0.011687769554555416, + 0.008227353915572166, + 0.02741658128798008, + 0.018297020345926285, + 0.018514709547162056, + 0.017757032066583633, + 0.010966815054416656, + 0.007644461002200842, + 0.0112098203971982, + 0.003322908189147711, + 0.007075099740177393, + 0.005397473927587271, + 0.014847164042294025, + 0.010135948657989502, + 0.005133654456585646, + 0.024786725640296936, + 0.005998110864311457, + 0.0031842426396906376, + 0.0035103734117001295, + 0.006528645288199186, + 0.009067189879715443, + 0.021795818582177162, + 0.004766526166349649, + 0.002917665056884289, + 0.00430918438360095, + 0.00255928304977715, + 0.003995924256742001, + 0.004443390294909477, + 0.011601371690630913, + 0.007689321879297495, + 0.004762055352330208 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 2, + 6 + ], + "selected_reward": -0.07897523045539856 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "medium close-up shot, shot from a slightly upward angle, with an almost imperceptible slow push-in, gradually approaches a man in intense agony. He wears a charcoal-gray, crew-neck cotton T-shirt, his collar and temples soaked with sweat, causing his short, dark brown hair to stick to his forehead. He raises his hands in despair, pressing the heels of his palms against his temples, his fingers splayed out, digging deep into his hair, as if attempting to physically stifle the invisible torment raging in his mind. His eyes are tightly closed, his eyelids trembling slightly from the effort, and his brows are knitted into a painful knot, creating deep vertical lines between them. A single, warm-toned desk light from the right side of the frame casts a stark contrast of light and dark on his face, highlighting the tense jawline and the ridges of muscle in his cheeks from clenching his teeth. The background is a dim and blurry interior, with only the vague outline of the out-of-focus bookshelf able to be discerned. The entire scene is shrouded in a repressive, tense and extremely private atmosphere, conveying an indescribable physical or psychological pain.", + "seed": 26476, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_5/index_8392_seed_26476_dp_11_rank_22_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 5, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.048196995817124846, + "vlm_logprob": -0.048196995817124846, + "vlm_yes_no_margin": 3.5, + "vlm_token_credit": -0.048196995817124846, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01575709879398346 + ], + "opposite_logprob_tokens": [ + -5.140757083892822 + ], + "logprob_reward": -0.01575709879398346, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.01575709879398346, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.13780800998210907, + 0.18025276064872742, + 0.17202717065811157, + 0.027699701488018036, + 0.14445613324642181, + 0.02598331682384014, + 0.0751010999083519, + 0.06345158815383911, + 0.1091603934764862, + 0.0640597939491272 + ], + "credit_map": [ + 0.019450593739748, + 0.01691550202667713, + 0.012544996105134487, + 0.027293752878904343, + 0.018829725682735443, + 0.0053525823168456554, + 0.019821589812636375, + 0.01759926788508892, + 0.06971772015094757, + 0.00689088786020875, + 0.0062294164672493935, + 0.00905161164700985, + 0.012349489144980907, + 0.0140296695753932, + 0.03329344838857651, + 0.0286905225366354, + 0.030501771718263626, + 0.006501828320324421, + 0.007203291170299053, + 0.013037075288593769, + 0.009438315406441689, + 0.02208135835826397, + 0.033551137894392014, + 0.04971238970756531, + 0.006987342145293951, + 0.00200777524150908, + 0.0019457102753221989, + 0.002842132467776537, + 0.0029214743990451097, + 0.0023100110702216625, + 0.00502376351505518, + 0.0036614921409636736, + 0.017322860658168793, + 0.019039487466216087, + 0.013732928782701492, + 0.026806125417351723, + 0.01779385656118393, + 0.020796721801161766, + 0.014274923130869865, + 0.014689219184219837, + 0.005605209618806839, + 0.002030727220699191, + 0.0031583071686327457, + 0.003764988388866186, + 0.0035144395660609007, + 0.0024092881940305233, + 0.004038404673337936, + 0.0014619510620832443, + 0.009950492531061172, + 0.013074934482574463, + 0.00738182058557868, + 0.008450960740447044, + 0.011650832369923592, + 0.006407580338418484, + 0.00862103421241045, + 0.009563444182276726, + 0.005360934417694807, + 0.009514333680272102, + 0.006141837686300278, + 0.008146368898451328, + 0.010800258256494999, + 0.006882100831717253, + 0.009364436380565166, + 0.007241321727633476, + 0.00841524824500084, + 0.01556953601539135, + 0.01145140454173088, + 0.02457706816494465, + 0.014335866086184978, + 0.011986340396106243, + 0.01413855329155922, + 0.008686380460858345, + 0.007933863438665867, + 0.006419822108000517, + 0.006149467080831528, + 0.005627521779388189, + 0.008188792504370213, + 0.012660008855164051, + 0.008008105680346489, + 0.00907220970839262 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.01575709879398346 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.1129983589053154 + ], + "opposite_logprob_tokens": [ + -2.2379982471466064 + ], + "logprob_reward": -0.1129983589053154, + "yes_no_margin_reward": 2.125, + "token_credit_reward": -0.1129983589053154, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14732056856155396, + 0.1451585292816162, + 0.10567279160022736, + 0.029703309759497643, + 0.19399484992027283, + 0.13808274269104004, + 0.05689661577343941, + 0.11177977174520493, + 0.03635429963469505, + 0.03503650799393654 + ], + "credit_map": [ + 0.02175830490887165, + 0.018409373238682747, + 0.013459467329084873, + 0.02689586579799652, + 0.023218685761094093, + 0.006877972278743982, + 0.02230456843972206, + 0.014396331273019314, + 0.04755372554063797, + 0.007286768406629562, + 0.008015913888812065, + 0.013333946466445923, + 0.014004838652908802, + 0.007854988798499107, + 0.034295354038476944, + 0.012813003733754158, + 0.031815703958272934, + 0.004141055513173342, + 0.005044524557888508, + 0.006045286078006029, + 0.006245906464755535, + 0.010841672308743, + 0.02054695598781109, + 0.02099168673157692, + 0.004550631158053875, + 0.003227987326681614, + 0.003039515810087323, + 0.004010124132037163, + 0.003727833740413189, + 0.0037049665115773678, + 0.0035334702115505934, + 0.003908778075128794, + 0.023802412673830986, + 0.015609469264745712, + 0.03269265219569206, + 0.02038942649960518, + 0.019639093428850174, + 0.02623380348086357, + 0.02452683262526989, + 0.031101156026124954, + 0.022426431998610497, + 0.01564382202923298, + 0.015862340107560158, + 0.013626517727971077, + 0.023784950375556946, + 0.012694069184362888, + 0.019713550806045532, + 0.014331053011119366, + 0.003980484325438738, + 0.007522264029830694, + 0.003999250940978527, + 0.006924608256667852, + 0.009514779783785343, + 0.0071259490214288235, + 0.009784769266843796, + 0.008044512942433357, + 0.009338104166090488, + 0.015243636444211006, + 0.014082035981118679, + 0.011893968097865582, + 0.01822049729526043, + 0.014833224937319756, + 0.012625051662325859, + 0.015543254092335701, + 0.0037817226257175207, + 0.005523752421140671, + 0.003478716127574444, + 0.007733138278126717, + 0.0051129041239619255, + 0.0038046310655772686, + 0.004340276587754488, + 0.0025791616644710302, + 0.0048173461109399796, + 0.005638928152620792, + 0.004105185624212027, + 0.004269230179488659, + 0.0055333087220788, + 0.004476696252822876, + 0.003692307509481907, + 0.002503506373614073 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 1 + ], + "selected_reward": -0.1129983589053154 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014209638349711895 + ], + "opposite_logprob_tokens": [ + -4.264209747314453 + ], + "logprob_reward": -0.014209638349711895, + "yes_no_margin_reward": 4.25, + "token_credit_reward": -0.014209638349711895, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08249460160732269, + 0.1866958886384964, + 0.09605929255485535, + 0.08266562968492508, + 0.05924453213810921, + 0.08873552083969116, + 0.04509735107421875, + 0.08178860694169998, + 0.135445237159729, + 0.14177341759204865 + ], + "credit_map": [ + 0.011648111045360565, + 0.008840949274599552, + 0.011329549364745617, + 0.011074223555624485, + 0.013089816085994244, + 0.006725470069795847, + 0.011028519831597805, + 0.008757960982620716, + 0.07718893140554428, + 0.008869949728250504, + 0.005958832334727049, + 0.008952350355684757, + 0.012241605669260025, + 0.011830361559987068, + 0.038581736385822296, + 0.023072117939591408, + 0.02635200507938862, + 0.0031600352376699448, + 0.0028116528410464525, + 0.004496694076806307, + 0.004990759305655956, + 0.007900948636233807, + 0.03750470280647278, + 0.00884249061346054, + 0.015121052972972393, + 0.010411919094622135, + 0.006730415392667055, + 0.010822444222867489, + 0.014170931652188301, + 0.006891664117574692, + 0.012979489751160145, + 0.005537707824259996, + 0.007326291874051094, + 0.00683645810931921, + 0.006846271455287933, + 0.009221365675330162, + 0.008133519440889359, + 0.007910223677754402, + 0.005933424457907677, + 0.007036973722279072, + 0.009531611576676369, + 0.013627670705318451, + 0.010614143684506416, + 0.013055604882538319, + 0.011153705418109894, + 0.010005817748606205, + 0.012147788889706135, + 0.00859916489571333, + 0.0050055175088346004, + 0.0038423205260187387, + 0.00333465076982975, + 0.007068255450576544, + 0.005016710143536329, + 0.00653300154954195, + 0.006085959263145924, + 0.008210932835936546, + 0.00594074372202158, + 0.01116250641644001, + 0.007858707569539547, + 0.010114656761288643, + 0.012832178734242916, + 0.010311649180948734, + 0.010616756975650787, + 0.012951401993632317, + 0.021016178652644157, + 0.017493419349193573, + 0.010549526661634445, + 0.024799497798085213, + 0.017436577007174492, + 0.015530431643128395, + 0.01404917985200882, + 0.014570417813956738, + 0.013894440606236458, + 0.017521090805530548, + 0.014825893566012383, + 0.02124280482530594, + 0.019433803856372833, + 0.02222779206931591, + 0.018546123057603836, + 0.014081466943025589 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 9, + 8 + ], + "selected_reward": -0.014209638349711895 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.05496230721473694 + ], + "opposite_logprob_tokens": [ + -2.929962396621704 + ], + "logprob_reward": -0.05496230721473694, + "yes_no_margin_reward": 2.875, + "token_credit_reward": -0.05496230721473694, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.33675655722618103, + 0.04953892529010773, + 0.20078682899475098, + 0.01813693158328533, + 0.033009909093379974, + 0.01979799196124077, + 0.03868306055665016, + 0.15400484204292297, + 0.043855272233486176, + 0.10542961210012436 + ], + "credit_map": [ + 0.07701648771762848, + 0.025399096310138702, + 0.02169852890074253, + 0.024862978607416153, + 0.062147896736860275, + 0.0388292595744133, + 0.03922828286886215, + 0.04757407307624817, + 0.017599942162632942, + 0.002449341583997011, + 0.0021249607671052217, + 0.0024712334852665663, + 0.003961230628192425, + 0.004034543875604868, + 0.008185815997421741, + 0.0087118586525321, + 0.061790525913238525, + 0.0057106614112854, + 0.005414316430687904, + 0.00865430198609829, + 0.008564368821680546, + 0.01586214266717434, + 0.08292743563652039, + 0.011863097548484802, + 0.003201845334842801, + 0.0018673627637326717, + 0.0014521401608362794, + 0.0018739128718152642, + 0.00231121270917356, + 0.002070304239168763, + 0.003198013175278902, + 0.002162141026929021, + 0.005241738632321358, + 0.003118980210274458, + 0.0034002435859292746, + 0.004154381342232227, + 0.004754174035042524, + 0.00515166437253356, + 0.0037222756072878838, + 0.003466453868895769, + 0.0025591396261006594, + 0.0018450065981596708, + 0.001780867576599121, + 0.0023093719501048326, + 0.003788439556956291, + 0.0017976353410631418, + 0.0038995579816401005, + 0.0018179750768467784, + 0.0026241077575832605, + 0.004658723250031471, + 0.002810853999108076, + 0.004762222059071064, + 0.005656693130731583, + 0.006815542001277208, + 0.004924772772938013, + 0.006430151406675577, + 0.01202904898673296, + 0.02229306288063526, + 0.015832941979169846, + 0.016391923651099205, + 0.02792571671307087, + 0.019332807511091232, + 0.021790295839309692, + 0.018409064039587975, + 0.006979967467486858, + 0.005636322312057018, + 0.0034681204706430435, + 0.007783378474414349, + 0.005280886776745319, + 0.005286323372274637, + 0.004291738383471966, + 0.005128536373376846, + 0.01389333140105009, + 0.0160903949290514, + 0.011513234116137028, + 0.016189029440283775, + 0.017405355349183083, + 0.010394696146249771, + 0.01232914812862873, + 0.007614424452185631 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 7 + ], + "selected_reward": -0.05496230721473694 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the shot a medium close-up of a man?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the man wearing a charcoal-gray, crew-neck cotton T-shirt?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the man raising his hands with palms pressed against his temples and fingers splayed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a warm-toned desk light visible on the right side of the frame casting strong contrast?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Does the camera perform an almost imperceptible slow push-in?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.043057575821876526 + ], + "opposite_logprob_tokens": [ + -3.168057680130005 + ], + "logprob_reward": -0.043057575821876526, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.043057575821876526, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.0441623330116272, + 0.12526769936084747, + 0.053229402750730515, + 0.034244708716869354, + 0.048344507813453674, + 0.04014355689287186, + 0.06124386936426163, + 0.13502360880374908, + 0.09123504906892776, + 0.36710524559020996 + ], + "credit_map": [ + 0.007071380037814379, + 0.00526757538318634, + 0.006201207172125578, + 0.007861982099711895, + 0.00650445744395256, + 0.0022907559759914875, + 0.004641935229301453, + 0.004323036875575781, + 0.049010273069143295, + 0.004544216673821211, + 0.004442176781594753, + 0.006662347353994846, + 0.0074775260873138905, + 0.01141658890992403, + 0.025527890771627426, + 0.016186678782105446, + 0.012944373302161694, + 0.0041123335249722, + 0.0029810459818691015, + 0.0058109695091843605, + 0.006077165715396404, + 0.0069898744113743305, + 0.006969214417040348, + 0.007344422396272421, + 0.006860990077257156, + 0.003425267292186618, + 0.0025391795206815004, + 0.0032352639827877283, + 0.004254916217178106, + 0.003760982071980834, + 0.0037159468047320843, + 0.006452164147049189, + 0.00733273895457387, + 0.003143852576613426, + 0.005979064851999283, + 0.006371126510202885, + 0.007107690442353487, + 0.005208716262131929, + 0.006072117481380701, + 0.007129200734198093, + 0.005327836610376835, + 0.005962925963103771, + 0.003807616885751486, + 0.004387151915580034, + 0.006303480360656977, + 0.005345815792679787, + 0.0057390667498111725, + 0.003269659588113427, + 0.004189400002360344, + 0.005575477611273527, + 0.00430352333933115, + 0.00973350927233696, + 0.008684521540999413, + 0.010134351439774036, + 0.009445780888199806, + 0.00917730014771223, + 0.01066155917942524, + 0.01672237738966942, + 0.017365293577313423, + 0.015733569860458374, + 0.027636030688881874, + 0.016085265204310417, + 0.015194813720881939, + 0.015624690800905228, + 0.013895242474973202, + 0.00987375620752573, + 0.007286872714757919, + 0.01773049868643284, + 0.010915927588939667, + 0.010403464548289776, + 0.009651202708482742, + 0.01147808600217104, + 0.02947499416768551, + 0.03382426127791405, + 0.029809903353452682, + 0.030032293871045113, + 0.06412540376186371, + 0.0850570872426033, + 0.04923012852668762, + 0.045551203191280365 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 1 + ], + "selected_reward": -0.043057575821876526 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_tensor([9439, 9439])_group_6_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_tensor([9439, 9439])_group_6_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..3d1e3efde8d904bfcade1caeed35c6f17a1425d1 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_tensor([9439, 9439])_group_6_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "strikingly dynamic low-angle tracking shot follows a heavy main battle tank as it hurtles forward at breakneck speed across a muddy, war-torn battlefield. The tank's hull, covered in dark olive green and brown camouflage, is covered in scratches and aged stains, its turret pointed steadily forward. Beneath gloomy, low-hanging gray clouds, a cold, diffuse light accentuates the icy texture of the metal and the wet sheen of the ground. Its broad, heavy metal tracks churn frantically, each bite fiercely into the slippery earth, sending chunks of sticky, dark brown mud and broken grass high into the air. These chunks arc through the air before slamming back down to the ground or splattering against the tank's rear, creating a spectacular, chaotic wave of mud. The distant skyline is outlined by sparse, charred trees and a hazy cloud of gunpowder. The entire scene is imbued with a sense of raw power, tense dynamics, and a cold, tense atmosphere of war.", + "seed": 31597, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31597_dp_12_rank_24_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.00998044926673174, + "vlm_logprob": -0.00998044926673174, + "vlm_yes_no_margin": 5.025, + "vlm_token_credit": -0.00998044926673174, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.014617257751524448 + ], + "opposite_logprob_tokens": [ + -4.764617443084717 + ], + "logprob_reward": -0.014617257751524448, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.014617257751524448, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26663222908973694, + 0.10669292509555817, + 0.05287447199225426, + 0.09474305063486099, + 0.1076928898692131, + 0.06594845652580261, + 0.13733647763729095, + 0.04442465305328369, + 0.1085410788655281, + 0.0151138324290514 + ], + "credit_map": [ + 0.042441319674253464, + 0.0187652800232172, + 0.040957972407341, + 0.024892909452319145, + 0.02428922988474369, + 0.057715557515621185, + 0.025714104995131493, + 0.03185584396123886, + 0.021317271515727043, + 0.01556975394487381, + 0.01093493215739727, + 0.00734135415405035, + 0.013898897916078568, + 0.019771724939346313, + 0.007659289985895157, + 0.01019969116896391, + 0.008896943181753159, + 0.006568138021975756, + 0.006941782310605049, + 0.005487472750246525, + 0.0047876969911158085, + 0.00854403991252184, + 0.004239865578711033, + 0.007408531848341227, + 0.021526718512177467, + 0.005619370378553867, + 0.026370342820882797, + 0.009635710157454014, + 0.004361577797681093, + 0.005486585665494204, + 0.010785825550556183, + 0.010956911370158195, + 0.009649891406297684, + 0.006409940309822559, + 0.02393290400505066, + 0.022543201223015785, + 0.007032282184809446, + 0.009592127054929733, + 0.01477829646319151, + 0.013754239305853844, + 0.010447194799780846, + 0.00663314713165164, + 0.009706098586320877, + 0.0074728019535541534, + 0.007318261545151472, + 0.01171308383345604, + 0.006416465155780315, + 0.006241403054445982, + 0.014941837638616562, + 0.0183865949511528, + 0.01985396072268486, + 0.015161966904997826, + 0.022262251004576683, + 0.017816953361034393, + 0.01455740723758936, + 0.014355494640767574, + 0.007386605720967054, + 0.0031043756753206253, + 0.009681162424385548, + 0.004775863606482744, + 0.005124235991388559, + 0.004009561613202095, + 0.004424275830388069, + 0.005918568931519985, + 0.013202071189880371, + 0.013124457560479641, + 0.015763739123940468, + 0.012794770300388336, + 0.0105099156498909, + 0.012587325647473335, + 0.008600834757089615, + 0.02195795625448227, + 0.001124391914345324, + 0.0023265541531145573, + 0.0016381875611841679, + 0.0014961421256884933, + 0.0016858588205650449, + 0.0028124074451625347, + 0.0014293086715042591, + 0.0026009802240878344 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 8 + ], + "selected_reward": -0.014617257751524448 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020609866827726364 + ], + "opposite_logprob_tokens": [ + -3.8956098556518555 + ], + "logprob_reward": -0.020609866827726364, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020609866827726364, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.22871357202529907, + 0.10839680582284927, + 0.06155342236161232, + 0.0990888848900795, + 0.0764741376042366, + 0.07235446572303772, + 0.1624329686164856, + 0.07111243903636932, + 0.06670898199081421, + 0.0531642846763134 + ], + "credit_map": [ + 0.032626207917928696, + 0.02800036408007145, + 0.02395382523536682, + 0.01947140507400036, + 0.04744633287191391, + 0.03569764643907547, + 0.022013358771800995, + 0.019504433497786522, + 0.024121088907122612, + 0.01673143357038498, + 0.014382916502654552, + 0.008807564154267311, + 0.014545144513249397, + 0.014628334902226925, + 0.008347569964826107, + 0.006832754705101252, + 0.01579541712999344, + 0.004874621517956257, + 0.00625435309484601, + 0.00586837949231267, + 0.004954602103680372, + 0.008647267706692219, + 0.005638977512717247, + 0.009519803337752819, + 0.024567220360040665, + 0.005726871080696583, + 0.03354702144861221, + 0.005698670167475939, + 0.004576539155095816, + 0.004857602063566446, + 0.009519299492239952, + 0.010595656000077724, + 0.009250416420400143, + 0.007766286376863718, + 0.015136036090552807, + 0.01627911627292633, + 0.00470423698425293, + 0.0057057589292526245, + 0.008672628551721573, + 0.008959655649960041, + 0.008258148096501827, + 0.008503610268235207, + 0.011421114206314087, + 0.008042292669415474, + 0.006625146139413118, + 0.01272533368319273, + 0.00755528686568141, + 0.009223535656929016, + 0.01723865047097206, + 0.018446743488311768, + 0.020231785252690315, + 0.023393064737319946, + 0.014188515953719616, + 0.022579416632652283, + 0.02845807559788227, + 0.017896704375743866, + 0.005714362487196922, + 0.006404426414519548, + 0.021481813862919807, + 0.007322423625737429, + 0.007390606217086315, + 0.00659391051158309, + 0.0051433248445391655, + 0.011061572469770908, + 0.0065299151465296745, + 0.007887395098805428, + 0.009243916720151901, + 0.011125839315354824, + 0.007202566601336002, + 0.007340366020798683, + 0.006082079838961363, + 0.01129690557718277, + 0.003867761930450797, + 0.007395860273391008, + 0.004772737622261047, + 0.004492734558880329, + 0.007541321683675051, + 0.011764859780669212, + 0.004633649718016386, + 0.008695361204445362 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 6, + 1 + ], + "selected_reward": -0.020609866827726364 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004098587669432163 + ], + "opposite_logprob_tokens": [ + -5.504098415374756 + ], + "logprob_reward": -0.004098587669432163, + "yes_no_margin_reward": 5.5, + "token_credit_reward": -0.004098587669432163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2671687602996826, + 0.06292817741632462, + 0.10160273313522339, + 0.05706459656357765, + 0.17474433779716492, + 0.08537542819976807, + 0.10717295855283737, + 0.04265312850475311, + 0.061303310096263885, + 0.03998664766550064 + ], + "credit_map": [ + 0.03789343312382698, + 0.021707577630877495, + 0.028295256197452545, + 0.021434271708130836, + 0.03589147701859474, + 0.05772264674305916, + 0.024977341294288635, + 0.03924676030874252, + 0.010209844447672367, + 0.006113868206739426, + 0.010096286423504353, + 0.005101935472339392, + 0.009441038593649864, + 0.008710243739187717, + 0.006414877716451883, + 0.006840078625828028, + 0.020810553804039955, + 0.015697486698627472, + 0.01050542201846838, + 0.011839596554636955, + 0.009332823567092419, + 0.013275259174406528, + 0.009545340202748775, + 0.010596241801977158, + 0.007937331683933735, + 0.008274104446172714, + 0.006696311756968498, + 0.007755435071885586, + 0.00525533827021718, + 0.0073566678911447525, + 0.007041259668767452, + 0.006748137064278126, + 0.01735064946115017, + 0.00884659867733717, + 0.030413025990128517, + 0.05735284090042114, + 0.007361317053437233, + 0.008536345325410366, + 0.022534482181072235, + 0.022349057719111443, + 0.008721618913114071, + 0.009300739504396915, + 0.01541021279990673, + 0.013123894110321999, + 0.007604377809911966, + 0.008689629845321178, + 0.011158935725688934, + 0.01136600598692894, + 0.011410717852413654, + 0.015217947773635387, + 0.012775377370417118, + 0.010131528601050377, + 0.01944814994931221, + 0.010900881141424179, + 0.013246611692011356, + 0.014041724614799023, + 0.008199907839298248, + 0.0046240477822721004, + 0.003482284490019083, + 0.003982423804700375, + 0.009716078639030457, + 0.00412783632054925, + 0.00396084226667881, + 0.004559703636914492, + 0.007114277221262455, + 0.006131683476269245, + 0.009437404572963715, + 0.006350225768983364, + 0.00671696150675416, + 0.007012030575424433, + 0.0063069346360862255, + 0.012233790010213852, + 0.0038329316303133965, + 0.007781854830682278, + 0.0045284549705684185, + 0.003914552740752697, + 0.004657848738133907, + 0.00574870640411973, + 0.0039012671913951635, + 0.005621025804430246 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 6 + ], + "selected_reward": -0.004098587669432163 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001939084380865097 + ], + "opposite_logprob_tokens": [ + -6.251939296722412 + ], + "logprob_reward": -0.001939084380865097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001939084380865097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2376926690340042, + 0.09441935271024704, + 0.057628318667411804, + 0.024878107011318207, + 0.04004235565662384, + 0.02828783169388771, + 0.030524861067533493, + 0.052516162395477295, + 0.3984864056110382, + 0.0355239063501358 + ], + "credit_map": [ + 0.03659139573574066, + 0.018734171986579895, + 0.031034337356686592, + 0.020471984520554543, + 0.024211589246988297, + 0.058688078075647354, + 0.023186758160591125, + 0.02477436140179634, + 0.019746653735637665, + 0.008795016445219517, + 0.014585909433662891, + 0.007994452491402626, + 0.01188310794532299, + 0.013937831856310368, + 0.007498601917177439, + 0.009977773763239384, + 0.013102415017783642, + 0.005784429609775543, + 0.006642154883593321, + 0.007494389079511166, + 0.004776868503540754, + 0.009035812690854073, + 0.004705724772065878, + 0.006086522713303566, + 0.0036465758457779884, + 0.002880413318052888, + 0.004106495529413223, + 0.0028905197978019714, + 0.0025938930921256542, + 0.00367415277287364, + 0.0026675951667129993, + 0.002418460790067911, + 0.005492058582603931, + 0.0026642705779522657, + 0.0074309734627604485, + 0.010586007498204708, + 0.002245268551632762, + 0.002922785934060812, + 0.004316254518926144, + 0.004384735599160194, + 0.003041820600628853, + 0.00276546529494226, + 0.004960131831467152, + 0.004239720292389393, + 0.002687491476535797, + 0.0028757958207279444, + 0.003992374986410141, + 0.0037250304594635963, + 0.0031013013795018196, + 0.004662463441491127, + 0.005309860222041607, + 0.0037282612174749374, + 0.003932504914700985, + 0.002647379646077752, + 0.004275157582014799, + 0.0028679342940449715, + 0.005531041417270899, + 0.004966547712683678, + 0.010635409504175186, + 0.004430529661476612, + 0.0070769526064395905, + 0.006050383672118187, + 0.0038972606416791677, + 0.009928039275109768, + 0.047091107815504074, + 0.04865304380655289, + 0.05478958785533905, + 0.038635097444057465, + 0.04148949682712555, + 0.037991590797901154, + 0.03779233619570732, + 0.0920441523194313, + 0.002699346747249365, + 0.00410803547129035, + 0.002840818138793111, + 0.0019047373207286, + 0.006870316807180643, + 0.008993705734610558, + 0.003307319711893797, + 0.004799624439328909 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 1 + ], + "selected_reward": -0.001939084380865097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637449704110622 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637449704110622, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637449704110622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1198834478855133, + 0.08388388156890869, + 0.10405883193016052, + 0.14472484588623047, + 0.058046940714120865, + 0.16346852481365204, + 0.0600661039352417, + 0.07165790349245071, + 0.15582503378391266, + 0.03838452696800232 + ], + "credit_map": [ + 0.019379235804080963, + 0.007873701862990856, + 0.018327414989471436, + 0.010967331938445568, + 0.011588601395487785, + 0.02423129789531231, + 0.011604750528931618, + 0.015911102294921875, + 0.02039661630988121, + 0.007655926514416933, + 0.00619809003546834, + 0.007032557390630245, + 0.008536070585250854, + 0.02084670402109623, + 0.006963251158595085, + 0.006254657171666622, + 0.015596803277730942, + 0.011763589456677437, + 0.011473039165139198, + 0.013958443887531757, + 0.011340281926095486, + 0.012089412659406662, + 0.01032558549195528, + 0.017511676996946335, + 0.02662518247961998, + 0.014642583206295967, + 0.03480903431773186, + 0.014957284554839134, + 0.012209867127239704, + 0.013530120253562927, + 0.01252307090908289, + 0.01542770303785801, + 0.0068421256728470325, + 0.007225129287689924, + 0.008374884724617004, + 0.009588220156729221, + 0.0046003893949091434, + 0.00789017416536808, + 0.007177737075835466, + 0.006348276045173407, + 0.022151803597807884, + 0.015636911615729332, + 0.02822929248213768, + 0.021366246044635773, + 0.012988022528588772, + 0.023132633417844772, + 0.020178688690066338, + 0.019784921780228615, + 0.006581449881196022, + 0.008339774794876575, + 0.0056403689086437225, + 0.004398527555167675, + 0.012630716897547245, + 0.0059974705800414085, + 0.008055971935391426, + 0.008421818725764751, + 0.009930476546287537, + 0.006364346947520971, + 0.015042120590806007, + 0.006009163800626993, + 0.013971347361803055, + 0.007071330677717924, + 0.005366317927837372, + 0.00790279172360897, + 0.018842434510588646, + 0.01841016300022602, + 0.021806467324495316, + 0.017446912825107574, + 0.01517288014292717, + 0.016799217090010643, + 0.01332677248865366, + 0.03402017429471016, + 0.005068165250122547, + 0.004695812240242958, + 0.004343710374087095, + 0.004407424945384264, + 0.005178854335099459, + 0.006236434914171696, + 0.002928651636466384, + 0.0055254725739359856 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 3 + ], + "selected_reward": -0.008637449704110622 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic low-angle tracking shot follows a heavy main battle tank as it hurtles forward at breakneck speed across a muddy, war-torn battlefield. The tank's hull, covered in dark olive green and brown camouflage, is covered in scratches and aged stains, its turret pointed steadily forward. Beneath gloomy, low-hanging gray clouds, a cold, diffuse light accentuates the icy texture of the metal and the wet sheen of the ground. Its broad, heavy metal tracks churn frantically, each bite fiercely into the slippery earth, sending chunks of sticky, dark brown mud and broken grass high into the air. These chunks arc through the air before slamming back down to the ground or splattering against the tank's rear, creating a spectacular, chaotic wave of mud. The distant skyline is outlined by sparse, charred trees and a hazy cloud of gunpowder. The entire scene is imbued with a sense of raw power, tense dynamics, and a cold, tense atmosphere of war.", + "seed": 31598, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31598_dp_12_rank_24_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.1719813806936145, + "vlm_logprob": -0.1719813806936145, + "vlm_yes_no_margin": 2.725000023841858, + "vlm_token_credit": -0.1719813806936145, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6449062824249268 + ], + "opposite_logprob_tokens": [ + -0.7699062824249268 + ], + "logprob_reward": -0.6449062824249268, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6449062824249268, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1185956671833992, + 0.1518457531929016, + 0.10800427943468094, + 0.07989276200532913, + 0.05228033289313316, + 0.2910732924938202, + 0.06934547424316406, + 0.04124348610639572, + 0.04674644023180008, + 0.04097247123718262 + ], + "credit_map": [ + 0.01897859014570713, + 0.009347127750515938, + 0.017512457445263863, + 0.011056515388190746, + 0.015022610314190388, + 0.013102544471621513, + 0.00881604291498661, + 0.024759769439697266, + 0.019908225163817406, + 0.012931999750435352, + 0.02903427742421627, + 0.025064293295145035, + 0.014878908172249794, + 0.01660507172346115, + 0.01086919941008091, + 0.02255377732217312, + 0.012018337845802307, + 0.011324371211230755, + 0.014079174026846886, + 0.01566760428249836, + 0.016010509803891182, + 0.01237634290009737, + 0.011325135827064514, + 0.01520280446857214, + 0.012112000025808811, + 0.007122456096112728, + 0.016355885192751884, + 0.01201910711824894, + 0.0076272739097476006, + 0.008060550317168236, + 0.008331633172929287, + 0.008263859897851944, + 0.009539221413433552, + 0.005587867461144924, + 0.007338281720876694, + 0.005534529685974121, + 0.006750890985131264, + 0.005161751993000507, + 0.006223763804882765, + 0.006144025828689337, + 0.0625564232468605, + 0.028853530064225197, + 0.03164779394865036, + 0.03615009784698486, + 0.039622917771339417, + 0.03981879726052284, + 0.02357316203415394, + 0.028850579634308815, + 0.010362310335040092, + 0.006120292469859123, + 0.00902266800403595, + 0.008266588672995567, + 0.008622055873274803, + 0.007167537696659565, + 0.010314634069800377, + 0.009469381533563137, + 0.00726444972679019, + 0.004749996121972799, + 0.007061962969601154, + 0.0051345364190638065, + 0.002875466365367174, + 0.005387300159782171, + 0.0044890702702105045, + 0.004280704539269209, + 0.00521785719320178, + 0.004979596473276615, + 0.008323779329657555, + 0.006140668410807848, + 0.004275929182767868, + 0.005568988621234894, + 0.006235840264707804, + 0.006003781221807003, + 0.004612760152667761, + 0.003890265943482518, + 0.005190892610698938, + 0.004679327365010977, + 0.0055557554587721825, + 0.006754714995622635, + 0.005020996555685997, + 0.005267754662781954 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 0 + ], + "selected_reward": -0.6449062824249268 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02984067052602768 + ], + "opposite_logprob_tokens": [ + -3.5298407077789307 + ], + "logprob_reward": -0.02984067052602768, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.02984067052602768, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12486165761947632, + 0.05930778756737709, + 0.051114462316036224, + 0.08354093879461288, + 0.08155985921621323, + 0.11454690247774124, + 0.1100090816617012, + 0.17078587412834167, + 0.1600693315267563, + 0.04420403018593788 + ], + "credit_map": [ + 0.03182179853320122, + 0.011143895797431469, + 0.01388812530785799, + 0.009929699823260307, + 0.013060986064374447, + 0.017733395099639893, + 0.010442976839840412, + 0.016840795055031776, + 0.008442327380180359, + 0.004900133237242699, + 0.012303915806114674, + 0.009284092113375664, + 0.006033039651811123, + 0.0071508209221065044, + 0.004404880106449127, + 0.006788577884435654, + 0.0069212401285767555, + 0.006076160352677107, + 0.007441525813192129, + 0.0067990729585289955, + 0.005292599555104971, + 0.0054433648474514484, + 0.005648619029670954, + 0.00749187869951129, + 0.014332529157400131, + 0.005295525770634413, + 0.015397988259792328, + 0.014413651078939438, + 0.007103834301233292, + 0.010082830674946308, + 0.009364401921629906, + 0.007550184614956379, + 0.013291999697685242, + 0.012288052588701248, + 0.010759066790342331, + 0.009126236662268639, + 0.007839737460017204, + 0.009767439216375351, + 0.009567592293024063, + 0.008919741958379745, + 0.026135532185435295, + 0.01613066904246807, + 0.01083147618919611, + 0.012060999870300293, + 0.00839375052601099, + 0.014802487567067146, + 0.010475484654307365, + 0.015716515481472015, + 0.014370237477123737, + 0.017147645354270935, + 0.011217890307307243, + 0.011279361322522163, + 0.022399980574846268, + 0.012072265148162842, + 0.011351684108376503, + 0.010170022025704384, + 0.026626553386449814, + 0.019442405551671982, + 0.027163174003362656, + 0.020765122026205063, + 0.019964590668678284, + 0.01922762766480446, + 0.017152931541204453, + 0.020443499088287354, + 0.021426372230052948, + 0.02992449328303337, + 0.017194034531712532, + 0.015285423025488853, + 0.014043977484107018, + 0.023070013150572777, + 0.016378166154026985, + 0.022746853530406952, + 0.005410619545727968, + 0.008273433893918991, + 0.006127209402620792, + 0.006334797944873571, + 0.004816979635506868, + 0.005133365746587515, + 0.003187010297551751, + 0.004920619539916515 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 8, + 0 + ], + "selected_reward": -0.02984067052602768 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.14282670617103577 + ], + "opposite_logprob_tokens": [ + -2.017826795578003 + ], + "logprob_reward": -0.14282670617103577, + "yes_no_margin_reward": 1.8750001192092896, + "token_credit_reward": -0.14282670617103577, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14774100482463837, + 0.208733931183815, + 0.1514657437801361, + 0.057653747498989105, + 0.04045223444700241, + 0.2192343771457672, + 0.05056247487664223, + 0.0455702543258667, + 0.031427934765815735, + 0.04715826362371445 + ], + "credit_map": [ + 0.029453260824084282, + 0.01183660514652729, + 0.018376346677541733, + 0.013237007893621922, + 0.016911324113607407, + 0.017037374898791313, + 0.01340588927268982, + 0.027483195066452026, + 0.02464538998901844, + 0.02065892145037651, + 0.03712617605924606, + 0.02608495205640793, + 0.023900043219327927, + 0.022242382168769836, + 0.01598597690463066, + 0.03809010609984398, + 0.017649400979280472, + 0.014612076804041862, + 0.02106703259050846, + 0.02632119506597519, + 0.018904421478509903, + 0.016910230740904808, + 0.015689987689256668, + 0.020311405882239342, + 0.008447086438536644, + 0.004802091047167778, + 0.008632550947368145, + 0.011335649527609348, + 0.006204822566360235, + 0.005896608345210552, + 0.0065094418823719025, + 0.0058254967443645, + 0.007015915587544441, + 0.005129287950694561, + 0.0054610795341432095, + 0.005139720160514116, + 0.004917639307677746, + 0.003973008133471012, + 0.004674775991588831, + 0.0041408054530620575, + 0.04685909301042557, + 0.022056816145777702, + 0.022537020966410637, + 0.027978751808404922, + 0.029565313830971718, + 0.0302417054772377, + 0.016911635175347328, + 0.02308402955532074, + 0.007625691592693329, + 0.004647924564778805, + 0.0074719274416565895, + 0.00617173919454217, + 0.00673892954364419, + 0.005159527994692326, + 0.006727852858603001, + 0.0060188802890479565, + 0.006871618330478668, + 0.005009317770600319, + 0.004725387319922447, + 0.006511594168841839, + 0.005377472378313541, + 0.005449929740279913, + 0.00620102696120739, + 0.0054239085875451565, + 0.003768750699236989, + 0.004645925480872393, + 0.004440590273588896, + 0.002957534044981003, + 0.003114704042673111, + 0.005633159540593624, + 0.0023423032835125923, + 0.00452496949583292, + 0.006114172283560038, + 0.007192765362560749, + 0.005772050004452467, + 0.00685812346637249, + 0.005769946612417698, + 0.005451332777738571, + 0.004987356252968311, + 0.005012518726289272 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 1, + 2 + ], + "selected_reward": -0.14282670617103577 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.008637449704110622 + ], + "opposite_logprob_tokens": [ + -4.758637428283691 + ], + "logprob_reward": -0.008637449704110622, + "yes_no_margin_reward": 4.75, + "token_credit_reward": -0.008637449704110622, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.29683026671409607, + 0.10322012007236481, + 0.08471335470676422, + 0.32217398285865784, + 0.025119390338659286, + 0.06024729087948799, + 0.019419249147176743, + 0.028955798596143723, + 0.03109225444495678, + 0.028228286653757095 + ], + "credit_map": [ + 0.06276434659957886, + 0.03959879279136658, + 0.02915070578455925, + 0.03014427423477173, + 0.03169596940279007, + 0.04054905101656914, + 0.025679077953100204, + 0.03724806383252144, + 0.013647102750837803, + 0.008819595910608768, + 0.022531000897288322, + 0.013264075852930546, + 0.012349389493465424, + 0.011768215335905552, + 0.007768004667013884, + 0.013072739355266094, + 0.008484388701617718, + 0.004878102336078882, + 0.015236691571772099, + 0.01939363405108452, + 0.005428615491837263, + 0.008809414692223072, + 0.009864363819360733, + 0.01261813659220934, + 0.05679994076490402, + 0.020676614716649055, + 0.057077161967754364, + 0.051744334399700165, + 0.03273759037256241, + 0.03752337396144867, + 0.03633952885866165, + 0.029275432229042053, + 0.0046405247412621975, + 0.0030305355321615934, + 0.0030106245540082455, + 0.0024387871380895376, + 0.0025831705424934626, + 0.0034142860677093267, + 0.0031118865590542555, + 0.0028895754367113113, + 0.013540935702621937, + 0.008941279724240303, + 0.004646894987672567, + 0.006518178153783083, + 0.006899066269397736, + 0.008691119961440563, + 0.004192071966826916, + 0.006817743647843599, + 0.002502990886569023, + 0.002161364071071148, + 0.0026882714591920376, + 0.002715968992561102, + 0.0024215488228946924, + 0.00188068684656173, + 0.002427540486678481, + 0.0026208788622170687, + 0.0036128568463027477, + 0.003162237349897623, + 0.003651862032711506, + 0.005081071984022856, + 0.0033459097612649202, + 0.0029787833336740732, + 0.004083951469510794, + 0.0030391260515898466, + 0.004432624205946922, + 0.00504757696762681, + 0.0035907200071960688, + 0.00284397741779685, + 0.0028452749829739332, + 0.004300564527511597, + 0.003204949898645282, + 0.0048265657387673855, + 0.0031783292070031166, + 0.0033698079641908407, + 0.0031194945331662893, + 0.005007216706871986, + 0.003332851454615593, + 0.00285700592212379, + 0.004335617646574974, + 0.003027962287887931 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 1 + ], + "selected_reward": -0.008637449704110622 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.03369579464197159 + ], + "opposite_logprob_tokens": [ + -3.408695697784424 + ], + "logprob_reward": -0.03369579464197159, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.03369579464197159, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.09905064105987549, + 0.1284511834383011, + 0.0942775160074234, + 0.10821253061294556, + 0.036894477903842926, + 0.11238044500350952, + 0.09172597527503967, + 0.06274312734603882, + 0.07444245368242264, + 0.19182169437408447 + ], + "credit_map": [ + 0.013068771921098232, + 0.012802166864275932, + 0.016065210103988647, + 0.012469726614654064, + 0.010452315211296082, + 0.009029901586472988, + 0.008447747677564621, + 0.0167147945612669, + 0.018601009622216225, + 0.01312265545129776, + 0.025895070284605026, + 0.016160408034920692, + 0.011744828894734383, + 0.016117822378873825, + 0.010286569595336914, + 0.016522817313671112, + 0.009842436760663986, + 0.007920186035335064, + 0.014013562351465225, + 0.01990518346428871, + 0.008732798509299755, + 0.008935995399951935, + 0.011726812459528446, + 0.013200543820858002, + 0.013757857494056225, + 0.010460748337209225, + 0.01943979784846306, + 0.013998610898852348, + 0.012967475689947605, + 0.01132613979279995, + 0.011929714120924473, + 0.014332182705402374, + 0.007386325392872095, + 0.004261474125087261, + 0.004931021947413683, + 0.004235505126416683, + 0.004148116800934076, + 0.005033883266150951, + 0.0035714965779334307, + 0.0033266523387283087, + 0.020847540348768234, + 0.01090648677200079, + 0.011069683358073235, + 0.017147978767752647, + 0.01888333261013031, + 0.014207014814019203, + 0.009370419196784496, + 0.00994799379259348, + 0.01692703366279602, + 0.009617386385798454, + 0.01025384571403265, + 0.010272997431457043, + 0.008643515408039093, + 0.010630401782691479, + 0.011699050664901733, + 0.013681750744581223, + 0.00898014660924673, + 0.006932451855391264, + 0.008218860253691673, + 0.010794002562761307, + 0.008056500926613808, + 0.0064458949491381645, + 0.00816101673990488, + 0.005154257174581289, + 0.009568930603563786, + 0.01135527528822422, + 0.008249358274042606, + 0.008111201226711273, + 0.007022096309810877, + 0.012575426138937473, + 0.008619381114840508, + 0.008940789848566055, + 0.019879763945937157, + 0.01931646652519703, + 0.027408841997385025, + 0.043137889355421066, + 0.015194856561720371, + 0.012626783922314644, + 0.0321040078997612, + 0.022153081372380257 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.03369579464197159 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic low-angle tracking shot follows a heavy main battle tank as it hurtles forward at breakneck speed across a muddy, war-torn battlefield. The tank's hull, covered in dark olive green and brown camouflage, is covered in scratches and aged stains, its turret pointed steadily forward. Beneath gloomy, low-hanging gray clouds, a cold, diffuse light accentuates the icy texture of the metal and the wet sheen of the ground. Its broad, heavy metal tracks churn frantically, each bite fiercely into the slippery earth, sending chunks of sticky, dark brown mud and broken grass high into the air. These chunks arc through the air before slamming back down to the ground or splattering against the tank's rear, creating a spectacular, chaotic wave of mud. The distant skyline is outlined by sparse, charred trees and a hazy cloud of gunpowder. The entire scene is imbued with a sense of raw power, tense dynamics, and a cold, tense atmosphere of war.", + "seed": 31599, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31599_dp_13_rank_26_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.01605682554654777, + "vlm_logprob": -0.01605682554654777, + "vlm_yes_no_margin": 4.575, + "vlm_token_credit": -0.01605682554654777, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02586933597922325 + ], + "opposite_logprob_tokens": [ + -4.150869369506836 + ], + "logprob_reward": -0.02586933597922325, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.02586933597922325, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1450175940990448, + 0.06962603330612183, + 0.10248947143554688, + 0.29728081822395325, + 0.07770329713821411, + 0.10662058740854263, + 0.0832272320985794, + 0.029169103130698204, + 0.03914603590965271, + 0.049719780683517456 + ], + "credit_map": [ + 0.020877107977867126, + 0.012156415730714798, + 0.02203638292849064, + 0.01624498888850212, + 0.014669745229184628, + 0.023372014984488487, + 0.021264566108584404, + 0.01439638715237379, + 0.010504394769668579, + 0.007040469441562891, + 0.011469319462776184, + 0.006847175769507885, + 0.009000472724437714, + 0.010451056063175201, + 0.009175756946206093, + 0.005137391854077578, + 0.014344191178679466, + 0.013091711327433586, + 0.01650877855718136, + 0.011563004925847054, + 0.01012936607003212, + 0.011746710166335106, + 0.0128786014392972, + 0.012227114289999008, + 0.027834519743919373, + 0.025304710492491722, + 0.06247931718826294, + 0.029992101714015007, + 0.01951885037124157, + 0.03702032193541527, + 0.032956141978502274, + 0.06217488273978233, + 0.006699489429593086, + 0.006450582295656204, + 0.012060174718499184, + 0.010783793404698372, + 0.008536760695278645, + 0.007402404211461544, + 0.010474353097379208, + 0.015295743942260742, + 0.0127833541482687, + 0.008376898244023323, + 0.013131128624081612, + 0.015395882539451122, + 0.00983638409525156, + 0.009895249269902706, + 0.01811063289642334, + 0.019091065973043442, + 0.009166342206299305, + 0.006859245244413614, + 0.016559593379497528, + 0.013193910010159016, + 0.008818887174129486, + 0.006436798255890608, + 0.01078697107732296, + 0.011405490338802338, + 0.003367707598954439, + 0.0024807099252939224, + 0.004395611118525267, + 0.003498741891235113, + 0.003180933650583029, + 0.0033596609719097614, + 0.004886935465037823, + 0.003998801577836275, + 0.004686957690864801, + 0.0070669883862137794, + 0.009262452833354473, + 0.004821789916604757, + 0.003743422217667103, + 0.0029331545811146498, + 0.003505380591377616, + 0.0031258973758667707, + 0.004939843900501728, + 0.005526330322027206, + 0.005879044067114592, + 0.007690741214901209, + 0.0057441540993750095, + 0.008431182242929935, + 0.00564232375472784, + 0.005866167601197958 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 5 + ], + "selected_reward": -0.02586933597922325 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.026378124952316284 + ], + "opposite_logprob_tokens": [ + -3.6513781547546387 + ], + "logprob_reward": -0.026378124952316284, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.026378124952316284, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.12812429666519165, + 0.1604626625776291, + 0.10303559899330139, + 0.18761135637760162, + 0.06559468060731888, + 0.09459429234266281, + 0.06060725450515747, + 0.05356856808066368, + 0.07940517365932465, + 0.06699611246585846 + ], + "credit_map": [ + 0.017128951847553253, + 0.012447944842278957, + 0.01806027442216873, + 0.013731574639678001, + 0.022721227258443832, + 0.013672295026481152, + 0.018368873745203018, + 0.011993147432804108, + 0.028564361855387688, + 0.0152504350990057, + 0.02112702466547489, + 0.014238838106393814, + 0.022790279239416122, + 0.025905968621373177, + 0.018135104328393936, + 0.014450655318796635, + 0.012053204700350761, + 0.008177319541573524, + 0.019105924293398857, + 0.014993662014603615, + 0.010756568983197212, + 0.010330112650990486, + 0.016596991568803787, + 0.011021814309060574, + 0.015220566652715206, + 0.014930227771401405, + 0.03954271599650383, + 0.01908918470144272, + 0.012058805674314499, + 0.02322162315249443, + 0.021482903510332108, + 0.0420653335750103, + 0.0054425341077148914, + 0.006357920356094837, + 0.011455525644123554, + 0.010361193679273129, + 0.006897489540278912, + 0.005975681822746992, + 0.009052351117134094, + 0.01005198061466217, + 0.01253328938037157, + 0.0063015460036695, + 0.021263670176267624, + 0.009426062926650047, + 0.0063619534485042095, + 0.015537896193563938, + 0.009578424505889416, + 0.013591448776423931, + 0.006741792894899845, + 0.004559263586997986, + 0.016177229583263397, + 0.009794989600777626, + 0.004912656731903553, + 0.004986355546861887, + 0.006170907989144325, + 0.007264062762260437, + 0.005977016873657703, + 0.004738478921353817, + 0.008193514309823513, + 0.007368486374616623, + 0.007143449038267136, + 0.007571699563413858, + 0.005982352886348963, + 0.006593570578843355, + 0.009596114978194237, + 0.013928231783211231, + 0.01561190839856863, + 0.009551283903419971, + 0.00909426249563694, + 0.0050412616692483425, + 0.011161036789417267, + 0.005421068985015154, + 0.007414678577333689, + 0.008395351469516754, + 0.007262615021318197, + 0.010834620334208012, + 0.007703445386141539, + 0.011523488909006119, + 0.007930578663945198, + 0.005931330844759941 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 1, + 0 + ], + "selected_reward": -0.026378124952316284 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00463993102312088 + ], + "opposite_logprob_tokens": [ + -5.379640102386475 + ], + "logprob_reward": -0.00463993102312088, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.00463993102312088, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17850354313850403, + 0.1147889718413353, + 0.19940917193889618, + 0.17587248980998993, + 0.07290711253881454, + 0.031451959162950516, + 0.11189579218626022, + 0.04726853594183922, + 0.044456563889980316, + 0.023445919156074524 + ], + "credit_map": [ + 0.023142650723457336, + 0.017913933843374252, + 0.0329555980861187, + 0.01626008190214634, + 0.027429580688476562, + 0.02869240567088127, + 0.015954554080963135, + 0.016154728829860687, + 0.0257203858345747, + 0.01042712852358818, + 0.01718439906835556, + 0.010000642389059067, + 0.010463938117027283, + 0.019647473469376564, + 0.009126625023782253, + 0.012218371964991093, + 0.02756071649491787, + 0.015499391593039036, + 0.033209215849637985, + 0.02119951881468296, + 0.02637171745300293, + 0.023016750812530518, + 0.028171926736831665, + 0.024379922077059746, + 0.015687208622694016, + 0.015025214292109013, + 0.03730691224336624, + 0.015298307873308659, + 0.01325466763228178, + 0.02062499150633812, + 0.01959856227040291, + 0.039076611399650574, + 0.004686032421886921, + 0.007053170818835497, + 0.010816507041454315, + 0.011440359987318516, + 0.009338774718344212, + 0.006243794225156307, + 0.009334060363471508, + 0.013994411565363407, + 0.005370553582906723, + 0.0020958862733095884, + 0.004476213362067938, + 0.003720158478245139, + 0.0022782550659030676, + 0.006107674911618233, + 0.003507518209517002, + 0.0038956974167376757, + 0.012473833747208118, + 0.006549748592078686, + 0.027988489717245102, + 0.017300112172961235, + 0.009910853579640388, + 0.010686773806810379, + 0.013488995842635632, + 0.013496968895196915, + 0.004931780509650707, + 0.0047094919718801975, + 0.010699336417019367, + 0.006605622358620167, + 0.003990963101387024, + 0.005748785100877285, + 0.005481464322656393, + 0.005101090297102928, + 0.004709512926638126, + 0.005060522351413965, + 0.009820918552577496, + 0.006487421691417694, + 0.006447874940931797, + 0.0029015722684562206, + 0.004924022126942873, + 0.004104713909327984, + 0.00229523703455925, + 0.0032093713525682688, + 0.0039762272499501705, + 0.003163618966937065, + 0.0013946926919743419, + 0.0014693164266645908, + 0.002904948079958558, + 0.005032505374401808 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 3 + ], + "selected_reward": -0.00463993102312088 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0028169972356408834 + ], + "opposite_logprob_tokens": [ + -5.877817153930664 + ], + "logprob_reward": -0.0028169972356408834, + "yes_no_margin_reward": 5.875, + "token_credit_reward": -0.0028169972356408834, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11471719294786453, + 0.1401958465576172, + 0.28073787689208984, + 0.088646799325943, + 0.05447515472769737, + 0.10547452419996262, + 0.035608794540166855, + 0.09222185611724854, + 0.06858842074871063, + 0.01933351717889309 + ], + "credit_map": [ + 0.017085595056414604, + 0.01037578471004963, + 0.023059291765093803, + 0.01212430838495493, + 0.012273081578314304, + 0.017047833651304245, + 0.011029628105461597, + 0.011721678078174591, + 0.03198186680674553, + 0.0125089967623353, + 0.021903805434703827, + 0.011512394063174725, + 0.011880558915436268, + 0.026713743805885315, + 0.010538711212575436, + 0.013155772350728512, + 0.0359751358628273, + 0.019949447363615036, + 0.04429369792342186, + 0.028788553550839424, + 0.04189646244049072, + 0.0314747616648674, + 0.04166099801659584, + 0.03669882193207741, + 0.0054455772042274475, + 0.005234809592366219, + 0.02012617513537407, + 0.009359238669276237, + 0.0055328249000012875, + 0.010770954191684723, + 0.0104684354737401, + 0.021708780899643898, + 0.0041478886269032955, + 0.00559444073587656, + 0.00804053246974945, + 0.007080630864948034, + 0.007509248331189156, + 0.0054302942007780075, + 0.008301564492285252, + 0.008370558731257915, + 0.014327016659080982, + 0.0071686129085719585, + 0.023369546979665756, + 0.010252564214169979, + 0.007770108059048653, + 0.017074503004550934, + 0.009797954000532627, + 0.01571422442793846, + 0.004008043557405472, + 0.002382411155849695, + 0.010830936953425407, + 0.005135074257850647, + 0.002856994280591607, + 0.0031691540498286486, + 0.002863372443243861, + 0.004362810868769884, + 0.011422683484852314, + 0.011205160059034824, + 0.018498608842492104, + 0.012132717296481133, + 0.005626061465591192, + 0.012434869073331356, + 0.011365083046257496, + 0.009536677971482277, + 0.010831105522811413, + 0.016963297501206398, + 0.014662014320492744, + 0.008008109405636787, + 0.005187504459172487, + 0.004270509816706181, + 0.004130087327212095, + 0.004535801708698273, + 0.0018340203678235412, + 0.0022614621557295322, + 0.002455278066918254, + 0.0027533031534403563, + 0.0017747227102518082, + 0.002110381843522191, + 0.0022450026590377092, + 0.0038993479683995247 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 1, + 0 + ], + "selected_reward": -0.0028169972356408834 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020579738542437553 + ], + "opposite_logprob_tokens": [ + -3.8955798149108887 + ], + "logprob_reward": -0.020579738542437553, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020579738542437553, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11300094425678253, + 0.07198503613471985, + 0.2192680537700653, + 0.1788940578699112, + 0.09731366485357285, + 0.08419399708509445, + 0.059366144239902496, + 0.06705135107040405, + 0.07206899672746658, + 0.03685775026679039 + ], + "credit_map": [ + 0.013479327782988548, + 0.009280393831431866, + 0.01911567524075508, + 0.015550734475255013, + 0.01454690657556057, + 0.01261225901544094, + 0.012689744122326374, + 0.015725912526249886, + 0.011278652586042881, + 0.007091942243278027, + 0.010199506767094135, + 0.008025246672332287, + 0.010138101875782013, + 0.009396742098033428, + 0.009174617938697338, + 0.0066802240908145905, + 0.031801801174879074, + 0.023298194631934166, + 0.0335703119635582, + 0.024776998907327652, + 0.025818107649683952, + 0.023953793570399284, + 0.02754857949912548, + 0.028500281274318695, + 0.022450372576713562, + 0.013814922422170639, + 0.03794855624437332, + 0.01882011629641056, + 0.011985121294856071, + 0.02322905696928501, + 0.017983773723244667, + 0.03266216814517975, + 0.009804203175008297, + 0.008517864160239697, + 0.01552871335297823, + 0.01579228602349758, + 0.009177474305033684, + 0.009661664254963398, + 0.012286842800676823, + 0.01654462330043316, + 0.010875944048166275, + 0.006301127374172211, + 0.013488140888512135, + 0.010651132091879845, + 0.00704500637948513, + 0.009981799870729446, + 0.014184698462486267, + 0.01166615542024374, + 0.006558776833117008, + 0.004719092510640621, + 0.013729527592658997, + 0.0071022664196789265, + 0.007047576364129782, + 0.005085754208266735, + 0.007518730126321316, + 0.007604425773024559, + 0.007749549578875303, + 0.010330547578632832, + 0.012135001830756664, + 0.009718255139887333, + 0.006509561091661453, + 0.007441551424562931, + 0.006365996319800615, + 0.006800884380936623, + 0.007610682398080826, + 0.008464016020298004, + 0.012981591746211052, + 0.010493687354028225, + 0.007889476604759693, + 0.006743143778294325, + 0.011557267047464848, + 0.006329132243990898, + 0.003662522416561842, + 0.00558526162058115, + 0.007347021717578173, + 0.004250498488545418, + 0.0029651550576090813, + 0.002226981334388256, + 0.004543846473097801, + 0.006276463624089956 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 0 + ], + "selected_reward": -0.020579738542437553 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "strikingly dynamic low-angle tracking shot follows a heavy main battle tank as it hurtles forward at breakneck speed across a muddy, war-torn battlefield. The tank's hull, covered in dark olive green and brown camouflage, is covered in scratches and aged stains, its turret pointed steadily forward. Beneath gloomy, low-hanging gray clouds, a cold, diffuse light accentuates the icy texture of the metal and the wet sheen of the ground. Its broad, heavy metal tracks churn frantically, each bite fiercely into the slippery earth, sending chunks of sticky, dark brown mud and broken grass high into the air. These chunks arc through the air before slamming back down to the ground or splattering against the tank's rear, creating a spectacular, chaotic wave of mud. The distant skyline is outlined by sparse, charred trees and a hazy cloud of gunpowder. The entire scene is imbued with a sense of raw power, tense dynamics, and a cold, tense atmosphere of war.", + "seed": 31600, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_6/index_9439_seed_31600_dp_13_rank_26_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 6, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.04615958565846086, + "vlm_logprob": -0.04615958565846086, + "vlm_yes_no_margin": 3.7, + "vlm_token_credit": -0.04615958565846086, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.15484267473220825 + ], + "opposite_logprob_tokens": [ + -2.0298426151275635 + ], + "logprob_reward": -0.15484267473220825, + "yes_no_margin_reward": 1.875, + "token_credit_reward": -0.15484267473220825, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21991175413131714, + 0.06333042681217194, + 0.10526373237371445, + 0.15138758718967438, + 0.14777830243110657, + 0.09464500844478607, + 0.044304009526968, + 0.05509636551141739, + 0.07762160897254944, + 0.04066119343042374 + ], + "credit_map": [ + 0.02955988608300686, + 0.020789306610822678, + 0.02811535820364952, + 0.029381418600678444, + 0.03849201276898384, + 0.02776036597788334, + 0.024245906621217728, + 0.021567493677139282, + 0.007077155169099569, + 0.005420881323516369, + 0.01102240290492773, + 0.005706982221454382, + 0.0072057778015732765, + 0.011607292108237743, + 0.005234175827354193, + 0.010055764578282833, + 0.015177818946540356, + 0.009417437948286533, + 0.02278081700205803, + 0.009075158275663853, + 0.011105475015938282, + 0.01308103371411562, + 0.010088001377880573, + 0.014537985436618328, + 0.024705795571208, + 0.02869388647377491, + 0.01382473949342966, + 0.010560490190982819, + 0.016680007800459862, + 0.018507957458496094, + 0.0114417914301157, + 0.026972917839884758, + 0.03341418877243996, + 0.009993883781135082, + 0.009490028955042362, + 0.015701161697506905, + 0.015753159299492836, + 0.02237590029835701, + 0.013546657748520374, + 0.027503322809934616, + 0.030608084052801132, + 0.006416069343686104, + 0.01016015000641346, + 0.007559098303318024, + 0.005848342087119818, + 0.02080395631492138, + 0.005737615749239922, + 0.0075116935186088085, + 0.007242568768560886, + 0.005013565067201853, + 0.006409069057554007, + 0.006514108739793301, + 0.002713648835197091, + 0.005863289348781109, + 0.003975588362663984, + 0.006572171114385128, + 0.005601819138973951, + 0.0072160279378294945, + 0.012443904764950275, + 0.00792410783469677, + 0.003992687910795212, + 0.005032124929130077, + 0.007552304305136204, + 0.005333390086889267, + 0.0054357475601136684, + 0.017110537737607956, + 0.013621705584228039, + 0.007450215984135866, + 0.011161886155605316, + 0.00604524090886116, + 0.0071624997071921825, + 0.009633778594434261, + 0.0033432706259191036, + 0.003427942516282201, + 0.005979408975690603, + 0.008215382695198059, + 0.0033953513484448195, + 0.00393934128805995, + 0.006545075681060553, + 0.005815420299768448 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 4 + ], + "selected_reward": -0.15484267473220825 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018220899626612663 + ], + "opposite_logprob_tokens": [ + -4.018220901489258 + ], + "logprob_reward": -0.018220899626612663, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018220899626612663, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.17968668043613434, + 0.04132828488945961, + 0.04451717808842659, + 0.061466775834560394, + 0.34574267268180847, + 0.038137901574373245, + 0.022098612040281296, + 0.0356309600174427, + 0.17093777656555176, + 0.06045319512486458 + ], + "credit_map": [ + 0.03542179614305496, + 0.01475505717098713, + 0.02160661108791828, + 0.015188192948698997, + 0.021336780861020088, + 0.044086065143346786, + 0.013043380342423916, + 0.01424879115074873, + 0.004868170712143183, + 0.0041382797062397, + 0.008432186208665371, + 0.00471661938354373, + 0.0045038750395178795, + 0.005286735016852617, + 0.004013465251773596, + 0.005368951242417097, + 0.007767328992486, + 0.005368355195969343, + 0.009884001687169075, + 0.003055964596569538, + 0.004438533913344145, + 0.006209655664861202, + 0.0029432980809360743, + 0.004850039258599281, + 0.010068288072943687, + 0.010027251206338406, + 0.007048522122204304, + 0.005021506920456886, + 0.0070118079893291, + 0.008855212479829788, + 0.0053167822770774364, + 0.00811739545315504, + 0.08635023981332779, + 0.027618763968348503, + 0.02602098323404789, + 0.03527449443936348, + 0.029535286128520966, + 0.05694923922419548, + 0.029421331360936165, + 0.054572317749261856, + 0.00791619997471571, + 0.0044630966149270535, + 0.003802225925028324, + 0.0030160038731992245, + 0.0062827253714203835, + 0.0066202846355736256, + 0.003073323518037796, + 0.0029640397988259792, + 0.003873561741784215, + 0.0028109883423894644, + 0.003152141347527504, + 0.002608630107715726, + 0.0021702297963202, + 0.0032139711547642946, + 0.0016049330588430166, + 0.002664154162630439, + 0.006874988321214914, + 0.0045219468884170055, + 0.006131857167929411, + 0.004298749845474958, + 0.0029379825573414564, + 0.0045790718868374825, + 0.003343050368130207, + 0.002943310420960188, + 0.011369728483259678, + 0.033800747245550156, + 0.028892885893583298, + 0.016270361840724945, + 0.02756623551249504, + 0.013201605528593063, + 0.01661452278494835, + 0.02322167344391346, + 0.005361147224903107, + 0.008337629027664661, + 0.009446672163903713, + 0.00910857692360878, + 0.007414594758301973, + 0.008965815417468548, + 0.006058597005903721, + 0.005760159809142351 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 8 + ], + "selected_reward": -0.018220899626612663 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.023301584646105766 + ], + "opposite_logprob_tokens": [ + -3.773301601409912 + ], + "logprob_reward": -0.023301584646105766, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.023301584646105766, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.6519798636436462, + 0.04134763777256012, + 0.07660665363073349, + 0.021628130227327347, + 0.05360015109181404, + 0.0566428117454052, + 0.010059824213385582, + 0.02769755944609642, + 0.0292525514960289, + 0.031184833496809006 + ], + "credit_map": [ + 0.11490540951490402, + 0.0580870620906353, + 0.08370167016983032, + 0.09343989938497543, + 0.09518303722143173, + 0.10987752676010132, + 0.056461531668901443, + 0.040323708206415176, + 0.004487111698836088, + 0.0037899913731962442, + 0.008528067730367184, + 0.0036521286237984896, + 0.004044791683554649, + 0.006662955041974783, + 0.003500569611787796, + 0.006682022009044886, + 0.014216043055057526, + 0.007636205293238163, + 0.01477514486759901, + 0.005611563567072153, + 0.007478092797100544, + 0.01119568943977356, + 0.007153916638344526, + 0.008539996109902859, + 0.0030295851174741983, + 0.002036530990153551, + 0.002398685086518526, + 0.002270521828904748, + 0.0022091041319072247, + 0.002216090215370059, + 0.002002515597268939, + 0.005465097725391388, + 0.011228857561945915, + 0.004107901360839605, + 0.0062971776351332664, + 0.006706629879772663, + 0.005508678965270519, + 0.007186310365796089, + 0.005832466296851635, + 0.0067321257665753365, + 0.017561599612236023, + 0.004284125752747059, + 0.0057807257398962975, + 0.004835167899727821, + 0.004700010642409325, + 0.011679309420287609, + 0.003690807381644845, + 0.004111066926270723, + 0.0019639027304947376, + 0.001314595341682434, + 0.0012497419957071543, + 0.0012095201527699828, + 0.0008546598255634308, + 0.0014975926605984569, + 0.0008456956711597741, + 0.0011241162428632379, + 0.0024447874166071415, + 0.004307447001338005, + 0.005169725511223078, + 0.0033966032788157463, + 0.003079912858083844, + 0.002993775764480233, + 0.003826369298622012, + 0.0024789394810795784, + 0.002069217385724187, + 0.005989800672978163, + 0.0053893704898655415, + 0.0029836029279977083, + 0.004319791682064533, + 0.0022844800259917974, + 0.002612221520394087, + 0.003604067489504814, + 0.0021734721958637238, + 0.003959678113460541, + 0.005512671545147896, + 0.005948798265308142, + 0.0034414392430335283, + 0.0025738829281181097, + 0.003943395800888538, + 0.003631494240835309 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 5 + ], + "selected_reward": -0.023301584646105766 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004638150800019503 + ], + "opposite_logprob_tokens": [ + -5.379638195037842 + ], + "logprob_reward": -0.004638150800019503, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004638150800019503, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3105393350124359, + 0.13265450298786163, + 0.026821935549378395, + 0.14771199226379395, + 0.03689523786306381, + 0.11524663120508194, + 0.026271654292941093, + 0.11366258561611176, + 0.04204265773296356, + 0.048153407871723175 + ], + "credit_map": [ + 0.05913064628839493, + 0.02737405151128769, + 0.03827357292175293, + 0.03741927817463875, + 0.041917555034160614, + 0.06121060997247696, + 0.02500958740711212, + 0.020204048603773117, + 0.023138614371418953, + 0.006972692906856537, + 0.028055131435394287, + 0.008162576705217361, + 0.010816416703164577, + 0.023743776604533195, + 0.00836917944252491, + 0.023396115750074387, + 0.004510968923568726, + 0.003750892821699381, + 0.003953821025788784, + 0.002286555478349328, + 0.0030487049371004105, + 0.0038088257424533367, + 0.002032856922596693, + 0.003429310629144311, + 0.025141935795545578, + 0.025759059935808182, + 0.013406599871814251, + 0.012189827859401703, + 0.01705187000334263, + 0.019356414675712585, + 0.01319902017712593, + 0.02160726860165596, + 0.008620751090347767, + 0.0025872548576444387, + 0.004245338030159473, + 0.004344093147665262, + 0.002495786175131798, + 0.005486069247126579, + 0.0027833525091409683, + 0.00633259117603302, + 0.02725188620388508, + 0.007807084824889898, + 0.015506944619119167, + 0.012631246820092201, + 0.0074707805179059505, + 0.01823984831571579, + 0.010904273018240929, + 0.01543456595391035, + 0.0045675793662667274, + 0.0034590926952660084, + 0.003763606771826744, + 0.003036102745682001, + 0.0026125898584723473, + 0.003689961275085807, + 0.0020842577796429396, + 0.003058464964851737, + 0.016634460538625717, + 0.014927839860320091, + 0.022683339193463326, + 0.015582988038659096, + 0.010277191177010536, + 0.011995108798146248, + 0.012683242559432983, + 0.008878418244421482, + 0.0023692091926932335, + 0.00540199177339673, + 0.009688361547887325, + 0.005223558750003576, + 0.0058855777606368065, + 0.003115455387160182, + 0.003905252320691943, + 0.006453249603509903, + 0.003378385677933693, + 0.004654130898416042, + 0.00908712763339281, + 0.009982914663851261, + 0.004500323906540871, + 0.003192324424162507, + 0.007027450483292341, + 0.006330747622996569 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.004638150800019503 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the main subject a heavy battle tank moving forward at high speed?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the tank's hull covered in dark olive green and brown camouflage with visible scratches and stains?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Does the scene show the tank's metal tracks churning mud and grass into the air?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is the environment a muddy battlefield under gloomy, low-hanging gray clouds with a cold diffuse light?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot a low-angle tracking shot following the tank?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.029794618487358093 + ], + "opposite_logprob_tokens": [ + -3.529794692993164 + ], + "logprob_reward": -0.029794618487358093, + "yes_no_margin_reward": 3.5, + "token_credit_reward": -0.029794618487358093, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15443949401378632, + 0.041753970086574554, + 0.09368064254522324, + 0.3483741283416748, + 0.12195982784032822, + 0.036047399044036865, + 0.01780635491013527, + 0.03202342987060547, + 0.06025729700922966, + 0.09365741908550262 + ], + "credit_map": [ + 0.028728581964969635, + 0.011769058182835579, + 0.02124428004026413, + 0.013857532292604446, + 0.01714063063263893, + 0.03616062179207802, + 0.011865229345858097, + 0.013673576526343822, + 0.004134902264922857, + 0.002765998477116227, + 0.00773197365924716, + 0.004129526671022177, + 0.0030355851631611586, + 0.0061971265822649, + 0.003513579024001956, + 0.01024528406560421, + 0.014546994119882584, + 0.006366809830069542, + 0.017323946580290794, + 0.010409792885184288, + 0.011718927882611752, + 0.012506811879575253, + 0.007052749861031771, + 0.013754612766206264, + 0.062180470675230026, + 0.06288091838359833, + 0.03082825243473053, + 0.027909785509109497, + 0.040192391723394394, + 0.04564473778009415, + 0.030135981738567352, + 0.048601601272821426, + 0.03002948686480522, + 0.008334365673363209, + 0.00952878687530756, + 0.014931994490325451, + 0.010246609337627888, + 0.019316133111715317, + 0.009728013537824154, + 0.019844437018036842, + 0.00791257619857788, + 0.002963049802929163, + 0.005877289921045303, + 0.003990751225501299, + 0.003839237615466118, + 0.005160945001989603, + 0.002955975942313671, + 0.003347576828673482, + 0.0024765522684901953, + 0.002415555063635111, + 0.0026437328197062016, + 0.0021868799813091755, + 0.0015977922594174743, + 0.002111463574692607, + 0.0019174851477146149, + 0.0024568962398916483, + 0.003233799245208502, + 0.0038082762621343136, + 0.005116208456456661, + 0.005034574307501316, + 0.002592862118035555, + 0.0034349863417446613, + 0.0055039082653820515, + 0.0032988195307552814, + 0.004140641540288925, + 0.012008913792669773, + 0.006575192790478468, + 0.006349632050842047, + 0.0115530826151371, + 0.004529013764113188, + 0.005105840507894754, + 0.009994982741773129, + 0.0058392975479364395, + 0.009488126263022423, + 0.019489526748657227, + 0.019114965572953224, + 0.009102416224777699, + 0.0055904206819832325, + 0.012941394932568073, + 0.012091277167201042 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 0, + 4 + ], + "selected_reward": -0.029794618487358093 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_tensor([3917, 3917])_group_7_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_tensor([3917, 3917])_group_7_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..d9fc7c6727750d7e6f86bee90305788a29d07417 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_tensor([3917, 3917])_group_7_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "static, mid-shot shot, taken from a level angle, captures a moment of tranquility and warmth on a warm afternoon. In a quiet room within an ancient Taoist temple, an elderly Taoist priest, his hair and beard white, sits cross-legged on a thick, woven cattail mat. He wears a plain, washed-out, indigo robe. A beam of warm, golden sunlight slants in through a nearby carved wooden window, illuminating the slowly dancing dust particles in the air. The old Taoist's eyes are slightly closed, a serene, contented smile on his age-marked face. His other hand rests naturally on his knee. In his arms, a plump, orange tabby cat curls up in a ball, dozing soundly, its body rising and falling slightly with its even breathing. The old Taoist priest's hand, covered with wrinkles and with distinct knuckles, is stroking the cat from the top of the head all the way down its back in an extremely slow and gentle rhythm. His movements are full of infinite love and compassion. The whole picture presents a kind of secluded tranquility and harmony under the soft light and shadow, full of Zen and the warmth of life.", + "seed": 4018, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4018_dp_14_rank_28_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.0235046461224556, + "vlm_logprob": -0.0235046461224556, + "vlm_yes_no_margin": 4.4, + "vlm_token_credit": -0.0235046461224556, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.07960094511508942 + ], + "opposite_logprob_tokens": [ + -2.7046010494232178 + ], + "logprob_reward": -0.07960094511508942, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.07960094511508942, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1382046788930893, + 0.110702745616436, + 0.12279397994279861, + 0.09330350905656815, + 0.24185457825660706, + 0.10374613851308823, + 0.06802088022232056, + 0.042864225804805756, + 0.038221657276153564, + 0.040287602692842484 + ], + "credit_map": [ + 0.01639583520591259, + 0.019148146733641624, + 0.017001647502183914, + 0.016163934022188187, + 0.016540780663490295, + 0.014883564785122871, + 0.01765892468392849, + 0.02041185088455677, + 0.01743135415017605, + 0.015734443441033363, + 0.010075807571411133, + 0.012478601187467575, + 0.016433538869023323, + 0.01048047374933958, + 0.011042044498026371, + 0.017026478424668312, + 0.014558578841388226, + 0.01576378382742405, + 0.017597896978259087, + 0.017374729737639427, + 0.018358763307332993, + 0.008837773464620113, + 0.012930639088153839, + 0.01737181842327118, + 0.010835444554686546, + 0.01176299899816513, + 0.01198708824813366, + 0.015482415445148945, + 0.011567934416234493, + 0.007465705741196871, + 0.00979517586529255, + 0.014406743459403515, + 0.023330772295594215, + 0.019667675718665123, + 0.018508385866880417, + 0.03353392705321312, + 0.051083825528621674, + 0.025172283872961998, + 0.042792439460754395, + 0.027765268459916115, + 0.006664833519607782, + 0.006986839696764946, + 0.009414617903530598, + 0.012712161988019943, + 0.024140065535902977, + 0.010554645210504532, + 0.01969061605632305, + 0.013582359068095684, + 0.004741127602756023, + 0.005538946017622948, + 0.00598245020955801, + 0.005116401705890894, + 0.0177498459815979, + 0.006059780716896057, + 0.012143921107053757, + 0.010688399896025658, + 0.006977690849453211, + 0.007979587651789188, + 0.005298184230923653, + 0.004219400230795145, + 0.004879485350102186, + 0.002760902978479862, + 0.004452533554285765, + 0.006296439561992884, + 0.004152002744376659, + 0.0037579149939119816, + 0.0030335960909724236, + 0.0033526660408824682, + 0.007486098445951939, + 0.003535805968567729, + 0.007365697994828224, + 0.005537876859307289, + 0.005403314717113972, + 0.006762202829122543, + 0.0032836301252245903, + 0.005216042511165142, + 0.003540053265169263, + 0.005521037615835667, + 0.004399070516228676, + 0.006162251345813274 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 0, + 2 + ], + "selected_reward": -0.07960094511508942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006754660978913307 + ], + "opposite_logprob_tokens": [ + -5.0067548751831055 + ], + "logprob_reward": -0.006754660978913307, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006754660978913307, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23482684791088104, + 0.08013990521430969, + 0.1096637099981308, + 0.11074735969305038, + 0.07260196655988693, + 0.10424382984638214, + 0.08116485178470612, + 0.05465427786111832, + 0.054716289043426514, + 0.09724098443984985 + ], + "credit_map": [ + 0.025304222479462624, + 0.034510381519794464, + 0.023216499015688896, + 0.036769647151231766, + 0.02689659409224987, + 0.01974664255976677, + 0.026929598301649094, + 0.041453249752521515, + 0.010549678467214108, + 0.010776815004646778, + 0.007710921112447977, + 0.010867448523640633, + 0.010928715579211712, + 0.007848840206861496, + 0.009396781213581562, + 0.01206070277839899, + 0.017084673047065735, + 0.01545732282102108, + 0.010482913814485073, + 0.011006890796124935, + 0.01537926122546196, + 0.010153818875551224, + 0.013439489528536797, + 0.016659339889883995, + 0.008981761522591114, + 0.01069011352956295, + 0.0130816251039505, + 0.022922208532691002, + 0.016427481546998024, + 0.0066910674795508385, + 0.01465039886534214, + 0.017302708700299263, + 0.007881427183747292, + 0.008737697266042233, + 0.006208823993802071, + 0.011534073390066624, + 0.013256419450044632, + 0.0078731719404459, + 0.008146257139742374, + 0.008964097127318382, + 0.008589524775743484, + 0.007929885759949684, + 0.01222634594887495, + 0.02040821500122547, + 0.01651920937001705, + 0.0076599386520683765, + 0.017403872683644295, + 0.01350683905184269, + 0.010403643362224102, + 0.010836473666131496, + 0.00789338443428278, + 0.007017313037067652, + 0.010624635964632034, + 0.01109606958925724, + 0.012236538343131542, + 0.011056791059672832, + 0.008563087321817875, + 0.008757016621530056, + 0.007341346703469753, + 0.005951735656708479, + 0.0054989708587527275, + 0.004493484739214182, + 0.0062936171889305115, + 0.007755020167678595, + 0.006117654498666525, + 0.005386857781559229, + 0.007342582568526268, + 0.005565930157899857, + 0.01064752135425806, + 0.004539462737739086, + 0.007768317591398954, + 0.007347961887717247, + 0.012229513376951218, + 0.014176763594150543, + 0.015461309812963009, + 0.012834297493100166, + 0.008559122681617737, + 0.008778710849583149, + 0.010568683966994286, + 0.014632579870522022 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.006754660978913307 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.004641829524189234 + ], + "opposite_logprob_tokens": [ + -5.379642009735107 + ], + "logprob_reward": -0.004641829524189234, + "yes_no_margin_reward": 5.375, + "token_credit_reward": -0.004641829524189234, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.1359638273715973, + 0.06264885514974594, + 0.09470082074403763, + 0.06007593125104904, + 0.35699227452278137, + 0.14428934454917908, + 0.04131247103214264, + 0.035929687321186066, + 0.038807716220617294, + 0.029279107227921486 + ], + "credit_map": [ + 0.0158553384244442, + 0.016599712893366814, + 0.014359105378389359, + 0.01654648408293724, + 0.02294904738664627, + 0.013144589960575104, + 0.014971395023167133, + 0.02153814397752285, + 0.00805466528981924, + 0.0089870635420084, + 0.00655344408005476, + 0.006559968460351229, + 0.010854486376047134, + 0.00724388612434268, + 0.005139906425029039, + 0.009255427867174149, + 0.01573382504284382, + 0.01562809757888317, + 0.008875461295247078, + 0.0122464494779706, + 0.010880094952881336, + 0.008466578088700771, + 0.008990345522761345, + 0.013879972510039806, + 0.008124738931655884, + 0.006704479455947876, + 0.008733746595680714, + 0.008874624036252499, + 0.0068982988595962524, + 0.005271719302982092, + 0.006135164760053158, + 0.009333161637187004, + 0.03428233787417412, + 0.02255524881184101, + 0.04163272678852081, + 0.08495992422103882, + 0.043861377984285355, + 0.03331812098622322, + 0.053030725568532944, + 0.04335181042551994, + 0.012013777159154415, + 0.010816275142133236, + 0.019351914525032043, + 0.034975964576005936, + 0.018015405163168907, + 0.008884251117706299, + 0.02240139991044998, + 0.017830345779657364, + 0.005300214048475027, + 0.005179582629352808, + 0.0035077815409749746, + 0.004532338585704565, + 0.007067475467920303, + 0.004140374716371298, + 0.006454534363001585, + 0.0051301694475114346, + 0.004874750971794128, + 0.0048356736078858376, + 0.004507782869040966, + 0.004167770501226187, + 0.004460696596652269, + 0.0034967318642884493, + 0.00437984149903059, + 0.005206440109759569, + 0.004527847748249769, + 0.004162355791777372, + 0.005959405098110437, + 0.00384029233828187, + 0.006121518090367317, + 0.0034802621230483055, + 0.004507832694798708, + 0.006208204198628664, + 0.005030957981944084, + 0.005344064440578222, + 0.0032930667512118816, + 0.003111991100013256, + 0.002169380895793438, + 0.00237076124176383, + 0.0034482155460864305, + 0.004510669037699699 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 5, + 0 + ], + "selected_reward": -0.004641829524189234 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005951895844191313 + ], + "opposite_logprob_tokens": [ + -5.130951881408691 + ], + "logprob_reward": -0.005951895844191313, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005951895844191313, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19314593076705933, + 0.13141517341136932, + 0.12678682804107666, + 0.14103089272975922, + 0.18386627733707428, + 0.04897942766547203, + 0.07418328523635864, + 0.034312210977077484, + 0.026489105075597763, + 0.03979090228676796 + ], + "credit_map": [ + 0.01999017223715782, + 0.02379409782588482, + 0.029043618589639664, + 0.03836609050631523, + 0.02379642240703106, + 0.011597325094044209, + 0.020941220223903656, + 0.025616969913244247, + 0.020558159798383713, + 0.023407207801938057, + 0.012531845830380917, + 0.01529026310890913, + 0.01730560138821602, + 0.013100960291922092, + 0.011343184858560562, + 0.01787794567644596, + 0.015483153983950615, + 0.02153761312365532, + 0.01566450670361519, + 0.01436648704111576, + 0.016729198396205902, + 0.012943335808813572, + 0.01239759474992752, + 0.017664924263954163, + 0.01300330925732851, + 0.011795475147664547, + 0.015082309953868389, + 0.02799336053431034, + 0.021837500855326653, + 0.010770673863589764, + 0.02034994214773178, + 0.02019832283258438, + 0.018375631421804428, + 0.01827860251069069, + 0.013168446719646454, + 0.017405005171895027, + 0.04074801504611969, + 0.024957044050097466, + 0.02849651873111725, + 0.022437022998929024, + 0.004260356072336435, + 0.004753148183226585, + 0.004369241651147604, + 0.006814073771238327, + 0.009426369331777096, + 0.005860453471541405, + 0.006717687472701073, + 0.006778097711503506, + 0.00690991897135973, + 0.008769117295742035, + 0.005068064201623201, + 0.005401753820478916, + 0.014949128031730652, + 0.0075263651087880135, + 0.015119267627596855, + 0.010439666919410229, + 0.005364098586142063, + 0.00498270383104682, + 0.004310705233365297, + 0.003457593033090234, + 0.004264247138053179, + 0.0030933215748518705, + 0.003490503178909421, + 0.005349038168787956, + 0.004274283070117235, + 0.0036366002168506384, + 0.0028739101253449917, + 0.00259277387522161, + 0.0032135536894202232, + 0.0020337924361228943, + 0.003398942295461893, + 0.004465250298380852, + 0.004686359316110611, + 0.006165983155369759, + 0.0051474496722221375, + 0.003952191211283207, + 0.005989830940961838, + 0.003782013664022088, + 0.003943446557968855, + 0.006123627070337534 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 4, + 3 + ], + "selected_reward": -0.005951895844191313 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.020573899149894714 + ], + "opposite_logprob_tokens": [ + -3.895573854446411 + ], + "logprob_reward": -0.020573899149894714, + "yes_no_margin_reward": 3.875, + "token_credit_reward": -0.020573899149894714, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.10455804318189621, + 0.06332343071699142, + 0.23080994188785553, + 0.18484948575496674, + 0.16620199382305145, + 0.058871202170848846, + 0.05986807495355606, + 0.037363581359386444, + 0.04675891995429993, + 0.0473952479660511 + ], + "credit_map": [ + 0.009329531341791153, + 0.017653949558734894, + 0.009909309446811676, + 0.012410865165293217, + 0.015564081259071827, + 0.00934626068919897, + 0.01255338080227375, + 0.017790675163269043, + 0.00984494760632515, + 0.009380177594721317, + 0.007629913277924061, + 0.007180912885814905, + 0.008222202770411968, + 0.0073894113302230835, + 0.005361869931221008, + 0.008314002305269241, + 0.02158447727560997, + 0.023207204416394234, + 0.024995898827910423, + 0.05566483736038208, + 0.028517315164208412, + 0.014090451411902905, + 0.0274680033326149, + 0.03528176620602608, + 0.016805335879325867, + 0.022213496267795563, + 0.017338503152132034, + 0.02852436527609825, + 0.03262563422322273, + 0.016098642721772194, + 0.02728051133453846, + 0.02396302856504917, + 0.016511444002389908, + 0.013592648319900036, + 0.013348747976124287, + 0.01789131388068199, + 0.03787935525178909, + 0.01892644725739956, + 0.02805970422923565, + 0.01999235525727272, + 0.006314062979072332, + 0.007554498501121998, + 0.007297562435269356, + 0.006888172123581171, + 0.008241754025220871, + 0.004462065175175667, + 0.009664228186011314, + 0.008448866195976734, + 0.005852764938026667, + 0.00613159267231822, + 0.005497109144926071, + 0.0046948702074587345, + 0.00867509189993143, + 0.0059614889323711395, + 0.011614738032221794, + 0.011440427042543888, + 0.00634622061625123, + 0.006772882770746946, + 0.0036669676192104816, + 0.004345361143350601, + 0.004129115492105484, + 0.0027825625147670507, + 0.0034235557541251183, + 0.005896919872611761, + 0.006913481745868921, + 0.00689435051754117, + 0.005442525260150433, + 0.004514030646532774, + 0.005813681520521641, + 0.004220078233629465, + 0.005908271297812462, + 0.0070525058545172215, + 0.0058471523225307465, + 0.007146161049604416, + 0.006580989342182875, + 0.00472670141607523, + 0.0058059534057974815, + 0.004147985018789768, + 0.006196044385433197, + 0.006944263353943825 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 3, + 4 + ], + "selected_reward": -0.020573899149894714 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, captures a moment of tranquility and warmth on a warm afternoon. In a quiet room within an ancient Taoist temple, an elderly Taoist priest, his hair and beard white, sits cross-legged on a thick, woven cattail mat. He wears a plain, washed-out, indigo robe. A beam of warm, golden sunlight slants in through a nearby carved wooden window, illuminating the slowly dancing dust particles in the air. The old Taoist's eyes are slightly closed, a serene, contented smile on his age-marked face. His other hand rests naturally on his knee. In his arms, a plump, orange tabby cat curls up in a ball, dozing soundly, its body rising and falling slightly with its even breathing. The old Taoist priest's hand, covered with wrinkles and with distinct knuckles, is stroking the cat from the top of the head all the way down its back in an extremely slow and gentle rhythm. His movements are full of infinite love and compassion. The whole picture presents a kind of secluded tranquility and harmony under the soft light and shadow, full of Zen and the warmth of life.", + "seed": 4019, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4019_dp_14_rank_28_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.015759721118956804, + "vlm_logprob": -0.015759721118956804, + "vlm_yes_no_margin": 4.55, + "vlm_token_credit": -0.015759721118956804, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0482759103178978 + ], + "opposite_logprob_tokens": [ + -3.173275947570801 + ], + "logprob_reward": -0.0482759103178978, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.0482759103178978, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.18285471200942993, + 0.054103679955005646, + 0.06557975709438324, + 0.043819867074489594, + 0.03350847214460373, + 0.04950873926281929, + 0.08612913638353348, + 0.05475694313645363, + 0.37965136766433716, + 0.050087373703718185 + ], + "credit_map": [ + 0.029171979054808617, + 0.0329580120742321, + 0.02879675105214119, + 0.02001708559691906, + 0.02023773081600666, + 0.017649469897150993, + 0.019697854295372963, + 0.014325800351798534, + 0.008018380962312222, + 0.012518621981143951, + 0.0062011005356907845, + 0.0056081293150782585, + 0.006365923676639795, + 0.006400235462933779, + 0.004560787230730057, + 0.0044304924085736275, + 0.014021658338606358, + 0.020421365275979042, + 0.006537802051752806, + 0.005020126700401306, + 0.005992332007735968, + 0.0063485861755907536, + 0.002989724976941943, + 0.004248159937560558, + 0.010204676538705826, + 0.012507723644375801, + 0.005222710780799389, + 0.0033018309623003006, + 0.002694963477551937, + 0.003549393732100725, + 0.0034295343793928623, + 0.0029090275056660175, + 0.004778697155416012, + 0.006297016981989145, + 0.003033352317288518, + 0.005924006458371878, + 0.0031153387390077114, + 0.0030010980553925037, + 0.0043303631246089935, + 0.0030285946559160948, + 0.007159830071032047, + 0.010156589560210705, + 0.004440562333911657, + 0.007996482774615288, + 0.0038314529228955507, + 0.0040889824740588665, + 0.00732537591829896, + 0.004509459249675274, + 0.01190535631030798, + 0.010641586035490036, + 0.010974261909723282, + 0.020969072356820107, + 0.006052127107977867, + 0.009239851497113705, + 0.009333335794508457, + 0.0070135388523340225, + 0.008227507583796978, + 0.007186844013631344, + 0.007909941487014294, + 0.011692715808749199, + 0.0038246805779635906, + 0.004775916691869497, + 0.00659047719091177, + 0.004548854194581509, + 0.06430120766162872, + 0.047003839164972305, + 0.056686751544475555, + 0.0841573178768158, + 0.023635409772396088, + 0.029627330601215363, + 0.045248761773109436, + 0.028990715742111206, + 0.00988239049911499, + 0.012909320183098316, + 0.006517804227769375, + 0.005445987917482853, + 0.00361958472058177, + 0.004929400980472565, + 0.0032688642386347055, + 0.0035140172112733126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 8, + 0, + 6 + ], + "selected_reward": -0.0482759103178978 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.006751937326043844 + ], + "opposite_logprob_tokens": [ + -5.006752014160156 + ], + "logprob_reward": -0.006751937326043844, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.006751937326043844, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.31937938928604126, + 0.07871665060520172, + 0.08806628733873367, + 0.04243854805827141, + 0.04536355286836624, + 0.06869381666183472, + 0.06341473013162613, + 0.14407460391521454, + 0.06623297184705734, + 0.08361943811178207 + ], + "credit_map": [ + 0.06320692598819733, + 0.06424827873706818, + 0.03633851930499077, + 0.028931541368365288, + 0.026606230065226555, + 0.035089973360300064, + 0.039041489362716675, + 0.02591640315949917, + 0.013377701863646507, + 0.012530715204775333, + 0.008285533636808395, + 0.007684973068535328, + 0.009550791233778, + 0.00857579056173563, + 0.010313434526324272, + 0.008397713303565979, + 0.015278455801308155, + 0.019188575446605682, + 0.0074633765034377575, + 0.013192565180361271, + 0.007948755286633968, + 0.010541210882365704, + 0.0073186373338103294, + 0.007134710904210806, + 0.006692887749522924, + 0.007777318358421326, + 0.0038803298957645893, + 0.006701220758259296, + 0.003946620039641857, + 0.004520865622907877, + 0.004851378034800291, + 0.004067928064614534, + 0.006566379684954882, + 0.00802009180188179, + 0.004489072598516941, + 0.006344476714730263, + 0.005108550190925598, + 0.008021802641451359, + 0.0029368516989052296, + 0.003876324510201812, + 0.011279632337391376, + 0.012293851934373379, + 0.009378314018249512, + 0.008668006397783756, + 0.006717587821185589, + 0.006849491503089666, + 0.007328695151954889, + 0.006178241688758135, + 0.01158529706299305, + 0.015366954728960991, + 0.00651796068996191, + 0.006614524405449629, + 0.00551844947040081, + 0.005995857995003462, + 0.0055723427794873714, + 0.006243341136723757, + 0.023091202601790428, + 0.02010718360543251, + 0.022233841940760612, + 0.03175164386630058, + 0.009249303489923477, + 0.011217571794986725, + 0.014673102647066116, + 0.011750766076147556, + 0.009990140795707703, + 0.010772728361189365, + 0.009118775837123394, + 0.014017051085829735, + 0.004852711223065853, + 0.006677539553493261, + 0.005745314992964268, + 0.005058710463345051, + 0.01280808262526989, + 0.015791282057762146, + 0.010406214743852615, + 0.01426021009683609, + 0.006607099436223507, + 0.007735932245850563, + 0.008963511325418949, + 0.00704710278660059 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 7, + 2 + ], + "selected_reward": -0.006751937326043844 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.00674246484413743 + ], + "opposite_logprob_tokens": [ + -5.006742477416992 + ], + "logprob_reward": -0.00674246484413743, + "yes_no_margin_reward": 5.0, + "token_credit_reward": -0.00674246484413743, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.37194088101387024, + 0.14136557281017303, + 0.08112619072198868, + 0.04928645119071007, + 0.0839187428355217, + 0.05062216892838478, + 0.051413360983133316, + 0.06183617562055588, + 0.03596508502960205, + 0.07252535223960876 + ], + "credit_map": [ + 0.06252706795930862, + 0.06480683386325836, + 0.045928120613098145, + 0.045682769268751144, + 0.03661900758743286, + 0.03954659402370453, + 0.04075596481561661, + 0.036074552685022354, + 0.021037286147475243, + 0.02070697210729122, + 0.012355862185359001, + 0.023025691509246826, + 0.012042997404932976, + 0.012443467043340206, + 0.025041373446583748, + 0.01471193041652441, + 0.015596223995089531, + 0.01293884962797165, + 0.008669811300933361, + 0.01113014668226242, + 0.007347782142460346, + 0.00888409186154604, + 0.007494248915463686, + 0.009065034799277782, + 0.009441846050322056, + 0.0073130372911691666, + 0.005280355457216501, + 0.00588766485452652, + 0.005918937735259533, + 0.004475048743188381, + 0.00636579655110836, + 0.004603764042258263, + 0.016968222334980965, + 0.021415987983345985, + 0.009771493263542652, + 0.005038760136812925, + 0.008731856942176819, + 0.011151366867125034, + 0.004690673667937517, + 0.006150387693196535, + 0.007643694058060646, + 0.00929561723023653, + 0.005640513263642788, + 0.006924956105649471, + 0.005386672914028168, + 0.004944765940308571, + 0.005728885531425476, + 0.005057062022387981, + 0.011022524908185005, + 0.01243472471833229, + 0.00566153135150671, + 0.0058549088425934315, + 0.0038319334853440523, + 0.004776821471750736, + 0.003669639816507697, + 0.004161273129284382, + 0.012059716507792473, + 0.011798448860645294, + 0.00821241457015276, + 0.009058528579771519, + 0.004780864343047142, + 0.004595332778990269, + 0.005056793335825205, + 0.006274080835282803, + 0.005405000876635313, + 0.00745640741661191, + 0.0033905620221048594, + 0.004107279237359762, + 0.0034650869201868773, + 0.003939750604331493, + 0.004255420062690973, + 0.003945577889680862, + 0.014236487448215485, + 0.017664404585957527, + 0.008557135239243507, + 0.005549425259232521, + 0.006578749977052212, + 0.007897053845226765, + 0.006077760364860296, + 0.005964337848126888 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 4 + ], + "selected_reward": -0.00674246484413743 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005952843930572271 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.005952843930572271, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005952843930572271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.16440756618976593, + 0.16138485074043274, + 0.19692280888557434, + 0.07714701443910599, + 0.03954073414206505, + 0.09111185371875763, + 0.06393709778785706, + 0.1174875795841217, + 0.0205488670617342, + 0.06751153618097305 + ], + "credit_map": [ + 0.02774461731314659, + 0.03433016315102577, + 0.016285104677081108, + 0.015519442968070507, + 0.021764051169157028, + 0.021489402279257774, + 0.011082818731665611, + 0.016191981732845306, + 0.030016276985406876, + 0.029681123793125153, + 0.02114054188132286, + 0.028509967029094696, + 0.012820572592318058, + 0.013289285823702812, + 0.014983578585088253, + 0.010943514294922352, + 0.037828512489795685, + 0.04862618446350098, + 0.020671110600233078, + 0.0207365769892931, + 0.020722204819321632, + 0.01948464848101139, + 0.013375265523791313, + 0.01547832041978836, + 0.017000434920191765, + 0.02173125557601452, + 0.008324679918587208, + 0.007217904087156057, + 0.0061195287853479385, + 0.006852285470813513, + 0.004249659366905689, + 0.005651281215250492, + 0.00749443331733346, + 0.009055729024112225, + 0.004231967963278294, + 0.0060934568755328655, + 0.002864727983251214, + 0.0041553787887096405, + 0.0023656992707401514, + 0.0032793451100587845, + 0.01764995977282524, + 0.02387281134724617, + 0.009888439439237118, + 0.008883084170520306, + 0.008206242695450783, + 0.008839933201670647, + 0.006788772530853748, + 0.0069826156832277775, + 0.011294668540358543, + 0.010559489950537682, + 0.008237245492637157, + 0.012855537235736847, + 0.0043053035624325275, + 0.005269087851047516, + 0.006425628438591957, + 0.00499014463275671, + 0.022058915346860886, + 0.017503298819065094, + 0.018342245370149612, + 0.022907577455043793, + 0.00804173108190298, + 0.008586504496634007, + 0.011480605229735374, + 0.008566721342504025, + 0.0031451424583792686, + 0.0053270841017365456, + 0.002139680553227663, + 0.0018292113672941923, + 0.0020968811586499214, + 0.0024316394701600075, + 0.0014169138157740235, + 0.0021623168140649796, + 0.012337576597929, + 0.014260686933994293, + 0.008650310337543488, + 0.009206799790263176, + 0.006294589955359697, + 0.006124255713075399, + 0.0062500531785190105, + 0.004387269262224436 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 2, + 0, + 1 + ], + "selected_reward": -0.005952843930572271 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011075449176132679 + ], + "opposite_logprob_tokens": [ + -4.511075496673584 + ], + "logprob_reward": -0.011075449176132679, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011075449176132679, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2660652995109558, + 0.07073421776294708, + 0.028739921748638153, + 0.044889189302921295, + 0.08870534598827362, + 0.07431522011756897, + 0.06948567926883698, + 0.14821620285511017, + 0.1690666824579239, + 0.039782192558050156 + ], + "credit_map": [ + 0.044056061655282974, + 0.056555334478616714, + 0.02574547752737999, + 0.03169303014874458, + 0.024277908727526665, + 0.028374282643198967, + 0.026956936344504356, + 0.028406279161572456, + 0.012881818227469921, + 0.015778673812747, + 0.006910885684192181, + 0.005731379147619009, + 0.007627564482390881, + 0.006321292836219072, + 0.008914470672607422, + 0.00656813383102417, + 0.004679393954575062, + 0.0071197934448719025, + 0.0029430254362523556, + 0.003846309380605817, + 0.0023998271208256483, + 0.002926139160990715, + 0.002173778833821416, + 0.00265165395103395, + 0.006021762266755104, + 0.010594271123409271, + 0.004019663669168949, + 0.004740850999951363, + 0.004304202739149332, + 0.006781760137528181, + 0.004610401578247547, + 0.003816273994743824, + 0.013744979165494442, + 0.02183307707309723, + 0.008528266102075577, + 0.010902614332735538, + 0.008469642139971256, + 0.007787151727825403, + 0.009507646784186363, + 0.007931962609291077, + 0.01440869178622961, + 0.012039689347147942, + 0.011853978037834167, + 0.011291256174445152, + 0.0060044447891414165, + 0.0069436440244317055, + 0.00581421097740531, + 0.005959298461675644, + 0.00933406874537468, + 0.011370123364031315, + 0.007862195372581482, + 0.015994690358638763, + 0.005076895002275705, + 0.0070977904833853245, + 0.007257369346916676, + 0.005492544732987881, + 0.029781635850667953, + 0.01604442112147808, + 0.025869246572256088, + 0.030519288033246994, + 0.008610912598669529, + 0.010528497397899628, + 0.014292076230049133, + 0.01257011666893959, + 0.029288867488503456, + 0.018262682482600212, + 0.025261854752898216, + 0.03814137727022171, + 0.010800477117300034, + 0.014318619854748249, + 0.019992182031273842, + 0.013000630773603916, + 0.007385867647826672, + 0.008829151280224323, + 0.0050328392535448074, + 0.003018498420715332, + 0.003125094110146165, + 0.0065180640667676926, + 0.002709185006096959, + 0.003163490677252412 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 8, + 7 + ], + "selected_reward": -0.011075449176132679 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, captures a moment of tranquility and warmth on a warm afternoon. In a quiet room within an ancient Taoist temple, an elderly Taoist priest, his hair and beard white, sits cross-legged on a thick, woven cattail mat. He wears a plain, washed-out, indigo robe. A beam of warm, golden sunlight slants in through a nearby carved wooden window, illuminating the slowly dancing dust particles in the air. The old Taoist's eyes are slightly closed, a serene, contented smile on his age-marked face. His other hand rests naturally on his knee. In his arms, a plump, orange tabby cat curls up in a ball, dozing soundly, its body rising and falling slightly with its even breathing. The old Taoist priest's hand, covered with wrinkles and with distinct knuckles, is stroking the cat from the top of the head all the way down its back in an extremely slow and gentle rhythm. His movements are full of infinite love and compassion. The whole picture presents a kind of secluded tranquility and harmony under the soft light and shadow, full of Zen and the warmth of life.", + "seed": 4020, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4020_dp_15_rank_30_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.022418915131129324, + "vlm_logprob": -0.022418915131129324, + "vlm_yes_no_margin": 5.05, + "vlm_token_credit": -0.022418915131129324, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.08758450299501419 + ], + "opposite_logprob_tokens": [ + -2.7125844955444336 + ], + "logprob_reward": -0.08758450299501419, + "yes_no_margin_reward": 2.625, + "token_credit_reward": -0.08758450299501419, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.26433926820755005, + 0.20912779867649078, + 0.0535905696451664, + 0.17096613347530365, + 0.05514504760503769, + 0.04470803961157799, + 0.05306754261255264, + 0.04339727386832237, + 0.04396047815680504, + 0.06169784441590309 + ], + "credit_map": [ + 0.02749285101890564, + 0.03184659779071808, + 0.034986019134521484, + 0.023211196064949036, + 0.05686710402369499, + 0.034599419683218, + 0.02054096944630146, + 0.03479510918259621, + 0.015365910716354847, + 0.018725140020251274, + 0.04707968607544899, + 0.03168969973921776, + 0.023524397984147072, + 0.0316050685942173, + 0.012607255019247532, + 0.02853064425289631, + 0.006210698746144772, + 0.005889939609915018, + 0.0071663386188447475, + 0.01161155104637146, + 0.006948988419026136, + 0.00562469894066453, + 0.004123660735785961, + 0.006014695391058922, + 0.014621615409851074, + 0.014992604032158852, + 0.03135499358177185, + 0.02805047668516636, + 0.021097617223858833, + 0.02442163974046707, + 0.01173750776797533, + 0.0246896892786026, + 0.006344284396618605, + 0.006089902948588133, + 0.008251122198998928, + 0.00497008953243494, + 0.008698903024196625, + 0.008162197656929493, + 0.004714392591267824, + 0.007914155721664429, + 0.004487909376621246, + 0.005052133928984404, + 0.006574308965355158, + 0.004657921846956015, + 0.006043188739567995, + 0.007177564315497875, + 0.004421023651957512, + 0.006293990649282932, + 0.003452497301623225, + 0.006228683050721884, + 0.006025124806910753, + 0.004384747240692377, + 0.014337943866848946, + 0.007526131812483072, + 0.004766079131513834, + 0.006346337962895632, + 0.00414230627939105, + 0.005120177753269672, + 0.005361283663660288, + 0.005920208990573883, + 0.0054136160761117935, + 0.006379743572324514, + 0.0036734126042574644, + 0.007386526558548212, + 0.004416401032358408, + 0.005992763210088015, + 0.004787752870470285, + 0.005008482839912176, + 0.006686825770884752, + 0.007215752732008696, + 0.004258587956428528, + 0.005593912675976753, + 0.003996295388787985, + 0.005327239632606506, + 0.004637708887457848, + 0.004725124221295118, + 0.003516125725582242, + 0.018585005775094032, + 0.002452987479045987, + 0.018457362428307533 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.08758450299501419 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011104214936494827 + ], + "opposite_logprob_tokens": [ + -4.511104106903076 + ], + "logprob_reward": -0.011104214936494827, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011104214936494827, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.278169721364975, + 0.16029489040374756, + 0.09004420042037964, + 0.0672084391117096, + 0.056496769189834595, + 0.07324311137199402, + 0.05356761813163757, + 0.06625368446111679, + 0.07060449570417404, + 0.08411704748868942 + ], + "credit_map": [ + 0.020875530317425728, + 0.029367035254836082, + 0.04160814732313156, + 0.029137054458260536, + 0.05441880226135254, + 0.041303280740976334, + 0.024176843464374542, + 0.037283021956682205, + 0.010561482980847359, + 0.01454936247318983, + 0.028277352452278137, + 0.01321855653077364, + 0.037651583552360535, + 0.025769390165805817, + 0.011567976325750351, + 0.01869918592274189, + 0.0068840221501886845, + 0.009822878986597061, + 0.014809482730925083, + 0.020759141072630882, + 0.010571065358817577, + 0.010901832953095436, + 0.005213893949985504, + 0.011081887409090996, + 0.005469923373311758, + 0.006019745487719774, + 0.007241071201860905, + 0.007075728382915258, + 0.019580675289034843, + 0.009115420281887054, + 0.004456646740436554, + 0.00824922975152731, + 0.003496806602925062, + 0.005648468155413866, + 0.006518969777971506, + 0.0046739461831748486, + 0.015805359929800034, + 0.008217121474444866, + 0.004646634217351675, + 0.00748946238309145, + 0.005166448187083006, + 0.00673925643786788, + 0.010709629394114017, + 0.004373342730104923, + 0.019511474296450615, + 0.011868965812027454, + 0.0056897373870015144, + 0.009184262715280056, + 0.004462538752704859, + 0.004591078031808138, + 0.006928779184818268, + 0.0057265483774244785, + 0.010806183330714703, + 0.008848479948937893, + 0.005371715873479843, + 0.0068322946317493916, + 0.00417277030646801, + 0.005454182159155607, + 0.007286636158823967, + 0.00525765772908926, + 0.014353319071233273, + 0.012717774137854576, + 0.008416097611188889, + 0.00859525054693222, + 0.005825916770845652, + 0.004886690992861986, + 0.012488067150115967, + 0.00586027093231678, + 0.008828185498714447, + 0.014192592352628708, + 0.008504941128194332, + 0.010017831809818745, + 0.0053948876447975636, + 0.004220846574753523, + 0.008777172304689884, + 0.004883928224444389, + 0.008321985602378845, + 0.024762701243162155, + 0.006399599835276604, + 0.021355926990509033 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.011104214936494827 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0015185264637693763 + ], + "opposite_logprob_tokens": [ + -6.501518726348877 + ], + "logprob_reward": -0.0015185264637693763, + "yes_no_margin_reward": 6.5, + "token_credit_reward": -0.0015185264637693763, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11883838474750519, + 0.08753333240747452, + 0.16191014647483826, + 0.3286462426185608, + 0.0969652384519577, + 0.033896904438734055, + 0.03148068115115166, + 0.042213696986436844, + 0.03390734642744064, + 0.06460800021886826 + ], + "credit_map": [ + 0.011093996465206146, + 0.013134056702256203, + 0.013817660510540009, + 0.014517761766910553, + 0.02427764981985092, + 0.016168328002095222, + 0.010303677059710026, + 0.015525243245065212, + 0.0059739272110164165, + 0.009345890022814274, + 0.014342941343784332, + 0.012419076636433601, + 0.010616494342684746, + 0.012875675223767757, + 0.0103740394115448, + 0.011585288681089878, + 0.01273291278630495, + 0.016826968640089035, + 0.020415714010596275, + 0.027111107483506203, + 0.023962052538990974, + 0.019726881757378578, + 0.014271915890276432, + 0.02686259150505066, + 0.024109184741973877, + 0.026384970173239708, + 0.05330410972237587, + 0.06216028705239296, + 0.026694219559431076, + 0.053514525294303894, + 0.022027285769581795, + 0.06045164912939072, + 0.010141843929886818, + 0.009111562743782997, + 0.01893775165081024, + 0.015946026891469955, + 0.007954839617013931, + 0.013313692063093185, + 0.0070266141556203365, + 0.014532906003296375, + 0.0033008428290486336, + 0.0031511674169451, + 0.008086142130196095, + 0.005711423698812723, + 0.002993034664541483, + 0.0044887918047606945, + 0.0022737046238034964, + 0.003891798434779048, + 0.0025435322895646095, + 0.002338320715352893, + 0.003986922558397055, + 0.005869455169886351, + 0.003606473095715046, + 0.005283932201564312, + 0.0034651828464120626, + 0.004386860877275467, + 0.003555700648576021, + 0.0039982362650334835, + 0.0046131303533911705, + 0.00699625862762332, + 0.004545923788100481, + 0.0068835425190627575, + 0.004839893896132708, + 0.006781006697565317, + 0.004704367369413376, + 0.002785037737339735, + 0.006147085689008236, + 0.0028632436878979206, + 0.0029524057172238827, + 0.006064525339752436, + 0.0026366361416876316, + 0.005754046142101288, + 0.0034661423414945602, + 0.0034352350048720837, + 0.006045423913747072, + 0.0025195488706231117, + 0.0033906716853380203, + 0.02148757129907608, + 0.0030003483407199383, + 0.021263062953948975 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 3, + 2, + 0 + ], + "selected_reward": -0.0015185264637693763 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0008156548719853163 + ], + "opposite_logprob_tokens": [ + -7.1258158683776855 + ], + "logprob_reward": -0.0008156548719853163, + "yes_no_margin_reward": 7.125, + "token_credit_reward": -0.0008156548719853163, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.27444955706596375, + 0.12809540331363678, + 0.08473143726587296, + 0.20762695372104645, + 0.068221814930439, + 0.07910818606615067, + 0.03873990848660469, + 0.032460518181324005, + 0.03155817091464996, + 0.05500801280140877 + ], + "credit_map": [ + 0.03138301521539688, + 0.047456298023462296, + 0.042250856757164, + 0.03175624832510948, + 0.032113779336214066, + 0.03428890183568001, + 0.01895911619067192, + 0.036241382360458374, + 0.008866315707564354, + 0.012853466905653477, + 0.017832471057772636, + 0.022472554817795753, + 0.013032493181526661, + 0.01975935511291027, + 0.010856336914002895, + 0.0224224254488945, + 0.006961737759411335, + 0.008324225433170795, + 0.00994572788476944, + 0.022834274917840958, + 0.006206070072948933, + 0.010217308066785336, + 0.006158602889627218, + 0.014083502814173698, + 0.018154582008719444, + 0.016533708199858665, + 0.04031245410442352, + 0.03607833385467529, + 0.022800631821155548, + 0.02998506650328636, + 0.014739837497472763, + 0.02902236208319664, + 0.005523444153368473, + 0.008668852970004082, + 0.009264258667826653, + 0.005872663576155901, + 0.012981279753148556, + 0.009590569883584976, + 0.005281682591885328, + 0.01103908009827137, + 0.00612643314525485, + 0.010692408308386803, + 0.009158912114799023, + 0.008829331956803799, + 0.01270454004406929, + 0.0115415183827281, + 0.005219221115112305, + 0.01483582891523838, + 0.003319916781038046, + 0.004081759601831436, + 0.004643226973712444, + 0.00368120358325541, + 0.00899114366620779, + 0.004826181102544069, + 0.0029520555399358273, + 0.006244424264878035, + 0.0023440944496542215, + 0.0023244055919349194, + 0.00475601851940155, + 0.006208580452948809, + 0.0024496831465512514, + 0.004812376573681831, + 0.0027226926758885384, + 0.0068426718935370445, + 0.0025609631557017565, + 0.0025855633430182934, + 0.0047037191689014435, + 0.0040781195275485516, + 0.0031104779336601496, + 0.005448101554065943, + 0.0020334303844720125, + 0.007037802133709192, + 0.0038596868980675936, + 0.003192607080563903, + 0.004734861198812723, + 0.0031840617302805185, + 0.005527766887098551, + 0.016250023618340492, + 0.002243373543024063, + 0.016015637665987015 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 1 + ], + "selected_reward": -0.0008156548719853163 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.011071676388382912 + ], + "opposite_logprob_tokens": [ + -4.511071681976318 + ], + "logprob_reward": -0.011071676388382912, + "yes_no_margin_reward": 4.5, + "token_credit_reward": -0.011071676388382912, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21257151663303375, + 0.07747136056423187, + 0.07704722881317139, + 0.18700361251831055, + 0.10526414215564728, + 0.056997280567884445, + 0.04393530264496803, + 0.06541158258914948, + 0.057518281042575836, + 0.11677972972393036 + ], + "credit_map": [ + 0.019754990935325623, + 0.041259776800870895, + 0.03062910959124565, + 0.021194828674197197, + 0.02643527276813984, + 0.028299229219555855, + 0.019643135368824005, + 0.0253551434725523, + 0.007616337854415178, + 0.011182140558958054, + 0.009327612817287445, + 0.008581718429923058, + 0.009011602960526943, + 0.011306419037282467, + 0.007031797431409359, + 0.013413723558187485, + 0.005410308018326759, + 0.008020160719752312, + 0.013790478929877281, + 0.007517480291426182, + 0.014511591754853725, + 0.013058945536613464, + 0.005416675470769405, + 0.009321575053036213, + 0.016916999593377113, + 0.02122405357658863, + 0.028730735182762146, + 0.02464486099779606, + 0.03370664268732071, + 0.022867955267429352, + 0.014040162786841393, + 0.024872181937098503, + 0.011164788156747818, + 0.011792127974331379, + 0.013939032331109047, + 0.009260640479624271, + 0.023088769987225533, + 0.015159322880208492, + 0.007698607165366411, + 0.013160845264792442, + 0.002692410722374916, + 0.004684648476541042, + 0.006666668225079775, + 0.004874613601714373, + 0.011042690835893154, + 0.010368824936449528, + 0.0044359262101352215, + 0.012231491506099701, + 0.0032656819093972445, + 0.004177918657660484, + 0.004799738060683012, + 0.00428386777639389, + 0.008396636694669724, + 0.00685363681986928, + 0.004970403388142586, + 0.0071874153800308704, + 0.004606735426932573, + 0.006045419257134199, + 0.005642939358949661, + 0.009608443826436996, + 0.012235390953719616, + 0.011120776645839214, + 0.007171944249421358, + 0.008979929611086845, + 0.0032104423735290766, + 0.003586989128962159, + 0.007038106210529804, + 0.0036598630249500275, + 0.00872712954878807, + 0.01523654069751501, + 0.005655196495354176, + 0.010404009371995926, + 0.004624688532203436, + 0.002852914622053504, + 0.004791542422026396, + 0.0034860132727771997, + 0.008311202749609947, + 0.04394743964076042, + 0.004983886610716581, + 0.043782029300928116 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 9 + ], + "selected_reward": -0.011071676388382912 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "static, mid-shot shot, taken from a level angle, captures a moment of tranquility and warmth on a warm afternoon. In a quiet room within an ancient Taoist temple, an elderly Taoist priest, his hair and beard white, sits cross-legged on a thick, woven cattail mat. He wears a plain, washed-out, indigo robe. A beam of warm, golden sunlight slants in through a nearby carved wooden window, illuminating the slowly dancing dust particles in the air. The old Taoist's eyes are slightly closed, a serene, contented smile on his age-marked face. His other hand rests naturally on his knee. In his arms, a plump, orange tabby cat curls up in a ball, dozing soundly, its body rising and falling slightly with its even breathing. The old Taoist priest's hand, covered with wrinkles and with distinct knuckles, is stroking the cat from the top of the head all the way down its back in an extremely slow and gentle rhythm. His movements are full of infinite love and compassion. The whole picture presents a kind of secluded tranquility and harmony under the soft light and shadow, full of Zen and the warmth of life.", + "seed": 4021, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_7/index_3917_seed_4021_dp_15_rank_30_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 7, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.01983049705158919, + "vlm_logprob": -0.01983049705158919, + "vlm_yes_no_margin": 4.45, + "vlm_token_credit": -0.01983049705158919, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.04975660890340805 + ], + "opposite_logprob_tokens": [ + -3.1747565269470215 + ], + "logprob_reward": -0.04975660890340805, + "yes_no_margin_reward": 3.125, + "token_credit_reward": -0.04975660890340805, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23618696630001068, + 0.10568928718566895, + 0.1165461465716362, + 0.09043730795383453, + 0.09995007514953613, + 0.15241898596286774, + 0.06511953473091125, + 0.03804033622145653, + 0.05072804167866707, + 0.04488329589366913 + ], + "credit_map": [ + 0.02978380396962166, + 0.03779149428009987, + 0.03547180816531181, + 0.02930273301899433, + 0.024935029447078705, + 0.028494345024228096, + 0.019871970638632774, + 0.03053578920662403, + 0.013371163047850132, + 0.008904848247766495, + 0.010770265944302082, + 0.023774228990077972, + 0.010286620818078518, + 0.00977925956249237, + 0.014362519606947899, + 0.014440376311540604, + 0.011015194468200207, + 0.015888327732682228, + 0.016824839636683464, + 0.023489682003855705, + 0.01210299227386713, + 0.010091440752148628, + 0.01333039440214634, + 0.013803278096020222, + 0.00883398111909628, + 0.014201730489730835, + 0.014758353121578693, + 0.015074850991368294, + 0.008929885923862457, + 0.008277998305857182, + 0.008469917811453342, + 0.011890583671629429, + 0.01154317893087864, + 0.018720140680670738, + 0.01324600912630558, + 0.01468595676124096, + 0.008051535114645958, + 0.009766186587512493, + 0.009931976906955242, + 0.014005087316036224, + 0.021502865478396416, + 0.025540705770254135, + 0.025174491107463837, + 0.019588520750403404, + 0.01781703345477581, + 0.008994406089186668, + 0.011999525129795074, + 0.021801438182592392, + 0.009707771241664886, + 0.010023375041782856, + 0.010237965732812881, + 0.007855554111301899, + 0.005565769504755735, + 0.005740144290030003, + 0.005661700386554003, + 0.010327255353331566, + 0.004605857189744711, + 0.005848403088748455, + 0.00532919354736805, + 0.0049403076991438866, + 0.004699711222201586, + 0.004031335934996605, + 0.004043162800371647, + 0.004542365204542875, + 0.0062063513323664665, + 0.008953348733484745, + 0.007127073593437672, + 0.007469827309250832, + 0.004634684883058071, + 0.004808740224689245, + 0.004253889434039593, + 0.007274126634001732, + 0.004792429506778717, + 0.0055802264250814915, + 0.008016985841095448, + 0.00628420477733016, + 0.0031956485472619534, + 0.005361925810575485, + 0.004330520052462816, + 0.007321352604776621 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.04975660890340805 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01821574941277504 + ], + "opposite_logprob_tokens": [ + -4.018215656280518 + ], + "logprob_reward": -0.01821574941277504, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01821574941277504, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3061683475971222, + 0.07787904143333435, + 0.11857293546199799, + 0.14418169856071472, + 0.05813013017177582, + 0.058103084564208984, + 0.05587064474821091, + 0.026400074362754822, + 0.11838588118553162, + 0.036308154463768005 + ], + "credit_map": [ + 0.032065797597169876, + 0.03613828122615814, + 0.03649488091468811, + 0.05501735210418701, + 0.03176161274313927, + 0.04071378707885742, + 0.034122221171855927, + 0.039854440838098526, + 0.01309120748192072, + 0.01070486381649971, + 0.01347469724714756, + 0.006727204192429781, + 0.009731912985444069, + 0.007277356460690498, + 0.006705154199153185, + 0.010166648775339127, + 0.013474937528371811, + 0.01404056791216135, + 0.020826419815421104, + 0.020219560712575912, + 0.010713104158639908, + 0.011470332741737366, + 0.013370804488658905, + 0.014457209035754204, + 0.009469829499721527, + 0.015312423929572105, + 0.020264601334929466, + 0.033155523240566254, + 0.012308684177696705, + 0.01584821380674839, + 0.018971027806401253, + 0.018851397559046745, + 0.007093416061252356, + 0.008929282426834106, + 0.008911649696528912, + 0.008045081049203873, + 0.006918987724930048, + 0.006614289246499538, + 0.004712078254669905, + 0.006905348505824804, + 0.009256616234779358, + 0.005777379963546991, + 0.00653728237375617, + 0.008162355050444603, + 0.009302270598709583, + 0.005146307405084372, + 0.007182122208178043, + 0.006738749798387289, + 0.005859141703695059, + 0.0048936037346720695, + 0.006608427502214909, + 0.00885145366191864, + 0.008230110630393028, + 0.006556189619004726, + 0.007965737022459507, + 0.006905979942530394, + 0.0026002032682299614, + 0.0035815141163766384, + 0.00362908816896379, + 0.004533452447503805, + 0.0031388388015329838, + 0.002664739964529872, + 0.003089743433520198, + 0.003162493696436286, + 0.008353560231626034, + 0.012070941738784313, + 0.02007758617401123, + 0.027274511754512787, + 0.007531203795224428, + 0.012592490762472153, + 0.013827315531671047, + 0.01665826141834259, + 0.003751072334125638, + 0.004250765778124332, + 0.006237237714231014, + 0.005089023150503635, + 0.003288621548563242, + 0.004987342283129692, + 0.00362519477494061, + 0.0050788940861821175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 2 + ], + "selected_reward": -0.01821574941277504 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005956398788839579 + ], + "opposite_logprob_tokens": [ + -5.130956172943115 + ], + "logprob_reward": -0.005956398788839579, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005956398788839579, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21995285153388977, + 0.1507994681596756, + 0.2022514045238495, + 0.107183076441288, + 0.03474632278084755, + 0.03776056692004204, + 0.042324312031269073, + 0.031729310750961304, + 0.11818639934062958, + 0.05506625398993492 + ], + "credit_map": [ + 0.026884466409683228, + 0.01895139180123806, + 0.030265333130955696, + 0.03983500599861145, + 0.025979219004511833, + 0.02043934166431427, + 0.025617798790335655, + 0.03198030963540077, + 0.020328694954514503, + 0.021010445430874825, + 0.02306702546775341, + 0.023792684078216553, + 0.01798851601779461, + 0.010899463668465614, + 0.01569785736501217, + 0.018014783039689064, + 0.024892618879675865, + 0.021277427673339844, + 0.030358340591192245, + 0.04158203303813934, + 0.01285971887409687, + 0.01839900203049183, + 0.025352858006954193, + 0.027529427781701088, + 0.00896065030246973, + 0.009947258047759533, + 0.01695563830435276, + 0.021361881867051125, + 0.011943381279706955, + 0.011936075985431671, + 0.012676392681896687, + 0.01340180542320013, + 0.005603186320513487, + 0.004774598404765129, + 0.006018668878823519, + 0.003922394011169672, + 0.00410482008010149, + 0.003646750468760729, + 0.0023524665739387274, + 0.004323439206928015, + 0.004426927771419287, + 0.004747116006910801, + 0.005535077769309282, + 0.004677740857005119, + 0.005633169319480658, + 0.004006568808108568, + 0.004074974451214075, + 0.0046589989215135574, + 0.004824971780180931, + 0.005152702797204256, + 0.006387804169207811, + 0.005781035404652357, + 0.004920098464936018, + 0.005412640515714884, + 0.004217536188662052, + 0.005627528764307499, + 0.0030084147583693266, + 0.0035025514662265778, + 0.005723596550524235, + 0.004557168576866388, + 0.003658249508589506, + 0.0031522007193416357, + 0.0035425627138465643, + 0.0045845708809792995, + 0.008768703788518906, + 0.012361641973257065, + 0.018396714702248573, + 0.029255088418722153, + 0.006761773955076933, + 0.010618166998028755, + 0.014842567034065723, + 0.017181750386953354, + 0.006081882398575544, + 0.005885316990315914, + 0.007402198389172554, + 0.01180342212319374, + 0.005309402942657471, + 0.006085712928324938, + 0.005696549080312252, + 0.006801774259656668 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 2, + 1 + ], + "selected_reward": -0.005956398788839579 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.001942058908753097 + ], + "opposite_logprob_tokens": [ + -6.251942157745361 + ], + "logprob_reward": -0.001942058908753097, + "yes_no_margin_reward": 6.25, + "token_credit_reward": -0.001942058908753097, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.21218474209308624, + 0.25752317905426025, + 0.10280017554759979, + 0.11635962873697281, + 0.0835592970252037, + 0.05970228835940361, + 0.05855514854192734, + 0.048675164580345154, + 0.03010508045554161, + 0.030535263940691948 + ], + "credit_map": [ + 0.0343158096075058, + 0.0354514941573143, + 0.032361309975385666, + 0.029147664085030556, + 0.017397044226527214, + 0.021160872653126717, + 0.014890840277075768, + 0.027459703385829926, + 0.038910988718271255, + 0.048750340938568115, + 0.04435242339968681, + 0.03523854911327362, + 0.022061588242650032, + 0.015287907794117928, + 0.019276872277259827, + 0.03364452347159386, + 0.010330610908567905, + 0.024003738537430763, + 0.016605453565716743, + 0.013045912608504295, + 0.009461632929742336, + 0.00849893968552351, + 0.008291165344417095, + 0.012562722899019718, + 0.011940651573240757, + 0.023761294782161713, + 0.01972093991935253, + 0.017047440633177757, + 0.01008530706167221, + 0.00820003729313612, + 0.009344255551695824, + 0.016259701922535896, + 0.01724439114332199, + 0.013541841879487038, + 0.009997330605983734, + 0.011493238620460033, + 0.00661963177844882, + 0.005710242781788111, + 0.007154963910579681, + 0.011797654442489147, + 0.010000301524996758, + 0.009770178236067295, + 0.008298435248434544, + 0.007885250262916088, + 0.006647111382335424, + 0.004417875315994024, + 0.004893111996352673, + 0.00779002346098423, + 0.007817384786903858, + 0.00841846875846386, + 0.009957815520465374, + 0.005729698576033115, + 0.007438391447067261, + 0.006096459925174713, + 0.005495239980518818, + 0.007601686753332615, + 0.0058972337283194065, + 0.009482243098318577, + 0.007716539781540632, + 0.0058868201449513435, + 0.005069939885288477, + 0.004478973802179098, + 0.004549636505544186, + 0.005593782290816307, + 0.003833630122244358, + 0.005893824156373739, + 0.004940140061080456, + 0.0041602663695812225, + 0.0023744492791593075, + 0.0025179393123835325, + 0.0025505386292934418, + 0.0038342925254255533, + 0.0036577677819877863, + 0.004481876268982887, + 0.0049141631461679935, + 0.0044977315701544285, + 0.0023792656138539314, + 0.003272922709584236, + 0.0031695032957941294, + 0.00416203448548913 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.001942058908753097 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the scene set inside a quiet room of an ancient Taoist temple?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is an elderly Taoist priest with white hair and beard sitting cross-legged on a thick woven cattail mat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the priest wearing a plain, washed-out indigo robe while stroking a plump orange tabby cat?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Is a beam of warm golden sunlight slanting through a carved wooden window to illuminate dancing dust particles?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is the camera shot static, mid-shot, and taken from a level angle?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02328166924417019 + ], + "opposite_logprob_tokens": [ + -3.7732815742492676 + ], + "logprob_reward": -0.02328166924417019, + "yes_no_margin_reward": 3.75, + "token_credit_reward": -0.02328166924417019, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.3024231493473053, + 0.18863044679164886, + 0.08296564966440201, + 0.11259841173887253, + 0.09978669881820679, + 0.03759026154875755, + 0.0498088039457798, + 0.04288506880402565, + 0.044078484177589417, + 0.03923305869102478 + ], + "credit_map": [ + 0.04506731033325195, + 0.05734460428357124, + 0.04114781692624092, + 0.028308851644396782, + 0.02752496674656868, + 0.03628421947360039, + 0.027483684942126274, + 0.03926171734929085, + 0.039788659662008286, + 0.025157159194350243, + 0.02851703390479088, + 0.0209256149828434, + 0.015104757621884346, + 0.016345569863915443, + 0.018565228208899498, + 0.024226419627666473, + 0.0128885917365551, + 0.008576111868023872, + 0.01043578702956438, + 0.016437675803899765, + 0.006392715033143759, + 0.008312821388244629, + 0.009899399243295193, + 0.010022553615272045, + 0.01401387620717287, + 0.012248245067894459, + 0.016449330374598503, + 0.0215830709785223, + 0.008697687648236752, + 0.012106086127460003, + 0.01252709049731493, + 0.014973022975027561, + 0.015116121619939804, + 0.01354148332029581, + 0.014153811149299145, + 0.01400881726294756, + 0.009177097119390965, + 0.011866576969623566, + 0.009766797535121441, + 0.012155992910265923, + 0.00478409556671977, + 0.00566685339435935, + 0.004647626541554928, + 0.005995670333504677, + 0.003106331918388605, + 0.003816730808466673, + 0.004458698909729719, + 0.005114255007356405, + 0.008863129653036594, + 0.005314133130013943, + 0.008247483521699905, + 0.0072053950279951096, + 0.0042929318733513355, + 0.004924188833683729, + 0.004475914407521486, + 0.006485628429800272, + 0.005888587795197964, + 0.006616838742047548, + 0.0062940288335084915, + 0.005487984512001276, + 0.0038504856638610363, + 0.004428975284099579, + 0.004886954557150602, + 0.005431215278804302, + 0.00749175064265728, + 0.0049024769105017185, + 0.004812013823539019, + 0.007139296270906925, + 0.0058290413580834866, + 0.004438626579940319, + 0.00360863353125751, + 0.005856646224856377, + 0.006413745693862438, + 0.005334359593689442, + 0.00464927451685071, + 0.004567275755107403, + 0.005201008170843124, + 0.0032216496765613556, + 0.004683761391788721, + 0.005161984357982874 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 3 + ], + "selected_reward": -0.02328166924417019 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file diff --git a/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_tensor([2887, 2887])_group_9_rewards.json b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_tensor([2887, 2887])_group_9_rewards.json new file mode 100644 index 0000000000000000000000000000000000000000..7a8834bb16085168b6ae34e3cec3b551558de504 --- /dev/null +++ b/outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_tensor([2887, 2887])_group_9_rewards.json @@ -0,0 +1,2622 @@ +[ + { + "prompt": "dynamic, high-speed tracking shot, shot from above, follows a figure as it journeys through a vast expanse of emerald green bamboo. Dazzling golden sunlight streams from deep within the frame, creating a strong backlight effect that cuts the figure of the knight into a near-black silhouette. He wears a loose white silk robe, its long sleeves and sleeves billowing and flowing like clouds as he moves at full speed. His movements are incredibly swift, seemingly weightless, as he breezes past towering bamboos with a fluid, flowing grace. Each touch of his toes, like a dragonfly skimming the water, only slightly bends the tough bamboo tips. The intense light frames his entire form, from his flowing hair to the fluttering hem of his robe, with a distinct, dazzling golden edge, making him appear divine and distinct against the deep green background. Light filters through the gaps between the bamboo leaves, casting a dappled halo across the lens, imbuing the image with the grandeur and artistic atmosphere of a martial arts epic.", + "seed": 10385, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_2887_seed_10385_dp_18_rank_36_sample_0_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 0, + "avg": -0.3992248409427702, + "vlm_logprob": -0.3992248409427702, + "vlm_yes_no_margin": 2.45, + "vlm_token_credit": -0.3992248409427702, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.6268328428268433 + ], + "opposite_logprob_tokens": [ + -0.25183284282684326 + ], + "logprob_reward": -1.6268328428268433, + "yes_no_margin_reward": -1.375, + "token_credit_reward": -1.6268328428268433, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.022207822650671005, + 0.8903406262397766, + 0.017454136162996292, + 0.008195680566132069, + 0.031056692823767662, + 0.004782561678439379, + 0.011413773521780968, + 0.004354456439614296, + 0.004578291904181242, + 0.005615918431431055 + ], + "credit_map": [ + 0.001033116364851594, + 0.0015542368637397885, + 0.001185410306788981, + 0.0021806471049785614, + 0.005509836599230766, + 0.0012798466486856341, + 0.006706729531288147, + 0.0027580028399825096, + 0.015439813025295734, + 0.019157441332936287, + 0.030779268592596054, + 0.01301037147641182, + 0.08107639849185944, + 0.37894079089164734, + 0.03573848679661751, + 0.3161981403827667, + 0.0019289524061605334, + 0.0012339446693658829, + 0.0025187518913298845, + 0.0031096336897462606, + 0.0011198394931852818, + 0.002159920521080494, + 0.003251962596550584, + 0.002131132408976555, + 0.0007057078764773905, + 0.0013911568094044924, + 0.0006865827017463744, + 0.0006942783365957439, + 0.0012384315486997366, + 0.0009949853410944343, + 0.0007583474507555366, + 0.00172619114164263, + 0.002634283620864153, + 0.004754654131829739, + 0.002887419192120433, + 0.005780905485153198, + 0.004029966425150633, + 0.003994215279817581, + 0.003423953428864479, + 0.003551298985257745, + 0.00042906907037831843, + 0.00047227562754414976, + 0.0004517513152677566, + 0.0010918988846242428, + 0.00047841545892879367, + 0.0007063503144308925, + 0.0006927810027264059, + 0.00046002003364264965, + 0.002026517875492573, + 0.001288555795326829, + 0.0014704059576615691, + 0.001035100780427456, + 0.0015616848831996322, + 0.0011641174787655473, + 0.0018795727519318461, + 0.0009878188138827682, + 0.0006681237136945128, + 0.000373714865418151, + 0.0009060055599547923, + 0.0005834919866174459, + 0.0004430651606526226, + 0.0004503087839111686, + 0.0004900562926195562, + 0.00043969068792648613, + 0.0008225831552408636, + 0.00045735074672847986, + 0.0009148340323008597, + 0.0005316259921528399, + 0.00039226977969519794, + 0.0003209189744666219, + 0.0006972562987357378, + 0.00044145301217213273, + 0.000528806762304157, + 0.000541423331014812, + 0.0013070181012153625, + 0.000449583021691069, + 0.0007446938543580472, + 0.0010818863520398736, + 0.0004155563365202397, + 0.0005469513707794249 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 4, + 0 + ], + "selected_reward": -1.6268328428268433 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0161101296544075 + ], + "opposite_logprob_tokens": [ + -4.141109943389893 + ], + "logprob_reward": -0.0161101296544075, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.0161101296544075, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04124099016189575, + 0.789338231086731, + 0.03084556572139263, + 0.012942160479724407, + 0.02865125611424446, + 0.031309738755226135, + 0.03210589289665222, + 0.010055636055767536, + 0.01439585816115141, + 0.009114624932408333 + ], + "credit_map": [ + 0.003207261674106121, + 0.00490691140294075, + 0.0042603109031915665, + 0.006490841507911682, + 0.006967827677726746, + 0.0031264007557183504, + 0.005909809842705727, + 0.006371623836457729, + 0.01365897711366415, + 0.0175113957375288, + 0.027530867606401443, + 0.012113402597606182, + 0.07275855541229248, + 0.3354544937610626, + 0.0313466377556324, + 0.27896395325660706, + 0.003459284780547023, + 0.002131448592990637, + 0.0035137755330652, + 0.005403308663517237, + 0.002862136345356703, + 0.004034055862575769, + 0.005308740306645632, + 0.004132815636694431, + 0.0013177754590287805, + 0.001432683551684022, + 0.001390217337757349, + 0.0009282847167924047, + 0.0026608689222484827, + 0.0026683975011110306, + 0.0010969986906275153, + 0.0014469349989667535, + 0.0028809348586946726, + 0.004490881226956844, + 0.0025653145276010036, + 0.0052427793852984905, + 0.0038334864657372236, + 0.00362101080827415, + 0.0022762976586818695, + 0.003740551881492138, + 0.00304806767962873, + 0.00450689485296607, + 0.0035719124134629965, + 0.005658828187733889, + 0.004057087004184723, + 0.003858886193484068, + 0.003021699609234929, + 0.003586362348869443, + 0.006223586853593588, + 0.00375112215988338, + 0.003897130023688078, + 0.0028101149946451187, + 0.004658470395952463, + 0.003301498480141163, + 0.00475996034219861, + 0.002704005688428879, + 0.0013215566286817193, + 0.0011089089093729854, + 0.0019476560410112143, + 0.0012532524997368455, + 0.0011935221264138818, + 0.0008570992504246533, + 0.0014498985838145018, + 0.0009237421909347177, + 0.0017002741806209087, + 0.002195705659687519, + 0.0020851902663707733, + 0.002284848364070058, + 0.0013839659513905644, + 0.0019954037852585316, + 0.001598878181539476, + 0.0011515918886289, + 0.0009775689104571939, + 0.0010014970321208239, + 0.0017485868884250522, + 0.001095104031264782, + 0.0011243062326684594, + 0.0014814073219895363, + 0.0008339301566593349, + 0.0008522240095771849 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 6 + ], + "selected_reward": -0.0161101296544075 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033711932599544525 + ], + "opposite_logprob_tokens": [ + -3.4087119102478027 + ], + "logprob_reward": -0.033711932599544525, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033711932599544525, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.04138451814651489, + 0.7776488661766052, + 0.03192245960235596, + 0.03530542552471161, + 0.029193714261054993, + 0.025367196649312973, + 0.017013084143400192, + 0.01987161673605442, + 0.013961785472929478, + 0.00833125226199627 + ], + "credit_map": [ + 0.0026162578724324703, + 0.0038348305970430374, + 0.00329135125502944, + 0.005343699362128973, + 0.007131833117455244, + 0.002897877711802721, + 0.012006320990622044, + 0.004262353293597698, + 0.013584687374532223, + 0.016867289319634438, + 0.026921380311250687, + 0.01181341614574194, + 0.07067075371742249, + 0.3306872248649597, + 0.0314248763024807, + 0.2756793200969696, + 0.002876868937164545, + 0.0020420397631824017, + 0.004444308578968048, + 0.00582666601985693, + 0.0024625093210488558, + 0.003878560848534107, + 0.006368675269186497, + 0.004022836685180664, + 0.0026874167378991842, + 0.004026961512863636, + 0.0038979509845376015, + 0.003993882331997156, + 0.003292666980996728, + 0.01004570908844471, + 0.002349616028368473, + 0.005011224187910557, + 0.0029923273250460625, + 0.004061483778059483, + 0.00237970519810915, + 0.004922719672322273, + 0.004116347059607506, + 0.00404692767187953, + 0.003284584265202284, + 0.0033896230161190033, + 0.0017586242174729705, + 0.002991283778101206, + 0.0019615355413407087, + 0.002597183221951127, + 0.006641467101871967, + 0.003700029803439975, + 0.0025740605778992176, + 0.003143016714602709, + 0.002630560426041484, + 0.0015360531397163868, + 0.0023273122496902943, + 0.0018649442354217172, + 0.0019185857381671667, + 0.0016702987486496568, + 0.0031802705489099026, + 0.0018850619671866298, + 0.0024025491438806057, + 0.001878537004813552, + 0.003263640683144331, + 0.002256532898172736, + 0.0017872770549729466, + 0.001549965818412602, + 0.0046871621161699295, + 0.002045953879132867, + 0.002523947274312377, + 0.001320660812780261, + 0.0027543387841433287, + 0.0013670717598870397, + 0.001392541453242302, + 0.0011855264892801642, + 0.0017448004800826311, + 0.0016729009803384542, + 0.0007913713925518095, + 0.0010898341424763203, + 0.001341954804956913, + 0.0010702917352318764, + 0.001161426305770874, + 0.0012539156014099717, + 0.0007358077564276755, + 0.000886651745531708 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 3 + ], + "selected_reward": -0.033711932599544525 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.31351444125175476 + ], + "opposite_logprob_tokens": [ + -1.3135144710540771 + ], + "logprob_reward": -0.31351444125175476, + "yes_no_margin_reward": 1.0, + "token_credit_reward": -0.31351444125175476, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.006352328695356846, + 0.8560779094696045, + 0.012943176552653313, + 0.018026180565357208, + 0.02155928872525692, + 0.024472709745168686, + 0.030804025009274483, + 0.01210404559969902, + 0.009023003280162811, + 0.008637291379272938 + ], + "credit_map": [ + 0.0005907039740122855, + 0.0010142914252355695, + 0.0006338896346278489, + 0.0008512199856340885, + 0.0006277982611209154, + 0.000842994311824441, + 0.0010105414548888803, + 0.0007808898808434606, + 0.01483005192130804, + 0.018826516345143318, + 0.02958744950592518, + 0.012492802925407887, + 0.07876239717006683, + 0.3641851842403412, + 0.03416166082024574, + 0.3032318949699402, + 0.0010297055123373866, + 0.0008351493743248284, + 0.0017953626811504364, + 0.002624368527904153, + 0.0012614537263289094, + 0.0016291711945086718, + 0.0021347696892917156, + 0.001633195672184229, + 0.0014057722873985767, + 0.0028530736453831196, + 0.001832320704124868, + 0.0022481014020740986, + 0.0017110473709180951, + 0.004461375996470451, + 0.0013023882638663054, + 0.002212102757766843, + 0.0017941751284524798, + 0.002572479657828808, + 0.0020579418633133173, + 0.0038576836232095957, + 0.0032267379574477673, + 0.0029595186933875084, + 0.002496511908248067, + 0.0025942411739379168, + 0.0016794862458482385, + 0.0025193614419549704, + 0.0019054764416068792, + 0.002488807775080204, + 0.006468579173088074, + 0.0035885232500731945, + 0.0032902867533266544, + 0.0025321885477751493, + 0.006042870227247477, + 0.003552171401679516, + 0.0038076434284448624, + 0.002690540160983801, + 0.004333927761763334, + 0.003162467386573553, + 0.0044800955802202225, + 0.0027343095280230045, + 0.0015408260514959693, + 0.001860498683527112, + 0.001282373326830566, + 0.0012885824544355273, + 0.001701974542811513, + 0.0013491130666807294, + 0.0020661300513893366, + 0.0010145471896976233, + 0.0012438003905117512, + 0.001003942801617086, + 0.0013277180260047317, + 0.0015734676271677017, + 0.0008601223235018551, + 0.0008636003476567566, + 0.001089482568204403, + 0.001060868613421917, + 0.0008180552395060658, + 0.0011283416533842683, + 0.0013516313629224896, + 0.0008489197352901101, + 0.001224536681547761, + 0.0013762153685092926, + 0.0009971243562176824, + 0.0008924673893488944 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 6, + 5 + ], + "selected_reward": -0.31351444125175476 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005954858381301165 + ], + "opposite_logprob_tokens": [ + -5.130954742431641 + ], + "logprob_reward": -0.005954858381301165, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.005954858381301165, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.01459752768278122, + 0.9168562889099121, + 0.019587378948926926, + 0.0049825394526124, + 0.010511078871786594, + 0.005918945651501417, + 0.014507859945297241, + 0.004261915106326342, + 0.003382260212674737, + 0.005394247360527515 + ], + "credit_map": [ + 0.001251696259714663, + 0.0017089934553951025, + 0.0017029903829097748, + 0.0013926163082942367, + 0.0017611919902265072, + 0.0014716355362907052, + 0.00220627523958683, + 0.003102125832810998, + 0.016107812523841858, + 0.01973845809698105, + 0.031626857817173004, + 0.013177461922168732, + 0.08366643637418747, + 0.3904086649417877, + 0.03666310012340546, + 0.3254673480987549, + 0.0022821479942649603, + 0.0017749639227986336, + 0.0022684140130877495, + 0.003817004384472966, + 0.0020478300284594297, + 0.002499972702935338, + 0.0024413641076534986, + 0.0024556780699640512, + 0.0004532910534180701, + 0.0006624367088079453, + 0.0006617598119191825, + 0.000526168558280915, + 0.0005718590691685677, + 0.0011190937366336584, + 0.0004551274760160595, + 0.0005328018451109529, + 0.0008013145416043699, + 0.001249960740096867, + 0.0010334016988053918, + 0.0018915697000920773, + 0.001401690416969359, + 0.0013690913328900933, + 0.0012458462733775377, + 0.0015182020142674446, + 0.0007183130946941674, + 0.0007234307704493403, + 0.0005544593441300094, + 0.0009073193068616092, + 0.0008216623682528734, + 0.0007882122881710529, + 0.0008760181372053921, + 0.0005295289447531104, + 0.0024178391322493553, + 0.001570866908878088, + 0.0019512750441208482, + 0.0015747778816148639, + 0.0017983708530664444, + 0.0014105970039963722, + 0.0025060863699764013, + 0.0012780431425198913, + 0.0006396444514393806, + 0.00032660437864251435, + 0.0004898276529274881, + 0.0004979904624633491, + 0.0006045324262231588, + 0.000593662029132247, + 0.0006602788344025612, + 0.00044937425991520286, + 0.0005008089356124401, + 0.00041874099406413734, + 0.0004737594863399863, + 0.00038441564538516104, + 0.00036174547858536243, + 0.0003464634937699884, + 0.00044127460569143295, + 0.00045505064190365374, + 0.0004832078702747822, + 0.0005982810398563743, + 0.0008146765758283436, + 0.0007651285268366337, + 0.0007251923670992255, + 0.00067873828811571, + 0.0008022520923987031, + 0.0005267693195492029 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 0 + ], + "selected_reward": -0.005954858381301165 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, shot from above, follows a figure as it journeys through a vast expanse of emerald green bamboo. Dazzling golden sunlight streams from deep within the frame, creating a strong backlight effect that cuts the figure of the knight into a near-black silhouette. He wears a loose white silk robe, its long sleeves and sleeves billowing and flowing like clouds as he moves at full speed. His movements are incredibly swift, seemingly weightless, as he breezes past towering bamboos with a fluid, flowing grace. Each touch of his toes, like a dragonfly skimming the water, only slightly bends the tough bamboo tips. The intense light frames his entire form, from his flowing hair to the fluttering hem of his robe, with a distinct, dazzling golden edge, making him appear divine and distinct against the deep green background. Light filters through the gaps between the bamboo leaves, casting a dappled halo across the lens, imbuing the image with the grandeur and artistic atmosphere of a martial arts epic.", + "seed": 10386, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_2887_seed_10386_dp_18_rank_36_sample_1_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 0, + "sample_idx_in_group": 1, + "avg": -0.28138376912102103, + "vlm_logprob": -0.28138376912102103, + "vlm_yes_no_margin": 2.6, + "vlm_token_credit": -0.28138376912102103, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -0.9228776693344116 + ], + "opposite_logprob_tokens": [ + -0.5478776693344116 + ], + "logprob_reward": -0.9228776693344116, + "yes_no_margin_reward": -0.375, + "token_credit_reward": -0.9228776693344116, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.0963524878025055, + 0.23594896495342255, + 0.09709323197603226, + 0.07963365316390991, + 0.05506431683897972, + 0.04850493371486664, + 0.10053662210702896, + 0.10584639012813568, + 0.13789679110050201, + 0.04312257841229439 + ], + "credit_map": [ + 0.007919036783277988, + 0.00855601392686367, + 0.018190301954746246, + 0.010339654982089996, + 0.010434125550091267, + 0.018284842371940613, + 0.008534235879778862, + 0.014094275422394276, + 0.05166565626859665, + 0.0695294663310051, + 0.021339843049645424, + 0.01272080559283495, + 0.019954293966293335, + 0.023330839350819588, + 0.015471928752958775, + 0.021936144679784775, + 0.009967999532818794, + 0.015368754044175148, + 0.013291902840137482, + 0.006227182690054178, + 0.013219555839896202, + 0.013443767093122005, + 0.012413956224918365, + 0.01316011231392622, + 0.009019378572702408, + 0.013294672593474388, + 0.008982468396425247, + 0.011829441413283348, + 0.0059348647482693195, + 0.013013373129069805, + 0.0067123775370419025, + 0.010847076773643494, + 0.003220721846446395, + 0.00684799998998642, + 0.004829465411603451, + 0.0034847131464630365, + 0.0034547606483101845, + 0.013983279466629028, + 0.005179161671549082, + 0.01406421884894371, + 0.005108871031552553, + 0.005550657398998737, + 0.0032060130033642054, + 0.004413200542330742, + 0.005322133656591177, + 0.007705317810177803, + 0.009988296777009964, + 0.007210446987301111, + 0.005750139243900776, + 0.010786863975226879, + 0.009775999933481216, + 0.01662110909819603, + 0.011606930755078793, + 0.021142061799764633, + 0.012175530195236206, + 0.012677987106144428, + 0.005149039905518293, + 0.006283471826463938, + 0.005304072983562946, + 0.007625191006809473, + 0.00912508089095354, + 0.034609805792570114, + 0.006966257467865944, + 0.030783474445343018, + 0.007633892819285393, + 0.02015809342265129, + 0.011420921422541142, + 0.02595808357000351, + 0.015228214673697948, + 0.01689111813902855, + 0.018673967570066452, + 0.021932506933808327, + 0.004598057828843594, + 0.006412431597709656, + 0.0034300806000828743, + 0.004444779362529516, + 0.006329648662358522, + 0.006826246622949839, + 0.005240892991423607, + 0.005840441677719355 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 8, + 7 + ], + "selected_reward": -0.9228776693344116 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01611212268471718 + ], + "opposite_logprob_tokens": [ + -4.141112327575684 + ], + "logprob_reward": -0.01611212268471718, + "yes_no_margin_reward": 4.125, + "token_credit_reward": -0.01611212268471718, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2996758222579956, + 0.24701876938343048, + 0.08693324029445648, + 0.03631526231765747, + 0.0373835451900959, + 0.01627228409051895, + 0.05903191491961479, + 0.18173566460609436, + 0.0226433165371418, + 0.012990188784897327 + ], + "credit_map": [ + 0.024948272854089737, + 0.025587571784853935, + 0.07520658522844315, + 0.024341018870472908, + 0.03709479421377182, + 0.03382211551070213, + 0.030046751722693443, + 0.048628710210323334, + 0.05796205252408981, + 0.06826955825090408, + 0.02376762218773365, + 0.01094281580299139, + 0.021746890619397163, + 0.021992767229676247, + 0.011487046256661415, + 0.030850013718008995, + 0.008358359336853027, + 0.016872256994247437, + 0.01912301778793335, + 0.0068404595367610455, + 0.008422457613050938, + 0.007152910344302654, + 0.007003887556493282, + 0.013159886002540588, + 0.0032643473241478205, + 0.004398394376039505, + 0.005287344567477703, + 0.0038952655158936977, + 0.0046839118003845215, + 0.006620536558330059, + 0.0027178493328392506, + 0.00544761260971427, + 0.002654819982126355, + 0.005911136046051979, + 0.003389682387933135, + 0.002837707521393895, + 0.005916474852710962, + 0.006575142499059439, + 0.003925262484699488, + 0.006173320114612579, + 0.0016413717530667782, + 0.0014450885355472565, + 0.0010228828759863973, + 0.0013318349374458194, + 0.0023064843844622374, + 0.003350452985614538, + 0.0026901941746473312, + 0.0024839758407324553, + 0.005040050949901342, + 0.006499028764665127, + 0.005176317412406206, + 0.006690487265586853, + 0.006403799168765545, + 0.01373505499213934, + 0.005970726255327463, + 0.009516450576484203, + 0.010486237704753876, + 0.017970016226172447, + 0.014145499095320702, + 0.022399380803108215, + 0.014754161238670349, + 0.045088641345500946, + 0.01524850819259882, + 0.041643209755420685, + 0.0023858200293034315, + 0.003537106327712536, + 0.0022987734992057085, + 0.003901588497683406, + 0.003701069625094533, + 0.002583939116448164, + 0.00249763042666018, + 0.0017373890150338411, + 0.0014171581715345383, + 0.0020202596206218004, + 0.0011623217724263668, + 0.0010013036662712693, + 0.001395360566675663, + 0.002048323629423976, + 0.0019234898500144482, + 0.0020219713915139437 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 7 + ], + "selected_reward": -0.01611212268471718 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.033710431307554245 + ], + "opposite_logprob_tokens": [ + -3.408710479736328 + ], + "logprob_reward": -0.033710431307554245, + "yes_no_margin_reward": 3.375, + "token_credit_reward": -0.033710431307554245, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.2734021246433258, + 0.17392480373382568, + 0.15523189306259155, + 0.05695809796452522, + 0.09432046860456467, + 0.033132486045360565, + 0.03427930176258087, + 0.05453936383128166, + 0.05182191729545593, + 0.07238952815532684 + ], + "credit_map": [ + 0.023256666958332062, + 0.022494466975331306, + 0.06683846563100815, + 0.02143712341785431, + 0.03115955926477909, + 0.03897375613451004, + 0.02878471463918686, + 0.0404573529958725, + 0.036604657769203186, + 0.04544844105839729, + 0.014750538393855095, + 0.011452924460172653, + 0.015216146595776081, + 0.016721557825803757, + 0.01666140928864479, + 0.017069123685359955, + 0.018428819254040718, + 0.020984288305044174, + 0.044401951134204865, + 0.01606212556362152, + 0.008993219584226608, + 0.011351183988153934, + 0.016315989196300507, + 0.01869431883096695, + 0.006539138033986092, + 0.007116161752492189, + 0.006622886750847101, + 0.006567578297108412, + 0.010086306370794773, + 0.009378258138895035, + 0.004857915453612804, + 0.005789852701127529, + 0.00836148764938116, + 0.010158417746424675, + 0.008956627920269966, + 0.010264313779771328, + 0.008980836719274521, + 0.01879560388624668, + 0.012654312886297703, + 0.016148868948221207, + 0.004033179488033056, + 0.004310654476284981, + 0.00323141529224813, + 0.002644094405695796, + 0.0038643006701022387, + 0.005413180682808161, + 0.005354232154786587, + 0.0042814286425709724, + 0.0021714861504733562, + 0.003976404666900635, + 0.0030063807498663664, + 0.00481840455904603, + 0.004850287921726704, + 0.006803490221500397, + 0.0036355990450829268, + 0.005017248913645744, + 0.0031392511446028948, + 0.006799819879233837, + 0.003490789094939828, + 0.0071239173412323, + 0.005295373499393463, + 0.011242074891924858, + 0.004662799648940563, + 0.012785336002707481, + 0.0038243348244577646, + 0.008032271638512611, + 0.005488286260515451, + 0.008829036727547646, + 0.008319570682942867, + 0.006297806277871132, + 0.005771353840827942, + 0.005259254481643438, + 0.0088747413828969, + 0.0120397275313735, + 0.006221163086593151, + 0.007355209439992905, + 0.008718662895262241, + 0.010756543837487698, + 0.010350139811635017, + 0.00807334017008543 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 1, + 2 + ], + "selected_reward": -0.033710431307554245 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.4289599657058716 + ], + "opposite_logprob_tokens": [ + -1.0539599657058716 + ], + "logprob_reward": -0.4289599657058716, + "yes_no_margin_reward": 0.625, + "token_credit_reward": -0.4289599657058716, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.11733817309141159, + 0.11073751002550125, + 0.14890997111797333, + 0.043850336223840714, + 0.15614375472068787, + 0.0626830980181694, + 0.08909856528043747, + 0.11897919327020645, + 0.10284174233675003, + 0.0494176484644413 + ], + "credit_map": [ + 0.009748050943017006, + 0.010547732003033161, + 0.02722983807325363, + 0.01047863345593214, + 0.014810554683208466, + 0.012817378155887127, + 0.011306856758892536, + 0.020399149507284164, + 0.025436192750930786, + 0.030831029638648033, + 0.011144930496811867, + 0.005141876172274351, + 0.01038294192403555, + 0.010348501615226269, + 0.005206813104450703, + 0.012245234102010727, + 0.01367719005793333, + 0.03460779786109924, + 0.021301576867699623, + 0.00834541954100132, + 0.012907180935144424, + 0.010903483256697655, + 0.014076187275350094, + 0.033091139048337936, + 0.004246666561812162, + 0.0056908018887043, + 0.0057783788070082664, + 0.0057267057709395885, + 0.0069008455611765385, + 0.006054833065718412, + 0.004298025742173195, + 0.005154083017259836, + 0.008517300710082054, + 0.014180799014866352, + 0.016349725425243378, + 0.010583396069705486, + 0.019265789538621902, + 0.04107298329472542, + 0.010389981791377068, + 0.035783801227808, + 0.00561936292797327, + 0.007638220675289631, + 0.00548823457211256, + 0.008584403432905674, + 0.0077753327786922455, + 0.008200837299227715, + 0.01231754943728447, + 0.007059161551296711, + 0.005342734511941671, + 0.008282958529889584, + 0.010279105044901371, + 0.015877865254878998, + 0.012169303372502327, + 0.015546659007668495, + 0.009957456029951572, + 0.011642491444945335, + 0.00612152274698019, + 0.009930970147252083, + 0.008209098130464554, + 0.013433764688670635, + 0.009312417358160019, + 0.033229291439056396, + 0.00945480540394783, + 0.02928733266890049, + 0.00888108555227518, + 0.015966469421982765, + 0.005935467779636383, + 0.020191723480820656, + 0.019865231588482857, + 0.012496119365096092, + 0.010833890177309513, + 0.008671761490404606, + 0.004231589846313, + 0.00806316826492548, + 0.0031495518051087856, + 0.004515902139246464, + 0.0055880313739180565, + 0.009024981409311295, + 0.007440140005201101, + 0.007404285948723555 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 2, + 7 + ], + "selected_reward": -0.4289599657058716 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005258656572550535 + ], + "opposite_logprob_tokens": [ + -5.255258560180664 + ], + "logprob_reward": -0.005258656572550535, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005258656572550535, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.15136705338954926, + 0.1743084043264389, + 0.05977022647857666, + 0.16448648273944855, + 0.14644700288772583, + 0.03156890347599983, + 0.03662196546792984, + 0.07948698103427887, + 0.11542531102895737, + 0.04051769897341728 + ], + "credit_map": [ + 0.013222706504166126, + 0.01743103563785553, + 0.03467715531587601, + 0.012601761147379875, + 0.018664101138710976, + 0.01808822713792324, + 0.01785731315612793, + 0.01882474683225155, + 0.04122580215334892, + 0.0525905005633831, + 0.016108445823192596, + 0.007952922023832798, + 0.014461156912147999, + 0.014504685997962952, + 0.008442017249763012, + 0.019022848457098007, + 0.005866257473826408, + 0.008455818518996239, + 0.010167807340621948, + 0.005770369432866573, + 0.005978657864034176, + 0.007418687455356121, + 0.007887102663516998, + 0.008225522935390472, + 0.01612222194671631, + 0.01946221850812435, + 0.015881158411502838, + 0.027154620736837387, + 0.01117694191634655, + 0.03003799170255661, + 0.017828572541475296, + 0.026822742074728012, + 0.004346422851085663, + 0.007907196879386902, + 0.006315147038549185, + 0.006344105117022991, + 0.00872306153178215, + 0.052741143852472305, + 0.007298740092664957, + 0.05277117341756821, + 0.003925294149667025, + 0.004086629953235388, + 0.0026342379860579967, + 0.002427050145342946, + 0.0034023316111415625, + 0.006495378911495209, + 0.0030070000793784857, + 0.005590977147221565, + 0.0026574814692139626, + 0.003802959341555834, + 0.003333176951855421, + 0.005433910526335239, + 0.004734495654702187, + 0.007077901624143124, + 0.004879796411842108, + 0.004702243022620678, + 0.004453714936971664, + 0.007887632586061954, + 0.0051830536685884, + 0.008222641423344612, + 0.006273212376981974, + 0.02275768853724003, + 0.0042981551960110664, + 0.020410874858498573, + 0.013134218752384186, + 0.01743444800376892, + 0.011190777644515038, + 0.016784707084298134, + 0.019560908898711205, + 0.01352669857442379, + 0.01340854074805975, + 0.010385001078248024, + 0.004048368893563747, + 0.008681734092533588, + 0.0028923596255481243, + 0.003546625142917037, + 0.005908256862312555, + 0.005280980374664068, + 0.004934656899422407, + 0.005224715918302536 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 3, + 0 + ], + "selected_reward": -0.005258656572550535 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, shot from above, follows a figure as it journeys through a vast expanse of emerald green bamboo. Dazzling golden sunlight streams from deep within the frame, creating a strong backlight effect that cuts the figure of the knight into a near-black silhouette. He wears a loose white silk robe, its long sleeves and sleeves billowing and flowing like clouds as he moves at full speed. His movements are incredibly swift, seemingly weightless, as he breezes past towering bamboos with a fluid, flowing grace. Each touch of his toes, like a dragonfly skimming the water, only slightly bends the tough bamboo tips. The intense light frames his entire form, from his flowing hair to the fluttering hem of his robe, with a distinct, dazzling golden edge, making him appear divine and distinct against the deep green background. Light filters through the gaps between the bamboo leaves, casting a dappled halo across the lens, imbuing the image with the grandeur and artistic atmosphere of a martial arts epic.", + "seed": 10387, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_2887_seed_10387_dp_19_rank_38_sample_2_slot_0.mp4", + "sample_idx": 0, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 2, + "avg": -0.2008747291751206, + "vlm_logprob": -0.2008747291751206, + "vlm_yes_no_margin": 2.7749999880790712, + "vlm_token_credit": -0.2008747291751206, + "vlm_yes_ratio": 1.0, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.5667331218719482 + ], + "opposite_logprob_tokens": [ + -0.9417331218719482 + ], + "logprob_reward": -0.5667331218719482, + "yes_no_margin_reward": 0.375, + "token_credit_reward": -0.5667331218719482, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19680586457252502, + 0.07901968061923981, + 0.053148895502090454, + 0.18446898460388184, + 0.0990840345621109, + 0.05142426863312721, + 0.07657652348279953, + 0.11243018507957458, + 0.03969021886587143, + 0.1073513776063919 + ], + "credit_map": [ + 0.019504021853208542, + 0.028342995792627335, + 0.027649354189634323, + 0.031275469809770584, + 0.030693605542182922, + 0.027847616001963615, + 0.01696150004863739, + 0.014531295746564865, + 0.007604435086250305, + 0.00944997277110815, + 0.009004254825413227, + 0.009662209078669548, + 0.010986247099936008, + 0.0113843297585845, + 0.012503547593951225, + 0.008424685336649418, + 0.0037029406521469355, + 0.0070348866283893585, + 0.006875060964375734, + 0.006422991398721933, + 0.006947636604309082, + 0.008644125424325466, + 0.008734122850000858, + 0.004787128418684006, + 0.010485943406820297, + 0.021961934864521027, + 0.03924432396888733, + 0.008451586589217186, + 0.017701221629977226, + 0.05313832685351372, + 0.010514323599636555, + 0.02297130972146988, + 0.008909212425351143, + 0.010889279656112194, + 0.016078947111964226, + 0.005676785483956337, + 0.01591181568801403, + 0.023346010595560074, + 0.0059993029572069645, + 0.012272673659026623, + 0.00555012421682477, + 0.00630413880571723, + 0.009867231361567974, + 0.005276495590806007, + 0.005906394682824612, + 0.006474485155194998, + 0.006204226519912481, + 0.005841171368956566, + 0.010492604225873947, + 0.017138855531811714, + 0.01195854414254427, + 0.0067764949053525925, + 0.005875329952687025, + 0.010083893314003944, + 0.007372591178864241, + 0.006878206506371498, + 0.015200567431747913, + 0.016380753368139267, + 0.019859855994582176, + 0.008119450882077217, + 0.018755367025732994, + 0.01647697575390339, + 0.007496046833693981, + 0.010141170583665371, + 0.0042838649824261665, + 0.0058274962939321995, + 0.005106213502585888, + 0.0038651293143630028, + 0.00577903725206852, + 0.005665400996804237, + 0.00343884713947773, + 0.005724224727600813, + 0.010638811625540257, + 0.010415698401629925, + 0.008834525011479855, + 0.013353661634027958, + 0.019713561981916428, + 0.018115926533937454, + 0.008327157236635685, + 0.017952026799321175 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 3, + 7 + ], + "selected_reward": -0.5667331218719482 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.018233073875308037 + ], + "opposite_logprob_tokens": [ + -4.018233299255371 + ], + "logprob_reward": -0.018233073875308037, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.018233073875308037, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19498378038406372, + 0.05852292478084564, + 0.04447639733552933, + 0.04250331595540047, + 0.17129595577716827, + 0.04795348271727562, + 0.09005175530910492, + 0.19836874306201935, + 0.09782019257545471, + 0.05402347818017006 + ], + "credit_map": [ + 0.01625540293753147, + 0.012167670764029026, + 0.04878221079707146, + 0.038655154407024384, + 0.02689112350344658, + 0.0187057051807642, + 0.021419281139969826, + 0.012107215821743011, + 0.004327104426920414, + 0.004896702244877815, + 0.00596113596111536, + 0.010491072200238705, + 0.0068687680177390575, + 0.007235621567815542, + 0.01260012574493885, + 0.006142391357570887, + 0.0032381864730268717, + 0.004799658432602882, + 0.008260463364422321, + 0.0052748750895261765, + 0.007264475803822279, + 0.007225296460092068, + 0.004945802967995405, + 0.003467633156105876, + 0.003880996722728014, + 0.00818601530045271, + 0.004994255490601063, + 0.0041714259423315525, + 0.006004671100527048, + 0.006673742551356554, + 0.0038413656875491142, + 0.004750837571918964, + 0.011494715698063374, + 0.014174890704452991, + 0.044310662895441055, + 0.006403564941138029, + 0.01122838445007801, + 0.028227530419826508, + 0.006344894878566265, + 0.04911129176616669, + 0.0075588831678032875, + 0.008685649372637272, + 0.007153639104217291, + 0.003869142383337021, + 0.0056754290126264095, + 0.005793151445686817, + 0.0038445654790848494, + 0.005373018328100443, + 0.013286340050399303, + 0.01536310464143753, + 0.01412690244615078, + 0.010219375602900982, + 0.00702951243147254, + 0.011361897923052311, + 0.008865518495440483, + 0.00979909859597683, + 0.02156761661171913, + 0.02205260656774044, + 0.02629373036324978, + 0.02165628969669342, + 0.03866756334900856, + 0.027276428416371346, + 0.0230033490806818, + 0.017851147800683975, + 0.01632276363670826, + 0.008421543054282665, + 0.018400313332676888, + 0.005578736308962107, + 0.016702968627214432, + 0.014031521044671535, + 0.00753438426181674, + 0.010827954858541489, + 0.006874512881040573, + 0.004945215303450823, + 0.00574282044544816, + 0.007582072168588638, + 0.008378742262721062, + 0.009567760862410069, + 0.00429898826405406, + 0.006633360404521227 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 7, + 0, + 4 + ], + "selected_reward": -0.018233073875308037 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.02634967863559723 + ], + "opposite_logprob_tokens": [ + -3.6513497829437256 + ], + "logprob_reward": -0.02634967863559723, + "yes_no_margin_reward": 3.625, + "token_credit_reward": -0.02634967863559723, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.14914391934871674, + 0.040544647723436356, + 0.09233152121305466, + 0.10640507191419601, + 0.20540960133075714, + 0.08019141852855682, + 0.0506712906062603, + 0.14973218739032745, + 0.0647195503115654, + 0.06085073575377464 + ], + "credit_map": [ + 0.009147942997515202, + 0.01769157312810421, + 0.042514774948358536, + 0.03510678932070732, + 0.016605792567133904, + 0.009706900455057621, + 0.010714719071984291, + 0.0076554277911782265, + 0.004154696594923735, + 0.0037630924489349127, + 0.005235479678958654, + 0.006874324753880501, + 0.005312393885105848, + 0.005305257625877857, + 0.00648363446816802, + 0.003415769664570689, + 0.006828701123595238, + 0.01055917888879776, + 0.014038300141692162, + 0.010901442728936672, + 0.01613536849617958, + 0.013921952806413174, + 0.013646514154970646, + 0.006300065666437149, + 0.006069048307836056, + 0.010439112782478333, + 0.02152450755238533, + 0.007346275262534618, + 0.012989938259124756, + 0.02761206217110157, + 0.010899387300014496, + 0.009524736553430557, + 0.012559122405946255, + 0.015924517065286636, + 0.05418216064572334, + 0.009337249211966991, + 0.01849213056266308, + 0.0262436643242836, + 0.01155700534582138, + 0.05711375176906586, + 0.008297553285956383, + 0.009534516371786594, + 0.009217709302902222, + 0.008755447342991829, + 0.014107554219663143, + 0.009769458323717117, + 0.01245387177914381, + 0.008055298589169979, + 0.005435805302113295, + 0.007741821929812431, + 0.008308446034789085, + 0.004453690256923437, + 0.0054602185264229774, + 0.006367221008986235, + 0.007957453839480877, + 0.004946632776409388, + 0.013599205762147903, + 0.016383569687604904, + 0.020670445635914803, + 0.018611956387758255, + 0.027170894667506218, + 0.020296499133110046, + 0.018993524834513664, + 0.014006097801029682, + 0.0059195831418037415, + 0.015523640438914299, + 0.006734746973961592, + 0.006689718924462795, + 0.007428363896906376, + 0.009329290129244328, + 0.005866578780114651, + 0.007227631751447916, + 0.010930684395134449, + 0.00632079690694809, + 0.00637176726013422, + 0.007467308547347784, + 0.007699172478169203, + 0.007299405988305807, + 0.006670690141618252, + 0.008090910501778126 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 4, + 7, + 0 + ], + "selected_reward": -0.02634967863559723 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.3871048092842102 + ], + "opposite_logprob_tokens": [ + -1.1371047496795654 + ], + "logprob_reward": -0.3871048092842102, + "yes_no_margin_reward": 0.7499999403953552, + "token_credit_reward": -0.3871048092842102, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.05812392756342888, + 0.0382167249917984, + 0.04456114023923874, + 0.053248047828674316, + 0.11907090991735458, + 0.22558805346488953, + 0.03935083746910095, + 0.16323180496692657, + 0.19189360737800598, + 0.06671496480703354 + ], + "credit_map": [ + 0.00527070090174675, + 0.008558075875043869, + 0.012940543703734875, + 0.010537982918322086, + 0.006955181248486042, + 0.005244661122560501, + 0.005161376670002937, + 0.0034554023295640945, + 0.002304636873304844, + 0.002709067426621914, + 0.004927119705826044, + 0.006635910365730524, + 0.006962752901017666, + 0.004607134498655796, + 0.006814391817897558, + 0.0032557121012359858, + 0.0027502134907990694, + 0.0052071972750127316, + 0.008456501178443432, + 0.005230901297181845, + 0.007454045582562685, + 0.007649947423487902, + 0.004596971441060305, + 0.0032153648789972067, + 0.004152711946517229, + 0.008435461670160294, + 0.007027884479612112, + 0.00479037594050169, + 0.007918902672827244, + 0.006232286337763071, + 0.007818568497896194, + 0.006871855817735195, + 0.006414283532649279, + 0.006692234892398119, + 0.034815892577171326, + 0.006434435490518808, + 0.008302709087729454, + 0.011184030212461948, + 0.00829806737601757, + 0.036929260939359665, + 0.01855252869427204, + 0.021949119865894318, + 0.021600481122732162, + 0.02526567131280899, + 0.03909057751297951, + 0.029963083565235138, + 0.04256635904312134, + 0.026600206270813942, + 0.004494151566177607, + 0.00571294454857707, + 0.00482788635417819, + 0.0049035148695111275, + 0.0040045795030891895, + 0.004528605844825506, + 0.006560086738318205, + 0.004319067113101482, + 0.019388815388083458, + 0.015933968126773834, + 0.019292375072836876, + 0.021092532202601433, + 0.03021787665784359, + 0.020341210067272186, + 0.01998850516974926, + 0.01697651669383049, + 0.01731552928686142, + 0.04806198179721832, + 0.01855519786477089, + 0.0222132820636034, + 0.015222563408315182, + 0.02812054194509983, + 0.021782249212265015, + 0.020622260868549347, + 0.010019544512033463, + 0.007691795937716961, + 0.006353477481752634, + 0.00828641653060913, + 0.011356308124959469, + 0.009479195810854435, + 0.006193557288497686, + 0.007334667257964611 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 5, + 8, + 7 + ], + "selected_reward": -0.3871048092842102 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.0059529622085392475 + ], + "opposite_logprob_tokens": [ + -5.130952835083008 + ], + "logprob_reward": -0.0059529622085392475, + "yes_no_margin_reward": 5.125, + "token_credit_reward": -0.0059529622085392475, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.19004583358764648, + 0.31805893778800964, + 0.07258496433496475, + 0.05677773058414459, + 0.08903640508651733, + 0.054143358021974564, + 0.03632861375808716, + 0.0625394880771637, + 0.04325501620769501, + 0.07722970098257065 + ], + "credit_map": [ + 0.021924734115600586, + 0.02326548658311367, + 0.03623868152499199, + 0.0372539758682251, + 0.024165043607354164, + 0.01579686626791954, + 0.017867794260382652, + 0.013533256947994232, + 0.02822386473417282, + 0.030794546008110046, + 0.05047314614057541, + 0.03854275867342949, + 0.06632569432258606, + 0.03452759608626366, + 0.04035567119717598, + 0.02881566621363163, + 0.006173266563564539, + 0.00893260445445776, + 0.012781506404280663, + 0.008493551053106785, + 0.013037124648690224, + 0.009605281054973602, + 0.007768083829432726, + 0.005793547723442316, + 0.004047704860568047, + 0.007647080812603235, + 0.00704170810058713, + 0.005104569718241692, + 0.009877784177660942, + 0.008854934014379978, + 0.006852856837213039, + 0.007351091131567955, + 0.006972889415919781, + 0.006774861365556717, + 0.023805217817425728, + 0.005035138223320246, + 0.009046399034559727, + 0.01177312433719635, + 0.006624302826821804, + 0.019004473462700844, + 0.004234982188791037, + 0.006596370134502649, + 0.006617328152060509, + 0.0054497867822647095, + 0.008028593845665455, + 0.006947671994566917, + 0.010082210414111614, + 0.006186415441334248, + 0.006253281142562628, + 0.006784689612686634, + 0.004358368460088968, + 0.0039528412744402885, + 0.0029318276792764664, + 0.004472097381949425, + 0.00432982575148344, + 0.00324568385258317, + 0.006614078301936388, + 0.006211588624864817, + 0.008665196597576141, + 0.007927937433123589, + 0.009733485989272594, + 0.008887547068297863, + 0.0079418383538723, + 0.006557818036526442, + 0.005198900122195482, + 0.009008984081447124, + 0.00405483040958643, + 0.004672184120863676, + 0.004760192707180977, + 0.005727867130190134, + 0.0037475398276001215, + 0.006084515247493982, + 0.013900461606681347, + 0.006993802729994059, + 0.006721555721014738, + 0.009913532063364983, + 0.011271633207798004, + 0.011624224483966827, + 0.00817097257822752, + 0.008633517660200596 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 0, + 4 + ], + "selected_reward": -0.0059529622085392475 + } + ], + "vlm_decoded_outputs": [ + "Yes", + "Yes", + "Yes", + "Yes", + "Yes" + ] + }, + { + "prompt": "dynamic, high-speed tracking shot, shot from above, follows a figure as it journeys through a vast expanse of emerald green bamboo. Dazzling golden sunlight streams from deep within the frame, creating a strong backlight effect that cuts the figure of the knight into a near-black silhouette. He wears a loose white silk robe, its long sleeves and sleeves billowing and flowing like clouds as he moves at full speed. His movements are incredibly swift, seemingly weightless, as he breezes past towering bamboos with a fluid, flowing grace. Each touch of his toes, like a dragonfly skimming the water, only slightly bends the tough bamboo tips. The intense light frames his entire form, from his flowing hair to the fluttering hem of his robe, with a distinct, dazzling golden edge, making him appear divine and distinct against the deep green background. Light filters through the gaps between the bamboo leaves, casting a dappled halo across the lens, imbuing the image with the grandeur and artistic atmosphere of a martial arts epic.", + "seed": 10388, + "video_path": "./outputs/yw006-4-sage-qwen35-9b-token-credit-video-adv-window3/rl_samples/0000104/group_9/index_2887_seed_10388_dp_19_rank_38_sample_3_slot_1.mp4", + "sample_idx": 1, + "ranks_per_group": 2, + "group_idx": 9, + "rank_in_group": 1, + "sample_idx_in_group": 3, + "avg": -0.3604913683608174, + "vlm_logprob": -0.3604913683608174, + "vlm_yes_no_margin": 2.3, + "vlm_token_credit": -0.3604913683608174, + "vlm_yes_ratio": 0.8, + "vlm_num_questions": 5.0, + "reward_records": [ + { + "sample_idx": 0, + "question_id": "q1", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "No", + "normalized_pred_keywords": [ + "No" + ], + "target_logprob_tokens": [ + -1.0840893983840942 + ], + "opposite_logprob_tokens": [ + -0.45908939838409424 + ], + "logprob_reward": -1.0840893983840942, + "yes_no_margin_reward": -0.625, + "token_credit_reward": -1.0840893983840942, + "yes_ratio_reward": 0.0, + "frame_importance": [ + 0.20391684770584106, + 0.1405860334634781, + 0.05540347099304199, + 0.03900619596242905, + 0.056129712611436844, + 0.06950023770332336, + 0.06446345150470734, + 0.03620880842208862, + 0.061576228588819504, + 0.27320897579193115 + ], + "credit_map": [ + 0.027520867064595222, + 0.03324131667613983, + 0.02363857813179493, + 0.02695850282907486, + 0.03300371766090393, + 0.01849573850631714, + 0.019615260884165764, + 0.021442873403429985, + 0.01780013181269169, + 0.02144913375377655, + 0.023811696097254753, + 0.015079031698405743, + 0.01885724626481533, + 0.0184337068349123, + 0.007125326432287693, + 0.018029754981398582, + 0.005346212070435286, + 0.005071230232715607, + 0.006282862275838852, + 0.004297837149351835, + 0.006369936745613813, + 0.014315447770059109, + 0.005499108694493771, + 0.00822084117680788, + 0.0032893605530261993, + 0.00423947349190712, + 0.005005813669413328, + 0.004491446074098349, + 0.0030821720138192177, + 0.004638453479856253, + 0.007097742520272732, + 0.007161731366068125, + 0.008269406855106354, + 0.005341218318790197, + 0.00959223322570324, + 0.004219779744744301, + 0.005465859081596136, + 0.011913987807929516, + 0.004267147742211819, + 0.007060077972710133, + 0.006424283143132925, + 0.003536755917593837, + 0.02147161401808262, + 0.004745211452245712, + 0.005157791078090668, + 0.007710813544690609, + 0.0036631973925977945, + 0.016790568828582764, + 0.00760620879009366, + 0.004899461288005114, + 0.008691332302987576, + 0.007556871045380831, + 0.008125084452331066, + 0.010426728054881096, + 0.005887453444302082, + 0.011270314455032349, + 0.003324643475934863, + 0.0027995191048830748, + 0.004893552046269178, + 0.00473106699064374, + 0.003857546951621771, + 0.004233113490045071, + 0.004120107740163803, + 0.008249261416494846, + 0.007087394129484892, + 0.005064455326646566, + 0.005217752885073423, + 0.006303728092461824, + 0.008154631592333317, + 0.01323403138667345, + 0.00566204683855176, + 0.010852192528545856, + 0.018521586433053017, + 0.014616535045206547, + 0.008246256969869137, + 0.012280669063329697, + 0.01684695854783058, + 0.10135778784751892, + 0.008477549068629742, + 0.092861607670784 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 0, + 1 + ], + "selected_reward": -1.0840893983840942 + }, + { + "sample_idx": 0, + "question_id": "q2", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.01822078227996826 + ], + "opposite_logprob_tokens": [ + -4.018220901489258 + ], + "logprob_reward": -0.01822078227996826, + "yes_no_margin_reward": 4.0, + "token_credit_reward": -0.01822078227996826, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08905543386936188, + 0.22304338216781616, + 0.1394287347793579, + 0.04900483414530754, + 0.13213321566581726, + 0.12309406697750092, + 0.05067075416445732, + 0.041768189519643784, + 0.05803319811820984, + 0.09376824647188187 + ], + "credit_map": [ + 0.01242532953619957, + 0.011452917940914631, + 0.011646801605820656, + 0.01413004845380783, + 0.006921192165464163, + 0.014744206331670284, + 0.005726081784814596, + 0.0120088467374444, + 0.044273704290390015, + 0.029955429956316948, + 0.03369845822453499, + 0.020537959411740303, + 0.01669223979115486, + 0.032549526542425156, + 0.011339019984006882, + 0.03399702534079552, + 0.011556350626051426, + 0.02299473062157631, + 0.013362911529839039, + 0.012576411478221416, + 0.01472740713506937, + 0.02697814255952835, + 0.020163988694548607, + 0.017068782821297646, + 0.005420509725809097, + 0.0073896534740924835, + 0.00533217191696167, + 0.005103237461298704, + 0.004684411454945803, + 0.008031080476939678, + 0.006777195259928703, + 0.00626656599342823, + 0.01812150701880455, + 0.011474939063191414, + 0.02245253138244152, + 0.009515529498457909, + 0.011549610644578934, + 0.025023261085152626, + 0.01043452974408865, + 0.023561308160424232, + 0.014008489437401295, + 0.00518085015937686, + 0.03968461975455284, + 0.004824695643037558, + 0.006383184809237719, + 0.012157577089965343, + 0.004579135216772556, + 0.03627549856901169, + 0.007737981621176004, + 0.004197683185338974, + 0.007261908613145351, + 0.005060155410319567, + 0.003966551274061203, + 0.0075019910000264645, + 0.005502517335116863, + 0.009441963396966457, + 0.007238231133669615, + 0.0043246145360171795, + 0.006461923010647297, + 0.0038710914086550474, + 0.004355350509285927, + 0.006641853600740433, + 0.002908290596678853, + 0.005966829601675272, + 0.007287165615707636, + 0.0037862611934542656, + 0.005661229602992535, + 0.008549531921744347, + 0.0076531050726771355, + 0.009988156147301197, + 0.0041373539716005325, + 0.010970385745167732, + 0.007922635413706303, + 0.006835257168859243, + 0.003111886791884899, + 0.005693625193089247, + 0.0070191663689911366, + 0.030519889667630196, + 0.005723568145185709, + 0.026942212134599686 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 1, + 2, + 4 + ], + "selected_reward": -0.01822078227996826 + }, + { + "sample_idx": 0, + "question_id": "q3", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.06204604357481003 + ], + "opposite_logprob_tokens": [ + -2.8120460510253906 + ], + "logprob_reward": -0.06204604357481003, + "yes_no_margin_reward": 2.75, + "token_credit_reward": -0.06204604357481003, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.08253277838230133, + 0.16560417413711548, + 0.07572148740291595, + 0.042582057416439056, + 0.03227672725915909, + 0.1504940241575241, + 0.033847931772470474, + 0.07816265523433685, + 0.04079984501004219, + 0.29797831177711487 + ], + "credit_map": [ + 0.008289280347526073, + 0.01099433284252882, + 0.010598357766866684, + 0.011831920593976974, + 0.009036539122462273, + 0.011524766683578491, + 0.010543044656515121, + 0.00971453357487917, + 0.03965269774198532, + 0.021221816539764404, + 0.02800205536186695, + 0.014453689567744732, + 0.013262669555842876, + 0.018698202446103096, + 0.009037611074745655, + 0.021275432780385017, + 0.006689506117254496, + 0.007695491425693035, + 0.008656558580696583, + 0.007176707033067942, + 0.010084054432809353, + 0.014347132295370102, + 0.011231081560254097, + 0.009840954095125198, + 0.0039506335742771626, + 0.005938439629971981, + 0.004171703476458788, + 0.005813724361360073, + 0.005107380915433168, + 0.0041634696535766125, + 0.007588319946080446, + 0.005848383996635675, + 0.004391612485051155, + 0.0031389787327498198, + 0.005219590850174427, + 0.002507780445739627, + 0.0028661121614277363, + 0.007560212630778551, + 0.002138413256034255, + 0.004454026464372873, + 0.01552856620401144, + 0.006312268786132336, + 0.05051881819963455, + 0.008872272446751595, + 0.008883188478648663, + 0.015275170095264912, + 0.010517532005906105, + 0.03458619862794876, + 0.00446963869035244, + 0.003330828621983528, + 0.005083460360765457, + 0.002773558720946312, + 0.0039582750760018826, + 0.006224519107490778, + 0.0031821057200431824, + 0.004825548268854618, + 0.008305455558001995, + 0.005741690751165152, + 0.011588649824261665, + 0.007946905680000782, + 0.01704050973057747, + 0.008924790658056736, + 0.00785875041037798, + 0.010755900293588638, + 0.005247123539447784, + 0.0029465132392942905, + 0.004280005116015673, + 0.005994926206767559, + 0.004831729922443628, + 0.0060180905275046825, + 0.006116185337305069, + 0.005365272518247366, + 0.02070707641541958, + 0.01377088576555252, + 0.007995309308171272, + 0.009096936322748661, + 0.01857682503759861, + 0.1154559999704361, + 0.006489698775112629, + 0.10588555037975311 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 1, + 5 + ], + "selected_reward": -0.06204604357481003 + }, + { + "sample_idx": 0, + "question_id": "q4", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.6328450441360474 + ], + "opposite_logprob_tokens": [ + -0.7578450441360474 + ], + "logprob_reward": -0.6328450441360474, + "yes_no_margin_reward": 0.125, + "token_credit_reward": -0.6328450441360474, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.06551346927881241, + 0.06942658126354218, + 0.030395330861210823, + 0.033495042473077774, + 0.01822512224316597, + 0.044875942170619965, + 0.032170455902814865, + 0.12509185075759888, + 0.09492283314466476, + 0.48588332533836365 + ], + "credit_map": [ + 0.009318690747022629, + 0.009756619110703468, + 0.008217385970056057, + 0.008358444087207317, + 0.007847257889807224, + 0.007014485541731119, + 0.006608422379940748, + 0.008392171934247017, + 0.01063610054552555, + 0.007687087636440992, + 0.008373737335205078, + 0.006657006219029427, + 0.00790801364928484, + 0.011150022968649864, + 0.006223740987479687, + 0.010790880769491196, + 0.003596276743337512, + 0.00483957352116704, + 0.003738929983228445, + 0.0026380952913314104, + 0.00333641585893929, + 0.004951732698827982, + 0.0025759683921933174, + 0.004718341398984194, + 0.0031990897841751575, + 0.00416925922036171, + 0.0032015200704336166, + 0.0037614095490425825, + 0.0029900516383349895, + 0.006313373800367117, + 0.0057405573315918446, + 0.004119782708585262, + 0.0029429944697767496, + 0.0016498122131451964, + 0.00281965802423656, + 0.0014275700086727738, + 0.0017932138871401548, + 0.0034163238015025854, + 0.0013551468728110194, + 0.0028204061090946198, + 0.006968400441110134, + 0.0026551303453743458, + 0.011539440602064133, + 0.0032638360280543566, + 0.003227604553103447, + 0.0077133639715611935, + 0.0022131584119051695, + 0.007295014336705208, + 0.003938321489840746, + 0.0029902951791882515, + 0.004687297157943249, + 0.004827367141842842, + 0.0038793780840933323, + 0.0038131410256028175, + 0.0034876081626862288, + 0.004547052085399628, + 0.013514213263988495, + 0.008107651025056839, + 0.017372002825140953, + 0.01337369717657566, + 0.024103470146656036, + 0.012208153493702412, + 0.014830767177045345, + 0.02158191427588463, + 0.014428472146391869, + 0.006878761108964682, + 0.012463441118597984, + 0.013663748279213905, + 0.011566361412405968, + 0.01180967316031456, + 0.009009657427668571, + 0.015102721750736237, + 0.030735699459910393, + 0.02596205845475197, + 0.009064826183021069, + 0.014732414856553078, + 0.02822762355208397, + 0.1927798092365265, + 0.010440831072628498, + 0.17394009232521057 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 9, + 7, + 8 + ], + "selected_reward": -0.6328450441360474 + }, + { + "sample_idx": 0, + "question_id": "q5", + "question": "You are a careful video question-answering assistant.\nUse only visually grounded evidence from the video, and do not guess unseen details.\n\nGiven an AI generated video, answer the following atomic visual questions using only evidence from the video.\n\nRules:\n- Answer each question independently.\n- Use concise canonical phrases.\n- Do not add explanations.\n- For \"yes_no\" questions, answer with \"Yes\" or \"No\" only.\n- Return valid JSON only.\n\nOutput format:\n{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"...\"}]}\n\nSub-questions:\n- {\"id\": \"q1\", \"question\": \"Is the camera shot from above looking down at the scene?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q2\", \"question\": \"Is the environment a vast expanse of emerald green bamboo?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q3\", \"question\": \"Is the central figure a knight wearing a loose white silk robe?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q4\", \"question\": \"Does the figure move at full speed with swift, weightless movements?\", \"answer_type\": \"yes_no\"}\n- {\"id\": \"q5\", \"question\": \"Is there a strong backlight effect creating a near-black silhouette with a dazzling golden edge?\", \"answer_type\": \"yes_no\"}", + "ref_answer": "{\"sub_answers\": [{\"id\": \"q1\", \"answer\": \"Yes\"}, {\"id\": \"q2\", \"answer\": \"Yes\"}, {\"id\": \"q3\", \"answer\": \"Yes\"}, {\"id\": \"q4\", \"answer\": \"Yes\"}, {\"id\": \"q5\", \"answer\": \"Yes\"}]}", + "target_answer": "Yes", + "pred_sentence": "Yes", + "normalized_pred_keywords": [ + "Yes" + ], + "target_logprob_tokens": [ + -0.005255573429167271 + ], + "opposite_logprob_tokens": [ + -5.255255699157715 + ], + "logprob_reward": -0.005255573429167271, + "yes_no_margin_reward": 5.25, + "token_credit_reward": -0.005255573429167271, + "yes_ratio_reward": 1.0, + "frame_importance": [ + 0.23876458406448364, + 0.08664258569478989, + 0.11616583168506622, + 0.04375717043876648, + 0.061515435576438904, + 0.1449822634458542, + 0.06522927433252335, + 0.07794807851314545, + 0.09955088794231415, + 0.06544388085603714 + ], + "credit_map": [ + 0.019348207861185074, + 0.036659855395555496, + 0.034417275339365005, + 0.03328455612063408, + 0.024270126596093178, + 0.03754463791847229, + 0.024086328223347664, + 0.029153594747185707, + 0.014477125369012356, + 0.0077770608477294445, + 0.00975786242634058, + 0.007525150198489428, + 0.013753801584243774, + 0.012870264239609241, + 0.010309956036508083, + 0.010171365924179554, + 0.015051266178488731, + 0.013869772665202618, + 0.013241944834589958, + 0.009801465086638927, + 0.012272083200514317, + 0.024050762876868248, + 0.009287171997129917, + 0.01859135739505291, + 0.00414377311244607, + 0.006599953398108482, + 0.005576588679105043, + 0.0040066116489470005, + 0.0032859090715646744, + 0.006330851465463638, + 0.005667364224791527, + 0.008146118372678757, + 0.008190497756004333, + 0.006108622532337904, + 0.00836766604334116, + 0.00419241888448596, + 0.00704581942409277, + 0.010966976173222065, + 0.005970035679638386, + 0.010673396289348602, + 0.018914420157670975, + 0.006900552194565535, + 0.042046573013067245, + 0.011049413122236729, + 0.008344349451363087, + 0.014768274500966072, + 0.00809671450406313, + 0.034861963242292404, + 0.007922923192381859, + 0.00618883827701211, + 0.008944230154156685, + 0.005242190323770046, + 0.008484507910907269, + 0.012889510951936245, + 0.005446908995509148, + 0.010110152885317802, + 0.008868053555488586, + 0.006048816256225109, + 0.011012669652700424, + 0.0075515578500926495, + 0.013096358627080917, + 0.010456717573106289, + 0.007474842481315136, + 0.013439050875604153, + 0.012455751188099384, + 0.006663567386567593, + 0.0060647837817668915, + 0.00833740271627903, + 0.017372149974107742, + 0.018730055540800095, + 0.007146359421312809, + 0.02278081700205803, + 0.00891196634620428, + 0.0058695231564342976, + 0.005172635894268751, + 0.005827443674206734, + 0.006302686408162117, + 0.014636448584496975, + 0.004631382413208485, + 0.014091795310378075 + ], + "credit_map_shape": [ + 10, + 2, + 4 + ], + "top_frames": [ + 0, + 5, + 2 + ], + "selected_reward": -0.005255573429167271 + } + ], + "vlm_decoded_outputs": [ + "No", + "Yes", + "Yes", + "Yes", + "Yes" + ] + } +] \ No newline at end of file